From: ZhuangYanying <ann.zhuangyanying@huawei.com>

When spin_lock_irqsave() deadlock occurs inside the guest, vcpu threads,
other than the lock-holding one, would enter into S state because of
pvspinlock. Then inject NMI via libvirt API "inject-nmi", the NMI could
not be injected into vm.

The reason is:
1 It sets nmi_queued to 1 when calling ioctl KVM_NMI in qemu, and sets
cpu->kvm_vcpu_dirty to true in do_inject_external_nmi() meanwhile.
2 It sets nmi_queued to 0 in process_nmi(), before entering guest, because
cpu->kvm_vcpu_dirty is true.

It's not enough just to check nmi_queued to decide whether to stay in
vcpu_block() or not. NMI should be injected immediately at any situation.
Add checking nmi_pending, and testing KVM_REQ_NMI replaces nmi_queued
in vm_vcpu_has_events().

Do the same change for SMIs.

Signed-off-by: Zhuang Yanying <ann.zhuangyanying@huawei.com>
---
v1->v2
  - simplify message. The complete description is here:
    http://www.spinics.net/lists/kvm/msg150380.html
  - Testing KVM_REQ_NMI replaces nmi_pending.
  - Add Testing kvm_x86_ops->nmi_allowed(vcpu).
v2->v3
  - Testing KVM_REQ_NMI replaces nmi_queued, not nmi_pending.
  - Do the same change for SMIs. 
---
 arch/x86/kvm/x86.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 02363e3..a2cd099 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -8394,10 +8394,13 @@ static inline bool kvm_vcpu_has_events(struct kvm_vcpu *vcpu)
 	if (vcpu->arch.pv.pv_unhalted)
 		return true;
 
-	if (atomic_read(&vcpu->arch.nmi_queued))
+	if (kvm_test_request(KVM_REQ_NMI, vcpu) ||
+	    (vcpu->arch.nmi_pending &&
+	     kvm_x86_ops->nmi_allowed(vcpu)))
 		return true;
 
-	if (kvm_test_request(KVM_REQ_SMI, vcpu))
+	if (kvm_test_request(KVM_REQ_SMI, vcpu) ||
+	    (vcpu->arch.smi_pending && !is_smm(vcpu)))
 		return true;
 
 	if (kvm_arch_interrupt_allowed(vcpu) &&
-- 
1.8.3.1

2017-05-26 13:16+0800, Zhuangyanying:
> From: ZhuangYanying <ann.zhuangyanying@huawei.com>
> 
> When spin_lock_irqsave() deadlock occurs inside the guest, vcpu threads,
> other than the lock-holding one, would enter into S state because of
> pvspinlock. Then inject NMI via libvirt API "inject-nmi", the NMI could
> not be injected into vm.
> 
> The reason is:
> 1 It sets nmi_queued to 1 when calling ioctl KVM_NMI in qemu, and sets
> cpu->kvm_vcpu_dirty to true in do_inject_external_nmi() meanwhile.
> 2 It sets nmi_queued to 0 in process_nmi(), before entering guest, because
> cpu->kvm_vcpu_dirty is true.
> 
> It's not enough just to check nmi_queued to decide whether to stay in
> vcpu_block() or not. NMI should be injected immediately at any situation.
> Add checking nmi_pending, and testing KVM_REQ_NMI replaces nmi_queued
> in vm_vcpu_has_events().
> 
> Do the same change for SMIs.
> 
> Signed-off-by: Zhuang Yanying <ann.zhuangyanying@huawei.com>
> ---
> v1->v2
>   - simplify message. The complete description is here:
>     http://www.spinics.net/lists/kvm/msg150380.html
>   - Testing KVM_REQ_NMI replaces nmi_pending.
>   - Add Testing kvm_x86_ops->nmi_allowed(vcpu).
> v2->v3
>   - Testing KVM_REQ_NMI replaces nmi_queued, not nmi_pending.
>   - Do the same change for SMIs. 
> ---
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> @@ -8394,10 +8394,13 @@ static inline bool kvm_vcpu_has_events(struct kvm_vcpu *vcpu)
>  	if (vcpu->arch.pv.pv_unhalted)
>  		return true;
>  
> -	if (atomic_read(&vcpu->arch.nmi_queued))
> +	if (kvm_test_request(KVM_REQ_NMI, vcpu) ||
> +	    (vcpu->arch.nmi_pending &&

I think the logic should be

  	if ((kvm_test_request(KVM_REQ_NMI, vcpu) || vcpu->arch.nmi_pending) &&
  			kvm_x86_ops->nmi_allowed(vcpu))

because there is no reason to resume the VCPU if we cannot inject.

> +	     kvm_x86_ops->nmi_allowed(vcpu)))
>  		return true;
>  
> -	if (kvm_test_request(KVM_REQ_SMI, vcpu))
> +	if (kvm_test_request(KVM_REQ_SMI, vcpu) ||
> +	    (vcpu->arch.smi_pending && !is_smm(vcpu)))

Ditto.

>  		return true;
>  
>  	if (kvm_arch_interrupt_allowed(vcpu) &&

We'll then be consistent with other interrupts,

Thanks.

On 30/05/2017 15:36, Radim Krčmář wrote:
>> -	if (atomic_read(&vcpu->arch.nmi_queued))
>> +	if (kvm_test_request(KVM_REQ_NMI, vcpu) ||
>> +	    (vcpu->arch.nmi_pending &&
> I think the logic should be
> 
>   	if ((kvm_test_request(KVM_REQ_NMI, vcpu) || vcpu->arch.nmi_pending) &&
>   			kvm_x86_ops->nmi_allowed(vcpu))
> 
> because there is no reason to resume the VCPU if we cannot inject.

KVM_REQ_NMI would be processed anyway, and would clear nmi_queued.  Of
course, it would very soon go back to sleep.  Even before Yanying's
patch, nmi_queued > 0 would have woken up the vCPU in this manner.  So
I'm applying the patch.  Thanks!

Paolo