RelationTruncate() must set DELAY_CHKPT_START.

macdice · macdice · commit 7d0b91a28421 · 2024-12-03T10:12:56.000+13:00
Previously, it set only DELAY_CHKPT_COMPLETE. That was important, because it meant that if the XLOG_SMGR_TRUNCATE record preceded a XLOG_CHECKPOINT_ONLINE record in the WAL, then the truncation would also happen on disk before the XLOG_CHECKPOINT_ONLINE record was written. However, it didn't guarantee that the sync request for the truncation was processed before the XLOG_CHECKPOINT_ONLINE record was written. By setting DELAY_CHKPT_START, we guarantee that if an XLOG_SMGR_TRUNCATE record is written to WAL before the redo pointer of a concurrent checkpoint, the sync request queued by that operation must be processed by that checkpoint, rather than being left for the following one. This is a refinement of commit 412ad7a. Back-patch to all supported releases, like that commit. Author: Robert Haas <robertmhaas@gmail.com> Reported-by: Thomas Munro <thomas.munro@gmail.com> Discussion: https://postgr.es/m/CA%2BhUKG%2B-2rjGZC2kwqr2NMLBcEBp4uf59QT1advbWYF_uc%2B0Aw%40mail.gmail.com
diff --git a/src/backend/catalog/storage.c b/src/backend/catalog/storage.c
@@ -326,20 +326,35 @@ RelationTruncate(Relation rel, BlockNumber nblocks)
 	RelationPreTruncate(rel);
 
 	/*
-	 * Make sure that a concurrent checkpoint can't complete while truncation
-	 * is in progress.
+	 * The code which follows can interact with concurrent checkpoints in two
+	 * separate ways.
 	 *
-	 * The truncation operation might drop buffers that the checkpoint
-	 * otherwise would have flushed. If it does, then it's essential that
-	 * the files actually get truncated on disk before the checkpoint record
-	 * is written. Otherwise, if reply begins from that checkpoint, the
+	 * First, the truncation operation might drop buffers that the checkpoint
+	 * otherwise would have flushed. If it does, then it's essential that the
+	 * files actually get truncated on disk before the checkpoint record is
+	 * written. Otherwise, if reply begins from that checkpoint, the
 	 * to-be-truncated blocks might still exist on disk but have older
-	 * contents than expected, which can cause replay to fail. It's OK for
-	 * the blocks to not exist on disk at all, but not for them to have the
-	 * wrong contents.
+	 * contents than expected, which can cause replay to fail. It's OK for the
+	 * blocks to not exist on disk at all, but not for them to have the wrong
+	 * contents. For this reason, we need to set DELAY_CHKPT_COMPLETE while
+	 * this code executes.
+	 *
+	 * Second, the call to smgrtruncate() below will in turn call
+	 * RegisterSyncRequest(). We need the sync request created by that call to
+	 * be processed before the checkpoint completes. CheckPointGuts() will
+	 * call ProcessSyncRequests(), but if we register our sync request after
+	 * that happens, then the WAL record for the truncation could end up
+	 * preceding the checkpoint record, while the actual sync doesn't happen
+	 * until the next checkpoint. To prevent that, we need to set
+	 * DELAY_CHKPT_START here. That way, if the XLOG_SMGR_TRUNCATE precedes
+	 * the redo pointer of a concurrent checkpoint, we're guaranteed that the
+	 * corresponding sync request will be processed before the checkpoint
+	 * completes.
 	 */
+	Assert(!MyProc->delayChkpt);
+	MyProc->delayChkpt = true;			/* DELAY_CHKPT_START */
 	Assert(!MyProc->delayChkptEnd);
-	MyProc->delayChkptEnd = true;
+	MyProc->delayChkptEnd = true;		/* DELAY_CHKPT_COMPLETE */
 
 	/*
 	 * We WAL-log the truncation before actually truncating, which means
@@ -387,6 +402,7 @@ RelationTruncate(Relation rel, BlockNumber nblocks)
 	smgrtruncate(RelationGetSmgr(rel), forks, nforks, blocks);
 
 	/* We've done all the critical work, so checkpoints are OK now. */
+	MyProc->delayChkpt = false;
 	MyProc->delayChkptEnd = false;
 
 	/*
