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

Hi,

On 2022-02-11 09:10:38 +1300, Thomas Munro wrote:
> I was about to commit that, because the original Windows problem it
> solved is showing up occasionally in CI failures (that is, it already
> solves a live problem, albeit a different and non-data-corrupting
> one):

+1

> It seems like I should go ahead and do that today, and we can study
> further uses for PROCSIGNAL_BARRIER_SMGRRELEASE in follow-on work?

Yes.

I wonder whether we really should make the barriers be conditional on
defined(WIN32) || defined(USE_ASSERT_CHECKING) as done in that patch, even
leaving wraparound dangers aside. On !windows we still have the issues of the
space for the dropped / moved files not being released if there are processes
having them open. That can be a lot of space if there's long-lived connections
in a cluster that doesn't fit into s_b (because processes will have random fds
open for writing back dirty buffers). And we don't truncate the files before
unlinking when done as part of a DROP DATABASE...

But that's something we can fine-tune later as well...

Greetings,

Andres Freund