Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
References: <1584007547-4802-1-git-send-email-wanpengli@tencent.com> <87r1xxrhb0.fsf@vitty.brq.redhat.com>
In-Reply-To: <87r1xxrhb0.fsf@vitty.brq.redhat.com>
From:   Wanpeng Li <kernellwp@gmail.com>
Date:   Thu, 12 Mar 2020 19:05:04 +0800
Message-ID: <CANRm+Cwawew=Xygxmzr2jmgPAKqDxvkqxxzjvoxnRRjC_Jx9Xw@mail.gmail.com>
Subject: Re: [PATCH] KVM: VMX: Micro-optimize vmexit time when not exposing PMU
To:     Vitaly Kuznetsov <vkuznets@redhat.com>
Cc:     LKML <linux-kernel@vger.kernel.org>, kvm <kvm@vger.kernel.org>,
        Paolo Bonzini <pbonzini@redhat.com>,
        Sean Christopherson <sean.j.christopherson@intel.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, 12 Mar 2020 at 18:36, Vitaly Kuznetsov <vkuznets@redhat.com> wrote:
>
> Wanpeng Li <kernellwp@gmail.com> writes:
>
> > From: Wanpeng Li <wanpengli@tencent.com>
> >
> > PMU is not exposed to guest by most of cloud providers since the bad performance
> > of PMU emulation and security concern. However, it calls perf_guest_switch_get_msrs()
> > and clear_atomic_switch_msr() unconditionally even if PMU is not exposed to the
> > guest before each vmentry.
> >
> > ~1.28% vmexit time reduced can be observed by kvm-unit-tests/vmexit.flat on my
> > SKX server.
> >
> > Before patch:
> > vmcall 1559
> >
> > After patch:
> > vmcall 1539
> >
> > Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
> > ---
> >  arch/x86/kvm/vmx/vmx.c | 3 +++
> >  1 file changed, 3 insertions(+)
> >
> > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> > index 40b1e61..fd526c8 100644
> > --- a/arch/x86/kvm/vmx/vmx.c
> > +++ b/arch/x86/kvm/vmx/vmx.c
> > @@ -6441,6 +6441,9 @@ static void atomic_switch_perf_msrs(struct vcpu_vmx *vmx)
> >       int i, nr_msrs;
> >       struct perf_guest_switch_msr *msrs;
> >
> > +     if (!vcpu_to_pmu(&vmx->vcpu)->version)
> > +             return;
> > +
> >       msrs = perf_guest_get_msrs(&nr_msrs);
> >
> >       if (!msrs)
>
> Personally, I'd prefer this to be expressed as
>
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index 40b1e6138cd5..ace92076c90f 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -6567,7 +6567,9 @@ static void vmx_vcpu_run(struct kvm_vcpu *vcpu)
>
>         pt_guest_enter(vmx);
>
> -       atomic_switch_perf_msrs(vmx);
> +       if (vcpu_to_pmu(&vmx->vcpu)->version)
> +               atomic_switch_perf_msrs(vmx);
> +

I just hope the beautiful codes before, I testing this version before
sending out the patch, ~30 cycles can be saved which means that ~2%
vmexit time, will update in next version. Let's wait Paolo for other
opinions below.

    Wanpeng

>
> Also, (not knowing much about PMU), is
> "vcpu_to_pmu(&vmx->vcpu)->version" check correct?
>
> E.g. in intel_is_valid_msr() correct for Intel PMU or is it stated
> somewhere that it is generic rule?
>
> Also, speaking about cloud providers and the 'micro' nature of this
> optimization, would it rather make sense to introduce a static branch
> (the policy to disable vPMU is likely to be host wide, right)?
>
> --
> Vitaly
>