Re: [PATCH] KVM: x86/svm: Add module param to control PMU virtualization

[Jim Mattson]

On Mon, Jan 10, 2022 at 10:18 PM Like Xu <like.xu.linux@xxxxxxxxx> wrote:
>
> On 11/1/2022 11:24 am, Jim Mattson wrote:
> > On Mon, Jan 10, 2022 at 6:11 PM Like Xu <like.xu.linux@xxxxxxxxx> wrote:
> >>
> >> On 11/1/2022 2:13 am, Jim Mattson wrote:
> >>> On Sun, Jan 9, 2022 at 10:23 PM Like Xu <like.xu.linux@xxxxxxxxx> wrote:
> >>>>
> >>>> On 9/1/2022 9:23 am, Jim Mattson wrote:
> >>>>> On Fri, Dec 10, 2021 at 7:48 PM Jim Mattson <jmattson@xxxxxxxxxx> wrote:
> >>>>>>
> >>>>>> On Fri, Dec 10, 2021 at 6:15 PM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
> >>>>>>>
> >>>>>>> On 12/10/21 20:25, Jim Mattson wrote:
> >>>>>>>> In the long run, I'd like to be able to override this system-wide
> >>>>>>>> setting on a per-VM basis, for VMs that I trust. (Of course, this
> >>>>>>>> implies that I trust the userspace process as well.)
> >>>>>>>>
> >>>>>>>> How would you feel if we were to add a kvm ioctl to override this
> >>>>>>>> setting, for a particular VM, guarded by an appropriate permissions
> >>>>>>>> check, like capable(CAP_SYS_ADMIN) or capable(CAP_SYS_MODULE)?
> >>>>>>>
> >>>>>>> What's the rationale for guarding this with a capability check?  IIRC
> >>>>>>> you don't have such checks for perf_event_open (apart for getting kernel
> >>>>>>> addresses, which is not a problem for virtualization).
> >>>>>>
> >>>>>> My reasoning was simply that for userspace to override a mode 0444
> >>>>>> kernel module parameter, it should have the rights to reload the
> >>>>>> module with the parameter override. I wasn't thinking specifically
> >>>>>> about PMU capabilities.
> >>>>
> >>>> Do we have a precedent on any module parameter rewriting for privileger ?
> >>>>
> >>>> A further requirement is whether we can dynamically change this part of
> >>>> the behaviour when the guest is already booted up.
> >>>>
> >>>>>
> >>>>> Assuming that we trust userspace to decide whether or not to expose a
> >>>>> virtual PMU to a guest (as we do on the Intel side), perhaps we could
> >>>>> make use of the existing PMU_EVENT_FILTER to give us per-VM control,
> >>>>> rather than adding a new module parameter for per-host control. If
> >>>>
> >>>> Various granularities of control are required to support vPMU production
> >>>> scenarios, including per-host, per-VM, and dynamic-guest-alive control.
> >>>>
> >>>>> userspace calls KVM_SET_PMU_EVENT_FILTER with an action of
> >>>>> KVM_PMU_EVENT_ALLOW and an empty list of allowed events, KVM could
> >>>>> just disable the virtual PMU for that VM.
> >>>>
> >>>> AMD will also have "CPUID Fn8000_0022_EBX[NumCorePmc, 3:0]".
> >>>
> >>> Where do you see this? Revision 3.33 (November 2021) of the AMD APM,
> >>> volume 3, only goes as high as CPUID Fn8000_0021.
> >>
> >> Try APM Revision: 4.04 (November 2021),  page 1849/3273,
> >> "CPUID Fn8000_0022_EBX Extended Performance Monitoring and Debug".
> >
> > Is this a public document?
>
> The latest version of APM (40332) is revision v4.04, released on 12/1/2021.

LOL. I was misled by the table of contents for Appendix E.4, which
stops at E.4.19 Function 8000_0021—Extended Feature Identification 2.

> >
> >> Given the current ambiguity in this revision, the AMD folks will reveal more
> >> details bout this field in the next revision.
> >>
> >>>
> >>>>>
> >>>>> Today, the semantics of an empty allow list are quite different from
> >>>>> the proposed pmuv module parameter being false. However, it should be
> >>>>> an easy conversion. Would anyone be concerned about changing the
> >>>>> current semantics of an empty allow list? Is there a need for
> >>>>> disabling PMU virtualization for legacy userspace implementations that
> >>>>> can't be modified to ask for an empty allow list?
> >>>>>
> >>>>
> >>>> AFAI, at least one user-space agent has integrated with it plus additional
> >>>> "action"s.
> >>>>
> >>>> Once the API that the kernel presents to user space has been defined,
> >>>> it's best not to change it and instead fall into remorse.
> >>>
> >>> Okay.
> >>>
> >>> I propose the following:
> >>> 1) The new module parameter should apply to Intel as well as AMD, for
> >>> situations where userspace is not trusted.
> >>> 2) If the module parameter allows PMU virtualization, there should be
> >>> a new KVM_CAP whereby userspace can enable/disable PMU virtualization.
> >>> (Since you require a dynamic toggle, and there is a move afoot to
> >>> refuse guest CPUID changes once a guest is running, this new KVM_CAP
> >>> is needed on Intel as well as AMD).
> >>
> >> Both hands in favour. Do you need me as a labourer, or you have a ready-made one ?
> >
> > We could split the work. Since (1) is a modification of the change you
> > proposed in this thread, perhaps you could apply it to both AMD and
>
> We obviously need extra code to make the module parameters suitable for Intel
> since it
> affects other features (such as LBR and PEBS), we may not rush to draw this line
> clearly.
>
> > Intel in v2? We can find someone for (2).
>
> The ioctl_set_pmu_event_filter() interface is already practical for dynamic toggle,
> as not being able to program any events is the same as having none vPMU,
> w/o considering performance impact of traversing the list.

Not being able to program any events is actually not the same as
setting the pmuv module parameter to false. In the former case,
prohibited events simply don't advance the counters. In the latter
case, all accesses to the PMU MSRs raise #GP. That was what I meant
earlier, when I said that we would have to change the semantics of an
empty allow list if we want to match the behavior of pmuv=0.

> I am not sure if the maintainer will buy in another KVM_CAP for only per-VM,
> considering
> "CPUID Fn8000_0022_EBX[NumCorePmc, 3:0]" is a feature that will be available soon.

For existing guests, the future availability of that feature is
irrelevant. We're going to need another solution that doesn't involve
recompiling every guest kernel.

> >
> >>> 3) If the module parameter does not allow PMU virtualization, there
> >>> should be no userspace override, since we have no precedent for
> >>> authorizing that kind of override.
> >>
> >> Uh, I thought you (Google) had a lot of these (interesting) use cases internally.
> >
> > We have modified some module parameters so that they can be changed at
> > runtime, but we don't have any concept of a privileged userspace
> > overriding a module parameter restriction.
>
> Considering that the semantics of the different module parameters are different,
> allowing one of them to be overridden does not mean that such a generic framework
> is promising, but it makes sense for the community to see another case for it.
>
> >
> >>>
> >>>> "But I am not a decision maker. " :D
> >>>>
> >>>> Thanks,
> >>>> Like Xu
> >>>>
> >