The problem seems to be a deadlock between KILL command execution
and BF abort issued by an applier, where:
* KILL has locked victim's LOCK_thd_kill and LOCK_thd_data.
* Applier has innodb side global lock mutex and victim trx mutex.
* KILL is calling innobase_kill_query, and is blocked by innodb
global lock mutex.
* Applier is in wsrep_innobase_kill_one_trx and is blocked by
victim's LOCK_thd_kill.
The fix in this commit removes the TOI replication of KILL command
and makes KILL execution less intrusive operation. Aborting the
victim happens now by using awake_no_mutex() and ha_abort_transaction().
If the KILL happens when the transaction is committing, the
KILL operation is postponed to happen after the statement
has completed in order to avoid KILL to interrupt commit
processing.
Notable changes in this commit:
* wsrep client connections's error state may remain sticky after
client connection is closed. This error message will then pop
up for the next client session issuing first SQL statement.
This problem raised with test galera.galera_bf_kill.
The fix is to reset wsrep client error state, before a THD is
reused for next connetion.
* Release THD locks in wsrep_abort_transaction when locking
innodb mutexes. This guarantees same locking order as with applier
BF aborting.
* BF abort from MDL was changed to do BF abort on server/wsrep-lib
side first, and only then do the BF abort on InnoDB side. This
removes the need to call back from InnoDB for BF aborts which originate
from MDL and simplifies the locking.
* Removed wsrep_thd_set_wsrep_aborter() from service_wsrep.h.
The manipulation of the wsrep_aborter can be done solely on
server side. Moreover, it is now debug only variable and
could be excluded from optimized builds.
* Remove LOCK_thd_kill from wsrep_thd_LOCK/UNLOCK to allow more
fine grained locking for SR BF abort which may require locking
of victim LOCK_thd_kill. Added explicit call for
wsrep_thd_kill_LOCK/UNLOCK where appropriate.
* Wsrep-lib was updated to version which allows external
locking for BF abort calls.
Changes to MTR tests:
* Disable galera_bf_abort_group_commit. This test is going to
be removed (MDEV-30855).
* Record galera_gcache_recover_manytrx as result file was incomplete.
Trivial change.
* Make galera_create_table_as_select more deterministic:
Wait until CTAS execution has reached MDL wait for multi-master
conflict case. Expected error from multi-master conflict is
ER_QUERY_INTERRUPTED. This is because CTAS does not yet have open
wsrep transaction when it is waiting for MDL, query gets interrupted
instead of BF aborted. This should be addressed in separate task.
* A new test galera_kill_group_commit to verify correct behavior
when KILL is executed while the transaction is committing.
Co-authored-by: Seppo Jaakola <seppo.jaakola@iki.fi>
Co-authored-by: Jan Lindström <jan.lindstrom@galeracluster.com>
Signed-off-by: Julius Goryavsky <julius.goryavsky@mariadb.com>
Assertion `thd->mdl_context.is_lock_owner()` fires when a client is
disconnected, while transaction and and a table is opened through
`HANDLER` interface.
Reason for the assertion is that when a connection closes, its ongoing
transaction is eventually rolled back in
`Wsrep_client_state::bf_rollback()`. This method also releases explicit
which are expected to survive beyond the transaction lifetime.
This patch also removes calls to `mysql_ull_cleanup()`. User level
locks are not supported in combination with Galera, making these calls
unnecessary.
With binlogs enabled, debug assertion ut_ad(xid_seqno > wsrep_seqno)
fired in trx_rseg_update_wsrep_checkpoint() when an applier thread
synced the seqno out of order for write set which had failed
certification. This was caused by releasing commit
order too early when binlogs were on, allowing group
commit to run in parallel and commit following transactions
too early.
Fixed by extending the commit order critical section to cover
call to wsrep_set_SE_checkpoint() also when binlogs are on.
Signed-off-by: Julius Goryavsky <julius.goryavsky@mariadb.com>
We should not call mdl_context.release_explicit_locks() in
Wsrep_client_service::bf_rollback() if client is quiting because
it will be done again in THD::cleanup().
Note that problem with GET_LOCK() / RELEASE_LOCK() will be fixed
on MDEV-30473.
Created mtr test for reproducing the crash
Developed actual fix for the issue.
Setting THD::system_thread_info.rpl_sql_info for replayer thread,
same way as it is handled for appliers.
Recorded test result, with the fix
Reviewed-by: Jan Lindström <jan.lindstrom@mariadb.com>
There are separate flags DBUG_OFF for disabling the DBUG facility
and ENABLED_DEBUG_SYNC for enabling the DEBUG_SYNC facility.
Let us allow debug builds without DEBUG_SYNC.
Note: For CMAKE_BUILD_TYPE=Debug, CMakeLists.txt will continue to
define ENABLED_DEBUG_SYNC.
Making changes to wsrep_mysqld.h causes large parts of server code to
be recompiled. The reason is that wsrep_mysqld.h is included by
sql_class.h, even tough very little of wsrep_mysqld.h is needed in
sql_class.h. This commit introduces a new header file, wsrep_on.h,
which is meant to be included from sql_class.h, and contains only
macros and variable declarations used to determine whether wsrep is
enabled.
Also, header wsrep.h should only contain definitions that are also
used outside of sql/. Therefore, move WSREP_TO_ISOLATION* and
WSREP_SYNC_WAIT macros to wsrep_mysqld.h.
Reviewed-by: Jan Lindström <jan.lindstrom@mariadb.com>
When thread is BF aborted by high priority service, ULL (user level
locks need to be removed and released). Calling directly release of lock for
MDL_EXPLICIT type doesn't clear also `thd->ull_hash`. Method
`mysql_ull_cleanup` will properly clear all information about ULL locks
for thread.
Reviewed-by: Jan Lindström <jan.lindstrom@mariadb.com>
In a rebase of the merge, two preceding commits were accidentally reverted:
commit 112b23969a (MDEV-26308)
commit ac2857a5fb (MDEV-25717)
Thanks to Daniele Sciascia for noticing this.
Contains following fixes:
* allow TOI commands to timeout while trying to acquire TOI with
override lock_wait_timeout with a LONG_TIMEOUT only after
succesfully entering TOI
* only ignore lock_wait_timeout on TOI
* fix galera_split_brain test as TOI operation now returns ER_LOCK_WAIT_TIMEOUT after lock_wait_timeout
* explicitly test for TOI
Reviewed-by: Jan Lindström <jan.lindstrom@mariadb.com>
Handling of write sets, which fail in certification happens differently than
with write sets which pass certification. When certification fails, the
write set applying can be skipped and applier needs only to take care of
wsrep XID checkpointing. With current implementation, this can rush ahead
of wsrep XID checkpointing of successful write sets.
The fix in this PR registers wsrep XID checkpointing of certification failure cases in group commit,
which guarantees that XID ceckpointing order is synchronized with real committing transactions.
Reviewed-by: Jan Lindström <jan.lindstrom@mariadb.com>
If log_slave_updates==OFF, wsrep applier threads used to be configured
with option: thd->variables.option_bits&= ~(OPTION_BIN_LOG);
(i.e. like sql_log_bin=ON). And this was regardless of log-bin configuration.
With this, having configuration of: --log-bin && --log-slave-updates=OFF,
local threads used binlogging, but applier threads did not. And further:
local threads went through binlog group commit, while applier threads did
direct commits. This resulted in situation, where applier threads entered
earlier in wsrep XID checkpointing, and could sync their wsrep XID out of order.
Later local thread commit would see that higher seqno was already checkpointed,
and fire an assert because of this.
As a fix, applier threads are now forced to enable binlogging regardless of
log-slave-updates configuration.
This PR comes with new mtr test: galera.MDEV-24327, which causes a scenario
where applier transaction is applied and committed while earlier local transaction
is parked before commit order monitor enter. A buggy mariadb versoin would fail
for assertion because of wsrep XID checkpoint order violation.
If statement still contains changes it must be committed
before actual transaction is committed.
This assertion was found using randgen and happens only on
applier. No repeatable test case found.
Support for galera GTID consistency thru cluster. All nodes in cluster
should have same GTID for replicated events which are originating from cluster.
Cluster originating commands need to contain sequential WSREP GTID seqno
Ignore manual setting of gtid_seq_no=X.
In master-slave scenario where master is non galera node replicated GTID is
replicated and is preserved in all nodes.
To have this - domain_id, server_id and seqnos should be same on all nodes.
Node which bootstraps the cluster, to achieve this, sends domain_id and
server_id to other nodes and this combination is used to write GTID for events
that are replicated inside cluster.
Cluster nodes that are executing non replicated events are going to have different
GTID than replicated ones, difference will be visible in domain part of gtid.
With wsrep_gtid_domain_id you can set domain_id for WSREP cluster.
Functions WSREP_LAST_WRITTEN_GTID, WSREP_LAST_SEEN_GTID and
WSREP_SYNC_WAIT_UPTO_GTID now works with "native" GTID format.
Fixed galera tests to reflect this chances.
Add variable to manually update WSREP GTID seqno in cluster
Add variable to manipulate and change WSREP GTID seqno. Next command
originating from cluster and on same thread will have set seqno and
cluster should change their internal counter to it's value.
Behavior is same as using @@gtid_seq_no for non WSREP transaction.
A certification failure followed by a clean shutdown would cause an
inconsistency between the sequence number stored in innodb and the
sequence number stored in provider.
This happened both in the case of local certification failure, and in
the case where dummy writeset is applied.
The fix consists of:
- updating wsrep position after dummy writeset is delivered in
`Wsrep_high_priority_service::log_dummy_write_set()`
- updating wsrep position while releasing commit order in wsrep-lib
side
Added two tests which stress the situation where a server is shutdown
after a certification failure.
Assertion failed in wsrep-lib after transaction replay which
failed due to conflict in certification.
- Implemented reproducible test case MDEV-20793 to reproduce the crash.
- Fixed wsrep-lib to deal with certification error during replay.
- Fixes a situation in which a thread gets BF aborted and does not send the reply back to
the client, even though the connection is still alive. That caused
both sides to hang waiting for the next message. Now we explicitly
check that the connection is still alive.
- MTR test for the above
- Replaced thd->killed assignments to thd->reset_kill_query where applicable.
MariaDB 10.4 was crashing when thread-handling was set to
pool-of-threads and wsrep was enabled.
There were two apparent reasons for the crash:
- Connection handling in threadpool_common.cc was missing calls to
control wsrep client state.
- Thread specific storage which contains thread variables (THR_KEY_mysys)
was not handled appropriately by wsrep patch when pool-of-threads
was configured.
This patch addresses the above issues in the following way:
- Wsrep client state open/close was moved in thd_prepare_connection() and
end_connection() to have common handling for one-thread-per-connection
and pool-of-threads.
- Thread local storage handling in wsrep patch was reworked by introducing
set of wsrep_xxx_threadvars() calls which replace calls to
THD store_globals()/reset_globals() and deal with thread handling
specifics internally.
Wsrep-lib was updated to version which relaxes internal concurrency
related sanity checks.
Rollback code from wsrep_rollback_process() was extracted to separate calls
for better readability.
Post rollback thread was removed as it was completely unused.
* Collect and pass apply error data to provider
* Rollback failed transaction and continue operation if provider returns
SUCCESS
* MTR tests for inconsistency voting
