Re: pg_logical_slot_peek_changes crashes postgres when called from inside pl/pgsql

Andres Freund <andres(at)anarazel(dot)de> writes:
> I'm kinda surprised that this only causes problems with logical decoding
> and not elsewhere, this isn't a new issue.

The reason for that seems to be that ReorderBufferCommit and
ReorderBufferImmediateInvalidation are the only users of
BeginInternalSubTransaction other than the PLs. So otherwise we
never see a case of subtransaction abort where we don't want
to clean up the current operation in the surrounding SPI context.

> So far your option of allowing to opt in into additional cleanup in the
> CATCH seems the least ugly

I thought of a better way, as attached. The core problem is that we
need to know whether the SPI context's current executor operation is
"inside" or "outside" the subtransaction being abandoned. That info
is not tracked at the moment, but we can track it, at a cost that's
entirely trivial compared to the other stuff that a SPI operation
will do.

This seems back-patchable without much angst, although we might want
to add the new _SPI_connection field at the end in the back branches.
I'm not sure if anything outside core is looking at spi_priv.h.

regards, tom lane