Message-ID: <CALMp9eQOBAyCY9pC54rG4K2WvLnj1kw4oeNrNOPVDr1ChpbOXQ@mail.gmail.com>
Date:   Fri, 3 Nov 2023 07:50:59 -0700
From:   Jim Mattson <jmattson@...gle.com>
To:     "Mi, Dapeng" <dapeng1.mi@...ux.intel.com>
Cc:     "Liang, Kan" <kan.liang@...ux.intel.com>,
        Sean Christopherson <seanjc@...gle.com>,
        Paolo Bonzini <pbonzini@...hat.com>, kvm@...r.kernel.org,
        linux-kernel@...r.kernel.org,
        Zhenyu Wang <zhenyuw@...ux.intel.com>,
        Zhang Xiong <xiong.y.zhang@...el.com>,
        Mingwei Zhang <mizhang@...gle.com>,
        Like Xu <like.xu.linux@...il.com>,
        Dapeng Mi <dapeng1.mi@...el.com>, Like Xu <likexu@...cent.com>
Subject: Re: [Patch 1/2] KVM: x86/pmu: Add Intel CPUID-hinted TopDown slots event

On Fri, Nov 3, 2023 at 3:03 AM Mi, Dapeng <dapeng1.mi@...ux.intel.com> wrote:
>
>
> On 11/3/2023 1:45 AM, Jim Mattson wrote:
> > On Wed, Nov 1, 2023 at 7:07 PM Mi, Dapeng <dapeng1.mi@...ux.intel.com> wrote:
> >>
> >> On 11/1/2023 9:33 PM, Liang, Kan wrote:
> >>> On 2023-10-31 11:31 p.m., Mi, Dapeng wrote:
> >>>> On 11/1/2023 11:04 AM, Jim Mattson wrote:
> >>>>> On Tue, Oct 31, 2023 at 6:59 PM Mi, Dapeng
> >>>>> <dapeng1.mi@...ux.intel.com> wrote:
> >>>>>> On 11/1/2023 2:22 AM, Jim Mattson wrote:
> >>>>>>> On Tue, Oct 31, 2023 at 1:58 AM Dapeng Mi
> >>>>>>> <dapeng1.mi@...ux.intel.com> wrote:
> >>>>>>>> This patch adds support for the architectural topdown slots event
> >>>>>>>> which
> >>>>>>>> is hinted by CPUID.0AH.EBX.
> >>>>>>> Can't a guest already program an event selector to count event select
> >>>>>>> 0xa4, unit mask 1, unless the event is prohibited by
> >>>>>>> KVM_SET_PMU_EVENT_FILTER?
> >>>>>> Actually defining this new slots arch event is to do the sanity check
> >>>>>> for supported arch-events which is enumerated by CPUID.0AH.EBX.
> >>>>>> Currently vPMU would check if the arch event from guest is supported by
> >>>>>> KVM. If not, it would be rejected just like intel_hw_event_available()
> >>>>>> shows.
> >>>>>>
> >>>>>> If we don't add the slots event in the intel_arch_events[] array, guest
> >>>>>> may program the slots event and pass the sanity check of KVM on a
> >>>>>> platform which actually doesn't support slots event and program the
> >>>>>> event on a real GP counter and got an invalid count. This is not
> >>>>>> correct.
> >>>>> On physical hardware, it is possible to program a GP counter with the
> >>>>> event selector and unit mask of the slots event whether or not the
> >>>>> platform supports it. Isn't KVM wrong to disallow something that a
> >>>>> physical CPU allows?
> >>>> Yeah, I agree. But I'm not sure if this is a flaw on PMU driver. If an
> >>>> event is not supported by the hardware,  we can't predict the PMU's
> >>>> behavior and a meaningless count may be returned and this could mislead
> >>>> the user.
> >>> The user can program any events on the GP counter. The perf doesn't
> >>> limit it. For the unsupported event, 0 should be returned. Please keep
> >>> in mind, the event list keeps updating. If the kernel checks for each
> >>> event, it could be a disaster. I don't think it's a flaw.
> >>
> >> Thanks Kan, it would be ok as long as 0 is always returned for
> >> unsupported events. IMO, it's a nice to have feature that KVM does this
> >> sanity check for supported arch events, it won't break anything.
> > The hardware PMU most assuredly does not return 0 for unsupported events.
> >
> > For example, if I use host perf to sample event selector 0xa4 unit
> > mask 1 on a Broadwell host (406f1), I get...
> >
> > # perf stat -e r01a4 sleep 10
> >
> >   Performance counter stats for 'sleep 10':
> >
> >             386,964      r01a4
> >
> >        10.000907211 seconds time elapsed
> >
> > Broadwell does not advertise support for architectural event 7 in
> > CPUID.0AH:EBX, so KVM will refuse to measure this event inside a
> > guest. That seems broken to me.
>
>
> Yeah, I also saw similar results on Coffee Lake which doesn't support
> slots events and the return count seems to be a random and meaningless
> value. If so, this meaningless value may mislead the guest perf user.
>  From this point view it looks the sanity check in KVM is useful, but it
> indeed leads to different behavior between guest and host.

Calling this a "sanity check" is somewhat specious, since the guards
are based on the guest CPUID rather than the host CPUID. There is
nothing to prevent the hypervisor from setting guest
CPUID.0AH:EAX[31:24] to 32 and guest CPUID.0AH:EBX to 0, making 32
architectural events available to the guest. If it were really a
sanity check, it would prevent the guest from using architectural
events that are not supported by the host.

Moreover, I'm pretty sure that a "sanity check" would only apply to
top-down slots. Have there been any physical CPUs to date that support
architectural events, but do not support all of the first seven
architectural events?

> I'm neutral on either keeping or removing this check. How's other
> reviewers' opinion on this?
>
>
> >
> >>> Thanks,
> >>> Kan
> >>>> Add Kan to confirm this.
> >>>>
> >>>> Hi Kan,
> >>>>
> >>>> Have you any comments on this? Thanks.
> >>>>
> >>>>
> >>>>>>> AFAICT, this change just enables event filtering based on
> >>>>>>> CPUID.0AH:EBX[bit 7] (though it's not clear to me why two independent
> >>>>>>> mechanisms are necessary for event filtering).
> >>>>>> IMO, these are two different things. this change is just to enable the
> >>>>>> supported arch events check for slot events, the event filtering is
> >>>>>> another thing.
> >>>>> How is clearing CPUID.0AH:EBX[bit 7] any different from putting {event
> >>>>> select 0xa4, unit mask 1} in a deny list with the PMU event filter?
> >>>> I think there is no difference in the conclusion but with two different
> >>>> methods.
> >>>>
> >>>>

Powered by blists - more mailing lists