On 31.03.2022 01:20, Sean Christopherson wrote:
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.
Note that KVM currently always tries to re-execute the current instruction
when asked to re-inject a #BP or a #OF, even when nrips are enabled.
Also, #BP (and #OF, too) is returned as type SVM_EXITINTINFO_TYPE_EXEPT,
not as SVM_EXITINTINFO_TYPE_SOFT (soft interrupt), so it should be
re-injected accordingly.
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.
This would also need making the #BP or #OF current instruction
re-execution conditional on (at least) nrips support.
I am not sure, however, whether this won't introduce any regressions.
That's why this patch set changed the behavior here only for the
L1 -> L2 case.
Another issue is whether a L1 hypervisor can legally inject a #VC
into its L2 (since these are never re-injected).
We still need L1 -> L2 event injection detection to restore the NextRIP
field when re-injecting an event that uses it.
And for nrips=false, my vote is to either punt
and document it as a "KVM erratum", or straight up make nested require nrips.
A quick Internet search shows that the first CPUs with NextRIP were the
second-generation Family 10h CPUs (Phenom II, Athlon II, etc.).
They started being released in early 2009, so we probably don't need to
worry about the non-nrips case too much.
For the nested case, orthodox reading of the aforementioned APM sentence
would mean that a L1 hypervisor is not allowed either to make use of such
event injection in the non-nrips case.
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.
svm_queue_exception() uses SVM_EVTINJ_TYPE_EXEPT, which is correct even
for software exceptions (#BP or #OF).
It does work indeed, as the self test included in this patch set
demonstrates.
Thanks,
Maciej
