MDEV-23586 Mariabackup: GTID saved for replication in 10.4.14 is wrong

MDEV-21953 deadlock between BACKUP STAGE BLOCK_COMMIT and parallel
replication

Fixed by partly reverting MDEV-21953 to put back MDL_BACKUP_COMMIT locking
before log_and_order.

The original problem for MDEV-21953 was that while a thread was waiting in
for another threads to commit in 'log_and_order', it had the
MDL_BACKUP_COMMIT lock. The backup thread was waiting to get the
MDL_BACKUP_WAIT_COMMIT lock, which blocks all new MDL_BACKUP_COMMIT locks.
This causes a deadlock as the waited-for thread can never get past the
MDL_BACKUP_COMMIT lock in ha_commit_trans.

The main part of the bug fix is to release the MDL_BACKUP_COMMIT lock while
a thread is waiting for other 'previous' threads to commit. This ensures
that no transactional thread keeps MDL_BACKUP_COMMIT while waiting, which
ensures that there are no deadlocks anymore.

sql/handler.cc

@@ -114,7 +114,7 @@ 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, bool rw_trans);
+                              bool is_real_trans);
 
 
 static plugin_ref ha_default_plugin(THD *thd)
@@ -1490,9 +1490,40 @@ int ha_commit_trans(THD *thd, bool all)
   /* rw_trans is TRUE when we in a transaction changing data */
   bool rw_trans= is_real_trans &&
                  (rw_ha_count > (thd->is_current_stmt_binlog_disabled()?0U:1U));
+  MDL_request mdl_backup;
   DBUG_PRINT("info", ("is_real_trans: %d  rw_trans:  %d  rw_ha_count: %d",
                       is_real_trans, rw_trans, rw_ha_count));
 
+  if (rw_trans)
+  {
+    /*
+      Acquire a metadata lock which will ensure that COMMIT is blocked
+      by an active FLUSH TABLES WITH READ LOCK (and vice versa:
+      COMMIT in progress blocks FTWRL).
+
+      We allow the owner of FTWRL to COMMIT; we assume that it knows
+      what it does.
+    */
+    mdl_backup.init(MDL_key::BACKUP, "", "", MDL_BACKUP_COMMIT, MDL_EXPLICIT);
+
+    if (!WSREP(thd))
+    {
+      if (thd->mdl_context.acquire_lock(&mdl_backup,
+                                        thd->variables.lock_wait_timeout))
+      {
+        ha_rollback_trans(thd, all);
+        DBUG_RETURN(1);
+      }
+      thd->backup_commit_lock= &mdl_backup;
+    }
+    DEBUG_SYNC(thd, "ha_commit_trans_after_acquire_commit_lock");
+
+    /* Use shortcut as we already have the MDL_BACKUP_COMMIT lock */
+    ha_maria::implicit_commit(thd, TRUE);
+  }
+  else
+    ha_maria_implicit_commit(thd, TRUE);
+
   if (rw_trans &&
       opt_readonly &&
       !(thd->security_ctx->master_access & SUPER_ACL) &&
@@ -1532,7 +1563,7 @@ int ha_commit_trans(THD *thd, bool all)
       // Here, the call will not commit inside InnoDB. It is only working
       // around closing thd->transaction.stmt open by TR_table::open().
       if (all)
-        commit_one_phase_2(thd, false, &thd->transaction.stmt, false, false);
+        commit_one_phase_2(thd, false, &thd->transaction.stmt, false);
     }
   }
 #endif
@@ -1552,7 +1583,7 @@ int ha_commit_trans(THD *thd, bool all)
       goto wsrep_err;
     }
 #endif /* WITH_WSREP */
-    error= ha_commit_one_phase(thd, all, rw_trans);
+    error= ha_commit_one_phase(thd, all);
 #ifdef WITH_WSREP
     if (run_wsrep_hooks)
       error= error || wsrep_after_commit(thd, all);
@@ -1604,7 +1635,7 @@ int ha_commit_trans(THD *thd, bool all)
 
   if (!is_real_trans)
   {
-    error= commit_one_phase_2(thd, all, trans, is_real_trans, rw_trans);
+    error= commit_one_phase_2(thd, all, trans, is_real_trans);
     goto done;
   }
 #ifdef WITH_WSREP
@@ -1622,7 +1653,7 @@ int ha_commit_trans(THD *thd, bool all)
   DEBUG_SYNC(thd, "ha_commit_trans_after_log_and_order");
   DBUG_EXECUTE_IF("crash_commit_after_log", DBUG_SUICIDE(););
 
-  error= commit_one_phase_2(thd, all, trans, is_real_trans, rw_trans) ? 2 : 0;
+  error= commit_one_phase_2(thd, all, trans, is_real_trans) ? 2 : 0;
 #ifdef WITH_WSREP
   if (run_wsrep_hooks && (error || (error = wsrep_after_commit(thd, all))))
   {
@@ -1685,6 +1716,17 @@ err:
                 thd->rgi_slave->is_parallel_exec);
   }
 end:
+  if (mdl_backup.ticket)
+  {
+    /*
+      We do not always immediately release transactional locks
+      after ha_commit_trans() (see uses of ha_enable_transaction()),
+      thus we release the commit blocker lock as soon as it's
+      not needed.
+    */
+    thd->mdl_context.release_lock(mdl_backup.ticket);
+  }
+  thd->backup_commit_lock= 0;
 #ifdef WITH_WSREP
   if (wsrep_is_active(thd) && is_real_trans && !error &&
       (rw_ha_count == 0 || all) &&
@@ -1699,8 +1741,8 @@ end:
 
 /**
   @note
-  This function does not care about global read lock. A caller should.
+  This function does not care about global read lock or backup locks,
-  However backup locks are handled in commit_one_phase_2.
+  the caller should.
 
   @param[in]  all  Is set in case of explicit commit
                    (COMMIT statement), or implicit commit
@@ -1709,7 +1751,7 @@ end:
                    autocommit=1.
 */
 
-int ha_commit_one_phase(THD *thd, bool all, bool rw_trans)
+int ha_commit_one_phase(THD *thd, bool all)
 {
   THD_TRANS *trans=all ? &thd->transaction.all : &thd->transaction.stmt;
   /*
@@ -1735,50 +1777,21 @@ int ha_commit_one_phase(THD *thd, bool all, bool rw_trans)
     if ((res= thd->wait_for_prior_commit()))
       DBUG_RETURN(res);
   }
-  res= commit_one_phase_2(thd, all, trans, is_real_trans, rw_trans);
+  res= commit_one_phase_2(thd, all, trans, is_real_trans);
   DBUG_RETURN(res);
 }
 
 
 static int
-commit_one_phase_2(THD *thd, bool all, THD_TRANS *trans, bool is_real_trans,
+commit_one_phase_2(THD *thd, bool all, THD_TRANS *trans, bool is_real_trans)
-                   bool rw_trans)
 {
   int error= 0;
   uint count= 0;
   Ha_trx_info *ha_info= trans->ha_list, *ha_info_next;
-  MDL_request mdl_request;
   DBUG_ENTER("commit_one_phase_2");
   if (is_real_trans)
     DEBUG_SYNC(thd, "commit_one_phase_2");
 
-  if (rw_trans)
-  {
-    /*
-      Acquire a metadata lock which will ensure that COMMIT is blocked
-      by an active FLUSH TABLES WITH READ LOCK (and vice versa:
-      COMMIT in progress blocks FTWRL).
-
-      We allow the owner of FTWRL to COMMIT; we assume that it knows
-      what it does.
-    */
-    mdl_request.init(MDL_key::BACKUP, "", "", MDL_BACKUP_COMMIT, MDL_EXPLICIT);
-
-    if (!WSREP(thd) &&
-      thd->mdl_context.acquire_lock(&mdl_request,
-                                    thd->variables.lock_wait_timeout))
-    {
-      my_error(ER_ERROR_DURING_COMMIT, MYF(0), 1);
-      ha_rollback_trans(thd, all);
-      DBUG_RETURN(1);
-    }
-    DEBUG_SYNC(thd, "ha_commit_trans_after_acquire_commit_lock");
-  }
-
-#if defined(WITH_ARIA_STORAGE_ENGINE) && MYSQL_VERSION_ID < 100500
-  ha_maria::implicit_commit(thd, TRUE);
-#endif
-
   if (ha_info)
   {
     for (; ha_info; ha_info= ha_info_next)
@@ -1807,16 +1820,6 @@ commit_one_phase_2(THD *thd, bool all, THD_TRANS *trans, bool is_real_trans,
 #endif
     }
   }
-  if (mdl_request.ticket)
-  {
-    /*
-      We do not always immediately release transactional locks
-      after ha_commit_trans() (see uses of ha_enable_transaction()),
-      thus we release the commit blocker lock as soon as it's
-      not needed.
-    */
-    thd->mdl_context.release_lock(mdl_request.ticket);
-  }
 
   /* Free resources and perform other cleanup even for 'empty' transactions. */
   if (is_real_trans)

sql/handler.h

@@ -5024,7 +5024,7 @@ int ha_change_key_cache(KEY_CACHE *old_key_cache, KEY_CACHE *new_key_cache);
 /* transactions: interface to handlerton functions */
 int ha_start_consistent_snapshot(THD *thd);
 int ha_commit_or_rollback_by_xid(XID *xid, bool commit);
-int ha_commit_one_phase(THD *thd, bool all, bool rw_trans);
+int ha_commit_one_phase(THD *thd, bool all);
 int ha_commit_trans(THD *thd, bool all);
 int ha_rollback_trans(THD *thd, bool all);
 int ha_prepare(THD *thd);

sql/log.cc

@@ -6380,11 +6380,25 @@ bool MYSQL_BIN_LOG::write(Log_event *event_info, my_bool *with_annotate)
 
     if (direct)
     {
+      /* We come here only for incident events */
       int res;
       uint64 commit_id= 0;
+      MDL_request mdl_request;
       DBUG_PRINT("info", ("direct is set"));
+      DBUG_ASSERT(!thd->backup_commit_lock);
+
+      mdl_request.init(MDL_key::BACKUP, "", "", MDL_BACKUP_COMMIT, MDL_EXPLICIT);
+      thd->mdl_context.acquire_lock(&mdl_request,
+                                    thd->variables.lock_wait_timeout);
+      thd->backup_commit_lock= &mdl_request;
+
       if ((res= thd->wait_for_prior_commit()))
+      {
+        if (mdl_request.ticket)
+          thd->mdl_context.release_lock(mdl_request.ticket);
+        thd->backup_commit_lock= 0;
         DBUG_RETURN(res);
+      }
       file= &log_file;
       my_org_b_tell= my_b_tell(file);
       mysql_mutex_lock(&LOCK_log);
@@ -6399,7 +6413,11 @@ bool MYSQL_BIN_LOG::write(Log_event *event_info, my_bool *with_annotate)
                                              commit_name.length);
           commit_id= entry->val_int(&null_value);
         });
-      if (write_gtid_event(thd, true, using_trans, commit_id))
+      res= write_gtid_event(thd, true, using_trans, commit_id);
+      if (mdl_request.ticket)
+        thd->mdl_context.release_lock(mdl_request.ticket);
+      thd->backup_commit_lock= 0;
+      if (res)
         goto err;
     }
     else
@@ -7461,7 +7479,11 @@ MYSQL_BIN_LOG::queue_for_group_commit(group_commit_entry *orig_entry)
   group_commit_entry *entry, *orig_queue, *last;
   wait_for_commit *cur;
   wait_for_commit *wfc;
+  bool backup_lock_released= 0;
+  int result= 0;
+  THD *thd= orig_entry->thd;
   DBUG_ENTER("MYSQL_BIN_LOG::queue_for_group_commit");
+  DBUG_ASSERT(thd == current_thd);
 
   /*
     Check if we need to wait for another transaction to commit before us.
@@ -7493,6 +7515,21 @@ MYSQL_BIN_LOG::queue_for_group_commit(group_commit_entry *orig_entry)
     {
       PSI_stage_info old_stage;
 
+        /*
+          Release MDL_BACKUP_COMMIT LOCK while waiting for other threads to
+          commit.
+          This is needed to avoid deadlock between the other threads (which not
+          yet have the MDL_BACKUP_COMMIT_LOCK) and any threads using
+          BACKUP LOCK BLOCK_COMMIT.
+        */
+      if (thd->backup_commit_lock && thd->backup_commit_lock->ticket &&
+          !backup_lock_released)
+      {
+        backup_lock_released= 1;
+        thd->mdl_context.release_lock(thd->backup_commit_lock->ticket);
+        thd->backup_commit_lock->ticket= 0;
+      }
+
       /*
         By setting wfc->opaque_pointer to our own entry, we mark that we are
         ready to commit, but waiting for another transaction to commit before
@@ -7553,7 +7590,8 @@ MYSQL_BIN_LOG::queue_for_group_commit(group_commit_entry *orig_entry)
             wfc->wakeup_error= ER_QUERY_INTERRUPTED;
           my_message(wfc->wakeup_error,
                      ER_THD(orig_entry->thd, wfc->wakeup_error), MYF(0));
-          DBUG_RETURN(-1);
+          result= -1;
+          goto end;
         }
       }
       orig_entry->thd->EXIT_COND(&old_stage);
@@ -7567,12 +7605,13 @@ MYSQL_BIN_LOG::queue_for_group_commit(group_commit_entry *orig_entry)
     then there is nothing else to do.
   */
   if (orig_entry->queued_by_other)
-    DBUG_RETURN(0);
+    goto end;
 
   if (wfc && wfc->wakeup_error)
   {
     my_error(ER_PRIOR_COMMIT_FAILED, MYF(0));
-    DBUG_RETURN(-1);
+    result= -1;
+    goto end;
   }
 
   /* Now enqueue ourselves in the group commit queue. */
@@ -7733,7 +7772,13 @@ MYSQL_BIN_LOG::queue_for_group_commit(group_commit_entry *orig_entry)
 
   DBUG_PRINT("info", ("Queued for group commit as %s",
                       (orig_queue == NULL) ? "leader" : "participant"));
-  DBUG_RETURN(orig_queue == NULL);
+  result= orig_queue == NULL;
+
+end:
+  if (backup_lock_released)
+    thd->mdl_context.acquire_lock(thd->backup_commit_lock,
+                                  thd->variables.lock_wait_timeout);
+  DBUG_RETURN(result);
 }
 
 bool

sql/sql_class.cc

@@ -1271,6 +1271,7 @@ void THD::init()
   first_successful_insert_id_in_prev_stmt_for_binlog= 0;
   first_successful_insert_id_in_cur_stmt= 0;
   current_backup_stage= BACKUP_FINISHED;
+  backup_commit_lock= 0;
 #ifdef WITH_WSREP
   wsrep_last_query_id= 0;
   wsrep_xid.null();
@@ -1383,7 +1384,7 @@ void THD::init_for_queries()
   set_time(); 
   /*
     We don't need to call ha_enable_transaction() as we can't have
-    any active transactions that has to be commited
+    any active transactions that has to be committed
   */
   transaction.on= TRUE;
 
@@ -7386,16 +7387,33 @@ wait_for_commit::register_wait_for_prior_commit(wait_for_commit *waitee)
 }
 
 
-/*
+/**
-  Wait for commit of another transaction to complete, as already registered
+  Waits for commit of another transaction to complete, as already registered
   with register_wait_for_prior_commit(). If the commit already completed,
   returns immediately.
+
+  If thd->backup_commit_lock is set, release it while waiting for other threads
 */
+
 int
 wait_for_commit::wait_for_prior_commit2(THD *thd)
 {
   PSI_stage_info old_stage;
   wait_for_commit *loc_waitee;
+  bool backup_lock_released= 0;
+
+  /*
+    Release MDL_BACKUP_COMMIT LOCK while waiting for other threads to commit
+    This is needed to avoid deadlock between the other threads (which not
+    yet have the MDL_BACKUP_COMMIT_LOCK) and any threads using
+    BACKUP LOCK BLOCK_COMMIT.
+  */
+  if (thd->backup_commit_lock && thd->backup_commit_lock->ticket)
+  {
+    backup_lock_released= 1;
+    thd->mdl_context.release_lock(thd->backup_commit_lock->ticket);
+    thd->backup_commit_lock->ticket= 0;
+  }
 
   mysql_mutex_lock(&LOCK_wait_commit);
   DEBUG_SYNC(thd, "wait_for_prior_commit_waiting");
@@ -7445,10 +7463,16 @@ wait_for_commit::wait_for_prior_commit2(THD *thd)
     use within enter_cond/exit_cond.
   */
   DEBUG_SYNC(thd, "wait_for_prior_commit_killed");
+  if (backup_lock_released)
+    thd->mdl_context.acquire_lock(thd->backup_commit_lock,
+                                  thd->variables.lock_wait_timeout);
   return wakeup_error;
 
 end:
   thd->EXIT_COND(&old_stage);
+  if (backup_lock_released)
+    thd->mdl_context.acquire_lock(thd->backup_commit_lock,
+                                  thd->variables.lock_wait_timeout);
   return wakeup_error;
 }
 

sql/sql_class.h

@@ -2200,7 +2200,10 @@ public:
     rpl_io_thread_info *rpl_io_info;
     rpl_sql_thread_info *rpl_sql_info;
   } system_thread_info;
+  /* Used for BACKUP LOCK */
   MDL_ticket *mdl_backup_ticket, *mdl_backup_lock;
+  /* Used to register that thread has a MDL_BACKUP_WAIT_COMMIT lock */
+  MDL_request *backup_commit_lock;
 
   void reset_for_next_command(bool do_clear_errors= 1);
   /*

sql/xa.cc

@@ -582,7 +582,7 @@ bool trans_xa_commit(THD *thd)
     {
       DEBUG_SYNC(thd, "trans_xa_commit_after_acquire_commit_lock");
 
-      res= MY_TEST(ha_commit_one_phase(thd, 1, 1));
+      res= MY_TEST(ha_commit_one_phase(thd, 1));
       if (res)
         my_error(ER_XAER_RMERR, MYF(0));
     }