branches/zip

rb://130

Enable Group Commit functionality that was broken in 5.0 when
distributed transactions were introduced.

Reviewed by: Heikki

handler/ha_innodb.cc

@@ -2449,7 +2449,12 @@ retry:
 		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);
@@ -2463,6 +2468,8 @@ retry:
 			pthread_mutex_unlock(&prepare_commit_mutex);
 		}
 
+		/* Now do a write + flush of logs. */
+		trx_commit_complete_for_mysql(trx);
 		trx->active_trans = 0;
 
 	} else {
@@ -8934,33 +8941,6 @@ innobase_xa_prepare(
 
 	DBUG_ASSERT(hton == innodb_hton_ptr);
 
-	if (thd_sql_command(thd) != SQLCOM_XA_PREPARE &&
-	    (all || !thd_test_options(thd, OPTION_NOT_AUTOCOMMIT | OPTION_BEGIN)))
-	{
-
-		/* 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;
-	}
-
 	/* we use support_xa value as it was seen at transaction start
 	time, not the current session variable value. Any possible changes
 	to the session variable take effect only in the next transaction */
@@ -9013,6 +8993,33 @@ 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)))
+	{
+
+		/* 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);
 }
 

include/trx0trx.h

@@ -497,10 +497,12 @@ struct trx_struct{
 					FALSE, one can save CPU time and about
 					150 bytes in the undo log size as then
 					we skip XA steps */
-	unsigned	flush_log_later:1;/* when we commit the transaction
-					in MySQL's binlog write, we will
-					flush the log to disk later in
-					a separate call */
+	unsigned	flush_log_later:1;/* In 2PC, we hold the
+					prepare_commit mutex across
+					both phases. In that case, we
+					defer flush of the logs to disk
+					until after we release the
+					mutex. */
 	unsigned	must_flush_log_later:1;/* this flag is set to TRUE in
 					trx_commit_off_kernel() if
 					flush_log_later was TRUE, and there

trx/trx0trx.c

@@ -891,11 +891,11 @@ trx_commit_off_kernel(
 		there are > 2 users in the database. Then at least 2 users can
 		gather behind one doing the physical log write to disk.
 
-		If we are calling trx_commit() under MySQL's binlog mutex, we
+		If we are calling trx_commit() under prepare_commit_mutex, we
 		will delay possible log write and flush to a separate function
 		trx_commit_complete_for_mysql(), which is only called when the
-		thread has released the binlog mutex. This is to make the
-		group commit algorithm to work. Otherwise, the MySQL binlog
+		thread has released the mutex. This is to make the
+		group commit algorithm to work. Otherwise, the prepare_commit
 		mutex would serialize all commits and prevent a group of
 		transactions from gathering. */