Re: [RFC PATCH 0/6] KVM: arm64: Errata management for VM Live migration

[Marc Zyngier <maz@kernel.org>]

On Tue, 22 Oct 2024 11:58:02 +0100,
Cornelia Huck <cohuck@redhat.com> wrote:
> 
> On Fri, Oct 18 2024, Oliver Upton <oliver.upton@linux.dev> wrote:
> 
> > On Fri, Oct 18, 2024 at 02:17:17PM +0100, Marc Zyngier wrote:
> >> On Thu, 17 Oct 2024 18:16:04 +0100,
> >> Eric Auger <eric.auger@redhat.com> wrote:
> >> > 
> >> > Hi Shameer,
> >> > 
> >> > On 10/17/24 17:49, Cornelia Huck wrote:
> >> 
> >> > > Speaking of QEMU: Do you maybe already have some prototype code that
> >> > > tries to do something with the setup here? (I don't think QEMU currently
> >> > > mucks around with MIDR and friends when running with KVM; I wonder what
> >> > > it should provide to the guest and if it should care to set something as
> >> > > a base level that gives guests not using the hypercall a chance to work
> >> > > properly.)
> >> > >
> >> > As discussed during the KVM forum we are working on a qemu integration
> >> > for writable ID regs. The first goal is to be able to specialize the
> >> > host passthrough model (custom host model). Maybe this will trigger more
> >> > discussions on named models too. This is complementary to the
> >> > MIDR/REVIDR problematic and I hope we will be able to consolidate our
> >> > works at some point.
> >> 
> >> Complementary to the MIDR/REVIDR work, I would also like to make
> >> MIDR/REVIDR writable when this scheme is available. Ideally reporting
> >> a synthetic CPU description (with MIDR_EL1.Implementer returning 0,
> >> and the rest being VMM-specific, but with a clear definition for the
> >> IMPDEF fields so that we can version the ABI).
> >
> > When the VMM is using the hypercall mechanism to describe
> > implementations, 100% agree.
> >
> >> Thoughts?
> >
> > I think we should go a step further and allow userspace full control of
> > known fields in these registers, even for a nonzero Implementer code.
> >
> > We're already affording userspace full control of what implementation(s)
> > the guest sees anyway via hypercalls, so there isn't much left for KVM
> > to enforce.
> >
> > Ignoring errata, it'd let folks spoof an old implementation on newer
> > hardware for testing, especially when dealing w/ older software.
> 
> So basically, the VMM would have to make a choice which guests it wants
> to support? Modern guests, which are aware through the hypercall
> interface what is going on and are able to parse any information in our
> synthetic MIDR correctly, or older guests, which should be presented
> with a MIDR that makes them hopefully act in the way we want it to act?
>
> I guess that would mean that for full migration support between
> different hosts, we'd want enlightned guests, and for (some) other
> cases, we could generate a configuration that will hopefully work?

That's sums it up, I think. The way I think of it is that, at VM
creation, you decide what you want to support:

- homogeneous migration: that's basically what we have today, with
  some limited flexibility.

- heterogeneous migration: synthetic MIDR, discovery hypercall, the
  works.

> (Spoofing for testing things sounds useful regardless.)

My problem with that is that there is no difference in ABI between "I
want this for testing" and "I expect this to fully work". The latter
is obviously not achievable...

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.