Re: [Qemu-devel] [PATCH] Fix for crash after migration in virtio-rng on bi-endian targets

[Greg Kurz]

On Thu, 27 Nov 2014 14:38:42 +0530
Amit Shah <address@hidden> wrote:
> On (Thu) 27 Nov 2014 [16:48:10], David Gibson wrote:
> > VirtIO devices now remember which endianness they're operating in in order
> > to support targets which may have guests of either endianness, such as
> > powerpc.  This endianness state is transferred in a subsection of the
> > virtio device's information.
> > 
> > With virtio-rng this can lead to an abort after a loadvm hitting the
> > assert() in virtio_is_big_endian().  This can be reproduced by doing a
> > migrate and load from file on a bi-endian target with a virtio-rng device.
> > The actual guest state isn't particularly important to triggering this.
> > 
> > The cause is that virtio_rng_load_device() calls virtio_rng_process() which
> > accesses the ring and thus needs the endianness.  However,
> > virtio_rng_process() is called via virtio_load() before it loads the
> > subsections.  Essentially the ->load callback in VirtioDeviceClass should
> > only be used for actually reading the device state from the stream, not for
> > post-load re-initialization.
> 
> Agreed.
> 
> > This patch fixes the bug by moving the virtio_rng_process() after the call
> > to virtio_load().  Better yet would be to convert virtio to use vmsd and
> > have the virtio_rng_process() as a post_load callback, but that's a bigger
> > project for another day.
> > 

I remember discussions on IRC last spring where I agreed I would work on it. :)

> > This is bugfix, and should be considered for the 2.2 branch.
> 
> This is undoing most of 3902d49e13c2428bd6381cfdf183103ca4477c1f ,
> added Greg to CC list.
> 

This commit is indeed completely wrong: the load callback only makes sense
when there's something to read which is obviously not the case here... This
is definitely post load stuff :-\

Thanks ! :)

Reviewed-by: Greg Kurz <address@hidden>

--
Greg

> Did you try this on x86 guests, or with multiple rng devices?
> 
> (keeping context for Greg)
> 
> > Signed-off-by: David Gibson <address@hidden>
> > ---
> >  hw/virtio/virtio-rng.c | 15 ++++++++-------
> >  1 file changed, 8 insertions(+), 7 deletions(-)
> > 
> > diff --git a/hw/virtio/virtio-rng.c b/hw/virtio/virtio-rng.c
> > index e85a979..473c044 100644
> > --- a/hw/virtio/virtio-rng.c
> > +++ b/hw/virtio/virtio-rng.c
> > @@ -113,20 +113,22 @@ static void virtio_rng_save(QEMUFile *f, void *opaque)
> >  
> >  static int virtio_rng_load(QEMUFile *f, void *opaque, int version_id)
> >  {
> > +    VirtIORNG *vrng = opaque;
> > +    int ret;
> > +
> >      if (version_id != 1) {
> >          return -EINVAL;
> >      }
> > -    return virtio_load(VIRTIO_DEVICE(opaque), f, version_id);
> > -}
> > +    ret = virtio_load(VIRTIO_DEVICE(vrng), f, version_id);
> > +    if (ret != 0) {
> > +        return ret;
> > +    }
> >  
> > -static int virtio_rng_load_device(VirtIODevice *vdev, QEMUFile *f,
> > -                                  int version_id)
> > -{
> >      /* We may have an element ready but couldn't process it due to a quota
> >       * limit.  Make sure to try again after live migration when the quota 
> > may
> >       * have been reset.
> >       */
> > -    virtio_rng_process(VIRTIO_RNG(vdev));
> > +    virtio_rng_process(vrng);
> >  
> >      return 0;
> >  }
> > @@ -231,7 +233,6 @@ static void virtio_rng_class_init(ObjectClass *klass, 
> > void *data)
> >      vdc->realize = virtio_rng_device_realize;
> >      vdc->unrealize = virtio_rng_device_unrealize;
> >      vdc->get_features = get_features;
> > -    vdc->load = virtio_rng_load_device;
> >  }
> >  
> >  static void virtio_rng_initfn(Object *obj)
> 
> 
> Thanks,
> 
>               Amit
>