Re: [PATCH 4/4] locking: Introduce smp_cond_acquire()

[Will Deacon <will.deacon@arm.com>]

On Mon, Nov 02, 2015 at 10:08:24AM -0800, Linus Torvalds wrote:
> On Mon, Nov 2, 2015 at 5:29 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> > +#define smp_cond_acquire(cond) do {            \
> > +       while (!(cond))                         \
> > +               cpu_relax();                    \
> > +       smp_read_barrier_depends(); /* ctrl */  \
> > +       smp_rmb(); /* ctrl + rmb := acquire */  \
> > +} while (0)
> 
> This code makes absolutely no sense.
> 
> smp_read_barrier_depends() is about a memory barrier where there is a
> data dependency between two accesses. The "depends" is very much about
> the data dependency, and very much about *nothing* else.

Paul wasn't so sure, which I think is why smp_read_barrier_depends()
is already used in, for example, READ_ONCE_CTRL:

  http://lkml.kernel.org/r/20151007154003.GJ3910@linux.vnet.ibm.com

although I agree that this would pave the way for speculative stores on
Alpha and that seems like a heavy accusation to make.

> Your comment talks about control dependencies, but
> smp_read_barrier_depends() has absolutely nothing to do with a control
> dependency. In fact, it is explicitly a no-op on architectures like
> ARM and PowerPC that violate control dependencies.

In this case, control dependencies are only referring to READ -> WRITE
ordering, so they are honoured by ARM and PowerPC.

Will