Re: Memory ordering issue in LWLockRelease, WakeupWaiters, WALInsertSlotRelease

On 2014-02-14 10:26:07 -0500, Tom Lane wrote:
> Florian Pflug <fgp(at)phlo(dot)org> writes:
> > Another idea for a fix would be to conflate lwWaiting and lwWaitLink into one
> > field. We could replace "lwWaiting" by "lwWaitLink != NULL" everywhere it's
> > tested, and set lwWaitLink to some special non-NULL value (say 0x1) when we
> > enqueue a PGPROC, instead of setting it to NULL and setting lwWaiting to true.
>
> > We'd then depend on pointer-sized stores being atomic, which I think we depend
> > on in other places already.
>
> I don't believe that's true; neither that we depend on it now, nor that
> it would be safe to do so.

Yea, we currently rely on 4 byte stores being atomic (most notably for
xids), but we don't rely on anything bigger. I am not sure if there are
architectures with 64bit pointers that aren't accessed atomically when
aligned? Even if, that's certainly nothing that should be introduced
when backpatching.

Greetings,

Andres Freund

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