(MDEV-8617: Post-fix for 10.1)
* Reset THD's PS members before returning when node is
not ready
* Add CF_SKIP_WSREP_CHECK flag to COM_STMT_XXX commands
* Skip TO replication of COM_STMT_PREPAREs for MyISAM
* Updated tests
Problem: Not all permanent Item_direct_view_ref was in permanent list of used items of the view.
Solution: Detect creating permenent view/derived table reference and put them in the permanent list at once.
- 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.
Alternative fix that doesn't cause view.test crash in --ps:
Remember when Item_ref was fixed right in the constructor
and did not have a full Item_ref::fix_fields() call. Later
in PS/SP, after Item_ref::cleanup, we use this knowledge
to avoid doing full fix_fields() for items that were never
supposed to be fix_field'ed.
Simplify the test case.
- 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.
Fixed several optimizer issues relatied to GROUP BY:
a) Refering to a SELECT column in HAVING sometimes calculated it twice, which caused problems with non determinstic functions
b) Removing duplicate fields and constants from GROUP BY was done too late for "using index for group by" optimization to work
c) EXPLAIN SELECT ... GROUP BY did wrongly show 'Using filesort' in some cases involving "Using index for group-by"
a) was fixed by:
- Changed last argument to Item::split_sum_func2() from bool to int to allow more flags
- Added flag argument to Item::split_sum_func() to allow on to specify if the item was in the SELECT part
- Mark all split_sum_func() calls from SELECT with SPLIT_SUM_SELECT
- Changed split_sum_func2() to do nothing if called with an argument that is not a sum function and doesn't include sum functions, if we are not an argument to SELECT.
This ensures that in a case like
select a*sum(b) as f1 from t1 where a=1 group by c having f1 <= 10;
That 'a' in the SELECT part is stored as a reference in the temporary table togeher with sum(b) while the 'a' in having isn't (not needed as 'a' is already a reference to a column in the result)
b) was fixed by:
- Added an extra remove_const() pass for GROUP BY arguments before make_join_statistics() in case of one table SELECT.
This allowes get_best_group_min_max() to optimize things better.
c) was fixed by:
- Added test for group by optimization in JOIN::exec_inner for
select->quick->get_type() == QUICK_SELECT_I::QS_TYPE_GROUP_MIN_MAX
item.cc:
- Simplifed Item::split_sum_func2()
- Split test to make them faster and easier to read
- Changed last argument to Item::split_sum_func2() from bool to int to allow more flags
- Added flag argument to Item::split_sum_func() to allow on to specify if the item was in the SELECT part
- Changed split_sum_func2() to do nothing if called with an argument that is not a sum function and doesn't include sum functions, if we are not an argument to SELECT.
opt_range.cc:
- Simplified get_best_group_min_max() by calcuating first how many group_by elements.
- Use join->group instead of join->group_list to test if group by, as join->group_list may be NULL if everything was optimized away.
sql_select.cc:
- Added an extra remove_const() pass for GROUP BY arguments before make_join_statistics() in case of one table SELECT.
- Use group instead of group_list to test if group by, as group_list may be NULL if everything was optimized away.
- Moved printing of "Error in remove_const" to remove_const() instead of having it in caller.
- Simplified some if tests by re-ordering code.
- update_depend_map_for_order() and remove_const() fixed to handle the case where make_join_statistics() has not yet been called (join->join_tab is 0 in this case)
SELECT ... WHERE XX IN (SELECT YY)
this was transformed to something like:
SELECT ... WHERE IF_EXISTS(SELECT ... HAVING XX=YY)
The bug was that for normal execution XX was fixed in the original outer SELECT context while in PS it was fixed in the sub query context and this confused the optimizer.
Fixed by ensuring that XX is always fixed in the outer context.
Do not call handler::rebind_psi() and handler::unbind_psi() when performance
schema is compiled out.
Overhead change:
handler::rebind_psi 0.04% -> out of radar
handler::unbind_psi 0.03% -> out of radar
open_table 0.21% -> 0.18%
close_thread_table 0.05% -> 0.05%
sql_alloc() has additional costs compared to direct mem_root allocation:
- function call: it is defined in a separate translation unit and can't be
inlined
- it needs to call pthread_getspecific() to get THD::mem_root
It is called dozens of times implicitely at least by:
- List<>::push_back()
- List<>::push_front()
- new (for Sql_alloc derived classes)
- sql_memdup()
Replaced lots of implicit sql_alloc() calls with direct mem_root allocation,
passing through THD pointer whenever it is needed.
Number of sql_alloc() calls reduced 345 -> 41 per OLTP RO transaction.
pthread_getspecific() overhead dropped 0.76 -> 0.59
sql_alloc() overhed dropped 0.25 -> 0.06
delete_dynamic() was called 9-11x per OLTP RO query + 3x per BEGIN/COMMIT.
3 calls were performed by LEX_MASTER_INFO. Added condition to call those only
for CHANGE MASTER.
1 call was performed by lock_table_names()/Hash_set/my_hash_free(). Hash_set was
supposed to be used for DDL and LOCK TABLES to gather database names, while it
was initialized/freed for DML too. In fact Hash_set didn't do any useful job
here. Hash_set was removed and MDL requests are now added directly to the list.
The rest 5-7 calls are done by optimizer, mostly by Explain_query and friends.
Since dynamic arrays are used in most cases, they can hardly be optimized.
my_hash_free() overhead dropped 0.02 -> out of radar.
delete_dynamic() overhead dropped 0.12 -> 0.04.
AVOID DEADLOCK AFTER RESTORE
Analysis
--------
Accessing the restored NDB table in an active multi-statement
transaction was resulting in deadlock found error.
MySQL Server needs to discover metadata of NDB table from
data nodes after table is restored from backup. Metadata
discovery happens on the first access to restored table.
Current code mandates this statement to be the first one
in the transaction. This is because discover needs exclusive
metadata lock on the table. Lock upgrade at this point can
lead to MDL deadlock and the code was written at the time
when MDL deadlock detector was not present. In case when
discovery attempted in the statement other than the first
one in transaction ER_LOCK_DEADLOCK error is reported
pessimistically.
Fix:
---
Removed the constraint as any potential deadlock will be
handled by deadlock detector. Also changed code in discover
to keep metadata locks of active transaction.
Same issue was present in table auto repair scenario. Same
fix is added in repair path also.
This was a regression from the patch for MDEV-7668.
A test was incorrect, so the slave would not properly handle re-using
temporary tables, which lead to replication failure in this case.
It is possible for Item_field to have a NULL field_name. This is true if
the Item_field is created based on a field in a temporary table that has
no name. It is thus necessary to do a null check before attempting a
strcmp.
Make sure that in parallel replication, we execute wait_for_prior_commit()
before setting table->in_use for a temporary table. Otherwise we can end up
with two parallel replication worker threads competing with each other for
use of a temporary table.
Re-factor the use of find_temporary_table() to be able to handle errors
in the caller (as wait_for_prior_commit() can return error in case of
deadlock kill).
[This commit cherry-picked to be able to merge MDEV-7936, of which it
is a pre-requisite, into both 10.0 and 10.1.]
Parallel replication depends on locking (table locks, row locks, etc.) to
prevent two conflicting transactions from running and committing in parallel.
But temporary tables are designed to be visible only to one thread, and have
no such locking.
In the concrete issue, an intermediate master could commit a CREATE TEMPORARY
TABLE in the same group commit as in INSERT into that table. Thus, a
lower-level master could attempt to run them in parallel and get an error.
More generally, we need protection from parallel replication trying to run
transactions in parallel that access a common temporary table.
This patch simply causes use of a temporary table from parallel replication
to wait for all previous transactions to commit, serialising the replication
at that point.
(A more fine-grained locking could be added later, possibly. However,
using temporary tables in statement-based replication is in any case
normally undesirable; for example a restart of the server will lose
temporary tables and can break replication).
Note that row-based replication is not affected, as it does not do any
temporary tables on the slave-side.
This patch also cleans up the locking around protecting the list of
temporary tables in Relay_log_info. This used to take the
rli->data_lock at the end of every statement, which is very bad for
concurrency. With this patch, the lock is not taken unless temporary
tables (with statement-based binlogging) are in use on the slave.
Parallel replication depends on locking (table locks, row locks, etc.) to
prevent two conflicting transactions from running and committing in parallel.
But temporary tables are designed to be visible only to one thread, and have
no such locking.
In the concrete issue, an intermediate master could commit a CREATE TEMPORARY
TABLE in the same group commit as in INSERT into that table. Thus, a
lower-level master could attempt to run them in parallel and get an error.
More generally, we need protection from parallel replication trying to run
transactions in parallel that access a common temporary table.
This patch simply causes use of a temporary table from parallel replication
to wait for all previous transactions to commit, serialising the replication
at that point.
(A more fine-grained locking could be added later, possibly. However,
using temporary tables in statement-based replication is in any case
normally undesirable; for example a restart of the server will lose
temporary tables and can break replication).
Note that row-based replication is not affected, as it does not do any
temporary tables on the slave-side.
This patch also cleans up the locking around protecting the list of
temporary tables in Relay_log_info. This used to take the
rli->data_lock at the end of every statement, which is very bad for
concurrency. With this patch, the lock is not taken unless temporary
tables (with statement-based binlogging) are in use on the slave.
Do not use merge_for_insert for commands which use SELECT because optimizer can't work with such tables.
Fixes which makes multi-delete working with normally merged views.
* reset current_thd in THD::~THD, otherwise my_malloc_size_cb_func()
might access THD after it was destroyed.
* remove now redundant set_current_thd(0) calls that follow delete thd.
