This bug was seen when parallel replication experienced a deadlock between
transactions T1 and T2, where T2 has reached the commit phase and is waiting
for T1 to commit first. In this case, the deadlock is broken by sending a kill
to T2; that kill error is then later detected and converted to a deadlock
error, which causes T2 to be rolled back and retried.
The problem was that the kill caused ha_commit_trans() to errorneously call
wakeup_subsequent_commits() on T3, signalling it to abort because T2 failed
during commit. This is incorrect, because the error in T2 is only a temporary
error, which will be resolved by normal transaction retry. We should not
signal error to the next transaction until we have executed the code that
handles such temporary errors.
So this patch just removes the calls to wakeup_subsequent_commits() from
ha_commit_trans(). They are incorrect in this case, and they are not needed in
general, as wakeup_subsequent_commits() must in any case be called in
finish_event_group() to wakeup any transactions that may have started to wait
after ha_commit_trans(). And normally, wakeup will in fact have happened
earlier, either from the binlog group commit code, or (in case of no
binlogging) after the fast part of InnoDB/XtraDB group commit.
The symptom of this bug was that replication would break on some transaction
with "Commit failed due to failure of an earlier commit on which this one
depends", but with no such failure of an earlier commit visible anywhere.
The retry of an event group in parallel replication set the wrong value for
the end log position of the event that was retried
(qev->future_event_relay_log_pos). It was too large by the size of the event,
so it pointed into the middle of the following event.
If the retry happened in the very last event of the event group, _and_ the SQL
thread was stopped just after successfully retrying that event, then the SQL
threads's relay log position would be left incorrect. Restarting the SQL
thread could then try to read events from a garbage offset in the relay log,
usually leading to an error about not being able to read the event.
The code in binlog group commit around wait_for_commit that controls commit
order, did the wakeup of subsequent commits early, as soon as a following
transaction is put into the group commit queue, but before any such commit has
actually taken place. This causes problems with too early wakeup of
transactions that need to wait for prior to commit, but do not take part in
the binlog group commit for one reason or the other.
This patch solves the problem, by moving the wakeup to happen only after the
binlog group commit is completed.
This requires a new solution to ensure that transactions that arrive later
than the leader are still able to participate in group commit. This patch
introduces a flag wait_for_commit::commit_started. When this is set, a waiter
can queue up itself in the group commit queue.
This way, effectively the wait_for_prior_commit() is skipped only for
transactions that participate in group commit, so that skipping the wait is
safe. Other transactions still wait as needed for correctness.
The code that handles free lists of various objects passed to worker threads
in parallel replication handles freeing in batches, to avoid taking and
releasing LOCK_rpl_thread too often. However, it was possible for freeing to
be delayed to the point where one thread could stall the SQL driver thread due
to full queue, while other worker threads might be idle. This could
significantly degrade possible parallelism and thus performance.
Clean up the batch freeing code so that it is more robust and now able to
regularly free batches of object, so that normally the queue will not run full
unless the SQL driver thread is really far ahead of the worker threads.
The bug occured in parallel replication when re-trying transactions that
failed due to deadlock. In this case, the relay log file is re-opened and the
events are read out again. This reading requires a format description event of
the appropriate version. But the code was using a description event stored in
rli, which is not thread-safe. This could lead to various rare races if the
format description event was replaced by the SQL driver thread at the exact
moment where a worker thread was trying to use it.
The fix is to instead make the retry code create and maintain its own format
description event. When the relay log file is opened, we first read the format
description event from the start of the file, before seeking to the current
position. This now uses the same code as when the SQL driver threads starts
from a given relay log position. This also makes sure that the correct format
description event version will be used in cases where the version of the
binlog could change during replication.
In parallel replication, threads can do two different waits for a prior
transaction. One is for the prior transaction to start commit, the other is
for it to complete commit.
It turns out that the same PSI_stage_info message was errorneously used in
both cases (probably a merge error), causing SHOW PROCESSLIST to be
misleading.
Fix by using correct, distinct message in each case.
In parallel replication, the wait_for_commit facility is used to ensure that
events are written into the binlog in the correct order. This is handled in an
optimised way in the binlogging group commit code.
However, some statements, for example GRANT, are written directly into the
binlog, outside of the group commit code. There was a bug that this direct
write does not correctly wait for the prior transactions to have been written
first, which allows f.ex. GRANT to be written ahead of earlier transactions.
This patch adds the missing wait_for_prior_commit() before writing directly to
the binlog.
However, the problem is still there, although the race is much less likely to
occur now. The problem is that the optimised group commit code does wakeup of
following transactions early, before the binlog write is actually done. A
woken-up following transaction is then allowed to run ahead and queue up for
the group commit, which will ensure that binlog write happens in correct order
in the end. However, the code for directly written events currently bypass
this mechanism, so they get woken up and written too early.
This will be fixed properly in a later patch.
The real bug was that open_tables() returned error in case of
thd->killed() without properly calling thd->send_kill_message()
to set the correct error. This was fixed some time ago.
So remove the, now redundant, extra checks for thd->is_error(),
possibly allowing to catch in debug builds more incorrect
error handling cases.
In SAFE_MUTEX builds, reset the wait_for_commit mutex (destroy and
re-initialise), so that SAFE_MUTEX lock order check does not become
confused when the mutex is re-used for a different purpose.
setting of innodb_io_capacity_max
(a) Changed the behaviour so that if you set innodb_io_capacity to a
value > innodb_io_capacity_max that the value is accepted AND
that innodb_io_capacity_max = innodb_io_capacity * 2.
(b) If someone wants to reduce innodb_io_capacity_max and
reduce it below innodb_io_capacity then innodb_io_capacity
should be reduced to the same level as innodb_io_capacity_max.
In both cases give a warning to user.
Analysis: InnoDB error monitor is responsible to call every second
sync_arr_wake_threads_if_sema_free() to wake up possible hanging
threads if they are missed in mutex_signal_object. This is not
possible if error monitor itself is on mutex/semaphore wait. We
should avoid all unnecessary mutex/semaphore waits on error monitor.
Currently error monitor calls function buf_flush_stat_update()
that calls log_get_lsn() function and there we will try to get
log_sys mutex. Better, solution for error monitor is that in
buf_flush_stat_update() we will try to get lsn with
mutex_enter_nowait() and if we did not get mutex do not update
the stats.
Fix: Use log_get_lsn_nowait() function on buf_flush_stat_update()
function. If returned lsn is 0, we do not update flush stats.
log_get_lsn_nowait() will use mutex_enter_nowait() and if
we get mutex we return a correct lsn if not we return 0.
on work-amd64-valgrind.
Fixed issue by finding out first the current used priority
for both treads and using that seeing did we really change
the priority or not.
main.information_schema: added a condition to the query to exclude perfschema tables
main.information_schema_all_engines: added a call to the include file to check for the presence of perfschema
The bug is not very important per se, but it was helpful to move
Item_func_strcmp out of Item_bool_func2 (to Item_int_func),
for the purposes of "MDEV-4912 Add a plugin to field types (column types)".
The test runs a query in one thread, then in another queries the processlist
and expects to find the first thread in the COM_SLEEP state. The problem is
that the thread signals completion to the client before changing to COM_SLEEP
state, so there is a window where the other thread can see the wrong state.
A previous attempt to fix this was ineffective. It set a DEBUG_SYNC to handle
proper waiting, but unfortunately that DEBUG_SYNC point ended up triggering
already at the end of SET DEBUG_SYNC=xxx, so the wait was ineffective.
Fix it properly now (hopefully) by ensuring that we wait for the DEBUG_SYNC
point to trigger at the end of the SELECT SLEEP(), not just at the end of
SET DEBUG_SYNC=xxx.
(Backport to 5.3)
(Attempt #2)
- Don't attempt to use BKA for materialized derived tables. The
table is neither filled nor fully opened yet, so attempt to
call handler->multi_range_read_info() causes crash.
(Backport to 5.3)
(variant #2, with fixed coding style)
- Make Mrr_ordered_index_reader::resume_read() restore index position
only if it was saved before with Mrr_ordered_index_reader::interrupt_read().
- TABLE::create_key_part_by_field() should not set PART_KEY_FLAG in field->flags
= The reason is that it is used by hash join code which calls it to create a hash
table lookup structure. It doesn't create a real index.
= Another caller of the function is TABLE::add_tmp_key(). Made it to set the flag itself.
- The differences in join_cache.result could also be observed before this patch: one
could put "FLUSH TABLES" before the queries and get exactly the same difference.
Merged Facebook commit ec1aac68c74f3c1e558d057c4c9fcfe6edbbea93
authored by Steaphan Greene from https://github.com/facebook/mysql-5.6
In C++11, "" is not parsed as before. So "A""B" is not the same as "AB".
Instead, whitespace is required, like: "A" "B"
Merged Facebook commit dd2d11be7aaf3be270e740fb95cbc4eacb52f4d7
authored by Rongrong Zhong from https://github.com/facebook/mysql-5.6
This fixes MySQL Bug #68220 innodb_rows_updated is misleading on slave
http://bugs.mysql.com/bug.php?id=68220
Added innodb_system_rows_read/inserted/updated/deleted counters
that are the equivalent of innodb_rows_* but that only account for
changes made to system databases (mysql, information_schame and
preformance_schema). These counters will be used on slaves to
differentiated the updates made on system databases from those made on
user databases.
innodb_rows_* status counters are not updated when innodb_system_rows_*
are updated.
dd2d11be7a
Merged Facebook commit ecff018632c6db49bad73d9233c3cdc9f41430e9
authored by Steaphan Greene from https://github.com/facebook/mysql-5.6
This change is to fix: http://bugs.mysql.com/62534
This makes innodb_max_dirty_pages_pct a double with min,default,max values
0.001, 75, 99.999.
This also makes innodb_max_dirty_pages_pct_lwm and adaptive_flushing_lwm
doubles, as these sysvars are inter-dependent.
Added more to the BUFFER POOL AND MEMORY section of SHOW INNODB STATUS:
Percent pages dirty: X.X
This is all n_dirty_pages / used_pages
Percent all pages dirty: X.X
This is all n_dirty_pages / all-pages
Max dirty pages percent: X.X
This is innodb_max_dirty_pages_pct
Also changed all of buf from 2 to 3 digits of precision (%.2f -> %.3f).
Merge Facebook commit f981a51a47519b0ba527917887f8adc6df9ae147
authored by Steaphan Greene from https://github.com/facebook/mysql-5.6.
This just moves some structure definitions from inside a
single .cc file to a shared .h file, with a few tweaks to
allow these structures to be shared.
On its own, it should have no actual effect. This is needed later.
Merge Facebook commit 25295d003cb0c17aa8fb756523923c77250b3294
authored by Steaphan Greene from https://github.com/facebook/mysql-5.6
This adds a pointer to the trx to each mtr.
This allows the trx to be accessed in parts of the code
where it was otherwise not available. This is needed later.
compressed pages
buf_flush_LRU() returns the number of pages processed. There are
two types of processing that can happen. A page can get evicted or
a page can get flushed. These two numbers are quite distinct and
should not be mixed.
