[PATCH] arm64: barrier: implement wfe-base smp_cond_load_acquire

[Will Deacon]

On Fri, Jun 17, 2016 at 03:27:42PM +0100, Mark Rutland wrote:
> On Fri, Jun 17, 2016 at 02:12:15PM +0100, Will Deacon wrote:
> > diff --git a/arch/arm64/include/asm/cmpxchg.h b/arch/arm64/include/asm/cmpxchg.h
> > index 510c7b404454..84b83e521edc 100644
> > --- a/arch/arm64/include/asm/cmpxchg.h
> > +++ b/arch/arm64/include/asm/cmpxchg.h
> > @@ -224,4 +224,56 @@ __CMPXCHG_GEN(_mb)
> >  	__ret;								\
> >  })
> >  
> > +#define __CMPWAIT_CASE(w, sz, name, acq, cl)				\
> > +static inline void __cmpwait_case_##name(volatile void *ptr,		\
> > +					 unsigned long val)		\
> > +{									\
> > +	unsigned long tmp;						\
> > +									\
> > +	asm volatile(							\
> > +	"	ld" #acq "xr" #sz "\t%" #w "[tmp], %[v]\n"		\
> > +	"	eor	%" #w "[tmp], %" #w "[tmp], %" #w "[val]\n"	\
> > +	"	cbnz	%" #w "[tmp], 1f\n"				\
> > +	"	wfe\n"							\
> > +	"1:"								\
> > +	: [tmp] "=&r" (tmp), [v] "+Q" (*(unsigned long *)ptr)		\
> > +	: [val] "r" (val)						\
> > +	: cl);								\
> > +}
> > +
> > +__CMPWAIT_CASE(w, b, acq_1, a, "memory");
> > +__CMPWAIT_CASE(w, h, acq_2, a, "memory");
> > +__CMPWAIT_CASE(w,  , acq_4, a, "memory");
> > +__CMPWAIT_CASE( ,  , acq_8, a, "memory");
> > +
> > +#undef __CMPWAIT_CASE
> 
> From my understanding of the intent, I believe that the asm is correct.

Cheers for having a look.

> I'm guessing from the way this and __CMPWAIT_GEN are parameterised that
> there is a plan for variants of this with different ordering semantics,
> though I'm having difficulty envisioning them. Perhaps I lack
> imagination. ;)

Originally I also had a cmpwait_relaxed implementation, since Peter had
a generic cmpwait utility. That might be resurrected some day, so I
figured leaving the code like this makes it easily extensible to other
memory-ordering semantics without adding much in the way of complexity.

> Is there any case for waiting with anything other than acquire
> semantics? If not, we could fold the acq parameter and acq_ prefix from
> the name parameter.
> 
> Do we expect this to be used outside of smp_cond_load_acquire? If not,
> can't we rely on the prior smp_load_acquire for the acquire semantics
> (and use pain LDXR* here)? Then the acq part can go from the cmpwait
> cases entirely.

Yeah, I think we can rely on the read-after-read ordering here and
use __cmpwait_relaxed to the job for the inner loop.

> Is there any case where we wouldn't want the memory clobber?

I don't think you'd need it for cmpwait_relaxed, because the CPU could
reorder stuff anyway, so anything the compiler does is potentially futile.
So actually, I can respin this without the clobber. I'll simplify
the __CMPWAIT_CASE macro to drop the last two parameters as well.

Will