Re: Spinlocks and compiler/memory barriers

Hi,

On 2014-06-26 23:01:10 +0200, Andres Freund wrote:
> I think we should rework things so that we fall back to
> pg_write_barrier(), (*((volatile slock_t *) (lock)) = 0) instead of what
> we have right now.
> That'd require to make barrier.h independent from s_lock.h, but I think
> that'd be a pretty clear improvement. One open question is what happens
> with the SpinlockRelease() when barriers are implemented using spinlocks
> and we need a barrier for the SpinlockRelease().

Too tired to think about this any further today. Here's my current state
of this:
* barrier.h's spinlock implementation moved to s_lock.c, loosing the
s_lock.h include. That requires a S_UNLOCK definition which doesn't
again use a barrier. I've coined it S_UNLOCKED_UNLOCK
* s_lock.h now includes barrier.h to implement the generic S_UNLOCK
safely.
* gcc x86-64 had a superflous "cc" clobber. Likely copied from the 32bit
version which does additional operations.
* PPC was missing a compiler barrier
* alpha was missing a "cc" clobber.
* mips was missing a compiler barrier and a "cc" clobber
* I have no idea how to fix pa-risc's S_UNLOCK for !gcc. The referenced
spinlock paper calls a external function to avoid reordering.

Greetings,

Andres Freund

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