On Thu, Oct 28, 2021, Maxim Levitsky wrote:
> On Fri, 2021-10-08 at 19:12 -0700, Sean Christopherson wrote:
> > Signal the AVIC doorbell iff the vCPU is running in the guest. If the vCPU
> > is not IN_GUEST_MODE, it's guaranteed to pick up any pending IRQs on the
> > next VMRUN, which unconditionally processes the vIRR.
> >
> > Add comments to document the logic.
> >
> > Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> > ---
> > arch/x86/kvm/svm/avic.c | 14 ++++++++++++--
> > 1 file changed, 12 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
> > index 208c5c71e827..cbf02e7e20d0 100644
> > --- a/arch/x86/kvm/svm/avic.c
> > +++ b/arch/x86/kvm/svm/avic.c
> > @@ -674,7 +674,12 @@ int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
> > kvm_lapic_set_irr(vec, vcpu->arch.apic);
> > smp_mb__after_atomic();
> >
> > - if (avic_vcpu_is_running(vcpu)) {
> > + /*
> > + * Signal the doorbell to tell hardware to inject the IRQ if the vCPU
> > + * is in the guest. If the vCPU is not in the guest, hardware will
> > + * automatically process AVIC interrupts at VMRUN.
> > + */
> > + if (vcpu->mode == IN_GUEST_MODE) {
> > int cpu = READ_ONCE(vcpu->cpu);
> >
> > /*
> > @@ -687,8 +692,13 @@ int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
> > if (cpu != get_cpu())
> > wrmsrl(SVM_AVIC_DOORBELL, kvm_cpu_get_apicid(cpu));
> > put_cpu();
> > - } else
> > + } else {
> > + /*
> > + * Wake the vCPU if it was blocking. KVM will then detect the
> > + * pending IRQ when checking if the vCPU has a wake event.
> > + */
> > kvm_vcpu_wake_up(vcpu);
> > + }
> >
> > return 0;
> > }
>
> It makes sense indeed to avoid ringing the doorbell when the vCPU is not in
> the guest mode.
>
> I do wonder if we want to call kvm_vcpu_wake_up always otherwise, as the vCPU
> might be just outside of the guest mode and not scheduled out. I don't know
> how expensive is kvm_vcpu_wake_up in this case.
IIUC, you're asking if we should do something like:
if (vcpu->mode == IN_GUEST_MODE) {
<signal doorbell>
} else if (!is_vcpu_loaded(vcpu)) {
kvm_vcpu_wake_up();
}
The answer is that kvm_vcpu_wake_up(), which is effectively rcuwait_wake_up(),
is very cheap except for specific configurations that may or may not be valid for
production[*]. Practically speaking, is_vcpu_loaded() doesn't exist and should
never exist because it's inherently racy. The closest we have would be
else if (vcpu != kvm_get_running_vcpu()) {
kvm_vcpu_wake_up();
}
but that's extremely unlikely to be a net win because getting the current vCPU
requires atomics to disable/re-enable preemption, especially if rcuwait_wake_up()
is modified to avoid the rcu lock/unlock.
TL;DR: rcuwait_wake_up() is cheap, and if it's too expensive, a better optimization
would be to make it less expensive.
[*] https://lkml.kernel.org/r/20211020110638.797389-1-pbonzini@xxxxxxxxxx
> Before this patch, the avic_vcpu_is_running would only be false when the vCPU
> is scheduled out (e.g when vcpu_put was done on it)
>
> Best regards,
> Maxim Levitsky
>
