Re: [Qemu-devel] Can we improve virtio data structures with QOM?

[Anthony Liguori]

On 05/31/2012 12:06 PM, Evgeny Voevodin wrote:
> On 31.05.2012 03:02, Anthony Liguori wrote:
> > On 05/31/2012 12:28 AM, Markus Armbruster wrote:
> > > Stefan Hajnoczi<address@hidden> writes:
> > >
> > > > On Wed, May 30, 2012 at 1:01 PM, Markus Armbruster<address@hidden>
> > > > wrote:
> > > > > Ordinary device models have a single state struct. The first member is
> > > > > a DeviceState or a specialization of DeviceState, e.g. a PCIDevice.
> > > > > Simple enough.
> > > >
> > > > I think Evgeny's virtio mmio patches change all this. In the recent
> > > > virtio-pci thread we were discussing how the virtio transport (mmio,
> > > > pci) and virtio devices (net, blk, etc) fit together. The email
> > > > thread is "Virtio-pci issues" from Evgeny Voevodin
> > > > <address@hidden>.
> > >
> > > Thanks for the pointer.
> > >
> > > It's been a couple of weeks. Evgeny, are you still pursuing this?
> > >
> > > > It probably makes sense to first merge Evgeny's virtio refactoring and
> > > > then ensure it's nicely mapped to QOM.
> > >
> > > Yes, no good attempting to do too much in one series. Nevertheless,
> > > having a sufficiently developed idea of the final state in mind helps.
> >
> > Basically, it should look like:
> >
> > VirtioPCIDevice is-a PCIDevice
> >
> > VirtioPCIDevice has-a link<VirtioDevice>
> >
> > VirtioDevice is-a DeviceState
> >
> > VirtioBlk is-a VirtioDevice
> > VirtioNet is-a VirtioDevice
> > ...
> >
> > VirtioPCIBlk is-a VirtioPCIDevice
> > VirtioPCIBlk has-a child<VirtioBlk>
> >
> > This gives us backwards compat while also providing a cleaner model.
> > VirtioPCIBlk et al would be deprecated eventually (but not any time soon).
> >
> > Regards,
> >
> > Anthony Liguori
> >
> > >
>
> Adding Peter in CC.
>
> Do not forget that we also have (will have, I hope) VirtioMMIO.
> And the last point I've stopped at (because of a lot of Tizen work) is:
>
> Added transport devices and bus:
> VirtioMMIOTransport is SysBusDevice
> VirtioPCITransport is PCIDevice
>
> VirtioMMIOTransport and VirtioPCITransport have VIRTIO_TRANSPORT_BUS.

You mean:

VirtioMMIOTransport implements VirtioTransportBus
VirtioPCITransport implements VirtioTransportBus

That's the best way to do this IMHO.  We also need:

VirtioDevice has-a link<VirtioTransportBus>

I think Having the word "Transport" is overly verbose.  I think VirtioMMIO and 
VirtioPCI is fine.

Regards,

Anthony Liguori

> Added back-ends:
> VirtioNet, VirtioBlk, etc are DeviceState and all of them
> have .bus_info initialised to VIRTIO_TRANSPORT_BUS.
>
> So, functionally we don't need to create new VirtioDevice class
> because it seems that for now we have all we need, but for cleaner
> model (and maybe at QOM's point of view too) I think it's worth to.
>
> This work is not completed and have issues pointed out here
> http://lists.gnu.org/archive/html/qemu-devel/2012-05/msg01421.html
>
> Since I'm not sure if I have time to finish this until our project
> milestone I can send an RFC patch-set and describe all it's issues to
> let somebody continue if this matter is urgent. But I'm still planning
> to continue this as soon as possible )