RE: [PATCH] KVM: x86/cpuid: Exclude unpermitted xfeatures for vcpu->arch.guest_supported_xcr0

[Tian, Kevin]

> From: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> Sent: Tuesday, January 25, 2022 12:30 AM
> 
> On 1/24/22 17:23, Sean Christopherson wrote:
> > On Mon, Jan 24, 2022, Like Xu wrote:
> >> On 24/1/2022 3:06 pm, Tian, Kevin wrote:
> >>>> From: Like Xu <like.xu.linux@xxxxxxxxx>
> >>>> Sent: Sunday, January 23, 2022 1:50 PM
> >>>>
> >>>> From: Like Xu <likexu@xxxxxxxxxxx>
> >>>>
> >>>> A malicious user space can bypass xstate_get_guest_group_perm() in
> the
> >>>> KVM_GET_SUPPORTED_CPUID mechanism and obtain unpermitted
> xfeatures,
> >>>> since the validity check of xcr0 depends only on guest_supported_xcr0.
> >>>
> >>> Unpermitted xfeatures cannot pass kvm_check_cpuid()...
> >>
> >> Indeed, 5ab2f45bba4894a0db4af8567da3efd6228dd010.
> >>
> >> This part of logic is pretty fragile and fragmented due to semantic
> >> inconsistencies between supported_xcr0 and guest_supported_xcr0
> >> in other three places:
> >
> > There are no inconsistencies, at least not in the examples below, as the
> examples
> > are intended to work in host context.  guest_supported_xcr0 is about what
> the guest
> > is/isn't allowed to access, it has no bearing on what host userspace
> can/can't do.
> > Or are you talking about a different type of inconsistency?
> 
> The extra complication is that arch_prctl(ARCH_REQ_XCOMP_GUEST_PERM)
> changes what host userspace can/can't do.  It would be easier if we
> could just say that KVM_GET_SUPPORTED_CPUID returns "the most" that
> userspace can do, but we already have the contract that userspace can
> take KVM_GET_SUPPORTED_CPUID and pass it straight to KVM_SET_CPUID2.
> 
> Therefore,  KVM_GET_SUPPORTED_CPUID must limit its returned values to
> what has already been enabled.
> 
> While reviewing the QEMU part of AMX support (this morning), I also
> noticed that there is no equivalent for guest permissions of
> ARCH_GET_XCOMP_SUPP.  This needs to know KVM's supported_xcr0, so it's
> probably best realized as a new KVM_CHECK_EXTENSION rather than as an
> arch_prctl.
> 

Would that lead to a weird situation where although KVM says no support
of guest permissions while the user can still request them via prctl()?

I wonder whether it's cleaner to do it still via prctl() if we really want to
enhance this part. But as you said then it needs a mechanism to know 
KVM's supported_xcr0 (and if KVM is not loaded then no guest permission
support at all)...

Thanks
Kevin