Re: [PATCH] cpufreq/amd-pstate: Bail out early in virtual machines and don't warn

[Mario Limonciello <mario.limonciello@amd.com>]

On 7/15/26 15:32, Jason Andryuk wrote:
> On 2026-07-15 15:21, Rong Zhang wrote:
>> Hi Borislav,
>>
>> Thanks for reviewing the patch.
>>
>> On Wed, 2026-07-15 at 11:04 -0700, Borislav Petkov wrote:
>>> On Wed, Jul 15, 2026 at 12:46:24PM -0500, Mario Limonciello wrote:
>>>> Conceptually this makes sense to me, but I would like some comments 
>>>> from
>>>> others if this is the right way forward.
>>>>
>>>> Jason, Boris,
>>>>
>>>> comments please?
>>>
>>> We don't care if someone "passes-through a relatively new CPUID" - I 
>>> am not
>>> sure what that means -
>>>
>>
>> Sorry, typo here. I meant CPU model.
>>
>> I usually boot a KVM with `qemu-system-x86_64 -accel kvm -cpu host
>> <other_args...>' to test various patches. These messages is a bit
>> disturbing to me as they are KERN_WARNING messages.
>>
>> FYI, my device is Zen 5 based and always triggers both messages.
>>
>> I can also reproduce both messages with `-cpu EPYC-Genoa' and `-cpu
>> EPYC-Turin'. It's just due to how amd_cppc_supported() is implemented.
>>
>>> but we won't patch the upstream kernel just because it
>>> dumped an error message in the guest.
>>
>> Well, the reason why I decided to send this is not only about how noisy
>> it is, but also about the correctness.
>>
>> The first messages says:
>>
>>     The CPPC feature is supported but currently disabled by the BIOS.
>>     Please enable it if your BIOS has the CPPC option.
>>
>> The CPPC feature is neither supported in a KVM nor disabled by the VM's
>> BIOS.
>>
>> IIUC, it's disabled by kvm_cpu_cap_init(CPUID_8000_0008_EBX, ...), so
>> the upstream kernel both disables it as a hypervisor and warns about it
>> as a guest. What a paradox!
>>
>> At least the first message is nonsense in this case and the paradox
>> should really be fixed. If you prefer preserving the second message
>> ("the _CPC object is not present in SBIOS or ACPI disabled") as an
>> informative hint anyway, how about moving the virtualization check into
>> amd_cppc_supported() to only mute the first one? ;-)
> 
> I think changing amd_cppc_supported() to return false for 
> X86_FEATURE_HYPERVISOR makes sense and can remove the message.  The 
> heuristic there doesn't seem relevant when virtualized.
> 
> With Xen, the Xen hypervisor itself controls CPPC.  dom0 
> (xen_initial_domain()) doesn't need the amd-pstate driver.
> 
> Regards,
> Jason

OK, Rong - please send an updated patch.