Fix possible recovery trouble if TRUNCATE overlaps a checkpoint.

If TRUNCATE causes some buffers to be invalidated and thus the
checkpoint does not flush them, TRUNCATE must also ensure that the
corresponding files are truncated on disk. Otherwise, a replay
from the checkpoint might find that the buffers exist but have
the wrong contents, which may cause replay to fail.

Report by Teja Mupparti. Patch by Kyotaro Horiguchi, per a design
suggestion from Heikki Linnakangas, with some changes to the
comments by me. Review of this and a prior patch that approached
the issue differently by Heikki Linnakangas, Andres Freund, Álvaro
Herrera, Masahiko Sawada, and Tom Lane.

Discussion: http://postgr.es/m/BYAPR06MB6373BF50B469CA393C614257ABF00@BYAPR06MB6373.namprd06.prod.outlook.com

Author: robertmhaas

src/backend/access/transam/multixact.c

@@ -3071,8 +3071,8 @@ TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB)
 	 * crash/basebackup, even though the state of the data directory would
 	 * require it.
 	 */
-	Assert(!MyProc->delayChkpt);
-	MyProc->delayChkpt = true;
+	Assert((MyProc->delayChkpt & DELAY_CHKPT_START) == 0);
+	MyProc->delayChkpt |= DELAY_CHKPT_START;
 
 	/* WAL log truncation */
 	WriteMTruncateXlogRec(newOldestMultiDB,
@@ -3098,7 +3098,7 @@ TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB)
 	/* Then offsets */
 	PerformOffsetsTruncation(oldestMulti, newOldestMulti);
 
-	MyProc->delayChkpt = false;
+	MyProc->delayChkpt &= ~DELAY_CHKPT_START;
 
 	END_CRIT_SECTION();
 	LWLockRelease(MultiXactTruncationLock);

src/backend/access/transam/twophase.c

@@ -476,7 +476,7 @@ MarkAsPreparingGuts(GlobalTransaction gxact, TransactionId xid, const char *gid,
 	}
 	pgxact->xid = xid;
 	pgxact->xmin = InvalidTransactionId;
-	proc->delayChkpt = false;
+	proc->delayChkpt = 0;
 	pgxact->vacuumFlags = 0;
 	proc->pid = 0;
 	proc->databaseId = databaseid;
@@ -1170,7 +1170,8 @@ EndPrepare(GlobalTransaction gxact)
 
 	START_CRIT_SECTION();
 
-	MyProc->delayChkpt = true;
+	Assert((MyProc->delayChkpt & DELAY_CHKPT_START) == 0);
+	MyProc->delayChkpt |= DELAY_CHKPT_START;
 
 	XLogBeginInsert();
 	for (record = records.head; record != NULL; record = record->next)
@@ -1213,7 +1214,7 @@ EndPrepare(GlobalTransaction gxact)
 	 * checkpoint starting after this will certainly see the gxact as a
 	 * candidate for fsyncing.
 	 */
-	MyProc->delayChkpt = false;
+	MyProc->delayChkpt &= ~DELAY_CHKPT_START;
 
 	/*
 	 * Remember that we have this GlobalTransaction entry locked for us.  If
@@ -2286,7 +2287,8 @@ RecordTransactionCommitPrepared(TransactionId xid,
 	START_CRIT_SECTION();
 
 	/* See notes in RecordTransactionCommit */
-	MyProc->delayChkpt = true;
+	Assert((MyProc->delayChkpt & DELAY_CHKPT_START) == 0);
+	MyProc->delayChkpt |= DELAY_CHKPT_START;
 
 	/*
 	 * Emit the XLOG commit record. Note that we mark 2PC commits as
@@ -2334,7 +2336,7 @@ RecordTransactionCommitPrepared(TransactionId xid,
 	TransactionIdCommitTree(xid, nchildren, children);
 
 	/* Checkpoint can proceed now */
-	MyProc->delayChkpt = false;
+	MyProc->delayChkpt &= ~DELAY_CHKPT_START;
 
 	END_CRIT_SECTION();
 

src/backend/access/transam/xact.c

@@ -1308,8 +1308,9 @@ RecordTransactionCommit(void)
 		 * This makes checkpoint's determination of which xacts are delayChkpt
 		 * a bit fuzzy, but it doesn't matter.
 		 */
+		Assert((MyProc->delayChkpt & DELAY_CHKPT_START) == 0);
 		START_CRIT_SECTION();
-		MyProc->delayChkpt = true;
+		MyProc->delayChkpt |= DELAY_CHKPT_START;
 
 		SetCurrentTransactionStopTimestamp();
 
@@ -1410,7 +1411,7 @@ RecordTransactionCommit(void)
 	 */
 	if (markXidCommitted)
 	{
-		MyProc->delayChkpt = false;
+		MyProc->delayChkpt &= ~DELAY_CHKPT_START;
 		END_CRIT_SECTION();
 	}
 

src/backend/access/transam/xlog.c

@@ -9022,18 +9022,30 @@ CreateCheckPoint(int flags)
 	 * and we will correctly flush the update below.  So we cannot miss any
 	 * xacts we need to wait for.
 	 */
-	vxids = GetVirtualXIDsDelayingChkpt(&nvxids);
+	vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
 	if (nvxids > 0)
 	{
 		do
 		{
 			pg_usleep(10000L);	/* wait for 10 msec */
-		} while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids));
+		} while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
+											  DELAY_CHKPT_START));
 	}
 	pfree(vxids);
 
 	CheckPointGuts(checkPoint.redo, flags);
 
+	vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
+	if (nvxids > 0)
+	{
+		do
+		{
+			pg_usleep(10000L);	/* wait for 10 msec */
+		} while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
+											  DELAY_CHKPT_COMPLETE));
+	}
+	pfree(vxids);
+
 	/*
 	 * Take a snapshot of running transactions and write this to WAL. This
 	 * allows us to reconstruct the state of running transactions during

src/backend/access/transam/xloginsert.c

@@ -904,7 +904,7 @@ XLogSaveBufferForHint(Buffer buffer, bool buffer_std)
 	/*
 	 * Ensure no checkpoint can change our view of RedoRecPtr.
 	 */
-	Assert(MyProc->delayChkpt);
+	Assert((MyProc->delayChkpt & DELAY_CHKPT_START) != 0);
 
 	/*
 	 * Update RedoRecPtr so that we can make the right decision

src/backend/catalog/storage.c

@@ -325,6 +325,22 @@ RelationTruncate(Relation rel, BlockNumber nblocks)
 
 	RelationPreTruncate(rel);
 
+	/*
+	 * Make sure that a concurrent checkpoint can't complete while truncation
+	 * is in progress.
+	 *
+	 * 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.
+	 */
+	Assert((MyProc->delayChkpt & DELAY_CHKPT_COMPLETE) == 0);
+	MyProc->delayChkpt |= DELAY_CHKPT_COMPLETE;
+
 	/*
 	 * We WAL-log the truncation before actually truncating, which means
 	 * trouble if the truncation fails. If we then crash, the WAL replay
@@ -363,13 +379,24 @@ RelationTruncate(Relation rel, BlockNumber nblocks)
 			XLogFlush(lsn);
 	}
 
-	/* Do the real work to truncate relation forks */
+	/*
+	 * This will first remove any buffers from the buffer pool that should no
+	 * longer exist after truncation is complete, and then truncate the
+	 * corresponding files on disk.
+	 */
 	smgrtruncate(rel->rd_smgr, forks, nforks, blocks);
 
+	/* We've done all the critical work, so checkpoints are OK now. */
+	MyProc->delayChkpt &= ~DELAY_CHKPT_COMPLETE;
+
 	/*
 	 * Update upper-level FSM pages to account for the truncation. This is
 	 * important because the just-truncated pages were likely marked as
 	 * all-free, and would be preferentially selected.
+	 *
+	 * NB: There's no point in delaying checkpoints until this is done.
+	 * Because the FSM is not WAL-logged, we have to be prepared for the
+	 * possibility of corruption after a crash anyway.
 	 */
 	if (need_fsm_vacuum)
 		FreeSpaceMapVacuumRange(rel, nblocks, InvalidBlockNumber);

src/backend/storage/buffer/bufmgr.c

@@ -3647,7 +3647,9 @@ MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
 			 * essential that CreateCheckpoint waits for virtual transactions
 			 * rather than full transactionids.
 			 */
-			MyProc->delayChkpt = delayChkpt = true;
+			Assert((MyProc->delayChkpt & DELAY_CHKPT_START) == 0);
+			MyProc->delayChkpt |= DELAY_CHKPT_START;
+			delayChkpt = true;
 			lsn = XLogSaveBufferForHint(buffer, buffer_std);
 		}
 
@@ -3680,7 +3682,7 @@ MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
 		UnlockBufHdr(bufHdr, buf_state);
 
 		if (delayChkpt)
-			MyProc->delayChkpt = false;
+			MyProc->delayChkpt &= ~DELAY_CHKPT_START;
 
 		if (dirtied)
 		{

src/backend/storage/ipc/procarray.c

@@ -434,7 +434,10 @@ ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
 		pgxact->xmin = InvalidTransactionId;
 		/* must be cleared with xid/xmin: */
 		pgxact->vacuumFlags &= ~PROC_VACUUM_STATE_MASK;
-		proc->delayChkpt = false;	/* be sure this is cleared in abort */
+
+		/* be sure this is cleared in abort */
+		proc->delayChkpt = 0;
+
 		proc->recoveryConflictPending = false;
 
 		Assert(pgxact->nxids == 0);
@@ -456,7 +459,10 @@ ProcArrayEndTransactionInternal(PGPROC *proc, PGXACT *pgxact,
 	pgxact->xmin = InvalidTransactionId;
 	/* must be cleared with xid/xmin: */
 	pgxact->vacuumFlags &= ~PROC_VACUUM_STATE_MASK;
-	proc->delayChkpt = false;	/* be sure this is cleared in abort */
+
+	/* be sure this is cleared in abort */
+	proc->delayChkpt = 0;
+
 	proc->recoveryConflictPending = false;
 
 	/* Clear the subtransaction-XID cache too while holding the lock */
@@ -2272,7 +2278,8 @@ GetOldestSafeDecodingTransactionId(bool catalogOnly)
  * delaying checkpoint because they have critical actions in progress.
  *
  * Constructs an array of VXIDs of transactions that are currently in commit
- * critical sections, as shown by having delayChkpt set in their PGPROC.
+ * critical sections, as shown by having specified delayChkpt bits set in their
+ * PGPROC.
  *
  * Returns a palloc'd array that should be freed by the caller.
  * *nvxids is the number of valid entries.
@@ -2286,13 +2293,15 @@ GetOldestSafeDecodingTransactionId(bool catalogOnly)
  * for clearing of delayChkpt to propagate is unimportant for correctness.
  */
 VirtualTransactionId *
-GetVirtualXIDsDelayingChkpt(int *nvxids)
+GetVirtualXIDsDelayingChkpt(int *nvxids, int type)
 {
 	VirtualTransactionId *vxids;
 	ProcArrayStruct *arrayP = procArray;
 	int			count = 0;
 	int			index;
 
+	Assert(type != 0);
+
 	/* allocate what's certainly enough result space */
 	vxids = (VirtualTransactionId *)
 		palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
@@ -2304,7 +2313,7 @@ GetVirtualXIDsDelayingChkpt(int *nvxids)
 		int			pgprocno = arrayP->pgprocnos[index];
 		PGPROC	   *proc = &allProcs[pgprocno];
 
-		if (proc->delayChkpt)
+		if ((proc->delayChkpt & type) != 0)
 		{
 			VirtualTransactionId vxid;
 
@@ -2330,12 +2339,14 @@ GetVirtualXIDsDelayingChkpt(int *nvxids)
  * those numbers should be small enough for it not to be a problem.
  */
 bool
-HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids)
+HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type)
 {
 	bool		result = false;
 	ProcArrayStruct *arrayP = procArray;
 	int			index;
 
+	Assert(type != 0);
+
 	LWLockAcquire(ProcArrayLock, LW_SHARED);
 
 	for (index = 0; index < arrayP->numProcs; index++)
@@ -2346,7 +2357,8 @@ HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids)
 
 		GET_VXID_FROM_PGPROC(vxid, *proc);
 
-		if (proc->delayChkpt && VirtualTransactionIdIsValid(vxid))
+		if ((proc->delayChkpt & type) != 0 &&
+			VirtualTransactionIdIsValid(vxid))
 		{
 			int			i;
 

src/backend/storage/lmgr/proc.c

@@ -396,7 +396,7 @@ InitProcess(void)
 	MyProc->roleId = InvalidOid;
 	MyProc->tempNamespaceId = InvalidOid;
 	MyProc->isBackgroundWorker = IsBackgroundWorker;
-	MyProc->delayChkpt = false;
+	MyProc->delayChkpt = 0;
 	MyPgXact->vacuumFlags = 0;
 	/* NB -- autovac launcher intentionally does not set IS_AUTOVACUUM */
 	if (IsAutoVacuumWorkerProcess())
@@ -578,7 +578,7 @@ InitAuxiliaryProcess(void)
 	MyProc->roleId = InvalidOid;
 	MyProc->tempNamespaceId = InvalidOid;
 	MyProc->isBackgroundWorker = IsBackgroundWorker;
-	MyProc->delayChkpt = false;
+	MyProc->delayChkpt = 0;
 	MyPgXact->vacuumFlags = 0;
 	MyProc->lwWaiting = false;
 	MyProc->lwWaitMode = 0;

src/include/storage/proc.h

@@ -83,6 +83,41 @@ struct XidCache
  */
 #define INVALID_PGPROCNO		PG_INT32_MAX
 
+/*
+ * Flags for PGPROC.delayChkpt
+ *
+ * These flags can be used to delay the start or completion of a checkpoint
+ * for short periods. A flag is in effect if the corresponding bit is set in
+ * the PGPROC of any backend.
+ *
+ * For our purposes here, a checkpoint has three phases: (1) determine the
+ * location to which the redo pointer will be moved, (2) write all the
+ * data durably to disk, and (3) WAL-log the checkpoint.
+ *
+ * Setting DELAY_CHKPT_START prevents the system from moving from phase 1
+ * to phase 2. This is useful when we are performing a WAL-logged modification
+ * of data that will be flushed to disk in phase 2. By setting this flag
+ * before writing WAL and clearing it after we've both written WAL and
+ * performed the corresponding modification, we ensure that if the WAL record
+ * is inserted prior to the new redo point, the corresponding data changes will
+ * also be flushed to disk before the checkpoint can complete. (In the
+ * extremely common case where the data being modified is in shared buffers
+ * and we acquire an exclusive content lock on the relevant buffers before
+ * writing WAL, this mechanism is not needed, because phase 2 will block
+ * until we release the content lock and then flush the modified data to
+ * disk.)
+ *
+ * Setting DELAY_CHKPT_COMPLETE prevents the system from moving from phase 2
+ * to phase 3. This is useful if we are performing a WAL-logged operation that
+ * might invalidate buffers, such as relation truncation. In this case, we need
+ * to ensure that any buffers which were invalidated and thus not flushed by
+ * the checkpoint are actaully destroyed on disk. Replay can cope with a file
+ * or block that doesn't exist, but not with a block that has the wrong
+ * contents.
+ */
+#define DELAY_CHKPT_START		(1<<0)
+#define DELAY_CHKPT_COMPLETE	(1<<1)
+
 /*
  * Each backend has a PGPROC struct in shared memory.  There is also a list of
  * currently-unused PGPROC structs that will be reallocated to new backends.
@@ -149,7 +184,7 @@ struct PGPROC
 	LOCKMASK	heldLocks;		/* bitmask for lock types already held on this
 								 * lock object by this backend */
 
-	bool		delayChkpt;		/* true if this proc delays checkpoint start */
+	int			delayChkpt;		/* for DELAY_CHKPT_* flags */
 
 	/*
 	 * Info to allow us to wait for synchronous replication, if needed.

src/include/storage/procarray.h

@@ -92,8 +92,9 @@ extern TransactionId GetOldestXmin(Relation rel, int flags);
 extern TransactionId GetOldestActiveTransactionId(void);
 extern TransactionId GetOldestSafeDecodingTransactionId(bool catalogOnly);
 
-extern VirtualTransactionId *GetVirtualXIDsDelayingChkpt(int *nvxids);
-extern bool HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids);
+extern VirtualTransactionId *GetVirtualXIDsDelayingChkpt(int *nvxids, int type);
+extern bool HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids,
+										 int nvxids, int type);
 
 extern PGPROC *BackendPidGetProc(int pid);
 extern PGPROC *BackendPidGetProcWithLock(int pid);