[libvirt] [RFC] qemu: Use virtio-pci by default for mach-virt guests

[Andrea Bolognani]

On Mon, 2016-10-24 at 17:47 +0200, Pavel Hrdina wrote:
> > > I like that this makes pci truly the default in a simple manner, but 
> > > still allows switching back to mmio if necessary. On the other hand, it 
> > > puts the potential "switch" to decide whether or not to use mmio for all 
> > > devices down into the config of a single device, which is a bit weird to 
> > > explain. (On the other hand, how often will mmio be used in the future? 
> > > Maybe it doesn't matter if it's weird to explain...)
> > 
> > We really want to push for virtio-pci going forward as it has
> > a number of advantages, but there are still legacy guest OSs
> > out there that don't support virtio-pci at all yet we still
> > want to be able to run.
> 
> Changing the default is usually a tricky part and may break some users.
> I'm not sure that this will save the need to adapt management applications
> and users.  They will have to adapt in both cases to support legacy and
> new OSes based on libosinfo or another tool/database.  If we make virtio-pci
> the default one, which I think is the way it should be used, they will have
> to make sure that with new libvirt for the same OS they will fallback
> to virtio-mmio.
> 
> From what I can remember we've never done such change of default device
> model or default address and we always left it no management application
> or user to change the default to better model.  In case of management
> applications it's not an issue because they will make sure that the best
> device models and device address are used.
> 
> I'm not against changing the default to virtio-pci, but we may break
> things for some users and management tools like virt-manager unless they
> will adapt to this change.

You raise very good points, thanks for your input! :)

AFAICT the only use case that we'd risk breaking is installing
a legacy guest OS that doesn't support virtio-pci without
requiring the user to explicitly ask for virtio-mmio addresses.

Once libosinfo has learned about this, and virt-manager has
been updated to query libosinfo and switch to virtio-mmio
automatically if required, would you be okay with this change?

I think for aarch64 we're still in a phase where we can afford
to take some tradeoffs when it comes to compatibility, if
they're properly motivated: in this specific case, seeing as
basic stuff like device hotplug has simply never worked for
virtio-mmio, I'm fairly confident nobody will want to stick
with virtio-mmio for very long now that virtio-pci is finally
viable.

-- 
Andrea Bolognani / Red Hat / Virtualization