Re: [PATCH v5 2/2] x86/svm: Use the virtual NMI when available

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 26/02/2026 1:20 pm, Abdelkareem Abdelsaamad wrote:
> With the Virtual NMI (vNMI), the pending NMI is simply stuffed into the VMCB
> and handed off to the hardware. There is no need for the artificial tracking
> of the NMI handling completion with the IRET instruction interception.
>
> Adjust the svm_inject_nmi to rather inject the NMIs using the vNMI Hardware
> accelerated feature when the AMD platform supports the vNMI.
>
> Signed-off-by: Abdelkareem Abdelsaamad <abdelkareem.abdelsaamad@xxxxxxxxxx>
> ---
>  xen/arch/x86/hvm/svm/intr.c | 6 ++++++
>  xen/arch/x86/hvm/svm/svm.c  | 1 +
>  xen/arch/x86/hvm/svm/vmcb.c | 2 ++
>  3 files changed, 9 insertions(+)
>
> diff --git a/xen/arch/x86/hvm/svm/intr.c b/xen/arch/x86/hvm/svm/intr.c
> index 6453a46b85..112a393211 100644
> --- a/xen/arch/x86/hvm/svm/intr.c
> +++ b/xen/arch/x86/hvm/svm/intr.c
> @@ -33,6 +33,12 @@ static void svm_inject_nmi(struct vcpu *v)
>      u32 general1_intercepts = vmcb_get_general1_intercepts(vmcb);
>      intinfo_t event;
>  
> +    if ( vmcb->_vintr.fields.vnmi_enable )
> +    {
> +        vmcb->_vintr.fields.vnmi_pending = 1;
> +        return;
> +    }
> +
>      event.raw = 0;
>      event.v = true;
>      event.type = X86_ET_NMI;
> diff --git a/xen/arch/x86/hvm/svm/svm.c b/xen/arch/x86/hvm/svm/svm.c
> index 18ba837738..815565c33f 100644
> --- a/xen/arch/x86/hvm/svm/svm.c
> +++ b/xen/arch/x86/hvm/svm/svm.c
> @@ -2524,6 +2524,7 @@ const struct hvm_function_table * __init start_svm(void)
>      P(cpu_has_tsc_ratio, "TSC Rate MSR");
>      P(cpu_has_svm_sss, "NPT Supervisor Shadow Stack");
>      P(cpu_has_svm_spec_ctrl, "MSR_SPEC_CTRL virtualisation");
> +    P(cpu_has_svm_vnmi, "Virtual NMI");
>      P(cpu_has_svm_bus_lock, "Bus Lock Filter");
>  #undef P
>  
> diff --git a/xen/arch/x86/hvm/svm/vmcb.c b/xen/arch/x86/hvm/svm/vmcb.c
> index e583ef8548..e90bbac332 100644
> --- a/xen/arch/x86/hvm/svm/vmcb.c
> +++ b/xen/arch/x86/hvm/svm/vmcb.c
> @@ -184,6 +184,8 @@ static int construct_vmcb(struct vcpu *v)
>      if ( default_xen_spec_ctrl == SPEC_CTRL_STIBP )
>          v->arch.msrs->spec_ctrl.raw = SPEC_CTRL_STIBP;
>  
> +    vmcb->_vintr.fields.vnmi_enable = cpu_has_svm_vnmi;
> +
>      return 0;
>  }
>  

This works for the common case, but breaks emulation of IRET which needs
to be able to clear the NMI latch.

At this juncture, I'm going to submit a version which I think works
adequately.

~Andrew