Re: [RFC PATCH] sys_membarrier(): system/process-wide memory barrier (x86) (v12)

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Mar 17, 2015 at 09:49:40AM -0700, Paul E. McKenney wrote:
> On Tue, Mar 17, 2015 at 05:37:56PM +0100, Peter Zijlstra wrote:
> > On Tue, Mar 17, 2015 at 01:13:36PM +0000, Mathieu Desnoyers wrote:
> > > > Its basically: WMB + ACQUIRE, which theoretically can leak a read in,
> > > > but nobody sane _delays_ reads, you want to speculate reads, not
> > > > postpone.
> > > 
> > > If I believe the memory ordering table at
> > > https://en.wikipedia.org/wiki/Memory_ordering , there appears
> > > to be quite a few architectures that can reorder loads after loads,
> > > and loads after stores: Alpha, ARMv7, PA-RISC, SPARC RMO, x86 oostore
> > > and ia64. There may be subtle details that would allow us to
> > > do without the barriers in specific situations, but for that I'd
> > > very much like to hear what Paul has to say.
> > 
> > So I was starting to write that you can get load after load by one
> > speculating more than the other, but I suppose you can delay loads just
> > fine too.
> > 
> > Imagine getting a cache miss on a load, the OoO engine can then continue
> > execution until it hits a hard dependency, so you're effectively
> > delaying the load.
> > 
> > So yeah, if we want to be able to replace smp_rmb() with a
> > barrier+sys_membar() we need to promote the smp_mb__before_spinlock() to
> > smp_mb__after_unlock_lock() or so, that would only penalize PPC a bit.
> 
> Agreed, though if Mathieu is dropping the expedited version for the
> moment, this should not be required yet, right?

Indeed so.