Re: BUG #19026: ResourceOwnerForget can't find owner for invalid plancache

I wrote:
> In practice, fcache->cplan will never be not-null after successful
> completion of a SQL function, so one idea is to simply clear it
> unconditionally as soon as we know we're starting a fresh execution,
> more or less as in alternative-1 attached. However that leaves me
> a bit unsatisfied, because it doesn't protect against the case of
> erroring out of a set-returning function: if we come in and see
> eslist != NULL, we'll pick right back up attempting to execute
> plans that probably aren't there anymore. I think that that case
> is unreachable today because we don't allow any opclass support
> functions to be SRFs, and AFAIK there are no other cases where an
> FmgrInfo would be re-used after a failed query. Still, I'm inclined
> to go with something more like alternative-2, which feels a little
> more future-proof.

After closer inspection: Alexander's test case doesn't expose the full
scope of the problem. If the function suffers an error at run-time
rather than early in setup, then we will fall out with fcache->eslist
being non-null as well as fcache->cplan. Then on the next call,
init_sql_fcache will believe that it is resuming execution of a
set-returning function, and we'll merrily try to execute an executor
state tree that's not there anymore.

So we have to do something more like my alternative-2 than any of
the other proposals, and after reviewing the code I believe it
had better reset the tstore and shutdown_reg fields as well.
Hence, v3 attached, now with regression test case.

regards, tom lane