Re: "Verifying and Optimizing Compact NUMA-Aware Locks on Weak Memory Models"

From: Will Deacon
Date: Tue Sep 13 2022 - 07:24:31 EST


On Fri, Aug 26, 2022 at 01:42:19PM -0700, Paul E. McKenney wrote:
> On Fri, Aug 26, 2022 at 01:10:39PM -0400, Alan Stern wrote:
> > On Fri, Aug 26, 2022 at 06:23:24PM +0200, Peter Zijlstra wrote:
> > > On Fri, Aug 26, 2022 at 05:48:12AM -0700, Paul E. McKenney wrote:
> > > > Hello!
> > > >
> > > > I have not yet done more than glance at this one, but figured I should
> > > > send it along sooner rather than later.
> > > >
> > > > "Verifying and Optimizing Compact NUMA-Aware Locks on Weak
> > > > Memory Models", Antonio Paolillo, Hernán Ponce-de-León, Thomas
> > > > Haas, Diogo Behrens, Rafael Chehab, Ming Fu, and Roland Meyer.
> > > > https://arxiv.org/abs/2111.15240
> > > >
> > > > The claim is that the queued spinlocks implementation with CNA violates
> > > > LKMM but actually works on all architectures having a formal hardware
> > > > memory model.
> > > >
> > > > Thoughts?
> > >
> > > So the paper mentions the following defects:
> > >
> > > - LKMM doesn't carry a release-acquire chain across a relaxed op
> >
> > That's right, although I'm not so sure this should be considered a
> > defect...
> >
> > > - some babbling about a missing propagation -- ISTR Linux if stuffed
> > > full of them, specifically we require stores to auto propagate
> > > without help from barriers
> >
> > Not a missing propagation; a late one.
> >
> > Don't understand what you mean by "auto propagate without help from
> > barriers".
> >
> > > - some handoff that is CNA specific and I've not looked too hard at
> > > presently.
> > >
> > >
> > > I think we should address that first one in LKMM, it seems very weird to
> > > me a RmW would break the chain like that.
> >
> > An explicitly relaxed RMW (atomic_cmpxchg_relaxed(), to be precise).
> >
> > If the authors wanted to keep the release-acquire chain intact, why not
> > use a cmpxchg version that has release semantics instead of going out of
> > their way to use a relaxed version?
> >
> > To put it another way, RMW accesses and release-acquire accesses are
> > unrelated concepts. You can have one without the other (in principle,
> > anyway). So a relaxed RMW is just as capable of breaking a
> > release-acquire chain as any other relaxed operation is.
> >
> > > Is there actual hardware that
> > > doesn't behave?
> >
> > Not as far as I know, although that isn't very far. Certainly an
> > other-multicopy-atomic architecture would make the litmus test succeed.
> > But the LKMM does not require other-multicopy-atomicity.
>
> My first attempt with ppcmem suggests that powerpc does -not- behave
> this way. But that surprises me, just on general principles. Most likely
> I blew the litmus test shown below.
>
> Thoughts?
>
> Thanx, Paul
>
> ------------------------------------------------------------------------
>
> PPC MP+lwsyncs+atomic
> "LwSyncdWW Rfe LwSyncdRR Fre"
> Cycle=Rfe LwSyncdRR Fre LwSyncdWW
> {
> 0:r2=x; 0:r4=y;
> 1:r2=y; 1:r5=2;
> 2:r2=y; 2:r4=x;
> }
> P0 | P1 | P2 ;
> li r1,1 | lwarx r1,r0,r2 | lwz r1,0(r2) ;
> stw r1,0(r2) | stwcx. r5,r0,r2 | lwsync ;
> lwsync | | lwz r3,0(r4) ;
> li r3,1 | | ;
> stw r3,0(r4) | | ;
> exists (1:r1=1 /\ 2:r1=2 /\ 2:r3=0)

Just catching up on this, but one possible gotcha here is if you have an
architecture with native load-acquire on P2 and then you move P2 to the end
of P1. e.g. at a high-level:


P0 P1
Wx = 1 RmW(y) // xchg() 1 => 2
WyRel = 1 RyAcq = 2
Rx = 0

arm64 forbids this, but it's not "natural" to the hardware and I don't
know what e.g. risc-v would say about it.

Will