On 2018-11-11 11:29:54 +1300, Thomas Munro wrote:
> On Sat, Nov 10, 2018 at 6:01 PM Andres Freund <andres@anarazel.de> wrote:
> > I've replaced that with a write barrier / read barrier. I strongly
> > suspect this isn't a problem on the write side in practice (due to the
> > dependent read), but the read side looks much less clear to me. I think
> > explicitly using barriers is much saner these days.
> >
> > Comments?
>
> +1
>
> I said the same over here:
>
> https://www.postgresql.org/message-id/flat/CAEepm%3D1nff0x%3D7i3YQO16jLA2qw-F9O39YmUew4oq-xcBQBs0g%40mail.gmail.com
Hah! Sorry for missing that back then. I think your patch from back
then misses a few things that mine did. But I also think mine missed the
fact that XidCacheRemove is problematic - I only was thinking of the
*reads* of MyPgXact->nxids in XidCacheRemoveRunningXids(), but you
correctly observe that the write is the problematic case (the reads are
ok, because it's the same process as GetNewTransactionId()).
Greetings,
Andres Freund