From 52039bba820d633bbb69615994d7bb78a512ae4e Mon Sep 17 00:00:00 2001 From: Ajin Cherian Date: Thu, 24 Jul 2025 07:54:27 -0400 Subject: [PATCH v4] Fix a deadlock during ALTER SUBSCRIPTION... DROP PUBLICATION When user drops a publication from a subscription, this will result in a publication refresh on the subscriber which will try and drop any pending origins. Meanwhile the apply worker could also be trying to cleanup origins. There could be a deadlock if the order of locking of SubscriptionRelationId, SubscriptionRelRelationId and ReplicationOriginRelationId are not aligned between functions process_syncing_tables_for_apply() and AlterSubscription_refresh(). The fix is to get locks on SubscriptionRelationId, SubscriptionRelRelationId and ReplicationOriginRelationId in that order when dropping tracking origins. --- src/backend/catalog/pg_subscription.c | 12 ++++++---- src/backend/replication/logical/tablesync.c | 36 +++++++++++++++++++++++++---- src/include/catalog/pg_subscription_rel.h | 2 +- 3 files changed, 40 insertions(+), 10 deletions(-) diff --git a/src/backend/catalog/pg_subscription.c b/src/backend/catalog/pg_subscription.c index 29fc421..cd799d3 100644 --- a/src/backend/catalog/pg_subscription.c +++ b/src/backend/catalog/pg_subscription.c @@ -282,7 +282,7 @@ AddSubscriptionRelState(Oid subid, Oid relid, char state, */ void UpdateSubscriptionRelState(Oid subid, Oid relid, char state, - XLogRecPtr sublsn) + XLogRecPtr sublsn, bool already_locked) { Relation rel; HeapTuple tup; @@ -290,9 +290,13 @@ UpdateSubscriptionRelState(Oid subid, Oid relid, char state, Datum values[Natts_pg_subscription_rel]; bool replaces[Natts_pg_subscription_rel]; - LockSharedObject(SubscriptionRelationId, subid, 0, AccessShareLock); - - rel = table_open(SubscriptionRelRelationId, RowExclusiveLock); + if (already_locked) + rel = table_open(SubscriptionRelRelationId, NoLock); + else + { + LockSharedObject(SubscriptionRelationId, subid, 0, AccessShareLock); + rel = table_open(SubscriptionRelRelationId, RowExclusiveLock); + } /* Try finding existing mapping. */ tup = SearchSysCacheCopy2(SUBSCRIPTIONRELMAP, diff --git a/src/backend/replication/logical/tablesync.c b/src/backend/replication/logical/tablesync.c index 26b71de..2471609 100644 --- a/src/backend/replication/logical/tablesync.c +++ b/src/backend/replication/logical/tablesync.c @@ -300,7 +300,8 @@ process_syncing_tables_for_sync(XLogRecPtr current_lsn) UpdateSubscriptionRelState(MyLogicalRepWorker->subid, MyLogicalRepWorker->relid, MyLogicalRepWorker->relstate, - MyLogicalRepWorker->relstate_lsn); + MyLogicalRepWorker->relstate_lsn, + false); /* * End streaming so that LogRepWorkerWalRcvConn can be used to drop @@ -366,6 +367,7 @@ process_syncing_tables_for_apply(XLogRecPtr current_lsn) static HTAB *last_start_times = NULL; ListCell *lc; bool started_tx = false; + Relation rel = NULL; Assert(!IsTransactionState()); @@ -463,7 +465,17 @@ process_syncing_tables_for_apply(XLogRecPtr current_lsn) * refresh for the subscription where we remove the table * state and its origin and by this time the origin might be * already removed. So passing missing_ok = true. + * + * Lock SubscriptionRelationId with AccessShareLock and + * take RowExclusiveLock on SubscriptionRelRelationId to + * keep the same locking order as refresh from an + * user issued DDL command to prevent any deadlocks. */ + LockSharedObject(SubscriptionRelationId, MyLogicalRepWorker->subid, + 0, AccessShareLock); + if (!rel) + rel = table_open(SubscriptionRelRelationId, RowExclusiveLock); + ReplicationOriginNameForTablesync(MyLogicalRepWorker->subid, rstate->relid, originname, @@ -475,7 +487,7 @@ process_syncing_tables_for_apply(XLogRecPtr current_lsn) */ UpdateSubscriptionRelState(MyLogicalRepWorker->subid, rstate->relid, rstate->state, - rstate->lsn); + rstate->lsn, true); } } else @@ -526,7 +538,14 @@ process_syncing_tables_for_apply(XLogRecPtr current_lsn) * This is required to avoid any undetected deadlocks * due to any existing lock as deadlock detector won't * be able to detect the waits on the latch. + * + * Also close any tables prior to the commit. */ + if (rel) + { + table_close(rel, NoLock); + rel = NULL; + } CommitTransactionCommand(); pgstat_report_stat(false); } @@ -586,6 +605,12 @@ process_syncing_tables_for_apply(XLogRecPtr current_lsn) } } + /* Close table if opened */ + if (rel) + { + table_close(rel, NoLock); + } + if (started_tx) { CommitTransactionCommand(); @@ -1041,7 +1066,8 @@ LogicalRepSyncTableStart(XLogRecPtr *origin_startpos) UpdateSubscriptionRelState(MyLogicalRepWorker->subid, MyLogicalRepWorker->relid, MyLogicalRepWorker->relstate, - MyLogicalRepWorker->relstate_lsn); + MyLogicalRepWorker->relstate_lsn, + false); CommitTransactionCommand(); pgstat_report_stat(false); @@ -1136,8 +1162,8 @@ LogicalRepSyncTableStart(XLogRecPtr *origin_startpos) UpdateSubscriptionRelState(MyLogicalRepWorker->subid, MyLogicalRepWorker->relid, SUBREL_STATE_FINISHEDCOPY, - MyLogicalRepWorker->relstate_lsn); - + MyLogicalRepWorker->relstate_lsn, + false); CommitTransactionCommand(); copy_table_done: diff --git a/src/include/catalog/pg_subscription_rel.h b/src/include/catalog/pg_subscription_rel.h index ed94f57..7baeda5 100644 --- a/src/include/catalog/pg_subscription_rel.h +++ b/src/include/catalog/pg_subscription_rel.h @@ -84,7 +84,7 @@ typedef struct SubscriptionRelState extern void AddSubscriptionRelState(Oid subid, Oid relid, char state, XLogRecPtr sublsn); extern void UpdateSubscriptionRelState(Oid subid, Oid relid, char state, - XLogRecPtr sublsn); + XLogRecPtr sublsn, bool already_locked); extern char GetSubscriptionRelState(Oid subid, Oid relid, XLogRecPtr *sublsn); extern void RemoveSubscriptionRel(Oid subid, Oid relid); -- 1.8.3.1