Re: [PATCH 4/7] KVM: SVM: Report NMI not allowed when Guest busy handling VNMI

[Maxim Levitsky <mlevitsk@redhat.com>]

On Thu, 2022-07-21 at 15:01 +0530, Shukla, Santosh wrote:
> 
> On 7/10/2022 9:38 PM, Maxim Levitsky wrote:
> > On Fri, 2022-06-17 at 20:29 +0530, Shukla, Santosh wrote:
> > > On 6/7/2022 6:42 PM, Maxim Levitsky wrote:
> > > > On Tue, 2022-06-07 at 16:10 +0300, Maxim Levitsky wrote:
> > > > > On Thu, 2022-06-02 at 19:56 +0530, Santosh Shukla wrote:
> > > > > > In the VNMI case, Report NMI is not allowed when the processor set the
> > > > > > V_NMI_MASK to 1 which means the Guest is busy handling VNMI.
> > > > > > 
> > > > > > Signed-off-by: Santosh Shukla <santosh.shukla@amd.com>
> > > > > > ---
> > > > > >  arch/x86/kvm/svm/svm.c | 6 ++++++
> > > > > >  1 file changed, 6 insertions(+)
> > > > > > 
> > > > > > diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> > > > > > index d67a54517d95..a405e414cae4 100644
> > > > > > --- a/arch/x86/kvm/svm/svm.c
> > > > > > +++ b/arch/x86/kvm/svm/svm.c
> > > > > > @@ -3483,6 +3483,9 @@ bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
> > > > > >         struct vmcb *vmcb = svm->vmcb;
> > > > > >         bool ret;
> > > > > >  
> > > > > > +       if (is_vnmi_enabled(vmcb) && is_vnmi_mask_set(vmcb))
> > > > > > +               return true;
> > > > > 
> > > > > How does this interact with GIF? if the guest does clgi, will the
> > > > > CPU update the V_NMI_MASK on its own If vGIF is enabled?
> > > > > 
> > > Yes.
> > > 
> > > > > What happens if vGIF is disabled and vNMI is enabled? KVM then intercepts
> > > > > the stgi/clgi, and it should then update the V_NMI_MASK?
> > > > > 
> > > No.
> > > 
> > > For both case - HW takes the V_NMI event at the boundary of VMRUN instruction.
> > 
> > How that is possible? if vGIF is disabled in L1, then L1 can't execute STGI/CLGI - 
> > that means that the CPU can't update the V_NMI, as it never sees the STGI/CLGI
> > beeing executed.
> > 
> 
> If vGIF is disabled then HW will take the vnmi event at the boundary of vmrun instruction.


I think I understand now, if vGIF is enabled, and V_NMI_MASK is set, and the guest does STGI, then nothing
new should be injected.

If V_NMI_MASK is not set, then svm_nmi_blocked will respect the HF_GIF_MASK, and on STGI interception,
the new NMI will be injected on VM entry by setting the V_NMI_PENDING.

So looks like it should work.

Thanks,
	Best regards,
		Maxim Levitsky


> 
> Thanks,
> Santosh
> 
> > Best regards,
> > 	Maxim Levitsky
> > 
> > > > > 
> > > > > > +
> > > > > >         if (!gif_set(svm))
> > > > > >                 return true;
> > > > > >  
> > > > > > @@ -3618,6 +3621,9 @@ static void svm_enable_nmi_window(struct kvm_vcpu *vcpu)
> > > > > >  {
> > > > > >         struct vcpu_svm *svm = to_svm(vcpu);
> > > > > >  
> > > > > > +       if (is_vnmi_enabled(svm->vmcb) && is_vnmi_mask_set(svm->vmcb))
> > > > > > +               return;
> > > > > 
> > > > > This might have hidden assumption that we will only enable NMI window when vNMI is masked.
> > > > 
> > > > Also what if vNMI is already pending?
> > > > 
> > > If V_NMI_MASK set, that means V_NMI is pending, if so then inject another V_NMI in next VMRUN.
> > > 
> > > Thanks,
> > > Santosh
> > > 
> > > > Best regards,
> > > > 	Maxim Levitsky
> > > > > 
> > > > > > +
> > > > > >         if ((vcpu->arch.hflags & (HF_NMI_MASK | HF_IRET_MASK)) == HF_NMI_MASK)
> > > > > >                 return; /* IRET will cause a vm exit */
> > > > > >  
> > > > > 
> > > > > Best regards,
> > > > >         Maxim Levitsky