MDEV-14756 - Remove trx_sys_t::rw_trx_list

Let lock_validate_table_locks(), lock_rec_other_trx_holds_expl(),
lock_table_locks_lookup(), trx_recover_for_mysql(), trx_get_trx_by_xid(),
trx_roll_must_shutdown(), fetch_data_into_cache() iterate rw_trx_hash
instead of rw_trx_list.

storage/innobase/include/trx0sys.h

@@ -757,7 +757,20 @@ class rw_trx_hash_t
     @param argument    opque argument passed to action
 
     May return the same element multiple times if hash is under contention.
-    Elements can be added or removed while this method is being executed.
+    If caller doesn't like to see the same transaction multiple times, it has
+    to call iterate_no_dups() instead.
+
+    May return element with committed transaction. If caller doesn't like to
+    see committed transactions, it has to skip those under element mutex:
+
+      mutex_enter(&element->mutex);
+      if (trx_t trx= element->trx)
+      {
+        // trx is protected against commit in this branch
+      }
+      mutex_exit(&element->mutex);
+
+    May miss concurrently inserted transactions.
 
     @return
       @retval 0 iteration completed successfully

storage/innobase/include/trx0trx.h

@@ -269,7 +269,7 @@ int
 trx_recover_for_mysql(
 /*==================*/
 	XID*	xid_list,	/*!< in/out: prepared transactions */
-	ulint	len);		/*!< in: number of slots in xid_list */
+	uint	len);		/*!< in: number of slots in xid_list */
 /*******************************************************************//**
 This function is used to find one X/Open XA distributed transaction
 which is in the prepared state

storage/innobase/lock/lock0lock.cc

@@ -1533,61 +1533,6 @@ lock_sec_rec_some_has_impl(
 	return(trx);
 }
 
-#ifdef UNIV_DEBUG
-/*********************************************************************//**
-Checks if some transaction, other than given trx_id, has an explicit
-lock on the given rec, in the given precise_mode.
-
-FIXME: if the current transaction holds implicit lock from INSERT, a
-subsequent locking read should not convert it to explicit. See also
-MDEV-11215.
-
-@return	the transaction, whose id is not equal to trx_id, that has an
-explicit lock on the given rec, in the given precise_mode or NULL.*/
-static
-trx_t*
-lock_rec_other_trx_holds_expl(
-/*==========================*/
-	ulint			precise_mode,	/*!< in: LOCK_S or LOCK_X
-						possibly ORed to LOCK_GAP or
-						LOCK_REC_NOT_GAP. */
-	trx_t*			trx,		/*!< in: trx holding implicit
-						lock on rec */
-	const rec_t*		rec,		/*!< in: user record */
-	const buf_block_t*	block)		/*!< in: buffer block
-						containing the record */
-{
-	ut_ad(!page_rec_is_default_row(rec));
-
-	trx_t* holds = NULL;
-	ulint heap_no = page_rec_get_heap_no(rec);
-
-	lock_mutex_enter();
-	mutex_enter(&trx_sys->mutex);
-
-	for (trx_t* t = UT_LIST_GET_FIRST(trx_sys->rw_trx_list);
-	     t != NULL;
-	     t = UT_LIST_GET_NEXT(trx_list, t)) {
-
-		lock_t* expl_lock = lock_rec_has_expl(
-			precise_mode, block, heap_no, t);
-
-		if (expl_lock && expl_lock->trx != trx) {
-			/* An explicit lock is held by trx other than
-			the trx holding the implicit lock. */
-			holds = expl_lock->trx;
-			break;
-		}
-	}
-
-	mutex_exit(&trx_sys->mutex);
-
-	lock_mutex_exit();
-
-	return(holds);
-}
-#endif /* UNIV_DEBUG */
-
 /*********************************************************************//**
 Return approximate number or record locks (bits set in the bitmap) for
 this transaction. Since delete-marked records may be removed, the
@@ -6464,45 +6409,6 @@ lock_rec_validate_page(
 	return(TRUE);
 }
 
-/*********************************************************************//**
-Validates the table locks.
-@return TRUE if ok */
-static
-ibool
-lock_validate_table_locks(
-/*======================*/
-	const trx_ut_list_t*	trx_list)	/*!< in: trx list */
-{
-	const trx_t*	trx;
-
-	ut_ad(lock_mutex_own());
-	ut_ad(trx_sys_mutex_own());
-
-	ut_ad(trx_list == &trx_sys->rw_trx_list);
-
-	for (trx = UT_LIST_GET_FIRST(*trx_list);
-	     trx != NULL;
-	     trx = UT_LIST_GET_NEXT(trx_list, trx)) {
-
-		const lock_t*	lock;
-
-		check_trx_state(trx);
-
-		for (lock = UT_LIST_GET_FIRST(trx->lock.trx_locks);
-		     lock != NULL;
-		     lock = UT_LIST_GET_NEXT(trx_locks, lock)) {
-
-			if (lock_get_type_low(lock) & LOCK_TABLE) {
-
-				lock_table_queue_validate(
-					lock->un_member.tab_lock.table);
-			}
-		}
-	}
-
-	return(TRUE);
-}
-
 /*********************************************************************//**
 Validate record locks up to a limit.
 @return lock at limit or NULL if no more locks in the hash bucket */
@@ -6590,6 +6496,29 @@ lock_rec_block_validate(
 	}
 }
 
+
+static my_bool lock_validate_table_locks(rw_trx_hash_element_t *element,
+                                         void *arg)
+{
+  ut_ad(lock_mutex_own());
+  ut_ad(trx_sys_mutex_own()); /* Do we really need this. */
+  mutex_enter(&element->mutex);
+  if (element->trx)
+  {
+    check_trx_state(element->trx);
+    for (const lock_t *lock= UT_LIST_GET_FIRST(element->trx->lock.trx_locks);
+         lock != NULL;
+         lock= UT_LIST_GET_NEXT(trx_locks, lock))
+    {
+      if (lock_get_type_low(lock) & LOCK_TABLE)
+        lock_table_queue_validate(lock->un_member.tab_lock.table);
+    }
+  }
+  mutex_exit(&element->mutex);
+  return 0;
+}
+
+
 /*********************************************************************//**
 Validates the lock system.
 @return TRUE if ok */
@@ -6609,7 +6538,9 @@ lock_validate()
 	lock_mutex_enter();
 	mutex_enter(&trx_sys->mutex);
 
-	ut_a(lock_validate_table_locks(&trx_sys->rw_trx_list));
+	/* Validate table locks */
+	trx_sys->rw_trx_hash.iterate(reinterpret_cast<my_hash_walk_action>
+				     (lock_validate_table_locks), 0);
 
 	/* Iterate over all the record locks and validate the locks. We
 	don't want to hog the lock_sys_t::mutex and the trx_sys_t::mutex.
@@ -6838,6 +6769,70 @@ lock_rec_convert_impl_to_expl_for_trx(
 	DEBUG_SYNC_C("after_lock_rec_convert_impl_to_expl_for_trx");
 }
 
+
+#ifdef UNIV_DEBUG
+struct lock_rec_other_trx_holds_expl_arg
+{
+  const ulint heap_no;
+  const buf_block_t * const block;
+  const trx_t *impl_trx;
+};
+
+
+static my_bool lock_rec_other_trx_holds_expl_callback(
+  rw_trx_hash_element_t *element,
+  lock_rec_other_trx_holds_expl_arg *arg)
+{
+  mutex_enter(&element->mutex);
+  if (element->trx)
+  {
+    lock_t *expl_lock= lock_rec_has_expl(LOCK_S | LOCK_REC_NOT_GAP, arg->block,
+                                         arg->heap_no, element->trx);
+    /*
+      An explicit lock is held by trx other than the trx holding the implicit
+      lock.
+    */
+    ut_ad(!expl_lock || expl_lock->trx == arg->impl_trx);
+  }
+  mutex_exit(&element->mutex);
+  return 0;
+}
+
+
+/**
+  Checks if some transaction, other than given trx_id, has an explicit
+  lock on the given rec.
+
+  FIXME: if the current transaction holds implicit lock from INSERT, a
+  subsequent locking read should not convert it to explicit. See also
+  MDEV-11215.
+
+  @param      caller_trx  trx of current thread
+  @param[in]  trx         trx holding implicit lock on rec
+  @param[in]  rec         user record
+  @param[in]  block       buffer block containing the record
+*/
+
+static void lock_rec_other_trx_holds_expl(trx_t *caller_trx, trx_t *trx,
+                                          const rec_t *rec,
+                                          const buf_block_t *block)
+{
+  if (trx)
+  {
+    ut_ad(!page_rec_is_default_row(rec));
+    lock_mutex_enter();
+    lock_rec_other_trx_holds_expl_arg arg= { page_rec_get_heap_no(rec), block,
+                                             trx };
+    trx_sys->rw_trx_hash.iterate(caller_trx,
+                                 reinterpret_cast<my_hash_walk_action>
+                                 (lock_rec_other_trx_holds_expl_callback),
+                                 &arg);
+    lock_mutex_exit();
+  }
+}
+#endif /* UNIV_DEBUG */
+
+
 /*********************************************************************//**
 If a transaction has an implicit x-lock on a record, but no explicit x-lock
 set on the record, sets one for it. */
@@ -6872,8 +6867,8 @@ lock_rec_convert_impl_to_expl(
 		trx = lock_sec_rec_some_has_impl(caller_trx, rec, index,
 						 offsets);
 
-		ut_ad(!trx || !lock_rec_other_trx_holds_expl(
-				LOCK_S | LOCK_REC_NOT_GAP, trx, rec, block));
+		ut_d(lock_rec_other_trx_holds_expl(caller_trx, trx, rec,
+						   block));
 	}
 
 	if (trx != 0) {
@@ -7771,52 +7766,38 @@ lock_table_get_n_locks(
 }
 
 #ifdef UNIV_DEBUG
-/*******************************************************************//**
-Do an exhaustive check for any locks (table or rec) against the table.
-@return lock if found */
-static
-const lock_t*
-lock_table_locks_lookup(
-/*====================*/
-	const dict_table_t*	table,		/*!< in: check if there are
-						any locks held on records in
-						this table or on the table
-						itself */
-	const trx_ut_list_t*	trx_list)	/*!< in: trx list to check */
-{
-	trx_t*			trx;
-
-	ut_a(table != NULL);
-	ut_ad(lock_mutex_own());
-	ut_ad(trx_sys_mutex_own());
-
-	for (trx = UT_LIST_GET_FIRST(*trx_list);
-	     trx != NULL;
-	     trx = UT_LIST_GET_NEXT(trx_list, trx)) {
-
-		const lock_t*	lock;
-
-		check_trx_state(trx);
-
-		for (lock = UT_LIST_GET_FIRST(trx->lock.trx_locks);
-		     lock != NULL;
-		     lock = UT_LIST_GET_NEXT(trx_locks, lock)) {
-
-			ut_a(lock->trx == trx);
-
-			if (lock_get_type_low(lock) == LOCK_REC) {
-				ut_ad(!dict_index_is_online_ddl(lock->index)
-				      || dict_index_is_clust(lock->index));
-				if (lock->index->table == table) {
-					return(lock);
-				}
-			} else if (lock->un_member.tab_lock.table == table) {
-				return(lock);
-			}
-		}
-	}
-
-	return(NULL);
+/**
+  Do an exhaustive check for any locks (table or rec) against the table.
+
+  @param[in]  table  check if there are any locks held on records in this table
+                     or on the table itself
+*/
+
+static my_bool lock_table_locks_lookup(rw_trx_hash_element_t *element,
+                                       const dict_table_t *table)
+{
+  ut_ad(lock_mutex_own());
+  mutex_enter(&element->mutex);
+  if (element->trx)
+  {
+    check_trx_state(element->trx);
+    for (const lock_t *lock= UT_LIST_GET_FIRST(element->trx->lock.trx_locks);
+         lock != NULL;
+         lock= UT_LIST_GET_NEXT(trx_locks, lock))
+    {
+      ut_ad(lock->trx == element->trx);
+      if (lock_get_type_low(lock) == LOCK_REC)
+      {
+        ut_ad(!dict_index_is_online_ddl(lock->index) ||
+              dict_index_is_clust(lock->index));
+        ut_ad(lock->index->table != table);
+      }
+      else
+        ut_ad(lock->un_member.tab_lock.table != table);
+    }
+  }
+  mutex_exit(&element->mutex);
+  return 0;
 }
 #endif /* UNIV_DEBUG */
 
@@ -7832,17 +7813,17 @@ lock_table_has_locks(
 {
 	ibool			has_locks;
 
+	ut_ad(table != NULL);
 	lock_mutex_enter();
 
 	has_locks = UT_LIST_GET_LEN(table->locks) > 0 || table->n_rec_locks > 0;
 
 #ifdef UNIV_DEBUG
 	if (!has_locks) {
-		mutex_enter(&trx_sys->mutex);
-
-		ut_ad(!lock_table_locks_lookup(table, &trx_sys->rw_trx_list));
-
-		mutex_exit(&trx_sys->mutex);
+		trx_sys->rw_trx_hash.iterate(
+			reinterpret_cast<my_hash_walk_action>
+			(lock_table_locks_lookup),
+			const_cast<dict_table_t*>(table));
 	}
 #endif /* UNIV_DEBUG */
 

storage/innobase/trx/trx0i_s.cc

@@ -1245,102 +1245,90 @@ trx_i_s_cache_clear(
 	ha_storage_empty(&cache->storage);
 }
 
-/*******************************************************************//**
-Fetches the data needed to fill the 3 INFORMATION SCHEMA tables into the
-table cache buffer. Cache must be locked for write. */
-static
-void
-fetch_data_into_cache_low(
-/*======================*/
-	trx_i_s_cache_t*	cache,		/*!< in/out: cache */
-	bool			read_write,	/*!< in: only read-write
-						transactions */
-	trx_ut_list_t*		trx_list)	/*!< in: trx list */
-{
-	const trx_t*		trx;
-	bool			rw_trx_list = trx_list == &trx_sys->rw_trx_list;
-
-	ut_ad(rw_trx_list || trx_list == &trx_sys->mysql_trx_list);
-
-	/* Iterate over the transaction list and add each one
-	to innodb_trx's cache. We also add all locks that are relevant
-	to each transaction into innodb_locks' and innodb_lock_waits'
-	caches. */
-
-	for (trx = UT_LIST_GET_FIRST(*trx_list);
-	     trx != NULL;
-	     trx =
-	     (rw_trx_list
-	      ? UT_LIST_GET_NEXT(trx_list, trx)
-	      : UT_LIST_GET_NEXT(mysql_trx_list, trx))) {
-
-		i_s_trx_row_t*		trx_row;
-		i_s_locks_row_t*	requested_lock_row;
-
-		/* Note: Read only transactions that modify temporary
-		tables an have a transaction ID */
-		if (!trx_is_started(trx)
-		    || (!rw_trx_list && trx->id != 0 && !trx->read_only)) {
 
-			continue;
-		}
-
-		assert_trx_nonlocking_or_in_list(trx);
-
-		ut_ad(trx->in_rw_trx_list == rw_trx_list);
-
-		if (!add_trx_relevant_locks_to_cache(cache, trx,
-						     &requested_lock_row)) {
-
-			cache->is_truncated = TRUE;
-			return;
-		}
-
-		trx_row = reinterpret_cast<i_s_trx_row_t*>(
-			table_cache_create_empty_row(
-				&cache->innodb_trx, cache));
-
-		/* memory could not be allocated */
-		if (trx_row == NULL) {
-
-			cache->is_truncated = TRUE;
-			return;
-		}
+/**
+  Add transactions to innodb_trx's cache.
 
-		if (!fill_trx_row(trx_row, trx, requested_lock_row, cache)) {
+  We also add all locks that are relevant to each transaction into
+  innodb_locks' and innodb_lock_waits' caches.
+*/
 
-			/* memory could not be allocated */
-			--cache->innodb_trx.rows_used;
-			cache->is_truncated = TRUE;
-			return;
-		}
-	}
+static void fetch_data_into_cache_low(trx_i_s_cache_t *cache, const trx_t *trx)
+{
+  i_s_locks_row_t *requested_lock_row;
+
+  assert_trx_nonlocking_or_in_list(trx);
+
+  if (add_trx_relevant_locks_to_cache(cache, trx, &requested_lock_row))
+  {
+    if (i_s_trx_row_t *trx_row= reinterpret_cast<i_s_trx_row_t*>(
+        table_cache_create_empty_row(&cache->innodb_trx, cache)))
+    {
+      if (fill_trx_row(trx_row, trx, requested_lock_row, cache))
+        return;
+      --cache->innodb_trx.rows_used;
+    }
+  }
+
+  /* memory could not be allocated */
+  cache->is_truncated= TRUE;
 }
 
-/*******************************************************************//**
-Fetches the data needed to fill the 3 INFORMATION SCHEMA tables into the
-table cache buffer. Cache must be locked for write. */
-static
-void
-fetch_data_into_cache(
-/*==================*/
-	trx_i_s_cache_t*	cache)	/*!< in/out: cache */
-{
-	ut_ad(lock_mutex_own());
-	ut_ad(trx_sys_mutex_own());
 
-	trx_i_s_cache_clear(cache);
+static my_bool fetch_data_into_cache_callback(
+  rw_trx_hash_element_t *element, trx_i_s_cache_t *cache)
+{
+  mutex_enter(&element->mutex);
+  if (element->trx)
+  {
+    assert_trx_in_rw_list(element->trx);
+    fetch_data_into_cache_low(cache, element->trx);
+  }
+  mutex_exit(&element->mutex);
+  return cache->is_truncated;
+}
 
-	/* Capture the state of the read-write transactions. This includes
-	internal transactions too. They are not on mysql_trx_list */
-	fetch_data_into_cache_low(cache, true, &trx_sys->rw_trx_list);
 
-	/* Capture the state of the read-only active transactions */
-	fetch_data_into_cache_low(cache, false, &trx_sys->mysql_trx_list);
+/**
+  Fetches the data needed to fill the 3 INFORMATION SCHEMA tables into the
+  table cache buffer. Cache must be locked for write.
+*/
 
-	cache->is_truncated = FALSE;
+static void fetch_data_into_cache(trx_i_s_cache_t *cache)
+{
+  ut_ad(lock_mutex_own());
+  trx_i_s_cache_clear(cache);
+
+  /*
+    Capture the state of the read-write transactions. This includes
+    internal transactions too. They are not on mysql_trx_list
+  */
+  trx_sys->rw_trx_hash.iterate_no_dups(reinterpret_cast<my_hash_walk_action>
+                                       (fetch_data_into_cache_callback),
+                                       cache);
+
+  /* Capture the state of the read-only active transactions */
+  trx_sys_mutex_enter();
+  for (const trx_t *trx= UT_LIST_GET_FIRST(trx_sys->mysql_trx_list);
+       trx != NULL;
+       trx= UT_LIST_GET_NEXT(mysql_trx_list, trx))
+  {
+    /*
+      Skip transactions that have trx->id > 0: they were added in previous
+      iteration. Although we may miss concurrently started transactions.
+    */
+    if (trx_is_started(trx) && trx->id == 0)
+    {
+      fetch_data_into_cache_low(cache, trx);
+      if (cache->is_truncated)
+        break;
+     }
+  }
+  trx_sys_mutex_exit();
+  cache->is_truncated= FALSE;
 }
 
+
 /*******************************************************************//**
 Update the transactions cache if it has not been read for some time.
 Called from handler/i_s.cc.
@@ -1358,13 +1346,7 @@ trx_i_s_possibly_fetch_data_into_cache(
 	/* We need to read trx_sys and record/table lock queues */
 
 	lock_mutex_enter();
-
-	trx_sys_mutex_enter();
-
 	fetch_data_into_cache(cache);
-
-	trx_sys_mutex_exit();
-
 	lock_mutex_exit();
 
 	/* update cache last read time */

storage/innobase/trx/trx0roll.cc

@@ -791,6 +791,33 @@ trx_rollback_resurrected(
 	goto func_exit;
 }
 
+
+struct trx_roll_count_callback_arg
+{
+  uint32_t n_trx;
+  uint64_t n_rows;
+  trx_roll_count_callback_arg(): n_trx(0), n_rows(0) {}
+};
+
+
+static my_bool trx_roll_count_callback(rw_trx_hash_element_t *element,
+                                       trx_roll_count_callback_arg *arg)
+{
+  mutex_enter(&element->mutex);
+  if (trx_t *trx= element->trx)
+  {
+    assert_trx_in_rw_list(trx);
+    if (trx->is_recovered && trx_state_eq(trx, TRX_STATE_ACTIVE))
+    {
+      arg->n_trx++;
+      arg->n_rows+= trx->undo_no;
+    }
+  }
+  mutex_exit(&element->mutex);
+  return 0;
+}
+
+
 /** Report progress when rolling back a row of a recovered transaction.
 @return	whether the rollback should be aborted due to pending shutdown */
 bool
@@ -808,31 +835,24 @@ trx_roll_must_shutdown()
 	}
 
 	ib_time_t time = ut_time();
-	mutex_enter(&trx_sys->mutex);
 	mutex_enter(&recv_sys->mutex);
 
 	if (recv_sys->report(time)) {
-		ulint n_trx = 0;
-		ulonglong n_rows = 0;
-		for (const trx_t* t = UT_LIST_GET_FIRST(trx_sys->rw_trx_list);
-		     t != NULL;
-		     t = UT_LIST_GET_NEXT(trx_list, t)) {
-
-			assert_trx_in_rw_list(t);
-			if (t->is_recovered
-			    && trx_state_eq(t, TRX_STATE_ACTIVE)) {
-				n_trx++;
-				n_rows += t->undo_no;
-			}
-		}
-		ib::info() << "To roll back: " << n_trx << " transactions, "
-			   << n_rows << " rows";
-		sd_notifyf(0, "STATUS=To roll back: " ULINTPF " transactions, "
-			   "%llu rows", n_trx, n_rows);
+		trx_roll_count_callback_arg arg;
+
+		/* Get number of recovered active transactions and number of
+		rows they modified. Numbers must be accurate, because only this
+		thread is allowed to touch recovered transactions. */
+		trx_sys->rw_trx_hash.iterate_no_dups(
+			reinterpret_cast<my_hash_walk_action>
+			(trx_roll_count_callback), &arg);
+		ib::info() << "To roll back: " << arg.n_trx
+			   << " transactions, " << arg.n_rows << " rows";
+		sd_notifyf(0, "STATUS=To roll back: " UINT32PF " transactions,"
+			   " " UINT64PF " rows", arg.n_trx, arg.n_rows);
 	}
 
 	mutex_exit(&recv_sys->mutex);
-	mutex_exit(&trx_sys->mutex);
 	return false;
 }
 

storage/innobase/trx/trx0trx.cc

@@ -2668,141 +2668,121 @@ trx_prepare_for_mysql(trx_t* trx)
 	return(DB_SUCCESS);
 }
 
-/**********************************************************************//**
-This function is used to find number of prepared transactions and
-their transaction objects for a recovery.
-@return number of prepared transactions stored in xid_list */
-int
-trx_recover_for_mysql(
-/*==================*/
-	XID*	xid_list,	/*!< in/out: prepared transactions */
-	ulint	len)		/*!< in: number of slots in xid_list */
-{
-	const trx_t*	trx;
-	ulint		count = 0;
-
-	ut_ad(xid_list);
-	ut_ad(len);
-
-	/* We should set those transactions which are in the prepared state
-	to the xid_list */
-
-	trx_sys_mutex_enter();
-
-	for (trx = UT_LIST_GET_FIRST(trx_sys->rw_trx_list);
-	     trx != NULL;
-	     trx = UT_LIST_GET_NEXT(trx_list, trx)) {
 
-		assert_trx_in_rw_list(trx);
-
-		/* The state of a read-write transaction cannot change
-		from or to NOT_STARTED while we are holding the
-		trx_sys->mutex. It may change to PREPARED, but not if
-		trx->is_recovered. It may also change to COMMITTED. */
-		if (trx_state_eq(trx, TRX_STATE_PREPARED)) {
-			xid_list[count] = *trx->xid;
-
-			if (count == 0) {
-				ib::info() << "Starting recovery for"
-					" XA transactions...";
-			}
+struct trx_recover_for_mysql_callback_arg
+{
+  XID *xid_list;
+  uint len;
+  uint count;
+};
 
-			ib::info() << "Transaction "
-				<< trx_get_id_for_print(trx)
-				<< " in prepared state after recovery";
 
-			ib::info() << "Transaction contains changes to "
-				<< trx->undo_no << " rows";
+static my_bool trx_recover_for_mysql_callback(rw_trx_hash_element_t *element,
+  trx_recover_for_mysql_callback_arg *arg)
+{
+  mutex_enter(&element->mutex);
+  if (trx_t *trx= element->trx)
+  {
+    assert_trx_in_rw_list(element->trx);
 
-			count++;
+    /*
+      The state of a read-write transaction can only change from ACTIVE to
+      PREPARED while we are holding the element->mutex. But since it is
+      executed at startup no state change should occur.
+    */
+    if (trx_state_eq(trx, TRX_STATE_PREPARED))
+    {
+      ut_ad(trx->is_recovered);
+      if (arg->count == 0)
+        ib::info() << "Starting recovery for XA transactions...";
+      ib::info() << "Transaction " << trx_get_id_for_print(trx)
+                 << " in prepared state after recovery";
+      ib::info() << "Transaction contains changes to " << trx->undo_no
+                 << " rows";
+      arg->xid_list[arg->count++]= *trx->xid;
+    }
+  }
+  mutex_exit(&element->mutex);
+  return arg->count == arg->len;
+}
 
-			if (count == len) {
-				break;
-			}
-		}
-	}
 
-	trx_sys_mutex_exit();
+/**
+  Find prepared transaction objects for recovery.
 
-	if (count > 0){
-		ib::info() << count << " transactions in prepared state"
-			" after recovery";
-	}
+  @param[out]  xid_list  prepared transactions
+  @param[in]   len       number of slots in xid_list
 
-	return(int (count));
-}
+  @return number of prepared transactions stored in xid_list
+*/
 
-/*******************************************************************//**
-This function is used to find one X/Open XA distributed transaction
-which is in the prepared state
-@return trx on match, the trx->xid will be invalidated;
-note that the trx may have been committed, unless the caller is
-holding lock_sys->mutex */
-static MY_ATTRIBUTE((warn_unused_result))
-trx_t*
-trx_get_trx_by_xid_low(
-/*===================*/
-	XID*	xid)		/*!< in: X/Open XA transaction
-					identifier */
+int trx_recover_for_mysql(XID *xid_list, uint len)
 {
-	trx_t*		trx;
-
-	ut_ad(trx_sys_mutex_own());
-
-	for (trx = UT_LIST_GET_FIRST(trx_sys->rw_trx_list);
-	     trx != NULL;
-	     trx = UT_LIST_GET_NEXT(trx_list, trx)) {
+  trx_recover_for_mysql_callback_arg arg= { xid_list, len, 0 };
+
+  ut_ad(xid_list);
+  ut_ad(len);
+
+  /* Fill xid_list with PREPARED transactions. */
+  trx_sys->rw_trx_hash.iterate_no_dups(reinterpret_cast<my_hash_walk_action>
+                                       (trx_recover_for_mysql_callback), &arg);
+  if (arg.count)
+    ib::info() << arg.count
+               << " transactions in prepared state after recovery";
+  return(arg.count);
+}
 
-		assert_trx_in_rw_list(trx);
 
-		/* Compare two X/Open XA transaction id's: their
-		length should be the same and binary comparison
-		of gtrid_length+bqual_length bytes should be
-		the same */
+struct trx_get_trx_by_xid_callback_arg
+{
+  XID *xid;
+  trx_t *trx;
+};
 
-		if (trx->is_recovered
-		    && trx_state_eq(trx, TRX_STATE_PREPARED)
-			&& xid->eq((XID*)trx->xid)) {
 
-			/* Invalidate the XID, so that subsequent calls
-			will not find it. */
-			trx->xid->null();
-			break;
-		}
-	}
-
-	return(trx);
+static my_bool trx_get_trx_by_xid_callback(rw_trx_hash_element_t *element,
+  trx_get_trx_by_xid_callback_arg *arg)
+{
+  my_bool found= 0;
+  mutex_enter(&element->mutex);
+  if (trx_t *trx= element->trx)
+  {
+    assert_trx_in_rw_list(element->trx);
+    if (trx->is_recovered && trx_state_eq(trx, TRX_STATE_PREPARED) &&
+        arg->xid->eq(reinterpret_cast<XID*>(trx->xid)))
+    {
+      /* Invalidate the XID, so that subsequent calls will not find it. */
+      trx->xid->null();
+      arg->trx= trx;
+      found= 1;
+    }
+  }
+  mutex_exit(&element->mutex);
+  return found;
 }
 
-/*******************************************************************//**
-This function is used to find one X/Open XA distributed transaction
-which is in the prepared state
-@return trx or NULL; on match, the trx->xid will be invalidated;
-note that the trx may have been committed, unless the caller is
-holding lock_sys->mutex */
-trx_t*
-trx_get_trx_by_xid(
-/*===============*/
-	XID*	xid)	/*!< in: X/Open XA transaction identifier */
-{
-	trx_t*	trx;
 
-	if (xid == NULL) {
+/**
+  Finds PREPARED XA transaction by xid.
 
-		return(NULL);
-	}
+  trx may have been committed, unless the caller is holding lock_sys->mutex.
 
-	trx_sys_mutex_enter();
+  @param[in]  xid  X/Open XA transaction identifier
 
-	/* Recovered/Resurrected transactions are always only on the
-	trx_sys_t::rw_trx_list. */
-	trx = trx_get_trx_by_xid_low((XID*)xid);
+  @return trx or NULL; on match, the trx->xid will be invalidated;
+*/
 
-	trx_sys_mutex_exit();
+trx_t *trx_get_trx_by_xid(XID *xid)
+{
+  trx_get_trx_by_xid_callback_arg arg= { xid, 0 };
 
-	return(trx);
+  if (xid)
+    trx_sys->rw_trx_hash.iterate(reinterpret_cast<my_hash_walk_action>
+                                 (trx_get_trx_by_xid_callback), &arg);
+  return arg.trx;
 }
 
+
 /*************************************************************//**
 Starts the transaction if it is not yet started. */
 void