On Mon, Apr 30, 2018 at 08:01:40PM -0400, Tom Lane wrote:
> It's clear from dory's results that something is causing a 4MB chunk
> of memory to get reserved in the process's address space, sometimes.
> It might happen during the main MapViewOfFileEx call, or during the
> preceding VirtualFree, or with my map/unmap dance in place, it might
> happen during that. Frequently it doesn't happen at all, at least not
> before the point where we've successfully done MapViewOfFileEx. But
> if it does happen, and the chunk happens to get put in a spot that
> overlaps where we want to put the shmem block, kaboom.
>
> What seems like a plausible theory at this point is that the apparent
> asynchronicity is due to the allocation being triggered by a different
> thread, and the fact that our added monitoring code seems to make the
> failure more likely can be explained by that code changing the timing.
> But what thread could it be? It doesn't really look to me like either
> the signal thread or the timer thread could eat 4MB. syslogger.c
> also spawns a thread, on Windows, but AFAICS that's not being used in
> this test configuration. Maybe the reason dory is showing the problem
> is something or other is spawning a thread we don't even know about?
Likely some privileged daemon is creating a thread in every new process. (On
Windows, it's not unusual for one process to create a thread in another
process.) We don't have good control over that.
> I'm at a loss for a reasonable way to fix it
> for real. Is there a way to seize control of a Windows process so that
> there are no other running threads?
I think not.
> Any other ideas?
PostgreSQL could retry the whole process creation, analogous to
internal_forkexec() retries. Have the failed process exit after recording the
fact that it couldn't attach. Make the postmaster notice and spawn a
replacement. Give up after 100 failed attempts.