Ensure snapshot is registered within ScanPgRelation().
authorAndres Freund <andres@anarazel.de>
Sat, 28 Mar 2020 18:52:11 +0000 (11:52 -0700)
committerAndres Freund <andres@anarazel.de>
Sat, 28 Mar 2020 19:26:46 +0000 (12:26 -0700)
In 9.4 I added support to use a historical snapshot in
ScanPgRelation(), while adding logical decoding. Unfortunately a
conflict with the concurrent removal of SnapshotNow was incorrectly
resolved, leading to an unregistered snapshot being used.

It is not correct to use an unregistered (or non-active) snapshot for
anything non-trivial, because catalog invalidations can cause the
snapshot to be invalidated.

Luckily it seems unlikely to actively cause problems in practice, as
ScanPgRelation() requires that we already have a lock on the relation,
we only look for a single row, and we don't appear to rely on the
result's tid to be correct. It however is clearly wrong and potential
negative consequences would likely be hard to find. So it seems worth
backpatching the fix, even without a concrete hazard.

Discussion: https://postgr.es/m/20200229052459.wzhqnbhrriezg4v2@alap3.anarazel.de
Backpatch: 9.5-

src/backend/utils/cache/relcache.c

index 76f41dbe36cc779edc73de1382a2180f9abd46c8..782af9aeed29be12cd1c713b8a73d88a131b4987 100644 (file)
@@ -313,7 +313,7 @@ ScanPgRelation(Oid targetRelId, bool indexOK, bool force_non_historic)
        Relation        pg_class_desc;
        SysScanDesc pg_class_scan;
        ScanKeyData key[1];
-       Snapshot        snapshot;
+       Snapshot        snapshot = NULL;
 
        /*
         * If something goes wrong during backend startup, we might find ourselves
@@ -343,12 +343,12 @@ ScanPgRelation(Oid targetRelId, bool indexOK, bool force_non_historic)
        /*
         * The caller might need a tuple that's newer than the one the historic
         * snapshot; currently the only case requiring to do so is looking up the
-        * relfilenode of non mapped system relations during decoding.
+        * relfilenode of non mapped system relations during decoding. That
+        * snapshot cant't change in the midst of a relcache build, so there's no
+        * need to register the snapshot.
         */
        if (force_non_historic)
                snapshot = GetNonHistoricCatalogSnapshot(RelationRelationId);
-       else
-               snapshot = GetCatalogSnapshot(RelationRelationId);
 
        pg_class_scan = systable_beginscan(pg_class_desc, ClassOidIndexId,
                                                                           indexOK && criticalRelcachesBuilt,