Re: [Xen-devel] [PATCHv4 5/8] xen: use ticket locks for spin locks

[Tim Deegan <tim@xxxxxxx>]

At 14:13 +0100 on 11 May (1431353629), David Vrabel wrote:
> On 08/05/15 10:36, Jan Beulich wrote:
> >>
> >> +        }
> >> +    }
> >>      smp_mb();
> >>  }
> > 
> > The old code had smp_mb() before _and_ after the check - is it really
> > correct to drop the one before (or effectively replace it by smp_rmb()
> > in observe_{lock,head}())?
> 
> Typical usage is:
> 
>     d->is_dying = DOMDYING_dying;
>     spin_barrier(&d->domain_lock);
> 
> So there needs to be a barrier before we check that the lock is
> released.  i.e., I removed the wrong smp_mb().
> 
> I don't see the need for the second barrier since there's no stores in
> _spin_barrier() and observe_lock() and observe_head() both have their
> required read barriers.

You need the second smp_mb() to make sure that any post-barrier writes
don't happen before those reads (i.e. before the barrier).
E.g. evtchn_destroy() uses a barrier before modifying some state
(though in that case I think the read barrier might be enough because
the writes depend on reads, but you get the idea).

Tim.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel