Jim Nasby wrote:
> >Here we have another case. prodesc is a global thing. And it is shared
> >between different operations. Problem was that there is no partcular
> >owner, and we have to wait when last operation which deals with it
> >would finish. It looks like perfect job for reference counting.
>
> I've just tried to wrap my head around what's going on with prodesc and
> failed... specifically, I don't understand this claim in the comment:
>
> * Add the proc description block to the hashtable. Note we do not
> * attempt to free any previously existing prodesc block. !!This is
> * annoying, but necessary since there could be active calls using
> * the old prodesc.!!
>
> What else could be referencing it? I realize it's stored in pltcl_proc_htab,
> but AFAICT that's backend-local. So I don't understand what else could be
> referencing it.
Try to open a cursor that uses the function, fetch a few tuples from it;
then change the function and fetch more rows from the cursor. I suppose
the open cursor could contain a reference to the function's prodesc.
Refcounting the prodesc would let it live until the cursor's closed,
then free it.
--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services