Re: [PATCH v3 2/4] KVM: X86: Fix loss of exception which has not yet injected

[Paolo Bonzini]

On 24/08/2017 11:34, Wanpeng Li wrote:
> 2017-08-24 17:13 GMT+08:00 Wanpeng Li <kernellwp@xxxxxxxxx>:
>> 2017-08-24 16:57 GMT+08:00 Paolo Bonzini <pbonzini@xxxxxxxxxx>:
>>> On 24/08/2017 08:52, Wanpeng Li wrote:
>>>>> @@ -6862,6 +6876,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
>>>>>                                 kvm_x86_ops->enable_nmi_window(vcpu);
>>>>>                         if (kvm_cpu_has_injectable_intr(vcpu) || req_int_win)
>>>>>                                 kvm_x86_ops->enable_irq_window(vcpu);
>>>>> +                       WARN_ON(vcpu->arch.exception.pending);
>>>>
>>>> This WARN_ON() is suggested during the review of last version,
>>>> however, there are many cases in inject_pending_event() can result in
>>>> return directly w/ vcpu->arch.exception.pending is true. Actually I
>>>> have already catched the warning several times during the testing. I
>>>> think we should remove it when committing.
>>>
>>> No, it's a good thing that it's failing, because it's finding a bug.
>>> There's no such thing as an "exception window", so at the very least it
>>
>> Good point, the code looks good, I will fold it in next version.
>> However, I still can observe the warning.
>>
>> Regards,
>> Wanpeng Li
> 
> I observed sometimes both vcpu->arch.exception.pending and
> vcpu->arch.expception.injected are true before executing below codes:
> 
> if (vcpu->arch.exception.injected) {
>     kvm_x86_ops->queue_exception(vcpu);
>     return 0;
> }

More missing pieces:

1) kvm_x86_ops->queue_exception must be called in the
vcpu->arch.exception.pending case.  Compare it with the others, which
are calling enter_smm, kvm_x86_ops->set_irq, kvm_x86_ops->set_nmi.
	
2) here:

        if (!vcpu->arch.exception.pending ||
                !vcpu->arch.exception.injected) {
        queue:
                if (has_error && !is_protmode(vcpu))
                        has_error = false;
                if (reinject)
                        vcpu->arch.exception.injected = true;
                else
                        vcpu->arch.exception.pending = true;

you need to reset the other field, because you can get here from the
double-fault case.  Likewise below:

        if ((class1 == EXCPT_CONTRIBUTORY && class2 == EXCPT_CONTRIBUTORY)
                || (class1 == EXCPT_PF && class2 != EXCPT_BENIGN)) {
                /* generate double fault per SDM Table 5-5 */
                vcpu->arch.exception.pending = true;

injected must be cleared.

> Regards,
> Wanpeng Li
> 
>>
>>> should set req_immediate_exit to true.
>>>
>>> Does this help?
>>>
>>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>>> index b698b2f135a2..76d5a192be6c 100644
>>> --- a/arch/x86/kvm/x86.c
>>> +++ b/arch/x86/kvm/x86.c
>>> @@ -6365,14 +6365,20 @@ static int inject_pending_event(struct kvm_vcpu *vcpu, bool req_int_win)
>>>                 return 0;
>>>         }
>>>
>>> -       if (vcpu->arch.nmi_injected) {
>>> -               kvm_x86_ops->set_nmi(vcpu);
>>> -               return 0;
>>> -       }
>>> +       /*
>>> +        * Exceptions must be injected immediately, or the exception
>>> +        * frame will have the address of the NMI or interrupt handler.
>>> +        */
>>> +       if (!vcpu->arch.exception.pending) {
>>> +               if (vcpu->arch.nmi_injected) {
>>> +                       kvm_x86_ops->set_nmi(vcpu);
>>> +                       return 0;
>>> +               }
>>>
>>> -       if (vcpu->arch.interrupt.pending) {
>>> -               kvm_x86_ops->set_irq(vcpu);
>>> -               return 0;
>>> +               if (vcpu->arch.interrupt.pending) {
>>> +                       kvm_x86_ops->set_irq(vcpu);
>>> +                       return 0;
>>> +               }
>>>         }
>>>
>>>         if (is_guest_mode(vcpu) && kvm_x86_ops->check_nested_events) {
>>>
>>> Paolo