Have multixact be truncated by checkpoint, not vacuum

Instead of truncating pg_multixact at vacuum time, do it only at
checkpoint time.  The reason for doing it this way is twofold: first, we
want it to delete only segments that we're certain will not be required
if there's a crash immediately after the removal; and second, we want to
do it relatively often so that older files are not left behind if
there's an untimely crash.

Per my proposal in
http://www.postgresql.org/message-id/20140626044519.GJ7340@eldon.alvh.no-ip.org
we now execute the truncation in the checkpointer process rather than as
part of vacuum.  Vacuum is in only charge of maintaining in shared
memory the value to which it's possible to truncate the files; that
value is stored as part of checkpoints also, and so upon recovery we can
reuse the same value to re-execute truncate and reset the
oldest-value-still-safe-to-use to one known to remain after truncation.

Per bug reported by Jeff Janes in the course of his tests involving
bug #8673.

While at it, update some comments that hadn't been updated since
multixacts were changed.

Backpatch to 9.3, where persistency of pg_multixact files was
introduced by commit 0ac5ad5.

Author: alvherre

src/backend/access/transam/multixact.c

@@ -45,14 +45,17 @@
  * anything we saw during replay.
  *
  * We are able to remove segments no longer necessary by carefully tracking
- * each table's used values: during vacuum, any multixact older than a
- * certain value is removed; the cutoff value is stored in pg_class.
- * The minimum value in each database is stored in pg_database, and the
- * global minimum is part of pg_control.  Any vacuum that is able to
- * advance its database's minimum value also computes a new global minimum,
- * and uses this value to truncate older segments.  When new multixactid
- * values are to be created, care is taken that the counter does not
- * fall within the wraparound horizon considering the global minimum value.
+ * each table's used values: during vacuum, any multixact older than a certain
+ * 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
+ * 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
+ * minimum value.
  *
  * Portions Copyright (c) 1996-2014, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
@@ -91,7 +94,7 @@
  * Note: because MultiXactOffsets are 32 bits and wrap around at 0xFFFFFFFF,
  * MultiXact page numbering also wraps around at
  * 0xFFFFFFFF/MULTIXACT_OFFSETS_PER_PAGE, and segment numbering at
- * 0xFFFFFFFF/MULTIXACT_OFFSETS_PER_PAGE/SLRU_SEGMENTS_PER_PAGE.    We need
+ * 0xFFFFFFFF/MULTIXACT_OFFSETS_PER_PAGE/SLRU_PAGES_PER_SEGMENT.  We need
  * take no explicit notice of that fact in this module, except when comparing
  * segment and page numbers in TruncateMultiXact (see
  * MultiXactOffsetPagePrecedes).
@@ -188,16 +191,20 @@ typedef struct MultiXactStateData
 	/* next-to-be-assigned offset */
 	MultiXactOffset nextOffset;
 
-	/* the Offset SLRU area was last truncated at this MultiXactId */
-	MultiXactId lastTruncationPoint;
-
 	/*
-	 * oldest multixact that is still on disk.  Anything older than this
-	 * should not be consulted.
+	 * Oldest multixact that is still on disk.  Anything older than this
+	 * should not be consulted.  These values are updated by vacuum.
 	 */
 	MultiXactId oldestMultiXactId;
 	Oid			oldestMultiXactDB;
 
+	/*
+	 * This is what the previous checkpoint stored as the truncate position.
+	 * This value is the oldestMultiXactId that was valid when a checkpoint
+	 * was last executed.
+	 */
+	MultiXactId lastCheckpointedOldest;
+
 	/* support for anti-wraparound measures */
 	MultiXactId multiVacLimit;
 	MultiXactId multiWarnLimit;
@@ -234,12 +241,20 @@ typedef struct MultiXactStateData
 	 * than its own OldestVisibleMXactId[] setting; this is necessary because
 	 * the checkpointer could truncate away such data at any instant.
 	 *
-	 * The checkpointer can compute the safe truncation point as the oldest
-	 * valid value among all the OldestMemberMXactId[] and
-	 * OldestVisibleMXactId[] entries, or nextMXact if none are valid.
-	 * Clearly, it is not possible for any later-computed OldestVisibleMXactId
-	 * value to be older than this, and so there is no risk of truncating data
-	 * that is still needed.
+	 * The oldest valid value among all of the OldestMemberMXactId[] and
+	 * OldestVisibleMXactId[] entries is considered by vacuum as the earliest
+	 * possible value still having any live member transaction.  Subtracting
+	 * vacuum_multixact_freeze_min_age from that value we obtain the freezing
+	 * point for multixacts for that table.  Any value older than that is
+	 * removed from tuple headers (or "frozen"; see FreezeMultiXactId.  Note
+	 * that multis that have member xids that are older than the cutoff point
+	 * for xids must also be frozen, even if the multis themselves are newer
+	 * than the multixid cutoff point).  Whenever a full table vacuum happens,
+	 * the freezing point so computed is used as the new pg_class.relminmxid
+	 * value.  The minimum of all those values in a database is stored as
+	 * pg_database.datminmxid.  In turn, the minimum of all of those values is
+	 * stored in pg_control and used as truncation point for pg_multixact.  At
+	 * checkpoint or restartpoint, unneeded segments are removed.
 	 */
 	MultiXactId perBackendXactIds[1];	/* VARIABLE LENGTH ARRAY */
 } MultiXactStateData;
@@ -1121,8 +1136,8 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
 	 * We check known limits on MultiXact before resorting to the SLRU area.
 	 *
 	 * An ID older than MultiXactState->oldestMultiXactId cannot possibly be
-	 * useful; it should have already been removed by vacuum.  We've truncated
-	 * the on-disk structures anyway.  Returning the wrong values could lead
+	 * useful; it has already been removed, or will be removed shortly, by
+	 * truncation.  Returning the wrong values could lead
 	 * to an incorrect visibility result.  However, to support pg_upgrade we
 	 * need to allow an empty set to be returned regardless, if the caller is
 	 * willing to accept it; the caller is expected to check that it's an
@@ -1932,14 +1947,14 @@ TrimMultiXact(void)
 	LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
 
 	/*
-	 * (Re-)Initialize our idea of the latest page number.
+	 * (Re-)Initialize our idea of the latest page number for offsets.
 	 */
 	pageno = MultiXactIdToOffsetPage(multi);
 	MultiXactOffsetCtl->shared->latest_page_number = pageno;
 
 	/*
 	 * Zero out the remainder of the current offsets page.  See notes in
-	 * StartupCLOG() for motivation.
+	 * TrimCLOG() for motivation.
 	 */
 	entryno = MultiXactIdToOffsetEntry(multi);
 	if (entryno != 0)
@@ -1962,7 +1977,7 @@ TrimMultiXact(void)
 	LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
 
 	/*
-	 * (Re-)Initialize our idea of the latest page number.
+	 * (Re-)Initialize our idea of the latest page number for members.
 	 */
 	pageno = MXOffsetToMemberPage(offset);
 	MultiXactMemberCtl->shared->latest_page_number = pageno;
@@ -2240,6 +2255,18 @@ MultiXactAdvanceOldest(MultiXactId oldestMulti, Oid oldestMultiDB)
 		SetMultiXactIdLimit(oldestMulti, oldestMultiDB);
 }
 
+/*
+ * Update the "safe truncation point".  This is the newest value of oldestMulti
+ * that is known to be flushed as part of a checkpoint record.
+ */
+void
+MultiXactSetSafeTruncate(MultiXactId safeTruncateMulti)
+{
+	LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
+	MultiXactState->lastCheckpointedOldest = safeTruncateMulti;
+	LWLockRelease(MultiXactGenLock);
+}
+
 /*
  * Make sure that MultiXactOffset has room for a newly-allocated MultiXactId.
  *
@@ -2478,25 +2505,31 @@ SlruScanDirCbFindEarliest(SlruCtl ctl, char *filename, int segpage, void *data)
  * Remove all MultiXactOffset and MultiXactMember segments before the oldest
  * ones still of interest.
  *
- * On a primary, this is called by vacuum after it has successfully advanced a
- * database's datminmxid value; the cutoff value we're passed is the minimum of
- * all databases' datminmxid values.
- *
- * During crash recovery, it's called from CreateRestartPoint() instead.  We
- * rely on the fact that xlog_redo() will already have called
- * MultiXactAdvanceOldest().  Our latest_page_number will already have been
- * initialized by StartupMultiXact() and kept up to date as new pages are
- * zeroed.
+ * On a primary, this is called by the checkpointer process after a checkpoint
+ * has been flushed; during crash recovery, it's called from
+ * CreateRestartPoint().  In the latter case, we rely on the fact that
+ * xlog_redo() will already have called MultiXactAdvanceOldest().  Our
+ * latest_page_number will already have been initialized by StartupMultiXact()
+ * and kept up to date as new pages are zeroed.
  */
 void
-TruncateMultiXact(MultiXactId oldestMXact)
+TruncateMultiXact(void)
 {
+	MultiXactId		oldestMXact;
 	MultiXactOffset oldestOffset;
 	MultiXactOffset nextOffset;
 	mxtruncinfo trunc;
 	MultiXactId earliest;
 	MembersLiveRange range;
 
+	Assert(AmCheckpointerProcess() || AmStartupProcess() ||
+		   !IsPostmasterEnvironment);
+
+	LWLockAcquire(MultiXactGenLock, LW_SHARED);
+	oldestMXact = MultiXactState->lastCheckpointedOldest;
+	LWLockRelease(MultiXactGenLock);
+	Assert(MultiXactIdIsValid(oldestMXact));
+
 	/*
 	 * Note we can't just plow ahead with the truncation; it's possible that
 	 * there are no segments to truncate, which is a problem because we are
@@ -2507,15 +2540,16 @@ TruncateMultiXact(MultiXactId oldestMXact)
 	trunc.earliestExistingPage = -1;
 	SlruScanDirectory(MultiXactOffsetCtl, SlruScanDirCbFindEarliest, &trunc);
 	earliest = trunc.earliestExistingPage * MULTIXACT_OFFSETS_PER_PAGE;
+	if (earliest < FirstMultiXactId)
+		earliest = FirstMultiXactId;
 
 	/* nothing to do */
 	if (MultiXactIdPrecedes(oldestMXact, earliest))
 		return;
 
 	/*
 	 * First, compute the safe truncation point for MultiXactMember. This is
-	 * the starting offset of the multixact we were passed as MultiXactOffset
-	 * cutoff.
+	 * the starting offset of the oldest multixact.
 	 */
 	{
 		int			pageno;
@@ -2538,10 +2572,6 @@ TruncateMultiXact(MultiXactId oldestMXact)
 		LWLockRelease(MultiXactOffsetControlLock);
 	}
 
-	/* truncate MultiXactOffset */
-	SimpleLruTruncate(MultiXactOffsetCtl,
-					  MultiXactIdToOffsetPage(oldestMXact));
-
 	/*
 	 * To truncate MultiXactMembers, we need to figure out the active page
 	 * range and delete all files outside that range.  The start point is the
@@ -2559,6 +2589,11 @@ TruncateMultiXact(MultiXactId oldestMXact)
 	range.rangeEnd = MXOffsetToMemberPage(nextOffset);
 
 	SlruScanDirectory(MultiXactMemberCtl, SlruScanDirCbRemoveMembers, &range);
+
+	/* Now we can truncate MultiXactOffset */
+	SimpleLruTruncate(MultiXactOffsetCtl,
+					  MultiXactIdToOffsetPage(oldestMXact));
+
 }
 
 /*

src/backend/access/transam/xlog.c

@@ -6264,6 +6264,7 @@ StartupXLOG(void)
 	MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
 	SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
 	SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
+	MultiXactSetSafeTruncate(checkPoint.oldestMulti);
 	XLogCtl->ckptXidEpoch = checkPoint.nextXidEpoch;
 	XLogCtl->ckptXid = checkPoint.nextXid;
 
@@ -8272,6 +8273,12 @@ 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).
 	 */
@@ -8305,6 +8312,11 @@ 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);
 
@@ -8578,21 +8590,6 @@ CreateRestartPoint(int flags)
 	}
 	LWLockRelease(ControlFileLock);
 
-	/*
-	 * Due to an 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 need to access those multis.
-	 *
-	 * It's probably worth improving this.
-	 */
-	TruncateMultiXact(lastCheckPoint.oldestMulti);
-
 	/*
 	 * Delete old log files (those no longer needed even for previous
 	 * checkpoint/restartpoint) to prevent the disk holding the xlog from
@@ -8651,6 +8648,21 @@ CreateRestartPoint(int flags)
 			ThisTimeLineID = 0;
 	}
 
+	/*
+	 * Due to an 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
@@ -9117,6 +9129,7 @@ xlog_redo(XLogRecPtr lsn, XLogRecord *record)
 							  checkPoint.nextMultiOffset);
 		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
@@ -9217,6 +9230,7 @@ xlog_redo(XLogRecPtr lsn, XLogRecord *record)
 								  checkPoint.oldestXidDB);
 		MultiXactAdvanceOldest(checkPoint.oldestMulti,
 							   checkPoint.oldestMultiDB);
+		MultiXactSetSafeTruncate(checkPoint.oldestMulti);
 
 		/* ControlFile->checkPointCopy always tracks the latest ckpt XID */
 		ControlFile->checkPointCopy.nextXidEpoch = checkPoint.nextXidEpoch;

src/backend/commands/vacuum.c

@@ -969,9 +969,11 @@ vac_truncate_clog(TransactionId frozenXID, MultiXactId minMulti)
 		return;
 	}
 
-	/* Truncate CLOG and Multi to the oldest computed value */
+	/*
+	 * Truncate CLOG to the oldest computed value.  Note we don't truncate
+	 * multixacts; that will be done by the next checkpoint.
+	 */
 	TruncateCLOG(frozenXID);
-	TruncateMultiXact(minMulti);
 
 	/*
 	 * Update the wrap limit for GetNewTransactionId and creation of new
@@ -980,7 +982,7 @@ vac_truncate_clog(TransactionId frozenXID, MultiXactId minMulti)
 	 * signalling twice?
 	 */
 	SetTransactionIdLimit(frozenXID, oldestxid_datoid);
-	MultiXactAdvanceOldest(minMulti, minmulti_datoid);
+	SetMultiXactIdLimit(minMulti, minmulti_datoid);
 }
 
 

src/include/access/multixact.h

@@ -119,12 +119,13 @@ extern void MultiXactGetCheckptMulti(bool is_shutdown,
 						 Oid *oldestMultiDB);
 extern void CheckPointMultiXact(void);
 extern MultiXactId GetOldestMultiXactId(void);
-extern void TruncateMultiXact(MultiXactId cutoff_multi);
+extern void TruncateMultiXact(void);
 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 void multixact_twophase_recover(TransactionId xid, uint16 info,
 						   void *recdata, uint32 len);