Re: [PATCH RFC 3/5] x86: KVM: svm: clear interrupt shadow on all paths in skip_emulated_instruction()

[Sean Christopherson <sean.j.christopherson@intel.com>]

On Wed, Jul 31, 2019 at 01:27:53PM -0700, Jim Mattson wrote:
> On Wed, Jul 31, 2019 at 9:37 AM Paolo Bonzini <pbonzini@redhat.com> wrote:
> >
> > On 31/07/19 15:50, Vitaly Kuznetsov wrote:
> > > Jim Mattson <jmattson@google.com> writes:
> > >
> > >> On Thu, Jun 20, 2019 at 4:02 AM Vitaly Kuznetsov <vkuznets@redhat.com> wrote:
> > >>>
> > >>> Regardless of the way how we skip instruction, interrupt shadow needs to be
> > >>> cleared.
> > >>
> > >> This change is definitely an improvement, but the existing code seems
> > >> to assume that we never call skip_emulated_instruction on a
> > >> POP-SS/MOV-to-SS/STI. Is that enforced anywhere?
> > >
> > > (before I send v1 of the series) I looked at the current code and I
> > > don't think it is enforced, however, VMX version does the same and
> > > honestly I can't think of a situation when we would be doing 'skip' for
> > > such an instruction.... and there's nothing we can easily enforce from
> > > skip_emulated_instruction() as we have no idea what the instruction
> > > is...
> 
> Can't we still coerce kvm into emulating any instruction by leveraging
> a stale ITLB entry? The 'emulator' kvm-unit-test did this before the
> KVM forced emulation prefix was introduced, but I haven't checked to
> see if the original (admittedly fragile) approach still works. Also,
> for POP-SS, you could always force emulation by mapping the %rsp
> address beyond guest physical memory. The hypervisor would then have
> to emulate the instruction to provide bus-error semantics.
> 
> > I agree, I think a comment is worthwhile but we can live with the
> > limitation.
> 
> I think we can live with the limitation, but I'd really prefer to see
> a KVM exit with KVM_INTERNAL_ERROR_EMULATION for an instruction that
> kvm doesn't emulate properly. That seems better than just a comment
> that the virtual CPU doesn't behave as architected. (I realize that I
> am probably in the minority here.)

At a glance, the full emulator models behavior correctly, e.g. see
toggle_interruptibility() and setters of ctxt->interruptibility.

I'm pretty sure that leaves the EPT misconfig MMIO and APIC access EOI
fast paths as the only (VMX) path that would incorrectly handle a
MOV/POP SS.  Reading the guest's instruction stream to detect MOV/POP SS
would defeat the whole "fast path" thing, not to mention both paths aren't
exactly architecturally compliant in the first place.