Re: [PATCH 3/6] KVM: SVM: fix AVIC race of host->guest IPI delivery vs AVIC inhibition

[Sean Christopherson <seanjc@google.com>]

On Thu, Dec 09, 2021, Maxim Levitsky wrote:
> On Thu, 2021-12-09 at 15:11 +0100, Paolo Bonzini wrote:
> > On 12/9/21 12:54, Maxim Levitsky wrote:
> > > If svm_deliver_avic_intr is called just after the target vcpu's AVIC got
> > > inhibited, it might read a stale value of vcpu->arch.apicv_active
> > > which can lead to the target vCPU not noticing the interrupt.
> > > 
> > > Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
> > > ---
> > >   arch/x86/kvm/svm/avic.c | 16 +++++++++++++---
> > >   1 file changed, 13 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
> > > index 859ad2dc50f1..8c1b934bfa9b 100644
> > > --- a/arch/x86/kvm/svm/avic.c
> > > +++ b/arch/x86/kvm/svm/avic.c
> > > @@ -691,6 +691,15 @@ int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
> > >   	 * automatically process AVIC interrupts at VMRUN.
> > >   	 */
> > >   	if (vcpu->mode == IN_GUEST_MODE) {
> > > +
> > > +		/*
> > > +		 * At this point we had read the vcpu->arch.apicv_active == true
> > > +		 * and the vcpu->mode == IN_GUEST_MODE.
> > > +		 * Since we have a memory barrier after setting IN_GUEST_MODE,
> > > +		 * it ensures that AVIC inhibition is complete and thus
> > > +		 * the target is really running with AVIC enabled.
> > > +		 */
> > > +
> > >   		int cpu = READ_ONCE(vcpu->cpu);
> > 
> > I don't think it's correct.  The vCPU has apicv_active written (in 
> > kvm_vcpu_update_apicv) before vcpu->mode.
> 
> I thought that we have a full memory barrier just prior to setting IN_GUEST_MODE
> thus if I see vcpu->mode == IN_GUEST_MODE then I'll see correct apicv_active value.
> But apparently the memory barrier is after setting vcpu->mode.
> 
> 
> > 
> > For the acquire/release pair to work properly you need to 1) read 
> > apicv_active *after* vcpu->mode here 2) use store_release and 
> > load_acquire for vcpu->mode, respectively in vcpu_enter_guest and here.
> 
> store_release for vcpu->mode in vcpu_enter_guest means a write barrier just before setting it,
> which I expected to be there.
> 
> And yes I see now, I need a read barrier here as well. I am still learning this.

Sans barriers and comments, can't this be written as returning an "error" if the
vCPU is not IN_GUEST_MODE?  Effectively the same thing, but a little more precise
and it avoids duplicating the lapic.c code.

diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
index 26ed5325c593..cddf7a8da3ea 100644
--- a/arch/x86/kvm/svm/avic.c
+++ b/arch/x86/kvm/svm/avic.c
@@ -671,7 +671,7 @@ void svm_load_eoi_exitmap(struct kvm_vcpu *vcpu, u64 *eoi_exit_bitmap)

 int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
 {
-       if (!vcpu->arch.apicv_active)
+       if (vcpu->mode != IN_GUEST_MODE || !vcpu->arch.apicv_active)
                return -1;

        kvm_lapic_set_irr(vec, vcpu->arch.apic);
@@ -706,8 +706,9 @@ int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
                put_cpu();
        } else {
                /*
-                * Wake the vCPU if it was blocking.  KVM will then detect the
-                * pending IRQ when checking if the vCPU has a wake event.
+                * Wake the vCPU if it is blocking.  If the vCPU exited the
+                * guest since the previous vcpu->mode check, it's guaranteed
+                * to see the event before re-enterring the guest.
                 */
                kvm_vcpu_wake_up(vcpu);
        }