4351 kevin.lewis@stripped 2012-09-27
14606694 - ASSERT,RECORD LOCK WAIT IN DICT OPERATION,
DROP TEMP TABLE WHILE IN CREATE INDEX
The assert that is hit in this bug says that if the current transaction
holds dict_operation_lock, it should not have to wait on any blocks in
a SYS_* dictionary table. But this assert was hit;
default:
/* There should never be a lock wait when the
dictionary latch is reserved in X mode. Dictionary
transactions should only acquire locks on dictionary
tables, not other tables. All access to dictionary
tables should be covered by dictionary
transactions. */
ut_error;
There are two functions that need to assert that dict_operations_lock is
x-owned; dict_insert_tablespace_and_filepath() and dict_update_filepath().
These two functions are called from fil_open_single_table_tablespace()
which is called from 3 places;
1) dict_check_tablespaces_and_store_max_id()
Run during single threaded startup. But this code does not get
dict_operations_lock. That is added in this patch, but since
this is single threaded, this cannot be the source of the assert
in this bug.
2) dict_load_table()
Only had dict_sys->mutex, not dict_operations_lock. But this call
to fil_open_single_table_tablespace() sets 'fix_dict = false'
so it does not call the dictionary functions.
3) row_import_for_mysql().
This calls row_mysql_lock_data_dictionary(trx); so it has both
locks already.
The actual way in which this assert was hit was probably that a
transaction created in calls 1 or 3 above needed to be rolled back
for some reason. If that happens, the UNDO code will not
get dict_operations_lock unless those transactions are identified
as dictionary operations. This patch does that by starting the
dictionary transaction in these two functions as TRX_DICT_OP_INDEX;
This patch also identifies the sub-transactions in these two functions
as owning an X-lock on dict_operation_lock.
Approved by marko in RB:1311
modified:
storage/innobase/dict/dict0dict.cc
storage/innobase/dict/dict0load.cc
storage/innobase/fil/fil0fil.cc
storage/innobase/row/row0import.cc
4350 Dmitry Lenev 2012-09-27
Fix for bug #14569140 "MYSQL 5.6.6-M9 HAS MORE MUTEX CONTENTION
THAN MYSQL 5.1".
MySQL 5.6 has shown somewhat worse performance than MySQL 5.1
for SysBench read-only/InnoDB workload against several tables
for low number of connections.
This issue was caused by the contention in MDL subsystem
introduced in 5.5. The particular source of contention was
MDL_map::m_mutex which protects hash containing all MDL_lock
objects in the system.
This patch addresses the problem by partitioning this hash
and corresponding mutex by key for MDL_lock objects, thus
allowing connections accessing different tables/objects to
work with different partitions and therefore avoid contention.
New start-up parameter --metadata-locks-hash-instances is
introduced which allows to specify number of such partitions
(with 8 partitions as default value).
added:
mysql-test/suite/sys_vars/r/metadata_locks_hash_instances_basic.result
mysql-test/suite/sys_vars/t/metadata_locks_hash_instances_basic.test
modified:
mysql-test/r/mysqld--help-notwin.result
mysql-test/r/mysqld--help-win.result
sql/mdl.cc
sql/mdl.h
sql/sys_vars.cc
unittest/gunit/mdl-t.cc
unittest/gunit/mdl_mytap-t.cc
=== modified file 'storage/innobase/dict/dict0dict.cc'
--- a/storage/innobase/dict/dict0dict.cc revid:dmitry.lenev@stripped
+++ b/storage/innobase/dict/dict0dict.cc revid:kevin.lewis@stripped
@@ -394,7 +394,8 @@ dict_table_stats_unlock(
}
/**********************************************************************//**
-Try to drop any indexes after an aborted index creation. */
+Try to drop any indexes after an aborted index creation.
+This can also be after a server kill during DROP INDEX. */
static
void
dict_table_try_drop_aborted(
@@ -407,6 +408,7 @@ dict_table_try_drop_aborted(
trx_t* trx;
trx = trx_allocate_for_background();
+ trx->op_info = "try to drop any indexes after an aborted index creation";
row_mysql_lock_data_dictionary(trx);
trx_set_dict_operation(trx, TRX_DICT_OP_INDEX);
@@ -1779,6 +1781,7 @@ dict_table_remove_from_cache_low(
ut_d(table->n_ref_count--);
ut_ad(table->n_ref_count == 0);
trx_commit_for_mysql(trx);
+ trx->dict_operation_lock_mode = 0;
trx_free_for_background(trx);
}
=== modified file 'storage/innobase/dict/dict0load.cc'
--- a/storage/innobase/dict/dict0load.cc revid:dmitry.lenev@stripped
+++ b/storage/innobase/dict/dict0load.cc revid:kevin.lewis@stripped
@@ -861,10 +861,15 @@ dict_update_filepath(
dberr_t err = DB_SUCCESS;
trx_t* trx;
+#ifdef UNIV_SYNC_DEBUG
+ ut_ad(rw_lock_own(&dict_operation_lock, RW_LOCK_EX));
+#endif /* UNIV_SYNC_DEBUG */
ut_ad(mutex_own(&(dict_sys->mutex)));
trx = trx_allocate_for_background();
- trx_start_if_not_started(trx);
+ trx->op_info = "update filepath";
+ trx->dict_operation_lock_mode = RW_X_LATCH;
+ trx_start_for_ddl(trx, TRX_DICT_OP_INDEX);
pars_info_t* info = pars_info_create();
@@ -880,6 +885,7 @@ dict_update_filepath(
"END;\n", FALSE, trx);
trx_commit_for_mysql(trx);
+ trx->dict_operation_lock_mode = 0;
trx_free_for_background(trx);
if (err == DB_SUCCESS) {
@@ -914,11 +920,16 @@ dict_insert_tablespace_and_filepath(
dberr_t err = DB_SUCCESS;
trx_t* trx;
+#ifdef UNIV_SYNC_DEBUG
+ ut_ad(rw_lock_own(&dict_operation_lock, RW_LOCK_EX));
+#endif /* UNIV_SYNC_DEBUG */
ut_ad(mutex_own(&(dict_sys->mutex)));
ut_ad(filepath);
trx = trx_allocate_for_background();
- trx_start_if_not_started(trx);
+ trx->op_info = "insert tablespace and filepath";
+ trx->dict_operation_lock_mode = RW_X_LATCH;
+ trx_start_for_ddl(trx, TRX_DICT_OP_INDEX);
/* A record for this space ID was not found in
SYS_DATAFILES. Assume the record is also missing in
@@ -927,6 +938,7 @@ dict_insert_tablespace_and_filepath(
space, name, fsp_flags, filepath, trx, false);
trx_commit_for_mysql(trx);
+ trx->dict_operation_lock_mode = 0;
trx_free_for_background(trx);
return(err);
@@ -957,6 +969,7 @@ dict_check_tablespaces_and_store_max_id(
ulint max_space_id;
mtr_t mtr;
+ rw_lock_x_lock(&dict_operation_lock);
mutex_enter(&(dict_sys->mutex));
mtr_start(&mtr);
@@ -991,6 +1004,7 @@ loop:
fil_set_max_space_id_if_bigger(max_space_id);
mutex_exit(&(dict_sys->mutex));
+ rw_lock_x_unlock(&dict_operation_lock);
return;
}
@@ -1093,7 +1107,12 @@ loop:
space_id, name);
}
- /* filepath can be NULL in this call. */
+ /* We set the 2nd param (fix_dict = true)
+ here because we already have an x-lock on
+ dict_operation_lock and dict_sys->mutex. Besides,
+ this is at startup and we are now single threaded.
+ If the filepath is not known, it will need to
+ be discovered. */
dberr_t err = fil_open_single_table_tablespace(
false, srv_read_only_mode ? false : true,
space_id, dict_tf_to_fsp_flags(flags),
@@ -2312,7 +2331,9 @@ err_exit:
}
}
- /* Try to open the tablespace */
+ /* Try to open the tablespace. We set the
+ 2nd param (fix_dict = false) here because we
+ do not have an x-lock on dict_operation_lock */
err = fil_open_single_table_tablespace(
true, false, table->space,
dict_tf_to_fsp_flags(table->flags),
=== modified file 'storage/innobase/fil/fil0fil.cc'
--- a/storage/innobase/fil/fil0fil.cc revid:dmitry.lenev@stripped
+++ b/storage/innobase/fil/fil0fil.cc revid:kevin.lewis@stripped
@@ -3507,6 +3507,11 @@ fil_open_single_table_tablespace(
ulint tablespaces_found = 0;
ulint valid_tablespaces_found = 0;
+#ifdef UNIV_SYNC_DEBUG
+ ut_ad(!fix_dict || rw_lock_own(&dict_operation_lock, RW_LOCK_EX));
+#endif /* UNIV_SYNC_DEBUG */
+ ut_ad(!fix_dict || mutex_own(&(dict_sys->mutex)));
+
if (!fsp_flags_is_valid(flags)) {
return(DB_CORRUPTION);
}
=== modified file 'storage/innobase/row/row0import.cc'
--- a/storage/innobase/row/row0import.cc revid:dmitry.lenev@stripped
+++ b/storage/innobase/row/row0import.cc revid:kevin.lewis@stripped
@@ -3051,7 +3051,9 @@ row_import_for_mysql(
}
ut_a(filepath);
- /* Open the tablespace so that we can access via the buffer pool. */
+ /* Open the tablespace so that we can access via the buffer pool.
+ We set the 2nd param (fix_dict = true) here because we already
+ have an x-lock on dict_operation_lock and dict_sys->mutex. */
err = fil_open_single_table_tablespace(
true, true, table->space,
No bundle (reason: useless for push emails).
| Thread |
|---|
| • bzr push into mysql-5.6 branch (kevin.lewis:4350 to 4351) | kevin.lewis | 28 Sep |
