[libvirt] [PATCH 15/18] qemu: Support input model=virtio-{non-}transitional

[Andrea Bolognani]

On Thu, 2019-01-24 at 12:05 -0500, Cole Robinson wrote:
> On 01/24/2019 09:48 AM, Andrea Bolognani wrote:
> > Now that you mention it, for devices such as <input> and <disk>,
> > where we're introducing the model attribute just now, we should make
> > sure that we expand the default correctly, such that eg.
> > 
> >   <disk type='file' device='disk'>
> >     <target dev='vda' bus='virtio'/>
> >   </disk>
> > 
> > is formatted back as
> > 
> >   <disk type='file' device='disk' model='virtio'>
> >     <target dev='vda' bus='virtio'/>
> >   </disk>
> > 
> > We might have to be careful when formatting the XML for migration
> > and scrub model='virtio' in that case in order not to break
> > backwards migration, as we already do for plenty of other cases.
> > 
> > Once that's done, then you'll be guaranteed the model above is
> > never going to be DEFAULT.
> 
> IMO extending the XML we generate by default should only be done for
> very good reasons, the migration compat issue is not a trivial extension
> and comes with issues of its own.
> 
> For example, say we turn model=DEFAULT into model=VIRTIO in the qemu
> driver. In generic domain format code, we would do:
> 
> if (!MIGRATABLE || disk->model != VIRTIO)
>     <format disk->model in the XML>
> 
> But the MIGRATABLE checking is very crude. It doesn't consider version
> or capabilities of the remote libvirt. We could be migrating to an
> identical qemu and libvirt version, but we still drop the model=VIRTIO
> from the XML. Basically for the indefinite future, disk->model=VIRTIO in
> the XML is never going to be accounted for in migration. This can cause
> theoretical future issues if non-qemu drivers (or other qemu configs)
> support model=VIRTIO but it's not their default. For those non-qemu
> drivers, if disk->model=VIRTIO is present in the XML, the user
> explicitly requested it, but generic formatting code will never send the
> value during migration.
> 
> Let's say we skip the MIGRATABLE check. In fact for most of the cases in
> this patch series (except maybe for <controller> devs), this would work
> fine. If we are migrating to same or newer libvirt+qemu, all is good. If
> we migrate to older libvirt, it won't know to check disk->model at all
> so won't complain, maybe migration would fail due to qemu compat issues
> but that's unavoidable.
> 
> However, the NEXT time we make a similar change, let's say filling in
> disk->model=FOO for bus=ide, or more likely, input->model=QEMUTABLET for
> input->bus=usb + input->type=tablet. Now we have to blacklist
> model=QEMUTABLET from the migratable XML for the indefinite future,
> because older libvirt _may_ know enough to parse input->model, but
> doesn't have QEMUTABLET in the enum, and parsing fails. Okay, this is
> what the MIGRATABLE flag is all about. But it's exposing us to the issue
> with other drivers mentioned above.

We have qemuDomainDefFormatBufInternal(), where at least some of
the MIGRATABLE tweaks (PCI and USB controllers, serial devices) are
performed. I haven't looked into the call graph but I think it's
fair to assume that what's safe for pci-root is safe for pretty
much anything else :) So that should rule out at least the concern
with messing with other drivers.

> And this model=QEMUTABLET MIGRATABLE issue is theoretical but it's an
> exact case that has happened before, with controller models at least.
> Basically once you open that door of adding new default output to the
> XML for common cases, you can't ever close it again for that enum property.
> 
> Plus there's other downsides for outputting new properties by default:
> 
> * massive test suite churn: every single virtio disk in the XML2XML
> tests will now spit out model='virtio', plus all the other devices. This
> makes backporting suck

We cause test suite churn all the time, and while it's true that it
makes backporting more annoying most of the time it's just a matter
of using REGENERATE_OUTPUT=1 and notifying the reviewer.

> * libvirt downgrade issues. Previous times we added new output, like
> input type=keyboard, there were lots of complaints like: I upgraded
> libvirt via virt-preview to see if it fixed a bug, type=keyboard was
> added by default, I downgraded libvirt back to regular distro version,
> my VM disappeared, because the XML parser choked on unknown
> type=keyboard. Again in this particular case it likely won't cause those
> issues because model= is new everywhere, but the pattern basically can't
> be extended forward beyond this one instance.

People who install from virt-preview are going to be either
developers, who should have no problem dealing with the situation,
or users that are being guided through the analysis process in the
first place, so they won't be on their own when guests disappear.

> Here's the benefits I see of outputing model=virtio by default
> 
> * Reports more accurate XML. Minor IMO.

I remember various occasions in which not formatting a value caused
grief down the line, but I can't recall a single time when the
opposite was true. So maybe I'm focusing on this more than I should,
but past experiences have certainly colored my perception :)

> * Better chance of maintaining qemu compat across libvirt/qemu upgrades.
> Encoding model=virtio means we will always try that. Hypothetically 5
> years from now libvirt starts defaulting to
> model=virtio-non-transitional, you might upgrade libvirt and your
> windows virtio VM doesn't boot. Yeah that's not nice, but it's fixable
> (user sets model=virtio in the XML). Same reasoning basically applies to
> migration compat as well

If we don't format the value from day one, then we *cannot ever*
change the default. But I guess that ship has sailed because the
XML missing model='...' might come from an old libvirt version, so
we can't have that mean anything but model='virtio'.

> I don't think the benefits outweigh the downsides. Of course I may be
> missing plenty of things so please correct me if I'm wrong

I still think we should format the model, but let Dan (now CC'd)
weigh in as well :)

-- 
Andrea Bolognani / Red Hat / Virtualization