Re: better atomics - v0.5

On 2014-06-26 07:44:14 -0500, Merlin Moncure wrote:
> On Thu, Jun 26, 2014 at 5:20 AM, Andres Freund <andres(at)2ndquadrant(dot)com> wrote:
> > On 2014-06-25 20:16:08 -0400, Robert Haas wrote:
> >> On Wed, Jun 25, 2014 at 4:36 PM, Andres Freund <andres(at)2ndquadrant(dot)com> wrote:
> >> > Since it better be legal to manipulate a atomic variable while holding a
> >> > spinlock we cannot simply use an arbitrary spinlock as backing for
> >> > atomics. That'd possibly cause us to wait on ourselves or cause
> >> > deadlocks.
> >>
> >> I think that's going to fall afoul of Tom's previously-articulated "no
> >> loops inside spinlocks" rule. Most atomics, by nature, are
> >> loop-until-it-works.
> >
> > Well, so is TAS itself :).
> >
> > More seriously, I think we're not going to have much fun if we're making
> > up the rule that you can't do an atomic add/sub while a spinlock is
> > held. That just precludes to many use cases and will make the code much
> > harder to understand. I don't think we're going to end up having many
> > problems if we allow atomic read/add/sub/write in there.
>
> That rule seems reasonable -- why would you ever want to do this?

Are you wondering why you'd ever manipulate an atomic op inside a
spinlock?
There's enough algorithms where a slowpath is done under a spinlock but
the fastpath is done without. You can't simply use nonatomic operations
for manipulation under the spinlock because the fastpaths might then
observe/cause bogus state.

Obviously I'm not advocating to do random stuff in spinlocks. W're
talking about single lock xadd;s (or the LL/SC) equivalent here.

Greetings,

Andres Freund

--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services