Re: [PATCH] ACPI: processor: Add acpi_processor_start() back to parse _CPC tables before CPU online

["zhenglifeng (A)" <zhenglifeng1@huawei.com>]

On 2026/1/28 2:00, Rafael J. Wysocki wrote:
> +linux-pm and Viresh
> 
> On Tue, Jan 27, 2026 at 5:58 PM Jonathan Cameron
> <jonathan.cameron@huawei.com> wrote:
>>
>> On Tue, 27 Jan 2026 15:42:16 +0100
>> "Rafael J. Wysocki" <rafael@kernel.org> wrote:
>>
>>> On Tue, Jan 20, 2026 at 12:33 PM Lifeng Zheng <zhenglifeng1@huawei.com> wrote:
>>>>
>>>> Currently, if boot with maxcpus less than NR_CPUS, the cppc_cpufreq driver
>>>> will fail to register. Because it requires the domain information of all
>>>> possible CPUs to construct shared_cpu_map, which shows the CPUs that share
>>>> the same domain.
>>>>
>>>> Commit c1385c1f0ba3 ("ACPI: processor: Simplify initial onlining to use
>>>> same path for cold and hotplug") removes probe() of acpi_processor_driver
>>>> and makes acpi_cppc_processor_probe() only being called the first time CPU
>>>> goes online. This means that CPUs that haven't yet gone online will not
>>>> have pre-parsed _CPC objects and causes cppc_cpufreq driver register fail.
>>>>
>>>> Add acpi_processor_start() back as the probe() callback of
>>>> acpi_processor_driver and call acpi_cppc_processor_probe() in it to make
>>>> sure all _CPC tables will be parsed when acpi_processor_driver registered.
>>>>
>>>> Fixes: c1385c1f0ba3 ("ACPI: processor: Simplify initial onlining to use same path for cold and hotplug")
>>>> Signed-off-by: Lifeng Zheng <zhenglifeng1@huawei.com>
>>>> ---
>>>>  drivers/acpi/processor_driver.c | 30 ++++++++++++++++++++++++++----
>>>>  1 file changed, 26 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/drivers/acpi/processor_driver.c b/drivers/acpi/processor_driver.c
>>>> index 65e779be64ff..c8b4daf580b0 100644
>>>> --- a/drivers/acpi/processor_driver.c
>>>> +++ b/drivers/acpi/processor_driver.c
>>>> @@ -33,6 +33,7 @@ MODULE_AUTHOR("Paul Diefenbaugh");
>>>>  MODULE_DESCRIPTION("ACPI Processor Driver");
>>>>  MODULE_LICENSE("GPL");
>>>>
>>>> +static int acpi_processor_start(struct device *dev);
>>>>  static int acpi_processor_stop(struct device *dev);
>>>>
>>>>  static const struct acpi_device_id processor_device_ids[] = {
>>>> @@ -46,6 +47,7 @@ static struct device_driver acpi_processor_driver = {
>>>>         .name = "processor",
>>>>         .bus = &cpu_subsys,
>>>>         .acpi_match_table = processor_device_ids,
>>>> +       .probe = acpi_processor_start,
>>>>         .remove = acpi_processor_stop,
>>>>  };
>>>>
>>>> @@ -162,10 +164,6 @@ static int __acpi_processor_start(struct acpi_device *device)
>>>>         if (!pr)
>>>>                 return -ENODEV;
>>>>
>>>> -       result = acpi_cppc_processor_probe(pr);
>>>> -       if (result && !IS_ENABLED(CONFIG_ACPI_CPU_FREQ_PSS))
>>>> -               dev_dbg(&device->dev, "CPPC data invalid or not present\n");
>>>> -
>>>>         if (!cpuidle_get_driver() || cpuidle_get_driver() == &acpi_idle_driver)
>>>>                 acpi_processor_power_init(pr);
>>>>
>>>> @@ -192,6 +190,30 @@ static int __acpi_processor_start(struct acpi_device *device)
>>>>         return result;
>>>>  }
>>>>
>>>> +static int acpi_processor_start(struct device *dev)
>>>> +{
>>>> +       struct acpi_device *device = ACPI_COMPANION(dev);
>>>> +       struct acpi_processor *pr;
>>>> +       int result;
>>>> +
>>>> +       if (!device)
>>>> +               return -ENODEV;
>>>> +
>>>> +       pr = acpi_driver_data(device);
>>>> +       if (!pr)
>>>> +               return -ENODEV;
>>>> +
>>>> +       /* Protect against concurrent CPU hotplug operations */
>>>> +       cpu_hotplug_disable();
>>>> +       result = acpi_cppc_processor_probe(pr);
>>>> +       cpu_hotplug_enable();
>>>
>>> This means that CPPC will be initialized for vCPUs that are not
>>> enabled on ARM if I'm not mistaken.
>>
>> If we are just talking powered down at boot it used to do that
>> so I assume it was fine. The corner case is ones we are explicitly
>> saying are not onlineable yet but marked online capable and will
>> turn up later.
>>
>>>
>>> I'm not sure if it is valid to do so.
>>
>> The conclusion of the following is I think this is fine but I'm not
>> entirely confident about it.
>>
>> I'm struggling to figure out the right answer to this and
>> it's not easy to test. I vaguely recall having some nasty emulation
>> hacks to poke some x86 related _CPC stuff a while back.
>> I might be able to hack something up for this as well and try to
>> create pathological corner cases.
>>
>> The short answer is CPPC + hotplug isn't a thing today in KVM + QEMU,
>> but that's not to say it never will be if someone virtualizes CPC for
>> a guest.  Let's consider that hypothetical virtualization / emulation.
>>
>> So the questions:
>> 1) Does simply making this acpi_cppc_processor_probe() result in any
>>    register accesses to the registers that might be found in _CPC or
>>    used via other ACPI methods?
>> 2) Can we rely on a a VMM not do something nasty if those are accessed
>>    on CPUs that haven't been instantiated yet?  e.g. Bus error.
>>    A related useful question is: Can we assume these registers are
>>    accessible on offlined CPUs?  If they can be unsafe to access from
>>    CPUs that are temporary powered down / offline then I think we are fine because
>>    the CPPC code must guarantee not to access them. (I'm relying on this!)
>>
>> For the particular case Lifeng has run into, I think the code that matters
>> (beyond instantiation of the infrastructure) is the creation of the
>> domain info in acpi_get_psd(). I think _PSD can only be static data so
>> shouldn't cause any register accesses to the powered down CPUs.
>>
>> So 'probably' fine + we'll not really know unless we get CPU HP and
>> CPC.
>>
>> Alternative much more complex change would be to separate the grabbing of
>> static data (done here) from setting up anything dynamic which would remain
>> in the hotplug handler.  If those registers haven't been discovered we definitely
>> can't access them from the cpu freq driver.
> 
> I'm thinking that maybe cppc_cpufreq should be updated instead.
> 
> I'm not really sure why it needs to collect information on offline
> CPUs.  Surely, they don't matter until they are brought online.

This information is collected in order to generate related_cpus. Without
doing so, a new policy will be created when the second CPU in the same
domain comes online, instead of reusing the existing policy. And this will
make a mess.

I can't find a good way to solve this problem in cppc_cpufreq or cpufreq.

>