crash in DROP INDEX CONCURRENTLY

On Wed, Nov 07, 2012 at 12:54:57PM -0300, Alvaro Herrera wrote:

I just observed that if I run src/test/regress' "make
installcheck-parallel" and concurrently src/test/isolation's "make
installcheck", I very reproducibly get a segfault (in an assert-enabled
build). Apparently the problem is that somebody is trying to transfer
predicate locks to a relation, but the index' relcache entry is pfreed.

Yes, the location of TransferPredicateLocksToHeapRelation is bad...

I think the fix is just to move the heap_open calls some lines upwards, as in
the attached patch.

diff --git a/src/backend/catalog/index.c b/src/backend/catalog/index.c
index 756f6d9..d2d91c1 100644
--- a/src/backend/catalog/index.c
+++ b/src/backend/catalog/index.c
@@ -1475,6 +1475,8 @@ index_drop(Oid indexId, bool concurrent)
* conflicts with existing predicate locks, so now is the time to move
* them to the heap relation.
*/
+		userHeapRelation = heap_open(heapId, ShareUpdateExclusiveLock);
+		userIndexRelation = index_open(indexId, ShareUpdateExclusiveLock);
TransferPredicateLocksToHeapRelation(userIndexRelation);

/*
@@ -1484,9 +1486,6 @@ index_drop(Oid indexId, bool concurrent)
*/
indexRelation = heap_open(IndexRelationId, RowExclusiveLock);

- userHeapRelation = heap_open(heapId, ShareUpdateExclusiveLock);
- userIndexRelation = index_open(indexId, ShareUpdateExclusiveLock);
-
tuple = SearchSysCacheCopy1(INDEXRELID,
ObjectIdGetDatum(indexId));
if (!HeapTupleIsValid(tuple))

Looks like the right thing to me.

Greetings,

Andres

Andres Freund wrote:

On Wed, Nov 07, 2012 at 12:54:57PM -0300, Alvaro Herrera wrote:

I think the fix is just to move the heap_open calls some lines upwards, as in
the attached patch.

Looks like the right thing to me.

Pushed.

--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services