On Tue, Jan 29, 2019 at 10:18:20AM -0800, Suren Baghdasaryan wrote:
> On Tue, Jan 29, 2019 at 4:38 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Thu, Jan 24, 2019 at 01:15:18PM -0800, Suren Baghdasaryan wrote:
> > > + atomic_set(&group->polling, polling);
> > > + /*
> > > + * Memory barrier is needed to order group->polling
> > > + * write before times[] read in collect_percpu_times()
> > > + */
> > > + smp_mb__after_atomic();
> >
> > That's broken, smp_mb__{before,after}_atomic() can only be used on
> > atomic RmW operations, something atomic_set() is _not_.
>
> Oh, I didn't realize that. After reading the following example from
> atomic_ops.txt
That document it woefully out of date (and I should double check, but I
think we can actually delete it now). Please see
Documentation/atomic_t.txt
> I was under impression that smp_mb__after_atomic()
> would make changes done by atomic_set() visible:
>
> /* All memory operations before this call will
> * be globally visible before the clear_bit().
> */
> smp_mb__before_atomic();
> clear_bit( ... );
> /* The clear_bit() will be visible before all
> * subsequent memory operations.
> */
> smp_mb__after_atomic();
>
> but I'm probably missing something. Is there a more detailed
> description of these rules anywhere else?
See atomic_t.txt; but the difference is that clear_bit() is a RmW, while
atomic_set() is just a plain store.
> Meanwhile I'll change smp_mb__after_atomic() into smp_mb(). Would that
> fix the ordering?
It would work here; but I'm still trying to actually understand all
this. So while the detail would be fine, I'm not ready to judge the
over-all thing.
