Another fix to relmapper race condition.

hlinnaka · hlinnaka · commit eb3bd243a245 · 2021-06-24T11:19:37.000+03:00
In previous commit, I missed that relmap_redo() was also not acquiring the RelationMappingLock. Thanks to Thomas Munro for pointing that out. Backpatch-through: 9.6, like previous commit. Discussion: https://www.postgresql.org/message-id/CA%2BhUKGLev%3DPpOSaL3WRZgOvgk217et%2BbxeJcRr4eR-NttP1F6Q%40mail.gmail.com
diff --git a/src/backend/utils/cache/relmapper.c b/src/backend/utils/cache/relmapper.c
@@ -949,12 +949,13 @@ relmap_redo(XLogReaderState *record)
 		 * preserve files, either.
 		 *
 		 * There shouldn't be anyone else updating relmaps during WAL replay,
-		 * so we don't bother to take the RelationMappingLock.  We would need
-		 * to do so if load_relmap_file needed to interlock against writers.
+		 * but grab the lock to interlock against load_relmap_file().
 		 */
+		LWLockAcquire(RelationMappingLock, LW_EXCLUSIVE);
 		write_relmap_file((xlrec->dbid == InvalidOid), &newmap,
 						  false, true, false,
 						  xlrec->dbid, xlrec->tsid, dbpath);
+		LWLockRelease(RelationMappingLock);
 
 		pfree(dbpath);
 	}
