[PATCH 0/1] Fix for a recent regression in kvm/queue (guest using 100% cpu time)

[PATCH 0/1] Fix for a recent regression in kvm/queue (guest using 100% cpu time)

[Maxim Levitsky]

I did a quick bisect yesterday after noticing that my VMs started
to take 100% cpu time.

Looks like we don't ignore SIPIs that are received while the CPU isn't
waiting for them, and that makes KVM think that CPU always has
pending events which makes it never enter an idle state.

Best regards,
	Maxim Levitsky

Maxim Levitsky (1):
  KVM: x86: ignore SIPIs that are received while not in wait-for-sipi
    state

 arch/x86/kvm/lapic.c | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

-- 
2.26.2

[PATCH 1/1] KVM: x86: ignore SIPIs that are received while not in wait-for-sipi state

[Maxim Levitsky]

In the commit 1c96dcceaeb3
("KVM: x86: fix apic_accept_events vs check_nested_events"),

we accidently started latching SIPIs that are received while the cpu is not
waiting for them.

This causes vCPUs to never enter a halted state.

Fixes: 1c96dcceaeb3 ("KVM: x86: fix apic_accept_events vs check_nested_events")
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
---
 arch/x86/kvm/lapic.c | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
index e3ee597ff5404..6a87623aa578e 100644
--- a/arch/x86/kvm/lapic.c
+++ b/arch/x86/kvm/lapic.c
@@ -2892,14 +2892,15 @@ void kvm_apic_accept_events(struct kvm_vcpu *vcpu)
 		else
 			vcpu->arch.mp_state = KVM_MP_STATE_INIT_RECEIVED;
 	}
-	if (test_bit(KVM_APIC_SIPI, &pe) &&
-	    vcpu->arch.mp_state == KVM_MP_STATE_INIT_RECEIVED) {
+	if (test_bit(KVM_APIC_SIPI, &pe)) {
 		clear_bit(KVM_APIC_SIPI, &apic->pending_events);
-		/* evaluate pending_events before reading the vector */
-		smp_rmb();
-		sipi_vector = apic->sipi_vector;
-		kvm_vcpu_deliver_sipi_vector(vcpu, sipi_vector);
-		vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
+		if (vcpu->arch.mp_state == KVM_MP_STATE_INIT_RECEIVED) {
+			/* evaluate pending_events before reading the vector */
+			smp_rmb();
+			sipi_vector = apic->sipi_vector;
+			kvm_vcpu_deliver_sipi_vector(vcpu, sipi_vector);
+			vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
+		}
 	}
 }
 
-- 
2.26.2

Re: [PATCH 1/1] KVM: x86: ignore SIPIs that are received while not in wait-for-sipi state

[Paolo Bonzini]

On 03/12/20 15:33, Maxim Levitsky wrote:
> In the commit 1c96dcceaeb3
> ("KVM: x86: fix apic_accept_events vs check_nested_events"),
> 
> we accidently started latching SIPIs that are received while the cpu is not
> waiting for them.
> 
> This causes vCPUs to never enter a halted state.
> 
> Fixes: 1c96dcceaeb3 ("KVM: x86: fix apic_accept_events vs check_nested_events")
> Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
> ---
>  arch/x86/kvm/lapic.c | 15 ++++++++-------


Indeed, it was done by the xchg before that commit, even if 
vcpu->arch.mp_state != KVM_MP_STATE_INIT_RECEIVED.

Paolo