Re: wrong fds used for refilenodes after pg_upgrade relfilenode changes Reply-To:

On Wed, May 4, 2022 at 2:23 PM Thomas Munro <thomas(dot)munro(at)gmail(dot)com> wrote:
> Assuming no
> objections or CI failures show up, I'll consider pushing the first two
> patches tomorrow.

Done. Time to watch the build farm.

It's possible that these changes will produce some blowback, now that
we're using PROCSIGNAL_BARRIER_SMGRRELEASE on common platforms.
Obviously the earlier work started down this path already, but that
was Windows-only, so it probably didn't get much attention from our
mostly Unix crowd.

For example, if there are bugs in the procsignal barrier system, or if
there are places that don't process interrupts at all or promptly, we
might hear complaints about that. That bug surface includes, for
example, background workers created by extensions. An example on my
mind is a place where we hold interrupts while cleaning up temporary
files (= a loop of arbitrary size with filesystem ops that might hang
on slow storage), so a concurrent DROP TABLESPACE will have to wait,
which is kinda strange because it would in fact be perfectly safe to
process this particular "interrupt". In that case we really just
don't want the kinds of interrupts that perform non-local exits and
prevent our cleanup work from running to completion, but we don't have
a way to say that. I think we'll probably also want to invent a way
to report which backend is holding up progress, since without that
it's practically impossible for an end user to understand why their
command is hanging.