Only one checkpoint may be in progress at a time.
The counter log_sys.n_pending_checkpoint_writes
was being protected by log_sys.mutex.
Let us replace it with the Boolean log_sys.checkpoint_pending.
If innodb_flush_method=O_DSYNC, log_sys.flushed_to_disk_lsn is changed
without 'flush_lock' protection inside log_write().
This leads to a race condition, if there are 2 threads running in parallel,
doing log_write_up_to() with different values for 'flush_to_disk'
In this case, log_write() and log_write_flush_to_disk_low() can execute at
the same time, and both would change flushed_lsn.
The fix is to remove special treatment of durable writes from log_write().
There is no apparent reason for this special treatment, log_write_flush_to_disk_low()
is already optimized for durable writes.
Nor there is an apparent reason to call log_flush_notify() more often in
for O_DSYNC.
In recv_sys_t::apply(), we were unnecessarily looking up pages
in buf_pool.page_hash and potentially waiting for exclusive page latches.
Before buf_page_get_low() would return an x-latched page,
that page will have to be read and buf_page_read_complete() would
have invoked recv_recover_page() to apply the log to the page.
Therefore, it suffices to invoke recv_read_in_area() to trigger
a transition from RECV_NOT_PROCESSED.
recv_read_in_area(): Take the iterator as a parameter, and remove
page_id lookups. Should the page already be in buf_pool.page_hash,
buf_page_init_for_read() will return nullptr to buf_read_page_low()
and buf_read_page_background().
recv_sys_t::apply(): Replace goto, remove dead code, and add assertions
to guarantee that the iteration will make progress.
Reviewed by: Vladislav Lesin
In MDEV-14425, an early plan was to introduce a separate log file
for file-level records and checkpoint information. The reasoning was
that fil_system.mutex contention would be reduced by not having to
maintain fil_system.named_spaces. The mutex contention was actually
fixed in MDEV-23855 by making some data fields in fil_space_t and
fil_node_t use std::atomic.
Using a single circular log file simplifies recovery and backup.
The problem was introduced by the removal of buf_pool.flush_rbt
in commit 46b1f50098 (MDEV-23399)
recv_sys_t::apply(): don't write to disc and fsync() the last batch.
Insead, sort it by oldest_modification for MariaDB server and some
mariabackup operations.
log_sort_flush_list(): a thread-safe function which sorts buf_pool::flush_list
InnoDB could sometimes hang when triggering a log checkpoint. This is
due to commit 7b1252c03d (MDEV-24278),
which introduced an untimed wait to buf_flush_page_cleaner().
The hang was noticed by occasional failures of IMPORT TABLESPACE tests,
such as innodb.innodb-wl5522, which would (unnecessarily) invoke
log_make_checkpoint() from row_import_cleanup().
The reason of the hang was that buf_flush_page_cleaner() would enter
untimed sleep despite buf_flush_sync_lsn being set. The exact failure
scenario is unclear, because buf_flush_sync_lsn should actually be
protected by buf_pool.flush_list_mutex. We prevent the hang by
invoking buf_pool.page_cleaner_set_idle(false) whenever we are
setting buf_flush_sync_lsn and signaling buf_pool.do_flush_list.
The bulk of these changes was originally developed as a preparation
for MDEV-26827, to invoke buf_flush_list() from fewer threads,
and tested on 10.6 by Matthias Leich.
This fix was tested by running 100 repetitions of 100 concurrent instances
of the test innodb.innodb-wl5522 on a RelWithDebInfo build, using ext4fs
and innodb_flush_method=O_DIRECT on a SATA SSD with 4096-byte block size.
During the test, the call to log_make_checkpoint() in row_import_cleanup()
was present.
buf_flush_list(): Make static.
buf_flush_wait(): Wait for buf_pool.get_oldest_modification()
to reach a target, by work done in the buf_flush_page_cleaner.
If buf_flush_sync_lsn is going to be set, we will invoke
buf_pool.page_cleaner_set_idle(false).
buf_flush_ahead(): If buf_flush_sync_lsn or buf_flush_async_lsn
is going to be set and the page cleaner woken up, we will invoke
buf_pool.page_cleaner_set_idle(false).
buf_flush_wait_flushed(): Invoke buf_flush_wait().
buf_flush_sync(): Invoke recv_sys.apply() at the start in case
crash recovery is active. Invoke buf_flush_wait().
buf_flush_sync_batch(): A lower-level variant of buf_flush_sync()
that is only called by recv_sys_t::apply().
buf_flush_sync_for_checkpoint(): Do not trigger log apply
or checkpoint during recovery.
buf_dblwr_t::create(): Only initiate a buffer pool flush, not
a checkpoint.
row_import_cleanup(): Do not unnecessarily invoke log_make_checkpoint().
Invoking buf_flush_list_space() before starting to generate redo log
for the imported tablespace should suffice.
srv_prepare_to_delete_redo_log_file():
Set recv_sys.recovery_on in order to prevent
buf_flush_sync_for_checkpoint() from initiating a checkpoint
while the log is inaccessible. Remove a wait loop that is already
part of buf_flush_sync().
Do not invoke fil_names_clear() if the log is being upgraded,
because the FILE_MODIFY record is specific to the latest format.
create_log_file(): Clear recv_sys.recovery_on only after calling
log_make_checkpoint(), to prevent buf_flush_page_cleaner from
invoking a checkpoint.
innodb_shutdown(): Simplify the logic in mariadb-backup --prepare.
os_aio_wait_until_no_pending_writes(): Update the function comment.
Apart from row_quiesce_table_start() during FLUSH TABLES...FOR EXPORT,
this is being called by buf_flush_list_space(), which is invoked
by ALTER TABLE...IMPORT TABLESPACE as well as some encryption operations.
During startup, InnoDB must write a FILE_CHECKPOINT record.
However, before MDEV-12353 (in MariaDB Server 10.2, 10.3, 10.4)
the corresponding record MLOG_CHECKPOINT was encoded in a different way.
When we are upgrading from a logically empty 10.2, 10.3, or 10.4 redo log,
we must not write anything to the old log file, because if the server were
killed during the upgrade, we would end up with a corrupted log file, and
both the old and the new server would refuse to start up.
On upgrade, we must simply create a new logically empty log file
and replace the old ib_logfile0 with that.
This is a low hanging fruit. Before this patch std::map::emplace() was
a ~50% of the whole recv_sys_t::parse() operation in by test.
After the fix it's only ~20%.
recv_sys_t::parse() recv_sys_t::pages is a collection of all pages
to recovery. Often, there are multiple changes for a single page.
Often, they go in a row and for such cases let's avoid
lookup in a std::map. cached_pages_it serves as a cache
of size 1.
recv_sys_t::add(): replace page_id argument with a std::map::iterator
At least since commit 055a3334ad
(MDEV-13564) the undo log truncation in InnoDB did not work correctly.
The main issue is that during the execution of
trx_purge_truncate_history() some pages of the newly truncated
undo tablespace could be discarded.
This is improved from commit 1cb218c37c
which was applied to earlier-version branches.
fsp_try_extend_data_file(): Apply the peculiar rounding of
fil_space_t::size_in_header only to the system tablespace,
whose size can be expressed in megabytes in a configuration parameter.
Other files may freely grow by a number of pages.
fseg_alloc_free_page_low(): Do allow the extension of undo tablespaces,
and mention the file name in the error message.
mtr_t::commit_shrink(): Implement crash-safe shrinking of a tablespace:
(1) durably write the log
(2) release the page latches of the rebuilt tablespace
(3) release the mutexes
(4) truncate the file
(5) release the tablespace latch
This is refactored from trx_purge_truncate_history().
log_write_and_flush_prepare(), log_write_and_flush(): New functions
to durably write log during mtr_t::commit_shrink().
It is implementation-defined whether alignment requirements
that are larger than std::max_align_t (typically 8 or 16 bytes)
will be honored by the compiler and linker.
It turns out that on IBM AIX, both alignas() and MY_ALIGNED()
only guarantees alignment up to 16 bytes.
For some data structures, specifying alignment to the CPU
cache line size (typically 64 or 128 bytes) is a mere performance
optimization, and we do not really care whether the requested
alignment is guaranteed.
But, for the correct operation of direct I/O, we do require that
the buffers be aligned at a block size boundary.
field_ref_zero: Define as a pointer, not an array.
For innochecksum, we can make this point to unaligned memory;
for anything else, we will allocate an aligned buffer from the heap.
This buffer will be used for overwriting freed data pages when
innodb_immediate_scrub_data_uncompressed=ON. And exactly that code
hit an assertion failure on AIX, in the test innodb.innodb_scrub.
log_sys.checkpoint_buf: Define as a pointer to aligned memory
that is allocated from heap.
log_t::file::write_header_durable(): Reuse log_sys.checkpoint_buf
instead of trying to allocate an aligned buffer from the stack.
init_crypt_key(): On failure, set info->key_version to
ENCRYPTION_KEY_VERSION_INVALID.
log_crypt_101_read_block(): Refuse to attempt decryption if
info->key_version is ENCRYPTION_KEY_VERSION_INVALID.
As pointed out by Andrei Elkin, the previous fix did not fix one
race condition that may have caused the observed hang.
innodb_log_flush_request(): If we are enqueueing the very first
request at the same time the log write is being completed,
we must ensure that a near-concurrent call to log_flush_notify()
will not result in a missed notification. We guarantee this by
release-acquire operations on log_requests.start and
log_sys.flushed_to_disk_lsn.
log_flush_notify_and_unlock(): Cleanup: Always release the mutex.
log_sys_t::get_flushed_lsn(): Use acquire memory order.
log_sys_t::set_flushed_lsn(): Use release memory order.
log_sys_t::set_lsn(): Use release memory order.
log_sys_t::get_lsn(): Use relaxed memory order by default, and
allow the caller to specify acquire memory order explicitly.
Whenever the log_sys.mutex is being held or when log writes are
prohibited during startup, we can use a relaxed load. Likewise,
in some assertions where reading a stale value of log_sys.lsn
should not matter, we can use a relaxed load.
This will cause some additional instructions to be emitted on
architectures that do not implement Total Store Ordering (TSO),
such as POWER, ARM, and RISC-V Weak Memory Ordering (RVWMO).
Starting with MariaDB 10.5, roughly after MDEV-23855 was fixed,
we are observing sporadic hangs during the execution of the
RESET MASTER statement. We are hoping to fix the hangs with these
changes, but due to the rather infrequent occurrence of the hangs
and our inability to reliably reproduce the hangs, we cannot be
sure of this.
What we do know is that innodb_force_recovery=2 (or a larger setting)
will prevent srv_master_callback (the former srv_master_thread) from
running. In that mode, periodic log flushes would never occur and
RESET MASTER could hang indefinitely. That is demonstrated by the new
test case that was developed by Andrei Elkin. We fix this case by
implementing a special case for it.
This also includes some code cleanup and renames of misleadingly
named code. The interface has nothing to do with log checkpoints in
the storage engine; it is only about requesting log writes to be
persistent.
handlerton::commit_checkpoint_request,
commit_checkpoint_notify_ha(): Remove the unused parameter hton.
log_requests.start: Replaces pending_checkpoint_list.
log_requests.end: Replaces pending_checkpoint_list_end.
log_requests.mutex: Replaces pending_checkpoint_mutex.
log_flush_notify_and_unlock(), log_flush_notify(): Replaces
innobase_mysql_log_notify(). The new implementation should be
functionally equivalent to the old one.
innodb_log_flush_request(): Replaces innobase_checkpoint_request().
Implement a fast path for common cases, and reduce the mutex hold time.
POSSIBLE FIX OF THE HANG: We will invoke commit_checkpoint_notify_ha()
for the current request if it is already satisfied, as well as invoke
log_flush_notify_and_unlock() for any satisfied requests.
log_write(): Invoke log_flush_notify() when the write is already durable.
This was missing WITH_PMEM when the log is in persistent memory.
Reviewed by: Vladislav Vaintroub
fil_op_replay_rename(): Remove.
fil_rename_tablespace_check(): Remove a parameter is_discarded=false.
recv_sys_t::parse(): Instead of applying FILE_RENAME operations,
buffer the operations in renamed_spaces.
recv_sys_t::apply(): In the last_batch, apply renamed_spaces.
In commit 118e258aaa (part of MDEV-23855)
we inadvertently broke crash recovery, reintroducing MDEV-11556.
fil_system_t::extend_to_recv_size(): Extend all open tablespace files
to the recovered size.
recv_sys_t::apply(): Invoke fil_system.extend_to_recv_size() at the
start of each batch. In this way, any fil_space_t::recv_size
changes that were parsed after the file was opened will be applied.
The test case encryption.innodb_encrypt_freed was failing in
MemorySanitizer builds.
recv_recover_page(): Mark non-recovered pages as freed.
fil_crypt_rotate_page(): Before comparing the block->frame contents,
check if the block was marked as freed.
Other places: Whenever using BUF_GET_POSSIBLY_FREED, check the
block->page.status before accessing the page frame.
(Both uses of BUF_GET_IF_IN_POOL should be correct now.)
- The commit 5fd3c7471e3e0673b50d309567c9747d36f09412(MDEV-24709)
resets the recv_no_ibuf_operations in
recv_recovery_from_checkpoint_start(), but InnoDB fails to reset
the variable recv_no_log_write() during that time and that leads
to the assert failure.
btr_cur_upd_rec_in_place(): Prefer WRITE to MEMSET for a single-byte
operation.
log_phys_t::apply(): Relax the assertion to allow a single-byte MEMSET,
even though it is 1 byte longer than a WRITE record.
The condition variables that were introduced in
commit 7cffb5f6e8 (MDEV-23399)
are never instrumented with PERFORMANCE_SCHEMA.
Let us avoid the storage overhead and dead code.
recv_recovery_from_checkpoint_start(): Clear the recv_no_ibuf_operations
flag at the same time when we enabled writes to the log.
The failure to clear the flag might have caused some missed
change buffer merges, at least to the secondary index of SYS_TABLES
that were accessed by trx_resurrect_table_locks() while the last
recovery batch was in progress.
Thanks to Thirunarayanan Balathandayuthapani for suggesting this fix.
This corresponds to 10.5 commit 39378e1366.
With a patched version of the test innodb.ibuf_not_empty (so that
it would trigger crash recovery after using the change buffer),
and patched code that would modify the os_thread_sleep() in
recv_apply_hashed_log_recs() to be 1ms as well as add a sleep of
the same duration to the end of recv_recover_page() when
recv_sys->n_addrs=0, we can demonstrate a race condition.
After disabling some debug checks in buf_all_freed_instance(),
buf_pool_invalidate_instance() and buf_validate(), we managed to
trigger an assertion failure in fseg_free_step(), on the XDES_FREE_BIT.
In other words, an trx_undo_seg_free() call during
trx_rollback_resurrected() was attempting a double-free of a page.
This was repeated about once in 400 to 500 test runs. With the fix
applied, the test passed 2,000 runs.
recv_apply_hashed_log_recs(): Do not only wait for recv_sys->n_addrs
to reach 0, but also wait for buf_get_n_pending_read_ios() to reach 0,
to guarantee that buf_page_io_complete() will not be executing
ibuf_merge_or_delete_for_page().
We are seeing !buf_pool.any_io_pending() assertion failures
in srv_start() ever since MDEV-21452 in 10.6. But, the
problem appears to be older. In 10.5 since MDEV-19514 removed
writes from the precursor of buf_page_read_complete(), it
seems that the debug assertion failure could have been harmless.
recv_sys_t::apply(): At the end of each batch, wait not only
for all log records to have been processed, but also for all
pending reads to complete, so that the buffer pool will be in
an idle state.
A side effect of MDEV-16264 is that a large number of threads will
be created at server startup, to be destroyed after a minute or two.
One source of such thread creation is srv_start_periodic_timer().
InnoDB is creating 3 periodic tasks: srv_master_callback (1Hz)
srv_error_monitor_task (1Hz), and srv_monitor_task (0.2Hz).
It appears that we can merge srv_error_monitor_task and srv_monitor_task
and have them invoked 4 times per minute (every 15 seconds). This will
affect our ability to enforce innodb_fatal_semaphore_wait_threshold and
some computations around BUF_LRU_STAT_N_INTERVAL.
We could remove srv_master_callback along with the DROP TABLE queue
at some point of time in the future. We must keep it independent
of the innodb_fatal_semaphore_wait_threshold detection, because
the background DROP TABLE queue could get stuck due to dict_sys
being locked by another thread. For now, srv_master_callback
must be invoked once per second, so that
innodb_flush_log_at_timeout=1 can work.
BUF_LRU_STAT_N_INTERVAL: Reduce the precision and extend the time
from 50*1 second to 4*15 seconds.
srv_error_monitor_timer: Remove.
MAX_MUTEX_NOWAIT: Increase from 20*1 second to 2*15 seconds.
srv_refresh_innodb_monitor_stats(): Avoid a repeated call to time(NULL).
Change the interval to less than 60 seconds.
srv_monitor(): Renamed from srv_monitor_task.
srv_monitor_task(): Renamed from srv_error_monitor_task().
Invoked only once in 15 seconds. Invoke also srv_monitor().
Increase the fatal_cnt threshold from 10*1 second to 1*15 seconds.
sync_array_print_long_waits_low(): Invoke time(NULL) only once.
Remove a bogus message about printouts for 30 seconds. Those
printouts were effectively already disabled in MDEV-16264
(commit 5e62b6a5e0).
