Date:   Thu, 5 Jul 2018 16:00:29 +0200
From:   Andrea Parri <andrea.parri@...rulasolutions.com>
To:     Will Deacon <will.deacon@....com>
Cc:     Alan Stern <stern@...land.harvard.edu>,
        LKMM Maintainers -- Akira Yokosawa <akiyks@...il.com>,
        Boqun Feng <boqun.feng@...il.com>,
        David Howells <dhowells@...hat.com>,
        Jade Alglave <j.alglave@....ac.uk>,
        Luc Maranget <luc.maranget@...ia.fr>,
        Nicholas Piggin <npiggin@...il.com>,
        "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Kernel development list <linux-kernel@...r.kernel.org>,
        dlustig@...dia.com
Subject: Re: [PATCH 2/2] tools/memory-model: Add write ordering by
 release-acquire and by locks

On Wed, Jul 04, 2018 at 01:11:04PM +0100, Will Deacon wrote:
> Hi Alan,
> 
> On Tue, Jul 03, 2018 at 01:28:17PM -0400, Alan Stern wrote:
> > On Mon, 25 Jun 2018, Andrea Parri wrote:
> > 
> > > On Fri, Jun 22, 2018 at 07:30:08PM +0100, Will Deacon wrote:
> > > > > > I think the second example would preclude us using LDAPR for load-acquire,
> > > 
> > > > I don't think it's a moot point. We want new architectures to implement
> > > > acquire/release efficiently, and it's not unlikely that they will have
> > > > acquire loads that are similar in semantics to LDAPR. This patch prevents
> > > > them from doing so,
> > > 
> > > By this same argument, you should not be a "big fan" of rfi-rel-acq in ppo ;)
> > > consider, e.g., the two litmus tests below: what am I missing?
> > 
> > This is an excellent point, which seems to have gotten lost in the 
> > shuffle.  I'd like to see your comments.
> 
> Yeah, sorry. Loads going on at the moment. You could ask herd instead of me
> though ;)
> 
> > In essence, if you're using release-acquire instructions that only
> > provide RCpc consistency, does store-release followed by load-acquire
> > of the same address provide read-read ordering?  In theory it doesn't
> > have to, because if the value from the store-release is forwarded to
> > the load-acquire then:
> > 
> > 	LOAD A
> > 	STORE-RELEASE X, v
> > 	LOAD-ACQUIRE X
> > 	LOAD B
> > 
> > could be executed by the CPU in the order:
> > 
> > 	LOAD-ACQUIRE X
> > 	LOAD B
> > 	LOAD A
> > 	STORE-RELEASE X, v
> > 
> > thereby accessing A and B out of program order without violating the
> > requirements on the release or the acquire.
> > 
> > Of course PPC doesn't allow this, but should we rule it out entirely?
> 
> This would be allowed if LOAD-ACQUIRE was implemented using LDAPR on Arm.
> I don't think we should be ruling out architectures using RCpc
> acquire/release primitives, because doing so just feels like an artifact of
> most architectures building these out of fences today.
> 
> It's funny really, because from an Arm-perspective I don't plan to stray
> outside of RCsc, but I feel like other weak architectures aren't being
> well represented here. If we just care about x86, Arm and Power (and assume
> that Power doesn't plan to implement RCpc acquire/release instructions)
> then we're good to tighten things up. But I fear that RISC-V should probably
> be more engaged (adding Daniel) and who knows about MIPS or these other
> random architectures popping up on linux-arch.
> 
> > > C MP+fencewmbonceonce+pooncerelease-rfireleaseacquire-poacquireonce
> > > 
> > > {}
> > > 
> > > P0(int *x, int *y)
> > > {
> > > 	WRITE_ONCE(*x, 1);
> > > 	smp_wmb();
> > > 	WRITE_ONCE(*y, 1);
> > > }
> > > 
> > > P1(int *x, int *y, int *z)
> > > {
> > > 	r0 = READ_ONCE(*y);
> > > 	smp_store_release(z, 1);
> > > 	r1 = smp_load_acquire(z);
> > > 	r2 = READ_ONCE(*x);
> > > }
> > > 
> > > exists (1:r0=1 /\ 1:r1=1 /\ 1:r2=0)
> > > 
> > > 
> > > AArch64 MP+dmb.st+popl-rfilq-poqp
> > > "DMB.STdWW Rfe PodRWPL RfiLQ PodRRQP Fre"
> > > Generator=diyone7 (version 7.49+02(dev))
> > > Prefetch=0:x=F,0:y=W,1:y=F,1:x=T
> > > Com=Rf Fr
> > > Orig=DMB.STdWW Rfe PodRWPL RfiLQ PodRRQP Fre
> > > {
> > > 0:X1=x; 0:X3=y;
> > > 1:X1=y; 1:X3=z; 1:X6=x;
> > > }
> > >  P0          | P1            ;
> > >  MOV W0,#1   | LDR W0,[X1]   ;
> > >  STR W0,[X1] | MOV W2,#1     ;
> > >  DMB ST      | STLR W2,[X3]  ;
> > >  MOV W2,#1   | LDAPR W4,[X3] ;
> > >  STR W2,[X3] | LDR W5,[X6]   ;
> > > exists
> > > (1:X0=1 /\ 1:X4=1 /\ 1:X5=0)
> 
> (you can also run this yourself, since 'Q' is supported in the .cat file
> I contributed to herdtools7)
> 
> Test MP+dmb.sy+popl-rfilq-poqp Allowed
> States 4
> 1:X0=0; 1:X4=1; 1:X5=0;
> 1:X0=0; 1:X4=1; 1:X5=1;
> 1:X0=1; 1:X4=1; 1:X5=0;
> 1:X0=1; 1:X4=1; 1:X5=1;
> Ok
> Witnesses
> Positive: 1 Negative: 3
> Condition exists (1:X0=1 /\ 1:X4=1 /\ 1:X5=0)
> Observation MP+dmb.sy+popl-rfilq-poqp Sometimes 1 3
> Time MP+dmb.sy+popl-rfilq-poqp 0.01
> Hash=61858b7b59a6310d869f99cd05718f96
> 
> > There's also read-write ordering, in the form of the LB pattern:
> > 
> > P0(int *x, int *y, int *z)
> > {
> > 	r0 = READ_ONCE(*x);
> > 	smp_store_release(z, 1);
> > 	r1 = smp_load_acquire(z);
> > 	WRITE_ONCE(*y, 1);
> > }
> > 
> > P1(int *x, int *y)
> > {
> > 	r2 = READ_ONCE(*y);
> > 	smp_mp();
> > 	WRITE_ONCE(*x, 1);
> > }
> > 
> > exists (0:r0=1 /\ 1:r2=1)
> 
> The access types are irrelevant to the acquire/release primitives, so yes
> that's also allowed.
> 
> > Would this be allowed if smp_load_acquire() was implemented with LDAPR?
> > If the answer is yes then we will have to remove the rfi-rel-acq and
> > rel-rf-acq-po relations from the memory model entirely.
> 
> I don't understand what you mean by "rfi-rel-acq-po", and I assume you mean
> rel-rfi-acq-po for the other? Sounds like I'm confused here.

[Your reply about 1/2 suggests that you've figured this out now, IAC ...]

"rfi-rel-acq" (as Alan wrote, and as I wrote before my question above...)
is defined and currently used in linux-kernel.cat (and, FWIW, it has been
so since when we upstreamed LKMM).

My point is that, as exemplified by the two tests I reported above, this
relation already prevents you from implementing your acquire with LDAPR;
so my/our question was not "can you run herd7 for me?" but rather "do you
think that changes are needed to the .cat file?".

This question goes back _at least_ to:

  http://lkml.kernel.org/r/1519301990-11766-1-git-send-email-parri.andrea@gmail.com

(see, in part., the "IMPORTANT" note at the bottom of the commit message)
and that discussion later resulted in:

  0123f4d76ca63b ("riscv/spinlock: Strengthen implementations with fences")
  5ce6c1f3535fa8 ("riscv/atomic: Strengthen implementations with fences")

(the latest _draft_ of the RISC-V specification, as pointed out by Daniel,

  https://github.com/riscv/riscv-isa-manual, Appendix A.5

 includes our "Linux mapping", although the currently-recommended mapping
 differs and involves a "fence.tso [+ any acquire, including RCpc .aq]").

My understanding is that your answer to this question is "Yes", but I am
not sure about the "How/Which changes?"; of course, an answer his question
_in the form_ of PATCHes would be appreciated! (but please also consider
that I'll be offline for most of the time until next Monday.)

  Andrea


> 
> Will

Powered by blists - more mailing lists