On Fri, Apr 26, 2024, Weijiang Yang wrote:
> On 4/26/2024 2:14 AM, Sean Christopherson wrote:
> > Extend KVM's suppression of failures due to a userspace access to an
> > unsupported, but advertised as a "to save" MSR to all MSRs, not just those
> > that happen to reach the default case statements in kvm_get_msr_common()
> > and kvm_set_msr_common(). KVM's soon-to-be-established ABI is that if an
> > MSR is advertised to userspace, then userspace is allowed to read the MSR,
> > and write back the value that was read, i.e. why an MSR is unsupported
> > doesn't change KVM's ABI.
> >
> > Practically speaking, this is very nearly a nop, as the only other paths
> > that return KVM_MSR_RET_UNSUPPORTED are {svm,vmx}_get_feature_msr(), and
> > it's unlikely, though not impossible, that userspace is using KVM_GET_MSRS
> > on unsupported MSRs.
> >
> > The primary goal of moving the suppression to common code is to allow
> > returning KVM_MSR_RET_UNSUPPORTED as appropriate throughout KVM, without
> > having to manually handle the "is userspace accessing an advertised"
> > waiver. I.e. this will allow formalizing KVM's ABI without incurring a
> > high maintenance cost.
> >
> > Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> > ---
> > arch/x86/kvm/x86.c | 27 +++++++++------------------
> > 1 file changed, 9 insertions(+), 18 deletions(-)
> >
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 04a5ae853774..4c91189342ff 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -527,6 +527,15 @@ static __always_inline int kvm_do_msr_access(struct kvm_vcpu *vcpu, u32 msr,
> > if (ret != KVM_MSR_RET_UNSUPPORTED)
> > return ret;
> > + /*
> > + * Userspace is allowed to read MSRs, and write '0' to MSRs, that KVM
> > + * reports as to-be-saved, even if an MSR isn't fully supported.
> > + * Simply check that @data is '0', which covers both the write '0' case
> > + * and all reads (in which case @data is zeroed on failure; see above).
> > + */
> > + if (host_initiated && !*data && kvm_is_msr_to_save(msr))
> > + return 0;
> > +
>
> IMHO, it's worth to document above phrase into virt/kvm/api.rst KVM_{GET,
> SET}_MSRS sections as a note because when users space reads/writes MSRs
> successfully, it doesn't necessarily mean the operation really took effect.
> Maybe it's just due to the fact they're exposed in "to-be-saved" list.
Agreed, though I think I'd prefer to wait to officially document the behavior
until we have fully converted KVM's internals to KVM_MSR_RET_UNSUPPORTED. I'm
99% certain the behavior won't actually be as simple as "userspace can write '0'",
e.g. I know of at least one case where KVM allows '0' _or_ the KVM's non-zero
default.
In the case that I'm aware of (MSR_AMD64_TSC_RATIO), the "default" is a hardcoded
KVM constant, e.g. won't change based on underlying hardware, but there me be
other special cases lurking, and we won't know until we complete the conversion :-(
