On 2016-04-07 16:50:44 +0300, Alexander Korotkov wrote:
> On Thu, Apr 7, 2016 at 4:41 PM, Andres Freund <andres@anarazel.de> wrote:
>
> > On 2016-03-31 20:21:02 +0300, Alexander Korotkov wrote:
> > > ! BEGIN_BUFSTATE_CAS_LOOP(bufHdr);
> > >
> > > ! Assert(BUF_STATE_GET_REFCOUNT(state) > 0);
> > > ! wasDirty = (state & BM_DIRTY) ? true : false;
> > > ! state |= BM_DIRTY | BM_JUST_DIRTIED;
> > > ! if (state == oldstate)
> > > ! break;
> >
> > I'm doubtful that this early exit is entirely safe. None of the
> > preceding operations imply a memory barrier. The buffer could previously
> > have been marked dirty, but cleaned since. It's pretty critical that we
> > re-set the dirty bit (there's no danger of loosing it with a barrier,
> > because we hold an exclusive content lock).
> >
>
> Oh, I get it.
>
>
> > Practically the risk seems fairly low, because acquiring the exclusive
> > content lock will have implied a barrier. But it seems unlikely to have
> > a measurable performance effect to me, so I'd rather not add the early
> > exit.
> >
>
> Ok, let's just remove it.
Here's my updated version of the patch. I've updated this on an
intercontinental flight, after a otherwise hectic week (moving from SF
to Berlin); so I'm planning to look over this once more before pushing (.
I've decided that the cas-loop macros are too obfuscating for my
taste. To avoid duplicating the wait part I've introduced
WaitBufHdrUnlocked().
As you can see in
http://archives.postgresql.org/message-id/CA%2BTgmoaeRbN%3DZ4oWENLvgGLeHEvGZ_S_Z3KGrdScyKiSvNt3oA%40mail.gmail.com
I'm planning to apply this sometime this weekend, after running some
tests and going over the patch again.
Any chance you could have a look over this?
Regards,
Andres
