On Tue, Sep 06, 2011 at 08:55:42AM +0200, Paolo Bonzini wrote:
> On 09/06/2011 05:12 AM, David Gibson wrote:
> >I'm not "fixing ppc". I'm fixing a fundamental flaw in the protocol
> >implementation._So far_ I've only observed the effects on ppc, but
> >that doesn't mean they don't exist.
>
> Actually Michael is right. The implementation is correct on x86,
> though wrong anywhere else (perhaps s390?). On those architectures
> you do not need rmb() and wmb().
>
> So doing the __sync_synchronize() on ppc only as he proposed is
> wrong; but not doing it (and leaving only a compiler barrier) on x86
> is correct. See http://g.oswego.edu/dl/jmm/cookbook.html under
> "Multiprocessors".
Well, in any case, I've realised that we ought to merge the barriers
in virtio with those in qemu-barrier.h. Which are also unsafe on
non-x86.
I have a draft two patch series to do that, which I intend to post
soon. Unfortunately, I haven't been able to test it on ppc kvm yet,
because there are other bugs which are making qemu not even boot the
pseries machine on a ppc64 host. Still trying to work out what the
hell's going on there, smells like bad extension or masking of CTR
(SLOF seems to get stuck in a bdnz loop with something close to 2^57
in CTR).
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_
| _way_ _around_!
http://www.ozlabs.org/~dgibson
