Re: [Qemu-devel] [PATCH for-4.0 v2] virtio: Provide version-specific variants of virtio PCI devices

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Nov 20, 2018 at 01:27:05PM +0100, Andrea Bolognani wrote:
> On Mon, 2018-11-19 at 14:14 -0500, Michael S. Tsirkin wrote:
> > On Mon, Nov 19, 2018 at 07:56:38PM +0100, Cornelia Huck wrote:
> > > On Mon, 19 Nov 2018 13:42:58 -0500 "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > > > We have this assumption that if we force a choice then people will
> > > > choose the right thing but in practice they will do what we all do, play
> > > > with it until it kind of works and leave well alone afterwards.
> > > > That's at best - at worst give up and use an easier tool.
> > > 
> > > That implies that we (the developers) need to care and make sure that
> > > "model=virtio" gets them the best possible transport (i.e. on s390x,
> > > that would be ccw unless the user explicitly requests pci; I'm not sure
> > > what the situation with mmio is -- probably "use pci whenever
> > > possible"?) I think that's what libvirt already gives us today (I hope.)
> 
> The interface at the libvirt level is exactly "model=virtio", with
> that ultimately translating to virtio-*-pci or virtio-*-ccw or
> virtio-*-device or whatever else based on the architecture, machine
> type and other information about the guest.
> 
> > > What makes it messy on the pci side is that the "best option" actually
> > > depends on what kind of guest the user wants to run (if the guest is
> > > too old, you're stuck with transitional; if you want to reap the
> > > benefits of PCIe, you need non-transitional...)
> > 
> > Well it works now - connect it to a bus and it figures out whether it
> > should do transitional or not. You can force transitional in PCIe anyway
> > but then you are limited to about 15 devices - probably sufficient for
> > most people ...
> 
> That's not how it works, though: current virtio-*-pci devices will
> be transitional (and thus support older guest OS) or not based on
> the kind of slot you plug them into.
> 
> >From the management point of view that's problematic, because libvirt
> (which takes care of the virtual hardware, including assigning PCI
> addresses to devices) has no knowledge of the guest OS running on
> said hardware, and management apps (which know about the guest OS and
> can figure out its capabilities using libosinfo) don't want to be in
> the business of assigning PCI addresses themselves.
> 
> Having separate transitional and non-transitional variants solves the
> issue because now management apps can query libosinfo to figure out
> whether the guest OS supports non-transitional virtio devices, and
> based on that they can ask libvirt to use either the transitional or
> non-transitional variant; from that, libvirt will be able to choose
> the correct slot for the device.
> 
> None of the above quite works if we have a single variant that
> morphs based on the slot, as we have today.

So can we get an ack on the patchset then?

> -- 
> Andrea Bolognani / Red Hat / Virtualization