BUG#26395: if crash during autocommit update to transactional table on master, slave fails

Now, every transaction (including autocommit transactions) starts with
a BEGIN and ends with a COMMIT/ROLLBACK in the binlog.
Added a test case, and updated lots of test case result files.


mysql-test/r/multi_update.result:
  Updated result file
mysql-test/r/sp_trans_log.result:
  Updated result file
mysql-test/suite/binlog/r/binlog_innodb.result:
  Updated result file
mysql-test/suite/binlog/r/binlog_multi_engine.result:
  Updated result file
mysql-test/suite/binlog/r/binlog_stm_blackhole.result:
  Updated result file
mysql-test/suite/binlog/r/binlog_stm_mix_innodb_myisam.result:
  Updated result file
mysql-test/suite/ndb/r/ndb_binlog_format.result:
  Updated result file
mysql-test/suite/rpl/r/rpl_innodb_mixed_dml.result:
  Updated result file
mysql-test/suite/rpl/r/rpl_row_charset_innodb.result:
  Updated result file
mysql-test/suite/rpl/r/rpl_row_create_table.result:
  Updated result file
mysql-test/suite/rpl/r/rpl_row_log_innodb.result:
  Updated result file
mysql-test/suite/rpl/r/rpl_switch_stm_row_mixed.result:
  Updated result file
mysql-test/suite/rpl/r/rpl_truncate_3innodb.result:
  Updated result file
mysql-test/suite/rpl/t/rpl_row_create_table.test:
  Updated result file
mysql-test/suite/rpl_ndb/r/rpl_ndb_stm_innodb.result:
  Updated result file
sql/log.cc:
   - Always write BEGIN and COMMIT around statements, even in autocommit
     mode.
   - Added comments for binlog_commit and binlog_rollback.
sql/log_event.cc:
  Added debug trigger to avoid writing xid events to the binlog.
mysql-test/suite/rpl_ndb/r/rpl_ndb_transaction.result:
  Results for new test case
mysql-test/suite/rpl_ndb/t/rpl_ndb_transaction-master.opt:
  Options for new test case
mysql-test/suite/rpl_ndb/t/rpl_ndb_transaction-slave.opt:
  Options for new test case
mysql-test/suite/rpl_ndb/t/rpl_ndb_transaction.test:
  Added new test case.

2 files changed, 82 insertions, 29 deletions

diff --git a/sql/log.cc b/sql/log.cc
index 9fdede9ef2c..b6916c855ab 100644
--- a/sql/log.cc
+++ b/sql/log.cc
@@ -1402,6 +1402,21 @@ static int binlog_prepare(handlerton *hton, THD *thd, bool all)
   return 0;
 }
 
+/**
+  This function is called once after each statement.
+
+  It has the responsibility to flush the transaction cache to the
+  binlog file on commits.
+
+  @param hton  The binlog handlerton.
+  @param thd   The client thread that executes the transaction.
+  @param all   true if this is the last statement before a COMMIT
+               statement; false if either this is a statement in a
+               transaction but not the last, or if this is a statement
+               not inside a BEGIN block and autocommit is on.
+
+  @see handlerton::commit
+*/
 static int binlog_commit(handlerton *hton, THD *thd, bool all)
 {
   DBUG_ENTER("binlog_commit");
@@ -1414,7 +1429,15 @@ static int binlog_commit(handlerton *hton, THD *thd, bool all)
     trx_data->reset();
     DBUG_RETURN(0);
   }
-  if (all)
+  /*
+    Write commit event if at least one of the following holds:
+     - the user sends an explicit COMMIT; or
+     - the autocommit flag is on, and we are not inside a BEGIN.
+    However, if the user has not sent an explicit COMMIT, and we are
+    either inside a BEGIN or run with autocommit off, then this is not
+    the end of a transaction and we should not write a commit event.
+  */
+  if (all || !(thd->options & (OPTION_NOT_AUTOCOMMIT | OPTION_BEGIN)))
   {
     Query_log_event qev(thd, STRING_WITH_LEN("COMMIT"), TRUE, FALSE);
     qev.error_code= 0; // see comment in MYSQL_LOG::write(THD, IO_CACHE)
@@ -1428,6 +1451,23 @@ static int binlog_commit(handlerton *hton, THD *thd, bool all)
   }
 }
 
+/**
+  This function is called when a transaction involving a transactional
+  table is rolled back.
+
+  It has the responsibility to flush the transaction cache to the
+  binlog file. However, if the transaction does not involve
+  non-transactional tables, nothing needs to be logged.
+
+  @param hton  The binlog handlerton.
+  @param thd   The client thread that executes the transaction.
+  @param all   true if this is the last statement before a COMMIT
+               statement; false if either this is a statement in a
+               transaction but not the last, or if this is a statement
+               not inside a BEGIN block and autocommit is on.
+
+  @see handlerton::rollback
+*/
 static int binlog_rollback(handlerton *hton, THD *thd, bool all)
 {
   DBUG_ENTER("binlog_rollback");
@@ -3967,9 +4007,11 @@ int MYSQL_BIN_LOG::write_cache(IO_CACHE *cache, bool lock_log, bool sync_log)
   IMPLEMENTATION
     - To support transaction over replication, we wrap the transaction
       with BEGIN/COMMIT or BEGIN/ROLLBACK in the binary log.
-      We want to write a BEGIN/ROLLBACK block when a non-transactional table
-      was updated in a transaction which was rolled back. This is to ensure
-      that the same updates are run on the slave.
+      If a transaction that only involves transactional tables is
+      rolled back, we do not binlog it. However, we write a
+      BEGIN/ROLLBACK block when a non-transactional table was updated
+      in a transaction which was rolled back. This is to ensure that
+      the same updates are run on the slave.
 */
 
 bool MYSQL_BIN_LOG::write(THD *thd, IO_CACHE *cache, Log_event *commit_event)
@@ -3990,32 +4032,42 @@ bool MYSQL_BIN_LOG::write(THD *thd, IO_CACHE *cache, Log_event *commit_event)
     if (my_b_tell(cache) > 0)
     {
       /*
-        Log "BEGIN" at the beginning of the transaction.
-        which may contain more than 1 SQL statement.
+        Log "BEGIN" at the beginning of every transaction.  Here, a
+        transaction is either a BEGIN..COMMIT block or a single
+        statement in autocommit mode.
       */
-      if (thd->options & (OPTION_NOT_AUTOCOMMIT | OPTION_BEGIN))
-      {
-        Query_log_event qinfo(thd, STRING_WITH_LEN("BEGIN"), TRUE, FALSE);
-        /*
-          Imagine this is rollback due to net timeout, after all statements of
-          the transaction succeeded. Then we want a zero-error code in BEGIN.
-          In other words, if there was a really serious error code it's already
-          in the statement's events, there is no need to put it also in this
-          internally generated event, and as this event is generated late it
-          would lead to false alarms.
-          This is safer than thd->clear_error() against kills at shutdown.
-        */
-        qinfo.error_code= 0;
-        /*
-          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 (qinfo.write(&log_file))
-          goto err;
-      }
+      Query_log_event qinfo(thd, STRING_WITH_LEN("BEGIN"), TRUE, FALSE);
+      /*
+        Imagine this is rollback due to net timeout, after all
+        statements of the transaction succeeded. Then we want a
+        zero-error code in BEGIN.  In other words, if there was a
+        really serious error code it's already in the statement's
+        events, there is no need to put it also in this internally
+        generated event, and as this event is generated late it would
+        lead to false alarms.
+
+        This is safer than thd->clear_error() against kills at shutdown.
+      */
+      qinfo.error_code= 0;
+      /*
+        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 (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;
diff --git a/sql/log_event.cc b/sql/log_event.cc
index f328f266c05..cdfd58535fb 100644
--- a/sql/log_event.cc
+++ b/sql/log_event.cc
@@ -4292,6 +4292,7 @@ Xid_log_event(const char* buf,
 #ifndef MYSQL_CLIENT
 bool Xid_log_event::write(IO_CACHE* file)
 {
+  DBUG_EXECUTE_IF("do_not_write_xid", return 0;);
   return write_header(file, sizeof(xid)) ||
          my_b_safe_write(file, (uchar*) &xid, sizeof(xid));
 }