Andres Freund <andres@2ndquadrant.com> writes:
> On 2015-02-26 13:53:18 -0500, Tom Lane wrote:
>> I thought that's what I was proposing in point #1.
> Sure, but my second point was to avoid duplicating that information into
> ->fcinfo and such and instead reference the typecache entry from
> there. So that if the typecache entry is being rebuilt (a new mechanism
> I'm afraid), whatever uses ->fcinfo gets the updated
> functions/definition.
The trick there would be that if we don't want to copy then we'd need a
reference counting mechanism, and FmgrInfos aren't used in a way that
would allow that to work easily. (Whatever the typcache is holding onto
clearly must be long-lived, but you do want an obsoleted one to go away
once there are no more FmgrInfos referencing it.)
I was just thinking though that we could possibly solve that if we went
ahead and invented the memory context reset callback mechanism that
I suggested in another thread a week or two back. Then you could imagine
that when a domain-checking function caches a reference to a "domain
constraints info" object in its FmgrInfo, it could increment a refcount
on the info object, and register a callback on the context containing the
FmgrInfo to release that refcount.
A different approach would be to try to use the ResourceOwner mechanism
to track these info objects. But I think that does not work nicely for
plpgsql; at least not unless it creates a ResourceOwner for every
function, which seems kinda messy.
regards, tom lane
