Re: [PATCH v2 5/5] KVM: SVM: allow AVIC to co-exist with a nested guest running

[Maxim Levitsky <mlevitsk@redhat.com>]

On Wed, 2022-01-05 at 21:56 +0000, Sean Christopherson wrote:
> On Mon, Dec 13, 2021, Maxim Levitsky wrote:
> > @@ -1486,6 +1485,12 @@ struct kvm_x86_ops {
> >  	int (*complete_emulated_msr)(struct kvm_vcpu *vcpu, int err);
> >  
> >  	void (*vcpu_deliver_sipi_vector)(struct kvm_vcpu *vcpu, u8 vector);
> > +
> > +	/*
> > +	 * Returns false if for some reason APICv (e.g guest mode)
> > +	 * must be inhibited on this vCPU
> 
> Comment is wrong, code returns 'true' if AVIC is inhibited due to is_guest_mode().
> Even better, rename the hook to something that's more self-documenting.
> 
> vcpu_is_apicv_inhibited() jumps to mind, but that's a bad name since it's not
> called by kvm_vcpu_apicv_active().  Maybe vcpu_has_apicv_inhibit_condition()?

Yep. I also kind of don't like the name, but I didn't though of anything better.
vcpu_has_apicv_inhibit_condition seems a good idea.

> 
> > +	 */
> > +	bool (*apicv_check_inhibit)(struct kvm_vcpu *vcpu);
> >  };
> >  
> >  struct kvm_x86_nested_ops {
> > diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
> > index 34f62da2fbadd..5a8304938f51e 100644
> > --- a/arch/x86/kvm/svm/avic.c
> > +++ b/arch/x86/kvm/svm/avic.c
> > @@ -734,6 +734,11 @@ int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
> >  	return 0;
> >  }
> >  
> > +bool avic_is_vcpu_inhibited(struct kvm_vcpu *vcpu)
> 
> This should follow whatever wording we decide on for the kvm_x86_ops hook.  In
> isolation, this name is too close to kvm_vcpu_apicv_active() and could be wrongly
> assumed to mean that APICv is not inhibited for _any_ reason on this vCPU if it
> returns false.
I will think of a better name.


> 
> > +{
> > +	return is_guest_mode(vcpu);
> > +}
> > +
> >  bool svm_dy_apicv_has_pending_interrupt(struct kvm_vcpu *vcpu)
> >  {
> >  	return false;
> 
> ...
> 
> > @@ -4486,6 +4493,7 @@ static struct kvm_x86_ops svm_x86_ops __initdata = {
> >  	.complete_emulated_msr = svm_complete_emulated_msr,
> >  
> >  	.vcpu_deliver_sipi_vector = svm_vcpu_deliver_sipi_vector,
> > +	.apicv_check_inhibit = avic_is_vcpu_inhibited,
> 
> This can technically be NULL if nested=0.
Good idea, now it is possible to after recent refactoring.

> 
> >  };
> >  
> >  /*
> > diff --git a/arch/x86/kvm/svm/svm.h b/arch/x86/kvm/svm/svm.h
> > index daa8ca84afccd..545684ea37353 100644
> > --- a/arch/x86/kvm/svm/svm.h
> > +++ b/arch/x86/kvm/svm/svm.h
> > @@ -590,6 +590,7 @@ void svm_load_eoi_exitmap(struct kvm_vcpu *vcpu, u64 *eoi_exit_bitmap);
> >  void svm_hwapic_irr_update(struct kvm_vcpu *vcpu, int max_irr);
> >  void svm_hwapic_isr_update(struct kvm_vcpu *vcpu, int max_isr);
> >  int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec);
> > +bool avic_is_vcpu_inhibited(struct kvm_vcpu *vcpu);
> >  bool svm_dy_apicv_has_pending_interrupt(struct kvm_vcpu *vcpu);
> >  int svm_update_pi_irte(struct kvm *kvm, unsigned int host_irq,
> >  		       uint32_t guest_irq, bool set);
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 81a74d86ee5eb..125599855af47 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -9161,6 +9161,10 @@ static int inject_pending_event(struct kvm_vcpu *vcpu, bool *req_immediate_exit)
> >  		r = kvm_check_nested_events(vcpu);
> >  		if (r < 0)
> >  			goto out;
> > +
> > +		/* Nested VM exit might need to update APICv status */
> > +		if (kvm_check_request(KVM_REQ_APICV_UPDATE, vcpu))
> > +			kvm_vcpu_update_apicv(vcpu);
> >  	}
> >  
> >  	/* try to inject new event if pending */
> > @@ -9538,6 +9542,10 @@ void kvm_vcpu_update_apicv(struct kvm_vcpu *vcpu)
> >  	down_read(&vcpu->kvm->arch.apicv_update_lock);
> >  
> >  	activate = kvm_apicv_activated(vcpu->kvm);
> > +
> > +	if (kvm_x86_ops.apicv_check_inhibit)
> > +		activate = activate && !kvm_x86_ops.apicv_check_inhibit(vcpu);
> 
> Might as well use Use static_call().  This would also be a candidate for
> DEFINE_STATIC_CALL_RET0, though that's overkill if this is the only call site.
This is also something that should be done, but I prefer to do this in one go.
There are several nested related functions that were not converted to static_call
(like .check_nested_events).

Also I recently found that we have KVM_X86_OP and KVM_X86_OP_NULL which are the
same thing - another thing for refactoring, so I prefer to refactor this
in one patch series.



> 
> > +
> >  	if (vcpu->arch.apicv_active == activate)
> >  		goto out;
> >  
> > @@ -9935,7 +9943,10 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
> >  		 * per-VM state, and responsing vCPUs must wait for the update
> >  		 * to complete before servicing KVM_REQ_APICV_UPDATE.
> >  		 */
> > -		WARN_ON_ONCE(kvm_apicv_activated(vcpu->kvm) != kvm_vcpu_apicv_active(vcpu));
> > +		if (!is_guest_mode(vcpu))
> > +			WARN_ON_ONCE(kvm_apicv_activated(vcpu->kvm) != kvm_vcpu_apicv_active(vcpu));
> > +		else
> > +			WARN_ON(kvm_vcpu_apicv_active(vcpu));
> 
> Won't this fire on VMX?

Yes it will! Good catch. It almost like I would like to have .apicv_is_avic boolean,
for such cases :-) I'll think of something.

Best regards,
	Maxim Levitsky

> 
> >  
> >  		exit_fastpath = static_call(kvm_x86_run)(vcpu);
> >  		if (likely(exit_fastpath != EXIT_FASTPATH_REENTER_GUEST))
> > -- 
> > 2.26.3
> >