Re: [PATCH v5 04/19] KVM:x86: Refresh CPUID on write to guest MSR_IA32_XSS

[Yang, Weijiang]

On 8/5/2023 2:27 AM, Sean Christopherson wrote:
> On Thu, Aug 03, 2023, Yang Weijiang wrote:
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 0b9033551d8c..5d6d6fa33e5b 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -3780,10 +3780,12 @@ int kvm_set_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
> >   		 * IA32_XSS[bit 8]. Guests have to use RDMSR/WRMSR rather than
> >   		 * XSAVES/XRSTORS to save/restore PT MSRs.
> >   		 */
> > -		if (data & ~kvm_caps.supported_xss)
> > +		if (data & ~vcpu->arch.guest_supported_xss)
> Hmm, this is arguably wrong for userspace-initiated writes, as it would prevent
> userspace from restoring MSRs before CPUID.
>
> And it would make the handling of MSR_IA32_XSS writes inconsistent just within
> this case statement.  The initial "can this MSR be written at all" check would
> *not* honor guest CPUID for host writes, but then the per-bit check *would* honor
> guest CPUID for host writes.
>
> But if we exempt host writes, then we'll end up with another mess, as exempting
> host writes for MSR_KVM_GUEST_SSP would let the guest coerce KVM into writing an
> illegal value by modifying SMRAM while in SMM.
>
> Blech.
>
> If we can get away with it, i.e. not break userspace, I think my preference is
> to enforce guest CPUID for host accesses to XSS, XFD, XFD_ERR, etc.  I'm 99%
> certain we can make that change, because there are many, many MSRs that do NOT
> exempt host writes, i.e. the only way this would be a breaking change is if
> userspace is writing things like XSS before KVM_SET_CPUID2, but other MSRs after
> KVM_SET_CPUID2.
>
> I'm pretty sure I've advocated for the exact opposite in the past, i.e. argued
> that KVM's ABI is to not enforce ordering between KVM_SET_CPUID2 and KVM_SET_MSR.
> But this is becoming untenable, juggling the dependencies in KVM is complex and
> is going to result in a nasty bug at some point.
>
> For this series, lets just tighten the rules for XSS, i.e. drop the host_initated
> exemption.  And in a parallel/separate series, try to do a wholesale cleanup of
> all the cases that essentially allow userspace to do KVM_SET_MSR before KVM_SET_CPUID2.
OK, will do it for this series and investigate for other MSRs.
Thanks!