Re: Triggers might be caching broken fmgr info

Peter Eisentraut <peter_e(at)gmx(dot)net> writes:
> My best guess is that the trigger fmgr lookup in trigger.c:846
> if (trigger->tgfunc.fn_oid == InvalidOid)
> fmgr_info(trigger->tgfoid, &trigger->tgfunc);
> might be reading out of an incompletely initialized trigger->tgfunc
> structure.

Yes, that's the problem.

> This is supported by the fact that if I move the
> finfo->fn_oid = functionId;
> assignment in fmgr_info() to the very end of that function, this passes
> gracefully (because the elog comes before the assignment).

I think this is an OK fix, since utils/fmgr/README documents as valid
the technique trigger.c is using:

fn_oid = InvalidOid can be used
to denote a not-yet-initialized FmgrInfo struct.

fmgr_info is clearly failing to live up to the implications of that
commitment. Please move the fn_oid assignment, and insert a note
that it must be last...

> This might even be a workable fix, but I'm wondering whether elog(ERROR)
> should not flush the trigger cache.

The cache in question is part of the relcache, and no we don't want to
flush it on every elog(ERROR).

regards, tom lane