2014-02-26 19:11:54 +01:00
|
|
|
/*****************************************************************************
|
|
|
|
|
|
2017-05-15 16:17:16 +02:00
|
|
|
Copyright (c) 1995, 2017, Oracle and/or its affiliates. All Rights Reserved.
|
2021-01-06 12:53:14 +01:00
|
|
|
Copyright (c) 2013, 2021, MariaDB Corporation.
|
2015-09-03 12:58:41 +02:00
|
|
|
Copyright (c) 2013, 2014, Fusion-io
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
This program is free software; you can redistribute it and/or modify it under
|
|
|
|
|
the terms of the GNU General Public License as published by the Free Software
|
|
|
|
|
Foundation; version 2 of the License.
|
|
|
|
|
|
|
|
|
|
This program is distributed in the hope that it will be useful, but WITHOUT
|
|
|
|
|
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
|
|
|
|
|
FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
|
|
|
|
|
|
|
|
|
|
You should have received a copy of the GNU General Public License along with
|
|
|
|
|
this program; if not, write to the Free Software Foundation, Inc.,
|
2019-05-11 18:25:02 +02:00
|
|
|
51 Franklin Street, Fifth Floor, Boston, MA 02110-1335 USA
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
*****************************************************************************/
|
|
|
|
|
|
|
|
|
|
/**************************************************//**
|
|
|
|
|
@file buf/buf0flu.cc
|
|
|
|
|
The database buffer buf_pool flush algorithm
|
|
|
|
|
|
|
|
|
|
Created 11/11/1995 Heikki Tuuri
|
|
|
|
|
*******************************************************/
|
|
|
|
|
|
2018-11-19 10:42:14 +01:00
|
|
|
#include "univ.i"
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
#include <my_service_manager.h>
|
2016-08-12 10:17:45 +02:00
|
|
|
#include <mysql/service_thd_wait.h>
|
2017-12-05 11:50:35 +01:00
|
|
|
#include <sql_class.h>
|
2016-08-12 10:17:45 +02:00
|
|
|
|
2014-02-26 19:11:54 +01:00
|
|
|
#include "buf0flu.h"
|
|
|
|
|
#include "buf0buf.h"
|
|
|
|
|
#include "buf0checksum.h"
|
2020-01-21 14:46:20 +01:00
|
|
|
#include "buf0dblwr.h"
|
2014-02-26 19:11:54 +01:00
|
|
|
#include "srv0start.h"
|
|
|
|
|
#include "page0zip.h"
|
|
|
|
|
#include "fil0fil.h"
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
#include "log0crypt.h"
|
2014-02-26 19:11:54 +01:00
|
|
|
#include "srv0mon.h"
|
2014-01-10 11:11:36 +01:00
|
|
|
#include "fil0pagecompress.h"
|
2016-08-12 10:17:45 +02:00
|
|
|
#ifdef UNIV_LINUX
|
|
|
|
|
/* include defs for CPU time priority settings */
|
|
|
|
|
#include <unistd.h>
|
|
|
|
|
#include <sys/syscall.h>
|
|
|
|
|
#include <sys/time.h>
|
|
|
|
|
#include <sys/resource.h>
|
|
|
|
|
#endif /* UNIV_LINUX */
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
#ifdef HAVE_LZO
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
# include "lzo/lzo1x.h"
|
|
|
|
|
#elif defined HAVE_SNAPPY
|
|
|
|
|
# include "snappy-c.h"
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
#endif
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
/** Number of pages flushed via LRU. Protected by buf_pool.mutex.
|
|
|
|
|
Also included in buf_flush_page_count. */
|
2019-07-03 16:31:20 +02:00
|
|
|
ulint buf_lru_flush_page_count;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
/** Number of pages flushed. Protected by buf_pool.mutex. */
|
|
|
|
|
ulint buf_flush_page_count;
|
|
|
|
|
|
|
|
|
|
/** Flag indicating if the page_cleaner is in active state. */
|
2017-06-12 16:43:07 +02:00
|
|
|
bool buf_page_cleaner_is_active;
|
2016-08-12 10:17:45 +02:00
|
|
|
|
|
|
|
|
/** Factor for scan length to determine n_pages for intended oldest LSN
|
|
|
|
|
progress */
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
static constexpr ulint buf_flush_lsn_scan_factor = 3;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2016-08-12 10:17:45 +02:00
|
|
|
/** Average redo generation rate */
|
|
|
|
|
static lsn_t lsn_avg_rate = 0;
|
|
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
/** Target oldest_modification for the page cleaner; writes are protected by
|
|
|
|
|
buf_pool.flush_list_mutex */
|
|
|
|
|
static Atomic_relaxed<lsn_t> buf_flush_sync_lsn;
|
2016-09-06 08:43:16 +02:00
|
|
|
|
2014-02-26 19:11:54 +01:00
|
|
|
#ifdef UNIV_PFS_THREAD
|
2016-08-12 10:17:45 +02:00
|
|
|
mysql_pfs_key_t page_cleaner_thread_key;
|
2014-02-26 19:11:54 +01:00
|
|
|
#endif /* UNIV_PFS_THREAD */
|
|
|
|
|
|
2020-02-12 13:45:21 +01:00
|
|
|
/** Page cleaner structure */
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
static struct
|
|
|
|
|
{
|
|
|
|
|
/** total elapsed time in adaptive flushing, in seconds */
|
|
|
|
|
ulint flush_time;
|
|
|
|
|
/** number of adaptive flushing passes */
|
|
|
|
|
ulint flush_pass;
|
|
|
|
|
} page_cleaner;
|
2014-11-06 12:17:11 +01:00
|
|
|
|
2016-09-06 08:43:16 +02:00
|
|
|
#ifdef UNIV_DEBUG
|
|
|
|
|
my_bool innodb_page_cleaner_disabled_debug;
|
|
|
|
|
#endif /* UNIV_DEBUG */
|
|
|
|
|
|
2014-02-26 19:11:54 +01:00
|
|
|
/** If LRU list of a buf_pool is less than this size then LRU eviction
|
|
|
|
|
should not happen. This is because when we do LRU flushing we also put
|
|
|
|
|
the blocks on free list. If LRU list is very small then we can end up
|
|
|
|
|
in thrashing. */
|
|
|
|
|
#define BUF_LRU_MIN_LEN 256
|
|
|
|
|
|
|
|
|
|
/* @} */
|
|
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
#ifdef UNIV_DEBUG
|
2020-02-12 13:45:21 +01:00
|
|
|
/** Validate the flush list. */
|
|
|
|
|
static void buf_flush_validate_low();
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-02-12 13:45:21 +01:00
|
|
|
/** Validates the flush list some of the time. */
|
|
|
|
|
static void buf_flush_validate_skip()
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
|
|
|
|
/** Try buf_flush_validate_low() every this many times */
|
|
|
|
|
# define BUF_FLUSH_VALIDATE_SKIP 23
|
|
|
|
|
|
|
|
|
|
/** The buf_flush_validate_low() call skip counter.
|
|
|
|
|
Use a signed type because of the race condition below. */
|
|
|
|
|
static int buf_flush_validate_count = BUF_FLUSH_VALIDATE_SKIP;
|
|
|
|
|
|
|
|
|
|
/* There is a race condition below, but it does not matter,
|
|
|
|
|
because this call is only for heuristic purposes. We want to
|
|
|
|
|
reduce the call frequency of the costly buf_flush_validate_low()
|
|
|
|
|
check in debug builds. */
|
|
|
|
|
if (--buf_flush_validate_count > 0) {
|
2020-02-12 13:45:21 +01:00
|
|
|
return;
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
buf_flush_validate_count = BUF_FLUSH_VALIDATE_SKIP;
|
2020-02-12 13:45:21 +01:00
|
|
|
buf_flush_validate_low();
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
#endif /* UNIV_DEBUG */
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-11-25 15:09:47 +01:00
|
|
|
/** Wake up the page cleaner if needed */
|
|
|
|
|
inline void buf_pool_t::page_cleaner_wakeup()
|
|
|
|
|
{
|
2021-01-06 12:53:14 +01:00
|
|
|
if (!page_cleaner_idle())
|
|
|
|
|
return;
|
|
|
|
|
double dirty_pct= double(UT_LIST_GET_LEN(buf_pool.flush_list)) * 100.0 /
|
|
|
|
|
double(UT_LIST_GET_LEN(buf_pool.LRU) + UT_LIST_GET_LEN(buf_pool.free));
|
|
|
|
|
double pct_lwm= srv_max_dirty_pages_pct_lwm;
|
|
|
|
|
if ((pct_lwm != 0.0 && pct_lwm <= dirty_pct) ||
|
|
|
|
|
srv_max_buf_pool_modified_pct <= dirty_pct)
|
2020-11-25 15:09:47 +01:00
|
|
|
{
|
|
|
|
|
page_cleaner_is_idle= false;
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_signal(&do_flush_list);
|
2020-11-25 15:09:47 +01:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-12 13:45:21 +01:00
|
|
|
/** Insert a modified block into the flush list.
|
|
|
|
|
@param[in,out] block modified block
|
|
|
|
|
@param[in] lsn oldest modification */
|
|
|
|
|
void buf_flush_insert_into_flush_list(buf_block_t* block, lsn_t lsn)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_assert_not_owner(&buf_pool.mutex);
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_assert_owner(&log_sys.flush_order_mutex);
|
2020-02-12 13:45:21 +01:00
|
|
|
ut_ad(lsn);
|
2020-12-09 08:22:13 +01:00
|
|
|
ut_ad(!fsp_is_system_temporary(block->page.id().space()));
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
block->page.set_oldest_modification(lsn);
|
2020-07-02 06:39:33 +02:00
|
|
|
MEM_CHECK_DEFINED(block->page.zip.data
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
? block->page.zip.data : block->frame,
|
|
|
|
|
block->physical_size());
|
|
|
|
|
buf_pool.stat.flush_list_bytes += block->physical_size();
|
|
|
|
|
ut_ad(buf_pool.stat.flush_list_bytes <= buf_pool.curr_pool_size);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-03-18 20:48:00 +01:00
|
|
|
UT_LIST_ADD_FIRST(buf_pool.flush_list, &block->page);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ut_d(buf_flush_validate_skip());
|
2020-11-25 15:09:47 +01:00
|
|
|
buf_pool.page_cleaner_wakeup();
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
2020-12-09 08:22:13 +01:00
|
|
|
/** Remove a block from buf_pool.flush_list */
|
|
|
|
|
static void buf_flush_remove_low(buf_page_t *bpage)
|
|
|
|
|
{
|
|
|
|
|
ut_ad(!fsp_is_system_temporary(bpage->id().space()));
|
|
|
|
|
mysql_mutex_assert_owner(&buf_pool.mutex);
|
|
|
|
|
mysql_mutex_assert_owner(&buf_pool.flush_list_mutex);
|
|
|
|
|
ut_ad(!bpage->oldest_modification());
|
|
|
|
|
buf_pool.flush_hp.adjust(bpage);
|
|
|
|
|
UT_LIST_REMOVE(buf_pool.flush_list, bpage);
|
|
|
|
|
buf_pool.stat.flush_list_bytes -= bpage->physical_size();
|
|
|
|
|
#ifdef UNIV_DEBUG
|
|
|
|
|
buf_flush_validate_skip();
|
|
|
|
|
#endif /* UNIV_DEBUG */
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-12 13:45:21 +01:00
|
|
|
/** Remove a block from the flush list of modified blocks.
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
@param[in,out] bpage block to be removed from the flush list */
|
|
|
|
|
static void buf_flush_remove(buf_page_t *bpage)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
2020-12-09 08:22:13 +01:00
|
|
|
bpage->clear_oldest_modification();
|
|
|
|
|
buf_flush_remove_low(bpage);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** Remove all dirty pages belonging to a given tablespace when we are
|
|
|
|
|
deleting the data file of that tablespace.
|
|
|
|
|
The pages still remain a part of LRU and are evicted from
|
|
|
|
|
the list as they age towards the tail of the LRU.
|
|
|
|
|
@param id tablespace identifier */
|
|
|
|
|
void buf_flush_remove_pages(ulint id)
|
|
|
|
|
{
|
|
|
|
|
const page_id_t first(id, 0), end(id + 1, 0);
|
|
|
|
|
ut_ad(id);
|
|
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
|
|
|
|
|
|
|
|
|
for (;;)
|
|
|
|
|
{
|
|
|
|
|
bool deferred= false;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
|
|
|
|
|
for (buf_page_t *bpage= UT_LIST_GET_LAST(buf_pool.flush_list); bpage; )
|
|
|
|
|
{
|
2020-10-26 16:07:17 +01:00
|
|
|
ut_d(const auto s= bpage->state());
|
|
|
|
|
ut_ad(s == BUF_BLOCK_ZIP_PAGE || s == BUF_BLOCK_FILE_PAGE ||
|
|
|
|
|
s == BUF_BLOCK_REMOVE_HASH);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
buf_page_t *prev= UT_LIST_GET_PREV(list, bpage);
|
|
|
|
|
|
|
|
|
|
const page_id_t bpage_id(bpage->id());
|
|
|
|
|
|
|
|
|
|
if (bpage_id < first || bpage_id >= end);
|
|
|
|
|
else if (bpage->io_fix() != BUF_IO_NONE)
|
|
|
|
|
deferred= true;
|
|
|
|
|
else
|
|
|
|
|
buf_flush_remove(bpage);
|
|
|
|
|
|
|
|
|
|
bpage= prev;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
|
|
|
|
|
if (!deferred)
|
|
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
|
|
|
|
os_thread_yield();
|
|
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
|
|
|
|
buf_flush_wait_batch_end(false);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** Try to flush all the dirty pages that belong to a given tablespace.
|
|
|
|
|
@param id tablespace identifier
|
|
|
|
|
@return number dirty pages that there were for this tablespace */
|
|
|
|
|
ulint buf_flush_dirty_pages(ulint id)
|
|
|
|
|
{
|
|
|
|
|
ulint n= 0;
|
|
|
|
|
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
|
|
|
|
|
for (buf_page_t *bpage= UT_LIST_GET_FIRST(buf_pool.flush_list); bpage;
|
|
|
|
|
bpage= UT_LIST_GET_NEXT(list, bpage))
|
|
|
|
|
{
|
2020-10-26 16:07:17 +01:00
|
|
|
ut_d(const auto s= bpage->state());
|
|
|
|
|
ut_ad(s == BUF_BLOCK_ZIP_PAGE || s == BUF_BLOCK_FILE_PAGE ||
|
|
|
|
|
s == BUF_BLOCK_REMOVE_HASH);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ut_ad(bpage->oldest_modification());
|
|
|
|
|
if (id == bpage->id().space())
|
|
|
|
|
n++;
|
|
|
|
|
}
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
if (n)
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
buf_flush_lists(srv_max_io_capacity, LSN_MAX);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
return n;
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*******************************************************************//**
|
|
|
|
|
Relocates a buffer control block on the flush_list.
|
|
|
|
|
Note that it is assumed that the contents of bpage have already been
|
|
|
|
|
copied to dpage.
|
|
|
|
|
IMPORTANT: When this function is called bpage and dpage are not
|
|
|
|
|
exact copies of each other. For example, they both will have different
|
|
|
|
|
::state. Also the ::list pointers in dpage may be stale. We need to
|
|
|
|
|
use the current list node (bpage) to do the list manipulation because
|
|
|
|
|
the list pointers could have changed between the time that we copied
|
|
|
|
|
the contents of bpage to the dpage and the flush list manipulation
|
|
|
|
|
below. */
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ATTRIBUTE_COLD
|
2014-02-26 19:11:54 +01:00
|
|
|
void
|
|
|
|
|
buf_flush_relocate_on_flush_list(
|
|
|
|
|
/*=============================*/
|
|
|
|
|
buf_page_t* bpage, /*!< in/out: control block being moved */
|
|
|
|
|
buf_page_t* dpage) /*!< in/out: destination block */
|
|
|
|
|
{
|
|
|
|
|
buf_page_t* prev;
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_assert_owner(&buf_pool.mutex);
|
2020-12-09 08:22:13 +01:00
|
|
|
ut_ad(!fsp_is_system_temporary(bpage->id().space()));
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
|
|
|
|
|
if (!bpage->oldest_modification()) {
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
/* FIXME: At this point we have both buf_pool and flush_list
|
|
|
|
|
mutexes. Theoretically removal of a block from flush list is
|
|
|
|
|
only covered by flush_list mutex but currently we do
|
|
|
|
|
have buf_pool mutex in buf_flush_remove() therefore this block
|
|
|
|
|
is guaranteed to be in the flush list. We need to check if
|
|
|
|
|
this will work without the assumption of block removing code
|
|
|
|
|
having the buf_pool mutex. */
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ut_ad(dpage->oldest_modification());
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2014-11-06 12:17:11 +01:00
|
|
|
/* Important that we adjust the hazard pointer before removing
|
|
|
|
|
the bpage from the flush list. */
|
2020-03-18 20:48:00 +01:00
|
|
|
buf_pool.flush_hp.adjust(bpage);
|
2014-11-06 12:17:11 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
bpage->clear_oldest_modification();
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
prev = UT_LIST_GET_PREV(list, bpage);
|
2020-03-18 20:48:00 +01:00
|
|
|
UT_LIST_REMOVE(buf_pool.flush_list, bpage);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
if (prev) {
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ut_ad(prev->oldest_modification());
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
UT_LIST_INSERT_AFTER(buf_pool.flush_list, prev, dpage);
|
2014-02-26 19:11:54 +01:00
|
|
|
} else {
|
2020-03-18 20:48:00 +01:00
|
|
|
UT_LIST_ADD_FIRST(buf_pool.flush_list, dpage);
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ut_d(buf_flush_validate_low());
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
/** Complete write of a file page from buf_pool.
|
2020-10-23 09:49:03 +02:00
|
|
|
@param request write request */
|
|
|
|
|
void buf_page_write_complete(const IORequest &request)
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
{
|
|
|
|
|
ut_ad(request.is_write());
|
2020-10-23 09:49:03 +02:00
|
|
|
ut_ad(!srv_read_only_mode/* ||
|
|
|
|
|
request.node->space->purpose == FIL_TYPE_TEMPORARY*/);
|
|
|
|
|
buf_page_t *bpage= request.bpage;
|
|
|
|
|
ut_ad(bpage);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ut_ad(bpage->in_file());
|
|
|
|
|
ut_ad(bpage->io_fix() == BUF_IO_WRITE);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ut_ad(!buf_dblwr.is_inside(bpage->id()));
|
2020-10-23 09:49:03 +02:00
|
|
|
bool dblwr;
|
|
|
|
|
if (bpage->status == buf_page_t::INIT_ON_FLUSH)
|
|
|
|
|
{
|
|
|
|
|
bpage->status= buf_page_t::NORMAL;
|
|
|
|
|
dblwr= false;
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
ut_ad(bpage->status == buf_page_t::NORMAL);
|
|
|
|
|
dblwr= request.node->space->use_doublewrite();
|
|
|
|
|
}
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
|
|
|
|
/* We do not need protect io_fix here by mutex to read it because
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
this and buf_page_read_complete() are the only functions where we can
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
change the value from BUF_IO_READ or BUF_IO_WRITE to some other
|
|
|
|
|
value, and our code ensures that this is the only thread that handles
|
|
|
|
|
the i/o for this block. */
|
|
|
|
|
if (bpage->slot)
|
|
|
|
|
{
|
|
|
|
|
bpage->slot->release();
|
|
|
|
|
bpage->slot= nullptr;
|
|
|
|
|
}
|
2017-02-17 09:32:21 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
if (UNIV_UNLIKELY(MONITOR_IS_ON(MONITOR_MODULE_BUF_PAGE)))
|
|
|
|
|
buf_page_monitor(bpage, BUF_IO_WRITE);
|
|
|
|
|
DBUG_PRINT("ib_buf", ("write page %u:%u",
|
|
|
|
|
bpage->id().space(), bpage->id().page_no()));
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ut_ad(request.is_LRU() ? buf_pool.n_flush_LRU : buf_pool.n_flush_list);
|
2020-12-09 08:22:13 +01:00
|
|
|
const bool temp= fsp_is_system_temporary(bpage->id().space());
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
|
|
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
bpage->set_io_fix(BUF_IO_NONE);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
2020-12-09 08:22:13 +01:00
|
|
|
ut_ad(!temp || bpage->oldest_modification() == 1);
|
|
|
|
|
bpage->clear_oldest_modification();
|
|
|
|
|
|
|
|
|
|
if (!temp)
|
|
|
|
|
buf_flush_remove_low(bpage);
|
|
|
|
|
else
|
|
|
|
|
ut_ad(request.is_LRU());
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
|
|
|
|
|
if (dblwr)
|
|
|
|
|
{
|
|
|
|
|
ut_ad(!fsp_is_system_temporary(bpage->id().space()));
|
|
|
|
|
buf_dblwr.write_completed();
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
if (bpage->state() == BUF_BLOCK_FILE_PAGE)
|
MDEV-24142: Replace InnoDB rw_lock_t with sux_lock
InnoDB buffer pool block and index tree latches depend on a
special kind of read-update-write lock that allows reentrant
(recursive) acquisition of the 'update' and 'write' locks
as well as an upgrade from 'update' lock to 'write' lock.
The 'update' lock allows any number of reader locks from
other threads, but no concurrent 'update' or 'write' lock.
If there were no requirement to support an upgrade from 'update'
to 'write', we could compose the lock out of two srw_lock
(implemented as any type of native rw-lock, such as SRWLOCK on
Microsoft Windows). Removing this requirement is very difficult,
so in commit f7e7f487d4b06695f91f6fbeb0396b9d87fc7bbf we
implemented an 'update' mode to our srw_lock.
Re-entrant or recursive locking is mostly needed when writing or
freeing BLOB pages, but also in crash recovery or when merging
buffered changes to an index page. The re-entrancy allows us to
attach a previously acquired page to a sub-mini-transaction that
will be committed before whatever else is holding the page latch.
The SUX lock supports Shared ('read'), Update, and eXclusive ('write')
locking modes. The S latches are not re-entrant, but a single S latch
may be acquired even if the thread already holds an U latch.
The idea of the U latch is to allow a write of something that concurrent
readers do not care about (such as the contents of BTR_SEG_LEAF,
BTR_SEG_TOP and other page allocation metadata structures, or
the MDEV-6076 PAGE_ROOT_AUTO_INC). (The PAGE_ROOT_AUTO_INC field
is only updated when a dict_table_t for the table exists, and only
read when a dict_table_t for the table is being added to dict_sys.)
block_lock::u_lock_try(bool for_io=true) is used in buf_flush_page()
to allow concurrent readers but no concurrent modifications while the
page is being written to the data file. That latch will be released
by buf_page_write_complete() in a different thread. Hence, we use
the special lock owner value FOR_IO.
The index_lock::u_lock() improves concurrency on operations that
involve non-leaf index pages.
The interface has been cleaned up a little. We will use
x_lock_recursive() instead of x_lock() when we know that a
lock is already held by the current thread. Similarly,
a lock upgrade from U to X is only allowed via u_x_upgrade()
or x_lock_upgraded() but not via x_lock().
We will disable the LatchDebug and sync_array interfaces to
InnoDB rw-locks.
The SEMAPHORES section of SHOW ENGINE INNODB STATUS output
will no longer include any information about InnoDB rw-locks,
only TTASEventMutex (cmake -DMUTEXTYPE=event) waits.
This will make a part of the 'innotop' script dead code.
The block_lock buf_block_t::lock will not be covered by any
PERFORMANCE_SCHEMA instrumentation.
SHOW ENGINE INNODB MUTEX and INFORMATION_SCHEMA.INNODB_MUTEXES
will no longer output source code file names or line numbers.
The dict_index_t::lock will be identified by index and table names,
which should be much more useful. PERFORMANCE_SCHEMA is lumping
information about all dict_index_t::lock together as
event_name='wait/synch/sxlock/innodb/index_tree_rw_lock'.
buf_page_free(): Remove the file,line parameters. The sux_lock will
not store such diagnostic information.
buf_block_dbg_add_level(): Define as empty macro, to be removed
in a subsequent commit.
Unless the build was configured with cmake -DPLUGIN_PERFSCHEMA=NO
the index_lock dict_index_t::lock will be instrumented via
PERFORMANCE_SCHEMA. Similar to
commit 1669c8890ca2e9092213626e5b047e58ca8b1e77
we will distinguish lock waits by registering shared_lock,exclusive_lock
events instead of try_shared_lock,try_exclusive_lock.
Actual 'try' operations will not be instrumented at all.
rw_lock_list: Remove. After MDEV-24167, this only covered
buf_block_t::lock and dict_index_t::lock. We will output their
information by traversing buf_pool or dict_sys.
2020-12-03 14:18:51 +01:00
|
|
|
reinterpret_cast<buf_block_t*>(bpage)->lock.u_unlock(true);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
buf_pool.stat.n_pages_written++;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (request.is_LRU())
|
|
|
|
|
{
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
buf_LRU_free_page(bpage, true);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (!--buf_pool.n_flush_LRU)
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_broadcast(&buf_pool.done_flush_LRU);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
if (!--buf_pool.n_flush_list)
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_broadcast(&buf_pool.done_flush_list);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
}
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
/** Calculate a ROW_FORMAT=COMPRESSED page checksum and update the page.
|
MDEV-18644: Support full_crc32 for page_compressed
This is a follow-up task to MDEV-12026, which introduced
innodb_checksum_algorithm=full_crc32 and a simpler page format.
MDEV-12026 did not enable full_crc32 for page_compressed tables,
which we will be doing now.
This is joint work with Thirunarayanan Balathandayuthapani.
For innodb_checksum_algorithm=full_crc32 we change the
page_compressed format as follows:
FIL_PAGE_TYPE: The most significant bit will be set to indicate
page_compressed format. The least significant bits will contain
the compressed page size, rounded up to a multiple of 256 bytes.
The checksum will be stored in the last 4 bytes of the page
(whether it is the full page or a page_compressed page whose
size is determined by FIL_PAGE_TYPE), covering all preceding
bytes of the page. If encryption is used, then the page will
be encrypted between compression and computing the checksum.
For page_compressed, FIL_PAGE_LSN will not be repeated at
the end of the page.
FSP_SPACE_FLAGS (already implemented as part of MDEV-12026):
We will store the innodb_compression_algorithm that may be used
to compress pages. Previously, the choice of algorithm was written
to each compressed data page separately, and one would be unable
to know in advance which compression algorithm(s) are used.
fil_space_t::full_crc32_page_compressed_len(): Determine if the
page_compressed algorithm of the tablespace needs to know the
exact length of the compressed data. If yes, we will reserve and
write an extra byte for this right before the checksum.
buf_page_is_compressed(): Determine if a page uses page_compressed
(in any innodb_checksum_algorithm).
fil_page_decompress(): Pass also fil_space_t::flags so that the
format can be determined.
buf_page_is_zeroes(): Check if a page is full of zero bytes.
buf_page_full_crc32_is_corrupted(): Renamed from
buf_encrypted_full_crc32_page_is_corrupted(). For full_crc32,
we always simply validate the checksum to the page contents,
while the physical page size is explicitly specified by an
unencrypted part of the page header.
buf_page_full_crc32_size(): Determine the size of a full_crc32 page.
buf_dblwr_check_page_lsn(): Make this a debug-only function, because
it involves potentially costly lookups of fil_space_t.
create_table_info_t::check_table_options(),
ha_innobase::check_if_supported_inplace_alter(): Do allow the creation
of SPATIAL INDEX with full_crc32 also when page_compressed is used.
commit_cache_norebuild(): Preserve the compression algorithm when
updating the page_compression_level.
dict_tf_to_fsp_flags(): Set the flags for page compression algorithm.
FIXME: Maybe there should be a table option page_compression_algorithm
and a session variable to back it?
2019-03-18 13:08:43 +01:00
|
|
|
@param[in,out] page page to update
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
@param[in] size compressed page size */
|
|
|
|
|
void buf_flush_update_zip_checksum(buf_frame_t *page, ulint size)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
ut_ad(size > 0);
|
|
|
|
|
mach_write_to_4(page + FIL_PAGE_SPACE_OR_CHKSUM,
|
|
|
|
|
page_zip_calc_checksum(page, size,
|
|
|
|
|
static_cast<srv_checksum_algorithm_t>
|
|
|
|
|
(srv_checksum_algorithm)));
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
MDEV-12026: Implement innodb_checksum_algorithm=full_crc32
MariaDB data-at-rest encryption (innodb_encrypt_tables)
had repurposed the same unused data field that was repurposed
in MySQL 5.7 (and MariaDB 10.2) for the Split Sequence Number (SSN)
field of SPATIAL INDEX. Because of this, MariaDB was unable to
support encryption on SPATIAL INDEX pages.
Furthermore, InnoDB page checksums skipped some bytes, and there
are multiple variations and checksum algorithms. By default,
InnoDB accepts all variations of all algorithms that ever existed.
This unnecessarily weakens the page checksums.
We hereby introduce two more innodb_checksum_algorithm variants
(full_crc32, strict_full_crc32) that are special in a way:
When either setting is active, newly created data files will
carry a flag (fil_space_t::full_crc32()) that indicates that
all pages of the file will use a full CRC-32C checksum over the
entire page contents (excluding the bytes where the checksum
is stored, at the very end of the page). Such files will always
use that checksum, no matter what the parameter
innodb_checksum_algorithm is assigned to.
For old files, the old checksum algorithms will continue to be
used. The value strict_full_crc32 will be equivalent to strict_crc32
and the value full_crc32 will be equivalent to crc32.
ROW_FORMAT=COMPRESSED tables will only use the old format.
These tables do not support new features, such as larger
innodb_page_size or instant ADD/DROP COLUMN. They may be
deprecated in the future. We do not want an unnecessary
file format change for them.
The new full_crc32() format also cleans up the MariaDB tablespace
flags. We will reserve flags to store the page_compressed
compression algorithm, and to store the compressed payload length,
so that checksum can be computed over the compressed (and
possibly encrypted) stream and can be validated without
decrypting or decompressing the page.
In the full_crc32 format, there no longer are separate before-encryption
and after-encryption checksums for pages. The single checksum is
computed on the page contents that is written to the file.
We do not make the new algorithm the default for two reasons.
First, MariaDB 10.4.2 was a beta release, and the default values
of parameters should not change after beta. Second, we did not
yet implement the full_crc32 format for page_compressed pages.
This will be fixed in MDEV-18644.
This is joint work with Marko Mäkelä.
2019-02-19 20:00:00 +01:00
|
|
|
/** Assign the full crc32 checksum for non-compressed page.
|
|
|
|
|
@param[in,out] page page to be updated */
|
|
|
|
|
void buf_flush_assign_full_crc32_checksum(byte* page)
|
|
|
|
|
{
|
MDEV-18644: Support full_crc32 for page_compressed
This is a follow-up task to MDEV-12026, which introduced
innodb_checksum_algorithm=full_crc32 and a simpler page format.
MDEV-12026 did not enable full_crc32 for page_compressed tables,
which we will be doing now.
This is joint work with Thirunarayanan Balathandayuthapani.
For innodb_checksum_algorithm=full_crc32 we change the
page_compressed format as follows:
FIL_PAGE_TYPE: The most significant bit will be set to indicate
page_compressed format. The least significant bits will contain
the compressed page size, rounded up to a multiple of 256 bytes.
The checksum will be stored in the last 4 bytes of the page
(whether it is the full page or a page_compressed page whose
size is determined by FIL_PAGE_TYPE), covering all preceding
bytes of the page. If encryption is used, then the page will
be encrypted between compression and computing the checksum.
For page_compressed, FIL_PAGE_LSN will not be repeated at
the end of the page.
FSP_SPACE_FLAGS (already implemented as part of MDEV-12026):
We will store the innodb_compression_algorithm that may be used
to compress pages. Previously, the choice of algorithm was written
to each compressed data page separately, and one would be unable
to know in advance which compression algorithm(s) are used.
fil_space_t::full_crc32_page_compressed_len(): Determine if the
page_compressed algorithm of the tablespace needs to know the
exact length of the compressed data. If yes, we will reserve and
write an extra byte for this right before the checksum.
buf_page_is_compressed(): Determine if a page uses page_compressed
(in any innodb_checksum_algorithm).
fil_page_decompress(): Pass also fil_space_t::flags so that the
format can be determined.
buf_page_is_zeroes(): Check if a page is full of zero bytes.
buf_page_full_crc32_is_corrupted(): Renamed from
buf_encrypted_full_crc32_page_is_corrupted(). For full_crc32,
we always simply validate the checksum to the page contents,
while the physical page size is explicitly specified by an
unencrypted part of the page header.
buf_page_full_crc32_size(): Determine the size of a full_crc32 page.
buf_dblwr_check_page_lsn(): Make this a debug-only function, because
it involves potentially costly lookups of fil_space_t.
create_table_info_t::check_table_options(),
ha_innobase::check_if_supported_inplace_alter(): Do allow the creation
of SPATIAL INDEX with full_crc32 also when page_compressed is used.
commit_cache_norebuild(): Preserve the compression algorithm when
updating the page_compression_level.
dict_tf_to_fsp_flags(): Set the flags for page compression algorithm.
FIXME: Maybe there should be a table option page_compression_algorithm
and a session variable to back it?
2019-03-18 13:08:43 +01:00
|
|
|
ut_d(bool compressed = false);
|
|
|
|
|
ut_d(bool corrupted = false);
|
|
|
|
|
ut_d(const uint size = buf_page_full_crc32_size(page, &compressed,
|
|
|
|
|
&corrupted));
|
|
|
|
|
ut_ad(!compressed);
|
|
|
|
|
ut_ad(!corrupted);
|
|
|
|
|
ut_ad(size == uint(srv_page_size));
|
|
|
|
|
const ulint payload = srv_page_size - FIL_PAGE_FCRC32_CHECKSUM;
|
|
|
|
|
mach_write_to_4(page + payload, ut_crc32(page, payload));
|
MDEV-12026: Implement innodb_checksum_algorithm=full_crc32
MariaDB data-at-rest encryption (innodb_encrypt_tables)
had repurposed the same unused data field that was repurposed
in MySQL 5.7 (and MariaDB 10.2) for the Split Sequence Number (SSN)
field of SPATIAL INDEX. Because of this, MariaDB was unable to
support encryption on SPATIAL INDEX pages.
Furthermore, InnoDB page checksums skipped some bytes, and there
are multiple variations and checksum algorithms. By default,
InnoDB accepts all variations of all algorithms that ever existed.
This unnecessarily weakens the page checksums.
We hereby introduce two more innodb_checksum_algorithm variants
(full_crc32, strict_full_crc32) that are special in a way:
When either setting is active, newly created data files will
carry a flag (fil_space_t::full_crc32()) that indicates that
all pages of the file will use a full CRC-32C checksum over the
entire page contents (excluding the bytes where the checksum
is stored, at the very end of the page). Such files will always
use that checksum, no matter what the parameter
innodb_checksum_algorithm is assigned to.
For old files, the old checksum algorithms will continue to be
used. The value strict_full_crc32 will be equivalent to strict_crc32
and the value full_crc32 will be equivalent to crc32.
ROW_FORMAT=COMPRESSED tables will only use the old format.
These tables do not support new features, such as larger
innodb_page_size or instant ADD/DROP COLUMN. They may be
deprecated in the future. We do not want an unnecessary
file format change for them.
The new full_crc32() format also cleans up the MariaDB tablespace
flags. We will reserve flags to store the page_compressed
compression algorithm, and to store the compressed payload length,
so that checksum can be computed over the compressed (and
possibly encrypted) stream and can be validated without
decrypting or decompressing the page.
In the full_crc32 format, there no longer are separate before-encryption
and after-encryption checksums for pages. The single checksum is
computed on the page contents that is written to the file.
We do not make the new algorithm the default for two reasons.
First, MariaDB 10.4.2 was a beta release, and the default values
of parameters should not change after beta. Second, we did not
yet implement the full_crc32 format for page_compressed pages.
This will be fixed in MDEV-18644.
This is joint work with Marko Mäkelä.
2019-02-19 20:00:00 +01:00
|
|
|
}
|
|
|
|
|
|
2016-08-12 10:17:45 +02:00
|
|
|
/** Initialize a page for writing to the tablespace.
|
MDEV-12026: Implement innodb_checksum_algorithm=full_crc32
MariaDB data-at-rest encryption (innodb_encrypt_tables)
had repurposed the same unused data field that was repurposed
in MySQL 5.7 (and MariaDB 10.2) for the Split Sequence Number (SSN)
field of SPATIAL INDEX. Because of this, MariaDB was unable to
support encryption on SPATIAL INDEX pages.
Furthermore, InnoDB page checksums skipped some bytes, and there
are multiple variations and checksum algorithms. By default,
InnoDB accepts all variations of all algorithms that ever existed.
This unnecessarily weakens the page checksums.
We hereby introduce two more innodb_checksum_algorithm variants
(full_crc32, strict_full_crc32) that are special in a way:
When either setting is active, newly created data files will
carry a flag (fil_space_t::full_crc32()) that indicates that
all pages of the file will use a full CRC-32C checksum over the
entire page contents (excluding the bytes where the checksum
is stored, at the very end of the page). Such files will always
use that checksum, no matter what the parameter
innodb_checksum_algorithm is assigned to.
For old files, the old checksum algorithms will continue to be
used. The value strict_full_crc32 will be equivalent to strict_crc32
and the value full_crc32 will be equivalent to crc32.
ROW_FORMAT=COMPRESSED tables will only use the old format.
These tables do not support new features, such as larger
innodb_page_size or instant ADD/DROP COLUMN. They may be
deprecated in the future. We do not want an unnecessary
file format change for them.
The new full_crc32() format also cleans up the MariaDB tablespace
flags. We will reserve flags to store the page_compressed
compression algorithm, and to store the compressed payload length,
so that checksum can be computed over the compressed (and
possibly encrypted) stream and can be validated without
decrypting or decompressing the page.
In the full_crc32 format, there no longer are separate before-encryption
and after-encryption checksums for pages. The single checksum is
computed on the page contents that is written to the file.
We do not make the new algorithm the default for two reasons.
First, MariaDB 10.4.2 was a beta release, and the default values
of parameters should not change after beta. Second, we did not
yet implement the full_crc32 format for page_compressed pages.
This will be fixed in MDEV-18644.
This is joint work with Marko Mäkelä.
2019-02-19 20:00:00 +01:00
|
|
|
@param[in] block buffer block; NULL if bypassing
|
|
|
|
|
the buffer pool
|
|
|
|
|
@param[in,out] page page frame
|
|
|
|
|
@param[in,out] page_zip_ compressed page, or NULL if
|
|
|
|
|
uncompressed
|
|
|
|
|
@param[in] use_full_checksum whether tablespace uses full checksum */
|
2014-02-26 19:11:54 +01:00
|
|
|
void
|
|
|
|
|
buf_flush_init_for_writing(
|
2016-08-12 10:17:45 +02:00
|
|
|
const buf_block_t* block,
|
|
|
|
|
byte* page,
|
|
|
|
|
void* page_zip_,
|
MDEV-12026: Implement innodb_checksum_algorithm=full_crc32
MariaDB data-at-rest encryption (innodb_encrypt_tables)
had repurposed the same unused data field that was repurposed
in MySQL 5.7 (and MariaDB 10.2) for the Split Sequence Number (SSN)
field of SPATIAL INDEX. Because of this, MariaDB was unable to
support encryption on SPATIAL INDEX pages.
Furthermore, InnoDB page checksums skipped some bytes, and there
are multiple variations and checksum algorithms. By default,
InnoDB accepts all variations of all algorithms that ever existed.
This unnecessarily weakens the page checksums.
We hereby introduce two more innodb_checksum_algorithm variants
(full_crc32, strict_full_crc32) that are special in a way:
When either setting is active, newly created data files will
carry a flag (fil_space_t::full_crc32()) that indicates that
all pages of the file will use a full CRC-32C checksum over the
entire page contents (excluding the bytes where the checksum
is stored, at the very end of the page). Such files will always
use that checksum, no matter what the parameter
innodb_checksum_algorithm is assigned to.
For old files, the old checksum algorithms will continue to be
used. The value strict_full_crc32 will be equivalent to strict_crc32
and the value full_crc32 will be equivalent to crc32.
ROW_FORMAT=COMPRESSED tables will only use the old format.
These tables do not support new features, such as larger
innodb_page_size or instant ADD/DROP COLUMN. They may be
deprecated in the future. We do not want an unnecessary
file format change for them.
The new full_crc32() format also cleans up the MariaDB tablespace
flags. We will reserve flags to store the page_compressed
compression algorithm, and to store the compressed payload length,
so that checksum can be computed over the compressed (and
possibly encrypted) stream and can be validated without
decrypting or decompressing the page.
In the full_crc32 format, there no longer are separate before-encryption
and after-encryption checksums for pages. The single checksum is
computed on the page contents that is written to the file.
We do not make the new algorithm the default for two reasons.
First, MariaDB 10.4.2 was a beta release, and the default values
of parameters should not change after beta. Second, we did not
yet implement the full_crc32 format for page_compressed pages.
This will be fixed in MDEV-18644.
This is joint work with Marko Mäkelä.
2019-02-19 20:00:00 +01:00
|
|
|
bool use_full_checksum)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-12026: Implement innodb_checksum_algorithm=full_crc32
MariaDB data-at-rest encryption (innodb_encrypt_tables)
had repurposed the same unused data field that was repurposed
in MySQL 5.7 (and MariaDB 10.2) for the Split Sequence Number (SSN)
field of SPATIAL INDEX. Because of this, MariaDB was unable to
support encryption on SPATIAL INDEX pages.
Furthermore, InnoDB page checksums skipped some bytes, and there
are multiple variations and checksum algorithms. By default,
InnoDB accepts all variations of all algorithms that ever existed.
This unnecessarily weakens the page checksums.
We hereby introduce two more innodb_checksum_algorithm variants
(full_crc32, strict_full_crc32) that are special in a way:
When either setting is active, newly created data files will
carry a flag (fil_space_t::full_crc32()) that indicates that
all pages of the file will use a full CRC-32C checksum over the
entire page contents (excluding the bytes where the checksum
is stored, at the very end of the page). Such files will always
use that checksum, no matter what the parameter
innodb_checksum_algorithm is assigned to.
For old files, the old checksum algorithms will continue to be
used. The value strict_full_crc32 will be equivalent to strict_crc32
and the value full_crc32 will be equivalent to crc32.
ROW_FORMAT=COMPRESSED tables will only use the old format.
These tables do not support new features, such as larger
innodb_page_size or instant ADD/DROP COLUMN. They may be
deprecated in the future. We do not want an unnecessary
file format change for them.
The new full_crc32() format also cleans up the MariaDB tablespace
flags. We will reserve flags to store the page_compressed
compression algorithm, and to store the compressed payload length,
so that checksum can be computed over the compressed (and
possibly encrypted) stream and can be validated without
decrypting or decompressing the page.
In the full_crc32 format, there no longer are separate before-encryption
and after-encryption checksums for pages. The single checksum is
computed on the page contents that is written to the file.
We do not make the new algorithm the default for two reasons.
First, MariaDB 10.4.2 was a beta release, and the default values
of parameters should not change after beta. Second, we did not
yet implement the full_crc32 format for page_compressed pages.
This will be fixed in MDEV-18644.
This is joint work with Marko Mäkelä.
2019-02-19 20:00:00 +01:00
|
|
|
if (block != NULL && block->frame != page) {
|
|
|
|
|
/* If page is encrypted in full crc32 format then
|
|
|
|
|
checksum stored already as a part of fil_encrypt_buf() */
|
|
|
|
|
ut_ad(use_full_checksum);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
2016-08-12 10:17:45 +02:00
|
|
|
ut_ad(block == NULL || block->frame == page);
|
|
|
|
|
ut_ad(block == NULL || page_zip_ == NULL
|
|
|
|
|
|| &block->page.zip == page_zip_);
|
2014-02-26 19:11:54 +01:00
|
|
|
ut_ad(page);
|
|
|
|
|
|
|
|
|
|
if (page_zip_) {
|
|
|
|
|
page_zip_des_t* page_zip;
|
2016-08-12 10:17:45 +02:00
|
|
|
ulint size;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
page_zip = static_cast<page_zip_des_t*>(page_zip_);
|
2016-08-12 10:17:45 +02:00
|
|
|
size = page_zip_get_size(page_zip);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2016-08-12 10:17:45 +02:00
|
|
|
ut_ad(size);
|
|
|
|
|
ut_ad(ut_is_2pow(size));
|
|
|
|
|
ut_ad(size <= UNIV_ZIP_SIZE_MAX);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2016-08-12 10:17:45 +02:00
|
|
|
switch (fil_page_get_type(page)) {
|
2014-02-26 19:11:54 +01:00
|
|
|
case FIL_PAGE_TYPE_ALLOCATED:
|
|
|
|
|
case FIL_PAGE_INODE:
|
|
|
|
|
case FIL_PAGE_IBUF_BITMAP:
|
|
|
|
|
case FIL_PAGE_TYPE_FSP_HDR:
|
|
|
|
|
case FIL_PAGE_TYPE_XDES:
|
|
|
|
|
/* These are essentially uncompressed pages. */
|
2016-08-12 10:17:45 +02:00
|
|
|
memcpy(page_zip->data, page, size);
|
2014-02-26 19:11:54 +01:00
|
|
|
/* fall through */
|
|
|
|
|
case FIL_PAGE_TYPE_ZBLOB:
|
|
|
|
|
case FIL_PAGE_TYPE_ZBLOB2:
|
|
|
|
|
case FIL_PAGE_INDEX:
|
2016-08-12 10:17:45 +02:00
|
|
|
case FIL_PAGE_RTREE:
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
buf_flush_update_zip_checksum(page_zip->data, size);
|
2014-02-26 19:11:54 +01:00
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
2016-08-12 10:17:45 +02:00
|
|
|
ib::error() << "The compressed page to be written"
|
|
|
|
|
" seems corrupt:";
|
|
|
|
|
ut_print_buf(stderr, page, size);
|
2014-02-26 19:11:54 +01:00
|
|
|
fputs("\nInnoDB: Possibly older version of the page:", stderr);
|
2016-08-12 10:17:45 +02:00
|
|
|
ut_print_buf(stderr, page_zip->data, size);
|
2014-02-26 19:11:54 +01:00
|
|
|
putc('\n', stderr);
|
|
|
|
|
ut_error;
|
|
|
|
|
}
|
|
|
|
|
|
MDEV-12026: Implement innodb_checksum_algorithm=full_crc32
MariaDB data-at-rest encryption (innodb_encrypt_tables)
had repurposed the same unused data field that was repurposed
in MySQL 5.7 (and MariaDB 10.2) for the Split Sequence Number (SSN)
field of SPATIAL INDEX. Because of this, MariaDB was unable to
support encryption on SPATIAL INDEX pages.
Furthermore, InnoDB page checksums skipped some bytes, and there
are multiple variations and checksum algorithms. By default,
InnoDB accepts all variations of all algorithms that ever existed.
This unnecessarily weakens the page checksums.
We hereby introduce two more innodb_checksum_algorithm variants
(full_crc32, strict_full_crc32) that are special in a way:
When either setting is active, newly created data files will
carry a flag (fil_space_t::full_crc32()) that indicates that
all pages of the file will use a full CRC-32C checksum over the
entire page contents (excluding the bytes where the checksum
is stored, at the very end of the page). Such files will always
use that checksum, no matter what the parameter
innodb_checksum_algorithm is assigned to.
For old files, the old checksum algorithms will continue to be
used. The value strict_full_crc32 will be equivalent to strict_crc32
and the value full_crc32 will be equivalent to crc32.
ROW_FORMAT=COMPRESSED tables will only use the old format.
These tables do not support new features, such as larger
innodb_page_size or instant ADD/DROP COLUMN. They may be
deprecated in the future. We do not want an unnecessary
file format change for them.
The new full_crc32() format also cleans up the MariaDB tablespace
flags. We will reserve flags to store the page_compressed
compression algorithm, and to store the compressed payload length,
so that checksum can be computed over the compressed (and
possibly encrypted) stream and can be validated without
decrypting or decompressing the page.
In the full_crc32 format, there no longer are separate before-encryption
and after-encryption checksums for pages. The single checksum is
computed on the page contents that is written to the file.
We do not make the new algorithm the default for two reasons.
First, MariaDB 10.4.2 was a beta release, and the default values
of parameters should not change after beta. Second, we did not
yet implement the full_crc32 format for page_compressed pages.
This will be fixed in MDEV-18644.
This is joint work with Marko Mäkelä.
2019-02-19 20:00:00 +01:00
|
|
|
if (use_full_checksum) {
|
2019-11-26 09:14:07 +01:00
|
|
|
static_assert(FIL_PAGE_FCRC32_END_LSN % 4 == 0, "aligned");
|
|
|
|
|
static_assert(FIL_PAGE_LSN % 4 == 0, "aligned");
|
|
|
|
|
memcpy_aligned<4>(page + srv_page_size
|
|
|
|
|
- FIL_PAGE_FCRC32_END_LSN,
|
|
|
|
|
FIL_PAGE_LSN + 4 + page, 4);
|
2019-12-30 17:59:02 +01:00
|
|
|
return buf_flush_assign_full_crc32_checksum(page);
|
MDEV-12026: Implement innodb_checksum_algorithm=full_crc32
MariaDB data-at-rest encryption (innodb_encrypt_tables)
had repurposed the same unused data field that was repurposed
in MySQL 5.7 (and MariaDB 10.2) for the Split Sequence Number (SSN)
field of SPATIAL INDEX. Because of this, MariaDB was unable to
support encryption on SPATIAL INDEX pages.
Furthermore, InnoDB page checksums skipped some bytes, and there
are multiple variations and checksum algorithms. By default,
InnoDB accepts all variations of all algorithms that ever existed.
This unnecessarily weakens the page checksums.
We hereby introduce two more innodb_checksum_algorithm variants
(full_crc32, strict_full_crc32) that are special in a way:
When either setting is active, newly created data files will
carry a flag (fil_space_t::full_crc32()) that indicates that
all pages of the file will use a full CRC-32C checksum over the
entire page contents (excluding the bytes where the checksum
is stored, at the very end of the page). Such files will always
use that checksum, no matter what the parameter
innodb_checksum_algorithm is assigned to.
For old files, the old checksum algorithms will continue to be
used. The value strict_full_crc32 will be equivalent to strict_crc32
and the value full_crc32 will be equivalent to crc32.
ROW_FORMAT=COMPRESSED tables will only use the old format.
These tables do not support new features, such as larger
innodb_page_size or instant ADD/DROP COLUMN. They may be
deprecated in the future. We do not want an unnecessary
file format change for them.
The new full_crc32() format also cleans up the MariaDB tablespace
flags. We will reserve flags to store the page_compressed
compression algorithm, and to store the compressed payload length,
so that checksum can be computed over the compressed (and
possibly encrypted) stream and can be validated without
decrypting or decompressing the page.
In the full_crc32 format, there no longer are separate before-encryption
and after-encryption checksums for pages. The single checksum is
computed on the page contents that is written to the file.
We do not make the new algorithm the default for two reasons.
First, MariaDB 10.4.2 was a beta release, and the default values
of parameters should not change after beta. Second, we did not
yet implement the full_crc32 format for page_compressed pages.
This will be fixed in MDEV-18644.
This is joint work with Marko Mäkelä.
2019-02-19 20:00:00 +01:00
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-01-03 15:15:40 +01:00
|
|
|
static_assert(FIL_PAGE_END_LSN_OLD_CHKSUM % 8 == 0, "aligned");
|
|
|
|
|
static_assert(FIL_PAGE_LSN % 8 == 0, "aligned");
|
|
|
|
|
memcpy_aligned<8>(page + srv_page_size - FIL_PAGE_END_LSN_OLD_CHKSUM,
|
|
|
|
|
FIL_PAGE_LSN + page, 8);
|
|
|
|
|
|
2018-03-09 12:43:32 +01:00
|
|
|
if (block && srv_page_size == 16384) {
|
|
|
|
|
/* The page type could be garbage in old files
|
|
|
|
|
created before MySQL 5.5. Such files always
|
|
|
|
|
had a page size of 16 kilobytes. */
|
|
|
|
|
ulint page_type = fil_page_get_type(page);
|
|
|
|
|
ulint reset_type = page_type;
|
|
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
switch (block->page.id().page_no() % 16384) {
|
2018-03-09 12:43:32 +01:00
|
|
|
case 0:
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
reset_type = block->page.id().page_no() == 0
|
2018-03-09 12:43:32 +01:00
|
|
|
? FIL_PAGE_TYPE_FSP_HDR
|
|
|
|
|
: FIL_PAGE_TYPE_XDES;
|
|
|
|
|
break;
|
|
|
|
|
case 1:
|
|
|
|
|
reset_type = FIL_PAGE_IBUF_BITMAP;
|
|
|
|
|
break;
|
|
|
|
|
case FSP_TRX_SYS_PAGE_NO:
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
if (block->page.id()
|
|
|
|
|
== page_id_t(TRX_SYS_SPACE, TRX_SYS_PAGE_NO)) {
|
2018-03-09 12:43:32 +01:00
|
|
|
reset_type = FIL_PAGE_TYPE_TRX_SYS;
|
2016-08-12 10:17:45 +02:00
|
|
|
break;
|
2018-03-09 12:43:32 +01:00
|
|
|
}
|
|
|
|
|
/* fall through */
|
|
|
|
|
default:
|
|
|
|
|
switch (page_type) {
|
|
|
|
|
case FIL_PAGE_INDEX:
|
2018-03-13 07:15:06 +01:00
|
|
|
case FIL_PAGE_TYPE_INSTANT:
|
2018-03-09 12:43:32 +01:00
|
|
|
case FIL_PAGE_RTREE:
|
|
|
|
|
case FIL_PAGE_UNDO_LOG:
|
|
|
|
|
case FIL_PAGE_INODE:
|
|
|
|
|
case FIL_PAGE_IBUF_FREE_LIST:
|
|
|
|
|
case FIL_PAGE_TYPE_ALLOCATED:
|
|
|
|
|
case FIL_PAGE_TYPE_SYS:
|
|
|
|
|
case FIL_PAGE_TYPE_TRX_SYS:
|
|
|
|
|
case FIL_PAGE_TYPE_BLOB:
|
|
|
|
|
case FIL_PAGE_TYPE_ZBLOB:
|
|
|
|
|
case FIL_PAGE_TYPE_ZBLOB2:
|
2016-08-12 10:17:45 +02:00
|
|
|
break;
|
2018-03-09 12:43:32 +01:00
|
|
|
case FIL_PAGE_TYPE_FSP_HDR:
|
|
|
|
|
case FIL_PAGE_TYPE_XDES:
|
|
|
|
|
case FIL_PAGE_IBUF_BITMAP:
|
|
|
|
|
/* These pages should have
|
|
|
|
|
predetermined page numbers
|
|
|
|
|
(see above). */
|
2016-08-12 10:17:45 +02:00
|
|
|
default:
|
2018-03-09 12:43:32 +01:00
|
|
|
reset_type = FIL_PAGE_TYPE_UNKNOWN;
|
|
|
|
|
break;
|
2016-08-12 10:17:45 +02:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2018-03-09 12:43:32 +01:00
|
|
|
if (UNIV_UNLIKELY(page_type != reset_type)) {
|
|
|
|
|
ib::info()
|
|
|
|
|
<< "Resetting invalid page "
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
<< block->page.id() << " type "
|
2018-03-09 12:43:32 +01:00
|
|
|
<< page_type << " to "
|
|
|
|
|
<< reset_type << " when flushing.";
|
|
|
|
|
fil_page_set_type(page, reset_type);
|
2016-08-12 10:17:45 +02:00
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
2018-07-05 10:41:55 +02:00
|
|
|
uint32_t checksum = BUF_NO_CHECKSUM_MAGIC;
|
2018-03-09 12:43:32 +01:00
|
|
|
|
|
|
|
|
switch (srv_checksum_algorithm_t(srv_checksum_algorithm)) {
|
|
|
|
|
case SRV_CHECKSUM_ALGORITHM_INNODB:
|
|
|
|
|
case SRV_CHECKSUM_ALGORITHM_STRICT_INNODB:
|
|
|
|
|
checksum = buf_calc_page_new_checksum(page);
|
|
|
|
|
mach_write_to_4(page + FIL_PAGE_SPACE_OR_CHKSUM,
|
|
|
|
|
checksum);
|
|
|
|
|
/* With the InnoDB checksum, we overwrite the first 4 bytes of
|
|
|
|
|
the end lsn field to store the old formula checksum. Since it
|
|
|
|
|
depends also on the field FIL_PAGE_SPACE_OR_CHKSUM, it has to
|
|
|
|
|
be calculated after storing the new formula checksum. */
|
|
|
|
|
checksum = buf_calc_page_old_checksum(page);
|
|
|
|
|
break;
|
MDEV-12026: Implement innodb_checksum_algorithm=full_crc32
MariaDB data-at-rest encryption (innodb_encrypt_tables)
had repurposed the same unused data field that was repurposed
in MySQL 5.7 (and MariaDB 10.2) for the Split Sequence Number (SSN)
field of SPATIAL INDEX. Because of this, MariaDB was unable to
support encryption on SPATIAL INDEX pages.
Furthermore, InnoDB page checksums skipped some bytes, and there
are multiple variations and checksum algorithms. By default,
InnoDB accepts all variations of all algorithms that ever existed.
This unnecessarily weakens the page checksums.
We hereby introduce two more innodb_checksum_algorithm variants
(full_crc32, strict_full_crc32) that are special in a way:
When either setting is active, newly created data files will
carry a flag (fil_space_t::full_crc32()) that indicates that
all pages of the file will use a full CRC-32C checksum over the
entire page contents (excluding the bytes where the checksum
is stored, at the very end of the page). Such files will always
use that checksum, no matter what the parameter
innodb_checksum_algorithm is assigned to.
For old files, the old checksum algorithms will continue to be
used. The value strict_full_crc32 will be equivalent to strict_crc32
and the value full_crc32 will be equivalent to crc32.
ROW_FORMAT=COMPRESSED tables will only use the old format.
These tables do not support new features, such as larger
innodb_page_size or instant ADD/DROP COLUMN. They may be
deprecated in the future. We do not want an unnecessary
file format change for them.
The new full_crc32() format also cleans up the MariaDB tablespace
flags. We will reserve flags to store the page_compressed
compression algorithm, and to store the compressed payload length,
so that checksum can be computed over the compressed (and
possibly encrypted) stream and can be validated without
decrypting or decompressing the page.
In the full_crc32 format, there no longer are separate before-encryption
and after-encryption checksums for pages. The single checksum is
computed on the page contents that is written to the file.
We do not make the new algorithm the default for two reasons.
First, MariaDB 10.4.2 was a beta release, and the default values
of parameters should not change after beta. Second, we did not
yet implement the full_crc32 format for page_compressed pages.
This will be fixed in MDEV-18644.
This is joint work with Marko Mäkelä.
2019-02-19 20:00:00 +01:00
|
|
|
case SRV_CHECKSUM_ALGORITHM_FULL_CRC32:
|
|
|
|
|
case SRV_CHECKSUM_ALGORITHM_STRICT_FULL_CRC32:
|
2018-03-09 12:43:32 +01:00
|
|
|
case SRV_CHECKSUM_ALGORITHM_CRC32:
|
|
|
|
|
case SRV_CHECKSUM_ALGORITHM_STRICT_CRC32:
|
|
|
|
|
/* In other cases we write the same checksum to both fields. */
|
|
|
|
|
checksum = buf_calc_page_crc32(page);
|
|
|
|
|
mach_write_to_4(page + FIL_PAGE_SPACE_OR_CHKSUM,
|
|
|
|
|
checksum);
|
|
|
|
|
break;
|
|
|
|
|
case SRV_CHECKSUM_ALGORITHM_NONE:
|
|
|
|
|
case SRV_CHECKSUM_ALGORITHM_STRICT_NONE:
|
|
|
|
|
mach_write_to_4(page + FIL_PAGE_SPACE_OR_CHKSUM,
|
|
|
|
|
checksum);
|
|
|
|
|
break;
|
|
|
|
|
/* no default so the compiler will emit a warning if
|
|
|
|
|
new enum is added and not handled here */
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2018-04-27 12:49:25 +02:00
|
|
|
mach_write_to_4(page + srv_page_size - FIL_PAGE_END_LSN_OLD_CHKSUM,
|
2014-02-26 19:11:54 +01:00
|
|
|
checksum);
|
|
|
|
|
}
|
|
|
|
|
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
/** Reserve a buffer for compression.
|
|
|
|
|
@param[in,out] slot reserved slot */
|
|
|
|
|
static void buf_tmp_reserve_compression_buf(buf_tmp_buffer_t* slot)
|
|
|
|
|
{
|
|
|
|
|
if (slot->comp_buf)
|
|
|
|
|
return;
|
|
|
|
|
/* Both Snappy and LZO compression methods require that the output
|
|
|
|
|
buffer be bigger than input buffer. Adjust the allocated size. */
|
|
|
|
|
ulint size= srv_page_size;
|
|
|
|
|
#ifdef HAVE_LZO
|
|
|
|
|
size+= LZO1X_1_15_MEM_COMPRESS;
|
|
|
|
|
#elif defined HAVE_SNAPPY
|
|
|
|
|
size= snappy_max_compressed_length(size);
|
|
|
|
|
#endif
|
|
|
|
|
slot->comp_buf= static_cast<byte*>(aligned_malloc(size, srv_page_size));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** Encrypt a buffer of temporary tablespace
|
|
|
|
|
@param[in] offset Page offset
|
|
|
|
|
@param[in] s Page to encrypt
|
|
|
|
|
@param[in,out] d Output buffer
|
|
|
|
|
@return encrypted buffer or NULL */
|
|
|
|
|
static byte* buf_tmp_page_encrypt(ulint offset, const byte* s, byte* d)
|
|
|
|
|
{
|
|
|
|
|
/* Calculate the start offset in a page */
|
2020-03-10 19:05:17 +01:00
|
|
|
uint srclen= static_cast<uint>(srv_page_size) -
|
|
|
|
|
(FIL_PAGE_FILE_FLUSH_LSN_OR_KEY_VERSION +
|
|
|
|
|
FIL_PAGE_FCRC32_CHECKSUM);
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
const byte* src= s + FIL_PAGE_FILE_FLUSH_LSN_OR_KEY_VERSION;
|
|
|
|
|
byte* dst= d + FIL_PAGE_FILE_FLUSH_LSN_OR_KEY_VERSION;
|
|
|
|
|
|
|
|
|
|
memcpy(d, s, FIL_PAGE_FILE_FLUSH_LSN_OR_KEY_VERSION);
|
|
|
|
|
|
|
|
|
|
if (!log_tmp_block_encrypt(src, srclen, dst, (offset * srv_page_size), true))
|
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
|
|
const ulint payload= srv_page_size - FIL_PAGE_FCRC32_CHECKSUM;
|
|
|
|
|
mach_write_to_4(d + payload, ut_crc32(d, payload));
|
|
|
|
|
|
|
|
|
|
srv_stats.pages_encrypted.inc();
|
|
|
|
|
srv_stats.n_temp_blocks_encrypted.inc();
|
|
|
|
|
return d;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** Encryption and page_compression hook that is called just before
|
|
|
|
|
a page is written to disk.
|
|
|
|
|
@param[in,out] space tablespace
|
|
|
|
|
@param[in,out] bpage buffer page
|
|
|
|
|
@param[in] s physical page frame that is being encrypted
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
@param[in,out] size payload size in bytes
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
@return page frame to be written to file
|
|
|
|
|
(may be src_frame or an encrypted/compressed copy of it) */
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
static byte *buf_page_encrypt(fil_space_t* space, buf_page_t* bpage, byte* s,
|
|
|
|
|
size_t *size)
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
{
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ut_ad(bpage->status != buf_page_t::FREED);
|
|
|
|
|
ut_ad(space->id == bpage->id().space());
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
|
|
|
|
|
ut_d(fil_page_type_validate(space, s));
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
const uint32_t page_no= bpage->id().page_no();
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
switch (page_no) {
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
case TRX_SYS_PAGE_NO:
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
if (bpage->id().space() != TRX_SYS_SPACE)
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
break;
|
|
|
|
|
/* The TRX_SYS page is neither encrypted nor compressed, because
|
|
|
|
|
it contains the address of the doublewrite buffer. */
|
|
|
|
|
/* fall through */
|
|
|
|
|
case 0:
|
|
|
|
|
/* Page 0 of a tablespace is not encrypted/compressed */
|
|
|
|
|
return s;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
fil_space_crypt_t *crypt_data= space->crypt_data;
|
|
|
|
|
bool encrypted, page_compressed;
|
|
|
|
|
if (space->purpose == FIL_TYPE_TEMPORARY)
|
|
|
|
|
{
|
|
|
|
|
ut_ad(!crypt_data);
|
|
|
|
|
encrypted= innodb_encrypt_temporary_tables;
|
|
|
|
|
page_compressed= false;
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
encrypted= crypt_data && !crypt_data->not_encrypted() &&
|
|
|
|
|
crypt_data->type != CRYPT_SCHEME_UNENCRYPTED &&
|
|
|
|
|
(!crypt_data->is_default_encryption() || srv_encrypt_tables);
|
|
|
|
|
page_compressed= space->is_compressed();
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
const bool full_crc32= space->full_crc32();
|
|
|
|
|
|
|
|
|
|
if (!encrypted && !page_compressed)
|
|
|
|
|
{
|
|
|
|
|
/* No need to encrypt or compress. Clear key-version & crypt-checksum. */
|
2019-11-26 09:14:07 +01:00
|
|
|
static_assert(FIL_PAGE_FCRC32_KEY_VERSION % 4 == 0, "alignment");
|
|
|
|
|
static_assert(FIL_PAGE_FILE_FLUSH_LSN_OR_KEY_VERSION % 4 == 2,
|
|
|
|
|
"not perfect alignment");
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
if (full_crc32)
|
2019-11-26 09:14:07 +01:00
|
|
|
memset_aligned<4>(s + FIL_PAGE_FCRC32_KEY_VERSION, 0, 4);
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
else
|
2019-11-26 09:14:07 +01:00
|
|
|
memset_aligned<2>(s + FIL_PAGE_FILE_FLUSH_LSN_OR_KEY_VERSION, 0, 8);
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
return s;
|
|
|
|
|
}
|
|
|
|
|
|
2019-11-26 09:14:07 +01:00
|
|
|
static_assert(FIL_PAGE_FCRC32_END_LSN % 4 == 0, "alignment");
|
|
|
|
|
static_assert(FIL_PAGE_LSN % 8 == 0, "alignment");
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
if (full_crc32)
|
2019-11-26 09:14:07 +01:00
|
|
|
memcpy_aligned<4>(s + srv_page_size - FIL_PAGE_FCRC32_END_LSN,
|
|
|
|
|
FIL_PAGE_LSN + 4 + s, 4);
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
|
|
|
|
|
ut_ad(!bpage->zip_size() || !page_compressed);
|
|
|
|
|
/* Find free slot from temporary memory array */
|
2020-03-18 20:48:00 +01:00
|
|
|
buf_tmp_buffer_t *slot= buf_pool.io_buf_reserve();
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
ut_a(slot);
|
|
|
|
|
slot->allocate();
|
|
|
|
|
slot->out_buf= NULL;
|
|
|
|
|
bpage->slot= slot;
|
|
|
|
|
|
|
|
|
|
byte *d= slot->crypt_buf;
|
|
|
|
|
|
|
|
|
|
if (!page_compressed)
|
|
|
|
|
{
|
|
|
|
|
not_compressed:
|
|
|
|
|
byte *tmp= space->purpose == FIL_TYPE_TEMPORARY
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
? buf_tmp_page_encrypt(page_no, s, d)
|
|
|
|
|
: fil_space_encrypt(space, page_no, s, d);
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
|
|
|
|
|
slot->out_buf= d= tmp;
|
|
|
|
|
|
|
|
|
|
ut_d(fil_page_type_validate(space, tmp));
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
ut_ad(space->purpose != FIL_TYPE_TEMPORARY);
|
|
|
|
|
/* First we compress the page content */
|
|
|
|
|
buf_tmp_reserve_compression_buf(slot);
|
|
|
|
|
byte *tmp= slot->comp_buf;
|
|
|
|
|
ulint len= fil_page_compress(s, tmp, space->flags,
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
fil_space_get_block_size(space, page_no),
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
encrypted);
|
|
|
|
|
|
|
|
|
|
if (!len)
|
|
|
|
|
goto not_compressed;
|
|
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
*size= len;
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
|
|
|
|
|
if (full_crc32)
|
|
|
|
|
{
|
|
|
|
|
ut_d(bool compressed = false);
|
|
|
|
|
len= buf_page_full_crc32_size(tmp,
|
|
|
|
|
#ifdef UNIV_DEBUG
|
|
|
|
|
&compressed,
|
|
|
|
|
#else
|
|
|
|
|
NULL,
|
|
|
|
|
#endif
|
|
|
|
|
NULL);
|
|
|
|
|
ut_ad(compressed);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Workaround for MDEV-15527. */
|
|
|
|
|
memset(tmp + len, 0 , srv_page_size - len);
|
|
|
|
|
ut_d(fil_page_type_validate(space, tmp));
|
|
|
|
|
|
|
|
|
|
if (encrypted)
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
tmp = fil_space_encrypt(space, page_no, tmp, d);
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
|
|
|
|
|
if (full_crc32)
|
|
|
|
|
{
|
2019-11-26 09:14:07 +01:00
|
|
|
static_assert(FIL_PAGE_FCRC32_CHECKSUM == 4, "alignment");
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
mach_write_to_4(tmp + len - 4, ut_crc32(tmp, len - 4));
|
|
|
|
|
ut_ad(!buf_page_is_corrupted(true, tmp, space->flags));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
slot->out_buf= d= tmp;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
ut_d(fil_page_type_validate(space, d));
|
|
|
|
|
return d;
|
|
|
|
|
}
|
|
|
|
|
|
MDEV-15528 Punch holes when pages are freed
When a InnoDB data file page is freed, its contents becomes garbage,
and any storage allocated in the data file is wasted. During flushing,
InnoDB initializes the page with zeros if scrubbing is enabled. If the
tablespace is compressed then InnoDB should punch a hole else ignore the
flushing of the freed page.
buf_page_t:
- Replaced the variable file_page_was_freed, init_on_flush in buf_page_t
with status enum variable.
- Changed all debug assert of file_page_was_freed to DBUG_ASSERT
of buf_page_t::status
Removed buf_page_set_file_page_was_freed(),
buf_page_reset_file_page_was_freed().
buf_page_free(): Newly added function which takes X-lock on the page
before marking the status as FREED. So that InnoDB flush handler can
avoid concurrent flush of the freed page. Also while flushing the page,
InnoDB make sure that redo log which does freeing of the page also written
to the disk. Currently, this function only marks the page as FREED if
it is in buffer pool
buf_flush_freed_page(): Newly added function which initializes zeros
asynchorously if innodb_immediate_scrub_data_uncompressed is enabled.
Punch a hole to the file synchorously if page_compressed is enabled.
Reset the io_fix to NORMAL. Release the block from flush list and
associated mutex before writing zeros or punch a hole to the file.
buf_flush_page(): Removed the unnecessary usage of temporary
variable "flush"
fil_io(): Introduce new parameter called punch_hole. It allows fil_io()
to punch the hole to the file for the given offset.
buf_page_create(): Let the callers assign buf_page_t::status.
Every caller should eventually invoke mtr_t::init().
fsp_page_create(): Remove the unused mtr_t parameter.
In all other callers of buf_page_create() except fsp_page_create(),
before invoking mtr_t::init(), invoke
mtr_t::sx_latch_at_savepoint() or mtr_t::x_latch_at_savepoint().
mtr_t::init(): Initialize buf_page_t::status also for the temporary
tablespace (when redo logging is disabled), to avoid assertion failures.
2020-03-09 08:55:33 +01:00
|
|
|
/** The following function deals with freed page during flushing.
|
|
|
|
|
i) Writing zeros to the file asynchronously if scrubbing is enabled
|
|
|
|
|
ii) Punch the hole to the file synchoronously if page_compressed is
|
|
|
|
|
enabled for the tablespace
|
|
|
|
|
This function also resets the IO_FIX to IO_NONE and making the
|
|
|
|
|
page status as NORMAL. It initiates the write to the file only after
|
|
|
|
|
releasing the page from flush list and its associated mutex.
|
2020-06-11 19:22:47 +02:00
|
|
|
@param[in,out] bpage freed buffer page */
|
|
|
|
|
static void buf_release_freed_page(buf_page_t *bpage)
|
MDEV-15528 Punch holes when pages are freed
When a InnoDB data file page is freed, its contents becomes garbage,
and any storage allocated in the data file is wasted. During flushing,
InnoDB initializes the page with zeros if scrubbing is enabled. If the
tablespace is compressed then InnoDB should punch a hole else ignore the
flushing of the freed page.
buf_page_t:
- Replaced the variable file_page_was_freed, init_on_flush in buf_page_t
with status enum variable.
- Changed all debug assert of file_page_was_freed to DBUG_ASSERT
of buf_page_t::status
Removed buf_page_set_file_page_was_freed(),
buf_page_reset_file_page_was_freed().
buf_page_free(): Newly added function which takes X-lock on the page
before marking the status as FREED. So that InnoDB flush handler can
avoid concurrent flush of the freed page. Also while flushing the page,
InnoDB make sure that redo log which does freeing of the page also written
to the disk. Currently, this function only marks the page as FREED if
it is in buffer pool
buf_flush_freed_page(): Newly added function which initializes zeros
asynchorously if innodb_immediate_scrub_data_uncompressed is enabled.
Punch a hole to the file synchorously if page_compressed is enabled.
Reset the io_fix to NORMAL. Release the block from flush list and
associated mutex before writing zeros or punch a hole to the file.
buf_flush_page(): Removed the unnecessary usage of temporary
variable "flush"
fil_io(): Introduce new parameter called punch_hole. It allows fil_io()
to punch the hole to the file for the given offset.
buf_page_create(): Let the callers assign buf_page_t::status.
Every caller should eventually invoke mtr_t::init().
fsp_page_create(): Remove the unused mtr_t parameter.
In all other callers of buf_page_create() except fsp_page_create(),
before invoking mtr_t::init(), invoke
mtr_t::sx_latch_at_savepoint() or mtr_t::x_latch_at_savepoint().
mtr_t::init(): Initialize buf_page_t::status also for the temporary
tablespace (when redo logging is disabled), to avoid assertion failures.
2020-03-09 08:55:33 +01:00
|
|
|
{
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ut_ad(bpage->in_file());
|
|
|
|
|
const bool uncompressed= bpage->state() == BUF_BLOCK_FILE_PAGE;
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
bpage->set_io_fix(BUF_IO_NONE);
|
2020-03-10 08:34:09 +01:00
|
|
|
bpage->status= buf_page_t::NORMAL;
|
2020-12-09 08:22:13 +01:00
|
|
|
const bool temp= fsp_is_system_temporary(bpage->id().space());
|
|
|
|
|
ut_ad(!temp || uncompressed);
|
|
|
|
|
ut_ad(!temp || bpage->oldest_modification() == 1);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
2020-12-09 08:22:13 +01:00
|
|
|
bpage->clear_oldest_modification();
|
|
|
|
|
if (!temp)
|
|
|
|
|
buf_flush_remove_low(bpage);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
2020-03-10 08:34:09 +01:00
|
|
|
|
|
|
|
|
if (uncompressed)
|
MDEV-24142: Replace InnoDB rw_lock_t with sux_lock
InnoDB buffer pool block and index tree latches depend on a
special kind of read-update-write lock that allows reentrant
(recursive) acquisition of the 'update' and 'write' locks
as well as an upgrade from 'update' lock to 'write' lock.
The 'update' lock allows any number of reader locks from
other threads, but no concurrent 'update' or 'write' lock.
If there were no requirement to support an upgrade from 'update'
to 'write', we could compose the lock out of two srw_lock
(implemented as any type of native rw-lock, such as SRWLOCK on
Microsoft Windows). Removing this requirement is very difficult,
so in commit f7e7f487d4b06695f91f6fbeb0396b9d87fc7bbf we
implemented an 'update' mode to our srw_lock.
Re-entrant or recursive locking is mostly needed when writing or
freeing BLOB pages, but also in crash recovery or when merging
buffered changes to an index page. The re-entrancy allows us to
attach a previously acquired page to a sub-mini-transaction that
will be committed before whatever else is holding the page latch.
The SUX lock supports Shared ('read'), Update, and eXclusive ('write')
locking modes. The S latches are not re-entrant, but a single S latch
may be acquired even if the thread already holds an U latch.
The idea of the U latch is to allow a write of something that concurrent
readers do not care about (such as the contents of BTR_SEG_LEAF,
BTR_SEG_TOP and other page allocation metadata structures, or
the MDEV-6076 PAGE_ROOT_AUTO_INC). (The PAGE_ROOT_AUTO_INC field
is only updated when a dict_table_t for the table exists, and only
read when a dict_table_t for the table is being added to dict_sys.)
block_lock::u_lock_try(bool for_io=true) is used in buf_flush_page()
to allow concurrent readers but no concurrent modifications while the
page is being written to the data file. That latch will be released
by buf_page_write_complete() in a different thread. Hence, we use
the special lock owner value FOR_IO.
The index_lock::u_lock() improves concurrency on operations that
involve non-leaf index pages.
The interface has been cleaned up a little. We will use
x_lock_recursive() instead of x_lock() when we know that a
lock is already held by the current thread. Similarly,
a lock upgrade from U to X is only allowed via u_x_upgrade()
or x_lock_upgraded() but not via x_lock().
We will disable the LatchDebug and sync_array interfaces to
InnoDB rw-locks.
The SEMAPHORES section of SHOW ENGINE INNODB STATUS output
will no longer include any information about InnoDB rw-locks,
only TTASEventMutex (cmake -DMUTEXTYPE=event) waits.
This will make a part of the 'innotop' script dead code.
The block_lock buf_block_t::lock will not be covered by any
PERFORMANCE_SCHEMA instrumentation.
SHOW ENGINE INNODB MUTEX and INFORMATION_SCHEMA.INNODB_MUTEXES
will no longer output source code file names or line numbers.
The dict_index_t::lock will be identified by index and table names,
which should be much more useful. PERFORMANCE_SCHEMA is lumping
information about all dict_index_t::lock together as
event_name='wait/synch/sxlock/innodb/index_tree_rw_lock'.
buf_page_free(): Remove the file,line parameters. The sux_lock will
not store such diagnostic information.
buf_block_dbg_add_level(): Define as empty macro, to be removed
in a subsequent commit.
Unless the build was configured with cmake -DPLUGIN_PERFSCHEMA=NO
the index_lock dict_index_t::lock will be instrumented via
PERFORMANCE_SCHEMA. Similar to
commit 1669c8890ca2e9092213626e5b047e58ca8b1e77
we will distinguish lock waits by registering shared_lock,exclusive_lock
events instead of try_shared_lock,try_exclusive_lock.
Actual 'try' operations will not be instrumented at all.
rw_lock_list: Remove. After MDEV-24167, this only covered
buf_block_t::lock and dict_index_t::lock. We will output their
information by traversing buf_pool or dict_sys.
2020-12-03 14:18:51 +01:00
|
|
|
reinterpret_cast<buf_block_t*>(bpage)->lock.u_unlock(true);
|
2020-03-10 08:34:09 +01:00
|
|
|
|
2020-06-11 19:22:47 +02:00
|
|
|
buf_LRU_free_page(bpage, true);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
MDEV-15528 Punch holes when pages are freed
When a InnoDB data file page is freed, its contents becomes garbage,
and any storage allocated in the data file is wasted. During flushing,
InnoDB initializes the page with zeros if scrubbing is enabled. If the
tablespace is compressed then InnoDB should punch a hole else ignore the
flushing of the freed page.
buf_page_t:
- Replaced the variable file_page_was_freed, init_on_flush in buf_page_t
with status enum variable.
- Changed all debug assert of file_page_was_freed to DBUG_ASSERT
of buf_page_t::status
Removed buf_page_set_file_page_was_freed(),
buf_page_reset_file_page_was_freed().
buf_page_free(): Newly added function which takes X-lock on the page
before marking the status as FREED. So that InnoDB flush handler can
avoid concurrent flush of the freed page. Also while flushing the page,
InnoDB make sure that redo log which does freeing of the page also written
to the disk. Currently, this function only marks the page as FREED if
it is in buffer pool
buf_flush_freed_page(): Newly added function which initializes zeros
asynchorously if innodb_immediate_scrub_data_uncompressed is enabled.
Punch a hole to the file synchorously if page_compressed is enabled.
Reset the io_fix to NORMAL. Release the block from flush list and
associated mutex before writing zeros or punch a hole to the file.
buf_flush_page(): Removed the unnecessary usage of temporary
variable "flush"
fil_io(): Introduce new parameter called punch_hole. It allows fil_io()
to punch the hole to the file for the given offset.
buf_page_create(): Let the callers assign buf_page_t::status.
Every caller should eventually invoke mtr_t::init().
fsp_page_create(): Remove the unused mtr_t parameter.
In all other callers of buf_page_create() except fsp_page_create(),
before invoking mtr_t::init(), invoke
mtr_t::sx_latch_at_savepoint() or mtr_t::x_latch_at_savepoint().
mtr_t::init(): Initialize buf_page_t::status also for the temporary
tablespace (when redo logging is disabled), to avoid assertion failures.
2020-03-09 08:55:33 +01:00
|
|
|
}
|
|
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
/** Write a flushable page from buf_pool to a file.
|
|
|
|
|
buf_pool.mutex must be held.
|
|
|
|
|
@param bpage buffer control block
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
@param lru true=buf_pool.LRU; false=buf_pool.flush_list
|
|
|
|
|
@param space tablespace
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
@return whether the page was flushed and buf_pool.mutex was released */
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
static bool buf_flush_page(buf_page_t *bpage, bool lru, fil_space_t *space)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ut_ad(bpage->in_file());
|
|
|
|
|
ut_ad(bpage->ready_for_flush());
|
2020-10-26 14:59:30 +01:00
|
|
|
ut_ad((space->purpose == FIL_TYPE_TEMPORARY) ==
|
|
|
|
|
(space == fil_system.temp_space));
|
|
|
|
|
ut_ad(space->purpose == FIL_TYPE_TABLESPACE ||
|
|
|
|
|
space->atomic_write_supported);
|
2020-10-26 15:04:12 +01:00
|
|
|
ut_ad(space->referenced());
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
MDEV-24142: Replace InnoDB rw_lock_t with sux_lock
InnoDB buffer pool block and index tree latches depend on a
special kind of read-update-write lock that allows reentrant
(recursive) acquisition of the 'update' and 'write' locks
as well as an upgrade from 'update' lock to 'write' lock.
The 'update' lock allows any number of reader locks from
other threads, but no concurrent 'update' or 'write' lock.
If there were no requirement to support an upgrade from 'update'
to 'write', we could compose the lock out of two srw_lock
(implemented as any type of native rw-lock, such as SRWLOCK on
Microsoft Windows). Removing this requirement is very difficult,
so in commit f7e7f487d4b06695f91f6fbeb0396b9d87fc7bbf we
implemented an 'update' mode to our srw_lock.
Re-entrant or recursive locking is mostly needed when writing or
freeing BLOB pages, but also in crash recovery or when merging
buffered changes to an index page. The re-entrancy allows us to
attach a previously acquired page to a sub-mini-transaction that
will be committed before whatever else is holding the page latch.
The SUX lock supports Shared ('read'), Update, and eXclusive ('write')
locking modes. The S latches are not re-entrant, but a single S latch
may be acquired even if the thread already holds an U latch.
The idea of the U latch is to allow a write of something that concurrent
readers do not care about (such as the contents of BTR_SEG_LEAF,
BTR_SEG_TOP and other page allocation metadata structures, or
the MDEV-6076 PAGE_ROOT_AUTO_INC). (The PAGE_ROOT_AUTO_INC field
is only updated when a dict_table_t for the table exists, and only
read when a dict_table_t for the table is being added to dict_sys.)
block_lock::u_lock_try(bool for_io=true) is used in buf_flush_page()
to allow concurrent readers but no concurrent modifications while the
page is being written to the data file. That latch will be released
by buf_page_write_complete() in a different thread. Hence, we use
the special lock owner value FOR_IO.
The index_lock::u_lock() improves concurrency on operations that
involve non-leaf index pages.
The interface has been cleaned up a little. We will use
x_lock_recursive() instead of x_lock() when we know that a
lock is already held by the current thread. Similarly,
a lock upgrade from U to X is only allowed via u_x_upgrade()
or x_lock_upgraded() but not via x_lock().
We will disable the LatchDebug and sync_array interfaces to
InnoDB rw-locks.
The SEMAPHORES section of SHOW ENGINE INNODB STATUS output
will no longer include any information about InnoDB rw-locks,
only TTASEventMutex (cmake -DMUTEXTYPE=event) waits.
This will make a part of the 'innotop' script dead code.
The block_lock buf_block_t::lock will not be covered by any
PERFORMANCE_SCHEMA instrumentation.
SHOW ENGINE INNODB MUTEX and INFORMATION_SCHEMA.INNODB_MUTEXES
will no longer output source code file names or line numbers.
The dict_index_t::lock will be identified by index and table names,
which should be much more useful. PERFORMANCE_SCHEMA is lumping
information about all dict_index_t::lock together as
event_name='wait/synch/sxlock/innodb/index_tree_rw_lock'.
buf_page_free(): Remove the file,line parameters. The sux_lock will
not store such diagnostic information.
buf_block_dbg_add_level(): Define as empty macro, to be removed
in a subsequent commit.
Unless the build was configured with cmake -DPLUGIN_PERFSCHEMA=NO
the index_lock dict_index_t::lock will be instrumented via
PERFORMANCE_SCHEMA. Similar to
commit 1669c8890ca2e9092213626e5b047e58ca8b1e77
we will distinguish lock waits by registering shared_lock,exclusive_lock
events instead of try_shared_lock,try_exclusive_lock.
Actual 'try' operations will not be instrumented at all.
rw_lock_list: Remove. After MDEV-24167, this only covered
buf_block_t::lock and dict_index_t::lock. We will output their
information by traversing buf_pool or dict_sys.
2020-12-03 14:18:51 +01:00
|
|
|
block_lock *rw_lock;
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
|
|
|
|
if (bpage->state() != BUF_BLOCK_FILE_PAGE)
|
|
|
|
|
rw_lock= nullptr;
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
rw_lock= &reinterpret_cast<buf_block_t*>(bpage)->lock;
|
MDEV-24142: Replace InnoDB rw_lock_t with sux_lock
InnoDB buffer pool block and index tree latches depend on a
special kind of read-update-write lock that allows reentrant
(recursive) acquisition of the 'update' and 'write' locks
as well as an upgrade from 'update' lock to 'write' lock.
The 'update' lock allows any number of reader locks from
other threads, but no concurrent 'update' or 'write' lock.
If there were no requirement to support an upgrade from 'update'
to 'write', we could compose the lock out of two srw_lock
(implemented as any type of native rw-lock, such as SRWLOCK on
Microsoft Windows). Removing this requirement is very difficult,
so in commit f7e7f487d4b06695f91f6fbeb0396b9d87fc7bbf we
implemented an 'update' mode to our srw_lock.
Re-entrant or recursive locking is mostly needed when writing or
freeing BLOB pages, but also in crash recovery or when merging
buffered changes to an index page. The re-entrancy allows us to
attach a previously acquired page to a sub-mini-transaction that
will be committed before whatever else is holding the page latch.
The SUX lock supports Shared ('read'), Update, and eXclusive ('write')
locking modes. The S latches are not re-entrant, but a single S latch
may be acquired even if the thread already holds an U latch.
The idea of the U latch is to allow a write of something that concurrent
readers do not care about (such as the contents of BTR_SEG_LEAF,
BTR_SEG_TOP and other page allocation metadata structures, or
the MDEV-6076 PAGE_ROOT_AUTO_INC). (The PAGE_ROOT_AUTO_INC field
is only updated when a dict_table_t for the table exists, and only
read when a dict_table_t for the table is being added to dict_sys.)
block_lock::u_lock_try(bool for_io=true) is used in buf_flush_page()
to allow concurrent readers but no concurrent modifications while the
page is being written to the data file. That latch will be released
by buf_page_write_complete() in a different thread. Hence, we use
the special lock owner value FOR_IO.
The index_lock::u_lock() improves concurrency on operations that
involve non-leaf index pages.
The interface has been cleaned up a little. We will use
x_lock_recursive() instead of x_lock() when we know that a
lock is already held by the current thread. Similarly,
a lock upgrade from U to X is only allowed via u_x_upgrade()
or x_lock_upgraded() but not via x_lock().
We will disable the LatchDebug and sync_array interfaces to
InnoDB rw-locks.
The SEMAPHORES section of SHOW ENGINE INNODB STATUS output
will no longer include any information about InnoDB rw-locks,
only TTASEventMutex (cmake -DMUTEXTYPE=event) waits.
This will make a part of the 'innotop' script dead code.
The block_lock buf_block_t::lock will not be covered by any
PERFORMANCE_SCHEMA instrumentation.
SHOW ENGINE INNODB MUTEX and INFORMATION_SCHEMA.INNODB_MUTEXES
will no longer output source code file names or line numbers.
The dict_index_t::lock will be identified by index and table names,
which should be much more useful. PERFORMANCE_SCHEMA is lumping
information about all dict_index_t::lock together as
event_name='wait/synch/sxlock/innodb/index_tree_rw_lock'.
buf_page_free(): Remove the file,line parameters. The sux_lock will
not store such diagnostic information.
buf_block_dbg_add_level(): Define as empty macro, to be removed
in a subsequent commit.
Unless the build was configured with cmake -DPLUGIN_PERFSCHEMA=NO
the index_lock dict_index_t::lock will be instrumented via
PERFORMANCE_SCHEMA. Similar to
commit 1669c8890ca2e9092213626e5b047e58ca8b1e77
we will distinguish lock waits by registering shared_lock,exclusive_lock
events instead of try_shared_lock,try_exclusive_lock.
Actual 'try' operations will not be instrumented at all.
rw_lock_list: Remove. After MDEV-24167, this only covered
buf_block_t::lock and dict_index_t::lock. We will output their
information by traversing buf_pool or dict_sys.
2020-12-03 14:18:51 +01:00
|
|
|
if (!rw_lock->u_lock_try(true))
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
return false;
|
|
|
|
|
}
|
MDEV-21132 Remove buf_page_t::newest_modification
At each mini-transaction commit, the log sequence number of the
mini-transaction must be written to each modified page, so that
it will be available in the FIL_PAGE_LSN field when the page is
being read in crash recovery.
InnoDB was unnecessarily allocating redundant storage for the
field, in buf_page_t::newest_modification. Let us access
FIL_PAGE_LSN directly.
Furthermore, on ALTER TABLE...IMPORT TABLESPACE, let us write
0 to FIL_PAGE_LSN instead of using log_sys.lsn.
buf_flush_init_for_writing(), buf_flush_update_zip_checksum(),
fil_encrypt_buf_for_full_crc32(), fil_encrypt_buf(),
fil_space_encrypt(): Remove the parameter lsn.
buf_page_get_newest_modification(): Merge with the only caller.
buf_tmp_reserve_compression_buf(), buf_tmp_page_encrypt(),
buf_page_encrypt(): Define static in the same compilation unit
with the only caller.
PageConverter::m_current_lsn: Remove. Write 0 to FIL_PAGE_LSN
on ALTER TABLE...IMPORT TABLESPACE.
2019-11-25 08:39:51 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
bpage->set_io_fix(BUF_IO_WRITE);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
buf_flush_page_count++;
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
|
|
|
|
mysql_mutex_assert_not_owner(&buf_pool.flush_list_mutex);
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
/* We are holding rw_lock = buf_block_t::lock in SX mode except if
|
|
|
|
|
this is a ROW_FORMAT=COMPRESSED page whose uncompressed page frame
|
|
|
|
|
has been evicted from the buffer pool.
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
Apart from possible rw_lock protection, bpage is also protected by
|
|
|
|
|
io_fix and oldest_modification()!=0. Thus, it cannot be relocated in
|
|
|
|
|
the buffer pool or removed from flush_list or LRU_list. */
|
2017-04-21 10:28:18 +02:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
DBUG_PRINT("ib_buf", ("%s %u page %u:%u",
|
|
|
|
|
lru ? "LRU" : "flush_list",
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
bpage->id().space(), bpage->id().page_no()));
|
|
|
|
|
ut_ad(bpage->io_fix() == BUF_IO_WRITE);
|
|
|
|
|
ut_ad(bpage->oldest_modification());
|
|
|
|
|
ut_ad(bpage->state() ==
|
|
|
|
|
(rw_lock ? BUF_BLOCK_FILE_PAGE : BUF_BLOCK_ZIP_PAGE));
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ut_ad(ULINT_UNDEFINED >
|
|
|
|
|
(lru ? buf_pool.n_flush_LRU : buf_pool.n_flush_list));
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
|
|
|
|
/* Because bpage->status can only be changed while buf_block_t
|
|
|
|
|
exists, it cannot be modified for ROW_FORMAT=COMPRESSED pages
|
|
|
|
|
without first allocating the uncompressed page frame. Such
|
|
|
|
|
allocation cannot be completed due to our io_fix. So, bpage->status
|
|
|
|
|
is protected even if !rw_lock. */
|
|
|
|
|
const auto status= bpage->status;
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
buf_block_t *block= reinterpret_cast<buf_block_t*>(bpage);
|
|
|
|
|
page_t *frame= bpage->zip.data;
|
|
|
|
|
|
|
|
|
|
if (UNIV_LIKELY(space->purpose == FIL_TYPE_TABLESPACE))
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
const lsn_t lsn= mach_read_from_8(my_assume_aligned<8>
|
|
|
|
|
(FIL_PAGE_LSN +
|
|
|
|
|
(frame ? frame : block->frame)));
|
|
|
|
|
ut_ad(lsn);
|
|
|
|
|
ut_ad(lsn >= bpage->oldest_modification());
|
|
|
|
|
ut_ad(!srv_read_only_mode);
|
|
|
|
|
if (UNIV_UNLIKELY(lsn > log_sys.get_flushed_lsn()))
|
|
|
|
|
{
|
|
|
|
|
if (rw_lock)
|
MDEV-24142: Replace InnoDB rw_lock_t with sux_lock
InnoDB buffer pool block and index tree latches depend on a
special kind of read-update-write lock that allows reentrant
(recursive) acquisition of the 'update' and 'write' locks
as well as an upgrade from 'update' lock to 'write' lock.
The 'update' lock allows any number of reader locks from
other threads, but no concurrent 'update' or 'write' lock.
If there were no requirement to support an upgrade from 'update'
to 'write', we could compose the lock out of two srw_lock
(implemented as any type of native rw-lock, such as SRWLOCK on
Microsoft Windows). Removing this requirement is very difficult,
so in commit f7e7f487d4b06695f91f6fbeb0396b9d87fc7bbf we
implemented an 'update' mode to our srw_lock.
Re-entrant or recursive locking is mostly needed when writing or
freeing BLOB pages, but also in crash recovery or when merging
buffered changes to an index page. The re-entrancy allows us to
attach a previously acquired page to a sub-mini-transaction that
will be committed before whatever else is holding the page latch.
The SUX lock supports Shared ('read'), Update, and eXclusive ('write')
locking modes. The S latches are not re-entrant, but a single S latch
may be acquired even if the thread already holds an U latch.
The idea of the U latch is to allow a write of something that concurrent
readers do not care about (such as the contents of BTR_SEG_LEAF,
BTR_SEG_TOP and other page allocation metadata structures, or
the MDEV-6076 PAGE_ROOT_AUTO_INC). (The PAGE_ROOT_AUTO_INC field
is only updated when a dict_table_t for the table exists, and only
read when a dict_table_t for the table is being added to dict_sys.)
block_lock::u_lock_try(bool for_io=true) is used in buf_flush_page()
to allow concurrent readers but no concurrent modifications while the
page is being written to the data file. That latch will be released
by buf_page_write_complete() in a different thread. Hence, we use
the special lock owner value FOR_IO.
The index_lock::u_lock() improves concurrency on operations that
involve non-leaf index pages.
The interface has been cleaned up a little. We will use
x_lock_recursive() instead of x_lock() when we know that a
lock is already held by the current thread. Similarly,
a lock upgrade from U to X is only allowed via u_x_upgrade()
or x_lock_upgraded() but not via x_lock().
We will disable the LatchDebug and sync_array interfaces to
InnoDB rw-locks.
The SEMAPHORES section of SHOW ENGINE INNODB STATUS output
will no longer include any information about InnoDB rw-locks,
only TTASEventMutex (cmake -DMUTEXTYPE=event) waits.
This will make a part of the 'innotop' script dead code.
The block_lock buf_block_t::lock will not be covered by any
PERFORMANCE_SCHEMA instrumentation.
SHOW ENGINE INNODB MUTEX and INFORMATION_SCHEMA.INNODB_MUTEXES
will no longer output source code file names or line numbers.
The dict_index_t::lock will be identified by index and table names,
which should be much more useful. PERFORMANCE_SCHEMA is lumping
information about all dict_index_t::lock together as
event_name='wait/synch/sxlock/innodb/index_tree_rw_lock'.
buf_page_free(): Remove the file,line parameters. The sux_lock will
not store such diagnostic information.
buf_block_dbg_add_level(): Define as empty macro, to be removed
in a subsequent commit.
Unless the build was configured with cmake -DPLUGIN_PERFSCHEMA=NO
the index_lock dict_index_t::lock will be instrumented via
PERFORMANCE_SCHEMA. Similar to
commit 1669c8890ca2e9092213626e5b047e58ca8b1e77
we will distinguish lock waits by registering shared_lock,exclusive_lock
events instead of try_shared_lock,try_exclusive_lock.
Actual 'try' operations will not be instrumented at all.
rw_lock_list: Remove. After MDEV-24167, this only covered
buf_block_t::lock and dict_index_t::lock. We will output their
information by traversing buf_pool or dict_sys.
2020-12-03 14:18:51 +01:00
|
|
|
rw_lock->u_unlock(true);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
|
|
|
|
bpage->set_io_fix(BUF_IO_NONE);
|
|
|
|
|
return false;
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
}
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
if (status == buf_page_t::FREED)
|
|
|
|
|
buf_release_freed_page(&block->page);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
else
|
|
|
|
|
{
|
2020-10-26 15:04:12 +01:00
|
|
|
space->reacquire();
|
2020-10-26 14:59:30 +01:00
|
|
|
ut_ad(status == buf_page_t::NORMAL || status == buf_page_t::INIT_ON_FLUSH);
|
2020-12-08 18:05:08 +01:00
|
|
|
size_t size;
|
|
|
|
|
#if defined HAVE_FALLOC_PUNCH_HOLE_AND_KEEP_SIZE || defined _WIN32
|
|
|
|
|
size_t orig_size;
|
|
|
|
|
#endif
|
2020-10-26 14:59:30 +01:00
|
|
|
IORequest::Type type= lru ? IORequest::WRITE_LRU : IORequest::WRITE_ASYNC;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
if (UNIV_UNLIKELY(!rw_lock)) /* ROW_FORMAT=COMPRESSED */
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
{
|
2020-10-26 14:59:30 +01:00
|
|
|
ut_ad(!space->full_crc32());
|
|
|
|
|
ut_ad(!space->is_compressed()); /* not page_compressed */
|
2020-12-08 18:05:08 +01:00
|
|
|
size= bpage->zip_size();
|
|
|
|
|
#if defined HAVE_FALLOC_PUNCH_HOLE_AND_KEEP_SIZE || defined _WIN32
|
|
|
|
|
orig_size= size;
|
|
|
|
|
#endif
|
2020-10-26 14:59:30 +01:00
|
|
|
buf_flush_update_zip_checksum(frame, size);
|
|
|
|
|
frame= buf_page_encrypt(space, bpage, frame, &size);
|
|
|
|
|
ut_ad(size == bpage->zip_size());
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
}
|
|
|
|
|
else
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
{
|
2020-10-26 14:59:30 +01:00
|
|
|
byte *page= block->frame;
|
2020-12-08 18:05:08 +01:00
|
|
|
size= block->physical_size();
|
|
|
|
|
#if defined HAVE_FALLOC_PUNCH_HOLE_AND_KEEP_SIZE || defined _WIN32
|
|
|
|
|
orig_size= size;
|
|
|
|
|
#endif
|
2020-10-26 14:59:30 +01:00
|
|
|
|
|
|
|
|
if (space->full_crc32())
|
|
|
|
|
{
|
|
|
|
|
/* innodb_checksum_algorithm=full_crc32 is not implemented for
|
|
|
|
|
ROW_FORMAT=COMPRESSED pages. */
|
|
|
|
|
ut_ad(!frame);
|
|
|
|
|
page= buf_page_encrypt(space, bpage, page, &size);
|
|
|
|
|
buf_flush_init_for_writing(block, page, nullptr, true);
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
buf_flush_init_for_writing(block, page, frame ? &bpage->zip : nullptr,
|
|
|
|
|
false);
|
|
|
|
|
page= buf_page_encrypt(space, bpage, frame ? frame : page, &size);
|
|
|
|
|
}
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
#if defined HAVE_FALLOC_PUNCH_HOLE_AND_KEEP_SIZE || defined _WIN32
|
2020-10-26 14:59:30 +01:00
|
|
|
if (size != orig_size && space->punch_hole)
|
|
|
|
|
type= lru ? IORequest::PUNCH_LRU : IORequest::PUNCH;
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
#endif
|
2020-10-26 14:59:30 +01:00
|
|
|
frame=page;
|
|
|
|
|
}
|
2016-08-12 10:17:45 +02:00
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
ut_ad(status == bpage->status);
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (lru)
|
|
|
|
|
buf_pool.n_flush_LRU++;
|
|
|
|
|
else
|
|
|
|
|
buf_pool.n_flush_list++;
|
2020-10-26 14:59:30 +01:00
|
|
|
if (status != buf_page_t::NORMAL || !space->use_doublewrite())
|
|
|
|
|
space->io(IORequest(type, bpage),
|
|
|
|
|
bpage->physical_offset(), size, frame, bpage);
|
|
|
|
|
else
|
2020-10-23 11:53:40 +02:00
|
|
|
buf_dblwr.add_to_batch(IORequest(bpage, space->chain.start, type), size);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
}
|
MDEV-15528 Punch holes when pages are freed
When a InnoDB data file page is freed, its contents becomes garbage,
and any storage allocated in the data file is wasted. During flushing,
InnoDB initializes the page with zeros if scrubbing is enabled. If the
tablespace is compressed then InnoDB should punch a hole else ignore the
flushing of the freed page.
buf_page_t:
- Replaced the variable file_page_was_freed, init_on_flush in buf_page_t
with status enum variable.
- Changed all debug assert of file_page_was_freed to DBUG_ASSERT
of buf_page_t::status
Removed buf_page_set_file_page_was_freed(),
buf_page_reset_file_page_was_freed().
buf_page_free(): Newly added function which takes X-lock on the page
before marking the status as FREED. So that InnoDB flush handler can
avoid concurrent flush of the freed page. Also while flushing the page,
InnoDB make sure that redo log which does freeing of the page also written
to the disk. Currently, this function only marks the page as FREED if
it is in buffer pool
buf_flush_freed_page(): Newly added function which initializes zeros
asynchorously if innodb_immediate_scrub_data_uncompressed is enabled.
Punch a hole to the file synchorously if page_compressed is enabled.
Reset the io_fix to NORMAL. Release the block from flush list and
associated mutex before writing zeros or punch a hole to the file.
buf_flush_page(): Removed the unnecessary usage of temporary
variable "flush"
fil_io(): Introduce new parameter called punch_hole. It allows fil_io()
to punch the hole to the file for the given offset.
buf_page_create(): Let the callers assign buf_page_t::status.
Every caller should eventually invoke mtr_t::init().
fsp_page_create(): Remove the unused mtr_t parameter.
In all other callers of buf_page_create() except fsp_page_create(),
before invoking mtr_t::init(), invoke
mtr_t::sx_latch_at_savepoint() or mtr_t::x_latch_at_savepoint().
mtr_t::init(): Initialize buf_page_t::status also for the temporary
tablespace (when redo logging is disabled), to avoid assertion failures.
2020-03-09 08:55:33 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
/* Increment the I/O operation count used for selecting LRU policy. */
|
|
|
|
|
buf_LRU_stat_inc_io();
|
|
|
|
|
return true;
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
/** Check whether a page can be flushed from the buf_pool.
|
|
|
|
|
@param id page identifier
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
@param fold id.fold()
|
|
|
|
|
@param lru true=buf_pool.LRU; false=buf_pool.flush_list
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
@return whether the page can be flushed */
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
static bool buf_flush_check_neighbor(const page_id_t id, ulint fold, bool lru)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_assert_owner(&buf_pool.mutex);
|
|
|
|
|
ut_ad(fold == id.fold());
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
buf_page_t *bpage= buf_pool.page_hash_get_low(id, fold);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
if (!bpage || buf_pool.watch_is_sentinel(*bpage))
|
|
|
|
|
return false;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
/* We avoid flushing 'non-old' blocks in an LRU flush, because the
|
|
|
|
|
flushed blocks are soon freed */
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
return (!lru || bpage->is_old()) && bpage->ready_for_flush();
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
/** Check which neighbors of a page can be flushed from the buf_pool.
|
|
|
|
|
@param space tablespace
|
|
|
|
|
@param id page identifier of a dirty page
|
2020-10-08 10:13:47 +02:00
|
|
|
@param contiguous whether to consider contiguous areas of pages
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
@param lru true=buf_pool.LRU; false=buf_pool.flush_list
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
@return last page number that can be flushed */
|
|
|
|
|
static page_id_t buf_flush_check_neighbors(const fil_space_t &space,
|
2020-10-08 10:13:47 +02:00
|
|
|
page_id_t &id, bool contiguous,
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
bool lru)
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
{
|
|
|
|
|
ut_ad(id.page_no() < space.size);
|
|
|
|
|
/* When flushed, dirty blocks are searched in neighborhoods of this
|
|
|
|
|
size, and flushed along with the original page. */
|
|
|
|
|
const ulint s= buf_pool.curr_size / 16;
|
|
|
|
|
const uint32_t read_ahead= buf_pool.read_ahead_area;
|
|
|
|
|
const uint32_t buf_flush_area= read_ahead > s
|
|
|
|
|
? static_cast<uint32_t>(s) : read_ahead;
|
|
|
|
|
page_id_t low= id - (id.page_no() % buf_flush_area);
|
|
|
|
|
page_id_t high= low + buf_flush_area;
|
2020-10-26 14:59:30 +01:00
|
|
|
high.set_page_no(std::min(high.page_no(), space.last_page_number()));
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
2020-10-08 10:13:47 +02:00
|
|
|
if (!contiguous)
|
|
|
|
|
{
|
|
|
|
|
high= std::max(id + 1, high);
|
|
|
|
|
id= low;
|
|
|
|
|
return high;
|
|
|
|
|
}
|
|
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
/* Determine the contiguous dirty area around id. */
|
|
|
|
|
const ulint id_fold= id.fold();
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
if (id > low)
|
|
|
|
|
{
|
|
|
|
|
ulint fold= id_fold;
|
|
|
|
|
for (page_id_t i= id - 1;; --i)
|
|
|
|
|
{
|
|
|
|
|
fold--;
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (!buf_flush_check_neighbor(i, fold, lru))
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
{
|
|
|
|
|
low= i + 1;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
if (i == low)
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
page_id_t i= id;
|
|
|
|
|
id= low;
|
|
|
|
|
ulint fold= id_fold;
|
|
|
|
|
while (++i < high)
|
|
|
|
|
{
|
|
|
|
|
++fold;
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (!buf_flush_check_neighbor(i, fold, lru))
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
break;
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
return i;
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
MY_ATTRIBUTE((nonnull))
|
2020-06-11 19:22:47 +02:00
|
|
|
/** Write punch-hole or zeroes of the freed ranges when
|
|
|
|
|
innodb_immediate_scrub_data_uncompressed from the freed ranges.
|
2020-10-26 14:59:30 +01:00
|
|
|
@param space tablespace which may contain ranges of freed pages */
|
2020-06-11 19:22:47 +02:00
|
|
|
static void buf_flush_freed_pages(fil_space_t *space)
|
|
|
|
|
{
|
2020-07-22 12:39:58 +02:00
|
|
|
const bool punch_hole= space->punch_hole;
|
|
|
|
|
if (!srv_immediate_scrub_data_uncompressed && !punch_hole)
|
2020-06-11 19:22:47 +02:00
|
|
|
return;
|
|
|
|
|
lsn_t flush_to_disk_lsn= log_sys.get_flushed_lsn();
|
|
|
|
|
|
|
|
|
|
std::unique_lock<std::mutex> freed_lock(space->freed_range_mutex);
|
|
|
|
|
if (space->freed_ranges.empty()
|
|
|
|
|
|| flush_to_disk_lsn < space->get_last_freed_lsn())
|
|
|
|
|
{
|
|
|
|
|
freed_lock.unlock();
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
range_set freed_ranges= std::move(space->freed_ranges);
|
|
|
|
|
freed_lock.unlock();
|
|
|
|
|
|
|
|
|
|
for (const auto &range : freed_ranges)
|
|
|
|
|
{
|
2020-10-26 14:59:30 +01:00
|
|
|
const ulint physical_size= space->physical_size();
|
2020-06-11 19:22:47 +02:00
|
|
|
|
|
|
|
|
if (punch_hole)
|
|
|
|
|
{
|
2020-10-26 15:04:12 +01:00
|
|
|
space->reacquire();
|
2020-10-26 14:59:30 +01:00
|
|
|
space->io(IORequest(IORequest::PUNCH_RANGE),
|
|
|
|
|
os_offset_t{range.first} * physical_size,
|
|
|
|
|
(range.last - range.first + 1) * physical_size,
|
|
|
|
|
nullptr);
|
2020-06-11 19:22:47 +02:00
|
|
|
}
|
|
|
|
|
else if (srv_immediate_scrub_data_uncompressed)
|
|
|
|
|
{
|
2020-10-26 14:59:30 +01:00
|
|
|
for (os_offset_t i= range.first; i <= range.last; i++)
|
2020-06-11 19:22:47 +02:00
|
|
|
{
|
2020-10-26 15:04:12 +01:00
|
|
|
space->reacquire();
|
2020-10-26 14:59:30 +01:00
|
|
|
space->io(IORequest(IORequest::WRITE_ASYNC),
|
|
|
|
|
i * physical_size, physical_size,
|
|
|
|
|
const_cast<byte*>(field_ref_zero));
|
2020-06-11 19:22:47 +02:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
buf_pool.stat.n_pages_written+= (range.last - range.first + 1);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** Flushes to disk all flushable pages within the flush area
|
|
|
|
|
and also write zeroes or punch the hole for the freed ranges of pages.
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
@param space tablespace
|
|
|
|
|
@param page_id page identifier
|
|
|
|
|
@param contiguous whether to consider contiguous areas of pages
|
|
|
|
|
@param lru true=buf_pool.LRU; false=buf_pool.flush_list
|
|
|
|
|
@param n_flushed number of pages flushed so far in this batch
|
|
|
|
|
@param n_to_flush maximum number of pages we are allowed to flush
|
2016-08-12 10:17:45 +02:00
|
|
|
@return number of pages flushed */
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
static ulint buf_flush_try_neighbors(fil_space_t *space,
|
|
|
|
|
const page_id_t page_id,
|
|
|
|
|
bool contiguous, bool lru,
|
|
|
|
|
ulint n_flushed, ulint n_to_flush)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ut_ad(space->id == page_id.space());
|
2020-10-08 10:13:47 +02:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ulint count= 0;
|
|
|
|
|
page_id_t id= page_id;
|
|
|
|
|
page_id_t high= buf_flush_check_neighbors(*space, id, contiguous, lru);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ut_ad(page_id >= id);
|
|
|
|
|
ut_ad(page_id < high);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
for (ulint id_fold= id.fold(); id < high && !space->is_stopping();
|
|
|
|
|
++id, ++id_fold)
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
{
|
|
|
|
|
if (count + n_flushed >= n_to_flush)
|
|
|
|
|
{
|
|
|
|
|
if (id > page_id)
|
|
|
|
|
break;
|
|
|
|
|
/* If the page whose neighbors we are flushing has not been
|
|
|
|
|
flushed yet, we must flush the page that we selected originally. */
|
|
|
|
|
id= page_id;
|
|
|
|
|
id_fold= id.fold();
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (buf_page_t *bpage= buf_pool.page_hash_get_low(id, id_fold))
|
|
|
|
|
{
|
|
|
|
|
ut_ad(bpage->in_file());
|
|
|
|
|
/* We avoid flushing 'non-old' blocks in an LRU flush,
|
|
|
|
|
because the flushed blocks are soon freed */
|
|
|
|
|
if (!lru || id == page_id || bpage->is_old())
|
|
|
|
|
{
|
2020-10-30 18:06:50 +01:00
|
|
|
if (!buf_pool.watch_is_sentinel(*bpage) &&
|
|
|
|
|
bpage->ready_for_flush() && buf_flush_page(bpage, lru, space))
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
{
|
|
|
|
|
++count;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
|
|
|
|
}
|
MDEV-17380 innodb_flush_neighbors=ON should be ignored on SSD
For tablespaces that do not reside on spinning storage, it does
not make sense to attempt to write nearby pages when writing out
dirty pages from the InnoDB buffer pool. It is actually detrimental
to performance and to the life span of flash ROM storage.
With this change, MariaDB will detect whether an InnoDB file resides
on solid-state storage. The detection has been implemented for Linux
and Microsoft Windows. For other systems, we will err on the safe side
and assume that files reside on SSD.
As part of this change, we will reduce the number of fstat() calls
when opening data files on POSIX systems and slightly clean up some
file I/O code.
FIXME: os_is_sparse_file_supported() on POSIX works in a destructive
manner. Thus, we can only invoke it when creating files, not when
opening them.
For diagnostics, we introduce the column ON_SSD to the table
INFORMATION_SCHEMA.INNODB_TABLESPACES_SCRUBBING. The table
INNODB_SYS_TABLESPACES might seem more appropriate, but its purpose
is to reflect the contents of the InnoDB system table SYS_TABLESPACES,
which we would like to remove at some point.
On Microsoft Windows, querying StorageDeviceSeekPenaltyProperty
sometimes returns ERROR_GEN_FAILURE instead of ERROR_INVALID_FUNCTION
or ERROR_NOT_SUPPORTED. We will silently ignore also this error,
and assume that the file does not reside on SSD.
On Linux, the detection will be based on the files
/sys/block/*/queue/rotational and /sys/block/*/dev.
Especially for USB storage, it is possible that
/sys/block/*/queue/rotational will wrongly report 1 instead of 0.
fil_node_t::on_ssd: Whether the InnoDB data file resides on
solid-state storage.
fil_system_t::ssd: Collection of Linux block devices that reside on
non-rotational storage.
fil_system_t::create(): Detect ssd on Linux based on the contents
of /sys/block/*/queue/rotational and /sys/block/*/dev.
fil_system_t::is_ssd(dev_t): Determine if a Linux block device is
non-rotational. Partitions will be identified with the containing
block device by assuming that the least significant 4 bits of the
minor number identify a partition, and that the "partition number"
of the entire device is 0.
2019-04-01 10:57:06 +02:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (auto n= count - 1)
|
|
|
|
|
{
|
|
|
|
|
MONITOR_INC_VALUE_CUMULATIVE(MONITOR_FLUSH_NEIGHBOR_TOTAL_PAGE,
|
|
|
|
|
MONITOR_FLUSH_NEIGHBOR_COUNT,
|
|
|
|
|
MONITOR_FLUSH_NEIGHBOR_PAGES, n);
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
return count;
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*******************************************************************//**
|
|
|
|
|
This utility moves the uncompressed frames of pages to the free list.
|
|
|
|
|
Note that this function does not actually flush any data to disk. It
|
|
|
|
|
just detaches the uncompressed frames from the compressed pages at the
|
|
|
|
|
tail of the unzip_LRU and puts those freed frames in the free list.
|
|
|
|
|
Note that it is a best effort attempt and it is not guaranteed that
|
|
|
|
|
after a call to this function there will be 'max' blocks in the free
|
|
|
|
|
list.
|
2020-02-12 13:45:21 +01:00
|
|
|
@param[in] max desired number of blocks in the free_list
|
2014-02-26 19:11:54 +01:00
|
|
|
@return number of blocks moved to the free list. */
|
2020-02-12 13:45:21 +01:00
|
|
|
static ulint buf_free_from_unzip_LRU_list_batch(ulint max)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
|
|
|
|
ulint scanned = 0;
|
|
|
|
|
ulint count = 0;
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_assert_owner(&buf_pool.mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-03-18 20:48:00 +01:00
|
|
|
buf_block_t* block = UT_LIST_GET_LAST(buf_pool.unzip_LRU);
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
while (block
|
2016-08-12 10:17:45 +02:00
|
|
|
&& count < max
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
&& UT_LIST_GET_LEN(buf_pool.free) < srv_LRU_scan_depth
|
|
|
|
|
&& UT_LIST_GET_LEN(buf_pool.unzip_LRU)
|
|
|
|
|
> UT_LIST_GET_LEN(buf_pool.LRU) / 10) {
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
++scanned;
|
|
|
|
|
if (buf_LRU_free_page(&block->page, false)) {
|
2020-03-18 20:48:00 +01:00
|
|
|
/* Block was freed. buf_pool.mutex potentially
|
2014-02-26 19:11:54 +01:00
|
|
|
released and reacquired */
|
|
|
|
|
++count;
|
2020-03-18 20:48:00 +01:00
|
|
|
block = UT_LIST_GET_LAST(buf_pool.unzip_LRU);
|
2014-02-26 19:11:54 +01:00
|
|
|
} else {
|
|
|
|
|
block = UT_LIST_GET_PREV(unzip_LRU, block);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_assert_owner(&buf_pool.mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
if (scanned) {
|
|
|
|
|
MONITOR_INC_VALUE_CUMULATIVE(
|
|
|
|
|
MONITOR_LRU_BATCH_SCANNED,
|
|
|
|
|
MONITOR_LRU_BATCH_SCANNED_NUM_CALL,
|
|
|
|
|
MONITOR_LRU_BATCH_SCANNED_PER_CALL,
|
|
|
|
|
scanned);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return(count);
|
|
|
|
|
}
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
/** Start writing out pages for a tablespace.
|
|
|
|
|
@param id tablespace identifier
|
|
|
|
|
@return tablespace
|
|
|
|
|
@retval nullptr if the pages for this tablespace should be discarded */
|
|
|
|
|
static fil_space_t *buf_flush_space(const uint32_t id)
|
|
|
|
|
{
|
2020-10-26 15:04:12 +01:00
|
|
|
fil_space_t *space= fil_space_t::get(id);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (space)
|
|
|
|
|
buf_flush_freed_pages(space);
|
|
|
|
|
return space;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct flush_counters_t
|
|
|
|
|
{
|
|
|
|
|
/** number of dirty pages flushed */
|
|
|
|
|
ulint flushed;
|
|
|
|
|
/** number of clean pages evicted */
|
|
|
|
|
ulint evicted;
|
|
|
|
|
};
|
2016-09-06 08:43:16 +02:00
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
/** Try to discard a dirty page.
|
|
|
|
|
@param bpage dirty page whose tablespace is not accessible */
|
|
|
|
|
static void buf_flush_discard_page(buf_page_t *bpage)
|
|
|
|
|
{
|
|
|
|
|
mysql_mutex_assert_owner(&buf_pool.mutex);
|
|
|
|
|
mysql_mutex_assert_not_owner(&buf_pool.flush_list_mutex);
|
|
|
|
|
ut_ad(bpage->in_file());
|
|
|
|
|
ut_ad(bpage->oldest_modification());
|
|
|
|
|
|
MDEV-24142: Replace InnoDB rw_lock_t with sux_lock
InnoDB buffer pool block and index tree latches depend on a
special kind of read-update-write lock that allows reentrant
(recursive) acquisition of the 'update' and 'write' locks
as well as an upgrade from 'update' lock to 'write' lock.
The 'update' lock allows any number of reader locks from
other threads, but no concurrent 'update' or 'write' lock.
If there were no requirement to support an upgrade from 'update'
to 'write', we could compose the lock out of two srw_lock
(implemented as any type of native rw-lock, such as SRWLOCK on
Microsoft Windows). Removing this requirement is very difficult,
so in commit f7e7f487d4b06695f91f6fbeb0396b9d87fc7bbf we
implemented an 'update' mode to our srw_lock.
Re-entrant or recursive locking is mostly needed when writing or
freeing BLOB pages, but also in crash recovery or when merging
buffered changes to an index page. The re-entrancy allows us to
attach a previously acquired page to a sub-mini-transaction that
will be committed before whatever else is holding the page latch.
The SUX lock supports Shared ('read'), Update, and eXclusive ('write')
locking modes. The S latches are not re-entrant, but a single S latch
may be acquired even if the thread already holds an U latch.
The idea of the U latch is to allow a write of something that concurrent
readers do not care about (such as the contents of BTR_SEG_LEAF,
BTR_SEG_TOP and other page allocation metadata structures, or
the MDEV-6076 PAGE_ROOT_AUTO_INC). (The PAGE_ROOT_AUTO_INC field
is only updated when a dict_table_t for the table exists, and only
read when a dict_table_t for the table is being added to dict_sys.)
block_lock::u_lock_try(bool for_io=true) is used in buf_flush_page()
to allow concurrent readers but no concurrent modifications while the
page is being written to the data file. That latch will be released
by buf_page_write_complete() in a different thread. Hence, we use
the special lock owner value FOR_IO.
The index_lock::u_lock() improves concurrency on operations that
involve non-leaf index pages.
The interface has been cleaned up a little. We will use
x_lock_recursive() instead of x_lock() when we know that a
lock is already held by the current thread. Similarly,
a lock upgrade from U to X is only allowed via u_x_upgrade()
or x_lock_upgraded() but not via x_lock().
We will disable the LatchDebug and sync_array interfaces to
InnoDB rw-locks.
The SEMAPHORES section of SHOW ENGINE INNODB STATUS output
will no longer include any information about InnoDB rw-locks,
only TTASEventMutex (cmake -DMUTEXTYPE=event) waits.
This will make a part of the 'innotop' script dead code.
The block_lock buf_block_t::lock will not be covered by any
PERFORMANCE_SCHEMA instrumentation.
SHOW ENGINE INNODB MUTEX and INFORMATION_SCHEMA.INNODB_MUTEXES
will no longer output source code file names or line numbers.
The dict_index_t::lock will be identified by index and table names,
which should be much more useful. PERFORMANCE_SCHEMA is lumping
information about all dict_index_t::lock together as
event_name='wait/synch/sxlock/innodb/index_tree_rw_lock'.
buf_page_free(): Remove the file,line parameters. The sux_lock will
not store such diagnostic information.
buf_block_dbg_add_level(): Define as empty macro, to be removed
in a subsequent commit.
Unless the build was configured with cmake -DPLUGIN_PERFSCHEMA=NO
the index_lock dict_index_t::lock will be instrumented via
PERFORMANCE_SCHEMA. Similar to
commit 1669c8890ca2e9092213626e5b047e58ca8b1e77
we will distinguish lock waits by registering shared_lock,exclusive_lock
events instead of try_shared_lock,try_exclusive_lock.
Actual 'try' operations will not be instrumented at all.
rw_lock_list: Remove. After MDEV-24167, this only covered
buf_block_t::lock and dict_index_t::lock. We will output their
information by traversing buf_pool or dict_sys.
2020-12-03 14:18:51 +01:00
|
|
|
block_lock *rw_lock;
|
2020-10-26 14:59:30 +01:00
|
|
|
|
|
|
|
|
if (bpage->state() != BUF_BLOCK_FILE_PAGE)
|
|
|
|
|
rw_lock= nullptr;
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
rw_lock= &reinterpret_cast<buf_block_t*>(bpage)->lock;
|
MDEV-24142: Replace InnoDB rw_lock_t with sux_lock
InnoDB buffer pool block and index tree latches depend on a
special kind of read-update-write lock that allows reentrant
(recursive) acquisition of the 'update' and 'write' locks
as well as an upgrade from 'update' lock to 'write' lock.
The 'update' lock allows any number of reader locks from
other threads, but no concurrent 'update' or 'write' lock.
If there were no requirement to support an upgrade from 'update'
to 'write', we could compose the lock out of two srw_lock
(implemented as any type of native rw-lock, such as SRWLOCK on
Microsoft Windows). Removing this requirement is very difficult,
so in commit f7e7f487d4b06695f91f6fbeb0396b9d87fc7bbf we
implemented an 'update' mode to our srw_lock.
Re-entrant or recursive locking is mostly needed when writing or
freeing BLOB pages, but also in crash recovery or when merging
buffered changes to an index page. The re-entrancy allows us to
attach a previously acquired page to a sub-mini-transaction that
will be committed before whatever else is holding the page latch.
The SUX lock supports Shared ('read'), Update, and eXclusive ('write')
locking modes. The S latches are not re-entrant, but a single S latch
may be acquired even if the thread already holds an U latch.
The idea of the U latch is to allow a write of something that concurrent
readers do not care about (such as the contents of BTR_SEG_LEAF,
BTR_SEG_TOP and other page allocation metadata structures, or
the MDEV-6076 PAGE_ROOT_AUTO_INC). (The PAGE_ROOT_AUTO_INC field
is only updated when a dict_table_t for the table exists, and only
read when a dict_table_t for the table is being added to dict_sys.)
block_lock::u_lock_try(bool for_io=true) is used in buf_flush_page()
to allow concurrent readers but no concurrent modifications while the
page is being written to the data file. That latch will be released
by buf_page_write_complete() in a different thread. Hence, we use
the special lock owner value FOR_IO.
The index_lock::u_lock() improves concurrency on operations that
involve non-leaf index pages.
The interface has been cleaned up a little. We will use
x_lock_recursive() instead of x_lock() when we know that a
lock is already held by the current thread. Similarly,
a lock upgrade from U to X is only allowed via u_x_upgrade()
or x_lock_upgraded() but not via x_lock().
We will disable the LatchDebug and sync_array interfaces to
InnoDB rw-locks.
The SEMAPHORES section of SHOW ENGINE INNODB STATUS output
will no longer include any information about InnoDB rw-locks,
only TTASEventMutex (cmake -DMUTEXTYPE=event) waits.
This will make a part of the 'innotop' script dead code.
The block_lock buf_block_t::lock will not be covered by any
PERFORMANCE_SCHEMA instrumentation.
SHOW ENGINE INNODB MUTEX and INFORMATION_SCHEMA.INNODB_MUTEXES
will no longer output source code file names or line numbers.
The dict_index_t::lock will be identified by index and table names,
which should be much more useful. PERFORMANCE_SCHEMA is lumping
information about all dict_index_t::lock together as
event_name='wait/synch/sxlock/innodb/index_tree_rw_lock'.
buf_page_free(): Remove the file,line parameters. The sux_lock will
not store such diagnostic information.
buf_block_dbg_add_level(): Define as empty macro, to be removed
in a subsequent commit.
Unless the build was configured with cmake -DPLUGIN_PERFSCHEMA=NO
the index_lock dict_index_t::lock will be instrumented via
PERFORMANCE_SCHEMA. Similar to
commit 1669c8890ca2e9092213626e5b047e58ca8b1e77
we will distinguish lock waits by registering shared_lock,exclusive_lock
events instead of try_shared_lock,try_exclusive_lock.
Actual 'try' operations will not be instrumented at all.
rw_lock_list: Remove. After MDEV-24167, this only covered
buf_block_t::lock and dict_index_t::lock. We will output their
information by traversing buf_pool or dict_sys.
2020-12-03 14:18:51 +01:00
|
|
|
if (!rw_lock->u_lock_try(false))
|
2020-10-26 14:59:30 +01:00
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
bpage->status= buf_page_t::NORMAL;
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
buf_flush_remove(bpage);
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
|
|
|
|
|
if (rw_lock)
|
MDEV-24142: Replace InnoDB rw_lock_t with sux_lock
InnoDB buffer pool block and index tree latches depend on a
special kind of read-update-write lock that allows reentrant
(recursive) acquisition of the 'update' and 'write' locks
as well as an upgrade from 'update' lock to 'write' lock.
The 'update' lock allows any number of reader locks from
other threads, but no concurrent 'update' or 'write' lock.
If there were no requirement to support an upgrade from 'update'
to 'write', we could compose the lock out of two srw_lock
(implemented as any type of native rw-lock, such as SRWLOCK on
Microsoft Windows). Removing this requirement is very difficult,
so in commit f7e7f487d4b06695f91f6fbeb0396b9d87fc7bbf we
implemented an 'update' mode to our srw_lock.
Re-entrant or recursive locking is mostly needed when writing or
freeing BLOB pages, but also in crash recovery or when merging
buffered changes to an index page. The re-entrancy allows us to
attach a previously acquired page to a sub-mini-transaction that
will be committed before whatever else is holding the page latch.
The SUX lock supports Shared ('read'), Update, and eXclusive ('write')
locking modes. The S latches are not re-entrant, but a single S latch
may be acquired even if the thread already holds an U latch.
The idea of the U latch is to allow a write of something that concurrent
readers do not care about (such as the contents of BTR_SEG_LEAF,
BTR_SEG_TOP and other page allocation metadata structures, or
the MDEV-6076 PAGE_ROOT_AUTO_INC). (The PAGE_ROOT_AUTO_INC field
is only updated when a dict_table_t for the table exists, and only
read when a dict_table_t for the table is being added to dict_sys.)
block_lock::u_lock_try(bool for_io=true) is used in buf_flush_page()
to allow concurrent readers but no concurrent modifications while the
page is being written to the data file. That latch will be released
by buf_page_write_complete() in a different thread. Hence, we use
the special lock owner value FOR_IO.
The index_lock::u_lock() improves concurrency on operations that
involve non-leaf index pages.
The interface has been cleaned up a little. We will use
x_lock_recursive() instead of x_lock() when we know that a
lock is already held by the current thread. Similarly,
a lock upgrade from U to X is only allowed via u_x_upgrade()
or x_lock_upgraded() but not via x_lock().
We will disable the LatchDebug and sync_array interfaces to
InnoDB rw-locks.
The SEMAPHORES section of SHOW ENGINE INNODB STATUS output
will no longer include any information about InnoDB rw-locks,
only TTASEventMutex (cmake -DMUTEXTYPE=event) waits.
This will make a part of the 'innotop' script dead code.
The block_lock buf_block_t::lock will not be covered by any
PERFORMANCE_SCHEMA instrumentation.
SHOW ENGINE INNODB MUTEX and INFORMATION_SCHEMA.INNODB_MUTEXES
will no longer output source code file names or line numbers.
The dict_index_t::lock will be identified by index and table names,
which should be much more useful. PERFORMANCE_SCHEMA is lumping
information about all dict_index_t::lock together as
event_name='wait/synch/sxlock/innodb/index_tree_rw_lock'.
buf_page_free(): Remove the file,line parameters. The sux_lock will
not store such diagnostic information.
buf_block_dbg_add_level(): Define as empty macro, to be removed
in a subsequent commit.
Unless the build was configured with cmake -DPLUGIN_PERFSCHEMA=NO
the index_lock dict_index_t::lock will be instrumented via
PERFORMANCE_SCHEMA. Similar to
commit 1669c8890ca2e9092213626e5b047e58ca8b1e77
we will distinguish lock waits by registering shared_lock,exclusive_lock
events instead of try_shared_lock,try_exclusive_lock.
Actual 'try' operations will not be instrumented at all.
rw_lock_list: Remove. After MDEV-24167, this only covered
buf_block_t::lock and dict_index_t::lock. We will output their
information by traversing buf_pool or dict_sys.
2020-12-03 14:18:51 +01:00
|
|
|
rw_lock->u_unlock();
|
2020-10-26 14:59:30 +01:00
|
|
|
|
|
|
|
|
buf_LRU_free_page(bpage, true);
|
|
|
|
|
}
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
/** Flush dirty blocks from the end of the LRU list.
|
|
|
|
|
@param max maximum number of blocks to make available in buf_pool.free
|
|
|
|
|
@param n counts of flushed and evicted pages */
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
static void buf_flush_LRU_list_batch(ulint max, flush_counters_t *n)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ulint scanned= 0;
|
|
|
|
|
ulint free_limit= srv_LRU_scan_depth;
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
|
|
|
|
|
mysql_mutex_assert_owner(&buf_pool.mutex);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
if (buf_pool.withdraw_target && buf_pool.curr_size < buf_pool.old_size)
|
|
|
|
|
free_limit+= buf_pool.withdraw_target - UT_LIST_GET_LEN(buf_pool.withdraw);
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
const auto neighbors= UT_LIST_GET_LEN(buf_pool.LRU) < BUF_LRU_OLD_MIN_LEN
|
|
|
|
|
? 0 : srv_flush_neighbors;
|
|
|
|
|
fil_space_t *space= nullptr;
|
2020-10-26 14:59:30 +01:00
|
|
|
uint32_t last_space_id= FIL_NULL;
|
|
|
|
|
static_assert(FIL_NULL > SRV_TMP_SPACE_ID, "consistency");
|
|
|
|
|
static_assert(FIL_NULL > SRV_SPACE_ID_UPPER_BOUND, "consistency");
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
for (buf_page_t *bpage= UT_LIST_GET_LAST(buf_pool.LRU);
|
|
|
|
|
bpage && n->flushed + n->evicted < max &&
|
|
|
|
|
UT_LIST_GET_LEN(buf_pool.LRU) > BUF_LRU_MIN_LEN &&
|
|
|
|
|
UT_LIST_GET_LEN(buf_pool.free) < free_limit;
|
|
|
|
|
++scanned, bpage= buf_pool.lru_hp.get())
|
|
|
|
|
{
|
|
|
|
|
buf_page_t *prev= UT_LIST_GET_PREV(LRU, bpage);
|
|
|
|
|
buf_pool.lru_hp.set(prev);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
if (bpage->ready_for_replace())
|
|
|
|
|
{
|
|
|
|
|
/* block is ready for eviction i.e., it is clean and is not
|
|
|
|
|
IO-fixed or buffer fixed. */
|
|
|
|
|
if (buf_LRU_free_page(bpage, true))
|
|
|
|
|
++n->evicted;
|
|
|
|
|
}
|
|
|
|
|
else if (bpage->ready_for_flush())
|
|
|
|
|
{
|
|
|
|
|
/* Block is ready for flush. Dispatch an IO request. The IO
|
|
|
|
|
helper thread will put it on free list in IO completion routine. */
|
|
|
|
|
const page_id_t page_id(bpage->id());
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
const uint32_t space_id= page_id.space();
|
|
|
|
|
if (!space || space->id != space_id)
|
|
|
|
|
{
|
2020-10-26 14:59:30 +01:00
|
|
|
if (last_space_id != space_id)
|
|
|
|
|
{
|
|
|
|
|
if (space)
|
2020-10-26 15:04:12 +01:00
|
|
|
space->release();
|
2020-10-26 14:59:30 +01:00
|
|
|
space= buf_flush_space(space_id);
|
|
|
|
|
last_space_id= space_id;
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
ut_ad(!space);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
}
|
2020-10-26 14:59:30 +01:00
|
|
|
else if (space->is_stopping())
|
|
|
|
|
{
|
2020-10-26 15:04:12 +01:00
|
|
|
space->release();
|
2020-10-26 14:59:30 +01:00
|
|
|
space= nullptr;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!space)
|
|
|
|
|
buf_flush_discard_page(bpage);
|
|
|
|
|
else if (neighbors && space->is_rotational())
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
{
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
|
|
|
|
n->flushed+= buf_flush_try_neighbors(space, page_id, neighbors == 1,
|
|
|
|
|
true, n->flushed, max);
|
|
|
|
|
reacquire_mutex:
|
|
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
|
|
|
|
}
|
|
|
|
|
else if (buf_flush_page(bpage, true, space))
|
|
|
|
|
{
|
|
|
|
|
++n->flushed;
|
|
|
|
|
goto reacquire_mutex;
|
|
|
|
|
}
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
/* Can't evict or dispatch this block. Go to previous. */
|
|
|
|
|
ut_ad(buf_pool.lru_hp.is_hp(prev));
|
|
|
|
|
}
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
buf_pool.lru_hp.set(nullptr);
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (space)
|
2020-10-26 15:04:12 +01:00
|
|
|
space->release();
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
/* We keep track of all flushes happening as part of LRU flush. When
|
|
|
|
|
estimating the desired rate at which flush_list should be flushed,
|
|
|
|
|
we factor in this value. */
|
|
|
|
|
buf_lru_flush_page_count+= n->flushed;
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_assert_owner(&buf_pool.mutex);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
|
|
|
|
if (scanned)
|
|
|
|
|
MONITOR_INC_VALUE_CUMULATIVE(MONITOR_LRU_BATCH_SCANNED,
|
|
|
|
|
MONITOR_LRU_BATCH_SCANNED_NUM_CALL,
|
|
|
|
|
MONITOR_LRU_BATCH_SCANNED_PER_CALL,
|
|
|
|
|
scanned);
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
2020-02-12 13:45:21 +01:00
|
|
|
/** Flush and move pages from LRU or unzip_LRU list to the free list.
|
|
|
|
|
Whether LRU or unzip_LRU is used depends on the state of the system.
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
@param max maximum number of blocks to make available in buf_pool.free
|
|
|
|
|
@return number of flushed pages */
|
|
|
|
|
static ulint buf_do_LRU_batch(ulint max)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
const ulint n_unzip_LRU_evicted= buf_LRU_evict_from_unzip_LRU()
|
|
|
|
|
? buf_free_from_unzip_LRU_list_batch(max)
|
|
|
|
|
: 0;
|
|
|
|
|
flush_counters_t n;
|
|
|
|
|
n.flushed= 0;
|
|
|
|
|
n.evicted= n_unzip_LRU_evicted;
|
|
|
|
|
buf_flush_LRU_list_batch(max, &n);
|
|
|
|
|
|
|
|
|
|
if (const ulint evicted= n.evicted - n_unzip_LRU_evicted)
|
|
|
|
|
{
|
|
|
|
|
MONITOR_INC_VALUE_CUMULATIVE(MONITOR_LRU_BATCH_EVICT_TOTAL_PAGE,
|
|
|
|
|
MONITOR_LRU_BATCH_EVICT_COUNT,
|
|
|
|
|
MONITOR_LRU_BATCH_EVICT_PAGES,
|
|
|
|
|
evicted);
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
return n.flushed;
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
2016-08-12 10:17:45 +02:00
|
|
|
/** This utility flushes dirty blocks from the end of the flush_list.
|
|
|
|
|
The calling thread is not allowed to own any latches on pages!
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
@param max_n maximum mumber of blocks to flush
|
|
|
|
|
@param lsn once an oldest_modification>=lsn is found, terminate the batch
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
@return number of blocks for which the write request was queued */
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
static ulint buf_do_flush_list_batch(ulint max_n, lsn_t lsn)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ulint count= 0;
|
|
|
|
|
ulint scanned= 0;
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_assert_owner(&buf_pool.mutex);
|
|
|
|
|
|
|
|
|
|
const auto neighbors= UT_LIST_GET_LEN(buf_pool.LRU) < BUF_LRU_OLD_MIN_LEN
|
|
|
|
|
? 0 : srv_flush_neighbors;
|
|
|
|
|
fil_space_t *space= nullptr;
|
2020-10-26 14:59:30 +01:00
|
|
|
uint32_t last_space_id= FIL_NULL;
|
|
|
|
|
static_assert(FIL_NULL > SRV_TMP_SPACE_ID, "consistency");
|
|
|
|
|
static_assert(FIL_NULL > SRV_SPACE_ID_UPPER_BOUND, "consistency");
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
|
|
|
|
/* Start from the end of the list looking for a suitable block to be
|
|
|
|
|
flushed. */
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
ulint len= UT_LIST_GET_LEN(buf_pool.flush_list);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
|
|
|
|
/* In order not to degenerate this scan to O(n*n) we attempt to
|
|
|
|
|
preserve pointer of previous block in the flush list. To do so we
|
|
|
|
|
declare it a hazard pointer. Any thread working on the flush list
|
|
|
|
|
must check the hazard pointer and if it is removing the same block
|
|
|
|
|
then it must reset it. */
|
|
|
|
|
for (buf_page_t *bpage= UT_LIST_GET_LAST(buf_pool.flush_list);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
bpage && len && count < max_n;
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
bpage= buf_pool.flush_hp.get(), ++scanned, len--)
|
|
|
|
|
{
|
|
|
|
|
const lsn_t oldest_modification= bpage->oldest_modification();
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (oldest_modification >= lsn)
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
break;
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ut_ad(oldest_modification);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
buf_page_t *prev= UT_LIST_GET_PREV(list, bpage);
|
|
|
|
|
buf_pool.flush_hp.set(prev);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ut_ad(bpage->in_file());
|
|
|
|
|
const bool flushed= bpage->ready_for_flush();
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
if (flushed)
|
|
|
|
|
{
|
|
|
|
|
const page_id_t page_id(bpage->id());
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
const uint32_t space_id= page_id.space();
|
|
|
|
|
if (!space || space->id != space_id)
|
|
|
|
|
{
|
2020-10-26 14:59:30 +01:00
|
|
|
if (last_space_id != space_id)
|
|
|
|
|
{
|
|
|
|
|
if (space)
|
2020-10-26 15:04:12 +01:00
|
|
|
space->release();
|
2020-10-26 14:59:30 +01:00
|
|
|
space= buf_flush_space(space_id);
|
|
|
|
|
last_space_id= space_id;
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
ut_ad(!space);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
}
|
2020-10-26 14:59:30 +01:00
|
|
|
else if (space->is_stopping())
|
|
|
|
|
{
|
2020-10-26 15:04:12 +01:00
|
|
|
space->release();
|
2020-10-26 14:59:30 +01:00
|
|
|
space= nullptr;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!space)
|
|
|
|
|
buf_flush_discard_page(bpage);
|
|
|
|
|
else if (neighbors && space->is_rotational())
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
{
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
|
|
|
|
count+= buf_flush_try_neighbors(space, page_id, neighbors == 1,
|
2020-10-26 14:59:30 +01:00
|
|
|
false, count, max_n);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
reacquire_mutex:
|
|
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
|
|
|
|
}
|
|
|
|
|
else if (buf_flush_page(bpage, false, space))
|
|
|
|
|
{
|
|
|
|
|
++count;
|
|
|
|
|
goto reacquire_mutex;
|
|
|
|
|
}
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
ut_ad(flushed || buf_pool.flush_hp.is_hp(prev));
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
buf_pool.flush_hp.set(nullptr);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
|
|
|
|
|
if (space)
|
2020-10-26 15:04:12 +01:00
|
|
|
space->release();
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
|
|
|
|
if (scanned)
|
|
|
|
|
MONITOR_INC_VALUE_CUMULATIVE(MONITOR_FLUSH_BATCH_SCANNED,
|
|
|
|
|
MONITOR_FLUSH_BATCH_SCANNED_NUM_CALL,
|
|
|
|
|
MONITOR_FLUSH_BATCH_SCANNED_PER_CALL,
|
|
|
|
|
scanned);
|
|
|
|
|
if (count)
|
|
|
|
|
MONITOR_INC_VALUE_CUMULATIVE(MONITOR_FLUSH_BATCH_TOTAL_PAGE,
|
|
|
|
|
MONITOR_FLUSH_BATCH_COUNT,
|
|
|
|
|
MONITOR_FLUSH_BATCH_PAGES,
|
|
|
|
|
count);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_assert_owner(&buf_pool.mutex);
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
return count;
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
/** Wait until a flush batch ends.
|
|
|
|
|
@param lru true=buf_pool.LRU; false=buf_pool.flush_list */
|
|
|
|
|
void buf_flush_wait_batch_end(bool lru)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
const auto &n_flush= lru ? buf_pool.n_flush_LRU : buf_pool.n_flush_list;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (n_flush)
|
|
|
|
|
{
|
|
|
|
|
auto cond= lru ? &buf_pool.done_flush_LRU : &buf_pool.done_flush_list;
|
|
|
|
|
tpool::tpool_wait_begin();
|
|
|
|
|
thd_wait_begin(nullptr, THD_WAIT_DISKIO);
|
|
|
|
|
do
|
2021-02-07 11:19:24 +01:00
|
|
|
my_cond_wait(cond, &buf_pool.mutex.m_mutex);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
while (n_flush);
|
|
|
|
|
tpool::tpool_wait_end();
|
|
|
|
|
thd_wait_end(nullptr);
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_broadcast(cond);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
/** Whether a background log flush is pending */
|
|
|
|
|
static std::atomic_flag log_flush_pending;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
/** Advance log_sys.get_flushed_lsn() */
|
|
|
|
|
static void log_flush(void *)
|
|
|
|
|
{
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
/* Between batches, we try to prevent I/O stalls by these calls.
|
|
|
|
|
This should not be needed for correctness. */
|
|
|
|
|
os_aio_wait_until_no_pending_writes();
|
|
|
|
|
fil_flush_file_spaces();
|
|
|
|
|
|
|
|
|
|
/* Guarantee progress for buf_flush_lists(). */
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
log_write_up_to(log_sys.get_lsn(), true);
|
|
|
|
|
log_flush_pending.clear();
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
static tpool::waitable_task log_flush_task(log_flush, nullptr, nullptr);
|
|
|
|
|
|
|
|
|
|
/** Write out dirty blocks from buf_pool.flush_list.
|
|
|
|
|
@param max_n wished maximum mumber of blocks flushed
|
|
|
|
|
@param lsn buf_pool.get_oldest_modification(LSN_MAX) target (0=LRU flush)
|
|
|
|
|
@return the number of processed pages
|
|
|
|
|
@retval 0 if a batch of the same type (lsn==0 or lsn!=0) is already running */
|
|
|
|
|
ulint buf_flush_lists(ulint max_n, lsn_t lsn)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
auto &n_flush= lsn ? buf_pool.n_flush_list : buf_pool.n_flush_LRU;
|
|
|
|
|
|
|
|
|
|
if (n_flush)
|
|
|
|
|
return 0;
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (log_sys.get_lsn() > log_sys.get_flushed_lsn())
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
log_flush_task.wait();
|
|
|
|
|
if (log_sys.get_lsn() > log_sys.get_flushed_lsn() &&
|
|
|
|
|
!log_flush_pending.test_and_set())
|
|
|
|
|
srv_thread_pool->submit_task(&log_flush_task);
|
|
|
|
|
#if defined UNIV_DEBUG || defined UNIV_IBUF_DEBUG
|
|
|
|
|
if (UNIV_UNLIKELY(ibuf_debug))
|
|
|
|
|
log_write_up_to(log_sys.get_lsn(), true);
|
|
|
|
|
#endif
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
auto cond= lsn ? &buf_pool.done_flush_list : &buf_pool.done_flush_LRU;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
|
|
|
|
const bool running= n_flush != 0;
|
|
|
|
|
/* FIXME: we are performing a dirty read of buf_pool.flush_list.count
|
|
|
|
|
while not holding buf_pool.flush_list_mutex */
|
2020-12-09 08:22:13 +01:00
|
|
|
if (running || (lsn && !UT_LIST_GET_LEN(buf_pool.flush_list)))
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
{
|
2020-10-26 14:59:30 +01:00
|
|
|
if (!running)
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_broadcast(cond);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
n_flush++;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ulint n_flushed= lsn
|
|
|
|
|
? buf_do_flush_list_batch(max_n, lsn)
|
|
|
|
|
: buf_do_LRU_batch(max_n);
|
|
|
|
|
|
|
|
|
|
const auto n_flushing= --n_flush;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
buf_pool.try_LRU_scan= true;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (!n_flushing)
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_broadcast(cond);
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
buf_dblwr.flush_buffered_writes();
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
DBUG_PRINT("ib_buf", ("%s completed, " ULINTPF " pages",
|
|
|
|
|
lsn ? "flush_list" : "LRU flush", n_flushed));
|
|
|
|
|
return n_flushed;
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
|
|
|
|
|
/** Initiate a log checkpoint, discarding the start of the log.
|
|
|
|
|
@param oldest_lsn the checkpoint LSN
|
|
|
|
|
@param end_lsn log_sys.get_lsn()
|
|
|
|
|
@return true if success, false if a checkpoint write was already running */
|
|
|
|
|
static bool log_checkpoint_low(lsn_t oldest_lsn, lsn_t end_lsn)
|
2016-08-12 10:17:45 +02:00
|
|
|
{
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ut_ad(!srv_read_only_mode);
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_assert_owner(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ut_ad(oldest_lsn <= end_lsn);
|
|
|
|
|
ut_ad(end_lsn == log_sys.get_lsn());
|
|
|
|
|
ut_ad(!recv_no_log_write);
|
|
|
|
|
|
|
|
|
|
ut_ad(oldest_lsn >= log_sys.last_checkpoint_lsn);
|
|
|
|
|
|
|
|
|
|
if (oldest_lsn > log_sys.last_checkpoint_lsn + SIZE_OF_FILE_CHECKPOINT)
|
|
|
|
|
/* Some log has been written since the previous checkpoint. */;
|
|
|
|
|
else if (srv_shutdown_state > SRV_SHUTDOWN_INITIATED)
|
|
|
|
|
/* MariaDB startup expects the redo log file to be logically empty
|
|
|
|
|
(not even containing a FILE_CHECKPOINT record) after a clean shutdown.
|
|
|
|
|
Perform an extra checkpoint at shutdown. */;
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
/* Do nothing, because nothing was logged (other than a
|
|
|
|
|
FILE_CHECKPOINT record) since the previous checkpoint. */
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_unlock(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
return true;
|
|
|
|
|
}
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
/* Repeat the FILE_MODIFY records after the checkpoint, in case some
|
|
|
|
|
log records between the checkpoint and log_sys.lsn need them.
|
|
|
|
|
Finally, write a FILE_CHECKPOINT record. Redo log apply expects to
|
|
|
|
|
see a FILE_CHECKPOINT after the checkpoint, except on clean
|
|
|
|
|
shutdown, where the log will be empty after the checkpoint.
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
It is important that we write out the redo log before any further
|
|
|
|
|
dirty pages are flushed to the tablespace files. At this point,
|
|
|
|
|
because we hold log_sys.mutex, mtr_t::commit() in other threads will
|
|
|
|
|
be blocked, and no pages can be added to the flush lists. */
|
|
|
|
|
lsn_t flush_lsn= oldest_lsn;
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (fil_names_clear(flush_lsn, oldest_lsn != end_lsn ||
|
|
|
|
|
srv_shutdown_state <= SRV_SHUTDOWN_INITIATED))
|
|
|
|
|
{
|
|
|
|
|
flush_lsn= log_sys.get_lsn();
|
|
|
|
|
ut_ad(flush_lsn >= end_lsn + SIZE_OF_FILE_CHECKPOINT);
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_unlock(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
log_write_up_to(flush_lsn, true, true);
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_lock(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (log_sys.last_checkpoint_lsn >= oldest_lsn)
|
|
|
|
|
{
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_unlock(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
ut_ad(oldest_lsn >= log_sys.last_checkpoint_lsn);
|
|
|
|
|
|
|
|
|
|
ut_ad(log_sys.get_flushed_lsn() >= flush_lsn);
|
|
|
|
|
|
|
|
|
|
if (log_sys.n_pending_checkpoint_writes)
|
|
|
|
|
{
|
|
|
|
|
/* A checkpoint write is running */
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_unlock(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
return false;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
log_sys.next_checkpoint_lsn= oldest_lsn;
|
|
|
|
|
log_write_checkpoint_info(end_lsn);
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_assert_not_owner(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
|
|
|
|
|
return true;
|
2016-08-12 10:17:45 +02:00
|
|
|
}
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
/** Make a checkpoint. Note that this function does not flush dirty
|
|
|
|
|
blocks from the buffer pool: it only checks what is lsn of the oldest
|
|
|
|
|
modification in the pool, and writes information about the lsn in
|
|
|
|
|
log file. Use log_make_checkpoint() to flush also the pool.
|
|
|
|
|
@retval true if the checkpoint was or had been made
|
|
|
|
|
@retval false if a checkpoint write was already running */
|
|
|
|
|
static bool log_checkpoint()
|
2016-08-12 10:17:45 +02:00
|
|
|
{
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (recv_recovery_is_on())
|
|
|
|
|
recv_sys.apply(true);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
switch (srv_file_flush_method) {
|
|
|
|
|
case SRV_NOSYNC:
|
|
|
|
|
case SRV_O_DIRECT_NO_FSYNC:
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
fil_flush_file_spaces();
|
|
|
|
|
}
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_lock(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
const lsn_t end_lsn= log_sys.get_lsn();
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_lock(&log_sys.flush_order_mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
const lsn_t oldest_lsn= buf_pool.get_oldest_modification(end_lsn);
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_unlock(&log_sys.flush_order_mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
return log_checkpoint_low(oldest_lsn, end_lsn);
|
|
|
|
|
}
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
/** Make a checkpoint. */
|
|
|
|
|
ATTRIBUTE_COLD void log_make_checkpoint()
|
|
|
|
|
{
|
|
|
|
|
buf_flush_wait_flushed(log_sys.get_lsn());
|
|
|
|
|
while (!log_checkpoint());
|
|
|
|
|
}
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
/** Wait until all persistent pages are flushed up to a limit.
|
|
|
|
|
@param sync_lsn buf_pool.get_oldest_modification(LSN_MAX) to wait for */
|
|
|
|
|
ATTRIBUTE_COLD void buf_flush_wait_flushed(lsn_t sync_lsn)
|
|
|
|
|
{
|
|
|
|
|
ut_ad(sync_lsn);
|
|
|
|
|
ut_ad(sync_lsn < LSN_MAX);
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_assert_not_owner(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ut_ad(!srv_read_only_mode);
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (recv_recovery_is_on())
|
|
|
|
|
recv_sys.apply(true);
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
2016-08-12 10:17:45 +02:00
|
|
|
|
2020-11-04 15:55:36 +01:00
|
|
|
if (buf_pool.get_oldest_modification(sync_lsn) < sync_lsn)
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
{
|
2020-11-04 15:55:36 +01:00
|
|
|
#if 1 /* FIXME: remove this, and guarantee that the page cleaner serves us */
|
|
|
|
|
if (UNIV_UNLIKELY(!buf_page_cleaner_is_active)
|
|
|
|
|
ut_d(|| innodb_page_cleaner_disabled_debug))
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
{
|
2020-11-04 15:55:36 +01:00
|
|
|
do
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
{
|
2020-11-04 15:55:36 +01:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
ulint n_pages= buf_flush_lists(srv_max_io_capacity, sync_lsn);
|
|
|
|
|
buf_flush_wait_batch_end_acquiring_mutex(false);
|
|
|
|
|
if (n_pages)
|
|
|
|
|
{
|
|
|
|
|
MONITOR_INC_VALUE_CUMULATIVE(MONITOR_FLUSH_SYNC_TOTAL_PAGE,
|
|
|
|
|
MONITOR_FLUSH_SYNC_COUNT,
|
|
|
|
|
MONITOR_FLUSH_SYNC_PAGES, n_pages);
|
|
|
|
|
}
|
|
|
|
|
MONITOR_INC(MONITOR_FLUSH_SYNC_WAITS);
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
}
|
2020-11-04 15:55:36 +01:00
|
|
|
while (buf_pool.get_oldest_modification(sync_lsn) < sync_lsn);
|
|
|
|
|
|
|
|
|
|
goto try_checkpoint;
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
if (buf_flush_sync_lsn < sync_lsn)
|
|
|
|
|
{
|
|
|
|
|
buf_flush_sync_lsn= sync_lsn;
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_signal(&buf_pool.do_flush_list);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
}
|
|
|
|
|
|
2020-11-04 15:55:36 +01:00
|
|
|
do
|
|
|
|
|
{
|
|
|
|
|
tpool::tpool_wait_begin();
|
|
|
|
|
thd_wait_begin(nullptr, THD_WAIT_DISKIO);
|
2021-02-07 11:19:24 +01:00
|
|
|
my_cond_wait(&buf_pool.done_flush_list,
|
|
|
|
|
&buf_pool.flush_list_mutex.m_mutex);
|
2020-11-04 15:55:36 +01:00
|
|
|
thd_wait_end(nullptr);
|
|
|
|
|
tpool::tpool_wait_end();
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
|
2020-11-04 15:55:36 +01:00
|
|
|
MONITOR_INC(MONITOR_FLUSH_SYNC_WAITS);
|
|
|
|
|
}
|
|
|
|
|
while (buf_pool.get_oldest_modification(sync_lsn) < sync_lsn);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
}
|
|
|
|
|
|
2020-11-04 15:55:36 +01:00
|
|
|
try_checkpoint:
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
2020-11-04 15:55:36 +01:00
|
|
|
|
|
|
|
|
if (UNIV_UNLIKELY(log_sys.last_checkpoint_lsn < sync_lsn))
|
|
|
|
|
log_checkpoint();
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** If innodb_flush_sync=ON, initiate a furious flush.
|
|
|
|
|
@param lsn buf_pool.get_oldest_modification(LSN_MAX) target */
|
|
|
|
|
void buf_flush_ahead(lsn_t lsn)
|
|
|
|
|
{
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_assert_not_owner(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ut_ad(!srv_read_only_mode);
|
|
|
|
|
|
|
|
|
|
if (recv_recovery_is_on())
|
|
|
|
|
recv_sys.apply(true);
|
|
|
|
|
|
2020-11-04 15:55:36 +01:00
|
|
|
if (buf_flush_sync_lsn < lsn)
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
{
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
if (buf_flush_sync_lsn < lsn)
|
|
|
|
|
{
|
|
|
|
|
buf_flush_sync_lsn= lsn;
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_signal(&buf_pool.do_flush_list);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
}
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
}
|
2016-08-12 10:17:45 +02:00
|
|
|
}
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
/** Wait for pending flushes to complete. */
|
|
|
|
|
void buf_flush_wait_batch_end_acquiring_mutex(bool lru)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
if (lru ? buf_pool.n_flush_LRU : buf_pool.n_flush_list)
|
|
|
|
|
{
|
|
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
|
|
|
|
buf_flush_wait_batch_end(lru);
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
/** Conduct checkpoint-related flushing for innodb_flush_sync=ON,
|
|
|
|
|
and try to initiate checkpoints until the target is met.
|
|
|
|
|
@param lsn minimum value of buf_pool.get_oldest_modification(LSN_MAX) */
|
|
|
|
|
ATTRIBUTE_COLD static void buf_flush_sync_for_checkpoint(lsn_t lsn)
|
|
|
|
|
{
|
|
|
|
|
ut_ad(!srv_read_only_mode);
|
|
|
|
|
|
|
|
|
|
for (;;)
|
|
|
|
|
{
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
|
|
|
|
|
if (ulint n_flushed= buf_flush_lists(srv_max_io_capacity, lsn))
|
|
|
|
|
{
|
|
|
|
|
MONITOR_INC_VALUE_CUMULATIVE(MONITOR_FLUSH_SYNC_TOTAL_PAGE,
|
|
|
|
|
MONITOR_FLUSH_SYNC_COUNT,
|
|
|
|
|
MONITOR_FLUSH_SYNC_PAGES, n_flushed);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Attempt to perform a log checkpoint upon completing each batch. */
|
|
|
|
|
if (recv_recovery_is_on())
|
|
|
|
|
recv_sys.apply(true);
|
|
|
|
|
|
|
|
|
|
switch (srv_file_flush_method) {
|
|
|
|
|
case SRV_NOSYNC:
|
|
|
|
|
case SRV_O_DIRECT_NO_FSYNC:
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
fil_flush_file_spaces();
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_lock(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
const lsn_t newest_lsn= log_sys.get_lsn();
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_lock(&log_sys.flush_order_mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
lsn_t measure= buf_pool.get_oldest_modification(0);
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_unlock(&log_sys.flush_order_mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
const lsn_t checkpoint_lsn= measure ? measure : newest_lsn;
|
|
|
|
|
|
|
|
|
|
if (checkpoint_lsn > log_sys.last_checkpoint_lsn + SIZE_OF_FILE_CHECKPOINT)
|
|
|
|
|
{
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
log_checkpoint_low(checkpoint_lsn, newest_lsn);
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
measure= buf_pool.get_oldest_modification(LSN_MAX);
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
{
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_unlock(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (!measure)
|
|
|
|
|
measure= LSN_MAX;
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-26 15:43:52 +01:00
|
|
|
mysql_mutex_assert_not_owner(&log_sys.mutex);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
|
|
|
|
|
/* After attempting log checkpoint, check if we have reached our target. */
|
|
|
|
|
const lsn_t target= buf_flush_sync_lsn;
|
|
|
|
|
|
|
|
|
|
if (measure >= target)
|
|
|
|
|
buf_flush_sync_lsn= 0;
|
|
|
|
|
|
|
|
|
|
/* wake up buf_flush_wait_flushed() */
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_broadcast(&buf_pool.done_flush_list);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
|
|
|
|
|
lsn= std::max(lsn, target);
|
|
|
|
|
|
|
|
|
|
if (measure >= lsn)
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2014-02-26 19:11:54 +01:00
|
|
|
/*********************************************************************//**
|
|
|
|
|
Calculates if flushing is required based on redo generation rate.
|
|
|
|
|
@return percent of io_capacity to flush to manage redo space */
|
|
|
|
|
static
|
|
|
|
|
ulint
|
|
|
|
|
af_get_pct_for_lsn(
|
|
|
|
|
/*===============*/
|
|
|
|
|
lsn_t age) /*!< in: current age of LSN. */
|
|
|
|
|
{
|
2020-03-10 19:05:17 +01:00
|
|
|
lsn_t af_lwm = static_cast<lsn_t>(
|
|
|
|
|
srv_adaptive_flushing_lwm
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
* static_cast<double>(log_sys.log_capacity) / 100);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
if (age < af_lwm) {
|
|
|
|
|
/* No adaptive flushing. */
|
|
|
|
|
return(0);
|
|
|
|
|
}
|
|
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
lsn_t lsn_age_factor = (age * 100) / log_sys.max_modified_age_async;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
ut_ad(srv_max_io_capacity >= srv_io_capacity);
|
2020-03-10 19:05:17 +01:00
|
|
|
return static_cast<ulint>(
|
|
|
|
|
(static_cast<double>(srv_max_io_capacity / srv_io_capacity
|
|
|
|
|
* lsn_age_factor)
|
|
|
|
|
* sqrt(static_cast<double>(lsn_age_factor))
|
|
|
|
|
/ 7.5));
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
/** This function is called approximately once every second by the
|
|
|
|
|
page_cleaner thread if innodb_adaptive_flushing=ON.
|
|
|
|
|
Based on various factors it decides if there is a need to do flushing.
|
2016-08-12 10:17:45 +02:00
|
|
|
@return number of pages recommended to be flushed
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
@param last_pages_in number of pages flushed in previous batch
|
|
|
|
|
@param oldest_lsn buf_pool.get_oldest_modification(0)
|
2020-12-08 13:57:51 +01:00
|
|
|
@param dirty_blocks UT_LIST_GET_LEN(buf_pool.flush_list)
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
@param dirty_pct 100*flush_list.count / (LRU.count + free.count) */
|
|
|
|
|
static ulint page_cleaner_flush_pages_recommendation(ulint last_pages_in,
|
|
|
|
|
lsn_t oldest_lsn,
|
2020-12-08 13:57:51 +01:00
|
|
|
ulint dirty_blocks,
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
double dirty_pct)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
|
|
|
|
static lsn_t prev_lsn = 0;
|
|
|
|
|
static ulint sum_pages = 0;
|
|
|
|
|
static ulint avg_page_rate = 0;
|
|
|
|
|
static ulint n_iterations = 0;
|
2016-08-12 10:17:45 +02:00
|
|
|
static time_t prev_time;
|
2014-02-26 19:11:54 +01:00
|
|
|
lsn_t lsn_rate;
|
|
|
|
|
ulint n_pages = 0;
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
const lsn_t cur_lsn = log_sys.get_lsn();
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ut_ad(oldest_lsn <= cur_lsn);
|
|
|
|
|
ulint pct_for_lsn = af_get_pct_for_lsn(cur_lsn - oldest_lsn);
|
|
|
|
|
time_t curr_time = time(nullptr);
|
2020-12-08 13:57:51 +01:00
|
|
|
const double max_pct = srv_max_buf_pool_modified_pct;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (!prev_lsn || !pct_for_lsn) {
|
|
|
|
|
prev_time = curr_time;
|
2014-02-26 19:11:54 +01:00
|
|
|
prev_lsn = cur_lsn;
|
2020-12-08 13:57:51 +01:00
|
|
|
if (max_pct > 0.0) {
|
|
|
|
|
dirty_pct /= max_pct;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
n_pages = ulint(dirty_pct * double(srv_io_capacity));
|
|
|
|
|
if (n_pages < dirty_blocks) {
|
|
|
|
|
n_pages= std::min<ulint>(srv_io_capacity, dirty_blocks);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return n_pages;
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
|
2016-08-12 10:17:45 +02:00
|
|
|
sum_pages += last_pages_in;
|
|
|
|
|
|
|
|
|
|
double time_elapsed = difftime(curr_time, prev_time);
|
|
|
|
|
|
2014-02-26 19:11:54 +01:00
|
|
|
/* We update our variables every srv_flushing_avg_loops
|
|
|
|
|
iterations to smooth out transition in workload. */
|
2016-08-12 10:17:45 +02:00
|
|
|
if (++n_iterations >= srv_flushing_avg_loops
|
2020-03-10 19:05:17 +01:00
|
|
|
|| time_elapsed >= static_cast<double>(srv_flushing_avg_loops)) {
|
2016-08-12 10:17:45 +02:00
|
|
|
|
|
|
|
|
if (time_elapsed < 1) {
|
|
|
|
|
time_elapsed = 1;
|
|
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2016-08-12 10:17:45 +02:00
|
|
|
avg_page_rate = static_cast<ulint>(
|
|
|
|
|
((static_cast<double>(sum_pages)
|
|
|
|
|
/ time_elapsed)
|
2020-03-10 19:05:17 +01:00
|
|
|
+ static_cast<double>(avg_page_rate)) / 2);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
/* How much LSN we have generated since last call. */
|
2016-08-12 10:17:45 +02:00
|
|
|
lsn_rate = static_cast<lsn_t>(
|
|
|
|
|
static_cast<double>(cur_lsn - prev_lsn)
|
|
|
|
|
/ time_elapsed);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
lsn_avg_rate = (lsn_avg_rate + lsn_rate) / 2;
|
|
|
|
|
|
2018-01-24 09:46:04 +01:00
|
|
|
ulint flush_tm = page_cleaner.flush_time;
|
|
|
|
|
ulint flush_pass = page_cleaner.flush_pass;
|
2016-08-12 10:17:45 +02:00
|
|
|
|
2018-01-24 09:46:04 +01:00
|
|
|
page_cleaner.flush_time = 0;
|
|
|
|
|
page_cleaner.flush_pass = 0;
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (flush_pass) {
|
|
|
|
|
flush_tm /= flush_pass;
|
2016-08-12 10:17:45 +02:00
|
|
|
}
|
|
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
MONITOR_SET(MONITOR_FLUSH_ADAPTIVE_AVG_TIME, flush_tm);
|
|
|
|
|
MONITOR_SET(MONITOR_FLUSH_ADAPTIVE_AVG_PASS, flush_pass);
|
2016-08-12 10:17:45 +02:00
|
|
|
|
2014-02-26 19:11:54 +01:00
|
|
|
prev_lsn = cur_lsn;
|
2016-08-12 10:17:45 +02:00
|
|
|
prev_time = curr_time;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
n_iterations = 0;
|
|
|
|
|
|
|
|
|
|
sum_pages = 0;
|
|
|
|
|
}
|
|
|
|
|
|
2021-01-06 12:53:14 +01:00
|
|
|
const ulint pct_for_dirty = srv_max_dirty_pages_pct_lwm == 0
|
|
|
|
|
? (dirty_pct >= max_pct ? 100 : 0)
|
|
|
|
|
: static_cast<ulint>
|
2020-12-08 13:57:51 +01:00
|
|
|
(max_pct > 0.0 ? dirty_pct / max_pct : dirty_pct);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ulint pct_total = std::max(pct_for_dirty, pct_for_lsn);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2016-08-12 10:17:45 +02:00
|
|
|
/* Estimate pages to be flushed for the lsn progress */
|
|
|
|
|
lsn_t target_lsn = oldest_lsn
|
2020-02-12 13:45:21 +01:00
|
|
|
+ lsn_avg_rate * buf_flush_lsn_scan_factor;
|
|
|
|
|
ulint pages_for_lsn = 0;
|
|
|
|
|
|
2020-12-08 13:57:51 +01:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
|
2020-03-18 20:48:00 +01:00
|
|
|
for (buf_page_t* b = UT_LIST_GET_LAST(buf_pool.flush_list);
|
2020-02-12 13:45:21 +01:00
|
|
|
b != NULL;
|
|
|
|
|
b = UT_LIST_GET_PREV(list, b)) {
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
if (b->oldest_modification() > target_lsn) {
|
2020-02-12 13:45:21 +01:00
|
|
|
break;
|
2016-08-12 10:17:45 +02:00
|
|
|
}
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (++pages_for_lsn >= srv_max_io_capacity) {
|
|
|
|
|
break;
|
|
|
|
|
}
|
2016-08-12 10:17:45 +02:00
|
|
|
}
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
2020-02-12 13:45:21 +01:00
|
|
|
|
|
|
|
|
pages_for_lsn /= buf_flush_lsn_scan_factor;
|
|
|
|
|
if (pages_for_lsn < 1) {
|
|
|
|
|
pages_for_lsn = 1;
|
2016-08-12 10:17:45 +02:00
|
|
|
}
|
|
|
|
|
|
2020-03-10 19:05:17 +01:00
|
|
|
n_pages = (ulint(double(srv_io_capacity) * double(pct_total) / 100.0)
|
|
|
|
|
+ avg_page_rate + pages_for_lsn) / 3;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
if (n_pages > srv_max_io_capacity) {
|
|
|
|
|
n_pages = srv_max_io_capacity;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
MONITOR_SET(MONITOR_FLUSH_N_TO_FLUSH_REQUESTED, n_pages);
|
|
|
|
|
|
2020-02-12 13:45:21 +01:00
|
|
|
MONITOR_SET(MONITOR_FLUSH_N_TO_FLUSH_BY_AGE, pages_for_lsn);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
MONITOR_SET(MONITOR_FLUSH_AVG_PAGE_RATE, avg_page_rate);
|
|
|
|
|
MONITOR_SET(MONITOR_FLUSH_LSN_AVG_RATE, lsn_avg_rate);
|
|
|
|
|
MONITOR_SET(MONITOR_FLUSH_PCT_FOR_DIRTY, pct_for_dirty);
|
|
|
|
|
MONITOR_SET(MONITOR_FLUSH_PCT_FOR_LSN, pct_for_lsn);
|
|
|
|
|
|
|
|
|
|
return(n_pages);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/******************************************************************//**
|
|
|
|
|
page_cleaner thread tasked with flushing dirty pages from the buffer
|
2016-08-12 10:17:45 +02:00
|
|
|
pools. As of now we'll have only one coordinator.
|
2014-02-26 19:11:54 +01:00
|
|
|
@return a dummy parameter */
|
2020-02-12 13:45:21 +01:00
|
|
|
static os_thread_ret_t DECLARE_THREAD(buf_flush_page_cleaner)(void*)
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
my_thread_init();
|
2014-02-26 19:11:54 +01:00
|
|
|
#ifdef UNIV_PFS_THREAD
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
pfs_register_thread(page_cleaner_thread_key);
|
2014-02-26 19:11:54 +01:00
|
|
|
#endif /* UNIV_PFS_THREAD */
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ut_ad(!srv_read_only_mode);
|
|
|
|
|
ut_ad(buf_page_cleaner_is_active);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
#ifdef UNIV_DEBUG_THREAD_CREATION
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ib::info() << "page_cleaner thread running, id "
|
2020-11-30 10:15:31 +01:00
|
|
|
<< os_thread_get_curr_id();
|
2014-02-26 19:11:54 +01:00
|
|
|
#endif /* UNIV_DEBUG_THREAD_CREATION */
|
2016-08-12 10:17:45 +02:00
|
|
|
#ifdef UNIV_LINUX
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
/* linux might be able to set different setting for each thread.
|
|
|
|
|
worth to try to set high priority for the page cleaner thread */
|
|
|
|
|
const pid_t tid= static_cast<pid_t>(syscall(SYS_gettid));
|
|
|
|
|
setpriority(PRIO_PROCESS, tid, -20);
|
|
|
|
|
if (getpriority(PRIO_PROCESS, tid) != -20)
|
|
|
|
|
ib::info() << "If the mysqld execution user is authorized,"
|
|
|
|
|
" page cleaner thread priority can be changed."
|
|
|
|
|
" See the man page of setpriority().";
|
2016-08-12 10:17:45 +02:00
|
|
|
#endif /* UNIV_LINUX */
|
|
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ulint last_pages= 0;
|
|
|
|
|
timespec abstime;
|
|
|
|
|
set_timespec(abstime, 1);
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
lsn_t lsn_limit;
|
2016-08-12 10:17:45 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
for (;;)
|
|
|
|
|
{
|
|
|
|
|
lsn_limit= buf_flush_sync_lsn;
|
2016-09-06 08:43:16 +02:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (UNIV_UNLIKELY(lsn_limit != 0))
|
|
|
|
|
{
|
|
|
|
|
furious_flush:
|
2020-11-04 15:55:36 +01:00
|
|
|
if (UNIV_LIKELY(srv_flush_sync))
|
|
|
|
|
{
|
|
|
|
|
buf_flush_sync_for_checkpoint(lsn_limit);
|
|
|
|
|
last_pages= 0;
|
|
|
|
|
set_timespec(abstime, 1);
|
|
|
|
|
continue;
|
|
|
|
|
}
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
}
|
2020-11-04 15:55:36 +01:00
|
|
|
else if (srv_shutdown_state > SRV_SHUTDOWN_INITIATED)
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
break;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-11-25 15:09:47 +01:00
|
|
|
if (buf_pool.page_cleaner_idle())
|
2021-02-07 11:19:24 +01:00
|
|
|
my_cond_wait(&buf_pool.do_flush_list,
|
|
|
|
|
&buf_pool.flush_list_mutex.m_mutex);
|
2020-11-25 15:09:47 +01:00
|
|
|
else
|
2021-02-07 11:19:24 +01:00
|
|
|
my_cond_timedwait(&buf_pool.do_flush_list,
|
|
|
|
|
&buf_pool.flush_list_mutex.m_mutex, &abstime);
|
2020-11-25 15:09:47 +01:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
set_timespec(abstime, 1);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
lsn_limit= buf_flush_sync_lsn;
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (UNIV_UNLIKELY(lsn_limit != 0))
|
2020-11-04 15:55:36 +01:00
|
|
|
{
|
|
|
|
|
if (UNIV_LIKELY(srv_flush_sync))
|
|
|
|
|
goto furious_flush;
|
|
|
|
|
}
|
|
|
|
|
else if (srv_shutdown_state > SRV_SHUTDOWN_INITIATED)
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
const ulint dirty_blocks= UT_LIST_GET_LEN(buf_pool.flush_list);
|
|
|
|
|
|
|
|
|
|
if (!dirty_blocks)
|
2020-11-04 15:55:36 +01:00
|
|
|
{
|
|
|
|
|
if (UNIV_UNLIKELY(lsn_limit != 0))
|
|
|
|
|
{
|
|
|
|
|
buf_flush_sync_lsn= 0;
|
|
|
|
|
/* wake up buf_flush_wait_flushed() */
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_broadcast(&buf_pool.done_flush_list);
|
2020-11-04 15:55:36 +01:00
|
|
|
}
|
2020-11-25 15:09:47 +01:00
|
|
|
unemployed:
|
|
|
|
|
buf_pool.page_cleaner_set_idle(true);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
continue;
|
2020-11-04 15:55:36 +01:00
|
|
|
}
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
|
|
|
|
|
/* We perform dirty reads of the LRU+free list lengths here.
|
|
|
|
|
Division by zero is not possible, because buf_pool.flush_list is
|
|
|
|
|
guaranteed to be nonempty, and it is a subset of buf_pool.LRU. */
|
|
|
|
|
const double dirty_pct= double(dirty_blocks) * 100.0 /
|
|
|
|
|
double(UT_LIST_GET_LEN(buf_pool.LRU) + UT_LIST_GET_LEN(buf_pool.free));
|
|
|
|
|
|
2021-01-06 12:53:14 +01:00
|
|
|
if (lsn_limit);
|
2021-02-18 17:20:50 +01:00
|
|
|
else if (srv_max_dirty_pages_pct_lwm != 0.0)
|
|
|
|
|
{
|
|
|
|
|
if (dirty_pct < srv_max_dirty_pages_pct_lwm)
|
|
|
|
|
goto unemployed;
|
|
|
|
|
}
|
2021-01-06 12:53:14 +01:00
|
|
|
else if (dirty_pct < srv_max_buf_pool_modified_pct)
|
|
|
|
|
goto unemployed;
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
|
2020-12-09 08:22:13 +01:00
|
|
|
const lsn_t oldest_lsn= buf_pool.get_oldest_modified()
|
|
|
|
|
->oldest_modification();
|
|
|
|
|
ut_ad(oldest_lsn);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
|
2020-11-04 15:55:36 +01:00
|
|
|
if (UNIV_UNLIKELY(lsn_limit != 0) && oldest_lsn >= lsn_limit)
|
|
|
|
|
buf_flush_sync_lsn= 0;
|
|
|
|
|
|
2020-11-25 15:09:47 +01:00
|
|
|
buf_pool.page_cleaner_set_idle(false);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
|
|
|
|
|
ulint n_flushed;
|
|
|
|
|
|
2020-11-04 15:55:36 +01:00
|
|
|
if (UNIV_UNLIKELY(lsn_limit != 0))
|
|
|
|
|
{
|
|
|
|
|
n_flushed= buf_flush_lists(srv_max_io_capacity, lsn_limit);
|
|
|
|
|
/* wake up buf_flush_wait_flushed() */
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_broadcast(&buf_pool.done_flush_list);
|
2020-11-04 15:55:36 +01:00
|
|
|
goto try_checkpoint;
|
|
|
|
|
}
|
|
|
|
|
else if (!srv_adaptive_flushing)
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
{
|
|
|
|
|
n_flushed= buf_flush_lists(srv_io_capacity, LSN_MAX);
|
2020-11-04 15:55:36 +01:00
|
|
|
try_checkpoint:
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (n_flushed)
|
|
|
|
|
{
|
|
|
|
|
MONITOR_INC_VALUE_CUMULATIVE(MONITOR_FLUSH_BACKGROUND_TOTAL_PAGE,
|
|
|
|
|
MONITOR_FLUSH_BACKGROUND_COUNT,
|
|
|
|
|
MONITOR_FLUSH_BACKGROUND_PAGES,
|
|
|
|
|
n_flushed);
|
|
|
|
|
do_checkpoint:
|
|
|
|
|
/* The periodic log_checkpoint() call here makes it harder to
|
|
|
|
|
reproduce bugs in crash recovery or mariabackup --prepare, or
|
|
|
|
|
in code that writes the redo log records. Omitting the call
|
|
|
|
|
here should not affect correctness, because log_free_check()
|
|
|
|
|
should still be invoking checkpoints when needed. */
|
|
|
|
|
DBUG_EXECUTE_IF("ib_log_checkpoint_avoid", goto next;);
|
|
|
|
|
|
|
|
|
|
if (!recv_recovery_is_on() && srv_operation == SRV_OPERATION_NORMAL)
|
|
|
|
|
log_checkpoint();
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
else if (ulint n= page_cleaner_flush_pages_recommendation(last_pages,
|
|
|
|
|
oldest_lsn,
|
2020-12-08 13:57:51 +01:00
|
|
|
dirty_blocks,
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
dirty_pct))
|
|
|
|
|
{
|
|
|
|
|
page_cleaner.flush_pass++;
|
|
|
|
|
const ulint tm= ut_time_ms();
|
|
|
|
|
last_pages= n_flushed= buf_flush_lists(n, LSN_MAX);
|
|
|
|
|
page_cleaner.flush_time+= ut_time_ms() - tm;
|
|
|
|
|
|
|
|
|
|
if (n_flushed)
|
|
|
|
|
{
|
|
|
|
|
MONITOR_INC_VALUE_CUMULATIVE(MONITOR_FLUSH_ADAPTIVE_TOTAL_PAGE,
|
|
|
|
|
MONITOR_FLUSH_ADAPTIVE_COUNT,
|
|
|
|
|
MONITOR_FLUSH_ADAPTIVE_PAGES,
|
|
|
|
|
n_flushed);
|
|
|
|
|
goto do_checkpoint;
|
|
|
|
|
}
|
|
|
|
|
}
|
2020-11-25 15:09:47 +01:00
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
goto unemployed;
|
|
|
|
|
}
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
|
|
|
|
|
#ifdef UNIV_DEBUG
|
|
|
|
|
while (innodb_page_cleaner_disabled_debug && !buf_flush_sync_lsn &&
|
|
|
|
|
srv_shutdown_state == SRV_SHUTDOWN_NONE)
|
|
|
|
|
os_thread_sleep(100000);
|
|
|
|
|
#endif /* UNIV_DEBUG */
|
|
|
|
|
|
|
|
|
|
#ifndef DBUG_OFF
|
|
|
|
|
next:
|
|
|
|
|
#endif /* !DBUG_OFF */
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
if (srv_fast_shutdown != 2)
|
|
|
|
|
{
|
|
|
|
|
buf_flush_wait_batch_end_acquiring_mutex(true);
|
|
|
|
|
buf_flush_wait_batch_end_acquiring_mutex(false);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
log_flush_task.wait();
|
|
|
|
|
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
lsn_limit= buf_flush_sync_lsn;
|
|
|
|
|
if (UNIV_UNLIKELY(lsn_limit != 0))
|
|
|
|
|
goto furious_flush;
|
|
|
|
|
buf_page_cleaner_is_active= false;
|
2021-02-07 11:19:24 +01:00
|
|
|
pthread_cond_broadcast(&buf_pool.done_flush_list);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
|
|
|
|
|
my_thread_end();
|
|
|
|
|
/* We count the number of threads in os_thread_exit(). A created
|
|
|
|
|
thread should always use that to exit and not use return() to exit. */
|
|
|
|
|
os_thread_exit();
|
|
|
|
|
|
|
|
|
|
OS_THREAD_DUMMY_RETURN;
|
|
|
|
|
}
|
2019-10-29 22:37:12 +01:00
|
|
|
|
2020-02-12 13:45:21 +01:00
|
|
|
/** Initialize page_cleaner. */
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ATTRIBUTE_COLD void buf_flush_page_cleaner_init()
|
2017-09-09 10:34:12 +02:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
ut_ad(!buf_page_cleaner_is_active);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
ut_ad(srv_operation == SRV_OPERATION_NORMAL ||
|
|
|
|
|
srv_operation == SRV_OPERATION_RESTORE ||
|
|
|
|
|
srv_operation == SRV_OPERATION_RESTORE_EXPORT);
|
|
|
|
|
buf_flush_sync_lsn= 0;
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
buf_page_cleaner_is_active= true;
|
|
|
|
|
os_thread_create(buf_flush_page_cleaner);
|
2016-08-12 10:17:45 +02:00
|
|
|
}
|
|
|
|
|
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
/** @return the number of dirty pages in the buffer pool */
|
|
|
|
|
static ulint buf_flush_list_length()
|
|
|
|
|
{
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
const ulint len= UT_LIST_GET_LEN(buf_pool.flush_list);
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
return len;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** Flush the buffer pool on shutdown. */
|
|
|
|
|
ATTRIBUTE_COLD void buf_flush_buffer_pool()
|
|
|
|
|
{
|
|
|
|
|
ut_ad(!buf_page_cleaner_is_active);
|
|
|
|
|
ut_ad(!buf_flush_sync_lsn);
|
|
|
|
|
|
|
|
|
|
service_manager_extend_timeout(INNODB_EXTEND_TIMEOUT_INTERVAL,
|
|
|
|
|
"Waiting to flush the buffer pool");
|
|
|
|
|
|
|
|
|
|
while (buf_pool.n_flush_list || buf_flush_list_length())
|
|
|
|
|
{
|
|
|
|
|
buf_flush_lists(srv_max_io_capacity, LSN_MAX);
|
|
|
|
|
timespec abstime;
|
|
|
|
|
|
|
|
|
|
if (buf_pool.n_flush_list)
|
|
|
|
|
{
|
|
|
|
|
service_manager_extend_timeout(INNODB_EXTEND_TIMEOUT_INTERVAL,
|
|
|
|
|
"Waiting to flush " ULINTPF " pages",
|
|
|
|
|
buf_flush_list_length());
|
|
|
|
|
set_timespec(abstime, INNODB_EXTEND_TIMEOUT_INTERVAL / 2);
|
|
|
|
|
mysql_mutex_lock(&buf_pool.mutex);
|
|
|
|
|
while (buf_pool.n_flush_list)
|
2021-02-07 11:19:24 +01:00
|
|
|
my_cond_timedwait(&buf_pool.done_flush_list, &buf_pool.mutex.m_mutex,
|
|
|
|
|
&abstime);
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
mysql_mutex_unlock(&buf_pool.mutex);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
ut_ad(!buf_pool.any_io_pending());
|
|
|
|
|
log_flush_task.wait();
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-12 13:45:21 +01:00
|
|
|
/** Synchronously flush dirty blocks.
|
|
|
|
|
NOTE: The calling thread is not allowed to hold any buffer page latches! */
|
|
|
|
|
void buf_flush_sync()
|
2016-08-12 10:17:45 +02:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
for (;;)
|
|
|
|
|
{
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
const ulint n_flushed= buf_flush_lists(srv_max_io_capacity, LSN_MAX);
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
buf_flush_wait_batch_end_acquiring_mutex(false);
|
|
|
|
|
if (!n_flushed)
|
|
|
|
|
{
|
|
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
const auto len= UT_LIST_GET_LEN(buf_pool.flush_list);
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
|
|
|
|
if (!len)
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
}
|
2016-08-12 10:17:45 +02:00
|
|
|
}
|
2014-02-26 19:11:54 +01:00
|
|
|
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
#ifdef UNIV_DEBUG
|
2014-02-26 19:11:54 +01:00
|
|
|
/** Functor to validate the flush list. */
|
|
|
|
|
struct Check {
|
2019-04-29 13:33:46 +02:00
|
|
|
void operator()(const buf_page_t* elem) const
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
2020-12-09 08:22:13 +01:00
|
|
|
ut_ad(elem->oldest_modification());
|
|
|
|
|
ut_ad(!fsp_is_system_temporary(elem->id().space()));
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
};
|
|
|
|
|
|
2020-02-12 13:45:21 +01:00
|
|
|
/** Validate the flush list. */
|
|
|
|
|
static void buf_flush_validate_low()
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
|
|
|
|
buf_page_t* bpage;
|
|
|
|
|
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_assert_owner(&buf_pool.flush_list_mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-03-18 20:48:00 +01:00
|
|
|
ut_list_validate(buf_pool.flush_list, Check());
|
2014-02-26 19:11:54 +01:00
|
|
|
|
2020-03-18 20:48:00 +01:00
|
|
|
bpage = UT_LIST_GET_FIRST(buf_pool.flush_list);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
while (bpage != NULL) {
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
const lsn_t om = bpage->oldest_modification();
|
2020-03-18 20:48:00 +01:00
|
|
|
/* A page in buf_pool.flush_list can be in
|
2014-02-26 19:11:54 +01:00
|
|
|
BUF_BLOCK_REMOVE_HASH state. This happens when a page
|
|
|
|
|
is in the middle of being relocated. In that case the
|
|
|
|
|
original descriptor can have this state and still be
|
|
|
|
|
in the flush list waiting to acquire the
|
2020-03-18 20:48:00 +01:00
|
|
|
buf_pool.flush_list_mutex to complete the relocation. */
|
2020-10-26 16:07:17 +01:00
|
|
|
ut_d(const auto s= bpage->state());
|
|
|
|
|
ut_ad(s == BUF_BLOCK_ZIP_PAGE || s == BUF_BLOCK_FILE_PAGE
|
|
|
|
|
|| s == BUF_BLOCK_REMOVE_HASH);
|
2020-10-15 09:27:25 +02:00
|
|
|
ut_ad(om > 0);
|
2014-02-26 19:11:54 +01:00
|
|
|
|
|
|
|
|
bpage = UT_LIST_GET_NEXT(list, bpage);
|
2020-10-15 09:27:25 +02:00
|
|
|
ut_ad(!bpage || recv_recovery_is_on()
|
|
|
|
|
|| om >= bpage->oldest_modification());
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-12 13:45:21 +01:00
|
|
|
/** Validate the flush list. */
|
|
|
|
|
void buf_flush_validate()
|
2014-02-26 19:11:54 +01:00
|
|
|
{
|
MDEV-23399: Performance regression with write workloads
The buffer pool refactoring in MDEV-15053 and MDEV-22871 shifted
the performance bottleneck to the page flushing.
The configuration parameters will be changed as follows:
innodb_lru_flush_size=32 (new: how many pages to flush on LRU eviction)
innodb_lru_scan_depth=1536 (old: 1024)
innodb_max_dirty_pages_pct=90 (old: 75)
innodb_max_dirty_pages_pct_lwm=75 (old: 0)
Note: The parameter innodb_lru_scan_depth will only affect LRU
eviction of buffer pool pages when a new page is being allocated. The
page cleaner thread will no longer evict any pages. It used to
guarantee that some pages will remain free in the buffer pool. Now, we
perform that eviction 'on demand' in buf_LRU_get_free_block().
The parameter innodb_lru_scan_depth(srv_LRU_scan_depth) is used as follows:
* When the buffer pool is being shrunk in buf_pool_t::withdraw_blocks()
* As a buf_pool.free limit in buf_LRU_list_batch() for terminating
the flushing that is initiated e.g., by buf_LRU_get_free_block()
The parameter also used to serve as an initial limit for unzip_LRU
eviction (evicting uncompressed page frames while retaining
ROW_FORMAT=COMPRESSED pages), but now we will use a hard-coded limit
of 100 or unlimited for invoking buf_LRU_scan_and_free_block().
The status variables will be changed as follows:
innodb_buffer_pool_pages_flushed: This includes also the count of
innodb_buffer_pool_pages_LRU_flushed and should work reliably,
updated one by one in buf_flush_page() to give more real-time
statistics. The function buf_flush_stats(), which we are removing,
was not called in every code path. For both counters, we will use
regular variables that are incremented in a critical section of
buf_pool.mutex. Note that show_innodb_vars() directly links to the
variables, and reads of the counters will *not* be protected by
buf_pool.mutex, so you cannot get a consistent snapshot of both variables.
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be
removed, because the page cleaner no longer deals with writing or
evicting least recently used pages, and because the single-page writes
have been removed:
* buffer_LRU_batch_flush_avg_time_slot
* buffer_LRU_batch_flush_avg_time_thread
* buffer_LRU_batch_flush_avg_time_est
* buffer_LRU_batch_flush_avg_pass
* buffer_LRU_single_flush_scanned
* buffer_LRU_single_flush_num_scan
* buffer_LRU_single_flush_scanned_per_call
When moving to a single buffer pool instance in MDEV-15058, we missed
some opportunity to simplify the buf_flush_page_cleaner thread. It was
unnecessarily using a mutex and some complex data structures, even
though we always have a single page cleaner thread.
Furthermore, the buf_flush_page_cleaner thread had separate 'recovery'
and 'shutdown' modes where it was waiting to be triggered by some
other thread, adding unnecessary latency and potential for hangs in
relatively rarely executed startup or shutdown code.
The page cleaner was also running two kinds of batches in an
interleaved fashion: "LRU flush" (writing out some least recently used
pages and evicting them on write completion) and the normal batches
that aim to increase the MIN(oldest_modification) in the buffer pool,
to help the log checkpoint advance.
The buf_pool.flush_list flushing was being blocked by
buf_block_t::lock for no good reason. Furthermore, if the FIL_PAGE_LSN
of a page is ahead of log_sys.get_flushed_lsn(), that is, what has
been persistently written to the redo log, we would trigger a log
flush and then resume the page flushing. This would unnecessarily
limit the performance of the page cleaner thread and trigger the
infamous messages "InnoDB: page_cleaner: 1000ms intended loop took 4450ms.
The settings might not be optimal" that were suppressed in
commit d1ab89037a518fcffbc50c24e4bd94e4ec33aed0 unless log_warnings>2.
Our revised algorithm will make log_sys.get_flushed_lsn() advance at
the start of buf_flush_lists(), and then execute a 'best effort' to
write out all pages. The flush batches will skip pages that were modified
since the log was written, or are are currently exclusively locked.
The MDEV-13670 message "page_cleaner: 1000ms intended loop took" message
will be removed, because by design, the buf_flush_page_cleaner() should
not be blocked during a batch for extended periods of time.
We will remove the single-page flushing altogether. Related to this,
the debug parameter innodb_doublewrite_batch_size will be removed,
because all of the doublewrite buffer will be used for flushing
batches. If a page needs to be evicted from the buffer pool and all
100 least recently used pages in the buffer pool have unflushed
changes, buf_LRU_get_free_block() will execute buf_flush_lists() to
write out and evict innodb_lru_flush_size pages. At most one thread
will execute buf_flush_lists() in buf_LRU_get_free_block(); other
threads will wait for that LRU flushing batch to finish.
To improve concurrency, we will replace the InnoDB ib_mutex_t and
os_event_t native mutexes and condition variables in this area of code.
Most notably, this means that the buffer pool mutex (buf_pool.mutex)
is no longer instrumented via any InnoDB interfaces. It will continue
to be instrumented via PERFORMANCE_SCHEMA.
For now, both buf_pool.flush_list_mutex and buf_pool.mutex will be
declared with MY_MUTEX_INIT_FAST (PTHREAD_MUTEX_ADAPTIVE_NP). The critical
sections of buf_pool.flush_list_mutex should be shorter than those for
buf_pool.mutex, because in the worst case, they cover a linear scan of
buf_pool.flush_list, while the worst case of a critical section of
buf_pool.mutex covers a linear scan of the potentially much longer
buf_pool.LRU list.
mysql_mutex_is_owner(), safe_mutex_is_owner(): New predicate, usable
with SAFE_MUTEX. Some InnoDB debug assertions need this predicate
instead of mysql_mutex_assert_owner() or mysql_mutex_assert_not_owner().
buf_pool_t::n_flush_LRU, buf_pool_t::n_flush_list:
Replaces buf_pool_t::init_flush[] and buf_pool_t::n_flush[].
The number of active flush operations.
buf_pool_t::mutex, buf_pool_t::flush_list_mutex: Use mysql_mutex_t
instead of ib_mutex_t, to have native mutexes with PERFORMANCE_SCHEMA
and SAFE_MUTEX instrumentation.
buf_pool_t::done_flush_LRU: Condition variable for !n_flush_LRU.
buf_pool_t::done_flush_list: Condition variable for !n_flush_list.
buf_pool_t::do_flush_list: Condition variable to wake up the
buf_flush_page_cleaner when a log checkpoint needs to be written
or the server is being shut down. Replaces buf_flush_event.
We will keep using timed waits (the page cleaner thread will wake
_at least_ once per second), because the calculations for
innodb_adaptive_flushing depend on fixed time intervals.
buf_dblwr: Allocate statically, and move all code to member functions.
Use a native mutex and condition variable. Remove code to deal with
single-page flushing.
buf_dblwr_check_block(): Make the check debug-only. We were spending
a significant amount of execution time in page_simple_validate_new().
flush_counters_t::unzip_LRU_evicted: Remove.
IORequest: Make more members const. FIXME: m_fil_node should be removed.
buf_flush_sync_lsn: Protect by std::atomic, not page_cleaner.mutex
(which we are removing).
page_cleaner_slot_t, page_cleaner_t: Remove many redundant members.
pc_request_flush_slot(): Replaces pc_request() and pc_flush_slot().
recv_writer_thread: Remove. Recovery works just fine without it, if we
simply invoke buf_flush_sync() at the end of each batch in
recv_sys_t::apply().
recv_recovery_from_checkpoint_finish(): Remove. We can simply call
recv_sys.debug_free() directly.
srv_started_redo: Replaces srv_start_state.
SRV_SHUTDOWN_FLUSH_PHASE: Remove. logs_empty_and_mark_files_at_shutdown()
can communicate with the normal page cleaner loop via the new function
flush_buffer_pool().
buf_flush_remove(): Assert that the calling thread is holding
buf_pool.flush_list_mutex. This removes unnecessary mutex operations
from buf_flush_remove_pages() and buf_flush_dirty_pages(),
which replace buf_LRU_flush_or_remove_pages().
buf_flush_lists(): Renamed from buf_flush_batch(), with simplified
interface. Return the number of flushed pages. Clarified comments and
renamed min_n to max_n. Identify LRU batch by lsn=0. Merge all the functions
buf_flush_start(), buf_flush_batch(), buf_flush_end() directly to this
function, which was their only caller, and remove 2 unnecessary
buf_pool.mutex release/re-acquisition that we used to perform around
the buf_flush_batch() call. At the start, if not all log has been
durably written, wait for a background task to do it, or start a new
task to do it. This allows the log write to run concurrently with our
page flushing batch. Any pages that were skipped due to too recent
FIL_PAGE_LSN or due to them being latched by a writer should be flushed
during the next batch, unless there are further modifications to those
pages. It is possible that a page that we must flush due to small
oldest_modification also carries a recent FIL_PAGE_LSN or is being
constantly modified. In the worst case, all writers would then end up
waiting in log_free_check() to allow the flushing and the checkpoint
to complete.
buf_do_flush_list_batch(): Clarify comments, and rename min_n to max_n.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_flush_space(): Auxiliary function to look up a tablespace for
page flushing.
buf_flush_page(): Defer the computation of space->full_crc32(). Never
call log_write_up_to(), but instead skip persistent pages whose latest
modification (FIL_PAGE_LSN) is newer than the redo log. Also skip
pages on which we cannot acquire a shared latch without waiting.
buf_flush_try_neighbors(): Do not bother checking buf_fix_count
because buf_flush_page() will no longer wait for the page latch.
Take the tablespace as a parameter, and only execute this function
when innodb_flush_neighbors>0. Avoid repeated calls of page_id_t::fold().
buf_flush_relocate_on_flush_list(): Declare as cold, and push down
a condition from the callers.
buf_flush_check_neighbor(): Take id.fold() as a parameter.
buf_flush_sync(): Ensure that the buf_pool.flush_list is empty,
because the flushing batch will skip pages whose modifications have
not yet been written to the log or were latched for modification.
buf_free_from_unzip_LRU_list_batch(): Remove redundant local variables.
buf_flush_LRU_list_batch(): Let the caller buf_do_LRU_batch() initialize
the counters, and report n->evicted.
Cache the last looked up tablespace. If neighbor flushing is not applicable,
invoke buf_flush_page() directly, avoiding a page lookup in between.
buf_do_LRU_batch(): Return the number of pages flushed.
buf_LRU_free_page(): Only release and re-acquire buf_pool.mutex if
adaptive hash index entries are pointing to the block.
buf_LRU_get_free_block(): Do not wake up the page cleaner, because it
will no longer perform any useful work for us, and we do not want it
to compete for I/O while buf_flush_lists(innodb_lru_flush_size, 0)
writes out and evicts at most innodb_lru_flush_size pages. (The
function buf_do_LRU_batch() may complete after writing fewer pages if
more than innodb_lru_scan_depth pages end up in buf_pool.free list.)
Eliminate some mutex release-acquire cycles, and wait for the LRU
flush batch to complete before rescanning.
buf_LRU_check_size_of_non_data_objects(): Simplify the code.
buf_page_write_complete(): Remove the parameter evict, and always
evict pages that were part of an LRU flush.
buf_page_create(): Take a pre-allocated page as a parameter.
buf_pool_t::free_block(): Free a pre-allocated block.
recv_sys_t::recover_low(), recv_sys_t::apply(): Preallocate the block
while not holding recv_sys.mutex. During page allocation, we may
initiate a page flush, which in turn may initiate a log flush, which
would require acquiring log_sys.mutex, which should always be acquired
before recv_sys.mutex in order to avoid deadlocks. Therefore, we must
not be holding recv_sys.mutex while allocating a buffer pool block.
BtrBulk::logFreeCheck(): Skip a redundant condition.
row_undo_step(): Do not invoke srv_inc_activity_count() for every row
that is being rolled back. It should suffice to invoke the function in
trx_flush_log_if_needed() during trx_t::commit_in_memory() when the
rollback completes.
sync_check_enable(): Remove. We will enable innodb_sync_debug from the
very beginning.
Reviewed by: Vladislav Vaintroub
2020-10-15 11:10:42 +02:00
|
|
|
mysql_mutex_lock(&buf_pool.flush_list_mutex);
|
|
|
|
|
buf_flush_validate_low();
|
|
|
|
|
mysql_mutex_unlock(&buf_pool.flush_list_mutex);
|
2014-02-26 19:11:54 +01:00
|
|
|
}
|
MDEV-15053 Reduce buf_pool_t::mutex contention
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eefe56b3e97e0edefb3df76dd7ef5e858
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b02ea707c6461353e6cfc15d7553bc6
and recently removed by us in
commit 1a6f708ec594ac0ae2dd30db926ab07b100fa24b (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6ef440a037011c11873b396923e93c5 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
2020-06-05 11:35:46 +02:00
|
|
|
#endif /* UNIV_DEBUG */
|