Replace the former method of determining snapshot xmax --- to wit, calling

ReadNewTransactionId from GetSnapshotData --- with a "latestCompletedXid"
variable that is updated during transaction commit or abort.  Since
latestCompletedXid is written only in places that had to lock ProcArrayLock
exclusively anyway, and is read only in places that had to lock ProcArrayLock
shared anyway, it adds no new locking requirements to the system despite being
cluster-wide.  Moreover, removing ReadNewTransactionId from snapshot
acquisition eliminates the need to take both XidGenLock and ProcArrayLock at
the same time.  Since XidGenLock is sometimes held across I/O this can be a
significant win.  Some preliminary benchmarking suggested that this patch has
no effect on average throughput but can significantly improve the worst-case
transaction times seen in pgbench.  Concept by Florian Pflug, implementation
by Tom Lane.

src/backend/commands/vacuum.c

@@ -13,7 +13,7 @@
  *
  *
  * IDENTIFICATION
- *	  $PostgreSQL: pgsql/src/backend/commands/vacuum.c,v 1.356 2007/09/05 18:10:47 tgl Exp $
+ *	  $PostgreSQL: pgsql/src/backend/commands/vacuum.c,v 1.357 2007/09/08 20:31:14 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -2437,14 +2437,15 @@ repair_frag(VRelStats *vacrelstats, Relation onerel,
 		 * For now, a quick hack: record status of current transaction as
 		 * committed, and continue.  We force the commit to be synchronous
 		 * so that it's down to disk before we truncate.  (Note: tqual.c
-		 * knows that VACUUM FULL always uses sync commit, too.)
+		 * knows that VACUUM FULL always uses sync commit, too.)  The
+		 * transaction continues to be shown as running in the ProcArray.
 		 *
 		 * 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"!
 		 */
 		ForceSyncCommit();
-		RecordTransactionCommit();
+		(void) RecordTransactionCommit();
 	}
 
 	/*