maria-developers team mailing list archive
-
maria-developers team
-
Mailing list archive
-
Message #03693
Re: MWL#116: Efficient group commit, ready for code review
Kristian Nielsen <knielsen@xxxxxxxxxxxxxxx> writes:
> Please find attached the patch for code review of MWL#116: Efficient group
> commit.
Of course I forgot to attach the patch, trying again.
- Kristian.
> I went through all your comments in the architecture review and updated the
> worklog description and the code accordingly. I hope I got everything,
> otherwise I hope you have the patience to point out any omissions once again!
>
> The patch is pushed here if you want to see the complete code:
>
> lp:~maria-captains/maria/mariadb-5.1-mwl116
>
> Once the code passes code review, I think the plan is to include this feature
> in MariaDB 5.3.
>
> Please let me know if you need anything else from me for the review.
=== added file 'mysql-test/r/group_commit.result'
--- mysql-test/r/group_commit.result 1970-01-01 00:00:00 +0000
+++ mysql-test/r/group_commit.result 2010-10-28 19:07:34 +0000
@@ -0,0 +1,63 @@
+CREATE TABLE t1 (a VARCHAR(10) PRIMARY KEY) ENGINE=innodb;
+SELECT variable_value INTO @commits FROM information_schema.global_status
+WHERE variable_name = 'binlog_commits';
+SELECT variable_value INTO @group_commits FROM information_schema.global_status
+WHERE variable_name = 'binlog_group_commits';
+SET DEBUG_SYNC= "commit_before_get_LOCK_commit_ordered SIGNAL group1_running WAIT_FOR group2_queued";
+INSERT INTO t1 VALUES ("con1");
+set DEBUG_SYNC= "now WAIT_FOR group1_running";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL group2_con2";
+SET DEBUG_SYNC= "commit_after_release_LOCK_log WAIT_FOR group3_committed";
+SET DEBUG_SYNC= "commit_after_group_run_commit_ordered SIGNAL group2_visible WAIT_FOR group2_checked";
+INSERT INTO t1 VALUES ("con2");
+SET DEBUG_SYNC= "now WAIT_FOR group2_con2";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL group2_con3";
+INSERT INTO t1 VALUES ("con3");
+SET DEBUG_SYNC= "now WAIT_FOR group2_con3";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL group2_con4";
+INSERT INTO t1 VALUES ("con4");
+SET DEBUG_SYNC= "now WAIT_FOR group2_con4";
+SET SESSION TRANSACTION ISOLATION LEVEL READ COMMITTED;
+SELECT * FROM t1 ORDER BY a;
+a
+SET DEBUG_SYNC= "now SIGNAL group2_queued";
+SELECT * FROM t1 ORDER BY a;
+a
+con1
+SET DEBUG_SYNC= "commit_before_get_LOCK_commit_ordered SIGNAL group3_con5";
+SET DEBUG_SYNC= "commit_after_get_LOCK_log SIGNAL con5_leader WAIT_FOR con6_queued";
+INSERT INTO t1 VALUES ("con5");
+SET DEBUG_SYNC= "now WAIT_FOR con5_leader";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL con6_queued";
+INSERT INTO t1 VALUES ("con6");
+SET DEBUG_SYNC= "now WAIT_FOR group3_con5";
+SELECT * FROM t1 ORDER BY a;
+a
+con1
+SET DEBUG_SYNC= "now SIGNAL group3_committed";
+SET DEBUG_SYNC= "now WAIT_FOR group2_visible";
+SELECT * FROM t1 ORDER BY a;
+a
+con1
+con2
+con3
+con4
+SET DEBUG_SYNC= "now SIGNAL group2_checked";
+SELECT * FROM t1 ORDER BY a;
+a
+con1
+con2
+con3
+con4
+con5
+con6
+SELECT variable_value - @commits FROM information_schema.global_status
+WHERE variable_name = 'binlog_commits';
+variable_value - @commits
+6
+SELECT variable_value - @group_commits FROM information_schema.global_status
+WHERE variable_name = 'binlog_group_commits';
+variable_value - @group_commits
+3
+SET DEBUG_SYNC= 'RESET';
+DROP TABLE t1;
=== added file 'mysql-test/r/group_commit_binlog_pos.result'
--- mysql-test/r/group_commit_binlog_pos.result 1970-01-01 00:00:00 +0000
+++ mysql-test/r/group_commit_binlog_pos.result 2010-10-28 19:01:40 +0000
@@ -0,0 +1,35 @@
+CREATE TABLE t1 (a INT PRIMARY KEY) ENGINE=innodb;
+INSERT INTO t1 VALUES (0);
+SET DEBUG_SYNC= "commit_after_get_LOCK_log SIGNAL con1_waiting WAIT_FOR con3_queued";
+SET DEBUG_SYNC= "commit_loop_entry_commit_ordered SIGNAL con1_loop WAIT_FOR con1_loop_cont EXECUTE 3";
+INSERT INTO t1 VALUES (1);
+SET DEBUG_SYNC= "now WAIT_FOR con1_waiting";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL con2_queued";
+INSERT INTO t1 VALUES (2);
+SET DEBUG_SYNC= "now WAIT_FOR con2_queued";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL con3_queued";
+INSERT INTO t1 VALUES (3);
+SET DEBUG_SYNC= "now WAIT_FOR con1_loop";
+SET DEBUG_SYNC= "now SIGNAL con1_loop_cont";
+SET DEBUG_SYNC= "now WAIT_FOR con1_loop";
+SET DEBUG_SYNC= "now SIGNAL con1_loop_cont";
+SET DEBUG_SYNC= "now WAIT_FOR con1_loop";
+SELECT * FROM t1 ORDER BY a;
+a
+0
+1
+2
+SET SESSION debug="+d,crash_dispatch_command_before";
+SELECT 1;
+Got one of the listed errors
+Got one of the listed errors
+Got one of the listed errors
+SELECT * FROM t1 ORDER BY a;
+a
+0
+1
+2
+3
+InnoDB: Last MySQL binlog file position 0 767, file name ./master-bin.000001
+SET DEBUG_SYNC= 'RESET';
+DROP TABLE t1;
=== added file 'mysql-test/suite/binlog/r/binlog_ioerr.result'
--- mysql-test/suite/binlog/r/binlog_ioerr.result 1970-01-01 00:00:00 +0000
+++ mysql-test/suite/binlog/r/binlog_ioerr.result 2010-10-01 08:16:44 +0000
@@ -0,0 +1,28 @@
+CALL mtr.add_suppression("Error writing file 'master-bin'");
+RESET MASTER;
+CREATE TABLE t1 (a INT PRIMARY KEY) ENGINE=innodb;
+INSERT INTO t1 VALUES(0);
+SET SESSION debug='+d,fail_binlog_write_1';
+INSERT INTO t1 VALUES(1);
+ERROR HY000: Error writing file 'master-bin' (errno: 22)
+INSERT INTO t1 VALUES(2);
+ERROR HY000: Error writing file 'master-bin' (errno: 22)
+SET SESSION debug='';
+INSERT INTO t1 VALUES(3);
+SELECT * FROM t1;
+a
+0
+3
+SHOW BINLOG EVENTS;
+Log_name Pos Event_type Server_id End_log_pos Info
+BINLOG POS Format_desc 1 ENDPOS Server ver: #, Binlog ver: #
+BINLOG POS Query 1 ENDPOS use `test`; CREATE TABLE t1 (a INT PRIMARY KEY) ENGINE=innodb
+BINLOG POS Query 1 ENDPOS BEGIN
+BINLOG POS Query 1 ENDPOS use `test`; INSERT INTO t1 VALUES(0)
+BINLOG POS Xid 1 ENDPOS COMMIT /* XID */
+BINLOG POS Query 1 ENDPOS BEGIN
+BINLOG POS Query 1 ENDPOS BEGIN
+BINLOG POS Query 1 ENDPOS BEGIN
+BINLOG POS Query 1 ENDPOS use `test`; INSERT INTO t1 VALUES(3)
+BINLOG POS Xid 1 ENDPOS COMMIT /* XID */
+DROP TABLE t1;
=== added file 'mysql-test/suite/binlog/t/binlog_ioerr.test'
--- mysql-test/suite/binlog/t/binlog_ioerr.test 1970-01-01 00:00:00 +0000
+++ mysql-test/suite/binlog/t/binlog_ioerr.test 2010-10-01 08:16:44 +0000
@@ -0,0 +1,29 @@
+source include/have_debug.inc;
+source include/have_innodb.inc;
+source include/have_log_bin.inc;
+source include/have_binlog_format_mixed_or_statement.inc;
+
+CALL mtr.add_suppression("Error writing file 'master-bin'");
+
+RESET MASTER;
+
+CREATE TABLE t1 (a INT PRIMARY KEY) ENGINE=innodb;
+INSERT INTO t1 VALUES(0);
+SET SESSION debug='+d,fail_binlog_write_1';
+--error ER_ERROR_ON_WRITE
+INSERT INTO t1 VALUES(1);
+--error ER_ERROR_ON_WRITE
+INSERT INTO t1 VALUES(2);
+SET SESSION debug='';
+INSERT INTO t1 VALUES(3);
+SELECT * FROM t1;
+
+# Actually the output from this currently shows a bug.
+# The injected IO error leaves partially written transactions in the binlog in
+# the form of stray "BEGIN" events.
+# These should disappear from the output if binlog error handling is improved.
+--replace_regex /\/\* xid=.* \*\//\/* XID *\// /Server ver: .*, Binlog ver: .*/Server ver: #, Binlog ver: #/ /table_id: [0-9]+/table_id: #/
+--replace_column 1 BINLOG 2 POS 5 ENDPOS
+SHOW BINLOG EVENTS;
+
+DROP TABLE t1;
=== added file 'mysql-test/t/group_commit.test'
--- mysql-test/t/group_commit.test 1970-01-01 00:00:00 +0000
+++ mysql-test/t/group_commit.test 2010-10-28 19:07:18 +0000
@@ -0,0 +1,115 @@
+--source include/have_debug_sync.inc
+--source include/have_innodb.inc
+--source include/have_log_bin.inc
+
+# Test some group commit code paths by using debug_sync to do controlled
+# commits of 6 transactions: first 1 alone, then 3 as a group, then 2 as a
+# group.
+#
+# Group 3 is allowed to race as far as possible ahead before group 2 finishes
+# to check some edge case for concurrency control.
+
+CREATE TABLE t1 (a VARCHAR(10) PRIMARY KEY) ENGINE=innodb;
+
+SELECT variable_value INTO @commits FROM information_schema.global_status
+ WHERE variable_name = 'binlog_commits';
+SELECT variable_value INTO @group_commits FROM information_schema.global_status
+ WHERE variable_name = 'binlog_group_commits';
+
+connect(con1,localhost,root,,);
+connect(con2,localhost,root,,);
+connect(con3,localhost,root,,);
+connect(con4,localhost,root,,);
+connect(con5,localhost,root,,);
+connect(con6,localhost,root,,);
+
+# Start group1 (with one thread) doing commit, waiting for
+# group2 to queue up before finishing.
+
+connection con1;
+SET DEBUG_SYNC= "commit_before_get_LOCK_commit_ordered SIGNAL group1_running WAIT_FOR group2_queued";
+send INSERT INTO t1 VALUES ("con1");
+
+# Make group2 (with three threads) queue up.
+# Make sure con2 is the group commit leader for group2.
+# Make group2 wait with running commit_ordered() until group3 has committed.
+
+connection con2;
+set DEBUG_SYNC= "now WAIT_FOR group1_running";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL group2_con2";
+SET DEBUG_SYNC= "commit_after_release_LOCK_log WAIT_FOR group3_committed";
+SET DEBUG_SYNC= "commit_after_group_run_commit_ordered SIGNAL group2_visible WAIT_FOR group2_checked";
+send INSERT INTO t1 VALUES ("con2");
+connection con3;
+SET DEBUG_SYNC= "now WAIT_FOR group2_con2";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL group2_con3";
+send INSERT INTO t1 VALUES ("con3");
+connection con4;
+SET DEBUG_SYNC= "now WAIT_FOR group2_con3";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL group2_con4";
+send INSERT INTO t1 VALUES ("con4");
+
+# When group2 is queued, let group1 continue and queue group3.
+
+connection default;
+SET DEBUG_SYNC= "now WAIT_FOR group2_con4";
+
+# At this point, trasaction 1 is still not visible as commit_ordered() has not
+# been called yet.
+SET SESSION TRANSACTION ISOLATION LEVEL READ COMMITTED;
+SELECT * FROM t1 ORDER BY a;
+
+SET DEBUG_SYNC= "now SIGNAL group2_queued";
+connection con1;
+reap;
+
+# Now transaction 1 is visible.
+connection default;
+SELECT * FROM t1 ORDER BY a;
+
+connection con5;
+SET DEBUG_SYNC= "commit_before_get_LOCK_commit_ordered SIGNAL group3_con5";
+SET DEBUG_SYNC= "commit_after_get_LOCK_log SIGNAL con5_leader WAIT_FOR con6_queued";
+send INSERT INTO t1 VALUES ("con5");
+
+connection con6;
+SET DEBUG_SYNC= "now WAIT_FOR con5_leader";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL con6_queued";
+send INSERT INTO t1 VALUES ("con6");
+
+connection default;
+SET DEBUG_SYNC= "now WAIT_FOR group3_con5";
+# Still only transaction 1 visible, as group2 have not yet run commit_ordered().
+SELECT * FROM t1 ORDER BY a;
+SET DEBUG_SYNC= "now SIGNAL group3_committed";
+SET DEBUG_SYNC= "now WAIT_FOR group2_visible";
+# Now transactions 1-4 visible.
+SELECT * FROM t1 ORDER BY a;
+SET DEBUG_SYNC= "now SIGNAL group2_checked";
+
+connection con2;
+reap;
+
+connection con3;
+reap;
+
+connection con4;
+reap;
+
+connection con5;
+reap;
+
+connection con6;
+reap;
+
+connection default;
+# Check all transactions finally visible.
+SELECT * FROM t1 ORDER BY a;
+
+SELECT variable_value - @commits FROM information_schema.global_status
+ WHERE variable_name = 'binlog_commits';
+SELECT variable_value - @group_commits FROM information_schema.global_status
+ WHERE variable_name = 'binlog_group_commits';
+
+SET DEBUG_SYNC= 'RESET';
+DROP TABLE t1;
=== added file 'mysql-test/t/group_commit_binlog_pos-master.opt'
--- mysql-test/t/group_commit_binlog_pos-master.opt 1970-01-01 00:00:00 +0000
+++ mysql-test/t/group_commit_binlog_pos-master.opt 2010-10-17 21:48:32 +0000
@@ -0,0 +1 @@
+--skip-stack-trace --skip-core-file
=== added file 'mysql-test/t/group_commit_binlog_pos.test'
--- mysql-test/t/group_commit_binlog_pos.test 1970-01-01 00:00:00 +0000
+++ mysql-test/t/group_commit_binlog_pos.test 2010-10-28 19:01:32 +0000
@@ -0,0 +1,85 @@
+--source include/have_debug_sync.inc
+--source include/have_innodb.inc
+--source include/have_log_bin.inc
+--source include/have_binlog_format_mixed_or_statement.inc
+
+# Need DBUG to crash the server intentionally
+--source include/have_debug.inc
+# Don't test this under valgrind, memory leaks will occur as we crash
+--source include/not_valgrind.inc
+
+# XtraDB stores the binlog position corresponding to the last commit, and
+# prints it during crash recovery.
+# Test that we get the correct position when we group commit several
+# transactions together.
+
+CREATE TABLE t1 (a INT PRIMARY KEY) ENGINE=innodb;
+INSERT INTO t1 VALUES (0);
+
+connect(con1,localhost,root,,);
+connect(con2,localhost,root,,);
+connect(con3,localhost,root,,);
+
+# Queue up three commits for group commit.
+
+connection con1;
+SET DEBUG_SYNC= "commit_after_get_LOCK_log SIGNAL con1_waiting WAIT_FOR con3_queued";
+SET DEBUG_SYNC= "commit_loop_entry_commit_ordered SIGNAL con1_loop WAIT_FOR con1_loop_cont EXECUTE 3";
+send INSERT INTO t1 VALUES (1);
+
+connection con2;
+SET DEBUG_SYNC= "now WAIT_FOR con1_waiting";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL con2_queued";
+send INSERT INTO t1 VALUES (2);
+
+connection con3;
+SET DEBUG_SYNC= "now WAIT_FOR con2_queued";
+SET DEBUG_SYNC= "commit_after_prepare_ordered SIGNAL con3_queued";
+send INSERT INTO t1 VALUES (3);
+
+connection default;
+SET DEBUG_SYNC= "now WAIT_FOR con1_loop";
+# At this point, no transactions are committed.
+SET DEBUG_SYNC= "now SIGNAL con1_loop_cont";
+SET DEBUG_SYNC= "now WAIT_FOR con1_loop";
+# At this point, 1 transaction is committed.
+SET DEBUG_SYNC= "now SIGNAL con1_loop_cont";
+SET DEBUG_SYNC= "now WAIT_FOR con1_loop";
+
+# At this point, 2 transactions are committed.
+SELECT * FROM t1 ORDER BY a;
+
+connection con2;
+reap;
+
+# Now crash the server with 1+2 in-memory committed, 3 only prepared.
+connection default;
+system echo wait-group_commit_binlog_pos.test >> $MYSQLTEST_VARDIR/tmp/mysqld.1.expect;
+SET SESSION debug="+d,crash_dispatch_command_before";
+--error 2006,2013
+SELECT 1;
+
+connection con1;
+--error 2006,2013
+reap;
+connection con3;
+--error 2006,2013
+reap;
+
+system echo restart-group_commit_binlog_pos.test >> $MYSQLTEST_VARDIR/tmp/mysqld.1.expect;
+
+connection default;
+--enable_reconnect
+--source include/wait_until_connected_again.inc
+
+# Crash recovery should recover all three transactions.
+SELECT * FROM t1 ORDER BY a;
+
+# Check that the binlog position reported by InnoDB is the correct one
+# for the end of the second transaction (as can be checked with
+# mysqlbinlog).
+let $MYSQLD_DATADIR= `SELECT @@datadir`;
+--exec grep 'InnoDB: Last MySQL binlog file position' $MYSQLD_DATADIR/../../log/mysqld.1.err | tail -1
+
+SET DEBUG_SYNC= 'RESET';
+DROP TABLE t1;
=== modified file 'sql/handler.cc'
--- sql/handler.cc 2010-10-20 10:58:43 +0000
+++ sql/handler.cc 2010-11-01 14:15:03 +0000
@@ -77,6 +77,8 @@ TYPELIB tx_isolation_typelib= {array_ele
static TYPELIB known_extensions= {0,"known_exts", NULL, NULL};
uint known_extensions_id= 0;
+static int commit_one_phase_2(THD *thd, bool all, THD_TRANS *trans,
+ bool is_real_trans);
static plugin_ref ha_default_plugin(THD *thd)
@@ -1075,7 +1077,7 @@ ha_check_and_coalesce_trx_read_only(THD
*/
int ha_commit_trans(THD *thd, bool all)
{
- int error= 0, cookie= 0;
+ int error= 0, cookie;
/*
'all' means that this is either an explicit commit issued by
user, or an implicit commit issued by a DDL.
@@ -1090,7 +1092,8 @@ int ha_commit_trans(THD *thd, bool all)
*/
bool is_real_trans= all || thd->transaction.all.ha_list == 0;
Ha_trx_info *ha_info= trans->ha_list;
- my_xid xid= thd->transaction.xid_state.xid.get_my_xid();
+ bool need_prepare_ordered, need_commit_ordered;
+ my_xid xid;
DBUG_ENTER("ha_commit_trans");
/*
@@ -1123,85 +1126,112 @@ int ha_commit_trans(THD *thd, bool all)
DBUG_RETURN(2);
}
#ifdef USING_TRANSACTIONS
- if (ha_info)
+ if (!ha_info)
{
- uint rw_ha_count;
- bool rw_trans;
+ /* Free resources and perform other cleanup even for 'empty' transactions. */
+ if (is_real_trans)
+ thd->transaction.cleanup();
+ DBUG_RETURN(0);
+ }
- DBUG_EXECUTE_IF("crash_commit_before", abort(););
+ DBUG_EXECUTE_IF("crash_commit_before", abort(););
- /* Close all cursors that can not survive COMMIT */
- if (is_real_trans) /* not a statement commit */
- thd->stmt_map.close_transient_cursors();
+ /* Close all cursors that can not survive COMMIT */
+ if (is_real_trans) /* not a statement commit */
+ thd->stmt_map.close_transient_cursors();
- rw_ha_count= ha_check_and_coalesce_trx_read_only(thd, ha_info, all);
- /* rw_trans is TRUE when we in a transaction changing data */
- rw_trans= is_real_trans && (rw_ha_count > 0);
+ uint rw_ha_count= ha_check_and_coalesce_trx_read_only(thd, ha_info, all);
+ /* rw_trans is TRUE when we in a transaction changing data */
+ bool rw_trans= is_real_trans && (rw_ha_count > 0);
- if (rw_trans &&
- wait_if_global_read_lock(thd, 0, 0))
- {
- ha_rollback_trans(thd, all);
- DBUG_RETURN(1);
- }
+ if (rw_trans &&
+ wait_if_global_read_lock(thd, 0, 0))
+ {
+ ha_rollback_trans(thd, all);
+ DBUG_RETURN(1);
+ }
- if (rw_trans &&
- opt_readonly &&
- !(thd->security_ctx->master_access & SUPER_ACL) &&
- !thd->slave_thread)
- {
- my_error(ER_OPTION_PREVENTS_STATEMENT, MYF(0), "--read-only");
- ha_rollback_trans(thd, all);
- error= 1;
- goto end;
- }
+ if (rw_trans &&
+ opt_readonly &&
+ !(thd->security_ctx->master_access & SUPER_ACL) &&
+ !thd->slave_thread)
+ {
+ my_error(ER_OPTION_PREVENTS_STATEMENT, MYF(0), "--read-only");
+ goto err;
+ }
- if (!trans->no_2pc && (rw_ha_count > 1))
- {
- for (; ha_info && !error; ha_info= ha_info->next())
- {
- int err;
- handlerton *ht= ha_info->ht();
- /*
- Do not call two-phase commit if this particular
- transaction is read-only. This allows for simpler
- implementation in engines that are always read-only.
- */
- if (! ha_info->is_trx_read_write())
- continue;
- /*
- Sic: we know that prepare() is not NULL since otherwise
- trans->no_2pc would have been set.
- */
- if ((err= ht->prepare(ht, thd, all)))
- {
- my_error(ER_ERROR_DURING_COMMIT, MYF(0), err);
- error= 1;
- }
- status_var_increment(thd->status_var.ha_prepare_count);
- }
- DBUG_EXECUTE_IF("crash_commit_after_prepare", DBUG_ABORT(););
- if (error || (is_real_trans && xid &&
- (error= !(cookie= tc_log->log_xid(thd, xid)))))
- {
- ha_rollback_trans(thd, all);
- error= 1;
- goto end;
- }
- DBUG_EXECUTE_IF("crash_commit_after_log", DBUG_ABORT(););
- }
- error=ha_commit_one_phase(thd, all) ? (cookie ? 2 : 1) : 0;
- DBUG_EXECUTE_IF("crash_commit_before_unlog", DBUG_ABORT(););
- if (cookie)
- tc_log->unlog(cookie, xid);
+ if (trans->no_2pc || (rw_ha_count <= 1))
+ {
+ error= ha_commit_one_phase(thd, all);
DBUG_EXECUTE_IF("crash_commit_after", DBUG_ABORT(););
-end:
- if (rw_trans)
- start_waiting_global_read_lock(thd);
+ goto end;
}
- /* Free resources and perform other cleanup even for 'empty' transactions. */
- else if (is_real_trans)
- thd->transaction.cleanup();
+
+ need_prepare_ordered= FALSE;
+ need_commit_ordered= FALSE;
+ xid= thd->transaction.xid_state.xid.get_my_xid();
+
+ for (Ha_trx_info *hi= ha_info; hi; hi= hi->next())
+ {
+ int err;
+ handlerton *ht= hi->ht();
+ /*
+ Do not call two-phase commit if this particular
+ transaction is read-only. This allows for simpler
+ implementation in engines that are always read-only.
+ */
+ if (! hi->is_trx_read_write())
+ continue;
+ /*
+ Sic: we know that prepare() is not NULL since otherwise
+ trans->no_2pc would have been set.
+ */
+ if ((err= ht->prepare(ht, thd, all)))
+ my_error(ER_ERROR_DURING_COMMIT, MYF(0), err);
+ status_var_increment(thd->status_var.ha_prepare_count);
+
+ if (err)
+ goto err;
+
+ if (ht->prepare_ordered)
+ need_prepare_ordered= TRUE;
+ if (ht->commit_ordered)
+ need_commit_ordered= TRUE;
+ }
+ DBUG_EXECUTE_IF("crash_commit_after_prepare", DBUG_ABORT(););
+
+ if (!is_real_trans)
+ {
+ error= commit_one_phase_2(thd, all, trans, is_real_trans);
+ DBUG_EXECUTE_IF("crash_commit_after", DBUG_ABORT(););
+ goto end;
+ }
+
+ cookie= tc_log->log_and_order(thd, xid, all, need_prepare_ordered,
+ need_commit_ordered);
+ if (!cookie)
+ goto err;
+
+ DBUG_EXECUTE_IF("crash_commit_after_log", DBUG_ABORT(););
+
+ error= commit_one_phase_2(thd, all, trans, is_real_trans) ? 2 : 0;
+ DBUG_EXECUTE_IF("crash_commit_after", DBUG_ABORT(););
+
+ DBUG_EXECUTE_IF("crash_commit_before_unlog", DBUG_ABORT(););
+ tc_log->unlog(cookie, xid);
+
+ DBUG_EXECUTE_IF("crash_commit_after", DBUG_ABORT(););
+ goto end;
+
+ /* Come here if error and we need to rollback. */
+err:
+ if (!error)
+ error= 1;
+ ha_rollback_trans(thd, all);
+
+end:
+ if (rw_trans)
+ start_waiting_global_read_lock(thd);
#endif /* USING_TRANSACTIONS */
DBUG_RETURN(error);
}
@@ -1212,7 +1242,6 @@ end:
*/
int ha_commit_one_phase(THD *thd, bool all)
{
- int error=0;
THD_TRANS *trans=all ? &thd->transaction.all : &thd->transaction.stmt;
/*
"real" is a nick name for a transaction for which a commit will
@@ -1222,8 +1251,16 @@ int ha_commit_one_phase(THD *thd, bool a
enclosing 'all' transaction is rolled back.
*/
bool is_real_trans=all || thd->transaction.all.ha_list == 0;
- Ha_trx_info *ha_info= trans->ha_list, *ha_info_next;
DBUG_ENTER("ha_commit_one_phase");
+ DBUG_RETURN(commit_one_phase_2(thd, all, trans, is_real_trans));
+}
+
+static int
+commit_one_phase_2(THD *thd, bool all, THD_TRANS *trans, bool is_real_trans)
+{
+ int error= 0;
+ Ha_trx_info *ha_info= trans->ha_list, *ha_info_next;
+ DBUG_ENTER("commit_one_phase_2");
#ifdef USING_TRANSACTIONS
if (ha_info)
{
@@ -1839,7 +1876,13 @@ int ha_start_consistent_snapshot(THD *th
{
bool warn= true;
+ /*
+ Holding the LOCK_commit_ordered mutex ensures that for any transaction
+ we either see it committed in all engines, or in none.
+ */
+ pthread_mutex_lock(&LOCK_commit_ordered);
plugin_foreach(thd, snapshot_handlerton, MYSQL_STORAGE_ENGINE_PLUGIN, &warn);
+ pthread_mutex_unlock(&LOCK_commit_ordered);
/*
Same idea as when one wants to CREATE TABLE in one engine which does not
=== modified file 'sql/handler.h'
--- sql/handler.h 2010-10-20 10:58:43 +0000
+++ sql/handler.h 2010-11-01 14:15:03 +0000
@@ -659,9 +659,100 @@ struct handlerton
NOTE 'all' is also false in auto-commit mode where 'end of statement'
and 'real commit' mean the same event.
*/
- int (*commit)(handlerton *hton, THD *thd, bool all);
+ int (*commit)(handlerton *hton, THD *thd, bool all);
+ /*
+ The commit_ordered() method is called prior to the commit() method, after
+ the transaction manager has decided to commit (not rollback) the
+ transaction. Unlike commit(), commit_ordered() is called only when the
+ full transaction is committed, not for each commit of statement
+ transaction in a multi-statement transaction.
+
+ Not that like prepare(), commit_ordered() is only called when 2-phase
+ commit takes place. Ie. when no binary log and only a single engine
+ participates in a transaction, one commit() is called, no
+ commit_orderd(). So engines must be prepared for this.
+
+ The calls to commit_ordered() in multiple parallel transactions is
+ guaranteed to happen in the same order in every participating
+ handler. This can be used to ensure the same commit order among multiple
+ handlers (eg. in table handler and binlog). So if transaction T1 calls
+ into commit_ordered() of handler A before T2, then T1 will also call
+ commit_ordered() of handler B before T2.
+
+ Engines that implement this method should during this call make the
+ transaction visible to other transactions, thereby making the order of
+ transaction commits be defined by the order of commit_ordered() calls.
+
+ The intension is that commit_ordered() should do the minimal amount of
+ work that needs to happen in consistent commit order among handlers. To
+ preserve ordering, calls need to be serialised on a global mutex, so
+ doing any time-consuming or blocking operations in commit_ordered() will
+ limit scalability.
+
+ Handlers can rely on commit_ordered() calls to be serialised (no two
+ calls can run in parallel, so no extra locking on the handler part is
+ required to ensure this).
+
+ Note that commit_ordered() can be called from a different thread than the
+ one handling the transaction! So it can not do anything that depends on
+ thread local storage, in particular it can not call my_error() and
+ friends (instead it can store the error code and delay the call of
+ my_error() to the commit() method).
+
+ Similarly, since commit_ordered() returns void, any return error code
+ must be saved and returned from the commit() method instead.
+
+ The commit_ordered method is optional, and can be left unset if not
+ needed in a particular handler (then there will be no ordering guarantees
+ wrt. other engines and binary log).
+ */
+ void (*commit_ordered)(handlerton *hton, THD *thd, bool all);
int (*rollback)(handlerton *hton, THD *thd, bool all);
int (*prepare)(handlerton *hton, THD *thd, bool all);
+ /*
+ The prepare_ordered method is optional. If set, it will be called after
+ successful prepare() in all handlers participating in 2-phase
+ commit. Like commit_ordered(), it is called only when the full
+ transaction is committed, not for each commit of statement transaction.
+
+ The calls to prepare_ordered() among multiple parallel transactions are
+ ordered consistently with calls to commit_ordered(). This means that
+ calls to prepare_ordered() effectively define the commit order, and that
+ each handler will see the same sequence of transactions calling into
+ prepare_ordered() and commit_ordered().
+
+ Thus, prepare_ordered() can be used to define commit order for handlers
+ that need to do this in the prepare step (like binlog). It can also be
+ used to release transaction's locks early in an order consistent with the
+ order transactions will be eventually committed.
+
+ Like commit_ordered(), prepare_ordered() calls are serialised to maintain
+ ordering, so the intension is that they should execute fast, with only
+ the minimal amount of work needed to define commit order. Handlers can
+ rely on this serialisation, and do not need to do any extra locking to
+ avoid two prepare_ordered() calls running in parallel.
+
+ Like commit_ordered(), prepare_ordered() is not guaranteed to be called
+ in the context of the thread handling the rest of the transaction. So it
+ cannot invoke code that relies on thread local storage, in particular it
+ cannot call my_error().
+
+ When prepare_ordered() is called, the transaction coordinator has already
+ decided to commit (not rollback) the transaction. So prepare_ordered()
+ cannot cause a rollback by returning an error, all possible errors must
+ be handled in prepare() (the prepare_ordered() method returns void). In
+ case of some fatal error, a record of the error must be made internally
+ by the engine and returned from commit() later.
+
+ Note that for user-level XA SQL commands, no consistent ordering among
+ prepare_ordered() and commit_ordered() is guaranteed (as that would
+ require blocking all other commits for an indefinite time).
+
+ When 2-phase commit is not used (eg. only one engine (and no binlog) in
+ transaction), prepare() is not called and in such cases prepare_ordered()
+ also is not called.
+ */
+ void (*prepare_ordered)(handlerton *hton, THD *thd, bool all);
int (*recover)(handlerton *hton, XID *xid_list, uint len);
int (*commit_by_xid)(handlerton *hton, XID *xid);
int (*rollback_by_xid)(handlerton *hton, XID *xid);
=== modified file 'sql/log.cc'
--- sql/log.cc 2010-10-19 13:58:35 +0000
+++ sql/log.cc 2010-11-01 14:15:03 +0000
@@ -38,6 +38,7 @@
#endif
#include <mysql/plugin.h>
+#include "debug_sync.h"
/* max size of the log message */
#define MAX_LOG_BUFFER_SIZE 1024
@@ -154,7 +155,7 @@ class binlog_trx_data {
public:
binlog_trx_data()
: at_least_one_stmt_committed(0), incident(FALSE), m_pending(0),
- before_stmt_pos(MY_OFF_T_UNDEF)
+ before_stmt_pos(MY_OFF_T_UNDEF), commit_bin_log_file_pos(0), using_xa(0)
{
trans_log.end_of_file= max_binlog_cache_size;
}
@@ -208,11 +209,13 @@ public:
completely.
*/
void reset() {
- if (!empty())
+ if (trans_log.type != WRITE_CACHE || !empty())
truncate(0);
before_stmt_pos= MY_OFF_T_UNDEF;
incident= FALSE;
trans_log.end_of_file= max_binlog_cache_size;
+ using_xa= FALSE;
+ commit_bin_log_file_pos= 0;
DBUG_ASSERT(empty());
}
@@ -257,6 +260,17 @@ public:
Binlog position before the start of the current statement.
*/
my_off_t before_stmt_pos;
+ /*
+ Binlog position after current commit, available to storage engines during
+ commit_ordered() and commit().
+ */
+ ulonglong commit_bin_log_file_pos;
+
+ /*
+ Flag set true if this transaction is committed with log_xid() as part of
+ XA, false if not.
+ */
+ bool using_xa;
};
handlerton *binlog_hton;
@@ -1394,103 +1408,131 @@ static int binlog_close_connection(handl
}
/*
- End a transaction.
+ End a transaction, writing events to the binary log.
SYNOPSIS
- binlog_end_trans()
+ binlog_flush_trx_cache()
thd The thread whose transaction should be ended
trx_data Pointer to the transaction data to use
- end_ev The end event to use, or NULL
- all True if the entire transaction should be ended, false if
- only the statement transaction should be ended.
+ end_ev The end event to use (COMMIT, ROLLBACK, or commit XID)
DESCRIPTION
End the currently open transaction. The transaction can be either
- a real transaction (if 'all' is true) or a statement transaction
- (if 'all' is false).
+ a real transaction or a statement transaction.
- If 'end_ev' is NULL, the transaction is a rollback of only
- transactional tables, so the transaction cache will be truncated
- to either just before the last opened statement transaction (if
- 'all' is false), or reset completely (if 'all' is true).
+ This can be to commit a transaction, with a COMMIT query event or an XA
+ commit XID event. But it can also be to rollback a transaction with a
+ ROLLBACK query event, used for rolling back transactions which also
+ contain updates to non-transactional tables.
*/
static int
-binlog_end_trans(THD *thd, binlog_trx_data *trx_data,
- Log_event *end_ev, bool all)
+binlog_flush_trx_cache(THD *thd, binlog_trx_data *trx_data,
+ Log_event *end_ev, bool all)
{
- DBUG_ENTER("binlog_end_trans");
- int error=0;
+ DBUG_ENTER("binlog_flush_trx_cache");
IO_CACHE *trans_log= &trx_data->trans_log;
- DBUG_PRINT("enter", ("transaction: %s end_ev: 0x%lx",
- all ? "all" : "stmt", (long) end_ev));
DBUG_PRINT("info", ("thd->options={ %s%s}",
FLAGSTR(thd->options, OPTION_NOT_AUTOCOMMIT),
FLAGSTR(thd->options, OPTION_BEGIN)));
+ if (thd->binlog_flush_pending_rows_event(TRUE))
+ DBUG_RETURN(1);
+
/*
- NULL denotes ROLLBACK with nothing to replicate: i.e., rollback of
- only transactional tables. If the transaction contain changes to
- any non-transactiona tables, we need write the transaction and log
- a ROLLBACK last.
- */
- if (end_ev != NULL)
- {
- if (thd->binlog_flush_pending_rows_event(TRUE))
- DBUG_RETURN(1);
- /*
- Doing a commit or a rollback including non-transactional tables,
- i.e., ending a transaction where we might write the transaction
- cache to the binary log.
-
- We can always end the statement when ending a transaction since
- transactions are not allowed inside stored functions. If they
- were, we would have to ensure that we're not ending a statement
- inside a stored function.
- */
- error= mysql_bin_log.write(thd, &trx_data->trans_log, end_ev,
- trx_data->has_incident());
- trx_data->reset();
+ Doing a commit or a rollback including non-transactional tables,
+ i.e., ending a transaction where we might write the transaction
+ cache to the binary log.
+
+ We can always end the statement when ending a transaction since
+ transactions are not allowed inside stored functions. If they
+ were, we would have to ensure that we're not ending a statement
+ inside a stored function.
+ */
+ int error= mysql_bin_log.write_transaction_to_binlog(thd, trx_data,
+ end_ev, all);
- statistic_increment(binlog_cache_use, &LOCK_status);
- if (trans_log->disk_writes != 0)
- {
- statistic_increment(binlog_cache_disk_use, &LOCK_status);
- trans_log->disk_writes= 0;
- }
- }
- else
+ trx_data->reset();
+
+ statistic_increment(binlog_cache_use, &LOCK_status);
+ if (trans_log->disk_writes != 0)
{
- /*
- If rolling back an entire transaction or a single statement not
- inside a transaction, we reset the transaction cache.
+ statistic_increment(binlog_cache_disk_use, &LOCK_status);
+ trans_log->disk_writes= 0;
+ }
- If rolling back a statement in a transaction, we truncate the
- transaction cache to remove the statement.
- */
- thd->binlog_remove_pending_rows_event(TRUE);
- if (all || !(thd->options & (OPTION_BEGIN | OPTION_NOT_AUTOCOMMIT)))
- {
- if (trx_data->has_incident())
- error= mysql_bin_log.write_incident(thd, TRUE);
- trx_data->reset();
- }
- else // ...statement
- trx_data->truncate(trx_data->before_stmt_pos);
+ DBUG_ASSERT(thd->binlog_get_pending_rows_event() == NULL);
+ DBUG_RETURN(error);
+}
+
+/*
+ Discard a transaction, ie. ROLLBACK with only transactional table updates.
+
+ SYNOPSIS
+ binlog_truncate_trx_cache()
+
+ thd The thread whose transaction should be ended
+ trx_data Pointer to the transaction data to use
+ all True if the entire transaction should be ended, false if
+ only the statement transaction should be ended.
+
+ DESCRIPTION
+
+ Rollback (and end) a transaction that only modifies transactional
+ tables. The transaction can be either a real transaction (if 'all' is
+ true) or a statement transaction (if 'all' is false).
+
+ The transaction cache will be truncated to either just before the last
+ opened statement transaction (if 'all' is false), or reset completely (if
+ 'all' is true).
+ */
+static int
+binlog_truncate_trx_cache(THD *thd, binlog_trx_data *trx_data, bool all)
+{
+ DBUG_ENTER("binlog_truncate_trx_cache");
+ int error= 0;
+ DBUG_PRINT("enter", ("transaction: %s", all ? "all" : "stmt"));
+ DBUG_PRINT("info", ("thd->options={ %s%s}",
+ FLAGSTR(thd->options, OPTION_NOT_AUTOCOMMIT),
+ FLAGSTR(thd->options, OPTION_BEGIN)));
+
+ /*
+ ROLLBACK with nothing to replicate: i.e., rollback of only transactional
+ tables.
+ */
+
+ /*
+ If rolling back an entire transaction or a single statement not
+ inside a transaction, we reset the transaction cache.
+
+ If rolling back a statement in a transaction, we truncate the
+ transaction cache to remove the statement.
+ */
+ thd->binlog_remove_pending_rows_event(TRUE);
+ if (all || !(thd->options & (OPTION_BEGIN | OPTION_NOT_AUTOCOMMIT)))
+ {
+ if (trx_data->has_incident())
+ error= mysql_bin_log.write_incident(thd);
+ trx_data->reset();
}
+ else // ...statement
+ trx_data->truncate(trx_data->before_stmt_pos);
DBUG_ASSERT(thd->binlog_get_pending_rows_event() == NULL);
DBUG_RETURN(error);
}
+static LEX_STRING const write_error_msg=
+ { C_STRING_WITH_LEN("error writing to the binary log") };
+
static int binlog_prepare(handlerton *hton, THD *thd, bool all)
{
/*
do nothing.
just pretend we can do 2pc, so that MySQL won't
switch to 1pc.
- real work will be done in MYSQL_BIN_LOG::log_xid()
+ real work will be done in MYSQL_BIN_LOG::log_and_order()
*/
return 0;
}
@@ -1541,8 +1583,8 @@ static int binlog_commit(handlerton *hto
(trans_has_no_stmt_committed(thd, all) &&
!stmt_has_updated_trans_table(thd) && stmt_has_updated_non_trans_table(thd)))
{
- Query_log_event qev(thd, STRING_WITH_LEN("COMMIT"), TRUE, TRUE, 0);
- error= binlog_end_trans(thd, trx_data, &qev, all);
+ Query_log_event end_ev(thd, STRING_WITH_LEN("COMMIT"), TRUE, TRUE, 0);
+ error= binlog_flush_trx_cache(thd, trx_data, &end_ev, all);
}
trx_data->at_least_one_stmt_committed = my_b_tell(&trx_data->trans_log) > 0;
@@ -1606,7 +1648,7 @@ static int binlog_rollback(handlerton *h
(thd->options & OPTION_KEEP_LOG)) &&
mysql_bin_log.check_write_error(thd))
trx_data->set_incident();
- error= binlog_end_trans(thd, trx_data, 0, all);
+ error= binlog_truncate_trx_cache(thd, trx_data, all);
}
else
{
@@ -1625,8 +1667,8 @@ static int binlog_rollback(handlerton *h
stmt_has_updated_non_trans_table(thd) &&
thd->current_stmt_binlog_row_based))
{
- Query_log_event qev(thd, STRING_WITH_LEN("ROLLBACK"), TRUE, TRUE, 0);
- error= binlog_end_trans(thd, trx_data, &qev, all);
+ Query_log_event end_ev(thd, STRING_WITH_LEN("ROLLBACK"), TRUE, TRUE, 0);
+ error= binlog_flush_trx_cache(thd, trx_data, &end_ev, all);
}
/*
Otherwise, we simply truncate the cache as there is no change on
@@ -1634,7 +1676,7 @@ static int binlog_rollback(handlerton *h
*/
else if (ending_trans(thd, all) ||
(!(thd->options & OPTION_KEEP_LOG) && !stmt_has_updated_non_trans_table(thd)))
- error= binlog_end_trans(thd, trx_data, 0, all);
+ error= binlog_truncate_trx_cache(thd, trx_data, all);
}
if (!all)
trx_data->before_stmt_pos = MY_OFF_T_UNDEF; // part of the stmt rollback
@@ -2468,6 +2510,7 @@ const char *MYSQL_LOG::generate_name(con
MYSQL_BIN_LOG::MYSQL_BIN_LOG()
:bytes_written(0), prepared_xids(0), file_id(1), open_count(1),
need_start_event(TRUE),
+ group_commit_queue(0), num_commits(0), num_group_commits(0),
is_relay_log(0),
description_event_for_exec(0), description_event_for_queue(0)
{
@@ -3946,6 +3989,10 @@ err:
}
+#ifndef DBUG_OFF
+static ulong opt_binlog_dbug_fsync_sleep= 0;
+#endif
+
bool MYSQL_BIN_LOG::flush_and_sync()
{
int err=0, fd=log_file.file;
@@ -3956,6 +4003,11 @@ bool MYSQL_BIN_LOG::flush_and_sync()
{
sync_binlog_counter= 0;
err=my_sync(fd, MYF(MY_WME));
+#ifndef DBUG_OFF
+ ulong usec_sleep= opt_binlog_dbug_fsync_sleep;
+ if (usec_sleep > 0)
+ my_sleep(usec_sleep);
+#endif
}
return err;
}
@@ -4254,44 +4306,41 @@ MYSQL_BIN_LOG::flush_and_set_pending_row
if (Rows_log_event* pending= trx_data->pending())
{
- IO_CACHE *file= &log_file;
-
/*
Decide if we should write to the log file directly or to the
transaction log.
*/
if (pending->get_cache_stmt() || my_b_tell(&trx_data->trans_log))
- file= &trx_data->trans_log;
-
- /*
- If we are not writing to the log file directly, we could avoid
- locking the log.
- */
- pthread_mutex_lock(&LOCK_log);
-
- /*
- Write pending event to log file or transaction cache
- */
- if (pending->write(file))
{
- pthread_mutex_unlock(&LOCK_log);
- set_write_error(thd);
- DBUG_RETURN(1);
+ /* Write to transaction log/cache. */
+ if (pending->write(&trx_data->trans_log))
+ {
+ set_write_error(thd);
+ DBUG_RETURN(1);
+ }
}
-
- delete pending;
-
- if (file == &log_file)
+ else
{
+ /* Write directly to log file. */
+ pthread_mutex_lock(&LOCK_log);
+ if (pending->write(&log_file))
+ {
+ pthread_mutex_unlock(&LOCK_log);
+ set_write_error(thd);
+ DBUG_RETURN(1);
+ }
+
error= flush_and_sync();
if (!error)
{
signal_update();
rotate_and_purge(RP_LOCK_LOG_IS_ALREADY_LOCKED);
}
+
+ pthread_mutex_unlock(&LOCK_log);
}
- pthread_mutex_unlock(&LOCK_log);
+ delete pending;
}
thd->binlog_set_pending_rows_event(event);
@@ -4321,11 +4370,6 @@ bool MYSQL_BIN_LOG::write(Log_event *eve
}
/*
- Flush the pending rows event to the transaction cache or to the
- log file. Since this function potentially aquire the LOCK_log
- mutex, we do this before aquiring the LOCK_log mutex in this
- function.
-
We only end the statement if we are in a top-level statement. If
we are inside a stored function, we do not end the statement since
this will close all tables on the slave.
@@ -4335,8 +4379,6 @@ bool MYSQL_BIN_LOG::write(Log_event *eve
if (thd->binlog_flush_pending_rows_event(end_stmt))
DBUG_RETURN(error);
- pthread_mutex_lock(&LOCK_log);
-
/*
In most cases this is only called if 'is_open()' is true; in fact this is
mostly called if is_open() *was* true a few instructions before, but it
@@ -4357,7 +4399,6 @@ bool MYSQL_BIN_LOG::write(Log_event *eve
thd->lex->sql_command != SQLCOM_SAVEPOINT &&
!binlog_filter->db_ok(local_db)))
{
- VOID(pthread_mutex_unlock(&LOCK_log));
DBUG_RETURN(0);
}
#endif /* HAVE_REPLICATION */
@@ -4399,15 +4440,11 @@ bool MYSQL_BIN_LOG::write(Log_event *eve
thd->binlog_start_trans_and_stmt();
file= trans_log;
}
- /*
- TODO as Mats suggested, for all the cases above where we write to
- trans_log, it sounds unnecessary to lock LOCK_log. We should rather
- test first if we want to write to trans_log, and if not, lock
- LOCK_log.
- */
}
#endif /* USING_TRANSACTIONS */
DBUG_PRINT("info",("event type: %d",event_info->get_type_code()));
+ if (file == &log_file)
+ pthread_mutex_lock(&LOCK_log);
/*
No check for auto events flag here - this write method should
@@ -4432,7 +4469,7 @@ bool MYSQL_BIN_LOG::write(Log_event *eve
Intvar_log_event e(thd,(uchar) LAST_INSERT_ID_EVENT,
thd->first_successful_insert_id_in_prev_stmt_for_binlog);
if (e.write(file))
- goto err;
+ goto err_unlock;
}
if (thd->auto_inc_intervals_in_cur_stmt_for_binlog.nb_elements() > 0)
{
@@ -4443,13 +4480,13 @@ bool MYSQL_BIN_LOG::write(Log_event *eve
thd->auto_inc_intervals_in_cur_stmt_for_binlog.
minimum());
if (e.write(file))
- goto err;
+ goto err_unlock;
}
if (thd->rand_used)
{
Rand_log_event e(thd,thd->rand_saved_seed1,thd->rand_saved_seed2);
if (e.write(file))
- goto err;
+ goto err_unlock;
}
if (thd->user_var_events.elements)
{
@@ -4464,7 +4501,7 @@ bool MYSQL_BIN_LOG::write(Log_event *eve
user_var_event->type,
user_var_event->charset_number);
if (e.write(file))
- goto err;
+ goto err_unlock;
}
}
}
@@ -4476,23 +4513,26 @@ bool MYSQL_BIN_LOG::write(Log_event *eve
if (event_info->write(file) ||
DBUG_EVALUATE_IF("injecting_fault_writing", 1, 0))
- goto err;
+ goto err_unlock;
if (file == &log_file) // we are writing to the real log (disk)
{
if (flush_and_sync())
- goto err;
+ goto err_unlock;
signal_update();
rotate_and_purge(RP_LOCK_LOG_IS_ALREADY_LOCKED);
}
error=0;
+err_unlock:
+ if (file == &log_file)
+ pthread_mutex_unlock(&LOCK_log);
+
err:
if (error)
set_write_error(thd);
}
- pthread_mutex_unlock(&LOCK_log);
DBUG_RETURN(error);
}
@@ -4614,18 +4654,14 @@ uint MYSQL_BIN_LOG::next_file_id()
SYNOPSIS
write_cache()
cache Cache to write to the binary log
- lock_log True if the LOCK_log mutex should be aquired, false otherwise
- sync_log True if the log should be flushed and sync:ed
DESCRIPTION
Write the contents of the cache to the binary log. The cache will
be reset as a READ_CACHE to be able to read the contents from it.
*/
-int MYSQL_BIN_LOG::write_cache(IO_CACHE *cache, bool lock_log, bool sync_log)
+int MYSQL_BIN_LOG::write_cache(IO_CACHE *cache)
{
- Mutex_sentry sentry(lock_log ? &LOCK_log : NULL);
-
if (reinit_io_cache(cache, READ_CACHE, 0, 0, 0))
return ER_ERROR_ON_WRITE;
uint length= my_b_bytes_in_cache(cache), group, carry, hdr_offs;
@@ -4736,6 +4772,7 @@ int MYSQL_BIN_LOG::write_cache(IO_CACHE
}
/* Write data to the binary log file */
+ DBUG_EXECUTE_IF("fail_binlog_write_1", return ER_ERROR_ON_WRITE;);
if (my_b_write(&log_file, cache->read_pos, length))
return ER_ERROR_ON_WRITE;
cache->read_pos=cache->read_end; // Mark buffer used up
@@ -4743,9 +4780,6 @@ int MYSQL_BIN_LOG::write_cache(IO_CACHE
DBUG_ASSERT(carry == 0);
- if (sync_log)
- flush_and_sync();
-
return 0; // All OK
}
@@ -4778,26 +4812,22 @@ int query_error_code(THD *thd, bool not_
return error;
}
-bool MYSQL_BIN_LOG::write_incident(THD *thd, bool lock)
+bool MYSQL_BIN_LOG::write_incident(THD *thd)
{
uint error= 0;
DBUG_ENTER("MYSQL_BIN_LOG::write_incident");
- LEX_STRING const write_error_msg=
- { C_STRING_WITH_LEN("error writing to the binary log") };
Incident incident= INCIDENT_LOST_EVENTS;
Incident_log_event ev(thd, incident, write_error_msg);
- if (lock)
- pthread_mutex_lock(&LOCK_log);
+
+ pthread_mutex_lock(&LOCK_log);
error= ev.write(&log_file);
- if (lock)
+ if (!error && !(error= flush_and_sync()))
{
- if (!error && !(error= flush_and_sync()))
- {
- signal_update();
- rotate_and_purge(RP_LOCK_LOG_IS_ALREADY_LOCKED);
- }
- pthread_mutex_unlock(&LOCK_log);
+ signal_update();
+ rotate_and_purge(RP_LOCK_LOG_IS_ALREADY_LOCKED);
}
+ pthread_mutex_unlock(&LOCK_log);
+
DBUG_RETURN(error);
}
@@ -4825,103 +4855,311 @@ bool MYSQL_BIN_LOG::write_incident(THD *
'cache' needs to be reinitialized after this functions returns.
*/
-bool MYSQL_BIN_LOG::write(THD *thd, IO_CACHE *cache, Log_event *commit_event,
- bool incident)
+bool
+MYSQL_BIN_LOG::write_transaction_to_binlog(THD *thd, binlog_trx_data *trx_data,
+ Log_event *end_ev, bool all)
{
- DBUG_ENTER("MYSQL_BIN_LOG::write(THD *, IO_CACHE *, Log_event *)");
+ group_commit_entry entry;
+ DBUG_ENTER("MYSQL_BIN_LOG::write_transaction_to_binlog");
+
+ entry.thd= thd;
+ entry.trx_data= trx_data;
+ entry.error= 0;
+ entry.all= all;
+
+ /*
+ Create the necessary events here, where we have the correct THD (and
+ thread context).
+
+ Due to group commit the actual writing to binlog may happen in a different
+ thread.
+ */
+ Query_log_event qinfo(thd, STRING_WITH_LEN("BEGIN"), TRUE, TRUE, 0);
+ entry.begin_event= &qinfo;
+ entry.end_event= end_ev;
+ if (trx_data->has_incident())
+ {
+ Incident_log_event inc_ev(thd, INCIDENT_LOST_EVENTS, write_error_msg);
+ entry.incident_event= &inc_ev;
+ DBUG_RETURN(write_transaction_to_binlog_events(&entry));
+ }
+ else
+ {
+ entry.incident_event= NULL;
+ DBUG_RETURN(write_transaction_to_binlog_events(&entry));
+ }
+}
+
+bool
+MYSQL_BIN_LOG::write_transaction_to_binlog_events(group_commit_entry *entry)
+{
+ /*
+ To facilitate group commit for the binlog, we first queue up ourselves in
+ the group commit queue. Then the first thread to enter the queue waits for
+ the LOCK_log mutex, and commits for everyone in the queue once it gets the
+ lock. Any other threads in the queue just wait for the first one to finish
+ the commit and wake them up.
+ */
+
+ entry->thd->clear_wakeup_ready();
+ pthread_mutex_lock(&LOCK_prepare_ordered);
+ group_commit_entry *orig_queue= group_commit_queue;
+ entry->next= orig_queue;
+ group_commit_queue= entry;
+
+ if (entry->trx_data->using_xa)
+ {
+ DEBUG_SYNC(entry->thd, "commit_before_prepare_ordered");
+ run_prepare_ordered(entry->thd, entry->all);
+ DEBUG_SYNC(entry->thd, "commit_after_prepare_ordered");
+ }
+ pthread_mutex_unlock(&LOCK_prepare_ordered);
+
+ /*
+ The first in the queue handle group commit for all; the others just wait
+ to be signalled when group commit is done.
+ */
+ if (orig_queue != NULL)
+ entry->thd->wait_for_wakeup_ready();
+ else
+ trx_group_commit_leader(entry);
+
+ if (!entry->error)
+ return 0;
+
+ switch (entry->error)
+ {
+ case ER_ERROR_ON_WRITE:
+ my_error(ER_ERROR_ON_WRITE, MYF(ME_NOREFRESH), name, entry->commit_errno);
+ break;
+ case ER_ERROR_ON_READ:
+ my_error(ER_ERROR_ON_READ, MYF(ME_NOREFRESH),
+ entry->trx_data->trans_log.file_name, entry->commit_errno);
+ break;
+ default:
+ /*
+ There are not (and should not be) any errors thrown not covered above.
+ But just in case one is added later without updating the above switch
+ statement, include a catch-all.
+ */
+ my_printf_error(entry->error,
+ "Error writing transaction to binary log: %d",
+ MYF(ME_NOREFRESH), entry->error);
+ }
+
+ /*
+ Since we return error, this transaction XID will not be committed, so
+ we need to mark it as not needed for recovery (unlog() is not called
+ for a transaction if log_xid() fails).
+ */
+ if (entry->trx_data->using_xa)
+ mark_xid_done();
+
+ return 1;
+}
+
+/*
+ Do binlog group commit as the lead thread.
+
+ This must be called when this thread/transaction is queued at the start of
+ the group_commit_queue. It will wait to obtain the LOCK_log mutex, then group
+ commit all the transactions in the queue (more may have entered while waiting
+ for LOCK_log). After commit is done, all other threads in the queue will be
+ signalled.
+
+ */
+void
+MYSQL_BIN_LOG::trx_group_commit_leader(group_commit_entry *leader)
+{
+ DBUG_ENTER("MYSQL_BIN_LOG::trx_group_commit_leader");
+ uint xid_count= 0;
+ uint write_count= 0;
+
+ /*
+ Lock the LOCK_log(), and once we get it, collect any additional writes
+ that queued up while we were waiting.
+ */
VOID(pthread_mutex_lock(&LOCK_log));
+ DEBUG_SYNC(leader->thd, "commit_after_get_LOCK_log");
+
+ pthread_mutex_lock(&LOCK_prepare_ordered);
+ group_commit_entry *current= group_commit_queue;
+ group_commit_queue= NULL;
+ pthread_mutex_unlock(&LOCK_prepare_ordered);
- /* NULL would represent nothing to replicate after ROLLBACK */
- DBUG_ASSERT(commit_event != NULL);
+ /* As the queue is in reverse order of entering, reverse it. */
+ group_commit_entry *queue= NULL;
+ while (current)
+ {
+ group_commit_entry *next= current->next;
+ current->next= queue;
+ queue= current;
+ current= next;
+ }
+ DBUG_ASSERT(leader == queue /* the leader should be first in queue */);
+ /* Now we have in queue the list of transactions to be committed in order. */
DBUG_ASSERT(is_open());
if (likely(is_open())) // Should always be true
{
/*
- We only bother to write to the binary log if there is anything
- to write.
- */
- if (my_b_tell(cache) > 0)
+ Commit every transaction in the queue.
+
+ Note that we are doing this in a different thread than the one running
+ the transaction! So we are limited in the operations we can do. In
+ particular, we cannot call my_error() on behalf of a transaction, as
+ that obtains the THD from thread local storage. Instead, we must set
+ current->error and let the thread do the error reporting itself once
+ we wake it up.
+ */
+ for (current= queue; current != NULL; current= current->next)
{
- /*
- Log "BEGIN" at the beginning of every transaction. Here, a
- transaction is either a BEGIN..COMMIT block or a single
- statement in autocommit mode.
- */
- Query_log_event qinfo(thd, STRING_WITH_LEN("BEGIN"), TRUE, TRUE, 0);
+ binlog_trx_data *trx_data= current->trx_data;
+ IO_CACHE *cache= &trx_data->trans_log;
+
+ /* Skip log_xid for transactions without xid, marked by NULL end_event. */
+ if (!current->end_event)
+ continue;
/*
- Now this Query_log_event has artificial log_pos 0. It must be
- adjusted to reflect the real position in the log. Not doing it
- would confuse the slave: it would prevent this one from
- knowing where he is in the master's binlog, which would result
- in wrong positions being shown to the user, MASTER_POS_WAIT
- undue waiting etc.
+ We only bother to write to the binary log if there is anything
+ to write.
*/
- if (qinfo.write(&log_file))
- goto err;
-
- DBUG_EXECUTE_IF("crash_before_writing_xid",
- {
- if ((write_error= write_cache(cache, false, true)))
- DBUG_PRINT("info", ("error writing binlog cache: %d",
- write_error));
- DBUG_PRINT("info", ("crashing before writing xid"));
- abort();
- });
-
- if ((write_error= write_cache(cache, false, false)))
- goto err;
+ if (my_b_tell(cache) > 0)
+ {
+ current->error= write_transaction(current);
+ if (current->error)
+ current->commit_errno= errno;
- if (commit_event && commit_event->write(&log_file))
- goto err;
+ write_count++;
+ }
- if (incident && write_incident(thd, FALSE))
- goto err;
+ trx_data->commit_bin_log_file_pos=
+ log_file.pos_in_file + (log_file.write_pos - log_file.write_buffer);
+ if (trx_data->using_xa)
+ xid_count++;
+ }
+ if (write_count > 0)
+ {
if (flush_and_sync())
- goto err;
- DBUG_EXECUTE_IF("half_binlogged_transaction", DBUG_ABORT(););
- if (cache->error) // Error on read
{
- sql_print_error(ER(ER_ERROR_ON_READ), cache->file_name, errno);
- write_error=1; // Don't give more errors
- goto err;
+ for (current= queue; current != NULL; current= current->next)
+ {
+ if (!current->error)
+ {
+ current->error= ER_ERROR_ON_WRITE;
+ current->commit_errno= errno;
+ }
+ }
+ }
+ else
+ {
+ signal_update();
}
- signal_update();
}
/*
- if commit_event is Xid_log_event, increase the number of
- prepared_xids (it's decreasd in ::unlog()). Binlog cannot be rotated
+ if any commit_events are Xid_log_event, increase the number of
+ prepared_xids (it's decreased in ::unlog()). Binlog cannot be rotated
if there're prepared xids in it - see the comment in new_file() for
an explanation.
- If the commit_event is not Xid_log_event (then it's a Query_log_event)
- rotate binlog, if necessary.
+ If no Xid_log_events (then it's all Query_log_event) rotate binlog,
+ if necessary.
*/
- if (commit_event && commit_event->get_type_code() == XID_EVENT)
+ if (xid_count > 0)
{
- pthread_mutex_lock(&LOCK_prep_xids);
- prepared_xids++;
- pthread_mutex_unlock(&LOCK_prep_xids);
+ mark_xids_active(xid_count);
}
else
rotate_and_purge(RP_LOCK_LOG_IS_ALREADY_LOCKED);
}
- VOID(pthread_mutex_unlock(&LOCK_log));
- DBUG_RETURN(0);
+ DEBUG_SYNC(leader->thd, "commit_before_get_LOCK_commit_ordered");
+ pthread_mutex_lock(&LOCK_commit_ordered);
+ /*
+ We cannot unlock LOCK_log until we have locked LOCK_commit_ordered;
+ otherwise scheduling could allow the next group commit to run ahead of us,
+ messing up the order of commit_ordered() calls. But as soon as
+ LOCK_commit_ordered is obtained, we can let the next group commit start.
+ */
+ pthread_mutex_unlock(&LOCK_log);
+ DEBUG_SYNC(leader->thd, "commit_after_release_LOCK_log");
+ ++num_group_commits;
-err:
- if (!write_error)
+ /*
+ Wakeup each participant waiting for our group commit, first calling the
+ commit_ordered() methods for any transactions doing 2-phase commit.
+ */
+ current= queue;
+ while (current != NULL)
{
- write_error= 1;
- sql_print_error(ER(ER_ERROR_ON_WRITE), name, errno);
+ DEBUG_SYNC(leader->thd, "commit_loop_entry_commit_ordered");
+ ++num_commits;
+ if (current->trx_data->using_xa && !current->error)
+ run_commit_ordered(current->thd, current->all);
+
+ /*
+ Careful not to access current->next after waking up the other thread! As
+ it may change immediately after wakeup.
+ */
+ group_commit_entry *next= current->next;
+ if (current != leader) // Don't wake up ourself
+ current->thd->signal_wakeup_ready();
+ current= next;
}
- VOID(pthread_mutex_unlock(&LOCK_log));
- DBUG_RETURN(1);
+ DEBUG_SYNC(leader->thd, "commit_after_group_run_commit_ordered");
+ pthread_mutex_unlock(&LOCK_commit_ordered);
+
+ DBUG_VOID_RETURN;
}
+int
+MYSQL_BIN_LOG::write_transaction(group_commit_entry *entry)
+{
+ binlog_trx_data *trx_data= entry->trx_data;
+ IO_CACHE *cache= &trx_data->trans_log;
+ /*
+ Log "BEGIN" at the beginning of every transaction. Here, a transaction is
+ either a BEGIN..COMMIT block or a single statement in autocommit mode. The
+ event was constructed in write_transaction_to_binlog(), in the thread
+ running the transaction.
+
+ Now this Query_log_event has artificial log_pos 0. It must be
+ adjusted to reflect the real position in the log. Not doing it
+ would confuse the slave: it would prevent this one from
+ knowing where he is in the master's binlog, which would result
+ in wrong positions being shown to the user, MASTER_POS_WAIT
+ undue waiting etc.
+ */
+ if (entry->begin_event->write(&log_file))
+ return ER_ERROR_ON_WRITE;
+
+ DBUG_EXECUTE_IF("crash_before_writing_xid",
+ {
+ if ((write_cache(cache)))
+ DBUG_PRINT("info", ("error writing binlog cache"));
+ else
+ flush_and_sync();
+
+ DBUG_PRINT("info", ("crashing before writing xid"));
+ abort();
+ });
+
+ if (write_cache(cache))
+ return ER_ERROR_ON_WRITE;
+
+ if (entry->end_event->write(&log_file))
+ return ER_ERROR_ON_WRITE;
+
+ if (entry->incident_event && entry->incident_event->write(&log_file))
+ return ER_ERROR_ON_WRITE;
+
+ if (cache->error) // Error on read
+ return ER_ERROR_ON_READ;
+
+ return 0;
+}
/**
Wait until we get a signal that the binary log has been updated.
@@ -5340,6 +5578,172 @@ void sql_print_information(const char *f
}
+static my_bool mutexes_inited;
+pthread_mutex_t LOCK_prepare_ordered;
+pthread_mutex_t LOCK_commit_ordered;
+
+void
+TC_init()
+{
+ my_pthread_mutex_init(&LOCK_prepare_ordered, MY_MUTEX_INIT_SLOW,
+ "LOCK_prepare_ordered", MYF(0));
+ my_pthread_mutex_init(&LOCK_commit_ordered, MY_MUTEX_INIT_SLOW,
+ "LOCK_commit_ordered", MYF(0));
+ mutexes_inited= TRUE;
+}
+
+void
+TC_destroy()
+{
+ if (mutexes_inited)
+ {
+ pthread_mutex_destroy(&LOCK_prepare_ordered);
+ pthread_mutex_destroy(&LOCK_commit_ordered);
+ mutexes_inited= FALSE;
+ }
+}
+
+void
+TC_LOG::run_prepare_ordered(THD *thd, bool all)
+{
+ Ha_trx_info *ha_info=
+ all ? thd->transaction.all.ha_list : thd->transaction.stmt.ha_list;
+
+ for (; ha_info; ha_info= ha_info->next())
+ {
+ handlerton *ht= ha_info->ht();
+ if (!ht->prepare_ordered)
+ continue;
+ safe_mutex_assert_owner(&LOCK_prepare_ordered);
+ ht->prepare_ordered(ht, thd, all);
+ }
+}
+
+void
+TC_LOG::run_commit_ordered(THD *thd, bool all)
+{
+ Ha_trx_info *ha_info=
+ all ? thd->transaction.all.ha_list : thd->transaction.stmt.ha_list;
+
+ for (; ha_info; ha_info= ha_info->next())
+ {
+ handlerton *ht= ha_info->ht();
+ if (!ht->commit_ordered)
+ continue;
+ safe_mutex_assert_owner(&LOCK_commit_ordered);
+ ht->commit_ordered(ht, thd, all);
+ DEBUG_SYNC(thd, "commit_after_run_commit_ordered");
+ }
+}
+
+int TC_LOG_MMAP::log_and_order(THD *thd, my_xid xid, bool all,
+ bool need_prepare_ordered,
+ bool need_commit_ordered)
+{
+ int cookie;
+ struct commit_entry entry;
+ bool is_group_commit_leader;
+ LINT_INIT(is_group_commit_leader);
+
+ if (need_prepare_ordered)
+ {
+ pthread_mutex_lock(&LOCK_prepare_ordered);
+ run_prepare_ordered(thd, all);
+ if (need_commit_ordered)
+ {
+ /*
+ Must put us in queue so we can run_commit_ordered() in same sequence
+ as we did run_prepare_ordered().
+ */
+ thd->clear_wakeup_ready();
+ entry.thd= thd;
+ commit_entry *previous_queue= commit_ordered_queue;
+ entry.next= previous_queue;
+ commit_ordered_queue= &entry;
+ is_group_commit_leader= (previous_queue == NULL);
+ }
+ pthread_mutex_unlock(&LOCK_prepare_ordered);
+ }
+
+ if (xid)
+ cookie= log_one_transaction(xid);
+ else
+ cookie= 0;
+
+ if (need_commit_ordered)
+ {
+ if (need_prepare_ordered)
+ {
+ /*
+ We did the run_prepare_ordered() serialised, then ran the log_xid() in
+ parallel. Now we have to do run_commit_ordered() serialised in the
+ same sequence as run_prepare_ordered().
+
+ We do this starting from the head of the queue, each thread doing
+ run_commit_ordered() and signalling the next in queue.
+ */
+ if (is_group_commit_leader)
+ {
+ /* The first in queue starts the ball rolling. */
+ pthread_mutex_lock(&LOCK_prepare_ordered);
+ while (commit_ordered_queue_busy)
+ pthread_cond_wait(&COND_queue_busy, &LOCK_prepare_ordered);
+ commit_entry *queue= commit_ordered_queue;
+ commit_ordered_queue= NULL;
+ /*
+ Mark the queue busy while we bounce it from one thread to the
+ next.
+ */
+ commit_ordered_queue_busy= true;
+ pthread_mutex_unlock(&LOCK_prepare_ordered);
+
+ /* Reverse the queue list so we get correct order. */
+ commit_entry *prev= NULL;
+ while (queue)
+ {
+ commit_entry *next= queue->next;
+ queue->next= prev;
+ prev= queue;
+ queue= next;
+ }
+ DBUG_ASSERT(prev == &entry && prev->thd == thd);
+ }
+ else
+ {
+ /* Not first in queue; just wait until previous thread wakes us up. */
+ thd->wait_for_wakeup_ready();
+ }
+ }
+
+ /* Only run commit_ordered() if log_xid was successful. */
+ if (cookie)
+ {
+ pthread_mutex_lock(&LOCK_commit_ordered);
+ run_commit_ordered(thd, all);
+ pthread_mutex_unlock(&LOCK_commit_ordered);
+ }
+
+ if (need_prepare_ordered)
+ {
+ commit_entry *next= entry.next;
+ if (next)
+ {
+ next->thd->signal_wakeup_ready();
+ }
+ else
+ {
+ pthread_mutex_lock(&LOCK_prepare_ordered);
+ commit_ordered_queue_busy= false;
+ pthread_cond_signal(&COND_queue_busy);
+ pthread_mutex_unlock(&LOCK_prepare_ordered);
+ }
+ }
+ }
+
+ return cookie;
+}
+
+
/********* transaction coordinator log for 2pc - mmap() based solution *******/
/*
@@ -5476,6 +5880,7 @@ int TC_LOG_MMAP::open(const char *opt_na
pthread_mutex_init(&LOCK_pool, MY_MUTEX_INIT_FAST);
pthread_cond_init(&COND_active, 0);
pthread_cond_init(&COND_pool, 0);
+ pthread_cond_init(&COND_queue_busy, 0);
inited=6;
@@ -5483,6 +5888,8 @@ int TC_LOG_MMAP::open(const char *opt_na
active=pages;
pool=pages+1;
pool_last=pages+npages-1;
+ commit_ordered_queue= NULL;
+ commit_ordered_queue_busy= false;
return 0;
@@ -5588,7 +5995,7 @@ int TC_LOG_MMAP::overflow()
to the position in memory where xid was logged to.
*/
-int TC_LOG_MMAP::log_xid(THD *thd, my_xid xid)
+int TC_LOG_MMAP::log_one_transaction(my_xid xid)
{
int err;
PAGE *p;
@@ -5757,6 +6164,8 @@ void TC_LOG_MMAP::close()
pthread_mutex_destroy(&LOCK_active);
pthread_mutex_destroy(&LOCK_pool);
pthread_cond_destroy(&COND_pool);
+ pthread_cond_destroy(&COND_active);
+ pthread_cond_destroy(&COND_queue_busy);
case 5:
data[0]='A'; // garble the first (signature) byte, in case my_delete fails
case 4:
@@ -5965,30 +6374,66 @@ void TC_LOG_BINLOG::close()
pthread_cond_destroy (&COND_prep_xids);
}
-/**
- @todo
- group commit
-
- @retval
- 0 error
- @retval
- 1 success
+/*
+ Do a binlog log_xid() for a group of transactions, linked through
+ thd->next_commit_ordered.
*/
-int TC_LOG_BINLOG::log_xid(THD *thd, my_xid xid)
+int
+TC_LOG_BINLOG::log_and_order(THD *thd, my_xid xid, bool all,
+ bool need_prepare_ordered __attribute__((unused)),
+ bool need_commit_ordered __attribute__((unused)))
{
- DBUG_ENTER("TC_LOG_BINLOG::log");
- Xid_log_event xle(thd, xid);
- binlog_trx_data *trx_data=
+ int err;
+ DBUG_ENTER("TC_LOG_BINLOG::log_and_order");
+
+ binlog_trx_data *const trx_data=
(binlog_trx_data*) thd_get_ha_data(thd, binlog_hton);
- /*
- We always commit the entire transaction when writing an XID. Also
- note that the return value is inverted.
- */
- DBUG_RETURN(!binlog_end_trans(thd, trx_data, &xle, TRUE));
+
+ trx_data->using_xa= TRUE;
+ if (xid)
+ {
+ Xid_log_event xid_event(thd, xid);
+ err= binlog_flush_trx_cache(thd, trx_data, &xid_event, all);
+ }
+ else
+ err= binlog_flush_trx_cache(thd, trx_data, NULL, all);
+
+ DBUG_RETURN(!err);
}
-void TC_LOG_BINLOG::unlog(ulong cookie, my_xid xid)
+/*
+ After an XID is logged, we need to hold on to the current binlog file until
+ it is fully committed in the storage engine. The reason is that crash
+ recovery only looks at the latest binlog, so we must make sure there are no
+ outstanding prepared (but not committed) transactions before rotating the
+ binlog.
+
+ To handle this, we keep a count of outstanding XIDs. This function is used
+ to increase this count when committing one or more transactions to the
+ binary log.
+*/
+void
+TC_LOG_BINLOG::mark_xids_active(uint xid_count)
+{
+ DBUG_ENTER("TC_LOG_BINLOG::mark_xids_active");
+ DBUG_PRINT("info", ("xid_count=%u", xid_count));
+ pthread_mutex_lock(&LOCK_prep_xids);
+ prepared_xids+= xid_count;
+ pthread_mutex_unlock(&LOCK_prep_xids);
+ DBUG_VOID_RETURN;
+}
+
+/*
+ Once an XID is committed, it is safe to rotate the binary log, as it can no
+ longer be needed during crash recovery.
+
+ This function is called to mark an XID this way. It needs to decrease the
+ count of pending XIDs, and signal the log rotator thread when it reaches zero.
+*/
+void
+TC_LOG_BINLOG::mark_xid_done()
{
+ DBUG_ENTER("TC_LOG_BINLOG::mark_xid_done");
pthread_mutex_lock(&LOCK_prep_xids);
DBUG_ASSERT(prepared_xids > 0);
if (--prepared_xids == 0) {
@@ -5996,7 +6441,16 @@ void TC_LOG_BINLOG::unlog(ulong cookie,
pthread_cond_signal(&COND_prep_xids);
}
pthread_mutex_unlock(&LOCK_prep_xids);
- rotate_and_purge(0); // as ::write() did not rotate
+ DBUG_VOID_RETURN;
+}
+
+void TC_LOG_BINLOG::unlog(ulong cookie, my_xid xid)
+{
+ DBUG_ENTER("TC_LOG_BINLOG::unlog");
+ if (xid)
+ mark_xid_done();
+ rotate_and_purge(0); // as ::write_transaction_to_binlog() did not rotate
+ DBUG_VOID_RETURN;
}
int TC_LOG_BINLOG::recover(IO_CACHE *log, Format_description_log_event *fdle)
@@ -6065,9 +6519,105 @@ ulonglong mysql_bin_log_file_pos(void)
{
return (ulonglong) mysql_bin_log.get_log_file()->pos_in_file;
}
+/*
+ Get the current position of the MySQL binlog for transaction currently being
+ committed.
+
+ This is valid to call from within storage engine commit_ordered() and
+ commit() methods only.
+
+ Since it stores the position inside THD, it is safe to call without any
+ locking.
+
+ Note that currently the binlog file name is not stored inside THD, but this
+ is still safe as it can only change when the log is rotated, and we never
+ rotate the binlog while commits are pending inside storage engines.
+*/
+void
+mysql_bin_log_commit_pos(THD *thd, ulonglong *out_pos, const char **out_file)
+{
+ binlog_trx_data *const trx_data=
+ (binlog_trx_data*) thd_get_ha_data(thd, binlog_hton);
+ if (trx_data)
+ {
+ *out_pos= trx_data->commit_bin_log_file_pos;
+ *out_file= mysql_bin_log.get_log_fname();
+ }
+ else
+ {
+ *out_pos= NULL;
+ *out_file= NULL;
+ }
+}
#endif /* INNODB_COMPATIBILITY_HOOKS */
+static ulonglong binlog_status_var_num_commits;
+static ulonglong binlog_status_var_num_group_commits;
+
+static SHOW_VAR binlog_status_vars_detail[]=
+{
+ {"commits",
+ (char *)&binlog_status_var_num_commits, SHOW_LONGLONG},
+ {"group_commits",
+ (char *)&binlog_status_var_num_group_commits, SHOW_LONGLONG},
+ {NullS, NullS, SHOW_LONG}
+};
+
+static int show_binlog_vars(THD *thd, SHOW_VAR *var, char *buff)
+{
+ mysql_bin_log.set_status_variables();
+ var->type= SHOW_ARRAY;
+ var->value= (char *)&binlog_status_vars_detail;
+ return 0;
+}
+
+static SHOW_VAR binlog_status_vars_top[]= {
+ {"binlog", (char *) &show_binlog_vars, SHOW_FUNC},
+ {NullS, NullS, SHOW_LONG}
+};
+
+#ifndef DBUG_OFF
+static MYSQL_SYSVAR_ULONG(
+ dbug_fsync_sleep,
+ opt_binlog_dbug_fsync_sleep,
+ PLUGIN_VAR_RQCMDARG,
+ "Extra sleep (in microseconds) to add to binlog fsync(), for debugging",
+ NULL,
+ NULL,
+ 0,
+ 0,
+ ULONG_MAX,
+ 0);
+
+static struct st_mysql_sys_var *binlog_sys_vars[]=
+{
+ MYSQL_SYSVAR(dbug_fsync_sleep),
+ NULL
+};
+#endif
+
+
+/*
+ Copy out current values of status variables, for SHOW STATUS or
+ information_schema.global_status.
+
+ This is called only under LOCK_status, so we can fill in a static array.
+*/
+void
+TC_LOG_BINLOG::set_status_variables()
+{
+ ulonglong num_commits, num_group_commits;
+
+ pthread_mutex_lock(&LOCK_commit_ordered);
+ num_commits= this->num_commits;
+ num_group_commits= this->num_group_commits;
+ pthread_mutex_unlock(&LOCK_commit_ordered);
+
+ binlog_status_var_num_commits= num_commits;
+ binlog_status_var_num_group_commits= num_group_commits;
+}
+
struct st_mysql_storage_engine binlog_storage_engine=
{ MYSQL_HANDLERTON_INTERFACE_VERSION };
@@ -6082,8 +6632,12 @@ mysql_declare_plugin(binlog)
binlog_init, /* Plugin Init */
NULL, /* Plugin Deinit */
0x0100 /* 1.0 */,
- NULL, /* status variables */
+ binlog_status_vars_top, /* status variables */
+#ifndef DBUG_OFF
+ binlog_sys_vars, /* system variables */
+#else
NULL, /* system variables */
+#endif
NULL /* config options */
}
mysql_declare_plugin_end;
=== modified file 'sql/log.h'
--- sql/log.h 2010-08-02 19:48:56 +0000
+++ sql/log.h 2010-10-28 20:00:51 +0000
@@ -38,17 +38,58 @@ class TC_LOG
virtual int open(const char *opt_name)=0;
virtual void close()=0;
- virtual int log_xid(THD *thd, my_xid xid)=0;
+ virtual int log_and_order(THD *thd, my_xid xid, bool all,
+ bool need_prepare_ordered,
+ bool need_commit_ordered) = 0;
virtual void unlog(ulong cookie, my_xid xid)=0;
+
+protected:
+ /*
+ These methods are meant to be invoked from log_and_order() implementations
+ to run any prepare_ordered() respectively commit_ordered() methods in
+ participating handlers.
+
+ They must be called using suitable thread syncronisation to ensure that
+ they are each called in the correct commit order among all
+ transactions. However, it is only necessary to call them if the
+ corresponding flag passed to log_and_order is set (it is safe, but not
+ required, to call them when the flag is false).
+
+ The caller must be holding LOCK_prepare_ordered respectively
+ LOCK_commit_ordered when calling these methods.
+ */
+ void run_prepare_ordered(THD *thd, bool all);
+ void run_commit_ordered(THD *thd, bool all);
};
+/*
+ Locks used to ensure serialised execution of TC_LOG::run_prepare_ordered()
+ and TC_LOG::run_commit_ordered(), or any other code that calls handler
+ prepare_ordered() or commit_ordered() methods.
+*/
+extern pthread_mutex_t LOCK_prepare_ordered;
+extern pthread_mutex_t LOCK_commit_ordered;
+
+extern void TC_init();
+extern void TC_destroy();
+
class TC_LOG_DUMMY: public TC_LOG // use it to disable the logging
{
public:
TC_LOG_DUMMY() {}
int open(const char *opt_name) { return 0; }
void close() { }
- int log_xid(THD *thd, my_xid xid) { return 1; }
+ /*
+ TC_LOG_DUMMY is only used when there are <= 1 XA-capable engines, and we
+ only use internal XA during commit when >= 2 XA-capable engines
+ participate.
+ */
+ int log_and_order(THD *thd, my_xid xid, bool all,
+ bool need_prepare_ordered, bool need_commit_ordered)
+ {
+ DBUG_ASSERT(0 /* Internal error - TC_LOG_DUMMY::log_and_order() called */);
+ return 1;
+ }
void unlog(ulong cookie, my_xid xid) { }
};
@@ -74,6 +115,13 @@ class TC_LOG_MMAP: public TC_LOG
pthread_cond_t cond; // to wait for a sync
} PAGE;
+ /* List of THDs for which to invoke commit_ordered(), in order. */
+ struct commit_entry
+ {
+ struct commit_entry *next;
+ THD *thd;
+ };
+
char logname[FN_REFLEN];
File fd;
my_off_t file_length;
@@ -88,16 +136,38 @@ class TC_LOG_MMAP: public TC_LOG
*/
pthread_mutex_t LOCK_active, LOCK_pool, LOCK_sync;
pthread_cond_t COND_pool, COND_active;
+ /*
+ Queue of threads that need to call commit_ordered().
+ Access to this queue must be protected by LOCK_prepare_ordered.
+ */
+ commit_entry *commit_ordered_queue;
+ /*
+ This flag and condition is used to reserve the queue while threads in it
+ each run the commit_ordered() methods one after the other. Only once the
+ last commit_ordered() in the queue is done can we start on a new queue
+ run.
+
+ Since we start this process in the first thread in the queue and finish in
+ the last (and possibly different) thread, we need a condition variable for
+ this (we cannot unlock a mutex in a different thread than the one who
+ locked it).
+
+ The condition is used together with the LOCK_prepare_ordered mutex.
+ */
+ my_bool commit_ordered_queue_busy;
+ pthread_cond_t COND_queue_busy;
public:
TC_LOG_MMAP(): inited(0) {}
int open(const char *opt_name);
void close();
- int log_xid(THD *thd, my_xid xid);
+ int log_and_order(THD *thd, my_xid xid, bool all,
+ bool need_prepare_ordered, bool need_commit_ordered);
void unlog(ulong cookie, my_xid xid);
int recover();
private:
+ int log_one_transaction(my_xid xid);
void get_active_from_pool();
int sync();
int overflow();
@@ -232,12 +302,35 @@ private:
time_t last_time;
};
+class binlog_trx_data;
class MYSQL_BIN_LOG: public TC_LOG, private MYSQL_LOG
{
private:
+ struct group_commit_entry
+ {
+ struct group_commit_entry *next;
+ THD *thd;
+ binlog_trx_data *trx_data;
+ /*
+ Extra events (BEGIN, COMMIT/ROLLBACK/XID, and possibly INCIDENT) to be
+ written during group commit. The incident_event is only valid if
+ trx_data->has_incident() is true.
+ */
+ Log_event *begin_event;
+ Log_event *end_event;
+ Log_event *incident_event;
+ /* Set during group commit to record any per-thread error. */
+ int error;
+ int commit_errno;
+ /* This is the `all' parameter for ha_commit_ordered(). */
+ bool all;
+ /* True if we come in through XA log_and_order(), false otherwise. */
+ };
+
/* LOCK_log and LOCK_index are inited by init_pthread_objects() */
pthread_mutex_t LOCK_index;
pthread_mutex_t LOCK_prep_xids;
+
pthread_cond_t COND_prep_xids;
pthread_cond_t update_cond;
ulonglong bytes_written;
@@ -276,6 +369,12 @@ class MYSQL_BIN_LOG: public TC_LOG, priv
In 5.0 it's 0 for relay logs too!
*/
bool no_auto_events;
+ /* Queue of transactions queued up to participate in group commit. */
+ group_commit_entry *group_commit_queue;
+ /* Total number of committed transactions. */
+ ulonglong num_commits;
+ /* Number of group commits done. */
+ ulonglong num_group_commits;
int write_to_file(IO_CACHE *cache);
/*
@@ -285,6 +384,11 @@ class MYSQL_BIN_LOG: public TC_LOG, priv
*/
void new_file_without_locking();
void new_file_impl(bool need_lock);
+ int write_transaction(group_commit_entry *entry);
+ bool write_transaction_to_binlog_events(group_commit_entry *entry);
+ void trx_group_commit_leader(group_commit_entry *leader);
+ void mark_xid_done();
+ void mark_xids_active(uint xid_count);
public:
MYSQL_LOG::generate_name;
@@ -313,7 +417,8 @@ public:
int open(const char *opt_name);
void close();
- int log_xid(THD *thd, my_xid xid);
+ int log_and_order(THD *thd, my_xid xid, bool all,
+ bool need_prepare_ordered, bool need_commit_ordered);
void unlog(ulong cookie, my_xid xid);
int recover(IO_CACHE *log, Format_description_log_event *fdle);
#if !defined(MYSQL_CLIENT)
@@ -358,10 +463,11 @@ public:
void reset_gathered_updates(THD *thd);
bool write(Log_event* event_info); // binary log write
- bool write(THD *thd, IO_CACHE *cache, Log_event *commit_event, bool incident);
+ bool write_transaction_to_binlog(THD *thd, binlog_trx_data *trx_data,
+ Log_event *end_ev, bool all);
- bool write_incident(THD *thd, bool lock);
- int write_cache(IO_CACHE *cache, bool lock_log, bool flush_and_sync);
+ bool write_incident(THD *thd);
+ int write_cache(IO_CACHE *cache);
void set_write_error(THD *thd);
bool check_write_error(THD *thd);
@@ -416,6 +522,7 @@ public:
inline void unlock_index() { pthread_mutex_unlock(&LOCK_index);}
inline IO_CACHE *get_index_file() { return &index_file;}
inline uint32 get_open_count() { return open_count; }
+ void set_status_variables();
};
class Log_event_handler
=== modified file 'sql/mysqld.cc'
--- sql/mysqld.cc 2010-10-20 19:42:17 +0000
+++ sql/mysqld.cc 2010-11-01 14:15:03 +0000
@@ -1373,6 +1373,7 @@ void clean_up(bool print_message)
ha_end();
if (tc_log)
tc_log->close();
+ TC_destroy();
xid_cache_free();
wt_end();
delete_elements(&key_caches, (void (*)(const char*, uchar*)) free_key_cache);
@@ -4195,6 +4196,8 @@ a file name for --log-bin-index option",
if (!errmesg[0][0])
unireg_abort(1);
+ TC_init();
+
/* We have to initialize the storage engines before CSV logging */
if (ha_init())
{
=== modified file 'sql/sql_class.cc'
--- sql/sql_class.cc 2010-08-27 14:12:44 +0000
+++ sql/sql_class.cc 2010-10-28 09:35:46 +0000
@@ -704,6 +704,8 @@ THD::THD()
active_vio = 0;
#endif
pthread_mutex_init(&LOCK_thd_data, MY_MUTEX_INIT_FAST);
+ pthread_mutex_init(&LOCK_wakeup_ready, MY_MUTEX_INIT_FAST);
+ pthread_cond_init(&COND_wakeup_ready, 0);
/* Variables with default values */
proc_info="login";
@@ -1035,6 +1037,8 @@ THD::~THD()
free_root(&transaction.mem_root,MYF(0));
#endif
mysys_var=0; // Safety (shouldn't be needed)
+ pthread_cond_destroy(&COND_wakeup_ready);
+ pthread_mutex_destroy(&LOCK_wakeup_ready);
pthread_mutex_destroy(&LOCK_thd_data);
#ifndef DBUG_OFF
dbug_sentry= THD_SENTRY_GONE;
@@ -4005,6 +4009,25 @@ int THD::binlog_query(THD::enum_binlog_q
DBUG_RETURN(0);
}
+void
+THD::wait_for_wakeup_ready()
+{
+ pthread_mutex_lock(&LOCK_wakeup_ready);
+ while (!wakeup_ready)
+ pthread_cond_wait(&COND_wakeup_ready, &LOCK_wakeup_ready);
+ pthread_mutex_unlock(&LOCK_wakeup_ready);
+}
+
+void
+THD::signal_wakeup_ready()
+{
+ pthread_mutex_lock(&LOCK_wakeup_ready);
+ wakeup_ready= true;
+ pthread_cond_signal(&COND_wakeup_ready);
+ pthread_mutex_unlock(&LOCK_wakeup_ready);
+}
+
+
bool Discrete_intervals_list::append(ulonglong start, ulonglong val,
ulonglong incr)
{
=== modified file 'sql/sql_class.h'
--- sql/sql_class.h 2010-10-19 13:58:35 +0000
+++ sql/sql_class.h 2010-11-01 14:15:03 +0000
@@ -2376,6 +2376,14 @@ public:
LEX_STRING get_invoker_user() { return invoker_user; }
LEX_STRING get_invoker_host() { return invoker_host; }
bool has_invoker() { return invoker_user.length > 0; }
+ void clear_wakeup_ready() { wakeup_ready= false; }
+ /*
+ Sleep waiting for others to wake us up with signal_wakeup_ready().
+ Must call clear_wakeup_ready() before waiting.
+ */
+ void wait_for_wakeup_ready();
+ /* Wake this thread up from wait_for_wakeup_ready(). */
+ void signal_wakeup_ready();
private:
/** The current internal error handler for this thread, or NULL. */
Internal_error_handler *m_internal_handler;
@@ -2414,6 +2422,16 @@ private:
*/
LEX_STRING invoker_user;
LEX_STRING invoker_host;
+ /*
+ Flag, mutex and condition for a thread to wait for a signal from another
+ thread.
+
+ Currently used to wait for group commit to complete, can also be used for
+ other purposes.
+ */
+ bool wakeup_ready;
+ pthread_mutex_t LOCK_wakeup_ready;
+ pthread_cond_t COND_wakeup_ready;
};
=== modified file 'sql/sql_parse.cc'
--- sql/sql_parse.cc 2010-10-19 13:58:35 +0000
+++ sql/sql_parse.cc 2010-11-01 14:15:03 +0000
@@ -1002,6 +1002,10 @@ bool dispatch_command(enum enum_server_c
DBUG_ENTER("dispatch_command");
DBUG_PRINT("info", ("command: %d", command));
+ DBUG_EXECUTE_IF("crash_dispatch_command_before",
+ { DBUG_PRINT("crash_dispatch_command_before", ("now"));
+ DBUG_ABORT(); });
+
thd->command=command;
/*
Commands which always take a long time are logged into
=== modified file 'storage/xtradb/handler/ha_innodb.cc'
--- storage/xtradb/handler/ha_innodb.cc 2010-10-19 15:03:26 +0000
+++ storage/xtradb/handler/ha_innodb.cc 2010-11-01 14:15:03 +0000
@@ -117,8 +117,6 @@ bool check_global_access(THD *thd, ulong
/** to protect innobase_open_files */
static pthread_mutex_t innobase_share_mutex;
-/** to force correct commit order in binlog */
-static pthread_mutex_t prepare_commit_mutex;
static ulong commit_threads = 0;
static pthread_mutex_t commit_threads_m;
static pthread_cond_t commit_cond;
@@ -223,6 +221,7 @@ static const char* innobase_change_buffe
static INNOBASE_SHARE *get_share(const char *table_name);
static void free_share(INNOBASE_SHARE *share);
static int innobase_close_connection(handlerton *hton, THD* thd);
+static void innobase_commit_ordered(handlerton *hton, THD* thd, bool all);
static int innobase_commit(handlerton *hton, THD* thd, bool all);
static int innobase_rollback(handlerton *hton, THD* thd, bool all);
static int innobase_rollback_to_savepoint(handlerton *hton, THD* thd,
@@ -1366,7 +1365,6 @@ innobase_trx_init(
trx_t* trx) /*!< in/out: InnoDB transaction handle */
{
DBUG_ENTER("innobase_trx_init");
- DBUG_ASSERT(EQ_CURRENT_THD(thd));
DBUG_ASSERT(thd == trx->mysql_thd);
trx->check_foreigns = !thd_test_options(
@@ -1425,8 +1423,6 @@ check_trx_exists(
{
trx_t*& trx = thd_to_trx(thd);
- ut_ad(EQ_CURRENT_THD(thd));
-
if (trx == NULL) {
trx = innobase_trx_allocate(thd);
} else if (UNIV_UNLIKELY(trx->magic_n != TRX_MAGIC_N)) {
@@ -1718,10 +1714,10 @@ innobase_query_caching_of_table_permitte
/* The call of row_search_.. will start a new transaction if it is
not yet started */
- if (trx->active_trans == 0) {
+ if ((trx->active_trans & TRX_ACTIVE_IN_MYSQL) == 0) {
innobase_register_trx_and_stmt(innodb_hton_ptr, thd);
- trx->active_trans = 1;
+ trx->active_trans |= TRX_ACTIVE_IN_MYSQL;
}
if (row_search_check_if_query_cache_permitted(trx, norm_name)) {
@@ -1991,11 +1987,11 @@ ha_innobase::init_table_handle_for_HANDL
/* Set the MySQL flag to mark that there is an active transaction */
- if (prebuilt->trx->active_trans == 0) {
+ if ((prebuilt->trx->active_trans & TRX_ACTIVE_IN_MYSQL) == 0) {
innobase_register_trx_and_stmt(ht, user_thd);
- prebuilt->trx->active_trans = 1;
+ prebuilt->trx->active_trans |= TRX_ACTIVE_IN_MYSQL;
}
/* We did the necessary inits in this function, no need to repeat them
@@ -2046,6 +2042,7 @@ innobase_init(
innobase_hton->savepoint_set=innobase_savepoint;
innobase_hton->savepoint_rollback=innobase_rollback_to_savepoint;
innobase_hton->savepoint_release=innobase_release_savepoint;
+ innobase_hton->commit_ordered=innobase_commit_ordered;
innobase_hton->commit=innobase_commit;
innobase_hton->rollback=innobase_rollback;
innobase_hton->prepare=innobase_xa_prepare;
@@ -2549,7 +2546,6 @@ skip_overwrite:
innobase_open_tables = hash_create(200);
pthread_mutex_init(&innobase_share_mutex, MY_MUTEX_INIT_FAST);
- pthread_mutex_init(&prepare_commit_mutex, MY_MUTEX_INIT_FAST);
pthread_mutex_init(&commit_threads_m, MY_MUTEX_INIT_FAST);
pthread_mutex_init(&commit_cond_m, MY_MUTEX_INIT_FAST);
pthread_cond_init(&commit_cond, NULL);
@@ -2604,7 +2600,6 @@ innobase_end(
my_free(internal_innobase_data_file_path,
MYF(MY_ALLOW_ZERO_PTR));
pthread_mutex_destroy(&innobase_share_mutex);
- pthread_mutex_destroy(&prepare_commit_mutex);
pthread_mutex_destroy(&commit_threads_m);
pthread_mutex_destroy(&commit_cond_m);
pthread_cond_destroy(&commit_cond);
@@ -2725,14 +2720,119 @@ innobase_start_trx_and_assign_read_view(
/* Set the MySQL flag to mark that there is an active transaction */
- if (trx->active_trans == 0) {
+ if ((trx->active_trans & TRX_ACTIVE_IN_MYSQL) == 0) {
innobase_register_trx_and_stmt(hton, thd);
- trx->active_trans = 1;
+ trx->active_trans |= TRX_ACTIVE_IN_MYSQL;
}
DBUG_RETURN(0);
}
+static
+void
+innobase_commit_ordered_2(
+/*============*/
+ trx_t* trx, /*!< in: Innodb transaction */
+ THD* thd) /*!< in: MySQL thread handle */
+{
+ ulonglong tmp_pos;
+ DBUG_ENTER("innobase_commit_ordered");
+
+ /* We need current binlog position for ibbackup to work.
+ Note, the position is current because commit_ordered is guaranteed
+ to be called in same sequenece as writing to binlog. */
+
+retry:
+ if (innobase_commit_concurrency > 0) {
+ pthread_mutex_lock(&commit_cond_m);
+ commit_threads++;
+
+ if (commit_threads > innobase_commit_concurrency) {
+ commit_threads--;
+ pthread_cond_wait(&commit_cond,
+ &commit_cond_m);
+ pthread_mutex_unlock(&commit_cond_m);
+ goto retry;
+ }
+ else {
+ pthread_mutex_unlock(&commit_cond_m);
+ }
+ }
+
+ mysql_bin_log_commit_pos(thd, &tmp_pos, &(trx->mysql_log_file_name));
+ trx->mysql_log_offset = (ib_int64_t) tmp_pos;
+
+ /* Don't do write + flush right now. For group commit
+ to work we want to do the flush in the innobase_commit()
+ method, which runs without holding any locks. */
+ trx->flush_log_later = TRUE;
+ innobase_commit_low(trx);
+ trx->flush_log_later = FALSE;
+
+ if (innobase_commit_concurrency > 0) {
+ pthread_mutex_lock(&commit_cond_m);
+ commit_threads--;
+ pthread_cond_signal(&commit_cond);
+ pthread_mutex_unlock(&commit_cond_m);
+ }
+
+ DBUG_VOID_RETURN;
+}
+
+/*****************************************************************//**
+Perform the first, fast part of InnoDB commit.
+
+Doing it in this call ensures that we get the same commit order here
+as in binlog and any other participating transactional storage engines.
+
+Note that we want to do as little as really needed here, as we run
+under a global mutex. The expensive fsync() is done later, in
+innobase_commit(), without a lock so group commit can take place.
+
+Note also that this method can be called from a different thread than
+the one handling the rest of the transaction. */
+static
+void
+innobase_commit_ordered(
+/*============*/
+ handlerton *hton, /*!< in: Innodb handlerton */
+ THD* thd, /*!< in: MySQL thread handle of the user for whom
+ the transaction should be committed */
+ bool all) /*!< in: TRUE - commit transaction
+ FALSE - the current SQL statement ended */
+{
+ trx_t* trx;
+ DBUG_ENTER("innobase_commit_ordered");
+ DBUG_ASSERT(hton == innodb_hton_ptr);
+
+ trx = check_trx_exists(thd);
+
+ /* Since we will reserve the kernel mutex, we have to release
+ the search system latch first to obey the latching order. */
+
+ if (trx->has_search_latch) {
+ trx_search_latch_release_if_reserved(trx);
+ }
+
+ if ((trx->active_trans & TRX_ACTIVE_IN_MYSQL) == 0
+ && trx->conc_state != TRX_NOT_STARTED) {
+ /* We cannot throw error here; instead we will catch this error
+ again in innobase_commit() and report it from there. */
+ DBUG_VOID_RETURN;
+ }
+
+ /* commit_ordered is only called when committing the whole transaction
+ (or an SQL statement when autocommit is on). */
+ DBUG_ASSERT(all ||
+ (!thd_test_options(thd, OPTION_NOT_AUTOCOMMIT | OPTION_BEGIN)));
+
+ innobase_commit_ordered_2(trx, thd);
+
+ trx->active_trans |= TRX_ACTIVE_COMMIT_ORDERED;
+
+ DBUG_VOID_RETURN;
+}
+
/*****************************************************************//**
Commits a transaction in an InnoDB database or marks an SQL statement
ended.
@@ -2758,11 +2858,12 @@ innobase_commit(
/* Since we will reserve the kernel mutex, we have to release
the search system latch first to obey the latching order. */
- if (trx->has_search_latch) {
+ if (trx->has_search_latch &&
+ (trx->active_trans & TRX_ACTIVE_COMMIT_ORDERED) == 0) {
trx_search_latch_release_if_reserved(trx);
}
- /* The flag trx->active_trans is set to 1 in
+ /* The flag TRX_ACTIVE_IN_MYSQL in trx->active_trans is set in
1. ::external_lock(),
2. ::start_stmt(),
@@ -2772,79 +2873,31 @@ innobase_commit(
6. innobase_start_trx_and_assign_read_view(),
7. ::transactional_table_lock()
- and it is only set to 0 in a commit or a rollback. If it is 0 we know
+ and it is only cleared in a commit or a rollback. If it is unset we know
there cannot be resources to be freed and we could return immediately.
For the time being, we play safe and do the cleanup though there should
be nothing to clean up. */
- if (trx->active_trans == 0
+ if ((trx->active_trans & TRX_ACTIVE_IN_MYSQL) == 0
&& trx->conc_state != TRX_NOT_STARTED) {
sql_print_error("trx->active_trans == 0, but"
" trx->conc_state != TRX_NOT_STARTED");
}
+
if (all
|| (!thd_test_options(thd, OPTION_NOT_AUTOCOMMIT | OPTION_BEGIN))) {
- /* We were instructed to commit the whole transaction, or
- this is an SQL statement end and autocommit is on */
-
- /* We need current binlog position for ibbackup to work.
- Note, the position is current because of
- prepare_commit_mutex */
-retry:
- if (innobase_commit_concurrency > 0) {
- pthread_mutex_lock(&commit_cond_m);
- commit_threads++;
-
- if (commit_threads > innobase_commit_concurrency) {
- commit_threads--;
- pthread_cond_wait(&commit_cond,
- &commit_cond_m);
- pthread_mutex_unlock(&commit_cond_m);
- goto retry;
- }
- else {
- pthread_mutex_unlock(&commit_cond_m);
- }
- }
-
- /* The following calls to read the MySQL binary log
- file name and the position return consistent results:
- 1) Other InnoDB transactions cannot intervene between
- these calls as we are holding prepare_commit_mutex.
- 2) Binary logging of other engines is not relevant
- to InnoDB as all InnoDB requires is that committing
- InnoDB transactions appear in the same order in the
- MySQL binary log as they appear in InnoDB logs.
- 3) A MySQL log file rotation cannot happen because
- MySQL protects against this by having a counter of
- transactions in prepared state and it only allows
- a rotation when the counter drops to zero. See
- LOCK_prep_xids and COND_prep_xids in log.cc. */
- trx->mysql_log_file_name = mysql_bin_log_file_name();
- trx->mysql_log_offset = (ib_int64_t) mysql_bin_log_file_pos();
-
- /* Don't do write + flush right now. For group commit
- to work we want to do the flush after releasing the
- prepare_commit_mutex. */
- trx->flush_log_later = TRUE;
- innobase_commit_low(trx);
- trx->flush_log_later = FALSE;
-
- if (innobase_commit_concurrency > 0) {
- pthread_mutex_lock(&commit_cond_m);
- commit_threads--;
- pthread_cond_signal(&commit_cond);
- pthread_mutex_unlock(&commit_cond_m);
+ /* Run the fast part of commit if we did not already. */
+ if ((trx->active_trans & TRX_ACTIVE_COMMIT_ORDERED) == 0) {
+ innobase_commit_ordered_2(trx, thd);
}
- if (trx->active_trans == 2) {
-
- pthread_mutex_unlock(&prepare_commit_mutex);
- }
+ /* We were instructed to commit the whole transaction, or
+ this is an SQL statement end and autocommit is on */
- /* Now do a write + flush of logs. */
+ /* We did the first part already in innobase_commit_ordered(),
+ Now finish by doing a write + flush of logs. */
trx_commit_complete_for_mysql(trx);
trx->active_trans = 0;
@@ -3063,7 +3116,7 @@ innobase_savepoint(
innobase_release_stat_resources(trx);
/* cannot happen outside of transaction */
- DBUG_ASSERT(trx->active_trans);
+ DBUG_ASSERT(trx->active_trans & TRX_ACTIVE_IN_MYSQL);
/* TODO: use provided savepoint data area to store savepoint data */
char name[64];
@@ -3093,7 +3146,7 @@ innobase_close_connection(
ut_a(trx);
- if (trx->active_trans == 0
+ if ((trx->active_trans & TRX_ACTIVE_IN_MYSQL) == 0
&& trx->conc_state != TRX_NOT_STARTED) {
sql_print_error("trx->active_trans == 0, but"
@@ -5017,7 +5070,7 @@ no_commit:
/* Altering to InnoDB format */
innobase_commit(ht, user_thd, 1);
/* Note that this transaction is still active. */
- prebuilt->trx->active_trans = 1;
+ prebuilt->trx->active_trans |= TRX_ACTIVE_IN_MYSQL;
/* We will need an IX lock on the destination table. */
prebuilt->sql_stat_start = TRUE;
} else {
@@ -5033,7 +5086,7 @@ no_commit:
locks, so they have to be acquired again. */
innobase_commit(ht, user_thd, 1);
/* Note that this transaction is still active. */
- prebuilt->trx->active_trans = 1;
+ prebuilt->trx->active_trans |= TRX_ACTIVE_IN_MYSQL;
/* Re-acquire the table lock on the source table. */
row_lock_table_for_mysql(prebuilt, src_table, mode);
/* We will need an IX lock on the destination table. */
@@ -8923,10 +8976,10 @@ ha_innobase::start_stmt(
trx->detailed_error[0] = '\0';
/* Set the MySQL flag to mark that there is an active transaction */
- if (trx->active_trans == 0) {
+ if ((trx->active_trans & TRX_ACTIVE_IN_MYSQL) == 0) {
innobase_register_trx_and_stmt(ht, thd);
- trx->active_trans = 1;
+ trx->active_trans |= TRX_ACTIVE_IN_MYSQL;
} else {
innobase_register_stmt(ht, thd);
}
@@ -9024,10 +9077,10 @@ ha_innobase::external_lock(
/* Set the MySQL flag to mark that there is an active
transaction */
- if (trx->active_trans == 0) {
+ if ((trx->active_trans & TRX_ACTIVE_IN_MYSQL) == 0) {
innobase_register_trx_and_stmt(ht, thd);
- trx->active_trans = 1;
+ trx->active_trans |= TRX_ACTIVE_IN_MYSQL;
} else if (trx->n_mysql_tables_in_use == 0) {
innobase_register_stmt(ht, thd);
}
@@ -9125,7 +9178,7 @@ ha_innobase::external_lock(
prebuilt->used_in_HANDLER = FALSE;
if (!thd_test_options(thd, OPTION_NOT_AUTOCOMMIT | OPTION_BEGIN)) {
- if (trx->active_trans != 0) {
+ if ((trx->active_trans & TRX_ACTIVE_IN_MYSQL) != 0) {
innobase_commit(ht, thd, TRUE);
}
} else {
@@ -9210,10 +9263,10 @@ ha_innobase::transactional_table_lock(
/* MySQL is setting a new transactional table lock */
/* Set the MySQL flag to mark that there is an active transaction */
- if (trx->active_trans == 0) {
+ if ((trx->active_trans & TRX_ACTIVE_IN_MYSQL) == 0) {
innobase_register_trx_and_stmt(ht, thd);
- trx->active_trans = 1;
+ trx->active_trans |= TRX_ACTIVE_IN_MYSQL;
}
if (THDVAR(thd, table_locks) && thd_in_lock_tables(thd)) {
@@ -10266,7 +10319,8 @@ innobase_xa_prepare(
innobase_release_stat_resources(trx);
- if (trx->active_trans == 0 && trx->conc_state != TRX_NOT_STARTED) {
+ if ((trx->active_trans & TRX_ACTIVE_IN_MYSQL) == 0 &&
+ trx->conc_state != TRX_NOT_STARTED) {
sql_print_error("trx->active_trans == 0, but trx->conc_state != "
"TRX_NOT_STARTED");
@@ -10278,7 +10332,7 @@ innobase_xa_prepare(
/* We were instructed to prepare the whole transaction, or
this is an SQL statement end and autocommit is on */
- ut_ad(trx->active_trans);
+ ut_ad(trx->active_trans & TRX_ACTIVE_IN_MYSQL);
error = (int) trx_prepare_for_mysql(trx);
} else {
@@ -10302,36 +10356,6 @@ innobase_xa_prepare(
srv_active_wake_master_thread();
- if (thd_sql_command(thd) != SQLCOM_XA_PREPARE &&
- (all || !thd_test_options(thd, OPTION_NOT_AUTOCOMMIT | OPTION_BEGIN)))
- {
- if (srv_enable_unsafe_group_commit && !THDVAR(thd, support_xa)) {
- /* choose group commit rather than binlog order */
- return(error);
- }
-
- /* For ibbackup to work the order of transactions in binlog
- and InnoDB must be the same. Consider the situation
-
- thread1> prepare; write to binlog; ...
- <context switch>
- thread2> prepare; write to binlog; commit
- thread1> ... commit
-
- To ensure this will not happen we're taking the mutex on
- prepare, and releasing it on commit.
-
- Note: only do it for normal commits, done via ha_commit_trans.
- If 2pc protocol is executed by external transaction
- coordinator, it will be just a regular MySQL client
- executing XA PREPARE and XA COMMIT commands.
- In this case we cannot know how many minutes or hours
- will be between XA PREPARE and XA COMMIT, and we don't want
- to block for undefined period of time. */
- pthread_mutex_lock(&prepare_commit_mutex);
- trx->active_trans = 2;
- }
-
return(error);
}
@@ -11619,11 +11643,6 @@ static MYSQL_SYSVAR_ENUM(adaptive_checkp
"Enable/Disable flushing along modified age. (none, reflex, [estimate])",
NULL, innodb_adaptive_checkpoint_update, 2, &adaptive_checkpoint_typelib);
-static MYSQL_SYSVAR_ULONG(enable_unsafe_group_commit, srv_enable_unsafe_group_commit,
- PLUGIN_VAR_RQCMDARG,
- "Enable/Disable unsafe group commit when support_xa=OFF and use with binlog or other XA storage engine.",
- NULL, NULL, 0, 0, 1, 0);
-
static MYSQL_SYSVAR_ULONG(expand_import, srv_expand_import,
PLUGIN_VAR_RQCMDARG,
"Enable/Disable converting automatically *.ibd files when import tablespace.",
@@ -11729,7 +11748,6 @@ static struct st_mysql_sys_var* innobase
MYSQL_SYSVAR(read_ahead),
MYSQL_SYSVAR(adaptive_checkpoint),
MYSQL_SYSVAR(flush_log_at_trx_commit_session),
- MYSQL_SYSVAR(enable_unsafe_group_commit),
MYSQL_SYSVAR(expand_import),
MYSQL_SYSVAR(extra_rsegments),
MYSQL_SYSVAR(dict_size_limit),
=== modified file 'storage/xtradb/handler/ha_innodb.h'
--- storage/xtradb/handler/ha_innodb.h 2010-09-03 17:42:54 +0000
+++ storage/xtradb/handler/ha_innodb.h 2010-10-17 21:39:52 +0000
@@ -239,16 +239,6 @@ LEX_STRING *thd_query_string(MYSQL_THD t
char **thd_query(MYSQL_THD thd);
#endif
-/** Get the file name of the MySQL binlog.
- * @return the name of the binlog file
- */
-const char* mysql_bin_log_file_name(void);
-
-/** Get the current position of the MySQL binlog.
- * @return byte offset from the beginning of the binlog
- */
-ulonglong mysql_bin_log_file_pos(void);
-
/**
Check if a user thread is a replication slave thread
@param thd user thread
@@ -289,6 +279,11 @@ bool thd_binlog_filter_ok(const MYSQL_TH
#endif /* MYSQL_VERSION_ID > 50140 */
}
+/** Get the file name and position of the MySQL binlog corresponding to the
+ * current commit.
+ */
+extern void mysql_bin_log_commit_pos(THD *thd, ulonglong *out_pos, const char **out_file);
+
typedef struct trx_struct trx_t;
/********************************************************************//**
@file handler/ha_innodb.h
=== modified file 'storage/xtradb/include/trx0trx.h'
--- storage/xtradb/include/trx0trx.h 2010-09-03 15:41:57 +0000
+++ storage/xtradb/include/trx0trx.h 2010-10-04 12:59:30 +0000
@@ -511,9 +511,10 @@ struct trx_struct{
in that case we must flush the log
in trx_commit_complete_for_mysql() */
ulint duplicates; /*!< TRX_DUP_IGNORE | TRX_DUP_REPLACE */
- ulint active_trans; /*!< 1 - if a transaction in MySQL
- is active. 2 - if prepare_commit_mutex
- was taken */
+ ulint active_trans; /*!< TRX_ACTIVE_IN_MYSQL - set if a
+ transaction in MySQL is active.
+ TRX_ACTIVE_COMMIT_ORDERED - set if
+ innobase_commit_ordered has run */
ulint has_search_latch;
/* TRUE if this trx has latched the
search system latch in S-mode */
@@ -824,6 +825,10 @@ Multiple flags can be combined with bitw
#define TRX_SIG_OTHER_SESS 1 /* sent by another session (which
must hold rights to this) */
+/* Flag bits for trx_struct.active_trans */
+#define TRX_ACTIVE_IN_MYSQL (1<<0)
+#define TRX_ACTIVE_COMMIT_ORDERED (1<<1)
+
/** Commit node states */
enum commit_node_state {
COMMIT_NODE_SEND = 1, /*!< about to send a commit signal to
Follow ups
References
