On 9/23/25 00:03, Neeraj Upadhyay wrote:
> The Secure AVIC hardware introduces a new model for handling Non-Maskable
> Interrupts (NMIs). This model differs significantly from standard SVM, as
> guest NMI state is managed by the hardware and is not visible to KVM.
>
> Consequently, KVM can no longer use the generic EVENT_INJ mechanism and
> must not track NMI masking state in software. Instead, it must adopt the
> vNMI (Virtual NMI) flow, which is the only mechanism supported by
> Secure AVIC.
>
> Enable NMI support by making three key changes:
>
> 1. Enable NMI in VMSA: Set the V_NMI_ENABLE_MASK bit in the VMSA's
> vintr_ctr field. This is a hardware prerequisite to enable the
> vNMI feature for the guest.
>
> 2. Use vNMI for Injection: Modify svm_inject_nmi() to use the vNMI
> flow for Secure AVIC guests. When an NMI is requested, set the
> V_NMI_PENDING_MASK in the VMCB instead of using EVENT_INJ.
>
> 3. Update NMI Windowing: Modify svm_nmi_allowed() to reflect that
> hardware now manages NMI blocking. KVM's only responsibility is to
> avoid queuing a new vNMI if one is already pending. The check is
> now simplified to whether V_NMI_PENDING_MASK is already set.
>
> Co-developed-by: Kishon Vijay Abraham I <kvijayab@amd.com>
> Signed-off-by: Kishon Vijay Abraham I <kvijayab@amd.com>
> Signed-off-by: Neeraj Upadhyay <Neeraj.Upadhyay@amd.com>
> ---
> arch/x86/kvm/svm/sev.c | 2 +-
> arch/x86/kvm/svm/svm.c | 56 ++++++++++++++++++++++++++----------------
> 2 files changed, 36 insertions(+), 22 deletions(-)
>
> diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c
> index 2dee210efb37..7c66aefe428a 100644
> --- a/arch/x86/kvm/svm/sev.c
> +++ b/arch/x86/kvm/svm/sev.c
> @@ -885,7 +885,7 @@ static int sev_es_sync_vmsa(struct vcpu_svm *svm)
> save->sev_features = sev->vmsa_features;
>
> if (sev_savic_active(vcpu->kvm))
> - save->vintr_ctrl |= V_GIF_MASK;
> + save->vintr_ctrl |= V_GIF_MASK | V_NMI_ENABLE_MASK;
>
> /*
> * Skip FPU and AVX setup with KVM_SEV_ES_INIT to avoid
> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> index fdd612c975ae..a945bc094c1a 100644
> --- a/arch/x86/kvm/svm/svm.c
> +++ b/arch/x86/kvm/svm/svm.c
> @@ -3635,27 +3635,6 @@ static int pre_svm_run(struct kvm_vcpu *vcpu)
> return 0;
> }
>
> -static void svm_inject_nmi(struct kvm_vcpu *vcpu)
> -{
> - struct vcpu_svm *svm = to_svm(vcpu);
> -
> - svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
> -
> - if (svm->nmi_l1_to_l2)
> - return;
> -
> - /*
> - * No need to manually track NMI masking when vNMI is enabled, hardware
> - * automatically sets V_NMI_BLOCKING_MASK as appropriate, including the
> - * case where software directly injects an NMI.
> - */
> - if (!is_vnmi_enabled(svm)) {
> - svm->nmi_masked = true;
> - svm_set_iret_intercept(svm);
> - }
> - ++vcpu->stat.nmi_injections;
> -}
A pre-patch that moves this function would make the changes you make to
it in this patch more obvious.
Thanks,
Tom
> -
> static bool svm_is_vnmi_pending(struct kvm_vcpu *vcpu)
> {
> struct vcpu_svm *svm = to_svm(vcpu);
> @@ -3689,6 +3668,33 @@ static bool svm_set_vnmi_pending(struct kvm_vcpu *vcpu)
> return true;
> }
>
> +static void svm_inject_nmi(struct kvm_vcpu *vcpu)
> +{
> + struct vcpu_svm *svm = to_svm(vcpu);
> +
> + if (sev_savic_active(vcpu->kvm)) {
> + svm_set_vnmi_pending(vcpu);
> + ++vcpu->stat.nmi_injections;
> + return;
> + }
> +
> + svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
> +
> + if (svm->nmi_l1_to_l2)
> + return;
> +
> + /*
> + * No need to manually track NMI masking when vNMI is enabled, hardware
> + * automatically sets V_NMI_BLOCKING_MASK as appropriate, including the
> + * case where software directly injects an NMI.
> + */
> + if (!is_vnmi_enabled(svm)) {
> + svm->nmi_masked = true;
> + svm_set_iret_intercept(svm);
> + }
> + ++vcpu->stat.nmi_injections;
> +}
> +
> static void svm_inject_irq(struct kvm_vcpu *vcpu, bool reinjected)
> {
> struct vcpu_svm *svm = to_svm(vcpu);
> @@ -3836,6 +3842,14 @@ bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
> static int svm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
> {
> struct vcpu_svm *svm = to_svm(vcpu);
> +
> + /* Secure AVIC only support V_NMI based NMI injection. */
> + if (sev_savic_active(vcpu->kvm)) {
> + if (svm->vmcb->control.int_ctl & V_NMI_PENDING_MASK)
> + return 0;
> + return 1;
> + }
> +
> if (svm->nested.nested_run_pending)
> return -EBUSY;
>