Re: KVM: x86/vPMU/AMD: Can we detect PMU is off for a VM?

[Dongli Zhang]

On 11/13/23 06:42, Denis V. Lunev wrote:
> On 11/13/23 15:14, Dongli Zhang wrote:
>> Hi Denis,
>>
>> On 11/13/23 01:31, Denis V. Lunev wrote:
>>> On 11/10/23 01:01, Dongli Zhang wrote:
>>>> On 11/9/23 3:46 PM, Denis V. Lunev wrote:
>>>>> On 11/9/23 23:52, Jim Mattson wrote:
>>>>>> On Thu, Nov 9, 2023 at 10:18 AM Konstantin Khorenko
>>>>>> <khorenko@xxxxxxxxxxxxx> wrote:
>>>>>>> Hi All,
>>>>>>>
>>>>>>> as a followup for my patch: i have noticed that
>>>>>>> currently Intel kernel code provides an ability to detect if PMU is totally
>>>>>>> disabled for a VM
>>>>>>> (pmu->version == 0 in this case), but for AMD code pmu->version is never 0,
>>>>>>> no matter if PMU is enabled or disabled for a VM (i mean <pmu state='off'/>
>>>>>>> in the VM config which
>>>>>>> results in "-cpu pmu=off" qemu option).
>>>>>>>
>>>>>>> So the question is - is it possible to enhance the code for AMD to also
>>>>>>> honor
>>>>>>> PMU VM setting or it is
>>>>>>> impossible by design?
>>>>>> The AMD architectural specification prior to AMD PMU v2 does not allow
>>>>>> one to describe a CPU (via CPUID or MSRs) that has fewer than 4
>>>>>> general purpose PMU counters. While AMD PMU v2 does allow one to
>>>>>> describe such a CPU, legacy software that knows nothing of AMD PMU v2
>>>>>> can expect four counters regardless.
>>>>>>
>>>>>> Having said that, KVM does provide a per-VM capability for disabling
>>>>>> the virtual PMU: KVM_CAP_PMU_CAPABILITY(KVM_PMU_CAP_DISABLE). See
>>>>>> section 8.35 in Documentation/virt/kvm/api.rst.
>>>>> But this means in particular that QEMU should immediately
>>>>> use this KVM_PMU_CAP_DISABLE if this capability is supported and PMU=off. I am
>>>>> not seeing this code thus I believe that we have missed this. I think that
>>>>> this
>>>>> change worth adding. We will measure the impact :-) Den
>>>>>
>>>> I used to have a patch to use KVM_PMU_CAP_DISABLE in QEMU, but that did not
>>>> draw
>>>> many developers' attention.
>>>>
>>>> https://urldefense.com/v3/__https://lore.kernel.org/qemu-devel/20230621013821.6874-2-dongli.zhang@xxxxxxxxxx/__;!!ACWV5N9M2RV99hQ!McSH2M-kuHmzAwTuXKxrjLkrdJoPqML6cY_Ndc-8k9LRQ7D1V9bSBRQPwHqtx9XCVLK3uzdsMaxyfwve$
>>>> It is time to first re-send that again.
>>>>
>>>> Dongli Zhang
>>> We have checked that setting KVM_PMU_CAP_DISABLE really helps. Konstantin has
>>> done this and this is good. On the other hand, looking into these patches I
>>> disagree with them. We should not introduce new option for QEMU. If PMU is
>>> disabled, i.e. we assume that pmu=off passed in the command line, we should set
>>> KVM_PMU_CAP_DISABLE for that virtual machine. Den
>> Can I assume you meant pmu=off, that is, cpu->enable_pmu in QEMU?
>>
>> In my opinion, cpu->enable_pmu indicates the option to control the cpu features.
>> It may be used by any accelerators, and it is orthogonal to the KVM cap.
>>
>>
>> The KVM_PMU_CAP_DISABLE is only specific to the KVM accelerator.
>>
>>
>> That's why I had introduced a new option, to allow to configure the VM in my
>> dimensions.
>>
>> It means one dimension to AMD, but two for Intel: to disable PMU via cpuid, or
>> KVM cap.
>>
>> Anyway, this is KVM mailing list, and I may initiate the discussion in QEMU list.
>>
>> Thank you very much!
>>
>> Dongli Zhang
> with the option pmu='off' it is expected that PMU should be
> off for the guest. At the moment (without this KVM capability)
> we can disable PMU for Intel only and thus have performance
> degradation on AMD.
> 
> This option disables PMU and thus normally when we are
> running KVM guest and wanting PMU to be off it would
> be required to
> * disable CPUID leaf for Intel
> * set KVM_PMU_CAP_DISABLE for both processors This would be quite natural and
> transparent for the libvirt. Alexander will prepare the patch today or tomorrow
> for the discussion. Den

That is what I had implemented in the v1 of patch.

https://lore.kernel.org/all/20221119122901.2469-3-dongli.zhang@xxxxxxxxxx/

However, I changed that after people suggested introduce a new property.

Dongli Zhang