Back-patch b1ffe3f into REL_13_STABLE.

macdice · macdice · commit a722252ef1e1 · 2025-01-09T10:14:11.000+13:00
This is a cherry pick of 4c8e00a from the 14 branch into the 13 branch. It avoids an assertion failure when ForwardSyncRequest() tries to allocate memory while trying to compact the queue, if run in a critical section. RelationTruncate() gained a critical section in 38c579b, and could fail in that way in the 13 branch. b1ffe3f originally fixed the same problem with TruncateMultiXact(), but for that case it only needed to go back as far as 14, where SLRUs started using the sync request queue. It also fixed a related deadlock risk that doesn't apply in this case (this case doesn't wait), but it might exist in theory and it doesn't hurt to keep the code the same as later branches. Author: Heikki Linnakangas <heikki.linnakangas@iki.fi> (original commit) Reviewed-by: Michael Paquier <michael@paquier.xyz> (in this new context) Reported-by: Yura Sokolov <y.sokolov@postgrespro.ru> Discussion: https://postgr.es/m/f98aaa79-80b5-47c9-832a-31416a3a528b%40postgrespro.ru
diff --git a/src/backend/access/transam/xlog.c b/src/backend/access/transam/xlog.c
@@ -9139,6 +9139,12 @@ CreateCheckPoint(int flags)
 	{
 		do
 		{
+			/*
+			 * Keep absorbing fsync requests while we wait. There could even
+			 * be a deadlock if we don't, if the process that prevents the
+			 * checkpoint is trying to add a request to the queue.
+			 */
+			AbsorbSyncRequests();
 			pg_usleep(10000L);	/* wait for 10 msec */
 		} while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids));
 	}
@@ -9151,6 +9157,7 @@ CreateCheckPoint(int flags)
 	{
 		do
 		{
+			AbsorbSyncRequests();
 			pg_usleep(10000L);	/* wait for 10 msec */
 		} while (HaveVirtualXIDsDelayingChkptEnd(vxids, nvxids));
 	}
diff --git a/src/backend/postmaster/checkpointer.c b/src/backend/postmaster/checkpointer.c
@@ -1144,6 +1144,10 @@ CompactCheckpointerRequestQueue(void)
 	/* must hold CheckpointerCommLock in exclusive mode */
 	Assert(LWLockHeldByMe(CheckpointerCommLock));
 
+	/* Avoid memory allocations in a critical section. */
+	if (CritSectionCount > 0)
+		return false;
+
 	/* Initialize skip_slot array */
 	skip_slot = palloc0(sizeof(bool) * CheckpointerShmem->num_requests);
 

Original file line number	Diff line number	Diff line change
`@@ -9139,6 +9139,12 @@ CreateCheckPoint(int flags)`
`9139`	`9139`	`{`
`9140`	`9140`	`do`
`9141`	`9141`	`{`
	`9142`	`+ /*`
	`9143`	`+ * Keep absorbing fsync requests while we wait. There could even`
	`9144`	`+ * be a deadlock if we don't, if the process that prevents the`
	`9145`	`+ * checkpoint is trying to add a request to the queue.`
	`9146`	`+ */`
	`9147`	`+ AbsorbSyncRequests();`
`9142`	`9148`	`pg_usleep(10000L); /* wait for 10 msec */`
`9143`	`9149`	`} while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids));`
`9144`	`9150`	`}`
`@@ -9151,6 +9157,7 @@ CreateCheckPoint(int flags)`
`9151`	`9157`	`{`
`9152`	`9158`	`do`
`9153`	`9159`	`{`
	`9160`	`+ AbsorbSyncRequests();`
`9154`	`9161`	`pg_usleep(10000L); /* wait for 10 msec */`
`9155`	`9162`	`} while (HaveVirtualXIDsDelayingChkptEnd(vxids, nvxids));`
`9156`	`9163`	`}`