Date: Fri, 29 Sep 2017 17:33:49 +0100
From: Will Deacon <will.deacon@arm.com>
To: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Michael Cree <mcree@orcon.net.nz>,
        Peter Zijlstra <peterz@infradead.org>, kirill.shutemov@linux.intel.com,
        linux-kernel@vger.kernel.org, ynorov@caviumnetworks.com,
        rruigrok@codeaurora.org, linux-arch@vger.kernel.org,
        akpm@linux-foundation.org, catalin.marinas@arm.com, rth@twiddle.net,
        ink@jurassic.park.msu.ru, mattst88@gmail.com,
        linux-alpha@vger.kernel.org
Subject: Re: [RFC PATCH 1/2] arm64: mm: Use READ_ONCE/WRITE_ONCE when
 accessing page tables
Message-ID: <20170929163348.GD2376@arm.com>
References: <1506527369-19535-1-git-send-email-will.deacon@arm.com>
 <1506527369-19535-2-git-send-email-will.deacon@arm.com>
 <20170928083801.m6rb4frbbgzgam2o@hirez.programming.kicks-ass.net>
 <20170928084535.GA19060@arm.com>
 <20170928154354.GK3521@linux.vnet.ibm.com>
 <20170928185909.kgwy36hkoqxqgyc3@tower>
 <20170929005830.GU3521@linux.vnet.ibm.com>
 <20170929090843.GB14791@arm.com>
 <20170929162939.GY3521@linux.vnet.ibm.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20170929162939.GY3521@linux.vnet.ibm.com>
User-Agent: Mutt/1.5.23 (2014-03-12)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3888
Lines: 79

On Fri, Sep 29, 2017 at 09:29:39AM -0700, Paul E. McKenney wrote:
> On Fri, Sep 29, 2017 at 10:08:43AM +0100, Will Deacon wrote:
> > On Thu, Sep 28, 2017 at 05:58:30PM -0700, Paul E. McKenney wrote:
> > > On Fri, Sep 29, 2017 at 07:59:09AM +1300, Michael Cree wrote:
> > > > On Thu, Sep 28, 2017 at 08:43:54AM -0700, Paul E. McKenney wrote:
> > > > > On Thu, Sep 28, 2017 at 09:45:35AM +0100, Will Deacon wrote:
> > > > > > On Thu, Sep 28, 2017 at 10:38:01AM +0200, Peter Zijlstra wrote:
> > > > > > > On Wed, Sep 27, 2017 at 04:49:28PM +0100, Will Deacon wrote:
> > > > > > > > In many cases, page tables can be accessed concurrently by either another
> > > > > > > > CPU (due to things like fast gup) or by the hardware page table walker
> > > > > > > > itself, which may set access/dirty bits. In such cases, it is important
> > > > > > > > to use READ_ONCE/WRITE_ONCE when accessing page table entries so that
> > > > > > > > entries cannot be torn, merged or subject to apparent loss of coherence.
> > > > > > > 
> > > > > > > In fact, we should use lockless_dereference() for many of them. Yes
> > > > > > > Alpha is the only one that cares about the difference between that and
> > > > > > > READ_ONCE() and they do have the extra barrier, but if we're going to do
> > > > > > > this, we might as well do it 'right' :-)
> > > > > > 
> > > > > > I know this sounds daft, but I think one of the big reasons why
> > > > > > lockless_dereference() doesn't get an awful lot of use is because it's
> > > > > > such a mouthful! Why don't we just move the smp_read_barrier_depends()
> > > > > > into READ_ONCE? Would anybody actually care about the potential impact on
> > > > > > Alpha (which, frankly, is treading on thin ice given the low adoption of
> > > > > > lockless_dereference())?
> > > > > 
> > > > > This is my cue to ask my usual question...  ;-)
> > > > > 
> > > > > Are people still running mainline kernels on Alpha?  (Added Alpha folks.)
> > > > 
> > > > Yes.  I run two Alpha build daemons that build the unofficial
> > > > debian-alpha port.  Debian popcon reports nine machines running
> > > > Alpha, which are likely to be running the 4.12.y kernel which
> > > > is currently in debian-alpha, (and presumably soon to be 4.13.y
> > > > which is now built on Alpha in experimental).
> > > 
> > > I salute your dedication to Alpha!  ;-)
> > 
> > Ok, but where does that leave us wrt my initial proposal of moving
> > smp_read_barrier_depends() into READ_ONCE and getting rid of
> > lockless_dereference?
> > 
> > Michael (or anybody else running mainline on SMP Alpha) -- would you be
> > able to give the diff below a spin and see whether there's a measurable
> > performance impact?
> 
> This will be a sensitive test.  The smp_read_barrier_depends() can be
> removed from lockless_dereference().  Without this removal Alpha will
> get two memory barriers from rcu_dereference() and friends.

Oh yes, good point. I was trying to keep the diff simple, but you're
right that this is packing too many barriers. Fixed diff below.

Thanks,

Will

--->8

diff --git a/include/linux/compiler.h b/include/linux/compiler.h
index e95a2631e545..c4ee9d6d8f2d 100644
--- a/include/linux/compiler.h
+++ b/include/linux/compiler.h
@@ -340,6 +340,7 @@ static __always_inline void __write_once_size(volatile void *p, void *res, int s
 		__read_once_size(&(x), __u.__c, sizeof(x));		\
 	else								\
 		__read_once_size_nocheck(&(x), __u.__c, sizeof(x));	\
+	smp_read_barrier_depends(); /* Enforce dependency ordering from x */ \
 	__u.__val;							\
 })
 #define READ_ONCE(x) __READ_ONCE(x, 1)
@@ -620,7 +621,6 @@ static __always_inline void __write_once_size(volatile void *p, void *res, int s
 ({ \
 	typeof(p) _________p1 = READ_ONCE(p); \
 	typeof(*(p)) *___typecheck_p __maybe_unused; \
-	smp_read_barrier_depends(); /* Dependency order vs. p above. */ \
 	(_________p1); \
 })