Rework the way multixact truncations work.

The fact that multixact truncations are not WAL logged has caused a fair
share of problems. Amongst others it requires to do computations during
recovery while the database is not in a consistent state, delaying
truncations till checkpoints, and handling members being truncated, but
offset not.

We tried to put bandaids on lots of these issues over the last years,
but it seems time to change course. Thus this patch introduces WAL
logging for multixact truncations.

This allows:
1) to perform the truncation directly during VACUUM, instead of delaying it
   to the checkpoint.
2) to avoid looking at the offsets SLRU for truncation during recovery,
   we can just use the master's values.
3) simplify a fair amount of logic to keep in memory limits straight,
   this has gotten much easier

During the course of fixing this a bunch of additional bugs had to be
fixed:
1) Data was not purged from memory the member's SLRU before deleting
   segments. This happened to be hard or impossible to hit due to the
   interlock between checkpoints and truncation.
2) find_multixact_start() relied on SimpleLruDoesPhysicalPageExist - but
   that doesn't work for offsets that haven't yet been flushed to
   disk. Add code to flush the SLRUs to fix. Not pretty, but it feels
   slightly safer to only make decisions based on actual on-disk state.
3) find_multixact_start() could be called concurrently with a truncation
   and thus fail. Via SetOffsetVacuumLimit() that could lead to a round
   of emergency vacuuming. The problem remains in
   pg_get_multixact_members(), but that's quite harmless.

For now this is going to only get applied to 9.5+, leaving the issues in
the older branches in place. It is quite possible that we need to
backpatch at a later point though.

For the case this gets backpatched we need to handle that an updated
standby may be replaying WAL from a not-yet upgraded primary. We have to
recognize that situation and use "old style" truncation (i.e. looking at
the SLRUs) during WAL replay. In contrast to before, this now happens in
the startup process, when replaying a checkpoint record, instead of the
checkpointer. Doing truncation in the restartpoint is incorrect, they
can happen much later than the original checkpoint, thereby leading to
wraparound.  To avoid "multixact_redo: unknown op code 48" errors
standbys would have to be upgraded before primaries.

A later patch will bump the WAL page magic, and remove the legacy
truncation codepaths. Legacy truncation support is just included to make
a possible future backpatch easier.

Discussion: 20150621192409.GA4797@alap3.anarazel.de
Reviewed-By: Robert Haas, Alvaro Herrera, Thomas Munro
Backpatch: 9.5 for now

src/backend/access/rmgrdesc/mxactdesc.c

@@ -70,6 +70,14 @@ multixact_desc(StringInfo buf, XLogReaderState *record)
 		for (i = 0; i < xlrec->nmembers; i++)
 			out_member(buf, &xlrec->members[i]);
 	}
+	else if (info == XLOG_MULTIXACT_TRUNCATE_ID)
+	{
+		xl_multixact_truncate *xlrec = (xl_multixact_truncate *) rec;
+
+		appendStringInfo(buf, "offsets [%u, %u), members [%u, %u)",
+						 xlrec->startTruncOff, xlrec->endTruncOff,
+						 xlrec->startTruncMemb, xlrec->endTruncMemb);
+	}
 }
 
 const char *
@@ -88,6 +96,9 @@ multixact_identify(uint8 info)
 		case XLOG_MULTIXACT_CREATE_ID:
 			id = "CREATE_ID";
 			break;
+		case XLOG_MULTIXACT_TRUNCATE_ID:
+			id = "TRUNCATE_ID";
+			break;
 	}
 
 	return id;

src/backend/access/transam/multixact.c

@@ -49,9 +49,7 @@
  * value is removed; the cutoff value is stored in pg_class.  The minimum value
  * across all tables in each database is stored in pg_database, and the global
  * minimum across all databases is part of pg_control and is kept in shared
- * memory.  At checkpoint time, after the value is known flushed in WAL, any
+ * memory.  Whenever that minimum is advanced, the SLRUs are truncated.
- * files that correspond to multixacts older than that value are removed.
- * (These files are also removed when a restartpoint is executed.)
  *
  * When new multixactid values are to be created, care is taken that the
  * counter does not fall within the wraparound horizon considering the global
@@ -83,6 +81,7 @@
 #include "postmaster/autovacuum.h"
 #include "storage/lmgr.h"
 #include "storage/pmsignal.h"
+#include "storage/proc.h"
 #include "storage/procarray.h"
 #include "utils/builtins.h"
 #include "utils/memutils.h"
@@ -109,6 +108,7 @@
 	((xid) / (MultiXactOffset) MULTIXACT_OFFSETS_PER_PAGE)
 #define MultiXactIdToOffsetEntry(xid) \
 	((xid) % (MultiXactOffset) MULTIXACT_OFFSETS_PER_PAGE)
+#define MultiXactIdToOffsetSegment(xid) (MultiXactIdToOffsetPage(xid) / SLRU_PAGES_PER_SEGMENT)
 
 /*
  * The situation for members is a bit more complex: we store one byte of
@@ -153,6 +153,7 @@
 
 /* page in which a member is to be found */
 #define MXOffsetToMemberPage(xid) ((xid) / (TransactionId) MULTIXACT_MEMBERS_PER_PAGE)
+#define MXOffsetToMemberSegment(xid) (MXOffsetToMemberPage(xid) / SLRU_PAGES_PER_SEGMENT)
 
 /* Location (byte offset within page) of flag word for a given member */
 #define MXOffsetToFlagsOffset(xid) \
@@ -212,19 +213,20 @@ typedef struct MultiXactStateData
 	Oid			oldestMultiXactDB;
 
 	/*
-	 * Oldest multixact offset that is potentially referenced by a
+	 * Oldest multixact offset that is potentially referenced by a multixact
-	 * multixact referenced by a relation.  We don't always know this value,
+	 * referenced by a relation.  We don't always know this value, so there's
-	 * so there's a flag here to indicate whether or not we currently do.
+	 * a flag here to indicate whether or not we currently do.
 	 */
 	MultiXactOffset oldestOffset;
 	bool		oldestOffsetKnown;
 
 	/*
-	 * This is what the previous checkpoint stored as the truncate position.
+	 * True if a multixact truncation WAL record was replayed since the last
-	 * This value is the oldestMultiXactId that was valid when a checkpoint
+	 * checkpoint. This is used to trigger 'legacy truncations', i.e. truncate
-	 * was last executed.
+	 * by looking at the data directory during WAL replay, when the primary is
+	 * too old to generate truncation records.
 	 */
-	MultiXactId lastCheckpointedOldest;
+	bool		sawTruncationInCkptCycle;
 
 	/* support for anti-wraparound measures */
 	MultiXactId multiVacLimit;
@@ -233,8 +235,7 @@ typedef struct MultiXactStateData
 	MultiXactId multiWrapLimit;
 
 	/* support for members anti-wraparound measures */
-	MultiXactOffset offsetStopLimit;
+	MultiXactOffset offsetStopLimit;	/* known if oldestOffsetKnown */
-	bool offsetStopLimitKnown;
 
 	/*
 	 * Per-backend data starts here.  We have two arrays stored in the area
@@ -364,12 +365,14 @@ static bool MultiXactOffsetPrecedes(MultiXactOffset offset1,
 						MultiXactOffset offset2);
 static void ExtendMultiXactOffset(MultiXactId multi);
 static void ExtendMultiXactMember(MultiXactOffset offset, int nmembers);
-static void DetermineSafeOldestOffset(MultiXactId oldestMXact);
 static bool MultiXactOffsetWouldWrap(MultiXactOffset boundary,
 						 MultiXactOffset start, uint32 distance);
-static bool SetOffsetVacuumLimit(bool finish_setup);
+static bool SetOffsetVacuumLimit(void);
 static bool find_multixact_start(MultiXactId multi, MultiXactOffset *result);
 static void WriteMZeroPageXlogRec(int pageno, uint8 info);
+static void WriteMTruncateXlogRec(Oid oldestMultiDB,
+					  MultiXactId startOff, MultiXactId endOff,
+					  MultiXactOffset startMemb, MultiXactOffset endMemb);
 
 
 /*
@@ -1099,7 +1102,7 @@ GetNewMultiXactId(int nmembers, MultiXactOffset *offset)
 	 *----------
 	 */
 #define OFFSET_WARN_SEGMENTS	20
-	if (MultiXactState->offsetStopLimitKnown &&
+	if (MultiXactState->oldestOffsetKnown &&
 		MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit, nextOffset,
 								 nmembers))
 	{
@@ -1139,7 +1142,7 @@ GetNewMultiXactId(int nmembers, MultiXactOffset *offset)
 			SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
 	}
 
-	if (MultiXactState->offsetStopLimitKnown &&
+	if (MultiXactState->oldestOffsetKnown &&
 		MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit,
 								 nextOffset,
 								 nmembers + MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT * OFFSET_WARN_SEGMENTS))
@@ -2010,20 +2013,24 @@ StartupMultiXact(void)
 
 /*
  * This must be called ONCE at the end of startup/recovery.
- *
- * We don't need any locks here, really; the SLRU locks are taken only because
- * slru.c expects to be called with locks held.
  */
 void
 TrimMultiXact(void)
 {
-	MultiXactId multi = MultiXactState->nextMXact;
+	MultiXactId nextMXact;
-	MultiXactOffset offset = MultiXactState->nextOffset;
+	MultiXactOffset offset;
 	MultiXactId oldestMXact;
+	Oid			oldestMXactDB;
 	int			pageno;
 	int			entryno;
 	int			flagsoff;
 
+	LWLockAcquire(MultiXactGenLock, LW_SHARED);
+	nextMXact = MultiXactState->nextMXact;
+	offset = MultiXactState->nextOffset;
+	oldestMXact = MultiXactState->oldestMultiXactId;
+	oldestMXactDB = MultiXactState->oldestMultiXactDB;
+	LWLockRelease(MultiXactGenLock);
 
 	/* Clean up offsets state */
 	LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
@@ -2031,20 +2038,20 @@ TrimMultiXact(void)
 	/*
 	 * (Re-)Initialize our idea of the latest page number for offsets.
 	 */
-	pageno = MultiXactIdToOffsetPage(multi);
+	pageno = MultiXactIdToOffsetPage(nextMXact);
 	MultiXactOffsetCtl->shared->latest_page_number = pageno;
 
 	/*
 	 * Zero out the remainder of the current offsets page.  See notes in
 	 * TrimCLOG() for motivation.
 	 */
-	entryno = MultiXactIdToOffsetEntry(multi);
+	entryno = MultiXactIdToOffsetEntry(nextMXact);
 	if (entryno != 0)
 	{
 		int			slotno;
 		MultiXactOffset *offptr;
 
-		slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
+		slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, nextMXact);
 		offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
 		offptr += entryno;
 
@@ -2093,12 +2100,13 @@ TrimMultiXact(void)
 
 	LWLockRelease(MultiXactMemberControlLock);
 
-	if (SetOffsetVacuumLimit(true) && IsUnderPostmaster)
+	/* signal that we're officially up */
-		SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
+	LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
-	LWLockAcquire(MultiXactGenLock, LW_SHARED);
+	MultiXactState->finishedStartup = true;
-	oldestMXact = MultiXactState->lastCheckpointedOldest;
 	LWLockRelease(MultiXactGenLock);
-	DetermineSafeOldestOffset(oldestMXact);
+
+	/* Now compute how far away the next members wraparound is. */
+	SetMultiXactIdLimit(oldestMXact, oldestMXactDB);
 }
 
 /*
@@ -2267,8 +2275,20 @@ SetMultiXactIdLimit(MultiXactId oldest_datminmxid, Oid oldest_datoid)
 	 (errmsg("MultiXactId wrap limit is %u, limited by database with OID %u",
 			 multiWrapLimit, oldest_datoid)));
 
+	/*
+	 * Computing the actual limits is only possible once the data directory is
+	 * in a consistent state. There's no need to compute the limits while
+	 * still replaying WAL - no decisions about new multis are made even
+	 * though multixact creations might be replayed. So we'll only do further
+	 * checks after TrimMultiXact() has been called.
+	 */
+	if (!MultiXactState->finishedStartup)
+		return;
+
+	Assert(!InRecovery);
+
 	/* Set limits for offset vacuum. */
-	needs_offset_vacuum = SetOffsetVacuumLimit(false);
+	needs_offset_vacuum = SetOffsetVacuumLimit();
 
 	/*
 	 * If past the autovacuum force point, immediately signal an autovac
@@ -2278,11 +2298,11 @@ SetMultiXactIdLimit(MultiXactId oldest_datminmxid, Oid oldest_datoid)
 	 * another iteration immediately if there are still any old databases.
 	 */
 	if ((MultiXactIdPrecedes(multiVacLimit, curMulti) ||
-		 needs_offset_vacuum) && IsUnderPostmaster && !InRecovery)
+		 needs_offset_vacuum) && IsUnderPostmaster)
 		SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
 
 	/* Give an immediate warning if past the wrap warn point */
-	if (MultiXactIdPrecedes(multiWarnLimit, curMulti) && !InRecovery)
+	if (MultiXactIdPrecedes(multiWarnLimit, curMulti))
 	{
 		char	   *oldest_datname;
 
@@ -2350,27 +2370,39 @@ MultiXactAdvanceNextMXact(MultiXactId minMulti,
 }
 
 /*
- * Update our oldestMultiXactId value, but only if it's more recent than
+ * Update our oldestMultiXactId value, but only if it's more recent than what
- * what we had.  However, even if not, always update the oldest multixact
+ * we had.
- * offset limit.
+ *
+ * This may only be called during WAL replay.
  */
 void
 MultiXactAdvanceOldest(MultiXactId oldestMulti, Oid oldestMultiDB)
 {
-	if (MultiXactIdPrecedes(MultiXactState->oldestMultiXactId, oldestMulti))
+	Assert(InRecovery);
-		SetMultiXactIdLimit(oldestMulti, oldestMultiDB);
-}
 
+	if (MultiXactIdPrecedes(MultiXactState->oldestMultiXactId, oldestMulti))
+	{
 		/*
- * Update the "safe truncation point".  This is the newest value of oldestMulti
+		 * If there has been a truncation on the master, detected by seeing a
- * that is known to be flushed as part of a checkpoint record.
+		 * moving oldestMulti, without a corresponding truncation record, we
+		 * know that the primary is still running an older version of postgres
+		 * that doesn't yet log multixact truncations. So perform the
+		 * truncation ourselves.
 		 */
-void
+		if (!MultiXactState->sawTruncationInCkptCycle)
-MultiXactSetSafeTruncate(MultiXactId safeTruncateMulti)
 		{
-	LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
+			ereport(LOG,
-	MultiXactState->lastCheckpointedOldest = safeTruncateMulti;
+					(errmsg("performing legacy multixact truncation"),
-	LWLockRelease(MultiXactGenLock);
+					 errdetail("Legacy truncations are sometimes performed when replaying WAL from an older primary."),
+					 errhint("Upgrade the primary, it is susceptible to data corruption.")));
+			TruncateMultiXact(oldestMulti, oldestMultiDB, true);
+		}
+
+		SetMultiXactIdLimit(oldestMulti, oldestMultiDB);
+	}
+
+	/* only looked at in the startup process, no lock necessary */
+	MultiXactState->sawTruncationInCkptCycle = false;
 }
 
 /*
@@ -2525,133 +2557,57 @@ GetOldestMultiXactId(void)
 	return oldestMXact;
 }
 
-/*
- * Based on the given oldest MultiXactId, determine what's the oldest member
- * offset and install the limit info in MultiXactState, where it can be used to
- * prevent overrun of old data in the members SLRU area.
- */
-static void
-DetermineSafeOldestOffset(MultiXactId oldestMXact)
-{
-	MultiXactOffset oldestOffset;
-	MultiXactOffset nextOffset;
-	MultiXactOffset offsetStopLimit;
-	MultiXactOffset prevOffsetStopLimit;
-	MultiXactId		nextMXact;
-	bool			finishedStartup;
-	bool			prevOffsetStopLimitKnown;
-
-	/* Fetch values from shared memory. */
-	LWLockAcquire(MultiXactGenLock, LW_SHARED);
-	finishedStartup = MultiXactState->finishedStartup;
-	nextMXact = MultiXactState->nextMXact;
-	nextOffset = MultiXactState->nextOffset;
-	prevOffsetStopLimit = MultiXactState->offsetStopLimit;
-	prevOffsetStopLimitKnown = MultiXactState->offsetStopLimitKnown;
-	LWLockRelease(MultiXactGenLock);
-
-	/* Don't worry about this until after we've started up. */
-	if (!finishedStartup)
-		return;
-
-	/*
-	 * Determine the offset of the oldest multixact.  Normally, we can read
-	 * the offset from the multixact itself, but there's an important special
-	 * case: if there are no multixacts in existence at all, oldestMXact
-	 * obviously can't point to one.  It will instead point to the multixact
-	 * ID that will be assigned the next time one is needed.
-	 *
-	 * NB: oldestMXact should be the oldest multixact that still exists in the
-	 * SLRU, unlike in SetOffsetVacuumLimit, where we do this same computation
-	 * based on the oldest value that might be referenced in a table.
-	 */
-	if (nextMXact == oldestMXact)
-		oldestOffset = nextOffset;
-	else
-	{
-		bool		oldestOffsetKnown;
-
-		oldestOffsetKnown = find_multixact_start(oldestMXact, &oldestOffset);
-		if (!oldestOffsetKnown)
-		{
-			ereport(LOG,
-					(errmsg("MultiXact member wraparound protections are disabled because oldest checkpointed MultiXact %u does not exist on disk",
-						oldestMXact)));
-			return;
-		}
-	}
-
-	/* move back to start of the corresponding segment */
-	offsetStopLimit = oldestOffset - (oldestOffset %
-		(MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT));
-	/* always leave one segment before the wraparound point */
-	offsetStopLimit -= (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT);
-
-	/* if nothing has changed, we're done */
-	if (prevOffsetStopLimitKnown && offsetStopLimit == prevOffsetStopLimit)
-		return;
-
-	LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
-	MultiXactState->offsetStopLimit = offsetStopLimit;
-	MultiXactState->offsetStopLimitKnown = true;
-	LWLockRelease(MultiXactGenLock);
-
-	if (!prevOffsetStopLimitKnown && IsUnderPostmaster)
-		ereport(LOG,
-				(errmsg("MultiXact member wraparound protections are now enabled")));
-	ereport(DEBUG1,
-			(errmsg("MultiXact member stop limit is now %u based on MultiXact %u",
-				offsetStopLimit, oldestMXact)));
-}
-
 /*
  * Determine how aggressively we need to vacuum in order to prevent member
  * wraparound.
  *
- * To determine the oldest multixact ID, we look at oldestMultiXactId, not
+ * To do so determine what's the oldest member offset and install the limit
- * lastCheckpointedOldest.  That's because vacuuming can't help with anything
+ * info in MultiXactState, where it can be used to prevent overrun of old data
- * older than oldestMultiXactId; anything older than that isn't referenced
+ * in the members SLRU area.
- * by any table.  Offsets older than oldestMultiXactId but not as old as
- * lastCheckpointedOldest will go away after the next checkpoint.
  *
  * The return value is true if emergency autovacuum is required and false
  * otherwise.
  */
 static bool
-SetOffsetVacuumLimit(bool finish_setup)
+SetOffsetVacuumLimit(void)
 {
 	MultiXactId oldestMultiXactId;
 	MultiXactId nextMXact;
-	bool		finishedStartup;
 	MultiXactOffset oldestOffset = 0;	/* placate compiler */
+	MultiXactOffset prevOldestOffset;
 	MultiXactOffset nextOffset;
 	bool		oldestOffsetKnown = false;
-	MultiXactOffset prevOldestOffset;
 	bool		prevOldestOffsetKnown;
+	MultiXactOffset offsetStopLimit = 0;
+
+	/*
+	 * NB: Have to prevent concurrent truncation, we might otherwise try to
+	 * lookup a oldestMulti that's concurrently getting truncated away.
+	 */
+	LWLockAcquire(MultiXactTruncationLock, LW_SHARED);
 
 	/* Read relevant fields from shared memory. */
 	LWLockAcquire(MultiXactGenLock, LW_SHARED);
 	oldestMultiXactId = MultiXactState->oldestMultiXactId;
 	nextMXact = MultiXactState->nextMXact;
 	nextOffset = MultiXactState->nextOffset;
-	finishedStartup = MultiXactState->finishedStartup;
-	prevOldestOffset = MultiXactState->oldestOffset;
 	prevOldestOffsetKnown = MultiXactState->oldestOffsetKnown;
+	prevOldestOffset = MultiXactState->oldestOffset;
+	Assert(MultiXactState->finishedStartup);
 	LWLockRelease(MultiXactGenLock);
 
-	/* Don't do this until after any recovery is complete. */
-	if (!finishedStartup && !finish_setup)
-		return false;
-
 	/*
-	 * If no multixacts exist, then oldestMultiXactId will be the next
+	 * Determine the offset of the oldest multixact.  Normally, we can read
-	 * multixact that will be created, rather than an existing multixact.
+	 * the offset from the multixact itself, but there's an important special
+	 * case: if there are no multixacts in existence at all, oldestMXact
+	 * obviously can't point to one.  It will instead point to the multixact
+	 * ID that will be assigned the next time one is needed.
 	 */
 	if (oldestMultiXactId == nextMXact)
 	{
 		/*
-		 * When the next multixact gets created, it will be stored at the
+		 * When the next multixact gets created, it will be stored at the next
-		 * next offset.
+		 * offset.
 		 */
 		oldestOffset = nextOffset;
 		oldestOffsetKnown = true;
@@ -2659,53 +2615,65 @@ SetOffsetVacuumLimit(bool finish_setup)
 	else
 	{
 		/*
-		 * Figure out where the oldest existing multixact's offsets are stored.
+		 * Figure out where the oldest existing multixact's offsets are
-		 * Due to bugs in early release of PostgreSQL 9.3.X and 9.4.X, the
+		 * stored. Due to bugs in early release of PostgreSQL 9.3.X and 9.4.X,
-		 * supposedly-earliest multixact might not really exist.  We are
+		 * the supposedly-earliest multixact might not really exist.  We are
 		 * careful not to fail in that case.
 		 */
 		oldestOffsetKnown =
 			find_multixact_start(oldestMultiXactId, &oldestOffset);
-	}
 
-	/*
-	 * Except when initializing the system for the first time, there's no
-	 * need to update anything if we don't know the oldest offset or if it
-	 * hasn't changed.
-	 */
-	if (finish_setup ||
-		(oldestOffsetKnown && !prevOldestOffsetKnown) ||
-		(oldestOffsetKnown && prevOldestOffset != oldestOffset))
-	{
-		/* Install the new limits. */
-		LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
-		MultiXactState->oldestOffset = oldestOffset;
-		MultiXactState->oldestOffsetKnown = oldestOffsetKnown;
-		MultiXactState->finishedStartup = true;
-		LWLockRelease(MultiXactGenLock);
-
-		/* Log the info */
 		if (oldestOffsetKnown)
 			ereport(DEBUG1,
 					(errmsg("oldest MultiXactId member is at offset %u",
 							oldestOffset)));
 		else
-			ereport(DEBUG1,
+			ereport(LOG,
-					(errmsg("oldest MultiXactId member offset unknown")));
+					(errmsg("MultiXact member wraparound protections are disabled because oldest checkpointed MultiXact %u does not exist on disk",
+							oldestMultiXactId)));
 	}
 
+	LWLockRelease(MultiXactTruncationLock);
+
 	/*
-	 * If we failed to get the oldest offset this time, but we have a value
+	 * If we can, compute limits (and install them MultiXactState) to prevent
-	 * from a previous pass through this function, assess the need for
+	 * overrun of old data in the members SLRU area. We can only do so if the
-	 * autovacuum based on that old value rather than automatically forcing
+	 * oldest offset is known though.
-	 * it.
 	 */
-	if (prevOldestOffsetKnown && !oldestOffsetKnown)
+	if (oldestOffsetKnown)
+	{
+		/* move back to start of the corresponding segment */
+		offsetStopLimit = oldestOffset - (oldestOffset %
+					  (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT));
+
+		/* always leave one segment before the wraparound point */
+		offsetStopLimit -= (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT);
+
+		if (!prevOldestOffsetKnown && IsUnderPostmaster)
+			ereport(LOG,
+					(errmsg("MultiXact member wraparound protections are now enabled")));
+		ereport(DEBUG1,
+		(errmsg("MultiXact member stop limit is now %u based on MultiXact %u",
+				offsetStopLimit, oldestMultiXactId)));
+	}
+	else if (prevOldestOffsetKnown)
 	{
+		/*
+		 * If we failed to get the oldest offset this time, but we have a
+		 * value from a previous pass through this function, use the old value
+		 * rather than automatically forcing it.
+		 */
 		oldestOffset = prevOldestOffset;
 		oldestOffsetKnown = true;
 	}
 
+	/* Install the computed values */
+	LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
+	MultiXactState->oldestOffset = oldestOffset;
+	MultiXactState->oldestOffsetKnown = oldestOffsetKnown;
+	MultiXactState->offsetStopLimit = offsetStopLimit;
+	LWLockRelease(MultiXactGenLock);
+
 	/*
 	 * Do we need an emergency autovacuum?	If we're not sure, assume yes.
 	 */
@@ -2720,7 +2688,7 @@ SetOffsetVacuumLimit(bool finish_setup)
  * boundary point, hence the name.  The reason we don't want to use the regular
  * 2^31-modulo arithmetic here is that we want to be able to use the whole of
  * the 2^32-1 space here, allowing for more multixacts that would fit
- * otherwise.  See also SlruScanDirCbRemoveMembers.
+ * otherwise.
  */
 static bool
 MultiXactOffsetWouldWrap(MultiXactOffset boundary, MultiXactOffset start,
@@ -2766,6 +2734,9 @@ MultiXactOffsetWouldWrap(MultiXactOffset boundary, MultiXactOffset start,
  *
  * Returns false if the file containing the multi does not exist on disk.
  * Otherwise, returns true and sets *result to the starting member offset.
+ *
+ * This function does not prevent concurrent truncation, so if that's
+ * required, the caller has to protect against that.
  */
 static bool
 find_multixact_start(MultiXactId multi, MultiXactOffset *result)
@@ -2776,9 +2747,22 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
 	int			slotno;
 	MultiXactOffset *offptr;
 
+	/* XXX: Remove || AmStartupProcess() after WAL page magic bump */
+	Assert(MultiXactState->finishedStartup || AmStartupProcess());
+
 	pageno = MultiXactIdToOffsetPage(multi);
 	entryno = MultiXactIdToOffsetEntry(multi);
 
+	/*
+	 * Flush out dirty data, so PhysicalPageExists can work correctly.
+	 * SimpleLruFlush() is a pretty big hammer for that.  Alternatively we
+	 * could add a in-memory version of page exists, but find_multixact_start
+	 * is called infrequently, and it doesn't seem bad to flush buffers to
+	 * disk before truncation.
+	 */
+	SimpleLruFlush(MultiXactOffsetCtl, true);
+	SimpleLruFlush(MultiXactMemberCtl, true);
+
 	if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
 		return false;
 
@@ -2884,65 +2868,6 @@ MultiXactMemberFreezeThreshold(void)
 	return multixacts - victim_multixacts;
 }
 
-/*
- * SlruScanDirectory callback.
- *		This callback deletes segments that are outside the range determined by
- *		the given page numbers.
- *
- * Both range endpoints are exclusive (that is, segments containing any of
- * those pages are kept.)
- */
-typedef struct MembersLiveRange
-{
-	int			rangeStart;
-	int			rangeEnd;
-} MembersLiveRange;
-
-static bool
-SlruScanDirCbRemoveMembers(SlruCtl ctl, char *filename, int segpage,
-						   void *data)
-{
-	MembersLiveRange *range = (MembersLiveRange *) data;
-	MultiXactOffset nextOffset;
-
-	if ((segpage == range->rangeStart) ||
-		(segpage == range->rangeEnd))
-		return false;			/* easy case out */
-
-	/*
-	 * To ensure that no segment is spuriously removed, we must keep track of
-	 * new segments added since the start of the directory scan; to do this,
-	 * we update our end-of-range point as we run.
-	 *
-	 * As an optimization, we can skip looking at shared memory if we know for
-	 * certain that the current segment must be kept.  This is so because
-	 * nextOffset never decreases, and we never increase rangeStart during any
-	 * one run.
-	 */
-	if (!((range->rangeStart > range->rangeEnd &&
-		   segpage > range->rangeEnd && segpage < range->rangeStart) ||
-		  (range->rangeStart < range->rangeEnd &&
-		   (segpage < range->rangeStart || segpage > range->rangeEnd))))
-		return false;
-
-	/*
-	 * Update our idea of the end of the live range.
-	 */
-	LWLockAcquire(MultiXactGenLock, LW_SHARED);
-	nextOffset = MultiXactState->nextOffset;
-	LWLockRelease(MultiXactGenLock);
-	range->rangeEnd = MXOffsetToMemberPage(nextOffset);
-
-	/* Recheck the deletion condition.  If it still holds, perform deletion */
-	if ((range->rangeStart > range->rangeEnd &&
-		 segpage > range->rangeEnd && segpage < range->rangeStart) ||
-		(range->rangeStart < range->rangeEnd &&
-		 (segpage < range->rangeStart || segpage > range->rangeEnd)))
-		SlruDeleteSegment(ctl, filename);
-
-	return false;				/* keep going */
-}
-
 typedef struct mxtruncinfo
 {
 	int			earliestExistingPage;
@@ -2966,37 +2891,115 @@ SlruScanDirCbFindEarliest(SlruCtl ctl, char *filename, int segpage, void *data)
 	return false;				/* keep going */
 }
 
+
+/*
+ * Delete members segments [oldest, newOldest)
+ *
+ * The members SLRU can, in contrast to the offsets one, be filled to almost
+ * the full range at once. This means SimpleLruTruncate() can't trivially be
+ * used - instead the to-be-deleted range is computed using the offsets
+ * SLRU. C.f. TruncateMultiXact().
+ */
+static void
+PerformMembersTruncation(MultiXactOffset oldestOffset, MultiXactOffset newOldestOffset)
+{
+	const int	maxsegment = MXOffsetToMemberSegment(MaxMultiXactOffset);
+	int			startsegment = MXOffsetToMemberSegment(oldestOffset);
+	int			endsegment = MXOffsetToMemberSegment(newOldestOffset);
+	int			segment = startsegment;
+
+	/*
+	 * Delete all the segments but the last one. The last segment can still
+	 * contain, possibly partially, valid data.
+	 */
+	while (segment != endsegment)
+	{
+		elog(DEBUG2, "truncating multixact members segment %x", segment);
+		SlruDeleteSegment(MultiXactMemberCtl, segment);
+
+		/* move to next segment, handling wraparound correctly */
+		if (segment == maxsegment)
+			segment = 0;
+		else
+			segment += 1;
+	}
+}
+
+/*
+ * Delete offsets segments [oldest, newOldest)
+ */
+static void
+PerformOffsetsTruncation(MultiXactId oldestMulti, MultiXactId newOldestMulti)
+{
+	/*
+	 * We step back one multixact to avoid passing a cutoff page that hasn't
+	 * been created yet in the rare case that oldestMulti would be the first
+	 * item on a page and oldestMulti == nextMulti.  In that case, if we
+	 * didn't subtract one, we'd trigger SimpleLruTruncate's wraparound
+	 * detection.
+	 */
+	SimpleLruTruncate(MultiXactOffsetCtl,
+			   MultiXactIdToOffsetPage(PreviousMultiXactId(newOldestMulti)));
+}
+
 /*
  * Remove all MultiXactOffset and MultiXactMember segments before the oldest
  * ones still of interest.
  *
- * On a primary, this is called by the checkpointer process after a checkpoint
+ * On a primary this is called as part of vacuum (via
- * has been flushed; during crash recovery, it's called from
+ * vac_truncate_clog()). During recovery truncation is normally done by
- * CreateRestartPoint().  In the latter case, we rely on the fact that
+ * replaying truncation WAL records instead of this routine; the exception is
- * xlog_redo() will already have called MultiXactAdvanceOldest().  Our
+ * when replaying records from an older primary that doesn't yet generate
- * latest_page_number will already have been initialized by StartupMultiXact()
+ * truncation WAL records. In that case truncation is triggered by
- * and kept up to date as new pages are zeroed.
+ * MultiXactAdvanceOldest().
+ *
+ * newOldestMulti is the oldest currently required multixact, newOldestMultiDB
+ * is one of the databases preventing newOldestMulti from increasing.
  */
 void
-TruncateMultiXact(void)
+TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB, bool in_recovery)
 {
-	MultiXactId oldestMXact;
+	MultiXactId oldestMulti;
+	MultiXactId nextMulti;
+	MultiXactOffset newOldestOffset;
 	MultiXactOffset oldestOffset;
-	MultiXactId		nextMXact;
 	MultiXactOffset nextOffset;
 	mxtruncinfo trunc;
 	MultiXactId earliest;
-	MembersLiveRange range;
 
-	Assert(AmCheckpointerProcess() || AmStartupProcess() ||
+	/*
-		   !IsPostmasterEnvironment);
+	 * Need to allow being called in recovery for backwards compatibility,
+	 * when an updated standby replays WAL generated by a non-updated primary.
+	 */
+	Assert(in_recovery || !RecoveryInProgress());
+	Assert(!in_recovery || AmStartupProcess());
+	Assert(in_recovery || MultiXactState->finishedStartup);
+
+	/*
+	 * We can only allow one truncation to happen at once. Otherwise parts of
+	 * members might vanish while we're doing lookups or similar. There's no
+	 * need to have an interlock with creating new multis or such, since those
+	 * are constrained by the limits (which only grow, never shrink).
+	 */
+	LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
 
 	LWLockAcquire(MultiXactGenLock, LW_SHARED);
-	oldestMXact = MultiXactState->lastCheckpointedOldest;
+	nextMulti = MultiXactState->nextMXact;
-	nextMXact = MultiXactState->nextMXact;
 	nextOffset = MultiXactState->nextOffset;
+	oldestMulti = MultiXactState->oldestMultiXactId;
 	LWLockRelease(MultiXactGenLock);
-	Assert(MultiXactIdIsValid(oldestMXact));
+	Assert(MultiXactIdIsValid(oldestMulti));
+
+	/*
+	 * Make sure to only attempt truncation if there's values to truncate
+	 * away. In normal processing values shouldn't go backwards, but there's
+	 * some corner cases (due to bugs) where that's possible.
+	 */
+	if (MultiXactIdPrecedesOrEquals(newOldestMulti, oldestMulti))
+	{
+		LWLockRelease(MultiXactTruncationLock);
+		return;
+	}
 
 	/*
 	 * Note we can't just plow ahead with the truncation; it's possible that
@@ -3004,6 +3007,9 @@ TruncateMultiXact(void)
 	 * going to attempt to read the offsets page to determine where to
 	 * truncate the members SLRU.  So we first scan the directory to determine
 	 * the earliest offsets page number that we can read without error.
+	 *
+	 * NB: It's also possible that the page that oldestMulti is on has already
+	 * been truncated away, and we crashed before updating oldestMulti.
 	 */
 	trunc.earliestExistingPage = -1;
 	SlruScanDirectory(MultiXactOffsetCtl, SlruScanDirCbFindEarliest, &trunc);
@@ -3011,19 +3017,10 @@ TruncateMultiXact(void)
 	if (earliest < FirstMultiXactId)
 		earliest = FirstMultiXactId;
 
-	/*
+	/* If there's nothing to remove, we can bail out early. */
-	 * If there's nothing to remove, we can bail out early.
+	if (MultiXactIdPrecedes(oldestMulti, earliest))
-	 *
-	 * Due to bugs in early releases of PostgreSQL 9.3.X and 9.4.X,
-	 * oldestMXact might point to a multixact that does not exist.
-	 * Autovacuum will eventually advance it to a value that does exist,
-	 * and we want to set a proper offsetStopLimit when that happens,
-	 * so call DetermineSafeOldestOffset here even if we're not actually
-	 * truncating.
-	 */
-	if (MultiXactIdPrecedes(oldestMXact, earliest))
 	{
-		DetermineSafeOldestOffset(oldestMXact);
+		LWLockRelease(MultiXactTruncationLock);
 		return;
 	}
 
@@ -3032,49 +3029,104 @@ TruncateMultiXact(void)
 	 * the starting offset of the oldest multixact.
 	 *
 	 * Hopefully, find_multixact_start will always work here, because we've
-	 * already checked that it doesn't precede the earliest MultiXact on
+	 * already checked that it doesn't precede the earliest MultiXact on disk.
-	 * disk.  But if it fails, don't truncate anything, and log a message.
+	 * But if it fails, don't truncate anything, and log a message.
 	 */
-	if (oldestMXact == nextMXact)
+	if (oldestMulti == nextMulti)
-		oldestOffset = nextOffset;		/* there are NO MultiXacts */
+	{
-	else if (!find_multixact_start(oldestMXact, &oldestOffset))
+		/* there are NO MultiXacts */
+		oldestOffset = nextOffset;
+	}
+	else if (!find_multixact_start(oldestMulti, &oldestOffset))
 	{
 		ereport(LOG,
 				(errmsg("oldest MultiXact %u not found, earliest MultiXact %u, skipping truncation",
-					oldestMXact, earliest)));
+						oldestMulti, earliest)));
+		LWLockRelease(MultiXactTruncationLock);
 		return;
 	}
 
 	/*
-	 * To truncate MultiXactMembers, we need to figure out the active page
+	 * Secondly compute up to where to truncate. Lookup the corresponding
-	 * range and delete all files outside that range.  The start point is the
+	 * member offset for newOldestMulti for that.
-	 * start of the segment containing the oldest offset; an end point of the
-	 * segment containing the next offset to use is enough.  The end point is
-	 * updated as MultiXactMember gets extended concurrently, elsewhere.
 	 */
-	range.rangeStart = MXOffsetToMemberPage(oldestOffset);
+	if (newOldestMulti == nextMulti)
-	range.rangeStart -= range.rangeStart % SLRU_PAGES_PER_SEGMENT;
+	{
+		/* there are NO MultiXacts */
+		newOldestOffset = nextOffset;
+	}
+	else if (!find_multixact_start(newOldestMulti, &newOldestOffset))
+	{
+		ereport(LOG,
+				(errmsg("cannot truncate up to MultiXact %u because it does not exist on disk, skipping truncation",
+						newOldestMulti)));
+		LWLockRelease(MultiXactTruncationLock);
+		return;
+	}
 
-	range.rangeEnd = MXOffsetToMemberPage(nextOffset);
+	elog(DEBUG1, "performing multixact truncation: "
+		 "offsets [%u, %u), offsets segments [%x, %x), "
+		 "members [%u, %u), members segments [%x, %x)",
+		 oldestMulti, newOldestMulti,
+		 MultiXactIdToOffsetSegment(oldestMulti),
+		 MultiXactIdToOffsetSegment(newOldestMulti),
+		 oldestOffset, newOldestOffset,
+		 MXOffsetToMemberSegment(oldestOffset),
+		 MXOffsetToMemberSegment(newOldestOffset));
 
-	SlruScanDirectory(MultiXactMemberCtl, SlruScanDirCbRemoveMembers, &range);
+	/*
+	 * Do truncation, and the WAL logging of the truncation, in a critical
+	 * section. That way offsets/members cannot get out of sync anymore, i.e.
+	 * once consistent the newOldestMulti will always exist in members, even
+	 * if we crashed in the wrong moment.
+	 */
+	START_CRIT_SECTION();
 
 	/*
-	 * Now we can truncate MultiXactOffset.  We step back one multixact to
+	 * Prevent checkpoints from being scheduled concurrently. This is critical
-	 * avoid passing a cutoff page that hasn't been created yet in the rare
+	 * because otherwise a truncation record might not be replayed after a
-	 * case that oldestMXact would be the first item on a page and oldestMXact
+	 * crash/basebackup, even though the state of the data directory would
-	 * == nextMXact.  In that case, if we didn't subtract one, we'd trigger
+	 * require it.  It's not possible (startup process doesn't have a PGXACT
-	 * SimpleLruTruncate's wraparound detection.
+	 * entry), and not needed, to do this during recovery, when performing an
+	 * old-style truncation, though. There the entire scheduling depends on
+	 * the replayed WAL records which be the same after a possible crash.
 	 */
-	SimpleLruTruncate(MultiXactOffsetCtl,
+	if (!in_recovery)
-				  MultiXactIdToOffsetPage(PreviousMultiXactId(oldestMXact)));
+	{
+		Assert(!MyPgXact->delayChkpt);
+		MyPgXact->delayChkpt = true;
+	}
+
+	/* WAL log truncation */
+	if (!in_recovery)
+		WriteMTruncateXlogRec(newOldestMultiDB,
+							  oldestMulti, newOldestMulti,
+							  oldestOffset, newOldestOffset);
 
 	/*
-	 * Now, and only now, we can advance the stop point for multixact members.
+	 * Update in-memory limits before performing the truncation, while inside
-	 * If we did it any sooner, the segments we deleted above might already
+	 * the critical section: Have to do it before truncation, to prevent
-	 * have been overwritten with new members.  That would be bad.
+	 * concurrent lookups of those values. Has to be inside the critical
+	 * section as otherwise a future call to this function would error out,
+	 * while looking up the oldest member in offsets, if our caller crashes
+	 * before updating the limits.
 	 */
-	DetermineSafeOldestOffset(oldestMXact);
+	LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
+	MultiXactState->oldestMultiXactId = newOldestMulti;
+	MultiXactState->oldestMultiXactDB = newOldestMultiDB;
+	LWLockRelease(MultiXactGenLock);
+
+	/* First truncate members */
+	PerformMembersTruncation(oldestOffset, newOldestOffset);
+
+	/* Then offsets */
+	PerformOffsetsTruncation(oldestMulti, newOldestMulti);
+
+	if (!in_recovery)
+		MyPgXact->delayChkpt = false;
+
+	END_CRIT_SECTION();
+	LWLockRelease(MultiXactTruncationLock);
 }
 
 /*
@@ -3170,6 +3222,34 @@ WriteMZeroPageXlogRec(int pageno, uint8 info)
 	(void) XLogInsert(RM_MULTIXACT_ID, info);
 }
 
+/*
+ * Write a TRUNCATE xlog record
+ *
+ * We must flush the xlog record to disk before returning --- see notes in
+ * TruncateCLOG().
+ */
+static void
+WriteMTruncateXlogRec(Oid oldestMultiDB,
+					  MultiXactId startTruncOff, MultiXactId endTruncOff,
+				MultiXactOffset startTruncMemb, MultiXactOffset endTruncMemb)
+{
+	XLogRecPtr	recptr;
+	xl_multixact_truncate xlrec;
+
+	xlrec.oldestMultiDB = oldestMultiDB;
+
+	xlrec.startTruncOff = startTruncOff;
+	xlrec.endTruncOff = endTruncOff;
+
+	xlrec.startTruncMemb = startTruncMemb;
+	xlrec.endTruncMemb = endTruncMemb;
+
+	XLogBeginInsert();
+	XLogRegisterData((char *) (&xlrec), SizeOfMultiXactTruncate);
+	recptr = XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_TRUNCATE_ID);
+	XLogFlush(recptr);
+}
+
 /*
  * MULTIXACT resource manager's routines
  */
@@ -3252,6 +3332,49 @@ multixact_redo(XLogReaderState *record)
 			LWLockRelease(XidGenLock);
 		}
 	}
+	else if (info == XLOG_MULTIXACT_TRUNCATE_ID)
+	{
+		xl_multixact_truncate xlrec;
+		int			pageno;
+
+		memcpy(&xlrec, XLogRecGetData(record),
+			   SizeOfMultiXactTruncate);
+
+		elog(DEBUG1, "replaying multixact truncation: "
+			 "offsets [%u, %u), offsets segments [%x, %x), "
+			 "members [%u, %u), members segments [%x, %x)",
+			 xlrec.startTruncOff, xlrec.endTruncOff,
+			 MultiXactIdToOffsetSegment(xlrec.startTruncOff),
+			 MultiXactIdToOffsetSegment(xlrec.endTruncOff),
+			 xlrec.startTruncMemb, xlrec.endTruncMemb,
+			 MXOffsetToMemberSegment(xlrec.startTruncMemb),
+			 MXOffsetToMemberSegment(xlrec.endTruncMemb));
+
+		/* should not be required, but more than cheap enough */
+		LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
+
+		/*
+		 * Advance the horizon values, so they're current at the end of
+		 * recovery.
+		 */
+		SetMultiXactIdLimit(xlrec.endTruncOff, xlrec.oldestMultiDB);
+
+		PerformMembersTruncation(xlrec.startTruncMemb, xlrec.endTruncMemb);
+
+		/*
+		 * During XLOG replay, latest_page_number isn't necessarily set up
+		 * yet; insert a suitable value to bypass the sanity test in
+		 * SimpleLruTruncate.
+		 */
+		pageno = MultiXactIdToOffsetPage(xlrec.endTruncOff);
+		MultiXactOffsetCtl->shared->latest_page_number = pageno;
+		PerformOffsetsTruncation(xlrec.startTruncOff, xlrec.endTruncOff);
+
+		LWLockRelease(MultiXactTruncationLock);
+
+		/* only looked at in the startup process, no lock necessary */
+		MultiXactState->sawTruncationInCkptCycle = true;
+	}
 	else
 		elog(PANIC, "multixact_redo: unknown op code %u", info);
 }

src/backend/access/transam/slru.c

@@ -134,6 +134,7 @@ static int	SlruSelectLRUPage(SlruCtl ctl, int pageno);
 
 static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
 						  int segpage, void *data);
+static void SlruInternalDeleteSegment(SlruCtl ctl, char *filename);
 
 /*
  * Initialization of shared memory
@@ -1075,7 +1076,7 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
  * Flush dirty pages to disk during checkpoint or database shutdown
  */
 void
-SimpleLruFlush(SlruCtl ctl, bool checkpoint)
+SimpleLruFlush(SlruCtl ctl, bool allow_redirtied)
 {
 	SlruShared	shared = ctl->shared;
 	SlruFlushData fdata;
@@ -1096,11 +1097,11 @@ SimpleLruFlush(SlruCtl ctl, bool checkpoint)
 		SlruInternalWritePage(ctl, slotno, &fdata);
 
 		/*
-		 * When called during a checkpoint, we cannot assert that the slot is
+		 * In some places (e.g. checkpoints), we cannot assert that the slot
-		 * clean now, since another process might have re-dirtied it already.
+		 * is clean now, since another process might have re-dirtied it
-		 * That's okay.
+		 * already.  That's okay.
 		 */
-		Assert(checkpoint ||
+		Assert(allow_redirtied ||
 			   shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
 			   (shared->page_status[slotno] == SLRU_PAGE_VALID &&
 				!shared->page_dirty[slotno]));
@@ -1210,8 +1211,14 @@ restart:;
 	(void) SlruScanDirectory(ctl, SlruScanDirCbDeleteCutoff, &cutoffPage);
 }
 
-void
+/*
-SlruDeleteSegment(SlruCtl ctl, char *filename)
+ * Delete an individual SLRU segment, identified by the filename.
+ *
+ * NB: This does not touch the SLRU buffers themselves, callers have to ensure
+ * they either can't yet contain anything, or have already been cleaned out.
+ */
+static void
+SlruInternalDeleteSegment(SlruCtl ctl, char *filename)
 {
 	char		path[MAXPGPATH];
 
@@ -1221,6 +1228,64 @@ SlruDeleteSegment(SlruCtl ctl, char *filename)
 	unlink(path);
 }
 
+/*
+ * Delete an individual SLRU segment, identified by the segment number.
+ */
+void
+SlruDeleteSegment(SlruCtl ctl, int segno)
+{
+	SlruShared	shared = ctl->shared;
+	int			slotno;
+	char		path[MAXPGPATH];
+	bool		did_write;
+
+	/* Clean out any possibly existing references to the segment. */
+	LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+restart:
+	did_write = false;
+	for (slotno = 0; slotno < shared->num_slots; slotno++)
+	{
+		int			pagesegno = shared->page_number[slotno] / SLRU_PAGES_PER_SEGMENT;
+
+		if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
+			continue;
+
+		/* not the segment we're looking for */
+		if (pagesegno != segno)
+			continue;
+
+		/* If page is clean, just change state to EMPTY (expected case). */
+		if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
+			!shared->page_dirty[slotno])
+		{
+			shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+			continue;
+		}
+
+		/* Same logic as SimpleLruTruncate() */
+		if (shared->page_status[slotno] == SLRU_PAGE_VALID)
+			SlruInternalWritePage(ctl, slotno, NULL);
+		else
+			SimpleLruWaitIO(ctl, slotno);
+
+		did_write = true;
+	}
+
+	/*
+	 * Be extra careful and re-check. The IO functions release the control
+	 * lock, so new pages could have been read in.
+	 */
+	if (did_write)
+		goto restart;
+
+	snprintf(path, MAXPGPATH, "%s/%04X", ctl->Dir, segno);
+	ereport(DEBUG2,
+			(errmsg("removing file \"%s\"", path)));
+	unlink(path);
+
+	LWLockRelease(shared->ControlLock);
+}
+
 /*
  * SlruScanDirectory callback
  *		This callback reports true if there's any segment prior to the one
@@ -1249,7 +1314,7 @@ SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename, int segpage, void *data)
 	int			cutoffPage = *(int *) data;
 
 	if (ctl->PagePrecedes(segpage, cutoffPage))
-		SlruDeleteSegment(ctl, filename);
+		SlruInternalDeleteSegment(ctl, filename);
 
 	return false;				/* keep going */
 }
@@ -1261,7 +1326,7 @@ SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename, int segpage, void *data)
 bool
 SlruScanDirCbDeleteAll(SlruCtl ctl, char *filename, int segpage, void *data)
 {
-	SlruDeleteSegment(ctl, filename);
+	SlruInternalDeleteSegment(ctl, filename);
 
 	return false;				/* keep going */
 }

src/backend/access/transam/xlog.c

@@ -6330,7 +6330,6 @@ StartupXLOG(void)
 	SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
 	SetCommitTsLimit(checkPoint.oldestCommitTs,
 					 checkPoint.newestCommitTs);
-	MultiXactSetSafeTruncate(checkPoint.oldestMulti);
 	XLogCtl->ckptXidEpoch = checkPoint.nextXidEpoch;
 	XLogCtl->ckptXid = checkPoint.nextXid;
 
@@ -6347,10 +6346,8 @@ StartupXLOG(void)
 	StartupReorderBuffer();
 
 	/*
-	 * Startup MultiXact.  We need to do this early for two reasons: one is
+	 * Startup MultiXact. We need to do this early to be able to replay
-	 * that we might try to access multixacts when we do tuple freezing, and
+	 * truncations.
-	 * the other is we need its state initialized because we attempt
-	 * truncation during restartpoints.
 	 */
 	StartupMultiXact();
 
@@ -8507,12 +8504,6 @@ CreateCheckPoint(int flags)
 	 */
 	END_CRIT_SECTION();
 
-	/*
-	 * Now that the checkpoint is safely on disk, we can update the point to
-	 * which multixact can be truncated.
-	 */
-	MultiXactSetSafeTruncate(checkPoint.oldestMulti);
-
 	/*
 	 * Let smgr do post-checkpoint cleanup (eg, deleting old files).
 	 */
@@ -8552,11 +8543,6 @@ CreateCheckPoint(int flags)
 	if (!RecoveryInProgress())
 		TruncateSUBTRANS(GetOldestXmin(NULL, false));
 
-	/*
-	 * Truncate pg_multixact too.
-	 */
-	TruncateMultiXact();
-
 	/* Real work is done, but log and update stats before releasing lock. */
 	LogCheckpointEnd(false);
 
@@ -8886,21 +8872,6 @@ CreateRestartPoint(int flags)
 			ThisTimeLineID = 0;
 	}
 
-	/*
-	 * Due to a historical accident multixact truncations are not WAL-logged,
-	 * but just performed everytime the mxact horizon is increased. So, unless
-	 * we explicitly execute truncations on a standby it will never clean out
-	 * /pg_multixact which obviously is bad, both because it uses space and
-	 * because we can wrap around into pre-existing data...
-	 *
-	 * We can only do the truncation here, after the UpdateControlFile()
-	 * above, because we've now safely established a restart point.  That
-	 * guarantees we will not need to access those multis.
-	 *
-	 * It's probably worth improving this.
-	 */
-	TruncateMultiXact();
-
 	/*
 	 * Truncate pg_subtrans if possible.  We can throw away all data before
 	 * the oldest XMIN of any running transaction.  No future transaction will
@@ -9261,9 +9232,14 @@ xlog_redo(XLogReaderState *record)
 		LWLockRelease(OidGenLock);
 		MultiXactSetNextMXact(checkPoint.nextMulti,
 							  checkPoint.nextMultiOffset);
+
+		/*
+		 * NB: This may perform multixact truncation when replaying WAL
+		 * generated by an older primary.
+		 */
+		MultiXactAdvanceOldest(checkPoint.oldestMulti,
+							   checkPoint.oldestMultiDB);
 		SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
-		SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
-		MultiXactSetSafeTruncate(checkPoint.oldestMulti);
 
 		/*
 		 * If we see a shutdown checkpoint while waiting for an end-of-backup
@@ -9353,14 +9329,17 @@ xlog_redo(XLogReaderState *record)
 		LWLockRelease(OidGenLock);
 		MultiXactAdvanceNextMXact(checkPoint.nextMulti,
 								  checkPoint.nextMultiOffset);
+
+		/*
+		 * NB: This may perform multixact truncation when replaying WAL
+		 * generated by an older primary.
+		 */
+		MultiXactAdvanceOldest(checkPoint.oldestMulti,
+							   checkPoint.oldestMultiDB);
 		if (TransactionIdPrecedes(ShmemVariableCache->oldestXid,
 								  checkPoint.oldestXid))
 			SetTransactionIdLimit(checkPoint.oldestXid,
 								  checkPoint.oldestXidDB);
-		MultiXactAdvanceOldest(checkPoint.oldestMulti,
-							   checkPoint.oldestMultiDB);
-		MultiXactSetSafeTruncate(checkPoint.oldestMulti);
-
 		/* ControlFile->checkPointCopy always tracks the latest ckpt XID */
 		ControlFile->checkPointCopy.nextXidEpoch = checkPoint.nextXidEpoch;
 		ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;

src/backend/commands/vacuum.c

@@ -1137,11 +1137,11 @@ vac_truncate_clog(TransactionId frozenXID,
 		return;
 
 	/*
-	 * Truncate CLOG and CommitTs to the oldest computed value. Note we don't
+	 * Truncate CLOG, multixact and CommitTs to the oldest computed value.
-	 * truncate multixacts; that will be done by the next checkpoint.
 	 */
 	TruncateCLOG(frozenXID);
 	TruncateCommitTs(frozenXID, true);
+	TruncateMultiXact(minMulti, minmulti_datoid, false);
 
 	/*
 	 * Update the wrap limit for GetNewTransactionId and creation of new

src/backend/storage/lmgr/lwlocknames.txt

@@ -45,3 +45,4 @@ ReplicationSlotControlLock			37
 CommitTsControlLock					38
 CommitTsLock						39
 ReplicationOriginLock				40
+MultiXactTruncationLock				41

src/include/access/multixact.h

@@ -71,6 +71,7 @@ typedef struct MultiXactMember
 #define XLOG_MULTIXACT_ZERO_OFF_PAGE	0x00
 #define XLOG_MULTIXACT_ZERO_MEM_PAGE	0x10
 #define XLOG_MULTIXACT_CREATE_ID		0x20
+#define XLOG_MULTIXACT_TRUNCATE_ID		0x30
 
 typedef struct xl_multixact_create
 {
@@ -82,6 +83,21 @@ typedef struct xl_multixact_create
 
 #define SizeOfMultiXactCreate (offsetof(xl_multixact_create, members))
 
+typedef struct xl_multixact_truncate
+{
+	Oid			oldestMultiDB;
+
+	/* to-be-truncated range of multixact offsets */
+	MultiXactId startTruncOff;	/* just for completeness' sake */
+	MultiXactId endTruncOff;
+
+	/* to-be-truncated range of multixact members */
+	MultiXactOffset startTruncMemb;
+	MultiXactOffset endTruncMemb;
+} xl_multixact_truncate;
+
+#define SizeOfMultiXactTruncate (sizeof(xl_multixact_truncate))
+
 
 extern MultiXactId MultiXactIdCreate(TransactionId xid1,
 				  MultiXactStatus status1, TransactionId xid2,
@@ -119,13 +135,12 @@ extern void MultiXactGetCheckptMulti(bool is_shutdown,
 						 Oid *oldestMultiDB);
 extern void CheckPointMultiXact(void);
 extern MultiXactId GetOldestMultiXactId(void);
-extern void TruncateMultiXact(void);
+extern void TruncateMultiXact(MultiXactId oldestMulti, Oid oldestMultiDB, bool in_recovery);
 extern void MultiXactSetNextMXact(MultiXactId nextMulti,
 					  MultiXactOffset nextMultiOffset);
 extern void MultiXactAdvanceNextMXact(MultiXactId minMulti,
 						  MultiXactOffset minMultiOffset);
 extern void MultiXactAdvanceOldest(MultiXactId oldestMulti, Oid oldestMultiDB);
-extern void MultiXactSetSafeTruncate(MultiXactId safeTruncateMulti);
 extern int	MultiXactMemberFreezeThreshold(void);
 
 extern void multixact_twophase_recover(TransactionId xid, uint16 info,

src/include/access/slru.h

@@ -143,14 +143,14 @@ extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
 extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
 						   TransactionId xid);
 extern void SimpleLruWritePage(SlruCtl ctl, int slotno);
-extern void SimpleLruFlush(SlruCtl ctl, bool checkpoint);
+extern void SimpleLruFlush(SlruCtl ctl, bool allow_redirtied);
 extern void SimpleLruTruncate(SlruCtl ctl, int cutoffPage);
 extern bool SimpleLruDoesPhysicalPageExist(SlruCtl ctl, int pageno);
 
 typedef bool (*SlruScanCallback) (SlruCtl ctl, char *filename, int segpage,
 											  void *data);
 extern bool SlruScanDirectory(SlruCtl ctl, SlruScanCallback callback, void *data);
-extern void SlruDeleteSegment(SlruCtl ctl, char *filename);
+extern void SlruDeleteSegment(SlruCtl ctl, int segno);
 
 /* SlruScanDirectory public callbacks */
 extern bool SlruScanDirCbReportPresence(SlruCtl ctl, char *filename,

src/tools/pgindent/typedefs.list

@@ -2750,6 +2750,7 @@ xl_invalid_page
 xl_invalid_page_key
 xl_multi_insert_tuple
 xl_multixact_create
+xl_multixact_truncate
 xl_parameter_change
 xl_relmap_update
 xl_replorigin_drop