Re: [PATCH v2] KVM: Fix simultaneous NMIs

[Marcelo Tosatti <mtosatti@redhat.com>]

On Tue, Sep 20, 2011 at 04:56:02PM +0300, Avi Kivity wrote:
> On 09/20/2011 04:25 PM, Marcelo Tosatti wrote:
> >>
> >>  @@ -2827,6 +2828,7 @@ static int kvm_vcpu_ioctl_x86_set_mce(struct kvm_vcpu *vcpu,
> >>   static void kvm_vcpu_ioctl_x86_get_vcpu_events(struct kvm_vcpu *vcpu,
> >>   					struct kvm_vcpu_events *events)
> >>   {
> >>  +	process_nmi(vcpu);
> >>   	events->exception.injected =
> >>   		vcpu->arch.exception.pending&&
> >>   		!kvm_exception_is_soft(vcpu->arch.exception.nr);
> >>  @@ -2844,7 +2846,7 @@ static void kvm_vcpu_ioctl_x86_get_vcpu_events(struct kvm_vcpu *vcpu,
> >>   			KVM_X86_SHADOW_INT_MOV_SS | KVM_X86_SHADOW_INT_STI);
> >>
> >>   	events->nmi.injected = vcpu->arch.nmi_injected;
> >>  -	events->nmi.pending = vcpu->arch.nmi_pending;
> >>  +	events->nmi.pending = vcpu->arch.nmi_pending != 0;
> >>   	events->nmi.masked = kvm_x86_ops->get_nmi_mask(vcpu);
> >>   	events->nmi.pad = 0;
> >
> >nmi_queued should also be saved and restored. Not sure if its necessary
> >though.
> >
> >Should at least reset nmi_queued somewhere (set_vcpu_events?).
> 
> Did you miss the call to process_nmi()?

It transfers nmi_queued to nmi_pending with capping. What i mean is that
upon system reset, nmi_queued (which refers to pre-reset system state)
should be zeroed.

> We do have a small issue.  If we exit during NMI-blocked-by-STI and
> nmi_pending == 2, then we lose the second interrupt.  Should rarely
> happen, since external interrupts never exit in that condition, but
> it's a wart.

And the above system reset case, you should be able to handle it by
saving/restoring nmi_queued (so that QEMU can zero it in vcpu_reset).

> >
> >>  @@ -2864,6 +2866,7 @@ static int kvm_vcpu_ioctl_x86_set_vcpu_events(struct kvm_vcpu *vcpu,
> >>   			| KVM_VCPUEVENT_VALID_SHADOW))
> >>   		return -EINVAL;
> >>
> >>  +	process_nmi(vcpu);
> >
> >This should be after nmi fields are set, not before?
> >
> 
> It's actually to clear queued NMIs in case we set nmi_pending (which
> should never happen unless the machine is completely quiet, since
> it's asynchronous to the vcpu; same as the IRR).

OK.