Re: [PATCH tip/locking/core v4 1/6] powerpc: atomic: Make *xchg and *cmpxchg a full barrier

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

[-- Attachment #1: Type: text/plain, Size: 4474 bytes --]

On Wed, Oct 14, 2015 at 11:04:19PM +0200, Peter Zijlstra wrote:
> On Wed, Oct 14, 2015 at 01:19:17PM -0700, Paul E. McKenney wrote:
> > Suppose we have something like the following, where "a" and "x" are both
> > initially zero:
> > 
> > 	CPU 0				CPU 1
> > 	-----				-----
> > 
> > 	WRITE_ONCE(x, 1);		WRITE_ONCE(a, 2);
> > 	r3 = xchg(&a, 1);		smp_mb();
> > 					r3 = READ_ONCE(x);
> > 
> > If xchg() is fully ordered, we should never observe both CPUs'
> > r3 values being zero, correct?
> > 
> > And wouldn't this be represented by the following litmus test?
> > 
> > 	PPC SB+lwsync-RMW2-lwsync+st-sync-leading
> > 	""
> > 	{
> > 	0:r1=1; 0:r2=x; 0:r3=3; 0:r10=0 ; 0:r11=0; 0:r12=a;
> > 	1:r1=2; 1:r2=x; 1:r3=3; 1:r10=0 ; 1:r11=0; 1:r12=a;
> > 	}
> > 	 P0                 | P1                 ;
> > 	 stw r1,0(r2)       | stw r1,0(r12)      ;
> > 	 lwsync             | sync               ;
> > 	 lwarx  r11,r10,r12 | lwz r3,0(r2)       ;
> > 	 stwcx. r1,r10,r12  | ;
> > 	 bne Fail0          | ;
> > 	 mr r3,r11          | ;
> > 	 Fail0:             | ;
> > 	exists
> > 	(0:r3=0 /\ a=2 /\ 1:r3=0)
> > 
> > I left off P0's trailing sync because there is nothing for it to order
> > against in this particular litmus test.  I tried adding it and verified
> > that it has no effect.
> > 
> > Am I missing something here?  If not, it seems to me that you need
> > the leading lwsync to instead be a sync.
> 
> So the scenario that would fail would be this one, right?
> 
> a = x = 0
> 
> 	CPU0				CPU1
> 
> 	r3 = load_locked (&a);
> 					a = 2;
> 					sync();
> 					r3 = x;
> 	x = 1;
> 	lwsync();
> 	if (!store_cond(&a, 1))
> 		goto again
> 
> 
> Where we hoist the load way up because lwsync allows this.

That scenario would end up with a==1 rather than a==2.

> I always thought this would fail because CPU1's store to @a would fail
> the store_cond() on CPU0 and we'd do the 'again' thing, re-issuing the
> load and now seeing the new value (2).

The stwcx. failure was one thing that prevented a number of other
misordering cases.  The problem is that we have to let go of the notion
of an implicit global clock.

To that end, the herd tool can make a diagram of what it thought
happened, and I have attached it.  I used this diagram to try and force
this scenario at https://www.cl.cam.ac.uk/~pes20/ppcmem/index.html#PPC,
and succeeded.  Here is the sequence of events:

o	Commit P0's write.  The model offers to propagate this write
	to the coherence point and to P1, but don't do so yet.

o	Commit P1's write.  Similar offers, but don't take them up yet.

o	Commit P0's lwsync.

o	Execute P0's lwarx, which reads a=0.  Then commit it.

o	Commit P0's stwcx. as successful.  This stores a=1.

o	Commit P0's branch (not taken).

o	Commit P0's final register-to-register move.

o	Commit P1's sync instruction.

o	There is now nothing that can happen in either processor.
	P0 is done, and P1 is waiting for its sync.  Therefore,
	propagate P1's a=2 write to the coherence point and to
	the other thread.

o	There is still nothing that can happen in either processor.
	So pick the barrier propagate, then the acknowledge sync.

o	P1 can now execute its read from x.  Because P0's write to
	x is still waiting to propagate to P1, this still reads
	x=0.  Execute and commit, and we now have both r3 registers
	equal to zero and the final value a=2.

o	Clean up by propagating the write to x everywhere, and
	propagating the lwsync.

And the "exists" clause really does trigger: 0:r3=0; 1:r3=0; [a]=2;

I am still not 100% confident of my litmus test.  It is quite possible
that I lost something in translation, but that is looking less likely.

> > Of course, if I am not missing something, then this applies also to the
> > value-returning RMW atomic operations that you pulled this pattern from.
> > If so, it would seem that I didn't think through all the possibilities
> > back when PPC_ATOMIC_EXIT_BARRIER moved to sync...  In fact, I believe
> > that I worried about the RMW atomic operation acting as a barrier,
> > but not as the load/store itself.  :-/
> 
> AARGH64 does something very similar; it does something like:
> 
> 	ll
> 	...
> 	sc-release
> 
> 	mb
> 
> Which I assumed worked for the same reason, any change to the variable
> would fail the sc, and we go for round 2, now observing the new value.

I have to defer to Will on this one.  You are right that ARM and PowerPC
do have similar memory models, but there are some differences.

							Thanx, Paul

[-- Attachment #2: SB+lwsync-RMW2-lwsync+st-sync-leading.pdf --]
[-- Type: application/pdf, Size: 21963 bytes --]