Re: [RFC PATCH 08/21] KVM: x86: Add kvm_x86_ops hook to short circuit emulation

From: Sean Christopherson
Date: Mon Aug 19 2019 - 18:01:53 EST

On Thu, Aug 15, 2019 at 05:47:12PM -0700, Andy Lutomirski wrote:
> >> On Jul 29, 2019, at 7:49 PM, Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote:
> >>
> >> On Sat, Jul 27, 2019 at 10:38:03AM -0700, Andy Lutomirski wrote:
> >> On Fri, Jul 26, 2019 at 10:52 PM Sean Christopherson
> >> <sean.j.christopherson@xxxxxxxxx> wrote:
> >>>
> >>> Similar to the existing AMD #NPF case where emulation of the current
> >>> instruction is not possible due to lack of information, virtualization
> >>> of Intel SGX will introduce a scenario where emulation is not possible
> >>> due to the VMExit occurring in an SGX enclave. And again similar to
> >>> the AMD case, emulation can be initiated by kvm_mmu_page_fault(), i.e.
> >>> outside of the control of the vendor-specific code.
> >>>
> >>> While the cause and architecturally visible behavior of the two cases
> >>> is different, e.g. Intel SGX will inject a #UD whereas AMD #NPF is a
> >>> clean resume or complete shutdown, the impact on the common emulation
> >>> code is identical: KVM must stop emulation immediately and resume the
> >>> guest.
> >>>
> >>> Replace the exisiting need_emulation_on_page_fault() with a more generic
> >>> is_emulatable() kvm_x86_ops callback, which is called unconditionally
> >>> by x86_emulate_instruction().
> >>
> >> Having recently noticed that emulate_ud() is broken when the guest's
> >> TF is set, I suppose I should ask: does your new code function
> >> sensibly when TF is set?
> >
> > Barring a VMX fault injection interaction I'm not thinking of, yes. The
> > SGX reaction to the #UD VM-Exit is to inject a #UD and resume the guest,
> > pending breakpoints shouldn't be affected in any way (unless some other
> > part of KVM mucks with them, e.g. when guest single-stepping is enabled).
> What I mean is: does the code actually do what you think it does if TF is
> set? Right now, as I understand it, the KVM emulation code has a bug in
> which some emulated faults also inject #DB despite the fact that the
> instruction faulted, and the #DB seems to take precedence over the original
> fault. This confuses the guest.

Yes. The proposed change is to inject the #UD instead of calling into the
emulator, and by inspection I've verified that all code that injects a #DB
is either contained within the emulator or is mutually exclusive with an
intercepted #UD. It's a qualified yes because I don't have an actual
testcase to verify my literacy. I'll look into adding a test, either to
the selftest/x86/sgx or to kvm-unit-tests.