Re: [PATCH v3 6/6] powerpc: atomic: Implement cmpxchg{,64}_* and atomic{,64}_cmpxchg_* variants

[Boqun Feng]

On Tue, Oct 13, 2015 at 10:32:59PM +0800, Boqun Feng wrote:
> On Tue, Oct 13, 2015 at 02:24:04PM +0100, Will Deacon wrote:
> > On Mon, Oct 12, 2015 at 10:14:06PM +0800, Boqun Feng wrote:
> > > Implement cmpxchg{,64}_relaxed and atomic{,64}_cmpxchg_relaxed, based on
> > > which _release variants can be built.
> > > 
> > > To avoid superfluous barriers in _acquire variants, we implement these
> > > operations with assembly code rather use __atomic_op_acquire() to build
> > > them automatically.
> > 
> > The "superfluous barriers" are for the case where the cmpxchg fails, right?
> 
> Yes.
> 
> > And you don't do the same thing for release, because you want to avoid a
> > barrier in the middle of the critical section?
> > 
> 
> Mostly because of the comments in include/linux/atomic.h:
> 
>  * For compound atomics performing both a load and a store, ACQUIRE
>  * semantics apply only to the load and RELEASE semantics only to the
>  * store portion of the operation. Note that a failed cmpxchg_acquire
>  * does -not- imply any memory ordering constraints.
> 
> so I thought only the barrier in cmpxchg_acquire() is conditional, and
> the barrier in cmpxchg_release() is not. Maybe we'd better call it out
> that cmpxchg *family* doesn't have any order guarantee if cmp fails, as
> a complement of
> 
> ed2de9f74ecb ("locking/Documentation: Clarify failed cmpxchg() memory ordering semantics")
> 
> Because it seems this commit only claims that the barriers in fully
> ordered version are conditional.
> 
> 
> If cmpxchg_release doesn't have order guarantee when failed, I guess I
> can implement it with a barrier in the middle as you mentioned:
> 
> 	unsigned int prev;
> 
> 	__asm__ __volatile__ (
> "1:	lwarx	%0,0,%2		
> 	cmpw	0,%0,%3\n\
> 	bne-	2f\n"
> 	PPC_RELEASE_BARRIER
> "	stwcx.	%4,0,%2\n\
> 	bne-	1b"
> 	"\n\
> 2:"
> 	: "=&r" (prev), "+m" (*p)
> 	: "r" (p), "r" (old), "r" (new)
> 	: "cc", "memory");
> 
> 	return prev;
> 
> 
> However, I need to check whether the architecture allows this and any
> other problem exists.
> 
> Besides, I don't think it's a good idea to do the "put barrier in the
> middle" thing in this patchset, because that seems a premature
> optimization and if we go further, I guess we can also replace the
> PPC_RELEASE_BARRIER above with a "sync" to implement a fully ordered

Correction: we can't just put the sync in the middle to implement a
fully ordered version. Sorry for that mistake..

Regards,
Boqun

> version cmpxchg(). Too much needs to investigate then..
> 
> > (just checking I understand your reasoning).
> > 
> 
> That actually helps me find a probably better implementation if allowed,
> thank you ;-)
> 
> Regards,
> Boqun

Attachment: signature.asc
Description: PGP signature