Ignore nextOid when replaying an ONLINE checkpoint.
authorTom Lane
Wed, 11 Apr 2018 22:11:29 +0000 (18:11 -0400)
committerTom Lane
Wed, 11 Apr 2018 22:11:29 +0000 (18:11 -0400)
The nextOid value is from the start of the checkpoint and may well be stale
compared to values from more recent XLOG_NEXTOID records.  Previously, we
adopted it anyway, allowing the OID counter to go backwards during a crash.
While this should be harmless, it contributed to the severity of the bug
fixed in commit 0408e1ed5, by allowing duplicate TOAST OIDs to be assigned
immediately following a crash.  Without this error, that issue would only
have arisen when TOAST objects just younger than a multiple of 2^32 OIDs
were deleted and then not vacuumed in time to avoid a conflict.

Pavan Deolasee

Discussion: https://postgr.es/m/CABOikdOgWT2hHkYG3Wwo2cyZJq2zfs1FH0FgX-=h4OLosXHf9w@mail.gmail.com

src/backend/access/transam/xlog.c

index a51df6b0b90df1f470875db2034eddd667a297b7..08dc9ba031ba935c3283881d8adcf965e01c0fbf 100644 (file)
@@ -9785,11 +9785,20 @@ xlog_redo(XLogReaderState *record)
                                  checkPoint.nextXid))
            ShmemVariableCache->nextXid = checkPoint.nextXid;
        LWLockRelease(XidGenLock);
-       /* ... but still treat OID counter as exact */
-       LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
-       ShmemVariableCache->nextOid = checkPoint.nextOid;
-       ShmemVariableCache->oidCount = 0;
-       LWLockRelease(OidGenLock);
+
+       /*
+        * We ignore the nextOid counter in an ONLINE checkpoint, preferring
+        * to track OID assignment through XLOG_NEXTOID records.  The nextOid
+        * counter is from the start of the checkpoint and might well be stale
+        * compared to later XLOG_NEXTOID records.  We could try to take the
+        * maximum of the nextOid counter and our latest value, but since
+        * there's no particular guarantee about the speed with which the OID
+        * counter wraps around, that's a risky thing to do.  In any case,
+        * users of the nextOid counter are required to avoid assignment of
+        * duplicates, so that a somewhat out-of-date value should be safe.
+        */
+
+       /* Handle multixact */
        MultiXactAdvanceNextMXact(checkPoint.nextMulti,
                                  checkPoint.nextMultiOffset);