Fix longstanding problems in VACUUM caused by untimely interruptions

In VACUUM FULL, an interrupt after the initial transaction has been recorded
as committed can cause postmaster to restart with the following error message:
PANIC: cannot abort transaction NNNN, it was already committed
This problem has been reported many times.

In lazy VACUUM, an interrupt after the table has been truncated by
lazy_truncate_heap causes other backends' relcache to still point to the
removed pages; this can cause future INSERT and UPDATE queries to error out
with the following error message:
could not read block XX of relation 1663/NNN/MMMM: read only 0 of 8192 bytes
The window to this race condition is extremely narrow, but it has been seen in
the wild involving a cancelled autovacuum process.

The solution for both problems is to inhibit interrupts in both operations
until after the respective transactions have been committed.  It's not a
complete solution, because the transaction could theoretically be aborted by
some other error, but at least fixes the most common causes of both problems.

Author: alvherre

src/backend/commands/vacuum.c

@@ -13,7 +13,7 @@
  *
  *
  * IDENTIFICATION
- *	  $PostgreSQL: pgsql/src/backend/commands/vacuum.c,v 1.394 2009/10/26 02:26:29 tgl Exp $
+ *	  $PostgreSQL: pgsql/src/backend/commands/vacuum.c,v 1.395 2009/11/10 18:00:06 alvherre Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -217,10 +217,10 @@ static List *get_rel_oids(Oid relid, const RangeVar *vacrel,
 static void vac_truncate_clog(TransactionId frozenXID);
 static void vacuum_rel(Oid relid, VacuumStmt *vacstmt, bool do_toast,
 		   bool for_wraparound, bool *scanned_all);
-static void full_vacuum_rel(Relation onerel, VacuumStmt *vacstmt);
+static bool full_vacuum_rel(Relation onerel, VacuumStmt *vacstmt);
 static void scan_heap(VRelStats *vacrelstats, Relation onerel,
 		  VacPageList vacuum_pages, VacPageList fraged_pages);
-static void repair_frag(VRelStats *vacrelstats, Relation onerel,
+static bool repair_frag(VRelStats *vacrelstats, Relation onerel,
 			VacPageList vacuum_pages, VacPageList fraged_pages,
 			int nindexes, Relation *Irel);
 static void move_chain_tuple(Relation rel,
@@ -1020,6 +1020,7 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, bool do_toast, bool for_wraparound,
 	Oid			toast_relid;
 	Oid			save_userid;
 	bool		save_secdefcxt;
+	bool		heldoff;
 
 	if (scanned_all)
 		*scanned_all = false;
@@ -1186,9 +1187,9 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, bool do_toast, bool for_wraparound,
 	 * Do the actual work --- either FULL or "lazy" vacuum
 	 */
 	if (vacstmt->full)
-		full_vacuum_rel(onerel, vacstmt);
+		heldoff = full_vacuum_rel(onerel, vacstmt);
 	else
-		lazy_vacuum_rel(onerel, vacstmt, vac_strategy, scanned_all);
+		heldoff = lazy_vacuum_rel(onerel, vacstmt, vac_strategy, scanned_all);
 
 	/* Restore userid */
 	SetUserIdAndContext(save_userid, save_secdefcxt);
@@ -1202,6 +1203,10 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, bool do_toast, bool for_wraparound,
 	PopActiveSnapshot();
 	CommitTransactionCommand();
 
+	/* now we can allow interrupts again, if disabled */
+	if (heldoff)
+		RESUME_INTERRUPTS();
+
 	/*
 	 * If the relation has a secondary toast rel, vacuum that too while we
 	 * still hold the session lock on the master table.  Note however that
@@ -1235,8 +1240,11 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, bool do_toast, bool for_wraparound,
  *
  *		At entry, we have already established a transaction and opened
  *		and locked the relation.
+ *
+ *		The return value indicates whether this function has held off
+ *		interrupts -- caller must RESUME_INTERRUPTS() after commit if true.
  */
-static void
+static bool
 full_vacuum_rel(Relation onerel, VacuumStmt *vacstmt)
 {
 	VacPageListData vacuum_pages;		/* List of pages to vacuum and/or
@@ -1247,6 +1255,7 @@ full_vacuum_rel(Relation onerel, VacuumStmt *vacstmt)
 	int			nindexes,
 				i;
 	VRelStats  *vacrelstats;
+	bool		heldoff = false;
 
 	vacuum_set_xid_limits(vacstmt->freeze_min_age, vacstmt->freeze_table_age,
 						  onerel->rd_rel->relisshared,
@@ -1298,7 +1307,7 @@ full_vacuum_rel(Relation onerel, VacuumStmt *vacstmt)
 	if (fraged_pages.num_pages > 0)
 	{
 		/* Try to shrink heap */
-		repair_frag(vacrelstats, onerel, &vacuum_pages, &fraged_pages,
+		heldoff = repair_frag(vacrelstats, onerel, &vacuum_pages, &fraged_pages,
 					nindexes, Irel);
 		vac_close_indexes(nindexes, Irel, NoLock);
 	}
@@ -1324,6 +1333,8 @@ full_vacuum_rel(Relation onerel, VacuumStmt *vacstmt)
 	/* report results to the stats collector, too */
 	pgstat_report_vacuum(RelationGetRelid(onerel), onerel->rd_rel->relisshared,
 						 true, vacstmt->analyze, vacrelstats->rel_tuples);
+
+	return heldoff;
 }
 
 
@@ -1874,8 +1885,11 @@ scan_heap(VRelStats *vacrelstats, Relation onerel,
  *		for them after committing (in hack-manner - without losing locks
  *		and freeing memory!) current transaction. It truncates relation
  *		if some end-blocks are gone away.
+ *
+ *		The return value indicates whether this function has held off
+ *		interrupts -- caller must RESUME_INTERRUPTS() after commit if true.
  */
-static void
+static bool
 repair_frag(VRelStats *vacrelstats, Relation onerel,
 			VacPageList vacuum_pages, VacPageList fraged_pages,
 			int nindexes, Relation *Irel)
@@ -1900,6 +1914,7 @@ repair_frag(VRelStats *vacrelstats, Relation onerel,
 	int			keep_tuples = 0;
 	int			keep_indexed_tuples = 0;
 	PGRUsage	ru0;
+	bool		heldoff = false;
 
 	pg_rusage_init(&ru0);
 
@@ -2705,8 +2720,12 @@ repair_frag(VRelStats *vacrelstats, Relation onerel,
 		 *
 		 * XXX This desperately needs to be revisited.	Any failure after this
 		 * point will result in a PANIC "cannot abort transaction nnn, it was
-		 * already committed"!
+		 * already committed"!  As a precaution, we prevent cancel interrupts
+		 * after this point to mitigate this problem; caller is responsible for
+		 * re-enabling them after committing the transaction.
 		 */
+		HOLD_INTERRUPTS();
+		heldoff = true;
 		ForceSyncCommit();
 		(void) RecordTransactionCommit();
 	}
@@ -2906,6 +2925,8 @@ repair_frag(VRelStats *vacrelstats, Relation onerel,
 		pfree(vacrelstats->vtlinks);
 
 	ExecContext_Finish(&ec);
+
+	return heldoff;
 }
 
 /*

src/backend/commands/vacuumlazy.c

@@ -29,7 +29,7 @@
  *
  *
  * IDENTIFICATION
- *	  $PostgreSQL: pgsql/src/backend/commands/vacuumlazy.c,v 1.122 2009/08/24 02:18:32 tgl Exp $
+ *	  $PostgreSQL: pgsql/src/backend/commands/vacuumlazy.c,v 1.123 2009/11/10 18:00:06 alvherre Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -140,8 +140,11 @@ static int	vac_cmp_itemptr(const void *left, const void *right);
  *
  *		At entry, we have already established a transaction and opened
  *		and locked the relation.
+ *
+ *		The return value indicates whether this function has held off
+ *		interrupts -- caller must RESUME_INTERRUPTS() after commit if true.
  */
-void
+bool
 lazy_vacuum_rel(Relation onerel, VacuumStmt *vacstmt,
 				BufferAccessStrategy bstrategy, bool *scanned_all)
 {
@@ -153,6 +156,7 @@ lazy_vacuum_rel(Relation onerel, VacuumStmt *vacstmt,
 	TimestampTz starttime = 0;
 	bool		scan_all;
 	TransactionId freezeTableLimit;
+	bool		heldoff = false;
 
 	pg_rusage_init(&ru0);
 
@@ -194,12 +198,22 @@ lazy_vacuum_rel(Relation onerel, VacuumStmt *vacstmt,
 	 *
 	 * Don't even think about it unless we have a shot at releasing a goodly
 	 * number of pages.  Otherwise, the time taken isn't worth it.
+	 *
+	 * Note that after we've truncated the heap, it's too late to abort the
+	 * transaction; doing so would lose the sinval messages needed to tell
+	 * the other backends about the table being shrunk.  We prevent interrupts
+	 * in that case; caller is responsible for re-enabling them after
+	 * committing the transaction.
 	 */
 	possibly_freeable = vacrelstats->rel_pages - vacrelstats->nonempty_pages;
 	if (possibly_freeable > 0 &&
 		(possibly_freeable >= REL_TRUNCATE_MINIMUM ||
 		 possibly_freeable >= vacrelstats->rel_pages / REL_TRUNCATE_FRACTION))
+	{
+		HOLD_INTERRUPTS();
+		heldoff = true;
 		lazy_truncate_heap(onerel, vacrelstats);
+	}
 
 	/* Vacuum the Free Space Map */
 	FreeSpaceMapVacuum(onerel);
@@ -246,6 +260,8 @@ lazy_vacuum_rel(Relation onerel, VacuumStmt *vacstmt,
 
 	if (scanned_all)
 		*scanned_all = vacrelstats->scanned_all;
+
+	return heldoff;
 }
 
 

src/include/commands/vacuum.h

@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2009, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $PostgreSQL: pgsql/src/include/commands/vacuum.h,v 1.85 2009/06/11 14:49:11 momjian Exp $
+ * $PostgreSQL: pgsql/src/include/commands/vacuum.h,v 1.86 2009/11/10 18:00:06 alvherre Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -146,7 +146,7 @@ extern bool vac_is_partial_index(Relation indrel);
 extern void vacuum_delay_point(void);
 
 /* in commands/vacuumlazy.c */
-extern void lazy_vacuum_rel(Relation onerel, VacuumStmt *vacstmt,
+extern bool lazy_vacuum_rel(Relation onerel, VacuumStmt *vacstmt,
 				BufferAccessStrategy bstrategy, bool *scanned_all);
 
 /* in commands/analyze.c */