Re: [Qemu-stable] [Qemu-devel] [PATCH 2/2] virtio-pci: Disable modern interface if backend without VIRTIO_F_VERSION_1

[Cornelia Huck]

On Fri, 9 Sep 2016 13:44:35 +0200
Maxime Coquelin <address@hidden> wrote:

> On 09/09/2016 01:20 PM, Cornelia Huck wrote:
> > On Fri, 9 Sep 2016 14:04:55 +0300
> > Marcel Apfelbaum <address@hidden> wrote:
> >
> >> On 09/09/2016 01:40 PM, Cornelia Huck wrote:
> >>> On Fri,  9 Sep 2016 12:14:32 +0200
> >>> Maxime Coquelin <address@hidden> wrote:
> >>>
> >>>> This patch makes pci devices plugging more robust, by not confusing
> >>>> guest with modern interface when the backend doesn't support
> >>>> VIRTIO_F_VERSION_1.
> >>>>
> >>>> Cc: Marcel Apfelbaum <address@hidden>
> >>>> Cc: Michael S. Tsirkin <address@hidden>
> >>>> Cc: address@hidden
> >>>> Signed-off-by: Maxime Coquelin <address@hidden>
> >>>> ---
> >>>>  hw/virtio/virtio-pci.c | 15 +++++++++++++++
> >>>>  hw/virtio/virtio-pci.h |  5 +++++
> >>>>  2 files changed, 20 insertions(+)
> >>>
> >>> Note that 11380b361 ("virtio: handle non-virtio-1-capable backend for
> >>> ccw") fixes this issue for ccw via the introduction of a
> >>> ->post_plugged() callback. Unfortunately, we did not find a good way to
> >>> make it work for pci back then.
> >>
> >> It seems that for ccw is enough to rewind dev->rev_max,
> >> sadly for pci we need to rewind a lot of settings/resources.
> >
> > Yes, that what I meant with 'more flexibility for ccw'.
> Maybe we could replace post_plugged with a pre_plugged approach?
> 
> In ->pre_plugged(), cww and pci would specify which features it can
> support using virtio_add_feature().
> Then we could call get_features() before ->device_plugged().

I think that would work for ccw (haven't looked at pci).

> 
> Doing this, both ccw and pci would have the needed information without
> having to rewind any settings.
> 
> Does that make sense?
> 
> But for now, I think it would be better to merge something in the spirit
> of this series (taking into account to remarks).
> Indeed, I think we want this fixed in stable, but the above proposal
> would be too huge for stable.

A 'just check for VERSION_1' approach would probably be best for stable.