Re: Spinlocks and compiler/memory barriers

On 2014-06-26 15:40:11 -0700, Tom Lane wrote:
> Andres Freund <andres@2ndquadrant.com> writes:
> > On 2014-06-26 14:13:07 -0700, Tom Lane wrote:
> >> Surely it had better be a read barrier as well?
> 
> > I don't immediately see why it has to be read barrier? Hoisting a load
> > from after the release into the locked area of code should be safe?
> 
> No doubt, but delaying a read till after the unlocking write would
> certainly not be safe.

Right. What we actually want for locking is what several systems
(e.g. C11, linux) coin acquire/release barriers.
Not sure whether we should introduce a separate set of acquire/release
macros, or "promote" our barriers to have stronger guarantees than the
name implies.

The definition as I understand it is:

A acquire barrier implies that:
* memory operations from after the barrier cannot appear to have happened before the barrier
* but: memory operations from *before* the barrier are *not* guaranteed to be finished

A finished release barrier implies:
* stores from before the barrier cannot be moved past
* loads from before the barrier cannot be moved past
* but: reads from *after* the barrier might occur *before* the barrier.

I believe that all our current barrier definitions (except maybe alpha
which I haven't bothered to check thoroughly) satisfy those
constraints. That's primarily because we don't have support for all that
many platforms and use full memory barriers for read/write barriers in
several places.

Greetings,

Andres Freund

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