On Thu, 30 Jul 2020 at 19:16, Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> wrote:
>
> Wanpeng Li <kernellwp@xxxxxxxxx> writes:
>
> > On Wed, 29 Jul 2020 at 20:21, Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> wrote:
> >>
> >> Wanpeng Li <kernellwp@xxxxxxxxx> writes:
> >>
> >> > From: Wanpeng Li <wanpengli@xxxxxxxxxxx>
> >> >
> >> > Commit 8566ac8b (KVM: SVM: Implement pause loop exit logic in SVM) drops
> >> > disable pause loop exit/pause filtering capability completely, I guess it
> >> > is a merge fault by Radim since disable vmexits capabilities and pause
> >> > loop exit for SVM patchsets are merged at the same time. This patch
> >> > reintroduces the disable pause loop exit/pause filtering capability
> >> > support.
> >> >
> >> > We can observe 2.9% hackbench improvement for a 92 vCPUs guest on AMD
> >> > Rome Server.
> >> >
> >> > Reported-by: Haiwei Li <lihaiwei@xxxxxxxxxxx>
> >> > Tested-by: Haiwei Li <lihaiwei@xxxxxxxxxxx>
> >> > Fixes: 8566ac8b (KVM: SVM: Implement pause loop exit logic in SVM)
> >> > Signed-off-by: Wanpeng Li <wanpengli@xxxxxxxxxxx>
> >> > ---
> >> > arch/x86/kvm/svm/svm.c | 9 ++++++---
> >> > 1 file changed, 6 insertions(+), 3 deletions(-)
> >> >
> >> > diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> >> > index c0da4dd..c20f127 100644
> >> > --- a/arch/x86/kvm/svm/svm.c
> >> > +++ b/arch/x86/kvm/svm/svm.c
> >> > @@ -1090,7 +1090,7 @@ static void init_vmcb(struct vcpu_svm *svm)
> >> > svm->nested.vmcb = 0;
> >> > svm->vcpu.arch.hflags = 0;
> >> >
> >> > - if (pause_filter_count) {
> >> > + if (pause_filter_count && !kvm_pause_in_guest(svm->vcpu.kvm)) {
> >> > control->pause_filter_count = pause_filter_count;
> >> > if (pause_filter_thresh)
> >> > control->pause_filter_thresh = pause_filter_thresh;
> >> > @@ -2693,7 +2693,7 @@ static int pause_interception(struct vcpu_svm *svm)
> >> > struct kvm_vcpu *vcpu = &svm->vcpu;
> >> > bool in_kernel = (svm_get_cpl(vcpu) == 0);
> >> >
> >> > - if (pause_filter_thresh)
> >> > + if (!kvm_pause_in_guest(vcpu->kvm))
> >> > grow_ple_window(vcpu);
> >> >
> >> > kvm_vcpu_on_spin(vcpu, in_kernel);
> >> > @@ -3780,7 +3780,7 @@ static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
> >> >
> >> > static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
> >> > {
> >> > - if (pause_filter_thresh)
> >> > + if (!kvm_pause_in_guest(vcpu->kvm))
> >> > shrink_ple_window(vcpu);
> >> > }
> >> >
> >> > @@ -3958,6 +3958,9 @@ static void svm_vm_destroy(struct kvm *kvm)
> >> >
> >> > static int svm_vm_init(struct kvm *kvm)
> >> > {
> >> > + if (!pause_filter_thresh)
> >> > + kvm->arch.pause_in_guest = true;
> >>
> >> Would it make sense to do
> >>
> >> if (!pause_filter_count || !pause_filter_thresh)
> >> kvm->arch.pause_in_guest = true;
> >>
> >> here and simplify the condition in init_vmcb()?
> >
> > kvm->arch.pause_in_guest can also be true when userspace sets the
> > KVM_CAP_X86_DISABLE_EXITS capability, so we can't simplify the
> > condition in init_vmcb().
> >
>
> I meant we simplify it to
>
> if (!kvm_pause_in_guest(svm->vcpu.kvm))
>
> as "!pause_filter_count" gets included.
Just do it in v3.
Wanpeng
