Mutex order violation when wsrep bf thread kills a conflicting trx,
the stack is
wsrep_thd_LOCK()
wsrep_kill_victim()
lock_rec_other_has_conflicting()
lock_clust_rec_read_check_and_lock()
row_search_mvcc()
ha_innobase::index_read()
ha_innobase::rnd_pos()
handler::ha_rnd_pos()
handler::rnd_pos_by_record()
handler::ha_rnd_pos_by_record()
Rows_log_event::find_row()
Update_rows_log_event::do_exec_row()
Rows_log_event::do_apply_event()
Log_event::apply_event()
wsrep_apply_events()
and mutexes are taken in the order
lock_sys->mutex -> victim_trx->mutex -> victim_thread->LOCK_thd_data
When a normal KILL statement is executed, the stack is
innobase_kill_query()
kill_handlerton()
plugin_foreach_with_mask()
ha_kill_query()
THD::awake()
kill_one_thread()
and mutexes are
victim_thread->LOCK_thd_data -> lock_sys->mutex -> victim_trx->mutex
This patch is the plan D variant for fixing potetial mutex locking
order exercised by BF aborting and KILL command execution.
In this approach, KILL command is replicated as TOI operation.
This guarantees total isolation for the KILL command execution
in the first node: there is no concurrent replication applying
and no concurrent DDL executing. Therefore there is no risk of
BF aborting to happen in parallel with KILL command execution
either. Potential mutex deadlocks between the different mutex
access paths with KILL command execution and BF aborting cannot
therefore happen.
TOI replication is used, in this approach, purely as means
to provide isolated KILL command execution in the first node.
KILL command should not (and must not) be applied in secondary
nodes. In this patch, we make this sure by skipping KILL
execution in secondary nodes, in applying phase, where we
bail out if applier thread is trying to execute KILL command.
This is effective, but skipping the applying of KILL command
could happen much earlier as well.
This also fixed unprotected calls to wsrep_thd_abort
that will use wsrep_abort_transaction. This is fixed
by holding THD::LOCK_thd_data while we abort transaction.
Reviewed-by: Jan Lindström <jan.lindstrom@mariadb.com>
make BACKUP STAGE behave as FTWRL, desyncing and pausing the node
to prevent BF threads (appliers) from interfering with blocking stages.
This is needed because BF threads don't respect BACKUP MDL locks.
Reviewed-by: Jan Lindström <jan.lindstrom@mariadb.com>
Reformulate mark_columns_used_by_index* function family in a more laconic
way:
mark_columns_used_by_index -> mark_index_columns
mark_columns_used_by_index_for_read_no_reset -> mark_index_columns_for_read
mark_columns_used_by_index_no_reset -> mark_index_columns_no_reset
static mark_index_columns -> do_mark_index_columns
In the code existed just before this patch binding of a table reference to
the specification of the corresponding CTE happens in the function
open_and_process_table(). If the table reference is not the first in the
query the specification is cloned in the same way as the specification of
a view is cloned for any reference of the view. This works fine for
standalone queries, but does not work for stored procedures / functions
for the following reason.
When the first call of a stored procedure/ function SP is processed the
body of SP is parsed. When a query of SP is parsed the info on each
encountered table reference is put into a TABLE_LIST object linked into
a global chain associated with the query. When parsing of the query is
finished the basic info on the table references from this chain except
table references to derived tables and information schema tables is put
in one hash table associated with SP. When parsing of the body of SP is
finished this hash table is used to construct TABLE_LIST objects for all
table references mentioned in SP and link them into the list of such
objects passed to a pre-locking process that calls open_and_process_table()
for each table from the list.
When a TABLE_LIST for a view is encountered the view is opened and its
specification is parsed. For any table reference occurred in
the specification a new TABLE_LIST object is created to be included into
the list for pre-locking. After all objects in the pre-locking have been
looked through the tables mentioned in the list are locked. Note that the
objects referenced CTEs are just skipped here as it is impossible to
resolve these references without any info on the context where they occur.
Now the statements from the body of SP are executed one by one that.
At the very beginning of the execution of a query the tables used in the
query are opened and open_and_process_table() now is called for each table
reference mentioned in the list of TABLE_LIST objects associated with the
query that was built when the query was parsed.
For each table reference first the reference is checked against CTEs
definitions in whose scope it occurred. If such definition is found the
reference is considered resolved and if this is not the first reference
to the found CTE the the specification of the CTE is re-parsed and the
result of the parsing is added to the parsing tree of the query as a
sub-tree. If this sub-tree contains table references to other tables they
are added to the list of TABLE_LIST objects associated with the query in
order the referenced tables to be opened. When the procedure that opens
the tables comes to the TABLE_LIST object created for a non-first
reference to a CTE it discovers that the referenced table instance is not
locked and reports an error.
Thus processing non-first table references to a CTE similar to how
references to view are processed does not work for queries used in stored
procedures / functions. And the main problem is that the current
pre-locking mechanism employed for stored procedures / functions does not
allow to save the context in which a CTE reference occur. It's not trivial
to save the info about the context where a CTE reference occurs while the
resolution of the table reference cannot be done without this context and
consequentially the specification for the table reference cannot be
determined.
This patch solves the above problem by moving resolution of all CTE
references at the parsing stage. More exactly references to CTEs occurred in
a query are resolved right after parsing of the query has finished. After
resolution any CTE reference it is marked as a reference to to derived
table. So it is excluded from the hash table created for pre-locking used
base tables and view when the first call of a stored procedure / function
is processed.
This solution required recursive calls of the parser. The function
THD::sql_parser() has been added specifically for recursive invocations of
the parser.
In the code existed just before this patch binding of a table reference to
the specification of the corresponding CTE happens in the function
open_and_process_table(). If the table reference is not the first in the
query the specification is cloned in the same way as the specification of
a view is cloned for any reference of the view. This works fine for
standalone queries, but does not work for stored procedures / functions
for the following reason.
When the first call of a stored procedure/ function SP is processed the
body of SP is parsed. When a query of SP is parsed the info on each
encountered table reference is put into a TABLE_LIST object linked into
a global chain associated with the query. When parsing of the query is
finished the basic info on the table references from this chain except
table references to derived tables and information schema tables is put
in one hash table associated with SP. When parsing of the body of SP is
finished this hash table is used to construct TABLE_LIST objects for all
table references mentioned in SP and link them into the list of such
objects passed to a pre-locking process that calls open_and_process_table()
for each table from the list.
When a TABLE_LIST for a view is encountered the view is opened and its
specification is parsed. For any table reference occurred in
the specification a new TABLE_LIST object is created to be included into
the list for pre-locking. After all objects in the pre-locking have been
looked through the tables mentioned in the list are locked. Note that the
objects referenced CTEs are just skipped here as it is impossible to
resolve these references without any info on the context where they occur.
Now the statements from the body of SP are executed one by one that.
At the very beginning of the execution of a query the tables used in the
query are opened and open_and_process_table() now is called for each table
reference mentioned in the list of TABLE_LIST objects associated with the
query that was built when the query was parsed.
For each table reference first the reference is checked against CTEs
definitions in whose scope it occurred. If such definition is found the
reference is considered resolved and if this is not the first reference
to the found CTE the the specification of the CTE is re-parsed and the
result of the parsing is added to the parsing tree of the query as a
sub-tree. If this sub-tree contains table references to other tables they
are added to the list of TABLE_LIST objects associated with the query in
order the referenced tables to be opened. When the procedure that opens
the tables comes to the TABLE_LIST object created for a non-first
reference to a CTE it discovers that the referenced table instance is not
locked and reports an error.
Thus processing non-first table references to a CTE similar to how
references to view are processed does not work for queries used in stored
procedures / functions. And the main problem is that the current
pre-locking mechanism employed for stored procedures / functions does not
allow to save the context in which a CTE reference occur. It's not trivial
to save the info about the context where a CTE reference occurs while the
resolution of the table reference cannot be done without this context and
consequentially the specification for the table reference cannot be
determined.
This patch solves the above problem by moving resolution of all CTE
references at the parsing stage. More exactly references to CTEs occurred in
a query are resolved right after parsing of the query has finished. After
resolution any CTE reference it is marked as a reference to to derived
table. So it is excluded from the hash table created for pre-locking used
base tables and view when the first call of a stored procedure / function
is processed.
This solution required recursive calls of the parser. The function
THD::sql_parser() has been added specifically for recursive invocations of
the parser.
So we are having a race condition of three of threads, resulting in a
deadlock backoff in purge, which is unexpected.
More precisely, the following happens:
T1: NOCOPY ALTER TABLE begins, and eventually it holds MDL_SHARED_NO_WRITE
lock;
T2: FLUSH TABLES begins. it sets share->tdc->flushed = true
T3: purge on a record with virtual column begins. it is going to open a
table. MDL_SHARED_READ lock is acquired therefore.
Since share->tdc->flushed is set, it waits for a TDC purge end.
T1: is going to elevate MDL LOCK to exclusive and therefore has to set
other waiters to back off.
T3: receives VICTIM status, reports a DEADLOCK, sets OT_BACKOFF_AND_RETRY
to Open_table_context::m_action
My fix is to allow opening table in purge while flushing. It is already
done the same way in other maintainance facilities like REPAIR TABLE.
Another way would be making an actual backoff, but Open_table_context
does not allow to distinguish it from other failure types, which still
seem to be unexpected. Making this would require hacking into
Open_table_context interface for no benefit, in comparison to passing
MYSQL_OPEN_IGNORE_FLUSH during table open.
The easiest way to compile and test the server with UBSAN is to run:
./BUILD/compile-pentium64-ubsan
and then run mysql-test-run.
After this commit, one should be able to run this without any UBSAN
warnings. There is still a few compiler warnings that should be fixed
at some point, but these do not expose any real bugs.
The 'special' cases where we disable, suppress or circumvent UBSAN are:
- ref10 source (as here we intentionally do some shifts that UBSAN
complains about.
- x86 version of optimized int#korr() methods. UBSAN do not like unaligned
memory access of integers. Fixed by using byte_order_generic.h when
compiling with UBSAN
- We use smaller thread stack with ASAN and UBSAN, which forced me to
disable a few tests that prints the thread stack size.
- Verifying class types does not work for shared libraries. I added
suppression in mysql-test-run.pl for this case.
- Added '#ifdef WITH_UBSAN' when using integer arithmetic where it is
safe to have overflows (two cases, in item_func.cc).
Things fixed:
- Don't left shift signed values
(byte_order_generic.h, mysqltest.c, item_sum.cc and many more)
- Don't assign not non existing values to enum variables.
- Ensure that bool and enum values are properly initialized in
constructors. This was needed as UBSAN checks that these types has
correct values when one copies an object.
(gcalc_tools.h, ha_partition.cc, item_sum.cc, partition_element.h ...)
- Ensure we do not called handler functions on unallocated objects or
deleted objects.
(events.cc, sql_acl.cc).
- Fixed bugs in Item_sp::Item_sp() where we did not call constructor
on Query_arena object.
- Fixed several cast of objects to an incompatible class!
(Item.cc, Item_buff.cc, item_timefunc.cc, opt_subselect.cc, sql_acl.cc,
sql_select.cc ...)
- Ensure we do not do integer arithmetic that causes over or underflows.
This includes also ++ and -- of integers.
(Item_func.cc, Item_strfunc.cc, item_timefunc.cc, sql_base.cc ...)
- Added JSON_VALUE_UNITIALIZED to json_value_types and ensure that
value_type is initialized to this instead of to -1, which is not a valid
enum value for json_value_types.
- Ensure we do not call memcpy() when second argument could be null.
- Fixed that Item_func_str::make_empty_result() creates an empty string
instead of a null string (safer as it ensures we do not do arithmetic
on null strings).
Other things:
- Changed struct st_position to an OBJECT and added an initialization
function to it to ensure that we do not copy or use uninitialized
members. The change to a class was also motived that we used "struct
st_position" and POSITION randomly trough the code which was
confusing.
- Notably big rewrite in sql_acl.cc to avoid using deleted objects.
- Changed in sql_partition to use '^' instead of '-'. This is safe as
the operator is either 0 or 0x8000000000000000ULL.
- Added check for select_nr < INT_MAX in JOIN::build_explain() to
avoid bug when get_select() could return NULL.
- Reordered elements in POSITION for better alignment.
- Changed sql_test.cc::print_plan() to use pointers instead of objects.
- Fixed bug in find_set() where could could execute '1 << -1'.
- Added variable have_sanitizer, used by mtr. (This variable was before
only in 10.5 and up). It can now have one of two values:
ASAN or UBSAN.
- Moved ~Archive_share() from ha_archive.cc to ha_archive.h and marked
it virtual. This was an effort to get UBSAN to work with loaded storage
engines. I kept the change as the new place is better.
- Added in CONNECT engine COLBLK::SetName(), to get around a wrong cast
in tabutil.cpp.
- Added HAVE_REPLICATION around usage of rgi_slave, to get embedded
server to compile with UBSAN. (Patch from Marko).
- Added #ifdef for powerpc64 to avoid a bug in old gcc versions related
to integer arithmetic.
Changes that should not be needed but had to be done to suppress warnings
from UBSAN:
- Added static_cast<<uint16_t>> around shift to get rid of a LOT of
compiler warnings when using UBSAN.
- Had to change some '/' of 2 base integers to shift to get rid of
some compile time warnings.
Reviewed by:
- Json changes: Alexey Botchkov
- Charset changes in ctype-uca.c: Alexander Barkov
- InnoDB changes & Embedded server: Marko Mäkelä
- sql_acl.cc changes: Vicențiu Ciorbaru
- build_explain() changes: Sergey Petrunia
* reuse the loop in THD::abort_current_cond_wait, don't duplicate it
* find_thread_by_id should return whatever it has found, it's the
caller's task not to kill COM_DAEMON (if the caller's a killer)
and other minor changes
Some DML operations on tables having unique secondary keys cause scanning
in the secondary index, for instance to find potential unique key violations
in the seconday index. This scanning may involve GAP locking in the index.
As this locking happens also when applying replication events in high priority
applier threads, there is a probabality for lock conflicts between two wsrep
high priority threads.
This PR avoids lock conflicts of high priority wsrep threads, which do
secondary index scanning e.g. for duplicate key detection.
The actual fix is the patch in sql_class.cc:thd_need_ordering_with(), where
we allow relaxed GAP locking protocol between wsrep high priority threads.
wsrep high priority threads (replication appliers, replayers and TOI processors)
are ordered by the replication provider, and they will not need serializability
support gained by secondary index GAP locks.
PR contains also a mtr test, which exercises a scenario where two replication
applier threads have a false positive conflict in GAP of unique secondary index.
The conflicting local committing transaction has to replay, and the test verifies
also that the replaying phase will not conflict with the latter repllication applier.
Commit also contains new test scenario for galera.galera_UK_conflict.test,
where replayer starts applying after a slave applier thread, with later seqno,
has advanced to commit phase. The applier and replayer have false positive GAP
lock conflict on secondary unique index, and replayer should ignore this.
This test scenario caused crash with earlier version in this PR, and to fix this,
the secondary index uniquenes checking has been relaxed even further.
Now innodb trx_t structure has new member: bool wsrep_UK_scan, which is set to
true, when high priority thread is performing unique secondary index scanning.
The member trx_t::wsrep_UK_scan is defined inside WITH_WSREP directive, to make
it possible to prepare a MariaDB build where this additional trx_t member is
not present and is not used in the code base. trx->wsrep_UK_scan is set to true
only for the duration of function call for: lock_rec_lock() trx->wsrep_UK_scan
is used only in lock_rec_has_to_wait() function to relax the need to wait if
wsrep_UK_scan is set and conflicting transaction is also high priority.
Reviewed-by: Jan Lindström <jan.lindstrom@mariadb.com>
Server part:
kill_handlerton() was accessing thd->ha_data[] for some other thd,
while it could be concurrently modified by its owner thd.
protect thd->ha_data[] modifications with a mutex.
require this mutex when accessing thd->ha_data[] from kill_handlerton.
InnoDB part:
on close_connection, detach trx from thd before freeing the trx
This reverts the server part of the commit 775fccea0
but keeps InnoDB part (which reverted MDEV-17092 5530a93f4).
So after this both MDEV-23536 and MDEV-17092 are reverted,
and the original bug is resurrected.
A race condition may occur between the execution of transaction commit,
and an execution of a KILL statement that would attempt to abort that
transaction.
MDEV-17092 worked around this race condition by modifying InnoDB code.
After that issue was closed, Sergey Vojtovich pointed out that this
race condition would better be fixed above the storage engine layer:
If you look carefully into the above, you can conclude that
thd->free_connection() can be called concurrently with
KILL/thd->awake(). Which is the bug. And it is partially fixed in
THD::~THD(), that is destructor waits for KILL completion:
Fix: Add necessary mutex operations to THD::free_connection()
and move WSREP specific code also there. This ensures that no
one is using THD while we do free_connection(). These mutexes
will also ensures that there can't be concurrent KILL/THD::awake().
innobase_kill_query
We can now remove usage of trx_sys_mutex introduced on MDEV-17092.
trx_t::free()
Poison trx->state and trx->mysql_thd
This patch is validated with an RQG run similar to the one that
reproduced MDEV-17092.
Analysis: select into outfile creates files everytime with 666 permission,
regardsless if umask environment variables and umask settings on OS level.
It seems hardcoded.
Fix: change 0666 to 0644 which will let anybody consume the file but not
change it.
The reason for the failure is that
thd->mdl_context.release_transactional_locks()
was called after commit & rollback even in cases where the current
transaction is still active.
For 10.2, 10.3 and 10.4 the fix is simple:
- Replace all calls to thd->mdl_context.release_transactional_locks() with
thd->release_transactional_locks(). The thd function will only call
the mdl_context function if there are no active transactional locks.
In 10.6 we will better fix where we will change the return value for
some trans_xxx() functions to indicate if transaction did close the
transaction or not. This will avoid the need of the indirect call.
Other things:
- trans_xa_commit() and trans_xa_rollback() will automatically
call release_transactional_locks() if the transaction is closed.
- We can't do that for the other functions as the caller of many of these
are doing additional work (like close_thread_tables) before calling
release_transactional_locks().
- Added missing abort_result_set() and missing DBUG_RETURN in
select_create::send_eof()
- Fixed wrong indentation in injector::transaction::commit()
This follows up commit
commit 94a520ddbe and
commit 7c5519c12d.
After these changes, the default test suites on a
cmake -DWITH_UBSAN=ON build no longer fail due to passing
null pointers as parameters that are declared to never be null,
but plenty of other runtime errors remain.
MDEV-20945: BACKUP UNLOCK + FTWRL assertion failure | SIGSEGV in I_P_List
from MDL_context::release_lock on INSERT w/ BACKUP LOCK (on optimized
builds) | Assertion `ticket->m_duration == MDL_EXPLICIT' failed
BACKUP LOCK behavior is modified so it won't be used wrong:
- BACKUP LOCK should commit any active transactions.
- BACKUP LOCK should not be allowed in stored procedures.
- When BACKUP LOCK is active, don't allow any DDL's for that connection.
- FTWRL is forbidden on the same connection while BACKUP LOCK is active.
Reviewed-by: monty@mariadb.com
MDEV-21953 deadlock between BACKUP STAGE BLOCK_COMMIT and parallel
replication
Fixed by partly reverting MDEV-21953 to put back MDL_BACKUP_COMMIT locking
before log_and_order.
The original problem for MDEV-21953 was that while a thread was waiting in
for another threads to commit in 'log_and_order', it had the
MDL_BACKUP_COMMIT lock. The backup thread was waiting to get the
MDL_BACKUP_WAIT_COMMIT lock, which blocks all new MDL_BACKUP_COMMIT locks.
This causes a deadlock as the waited-for thread can never get past the
MDL_BACKUP_COMMIT lock in ha_commit_trans.
The main part of the bug fix is to release the MDL_BACKUP_COMMIT lock while
a thread is waiting for other 'previous' threads to commit. This ensures
that no transactional thread keeps MDL_BACKUP_COMMIT while waiting, which
ensures that there are no deadlocks anymore.
- Adding optional qualifiers to data types:
CREATE TABLE t1 (a schema.DATE);
Qualifiers now work only for three pre-defined schemas:
mariadb_schema
oracle_schema
maxdb_schema
These schemas are virtual (hard-coded) for now, but may turn into real
databases on disk in the future.
- mariadb_schema.TYPE now always resolves to a true MariaDB data
type TYPE without sql_mode specific translations.
- oracle_schema.DATE translates to MariaDB DATETIME.
- maxdb_schema.TIMESTAMP translates to MariaDB DATETIME.
- Fixing SHOW CREATE TABLE to use a qualifier for a data type TYPE
if the current sql_mode translates TYPE to something else.
The above changes fix the reported problem, so this script:
SET sql_mode=ORACLE;
CREATE TABLE t2 AS SELECT mariadb_date_column FROM t1;
is now replicated as:
SET sql_mode=ORACLE;
CREATE TABLE t2 (mariadb_date_column mariadb_schema.DATE);
and the slave can unambiguously treat DATE as the true MariaDB DATE
without ORACLE specific translation to DATETIME.
Similar,
SET sql_mode=MAXDB;
CREATE TABLE t2 AS SELECT mariadb_timestamp_column FROM t1;
is now replicated as:
SET sql_mode=MAXDB;
CREATE TABLE t2 (mariadb_timestamp_column mariadb_schema.TIMESTAMP);
so the slave treats TIMESTAMP as the true MariaDB TIMESTAMP
without MAXDB specific translation to DATETIME.
The issue was:
T1, a parallel slave worker thread, is waiting for another worker thread to
commit. While waiting, it has the MDL_BACKUP_COMMIT lock.
T2, working for mariabackup, is doing BACKUP STAGE BLOCK_COMMIT and blocks
all commits.
This causes a deadlock as the thread T1 is waiting for can't commit.
Fixed by moving locking of MDL_BACKUP_COMMIT from ha_commit_trans() to
commit_one_phase_2()
Other things:
- Added a new argument to ha_comit_one_phase() to signal if the
transaction was a write transaction.
- Ensured that ha_maria::implicit_commit() is always called under
MDL_BACKUP_COMMIT. This code is not needed in 10.5
- Ensure that MDL_Request values 'type' and 'ticket' are always
initialized. This makes it easier to check the state of the MDL_Request.
- Moved thd->store_globals() earlier in handle_rpl_parallel_thread() as
thd->init_for_queries() could use a MDL that could crash if store_globals
where not called.
- Don't call ha_enable_transactions() in THD::init_for_queries() as this
is both slow (uses MDL locks) and not needed.
* Allocate items on thd->mem_root while refixing vcol exprs
* Make vcol tree changes register and roll them back after the statement is executed.
Explanation:
Due to collation implementation specifics an Item tree could change while fixing.
The tricky thing here is to make it on a proper arena.
It's usually not a problem when a field is deterministic, however, makes a pain vice-versa, during allocation allocating.
A non-deterministic field should be refixed on each statement, since it depends on the environment state.
Changing the tree will be temporary and therefore it should be reverted after the statement execution.
When high priority replication slave applier encounters lock conflict in innodb,
it will force the conflicting lock holder transaction (victim) to rollback.
This is a must in multi-master sychronous replication model to avoid cluster lock-up.
This high priority victim abort (aka "brute force" (BF) abort), is started
from innodb lock manager while holding the victim's transaction's (trx) mutex.
Depending on the execution state of the victim transaction, it may happen that the
BF abort will call for THD::awake() to wake up the victim transaction for the rollback.
Now, if BF abort requires THD::awake() to be called, then the applier thread executed
locking protocol of: victim trx mutex -> victim THD::LOCK_thd_data
If, at the same time another DBMS super user issues KILL command to abort the same victim,
it will execute locking protocol of: victim THD::LOCK_thd_data -> victim trx mutex.
These two locking protocol acquire mutexes in opposite order, hence unresolvable mutex locking
deadlock may occur.
The fix in this commit adds THD::wsrep_aborter flag to synchronize who can kill the victim
This flag is set both when BF is called for from innodb and by KILL command.
Either path of victim killing will bail out if victim's wsrep_killed is already
set to avoid mutex conflicts with the other aborter execution. THD::wsrep_aborter
records the aborter THD's ID. This is needed to preserve the right to kill
the victim from different locations for the same aborter thread.
It is also good error logging, to see who is reponsible for the abort.
A new test case was added in galera.galera_bf_kill_debug.test for scenario where
wsrep applier thread and manual KILL command try to kill same idle victim