On 2014-09-19 13:58:17 -0400, Robert Haas wrote:
> On Wed, Sep 17, 2014 at 7:45 AM, Andres Freund <andres@2ndquadrant.com> wrote:
> > I just tried this on my normal x86 workstation. I applied your lwlock
> > patch and ontop I removed most volatiles (there's a couple still
> > required) from xlog.c. Works for 100 seconds. Then I reverted the above
> > commits. Breaks within seconds:
> > master:
> > LOG: request to flush past end of generated WAL; request 2/E5EC3DE0, currpos 2/E5EC1E60
> > standby:
> > LOG: record with incorrect prev-link 4/684C3108 at 4/684C3108
> > and similar.
> >
> > So at least for x86 the compiler barriers are obviously required and
> > seemingly working.
>
> Oh, that's great. In that case I think I should go ahead and apply
> that patch in the hopes of turning up any systems where the barriers
> aren't working properly yet.
Agreed.
> Although it would be nice to know whether it breaks with *only* the lwlock.c patch.
It didn't, at least not visibly within the 1000s I let pgbench run.
> > I've attached the very quickly written xlog.c de-volatizing patch.
>
> I don't have time to go through this in detail, but I don't object to
> you applying it if you're confident you've done it carefully enough.
It's definitely not yet carefully enough checked. I wanted to get it
break fast and only made one pass through the file. But I think it
should be easy enough to get it into shape for that.
Greetings,
Andres Freund
-- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services