Messages in this thread | | | Date | Tue, 23 Apr 2024 16:24:45 +0800 | Subject | Re: [RFC PATCH 23/41] KVM: x86/pmu: Implement the save/restore of PMU state for Intel CPU | From | "Mi, Dapeng" <> |
| |
On 4/23/2024 3:10 PM, Mingwei Zhang wrote: > On Mon, Apr 22, 2024 at 11:45 PM Mi, Dapeng <dapeng1.mi@linux.intel.com> wrote: >> >> On 4/23/2024 2:08 PM, maobibo wrote: >>> >>> On 2024/4/23 下午12:23, Mingwei Zhang wrote: >>>> On Mon, Apr 22, 2024 at 8:55 PM maobibo <maobibo@loongson.cn> wrote: >>>>> >>>>> >>>>> On 2024/4/23 上午11:13, Mi, Dapeng wrote: >>>>>> On 4/23/2024 10:53 AM, maobibo wrote: >>>>>>> >>>>>>> On 2024/4/23 上午10:44, Mi, Dapeng wrote: >>>>>>>> On 4/23/2024 9:01 AM, maobibo wrote: >>>>>>>>> >>>>>>>>> On 2024/4/23 上午1:01, Sean Christopherson wrote: >>>>>>>>>> On Mon, Apr 22, 2024, maobibo wrote: >>>>>>>>>>> On 2024/4/16 上午6:45, Sean Christopherson wrote: >>>>>>>>>>>> On Mon, Apr 15, 2024, Mingwei Zhang wrote: >>>>>>>>>>>>> On Mon, Apr 15, 2024 at 10:38 AM Sean Christopherson >>>>>>>>>>>>> <seanjc@google.com> wrote: >>>>>>>>>>>>>> One my biggest complaints with the current vPMU code is that >>>>>>>>>>>>>> the roles and >>>>>>>>>>>>>> responsibilities between KVM and perf are poorly defined, >>>>>>>>>>>>>> which >>>>>>>>>>>>>> leads to suboptimal >>>>>>>>>>>>>> and hard to maintain code. >>>>>>>>>>>>>> >>>>>>>>>>>>>> Case in point, I'm pretty sure leaving guest values in PMCs >>>>>>>>>>>>>> _would_ leak guest >>>>>>>>>>>>>> state to userspace processes that have RDPMC permissions, as >>>>>>>>>>>>>> the PMCs might not >>>>>>>>>>>>>> be dirty from perf's perspective (see >>>>>>>>>>>>>> perf_clear_dirty_counters()). >>>>>>>>>>>>>> >>>>>>>>>>>>>> Blindly clearing PMCs in KVM "solves" that problem, but in >>>>>>>>>>>>>> doing so makes the >>>>>>>>>>>>>> overall code brittle because it's not clear whether KVM >>>>>>>>>>>>>> _needs_ >>>>>>>>>>>>>> to clear PMCs, >>>>>>>>>>>>>> or if KVM is just being paranoid. >>>>>>>>>>>>> So once this rolls out, perf and vPMU are clients directly to >>>>>>>>>>>>> PMU HW. >>>>>>>>>>>> I don't think this is a statement we want to make, as it opens a >>>>>>>>>>>> discussion >>>>>>>>>>>> that we won't win. Nor do I think it's one we *need* to make. >>>>>>>>>>>> KVM doesn't need >>>>>>>>>>>> to be on equal footing with perf in terms of owning/managing PMU >>>>>>>>>>>> hardware, KVM >>>>>>>>>>>> just needs a few APIs to allow faithfully and accurately >>>>>>>>>>>> virtualizing a guest PMU. >>>>>>>>>>>> >>>>>>>>>>>>> Faithful cleaning (blind cleaning) has to be the baseline >>>>>>>>>>>>> implementation, until both clients agree to a "deal" between >>>>>>>>>>>>> them. >>>>>>>>>>>>> Currently, there is no such deal, but I believe we could have >>>>>>>>>>>>> one via >>>>>>>>>>>>> future discussion. >>>>>>>>>>>> What I am saying is that there needs to be a "deal" in place >>>>>>>>>>>> before this code >>>>>>>>>>>> is merged. It doesn't need to be anything fancy, e.g. perf can >>>>>>>>>>>> still pave over >>>>>>>>>>>> PMCs it doesn't immediately load, as opposed to using >>>>>>>>>>>> cpu_hw_events.dirty to lazily >>>>>>>>>>>> do the clearing. But perf and KVM need to work together from >>>>>>>>>>>> the >>>>>>>>>>>> get go, ie. I >>>>>>>>>>>> don't want KVM doing something without regard to what perf does, >>>>>>>>>>>> and vice versa. >>>>>>>>>>>> >>>>>>>>>>> There is similar issue on LoongArch vPMU where vm can directly >>>>>>>>>>> pmu >>>>>>>>>>> hardware >>>>>>>>>>> and pmu hw is shard with guest and host. Besides context switch >>>>>>>>>>> there are >>>>>>>>>>> other places where perf core will access pmu hw, such as tick >>>>>>>>>>> timer/hrtimer/ipi function call, and KVM can only intercept >>>>>>>>>>> context switch. >>>>>>>>>> Two questions: >>>>>>>>>> >>>>>>>>>> 1) Can KVM prevent the guest from accessing the PMU? >>>>>>>>>> >>>>>>>>>> 2) If so, KVM can grant partial access to the PMU, or is it all >>>>>>>>>> or nothing? >>>>>>>>>> >>>>>>>>>> If the answer to both questions is "yes", then it sounds like >>>>>>>>>> LoongArch *requires* >>>>>>>>>> mediated/passthrough support in order to virtualize its PMU. >>>>>>>>> Hi Sean, >>>>>>>>> >>>>>>>>> Thank for your quick response. >>>>>>>>> >>>>>>>>> yes, kvm can prevent guest from accessing the PMU and grant partial >>>>>>>>> or all to access to the PMU. Only that if one pmu event is granted >>>>>>>>> to VM, host can not access this pmu event again. There must be pmu >>>>>>>>> event switch if host want to. >>>>>>>> PMU event is a software entity which won't be shared. did you >>>>>>>> mean if >>>>>>>> a PMU HW counter is granted to VM, then Host can't access the PMU HW >>>>>>>> counter, right? >>>>>>> yes, if PMU HW counter/control is granted to VM. The value comes from >>>>>>> guest, and is not meaningful for host. Host pmu core does not know >>>>>>> that it is granted to VM, host still think that it owns pmu. >>>>>> That's one issue this patchset tries to solve. Current new mediated >>>>>> x86 >>>>>> vPMU framework doesn't allow Host or Guest own the PMU HW resource >>>>>> simultaneously. Only when there is no !exclude_guest event on host, >>>>>> guest is allowed to exclusively own the PMU HW resource. >>>>>> >>>>>> >>>>>>> Just like FPU register, it is shared by VM and host during different >>>>>>> time and it is lately switched. But if IPI or timer interrupt uses >>>>>>> FPU >>>>>>> register on host, there will be the same issue. >>>>>> I didn't fully get your point. When IPI or timer interrupt reach, a >>>>>> VM-exit is triggered to make CPU traps into host first and then the >>>>>> host >>>>> yes, it is. >>>> This is correct. And this is one of the points that we had debated >>>> internally whether we should do PMU context switch at vcpu loop >>>> boundary or VM Enter/exit boundary. (host-level) timer interrupt can >>>> force VM Exit, which I think happens every 4ms or 1ms, depending on >>>> configuration. >>>> >>>> One of the key reasons we currently propose this is because it is the >>>> same boundary as the legacy PMU, i.e., it would be simple to propose >>>> from the perf subsystem perspective. >>>> >>>> Performance wise, doing PMU context switch at vcpu boundary would be >>>> way better in general. But the downside is that perf sub-system lose >>>> the capability to profile majority of the KVM code (functions) when >>>> guest PMU is enabled. >>>> >>>>>> interrupt handler is called. Or are you complaining the executing >>>>>> sequence of switching guest PMU MSRs and these interrupt handler? >>>>> In our vPMU implementation, it is ok if vPMU is switched in vm exit >>>>> path, however there is problem if vPMU is switched during vcpu thread >>>>> sched-out/sched-in path since IPI/timer irq interrupt access pmu >>>>> register in host mode. >>>> Oh, the IPI/timer irq handler will access PMU registers? I thought >>>> only the host-level NMI handler will access the PMU MSRs since PMI is >>>> registered under NMI. >>>> >>>> In that case, you should disable IRQ during vcpu context switch. For >>>> NMI, we prevent its handler from accessing the PMU registers. In >>>> particular, we use a per-cpu variable to guard that. So, the >>>> host-level PMI handler for perf sub-system will check the variable >>>> before proceeding. >>> perf core will access pmu hw in tick timer/hrtimer/ipi function call, >>> such as function perf_event_task_tick() is called in tick timer, there >>> are event_function_call(event, __perf_event_xxx, &value) in file >>> kernel/events/core.c. >>> >>> https://lore.kernel.org/lkml/20240417065236.500011-1-gaosong@loongson.cn/T/#m15aeb79fdc9ce72dd5b374edd6acdcf7a9dafcf4 >>> >> Just go through functions (not sure if all), whether >> perf_event_task_tick() or the callbacks of event_function_call() would >> check the event->state first, if the event is in >> PERF_EVENT_STATE_INACTIVE, the PMU HW MSRs would not be touched really. >> In this new proposal, all host events with exclude_guest attribute would >> be put on PERF_EVENT_STATE_INACTIVE sate if guest own the PMU HW >> resource. So I think it's fine. >> > Is there any event in the host still having PERF_EVENT_STATE_ACTIVE? > If so, hmm, it will reach perf_pmu_disable(event->pmu), which will > access the global ctrl MSR.
I don't think there is any event with PERF_EVENT_STATE_ACTIVE state on host when guest owns the PMU HW resource.
In current solution, VM would fail to create if there is any system-wide event without exclude_guest attribute. If VM is created successfully and when vm-entry happens, the helper perf_guest_enter() would put all host events with exclude_guest attribute into PERF_EVENT_STATE_INACTIVE state and block host to create system-wide events without exclude_guest attribute.
| |