Re: [PATCH 4/4] locking: Introduce smp_cond_acquire()

From: Will Deacon <will.deacon@arm.com>
To: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>,
	Oleg Nesterov <oleg@redhat.com>,
	Boqun Feng <boqun.feng@gmail.com>,
	mingo@kernel.org, linux-kernel@vger.kernel.org, corbet@lwn.net,
	mhocko@kernel.org, dhowells@redhat.com,
	torvalds@linux-foundation.org,
	Michael Ellerman <mpe@ellerman.id.au>,
	Benjamin Herrenschmidt <benh@kernel.crashing.org>,
	Paul Mackerras <paulus@samba.org>
Subject: Re: [PATCH 4/4] locking: Introduce smp_cond_acquire()
Date: Thu, 12 Nov 2015 21:25:35 +0000	[thread overview]
Message-ID: <20151112212535.GB23979@arm.com> (raw)
In-Reply-To: <20151112152042.GC3972@linux.vnet.ibm.com>

[sorry for the late reply, I'm away from my desk until Monday since I'm
 busy with family issues]

On Thu, Nov 12, 2015 at 07:20:42AM -0800, Paul E. McKenney wrote:
> On Thu, Nov 12, 2015 at 04:08:22PM +0100, Peter Zijlstra wrote:
> > On Thu, Nov 12, 2015 at 07:01:49AM -0800, Paul E. McKenney wrote:
> > > On Thu, Nov 12, 2015 at 03:50:13PM +0100, Peter Zijlstra wrote:
> > > > On Thu, Nov 12, 2015 at 06:40:04AM -0800, Paul E. McKenney wrote:
> > > > > I cannot resist suggesting that any lock that interacts with
> > > > > spin_unlock_wait() must have all relevant acquisitions followed by
> > > > > smp_mb__after_unlock_lock().
> > > > 
> > > > Ha! that would certainly help here. But it would mean that argh64v8 also
> > > > needs to define that, even though that is already RCsc.
> > > 
> > > Maybe.  It could also be that arm64 avoids the need somehow, for example
> > > via their RCsc behavior.  Their memory model is similar to PPC, but not
> > > exactly the same.
> > > 
> > > Will?
> > 
> > So when I spoke to Will earlier today, we agreed that LDAXR+STXR is
> > susceptible to the same problem. The STXR will allow loads to pass up
> > over that store.
> > 
> > On v8.1, which is using LDADDA, this is not an issue, for as the ACQUIRE
> > is part of the LOAD, the Read-Modify-Write is indivisible as a whole,
> > and therefore a subsequent load has nothing to pass over.
> 
> So one approach required for one level of hardware and another for the
> next level.  I can relate to that all too well...  :-/

Just to confirm, Peter's correct in that Boqun's litmus test is permitted
by the arm64 architecture when the ll/sc spinlock definitions are in use.

However, I don't think that strengthening smp_mb__after_unlock_lock is
the right way to solve this. I'll reply to the other part of the thread...

Will