Re: [RFC PATCH 0/2] target/i386: SEV: allow running SNP guests with "-cpu host"

[Paolo Bonzini <pbonzini@redhat.com>]

On Thu, Jul 4, 2024 at 2:26 AM Michael Roth <michael.roth@amd.com> wrote:
> > Michael, any ideas?  Is there a way for the host to retrieve the supported
> > CPUID bits for SEV-SNP guests?
>
> If we want to support -cpu host, then I don't really see a way around
> needing to maintain a filter of some sort sanitize what gets passed to
> firmware. Generally, every new CPU model is likely to have some features
> which might be a liability security-wise to allow in SNP guests, so the
> CPUID validation is sort of a whitelist of curated features that make
> sense for guests and can be enabled securely in the context of SNP.
>
> Everything else would need to be filtered out, so we'd need to keep that
> list constantly updated.

It would be per new model and right now there are only a handful of
bits that have to be blocked; so it wouldn't be particularly bad.

> I think that may be possible, but do we have a strong use-case for
> supporting -cpu host in conjunction with SNP guests that this would be
> a worthwhile endeavor?

It's a common way to launch a guest if you're not interested in
migration (which is obviously the case for SNP right now), so it's
more like "why not". :)

> > One possibility is to set up a fake guest---either in QEMU or when KVM
> > starts---to do a LAUNCH_UPDATE for the CPUID page, but even that is not
> > perfect.  For example, I got
>
> Yah, the firmware-provided responses are more of a debug tool and not
> something I think we can rely on to enumerate capabilities.
>
> You could in theory take the ruleset in the PPR (Chapter 2, CPUID Policy
> Enforcement), turn that into something programmatic, and apply that
> against the host's CPUID values, but the policies are a bit more
> specific in some cases, and the PPR is per-CPU-model so both the rules
> and inputs can change from one host to the next.

Yeah, and if you mix that with knowledge of what KVM can/cannot
virtualize that doesn't exist in the processor (which isn't that
much), then you end up with something a lot like patch 2

It would be nice if the policy enforcement were changed to allow the
TSC deadline timer and X2APIC bits (you probably don't want TSC
adjust, that's the right call; and virt SSBD is not accessible because
you use V_SPEC_CTRL instead). But then there would be no way to find
out if the change actually happened.

> So I don't see a great way to leverage that to make things easier here.
> The manually-maintained filter you've proposed here seems more reliable
> to me.

Yep, I think I'll include that patch as the maintainability doesn't seem bad.

Paolo