As galera node (slave) received query log events from an async
replication master, it partially wrote the updates made to replication
state table (mysql.gtid_slave_pos) to galera transaction writeset post
TOI. As a result, the transaction handle, thus created within galera,
was never freed/purged as the corresponding trx did not commit.
Thus, it kept piling up for every query log event and was only reclaimed
upon server shutdown when the transaction map object got destructed.
Fixed by making sure that updates in replication slave state table
are not written to galera transaction writeset and thus, not replicated
to other nodes.
The following left in semi-improved state to keep patch size reasonable:
- Field operator new: left thd_alloc(current_thd)
- Sql_alloc operator new: left thd_alloc(thd_get_current_thd())
- Item_args constructors: left thd_alloc(thd)
- Item_func_interval::fix_length_and_dec(): no THD arg, have to call current_thd
- Item_func_dyncol_exists::val_int(): same
- Item_dyncol_get::val_str(): same
- Item_dyncol_get::val_int(): same
- Item_dyncol_get::val_real(): same
- Item_dyncol_get::val_decimal(): same
- Item_singlerow_subselect::fix_length_and_dec(): same
Problem:
At the end of first execution select_lex->prep_where is pointing to
a runtime created object (temporary table field). As a result
server exits trying to access a invalid pointer during second
execution.
Analysis:
While optimizing the join conditions for the query, after the
permanent transformation, optimizer makes a copy of the new
where conditions in select_lex->prep_where. "prep_where" is what
is used as the "where condition" for the query at the start of execution.
W.r.t the query in question, "where" condition is actually pointing
to a field in the temporary table. As a result, for the second
execution the pointer is no more valid resulting in server exit.
Fix:
At the end of the first execution, select_lex->where will have the
original item of the where condition.
Make prep_where the new place where the original item of select->where
has to be rolled back.
Fixed in 5.7 with the wl#7082 - Move permanent transformations from
JOIN::optimize to JOIN::prepare
Patch for 5.5 includes the following backports from 5.6:
Bugfix for Bug12603141 - This makes the first execute statement in the testcase
pass in 5.5
However it was noted later in in Bug16163596 that the above bugfix needed to
be modified. Although Bug16163596 is reproducible only with changes done for
Bug12582849, we have decided include the fix.
Considering that Bug12582849 is related to Bug12603141, the fix is
also included here. However this results in Bug16317817, Bug16317685,
Bug16739050. So fix for the above three bugs is also part of this patch.
- Part 3: Adding mem_root to push_back() and push_front()
Other things:
- Added THD as an argument to some partition functions.
- Added memory overflow checking for XML tag's in read_xml()
- Added mem_root to all calls to new Item
- Added private method operator new(size_t size) to Item to ensure that
we always use a mem_root when creating an item.
This saves use once call to current_thd per Item creation
Added mandatory thd parameter to Item (and all derivative classes) constructor.
Added thd parameter to all routines that may create items.
Also removed "current_thd" from Item::Item. This reduced number of
pthread_getspecific() calls from 290 to 177 per OLTP RO transaction.
In galera cluster, when myisam replication is enabled
(wsrep_replicate_myisam=ON), DML statements are replicated
in open_tables(). However, in case of prepared statements,
for an INSERT, open_tables() gets invoked twice. Once for
COM_STMT_PREPARE (to validate and prepare INSERT) and later
for COM_STMT_EXECUTE. As a result, the command gets replicated
twice. Same happens for REPLACE, UPDATE and DELETE commands.
Fixed by adding a check to not replicate during 'prepare'
phase. Also changed the order of conditions to make it more
efficient. Lastly, in order to support wsrep_dirty_reads, made
changes to allow COM_STMT_XXX commands to continue past initial
check even when wsrep is not ready.
While sql_bin_log=1(0) is meant to control binary logging for the
current session so that the updates to do(not) get logged into the
binary log to be replicated to the async MariaDB slave. The same
should not affect galera replication.
That is, the updates should always get replicated to other galera
nodes regardless of sql_bin_log's value.
Fixed by making sure that the updates are written to binlog cache
irrespective of sql_bin_log.
Added test cases.
MariaDB Galera server's (MGC) handlerton & THD structures have
some extra wsrep members as compared to MariaDB server. As a
result, when plugins from packages, built along with MariaDB
server like Connect & OQGraph are loaded and accessed by MGC,
the objects may get initialized improperly.
Fixed by moving the wsrep-additions to the end of the structure.
The old code used pthread_setspecific() to store temporary data used by the thread.
This is not safe when used with thread pool, as the thread may change for the transaction.
The fix is to save the data in THD, which is guaranteed to be properly freed.
I also fixed the code so that we don't do a malloc() for every transaction.
- Cache variables.lc_messages->errmsgs->errmsgs in variables.errmsgs
This gives us 15 byte less code space and 2 memory references for any access to language dependent message,
of which there are 500 in the server..
- Changed ER(ER_...) to ER_THD(thd, ER_...) when thd was known or if there was many calls to current_thd in the same function.
- Changed ER(ER_..) to ER_THD_OR_DEFAULT(current_thd, ER...) in some places where current_thd is not necessary defined.
- Removing calls to current_thd when we have access to thd
Part of this is optimization (not calling current_thd when not needed),
but part is bug fixing for error condition when current_thd is not defined
(For example on startup and end of mysqld)
Notable renames done as otherwise a lot of functions would have to be changed:
- In JOIN structure renamed:
examined_rows -> join_examined_rows
record_count -> join_record_count
- In Field, renamed new_field() to make_new_field()
Other things:
- Added DBUG_ASSERT(thd == tmp_thd) in Item_singlerow_subselect() just to be safe.
- Removed old 'tab' prefix in JOIN_TAB::save_explain_data() and use members directly
- Added 'thd' as argument to a few functions to avoid calling current_thd.
When writing rows with a minimal row image, it is possible to receive
empty events. In that case m_curr_row and m_rows_end are the same,
however the event implies an insert into the table with the default
values associated for that table.
Pretend that CREATE TABLE and CREATE TEMPORARY TABLE are
two different commands internally. The user doesn't need
to know that they both are SQLCOM_CREATE_TABLE.
Same for DROP [TEMPORARY] TABLE
* Wait for aborted thd (victim) to release MDL locks
* Skip aborting an already aborted thd
* Defer setting OK status in case of CTAS
* Minor cosmetic changes
* Added a test case
Added THD argument to select_result and all derivative classes.
This reduces number of pthread_getspecific calls from 796 to 776 per OLTP RO
transaction.
THD::enter_stage() optimizations:
- stage backup code moved to THD::backup_stage(), saves one condition
- moved check for "new_stage" out to callers that actually need it
- remnants of enter_stage() moved to sql_class.h so it can be inlined
THD::enter_stage() overhead dropped 0.48% -> 0.07%.
PROFILING::status_change() optimizations:
- "status_arg" is now checked by QUERY_PROFILE::new_status()
- no need to check "enabled": !enabled && current is impossible
- remnants of status_change() moved to sql_profile.h so it can be inlined
PROFILING::status_change() overhead dropped 0.1% -> out of radar.
