MDEV-26: Global transaction ID

 - Add first basic mysql-test-run test case which tests switch to new master
   using MASTER_GTID_POS=AUTO.

 - When we connect with GTID, do not use any old relay logs, as they may
   contain the wrong events or be corrupt after crash.

 - Fix old bug that fails replication if we receive a heartbeat event
   immediately after an event was omitted in the stream from the master.

 - Fix rpl_end to clear Gtid_Pos_Auto, to keep check_testcase happy.

mysql-test/include/rpl_end.inc

@@ -79,6 +79,22 @@ while ($_rpl_server)
 --source include/rpl_sync.inc
 --source include/rpl_stop_slaves.inc
 
+if (!$rpl_debug)
+{
+  --disable_query_log
+}
+--let $_rpl_server= $rpl_server_count
+while ($_rpl_server)
+{
+  --let $rpl_connection_name= server_$_rpl_server
+  --source include/rpl_connection.inc
+
+  # Clear Gtid_Pos_Auto in SHOW SLAVE STATUS to keep check_testcase happy.
+  CHANGE MASTER TO master_log_file='';
+
+  --dec $_rpl_server
+}
+
 # mtr configures server 2 to be a slave before it runs the test. We
 # have to restore that state now, so we change topology to 1->2.
 --let $rpl_topology= none

mysql-test/suite/rpl/r/rpl_gtid_basic.result

@@ -44,5 +44,74 @@ a	b
 2	i2
 3	i3
 4	i4
+*** Now take out D, let it fall behind a bit, and then test re-attaching it to A ***
+include/stop_slave.inc
+INSERT INTO t1 VALUES (5, "m1a");
+INSERT INTO t2 VALUES (5, "i1a");
+CHANGE MASTER TO master_host = '127.0.0.1', master_port = MASTER_PORT,
+MASTER_GTID_POS=AUTO;
+include/start_slave.inc
+SELECT * FROM t1 ORDER BY a;
+a	b
+1	m1
+2	m2
+3	m3
+4	m4
+5	m1a
+SELECT * FROM t2 ORDER BY a;
+a	b
+1	i1
+2	i2
+3	i3
+4	i4
+5	i1a
+*** Now move B to D (C is still replicating from B) ***
+include/stop_slave.inc
+CHANGE MASTER TO master_host = '127.0.0.1', master_port = SERVER_MYPORT_4,
+MASTER_GTID_POS=AUTO;
+include/start_slave.inc
+UPDATE t2 SET b="j1a" WHERE a=5;
+SELECT * FROM t1 ORDER BY a;
+a	b
+1	m1
+2	m2
+3	m3
+4	m4
+5	m1a
+SELECT * FROM t2 ORDER BY a;
+a	b
+1	i1
+2	i2
+3	i3
+4	i4
+5	j1a
+*** Now move C to D, after letting it fall a little behind ***
+include/stop_slave.inc
+BEGIN;
+INSERT INTO t2 VALUES (6, "i6b");
+INSERT INTO t2 VALUES (7, "i7b");
+COMMIT;
+CHANGE MASTER TO master_host = '127.0.0.1', master_port = SERVER_MYPORT_4,
+MASTER_GTID_POS=AUTO;
+include/start_slave.inc
+SELECT * FROM t2 ORDER BY a;
+a	b
+1	i1
+2	i2
+3	i3
+4	i4
+5	j1a
+6	i6b
+7	i7b
+*** Now change everything back to what it was, to make rpl_end.inc happy
+include/stop_slave.inc
+CHANGE MASTER TO master_host = '127.0.0.1', master_port = MASTER_MYPORT;
+include/start_slave.inc
+include/stop_slave.inc
+CHANGE MASTER TO master_host = '127.0.0.1', master_port = SLAVE_MYPORT;
+include/start_slave.inc
+include/stop_slave.inc
+CHANGE MASTER TO master_host = '127.0.0.1', master_port = SERVER_MYPORT_3;
+include/start_slave.inc
 DROP TABLE t1,t2;
 include/rpl_end.inc

mysql-test/suite/rpl/t/rpl_gtid_basic.test

@@ -2,6 +2,11 @@
 --let $rpl_topology=1->2->3->4
 --source include/rpl_init.inc
 
+# Set up a 4-deep replication topology, then test various fail-overs
+# using GTID.
+#
+#  A -> B -> C -> D
+
 connection server_1;
 CREATE TABLE t1 (a INT PRIMARY KEY, b VARCHAR(10)) ENGINE=MyISAM;
 CREATE TABLE t2 (a INT PRIMARY KEY, b VARCHAR(10)) ENGINE=InnoDB;
@@ -31,6 +36,82 @@ sync_with_master;
 SELECT * FROM t1 ORDER BY a;
 SELECT * FROM t2 ORDER BY a;
 
+
+--echo *** Now take out D, let it fall behind a bit, and then test re-attaching it to A ***
+connection server_4;
+--source include/stop_slave.inc
+
+connection server_1;
+INSERT INTO t1 VALUES (5, "m1a");
+INSERT INTO t2 VALUES (5, "i1a");
+save_master_pos;
+
+connection server_4;
+--replace_result $MASTER_MYPORT MASTER_PORT
+eval CHANGE MASTER TO master_host = '127.0.0.1', master_port = $MASTER_MYPORT,
+     MASTER_GTID_POS=AUTO;
+--source include/start_slave.inc
+sync_with_master;
+SELECT * FROM t1 ORDER BY a;
+SELECT * FROM t2 ORDER BY a;
+
+--echo *** Now move B to D (C is still replicating from B) ***
+connection server_2;
+--source include/stop_slave.inc
+--replace_result $SERVER_MYPORT_4 SERVER_MYPORT_4
+eval CHANGE MASTER TO master_host = '127.0.0.1', master_port = $SERVER_MYPORT_4,
+     MASTER_GTID_POS=AUTO;
+--source include/start_slave.inc
+
+connection server_4;
+UPDATE t2 SET b="j1a" WHERE a=5;
+save_master_pos;
+
+connection server_2;
+sync_with_master;
+SELECT * FROM t1 ORDER BY a;
+SELECT * FROM t2 ORDER BY a;
+
+--echo *** Now move C to D, after letting it fall a little behind ***
+connection server_3;
+--source include/stop_slave.inc
+
+connection server_1;
+BEGIN;
+INSERT INTO t2 VALUES (6, "i6b");
+INSERT INTO t2 VALUES (7, "i7b");
+COMMIT;
+
+connection server_3;
+--replace_result $SERVER_MYPORT_4 SERVER_MYPORT_4
+eval CHANGE MASTER TO master_host = '127.0.0.1', master_port = $SERVER_MYPORT_4,
+     MASTER_GTID_POS=AUTO;
+--source include/start_slave.inc
+# This time, let's sync up without reference to binlog on D.
+--let $wait_condition= SELECT COUNT(*) = 7 FROM t2
+--source include/wait_condition.inc
+SELECT * FROM t2 ORDER BY a;
+
+--echo *** Now change everything back to what it was, to make rpl_end.inc happy
+# Also check that MASTER_GTID_POS=AUTO is still enabled.
+connection server_2;
+--source include/stop_slave.inc
+--replace_result $MASTER_MYPORT MASTER_MYPORT
+eval CHANGE MASTER TO master_host = '127.0.0.1', master_port = $MASTER_MYPORT;
+--source include/start_slave.inc
+
+connection server_3;
+--source include/stop_slave.inc
+--replace_result $SLAVE_MYPORT SLAVE_MYPORT
+eval CHANGE MASTER TO master_host = '127.0.0.1', master_port = $SLAVE_MYPORT;
+--source include/start_slave.inc
+
+connection server_4;
+--source include/stop_slave.inc
+--replace_result $SERVER_MYPORT_3 SERVER_MYPORT_3
+eval CHANGE MASTER TO master_host = '127.0.0.1', master_port = $SERVER_MYPORT_3;
+--source include/start_slave.inc
+
 connection server_1;
 DROP TABLE t1,t2;
 

sql/slave.cc

@@ -397,6 +397,33 @@ int init_recovery(Master_info* mi, const char** errmsg)
 
   DBUG_RETURN(0);
 }
+
+
+/*
+  When connecting a slave to a master with GTID, we reset the relay log
+  coordinates of the SQL thread and clear the master coordinates of SQL and IO
+  threads.
+
+  This way we ensure that we start from the correct place even after a change
+  to new master or a crash where relay log coordinates may be wrong (GTID
+  state is crash safe but master.info is not). And we get the correct master
+  coordinates set upon reading the initial fake rotate event sent from master.
+*/
+static void
+reset_coordinates_for_gtid(Master_info *mi, Relay_log_info *rli)
+{
+  mi->master_log_pos= 0;
+  mi->master_log_name[0]= 0;
+  rli->group_master_log_pos= 0;
+  rli->group_master_log_name[0]= 0;
+  rli->group_relay_log_pos= BIN_LOG_HEADER_SIZE;
+  strmake(rli->group_relay_log_name, rli->relay_log.get_log_fname(),
+          sizeof(rli->group_relay_log_name)-1);
+  rli->event_relay_log_pos= BIN_LOG_HEADER_SIZE;
+  strmake(rli->event_relay_log_name, rli->relay_log.get_log_fname(),
+          sizeof(mi->rli.event_relay_log_name)-1);
+}
+
  
 /**
   Convert slave skip errors bitmap into a printable string.
@@ -3385,6 +3412,8 @@ connected:
   if (ret == 1)
     /* Fatal error */
     goto err;
+  if (mi->gtid_pos_auto)
+    reset_coordinates_for_gtid(mi, rli);
 
   if (ret == 2) 
   { 
@@ -4694,16 +4723,18 @@ static int queue_event(Master_info* mi,const char* buf, ulong event_len)
        
        Heartbeat is sent only after an event corresponding to the corrdinates
        the heartbeat carries.
-       Slave can not have a difference in coordinates except in the only
+       Slave can not have a higher coordinate except in the only
        special case when mi->master_log_name, master_log_pos have never
        been updated by Rotate event i.e when slave does not have any history
        with the master (and thereafter mi->master_log_pos is NULL).
 
+       Slave can have lower coordinates, if some event from master was omitted.
+
        TODO: handling `when' for SHOW SLAVE STATUS' snds behind
     */
     if ((memcmp(mi->master_log_name, hb.get_log_ident(), hb.get_ident_len())
          && mi->master_log_name != NULL)
-        || mi->master_log_pos != hb.log_pos)
+        || mi->master_log_pos > hb.log_pos)
     {
       /* missed events of heartbeat from the past */
       error= ER_SLAVE_HEARTBEAT_FAILURE;