KVM: x86: Inject pending interrupt even if pending nmi exist
authorYuki Shibuya <shibuya.yk@ncos.nec.co.jp>
Thu, 24 Mar 2016 05:17:03 +0000 (05:17 +0000)
committerBen Hutchings <ben@decadent.org.uk>
Sat, 30 Apr 2016 22:05:22 +0000 (00:05 +0200)
commit 321c5658c5e9192dea0d58ab67cf1791e45b2b26 upstream.

Non maskable interrupts (NMI) are preferred to interrupts in current
implementation. If a NMI is pending and NMI is blocked by the result
of nmi_allowed(), pending interrupt is not injected and
enable_irq_window() is not executed, even if interrupts injection is

In old kernel (e.g. 2.6.32), schedule() is often called in NMI context.
In this case, interrupts are needed to execute iret that intends end
of NMI. The flag of blocking new NMI is not cleared until the guest
execute the iret, and interrupts are blocked by pending NMI. Due to
this, iret can't be invoked in the guest, and the guest is starved
until block is cleared by some events (e.g. canceling injection).

This patch injects pending interrupts, when it's allowed, even if NMI
is blocked. And, If an interrupts is pending after executing
inject_pending_event(), enable_irq_window() is executed regardless of
NMI pending counter.

Signed-off-by: Yuki Shibuya <shibuya.yk@ncos.nec.co.jp>
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
[bwh: Backported to 3.2:
 - vcpu_enter_guest() is simpler because inject_pending_event() can't fail
 - Adjust context]
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>

index 09dab5b..da1a126 100644 (file)
@@ -5634,12 +5634,10 @@ static void inject_pending_event(struct kvm_vcpu *vcpu)
        /* try to inject new event if pending */
-       if (vcpu->arch.nmi_pending) {
-               if (kvm_x86_ops->nmi_allowed(vcpu)) {
-                       --vcpu->arch.nmi_pending;
-                       vcpu->arch.nmi_injected = true;
-                       kvm_x86_ops->set_nmi(vcpu);
-               }
+       if (vcpu->arch.nmi_pending && kvm_x86_ops->nmi_allowed(vcpu)) {
+               --vcpu->arch.nmi_pending;
+               vcpu->arch.nmi_injected = true;
+               kvm_x86_ops->set_nmi(vcpu);
        } else if (kvm_cpu_has_interrupt(vcpu)) {
                if (kvm_x86_ops->interrupt_allowed(vcpu)) {
                        kvm_queue_interrupt(vcpu, kvm_cpu_get_interrupt(vcpu),
@@ -5742,7 +5740,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
                /* enable NMI/IRQ window open exits if needed */
                if (vcpu->arch.nmi_pending)
-               else if (kvm_cpu_has_interrupt(vcpu) || req_int_win)
+               if (kvm_cpu_has_interrupt(vcpu) || req_int_win)
                if (kvm_lapic_enabled(vcpu)) {