On Thu, Mar 31, 2022, Maciej S. Szmigiero wrote:
> On 30.03.2022 23:59, Sean Christopherson wrote:
> > On Thu, Mar 10, 2022, Maciej S. Szmigiero wrote:
> > > @@ -3627,6 +3632,14 @@ static void svm_complete_interrupts(struct kvm_vcpu *vcpu)
> > > if (!(exitintinfo & SVM_EXITINTINFO_VALID))
> > > return;
> > > + /* L1 -> L2 event re-injection needs a different handling */
> > > + if (is_guest_mode(vcpu) &&
> > > + exit_during_event_injection(svm, svm->nested.ctl.event_inj,
> > > + svm->nested.ctl.event_inj_err)) {
> > > + nested_svm_maybe_reinject(vcpu);
> >
> > Why is this manually re-injecting? More specifically, why does the below (out of
> > sight in the diff) code that re-queues the exception/interrupt not work? The
> > re-queued event should be picked up by nested_save_pending_event_to_vmcb12() and
> > propagatred to vmcb12.
>
> A L1 -> L2 injected event should either be re-injected until successfully
> injected into L2 or propagated to VMCB12 if there is a nested VMEXIT
> during its delivery.
>
> svm_complete_interrupts() does not do such re-injection in some cases
> (soft interrupts, soft exceptions, #VC) - it is trying to resort to
> emulation instead, which is incorrect in this case.
>
> I think it's better to split out this L1 -> L2 nested case to a
> separate function in nested.c rather than to fill
> svm_complete_interrupts() in already very large svm.c with "if" blocks
> here and there.
Ah, I see it now. WTF.
Ugh, commit 66fd3f7f901f ("KVM: Do not re-execute INTn instruction.") fixed VMX,
but left SVM broken.
Re-executing the INTn is wrong, the instruction has already completed decode and
execution. E.g. if there's there's a code breakpoint on the INTn, rewinding will
cause a spurious #DB.
KVM's INT3 shenanigans are bonkers, but I guess there's no better option given
that the APM says "Software interrupts cannot be properly injected if the processor
does not support the NextRIP field.". What a mess.
Anyways, for the common nrips=true case, I strongly prefer that we properly fix
the non-nested case and re-inject software interrupts, which should in turn
naturally fix this nested case. And for nrips=false, my vote is to either punt
and document it as a "KVM erratum", or straight up make nested require nrips.
Note, that also requires updating svm_queue_exception(), which assumes it will
only be handed hardware exceptions, i.e. hardcodes type EXEPT. That's blatantly
wrong, e.g. if userspace injects a software exception via KVM_SET_VCPU_EVENTS.
