Re: [PATCH v2 3/6] KVM: x86: Fold kvm_arch_sched_in() into kvm_arch_vcpu_load()

[Sean Christopherson]

On Fri, May 24, 2024, Kai Huang wrote:
> > @@ -1548,6 +1548,9 @@ static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
> >   	struct vcpu_svm *svm = to_svm(vcpu);
> >   	struct svm_cpu_data *sd = per_cpu_ptr(&svm_data, cpu);
> > +	if (vcpu->scheduled_out && !kvm_pause_in_guest(vcpu->kvm))
> > +		shrink_ple_window(vcpu);
> > +
> 
> [...]
> 
> > @@ -1517,6 +1517,9 @@ void vmx_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
> >   {
> >   	struct vcpu_vmx *vmx = to_vmx(vcpu);
> > +	if (vcpu->scheduled_out && !kvm_pause_in_guest(vcpu->kvm))
> > +		shrink_ple_window(vcpu);
> > +
> 
> Nit:  Perhaps we need a kvm_x86_ops::shrink_ple_window()?  :-)

Heh, that duplicate code annoys me too.  The problem is the "old" window value
comes from the VMCS/VMCB, so either we'd end up with multiple kvm_x86_ops, or
we'd only be able to consolidate the scheduled_out + kvm_pause_in_guest() code,
which isn't all that interesting.

Aha!  Actually, VMX already open codes the functionality provided by VCPU_EXREG_*,
e.g. has vmx->ple_window_dirty.  If we add VCPU_EXREG_PLE_WINDOW, then the info
get be made available to common x86 code without having to add new hooks.  And
that would also allow moving the guts of handle_pause()/pause_interception() to
common code, i.e. will also allow deduplicating the "grow" side of things.