Re: StrategyGetBuffer optimization, take 2

On 2013-08-19 15:17:44 -0700, Jeff Janes wrote:
> On Wed, Aug 7, 2013 at 7:40 AM, Merlin Moncure <mmoncure(at)gmail(dot)com> wrote:
>
> > I agree; at least then it's not unambiguously better. if you (in
> > effect) swap all contention on allocation from a lwlock to a spinlock
> > it's not clear if you're improving things; it would have to be proven
> > and I'm trying to keep things simple.
> >
> > Attached is a scaled down version of the patch that keeps the freelist
> > lock but still removes the spinlock during the clock sweep. This
> > still hits the major objectives of reducing the chance of scheduling
> > out while holding the BufFreelistLock and mitigating the worst case
> > impact of doing so if it does happen. An even more scaled down
> > version would keep the current logic exactly as is except for
> > replacing buffer lock in the clock sweep with a trylock (which is
> > IMNSHO a no-brainer).
>
> Since usage_count is unsigned, are you sure that changing the tests
> from "buf->usage_count == 0" to "buf->usage_count <= 0" accomplishes
> what you need it to? If usage_count gets decremented when it already
> zero, it will wrap around to 65,535, at least on some compilers some
> of the time, won't it?

Overflow of *unsigned* variables is actually defined and will always
wrap around. It's signed variables which don't have such a clear
behaviour.

Greetings,

Andres Freund

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