Fix reindexing of pg_class indexes some more.

Commits 3dbb317d3 et al failed under CLOBBER_CACHE_ALWAYS testing.
Investigation showed that to reindex pg_class_oid_index, we must
suppress accesses to the index (via SetReindexProcessing) before we call
RelationSetNewRelfilenode, or at least before we do CommandCounterIncrement
therein; otherwise, relcache reloads happening within the CCI may try to
fetch pg_class rows using the index's new relfilenode value, which is as
yet an empty file.

Of course, the point of 3dbb317d3 was that that ordering didn't work
either, because then RelationSetNewRelfilenode's own update of the index's
pg_class row cannot access the index, should it need to.

There are various ways we might have got around that, but Andres Freund
came up with a brilliant solution: for a mapped index, we can really just
skip the pg_class update altogether.  The only fields it was actually
changing were relpages etc, but it was just setting them to zeroes which
is useless make-work.  (Correct new values will be installed at the end
of index build.)  All pg_class indexes are mapped and probably always will
be, so this eliminates the problem by removing work rather than adding it,
always a pleasant outcome.  Having taught RelationSetNewRelfilenode to do
it that way, we can revert the code reordering in reindex_index.  (But
I left the moved setup code where it was; there seems no reason why it
has to run without use of the old index.  If you're trying to fix a
busted pg_class index, you'll have had to disable system index use
altogether to get this far.)

Moreover, this means we don't need RelationSetIndexList at all, because
reindex_relation's hacking to make "REINDEX TABLE pg_class" work is
likewise now unnecessary.  We'll leave that code in place in the back
branches, but a follow-on patch will remove it in HEAD.

In passing, do some minor cleanup for commit 5c1560606 (in HEAD only),
notably removing a duplicate newrnode assignment.

Patch by me, using a core idea due to Andres Freund.  Back-patch to all
supported branches, as 3dbb317d3 was.

Discussion: https://postgr.es/m/28926.1556664156@sss.pgh.pa.us

diff --git a/src/backend/catalog/index.c b/src/backend/catalog/index.c
index 071422c6b3266025724da8cf86c725118489ced0..a6b56d31007569ba64738c9f1613004e4efd900f 100644 (file)
--- a/src/backend/catalog/index.c
+++ b/src/backend/catalog/index.c
@@ -3377,21 +3377,16 @@ reindex_index(Oid indexId, bool skip_constraint_checks, char persistence,
                indexInfo->ii_ExclusionStrats = NULL;
        }
 
-       /*
-        * Build a new physical relation for the index. Need to do that before
-        * "officially" starting the reindexing with SetReindexProcessing -
-        * otherwise the necessary pg_class changes cannot be made with
-        * encountering assertions.
-        */
-       RelationSetNewRelfilenode(iRel, persistence, InvalidTransactionId,
-                                                         InvalidMultiXactId);
-
        /* ensure SetReindexProcessing state isn't leaked */
        PG_TRY();
        {
                /* Suppress use of the target index while rebuilding it */
                SetReindexProcessing(heapId, indexId);
 
+               /* Create a new physical relation for the index */
+               RelationSetNewRelfilenode(iRel, persistence, InvalidTransactionId,
+                                                                 InvalidMultiXactId);
+
                /* Initialize the index and rebuild */
                /* Note: we do not need to re-establish pkey setting */
                index_build(heapRelation, iRel, indexInfo, false, true);
@@ -3537,7 +3532,6 @@ reindex_relation(Oid relid, int flags, int options)
        Relation        rel;
        Oid                     toast_relid;
        List       *indexIds;
-       bool            is_pg_class;
        bool            result;
 
        /*
@@ -3556,37 +3550,8 @@ reindex_relation(Oid relid, int flags, int options)
         */
        indexIds = RelationGetIndexList(rel);
 
-       /*
-        * reindex_index will attempt to update the pg_class rows for the relation
-        * and index.  If we are processing pg_class itself, we want to make sure
-        * that the updates do not try to insert index entries into indexes we
-        * have not processed yet.  (When we are trying to recover from corrupted
-        * indexes, that could easily cause a crash.) We can accomplish this
-        * because CatalogTupleInsert/CatalogTupleUpdate will use the relcache's
-        * index list to know which indexes to update. We just force the index
-        * list to be only the stuff we've processed.
-        *
-        * It is okay to not insert entries into the indexes we have not processed
-        * yet because all of this is transaction-safe.  If we fail partway
-        * through, the updated rows are dead and it doesn't matter whether they
-        * have index entries.  Also, a new pg_class index will be created with a
-        * correct entry for its own pg_class row because we do
-        * RelationSetNewRelfilenode() before we do index_build().
-        *
-        * Note that we also clear pg_class's rd_oidindex until the loop is done,
-        * so that that index can't be accessed either.  This means we cannot
-        * safely generate new relation OIDs while in the loop; shouldn't be a
-        * problem.
-        */
-       is_pg_class = (RelationGetRelid(rel) == RelationRelationId);
-
-       /* Ensure rd_indexattr is valid; see comments for RelationSetIndexList */
-       if (is_pg_class)
-               (void) RelationGetIndexAttrBitmap(rel, INDEX_ATTR_BITMAP_ALL);
-
        PG_TRY();
        {
-               List       *doneIndexes;
                ListCell   *indexId;
                char            persistence;
 
@@ -3614,14 +3579,10 @@ reindex_relation(Oid relid, int flags, int options)
                        persistence = rel->rd_rel->relpersistence;
 
                /* Reindex all the indexes. */
-               doneIndexes = NIL;
                foreach(indexId, indexIds)
                {
                        Oid                     indexOid = lfirst_oid(indexId);
 
-                       if (is_pg_class)
-                               RelationSetIndexList(rel, doneIndexes, InvalidOid);
-
                        reindex_index(indexOid, !(flags & REINDEX_REL_CHECK_CONSTRAINTS),
                                                  persistence, options);
 
@@ -3629,9 +3590,6 @@ reindex_relation(Oid relid, int flags, int options)
 
                        /* Index should no longer be in the pending list */
                        Assert(!ReindexIsProcessingIndex(indexOid));
-
-                       if (is_pg_class)
-                               doneIndexes = lappend_oid(doneIndexes, indexOid);
                }
        }
        PG_CATCH();
@@ -3643,9 +3601,6 @@ reindex_relation(Oid relid, int flags, int options)
        PG_END_TRY();
        ResetReindexPending();
 
-       if (is_pg_class)
-               RelationSetIndexList(rel, indexIds, ClassOidIndexId);
-
        /*
         * Close rel, but continue to hold the lock.
         */

diff --git a/src/backend/utils/cache/relcache.c b/src/backend/utils/cache/relcache.c
index 9f49cf1b313c4c87e508237040142e62fff5552e..65b74cb6d8788c4975c8ae0076f09e8e0293625b 100644 (file)
--- a/src/backend/utils/cache/relcache.c
+++ b/src/backend/utils/cache/relcache.c
@@ -3501,38 +3501,67 @@ RelationSetNewRelfilenode(Relation relation, char persistence,
        RelationDropStorage(relation);
 
        /*
-        * Now update the pg_class row.  However, if we're dealing with a mapped
-        * index, pg_class.relfilenode doesn't change; instead we have to send the
-        * update to the relation mapper.
+        * If we're dealing with a mapped index, pg_class.relfilenode doesn't
+        * change; instead we have to send the update to the relation mapper.
+        *
+        * For mapped indexes, we don't actually change the pg_class entry at all;
+        * this is essential when reindexing pg_class itself.  That leaves us with
+        * possibly-inaccurate values of relpages etc, but those will be fixed up
+        * later.
         */
        if (RelationIsMapped(relation))
+       {
+               /* This case is only supported for indexes */
+               Assert(relation->rd_rel->relkind == RELKIND_INDEX);
+
+               /* Since we're not updating pg_class, these had better not change */
+               Assert(classform->relfrozenxid == freezeXid);
+               Assert(classform->relminmxid == minmulti);
+               Assert(classform->relpersistence == persistence);
+
+               /*
+                * In some code paths it's possible that the tuple update we'd
+                * otherwise do here is the only thing that would assign an XID for
+                * the current transaction.  However, we must have an XID to delete
+                * files, so make sure one is assigned.
+                */
+               (void) GetCurrentTransactionId();
+
+               /* Do the deed */
                RelationMapUpdateMap(RelationGetRelid(relation),
                                                         newrelfilenode,
                                                         relation->rd_rel->relisshared,
                                                         false);
+
+               /* Since we're not updating pg_class, must trigger inval manually */
+               CacheInvalidateRelcache(relation);
+       }
        else
+       {
+               /* Normal case, update the pg_class entry */
                classform->relfilenode = newrelfilenode;
 
-       /* These changes are safe even for a mapped relation */
-       if (relation->rd_rel->relkind != RELKIND_SEQUENCE)
-       {
-               classform->relpages = 0;        /* it's empty until further notice */
-               classform->reltuples = 0;
-               classform->relallvisible = 0;
-       }
-       classform->relfrozenxid = freezeXid;
-       classform->relminmxid = minmulti;
-       classform->relpersistence = persistence;
+               /* relpages etc. never change for sequences */
+               if (relation->rd_rel->relkind != RELKIND_SEQUENCE)
+               {
+                       classform->relpages = 0;        /* it's empty until further notice */
+                       classform->reltuples = 0;
+                       classform->relallvisible = 0;
+               }
+               classform->relfrozenxid = freezeXid;
+               classform->relminmxid = minmulti;
+               classform->relpersistence = persistence;
 
-       CatalogTupleUpdate(pg_class, &tuple->t_self, tuple);
+               CatalogTupleUpdate(pg_class, &tuple->t_self, tuple);
+       }
 
        heap_freetuple(tuple);
 
        heap_close(pg_class, RowExclusiveLock);
 
        /*
-        * Make the pg_class row change visible, as well as the relation map
-        * change if any.  This will cause the relcache entry to get updated, too.
+        * Make the pg_class row change or relation map change visible.  This will
+        * cause the relcache entry to get updated, too.
         */
        CommandCounterIncrement();