Merge branch 'gtid_table_garbage_rows_10.3' into 10.3

knielsen · knielsen · commit 00164ea4b18b · 2018-10-13T19:50:53.000+02:00
diff --git a/mysql-test/suite/rpl/r/rpl_parallel_optimistic.result b/mysql-test/suite/rpl/r/rpl_parallel_optimistic.result
@@ -648,4 +648,12 @@ SET GLOBAL slave_parallel_threads=@old_parallel_threads;
 include/start_slave.inc
 connection server_1;
 DROP TABLE t1, t2, t3;
+include/save_master_gtid.inc
+connection server_2;
+include/sync_with_master_gtid.inc
+Check that no more than the expected last four GTIDs are in mysql.gtid_slave_pos
+select count(4) <= 4 from mysql.gtid_slave_pos order by domain_id, sub_id;
+count(4) <= 4
+1
+connection server_1;
 include/rpl_end.inc
diff --git a/mysql-test/suite/rpl/t/rpl_parallel_optimistic.test b/mysql-test/suite/rpl/t/rpl_parallel_optimistic.test
@@ -553,5 +553,29 @@ SET GLOBAL slave_parallel_threads=@old_parallel_threads;
 
 --connection server_1
 DROP TABLE t1, t2, t3;
+--source include/save_master_gtid.inc
 
+--connection server_2
+--source include/sync_with_master_gtid.inc
+# Check for left-over rows in table mysql.gtid_slave_pos (MDEV-12147).
+#
+# There was a bug when a transaction got a conflict and was rolled back. It
+# might have also handled deletion of some old rows, and these deletions would
+# then also be rolled back. And since the deletes were never re-tried, old no
+# longer needed rows would accumulate in the table without limit.
+# 
+# The earlier part of this test file have plenty of transactions being rolled
+# back. But the last DROP TABLE statement runs on its own and should never
+# conflict, thus at this point the mysql.gtid_slave_pos table should be clean.
+#
+# To support @@gtid_pos_auto_engines, when a row is inserted in the table, it
+# is associated with the engine of the table at insertion time, and it will
+# only be deleted during record_gtid from a table of the same engine. Since we
+# alter the table from MyISAM to InnoDB at the start of this test, we should
+# end up with 4 rows: two left-over from when the table was MyISAM, and two
+# left-over from the InnoDB part.
+--echo Check that no more than the expected last four GTIDs are in mysql.gtid_slave_pos
+select count(4) <= 4 from mysql.gtid_slave_pos order by domain_id, sub_id;
+
+--connection server_1
 --source include/rpl_end.inc
diff --git a/sql/log_event.cc b/sql/log_event.cc
@@ -5565,7 +5565,7 @@ int Query_log_event::do_apply_event(rpl_group_info *rgi,
           gtid= rgi->current_gtid;
           if (unlikely(rpl_global_gtid_slave_state->record_gtid(thd, &gtid,
                                                                 sub_id,
-                                                                true, false,
+                                                                rgi, false,
                                                                 &hton)))
           {
             int errcode= thd->get_stmt_da()->sql_errno();
@@ -8362,7 +8362,7 @@ Gtid_list_log_event::do_apply_event(rpl_group_info *rgi)
     {
       if ((ret= rpl_global_gtid_slave_state->record_gtid(thd, &list[i],
                                                          sub_id_list[i],
-                                                         false, false, &hton)))
+                                                         NULL, false, &hton)))
         return ret;
       rpl_global_gtid_slave_state->update_state_hash(sub_id_list[i], &list[i],
                                                      hton, NULL);
@@ -8899,7 +8899,7 @@ int Xid_log_event::do_apply_event(rpl_group_info *rgi)
     rgi->gtid_pending= false;
 
     gtid= rgi->current_gtid;
-    err= rpl_global_gtid_slave_state->record_gtid(thd, &gtid, sub_id, true,
+    err= rpl_global_gtid_slave_state->record_gtid(thd, &gtid, sub_id, rgi,
                                                   false, &hton);
     if (unlikely(err))
     {
diff --git a/sql/rpl_gtid.cc b/sql/rpl_gtid.cc
@@ -79,7 +79,7 @@ rpl_slave_state::record_and_update_gtid(THD *thd, rpl_group_info *rgi)
     rgi->gtid_pending= false;
     if (rgi->gtid_ignore_duplicate_state!=rpl_group_info::GTID_DUPLICATE_IGNORE)
     {
-      if (record_gtid(thd, &rgi->current_gtid, sub_id, false, false, &hton))
+      if (record_gtid(thd, &rgi->current_gtid, sub_id, NULL, false, &hton))
         DBUG_RETURN(1);
       update_state_hash(sub_id, &rgi->current_gtid, hton, rgi);
     }
@@ -331,6 +331,8 @@ rpl_slave_state::update(uint32 domain_id, uint32 server_id, uint64 sub_id,
       }
     }
     rgi->gtid_ignore_duplicate_state= rpl_group_info::GTID_DUPLICATE_NULL;
+
+    rgi->pending_gtid_deletes_clear();
   }
 
   if (!(list_elem= (list_element *)my_malloc(sizeof(*list_elem), MYF(MY_WME))))
@@ -381,15 +383,24 @@ int
 rpl_slave_state::put_back_list(uint32 domain_id, list_element *list)
 {
   element *e;
+  int err= 0;
+
+  mysql_mutex_lock(&LOCK_slave_state);
   if (!(e= (element *)my_hash_search(&hash, (const uchar *)&domain_id, 0)))
-    return 1;
+  {
+    err= 1;
+    goto end;
+  }
   while (list)
   {
     list_element *next= list->next;
     e->add(list);
     list= next;
   }
-  return 0;
+
+end:
+  mysql_mutex_unlock(&LOCK_slave_state);
+  return err;
 }
 
 
@@ -559,20 +570,20 @@ rpl_slave_state::select_gtid_pos_table(THD *thd, LEX_CSTRING *out_tablename)
 /*
   Write a gtid to the replication slave state table.
 
-  Do it as part of the transaction, to get slave crash safety, or as a separate
-  transaction if !in_transaction (eg. MyISAM or DDL).
-
     gtid    The global transaction id for this event group.
     sub_id  Value allocated within the sub_id when the event group was
             read (sub_id must be consistent with commit order in master binlog).
+    rgi     rpl_group_info context, if we are recording the gtid transactionally
+            as part of replicating a transactional event. NULL if called from
+            outside of a replicated transaction.
 
   Note that caller must later ensure that the new gtid and sub_id is inserted
   into the appropriate HASH element with rpl_slave_state.add(), so that it can
   be deleted later. But this must only be done after COMMIT if in transaction.
 */
 int
 rpl_slave_state::record_gtid(THD *thd, const rpl_gtid *gtid, uint64 sub_id,
-                             bool in_transaction, bool in_statement,
+                             rpl_group_info *rgi, bool in_statement,
                              void **out_hton)
 {
   TABLE_LIST tlist;
@@ -671,7 +682,7 @@ rpl_slave_state::record_gtid(THD *thd, const rpl_gtid *gtid, uint64 sub_id,
   thd->wsrep_ignore_table= true;
 #endif
 
-  if (!in_transaction)
+  if (!rgi)
   {
     DBUG_PRINT("info", ("resetting OPTION_BEGIN"));
     thd->variables.option_bits&=
@@ -776,7 +787,8 @@ rpl_slave_state::record_gtid(THD *thd, const rpl_gtid *gtid, uint64 sub_id,
     table->file->print_error(err, MYF(0));
     goto end;
   }
-  while (delete_list)
+  cur = delete_list;
+  while (cur)
   {
     uchar key_buffer[4+8];
 
@@ -786,9 +798,9 @@ rpl_slave_state::record_gtid(THD *thd, const rpl_gtid *gtid, uint64 sub_id,
                       /* `break' does not work inside DBUG_EXECUTE_IF */
                       goto dbug_break; });
 
-    next= delete_list->next;
+    next= cur->next;
 
-    table->field[1]->store(delete_list->sub_id, true);
+    table->field[1]->store(cur->sub_id, true);
     /* domain_id is already set in table->record[0] from write_row() above. */
     key_copy(key_buffer, table->record[0], &table->key_info[0], 0, false);
     if (table->file->ha_index_read_map(table->record[1], key_buffer,
@@ -802,8 +814,7 @@ rpl_slave_state::record_gtid(THD *thd, const rpl_gtid *gtid, uint64 sub_id,
       not want to endlessly error on the same element in case of table
       corruption or such.
     */
-    my_free(delete_list);
-    delete_list= next;
+    cur= next;
     if (err)
       break;
   }
@@ -826,18 +837,31 @@ IF_DBUG(dbug_break:, )
       */
       if (delete_list)
       {
-        mysql_mutex_lock(&LOCK_slave_state);
         put_back_list(gtid->domain_id, delete_list);
-        mysql_mutex_unlock(&LOCK_slave_state);
+        delete_list = 0;
       }
 
       ha_rollback_trans(thd, FALSE);
     }
     close_thread_tables(thd);
-    if (in_transaction)
+    if (rgi)
+    {
       thd->mdl_context.release_statement_locks();
+      /*
+        Save the list of old gtid entries we deleted. If this transaction
+        fails later for some reason and is rolled back, the deletion of those
+        entries will be rolled back as well, and we will need to put them back
+        on the to-be-deleted list so we can re-do the deletion. Otherwise
+        redundant rows in mysql.gtid_slave_pos may accumulate if transactions
+        are rolled back and retried after record_gtid().
+      */
+      rgi->pending_gtid_deletes_save(gtid->domain_id, delete_list);
+    }
     else
+    {
       thd->mdl_context.release_transactional_locks();
+      rpl_group_info::pending_gtid_deletes_free(delete_list);
+    }
   }
   thd->lex->restore_backup_query_tables_list(&lex_backup);
   thd->variables.option_bits= thd_saved_option;
@@ -1221,7 +1245,7 @@ rpl_slave_state::load(THD *thd, const char *state_from_master, size_t len,
 
     if (gtid_parser_helper(&state_from_master, end, &gtid) ||
         !(sub_id= next_sub_id(gtid.domain_id)) ||
-        record_gtid(thd, &gtid, sub_id, false, in_statement, &hton) ||
+        record_gtid(thd, &gtid, sub_id, NULL, in_statement, &hton) ||
         update(gtid.domain_id, gtid.server_id, sub_id, gtid.seq_no, hton, NULL))
       return 1;
     if (state_from_master == end)
diff --git a/sql/rpl_gtid.h b/sql/rpl_gtid.h
@@ -233,7 +233,7 @@ struct rpl_slave_state
   int truncate_state_table(THD *thd);
   void select_gtid_pos_table(THD *thd, LEX_CSTRING *out_tablename);
   int record_gtid(THD *thd, const rpl_gtid *gtid, uint64 sub_id,
-                  bool in_transaction, bool in_statement, void **out_hton);
+                  rpl_group_info *rgi, bool in_statement, void **out_hton);
   uint64 next_sub_id(uint32 domain_id);
   int iterate(int (*cb)(rpl_gtid *, void *), void *data,
               rpl_gtid *extra_gtids, uint32 num_extra,
diff --git a/sql/rpl_rli.cc b/sql/rpl_rli.cc
@@ -2086,6 +2086,7 @@ rpl_group_info::reinit(Relay_log_info *rli)
   long_find_row_note_printed= false;
   did_mark_start_commit= false;
   gtid_ev_flags2= 0;
+  pending_gtid_delete_list= NULL;
   last_master_timestamp = 0;
   gtid_ignore_duplicate_state= GTID_DUPLICATE_NULL;
   speculation= SPECULATE_NO;
@@ -2216,6 +2217,12 @@ void rpl_group_info::cleanup_context(THD *thd, bool error)
       erroneously update the GTID position.
     */
     gtid_pending= false;
+
+    /*
+      Rollback will have undone any deletions of old rows we might have made
+      in mysql.gtid_slave_pos. Put those rows back on the list to be deleted.
+    */
+    pending_gtid_deletes_put_back();
   }
   m_table_map.clear_tables();
   slave_close_thread_tables(thd);
@@ -2441,6 +2448,78 @@ rpl_group_info::unmark_start_commit()
 }
 
 
+/*
+  When record_gtid() has deleted any old rows from the table
+  mysql.gtid_slave_pos as part of a replicated transaction, save the list of
+  rows deleted here.
+
+  If later the transaction fails (eg. optimistic parallel replication), the
+  deletes will be undone when the transaction is rolled back. Then we can
+  put back the list of rows into the rpl_global_gtid_slave_state, so that
+  we can re-do the deletes and avoid accumulating old rows in the table.
+*/
+void
+rpl_group_info::pending_gtid_deletes_save(uint32 domain_id,
+                                          rpl_slave_state::list_element *list)
+{
+  /*
+    We should never get to a state where we try to save a new pending list of
+    gtid deletes while we still have an old one. But make sure we handle it
+    anyway just in case, so we avoid leaving stray entries in the
+    mysql.gtid_slave_pos table.
+  */
+  DBUG_ASSERT(!pending_gtid_delete_list);
+  if (unlikely(pending_gtid_delete_list))
+    pending_gtid_deletes_put_back();
+
+  pending_gtid_delete_list= list;
+  pending_gtid_delete_list_domain= domain_id;
+}
+
+
+/*
+  Take the list recorded by pending_gtid_deletes_save() and put it back into
+  rpl_global_gtid_slave_state. This is needed if deletion of the rows was
+  rolled back due to transaction failure.
+*/
+void
+rpl_group_info::pending_gtid_deletes_put_back()
+{
+  if (pending_gtid_delete_list)
+  {
+    rpl_global_gtid_slave_state->put_back_list(pending_gtid_delete_list_domain,
+                                               pending_gtid_delete_list);
+    pending_gtid_delete_list= NULL;
+  }
+}
+
+
+/*
+  Free the list recorded by pending_gtid_deletes_save(). Done when the deletes
+  in the list have been permanently committed.
+*/
+void
+rpl_group_info::pending_gtid_deletes_clear()
+{
+  pending_gtid_deletes_free(pending_gtid_delete_list);
+  pending_gtid_delete_list= NULL;
+}
+
+
+void
+rpl_group_info::pending_gtid_deletes_free(rpl_slave_state::list_element *list)
+{
+  rpl_slave_state::list_element *next;
+
+  while (list)
+  {
+    next= list->next;
+    my_free(list);
+    list= next;
+  }
+}
+
+
 rpl_sql_thread_info::rpl_sql_thread_info(Rpl_filter *filter)
   : rpl_filter(filter)
 {
diff --git a/sql/rpl_rli.h b/sql/rpl_rli.h
@@ -757,6 +757,11 @@ struct rpl_group_info
   /* Needs room for "Gtid D-S-N\x00". */
   char gtid_info_buf[5+10+1+10+1+20+1];
 
+  /* List of not yet committed deletions in mysql.gtid_slave_pos. */
+  rpl_slave_state::list_element *pending_gtid_delete_list;
+  /* Domain associated with pending_gtid_delete_list. */
+  uint32 pending_gtid_delete_list_domain;
+
   /*
     The timestamp, from the master, of the commit event.
     Used to do delayed update of rli->last_master_timestamp, for getting
@@ -898,6 +903,12 @@ struct rpl_group_info
   char *gtid_info();
   void unmark_start_commit();
 
+  static void pending_gtid_deletes_free(rpl_slave_state::list_element *list);
+  void pending_gtid_deletes_save(uint32 domain_id,
+                                 rpl_slave_state::list_element *list);
+  void pending_gtid_deletes_put_back();
+  void pending_gtid_deletes_clear();
+
   longlong get_row_stmt_start_timestamp()
   {
     return row_stmt_start_timestamp;
