On Thu, Apr 13, 2017 at 07:10:27PM +0200, Peter Zijlstra wrote:
> On Thu, Apr 13, 2017 at 09:57:55AM -0700, Paul E. McKenney wrote:
> > On Thu, Apr 13, 2017 at 06:24:09PM +0200, Peter Zijlstra wrote:
> > > On Thu, Apr 13, 2017 at 09:10:42AM -0700, Paul E. McKenney wrote:
> > > > On Thu, Apr 13, 2017 at 11:18:32AM +0200, Peter Zijlstra wrote:
> > > > > On Wed, Apr 12, 2017 at 09:55:43AM -0700, Paul E. McKenney wrote:
> > > > > > However, a little future-proofing is a good thing,
> > > > > > especially given that smp_mb__before_atomic() is only required to
> > > > > > provide acquire semantics rather than full ordering. This commit
> > > > > > therefore adds smp_mb__after_atomic() after the atomic_long_inc()
> > > > > > in sync_exp_work_done().
> > > > >
> > > > > Oh!? As far as I'm away the smp_mb__{before,after}_atomic() really
> > > > > must
> > > > > provide full MB, no confusion about that.
> > > > >
> > > > > We have other primitives for acquire/release.
> > > >
> > > > Hmmm... Rechecking atomic_ops.txt, it does appear that you are quite
> > > > correct. Adding Will and Dmitry on CC, but dropping this patch for now.
> > >
> > > I'm afraid that document is woefully out dated. I'm surprised it says
> > > anything on the subject.
> >
> > And there is some difference of opinion. Some believe that the
> > smp_mb__before_atomic() only guarantees acquire and smp_mb__after_atomic()
> > only guarantees release, but all current architectures provide full
> > ordering, as you noted and as stated in atomic_ops.txt.
>
> Which 'some' think it only provides acquire/release ?
>
> I made very sure -- when I renamed/audited/wrote all this -- that they
> indeed do a full memory barrier.
>
> > How do we decide?
>
> I say its a full mb, always was.
>
> People used it to create acquire/release _like_ constructs, because we
> simply didn't have anything else.
>
> Also, I think Linus once opined that acquire/release is part of a
> store/load (hence smp_store_release/smp_load_acquire) and not a barrier.
>
> > Once we do decide, atomic_ops.txt of course needs to be updated accordingly.
>
> There was so much missing there that I didn't quite know where to start.
Well, if there are no objections, I will fix up the smp_mb__before_atomic()
and smp_mb__after_atomic() pieces.
I suppose that one alternative is the new variant of kerneldoc, though
very few of these functions have comment headers, let alone kerneldoc
headers. Which reminds me, the question of spin_unlock_wait() and
spin_is_locked() semantics came up a bit ago. Here is what I believe
to be the case. Does this match others' expectations?
o spin_unlock_wait() semantics:
1. Any access in any critical section prior to the
spin_unlock_wait() is visible to all code following
(in program order) the spin_unlock_wait().
2. Any access prior (in program order) to the
spin_unlock_wait() is visible to any critical
section following the spin_unlock_wait().
o spin_is_locked() semantics: Half of spin_unlock_wait(),
but only if it returns false:
1. Any access in any critical section prior to the
spin_unlock_wait() is visible to all code following
(in program order) the spin_unlock_wait().
Thanx, Paul
