Re: Latches and barriers

On 2015-01-12 12:44:56 -0500, Robert Haas wrote:
> On Mon, Jan 12, 2015 at 11:27 AM, Andres Freund <andres@2ndquadrant.com> wrote:
> > On 2015-01-12 11:03:42 -0500, Tom Lane wrote:
> >> Andres Freund <andres@2ndquadrant.com> writes:
> >> > While it might not be required for existing latch uses (I'm *not* sure
> >> > that's true)
> >
> > I think at least syncrep.c might not be correct. In SyncRepWakeQueue()
> > it sets PGPROC->syncRepState without the necessary barriers (via locks),
> > although it does use them in SyncRepWaitForLSN().
> >
> > It is, perhaps surprisingly to many, not sufficient to take a spinlock,
> > change the flag, release it and then set the latch - the release alone
> > doesn't guarantee a sufficient barrier unless looking at the flag is
> > also protected by the spinlock.
> 
> I thought we decided that a spinlock acquire or release should be a
> full barrier.

Acquire + release, yes. But not release alone? The x86 spinlock release
currently is just a store - that won't ever be a full barrier.

Greetings,

Andres Freund

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