I think there might be a case like this:
* ExecRescanHashJoin() decides it can't reuse the hash table for a
rescan, so it calls ExecHashTableDestroy(), clears HashJoinState's
hj_HashTable and sets hj_JoinState to HJ_BUILD_HASHTABLE
* the HashState node still has a reference to the pfree'd HashJoinTable!
* HJ_BUILD_HASHTABLE case reaches the empty-outer optimisation case so
it doesn't bother to build a new hash table
* EXPLAIN examines the HashState's pointer to a freed HashJoinTable struct
Yes, debugging with gdb shows this is exactly what happens.
According to the scenario above, here is a recipe that reproduces this
issue.
-- recipe start
create table a(i int, j int);
create table b(i int, j int);
create table c(i int, j int);
insert into a select 3,3;
insert into a select 2,2;
insert into a select 1,1;
insert into b select 3,3;
insert into c select 0,0;
analyze a;
analyze b;
analyze c;
set enable_nestloop to off;
set enable_mergejoin to off;
explain analyze
select exists(select * from b join c on a.i > c.i and a.i = b.i and b.j = c.j) from a;
-- recipe end
I tried this recipe on different PostgreSQL versions, starting from
current master and going backwards. I was able to reproduce this issue
on all versions above 8.4. In 8.4 version, we do not output information
on hash buckets/batches. But manual inspection with gdb shows in 8.4 we
also have the dangling pointer for HashState->hashtable. I didn't check
versions below 8.4 though.
Thanks
Richard