Re: [PATCH v1 5/9] ACPI: processor: idle: Add the verification of processor FFH LPI state

["lihuisong (C)" <lihuisong@huawei.com>]

在 2025/10/26 20:40, Rafael J. Wysocki 写道:
> On Fri, Oct 24, 2025 at 11:40 AM lihuisong (C) <lihuisong@huawei.com> wrote:
>>
>> 在 2025/10/23 18:35, Rafael J. Wysocki 写道:
>>> On Thu, Oct 23, 2025 at 12:17 PM lihuisong (C) <lihuisong@huawei.com> wrote:
>>>> 在 2025/10/22 3:42, Rafael J. Wysocki 写道:
>>>>> On Mon, Sep 29, 2025 at 11:38 AM Huisong Li <lihuisong@huawei.com> wrote:
>>>>>> Both ARM64 and RISCV architecture would validate Entry Method of LPI
>>>>>> state and SBI HSM or PSCI cpu suspend. Driver should return failure
>>>>>> if FFH of LPI state are not ok.
>>>>> First of all, I cannot parse this changelog, so I don't know the
>>>>> motivation for the change.
>>>> Sorry for your confusion.
>>>>> Second, if _LPI is ever used on x86, the
>>>>> acpi_processor_ffh_lpi_probe() in acpi_processor_get_power_info() will
>>>>> get in the way.
>>>> AFAICS, it's also ok if X86 platform use LPI.
>>> No, because it returns an error by default as it stands today.
>>>
>>>>> Why does the evaluation in acpi_processor_setup_cpuidle_dev() not work?
>>>> The acpi_processor_ffh_lpi_probe does verify the validity of LPI for ARM
>>>> and RISCV.
>>>> But the caller of the acpi_processor_setup_cpuidle_dev()don't verify the
>>>> return value.
>>>> In addition, from the name of acpi_processor_setup_cpuidle_dev(), its
>>>> main purpose is to setup cpudile device rather than to verify LPI.
>>> That's fair enough.
>>>
>>> Also, the list of idle states belongs to the cpuidle driver, not to a
>>> cpuidle device.
>>>
>>>> So I move it to a more prominent position and redefine the
>>>> acpi_processor_setup_cpuidle_dev to void in patch 9/9.
>>>>>> Fixes: a36a7fecfe60 ("ACPI / processor_idle: Add support for Low Power Idle(LPI) states")
>>>>>> Signed-off-by: Huisong Li <lihuisong@huawei.com>
>>>>>> ---
>>>>>>     drivers/acpi/processor_idle.c | 10 ++++++++--
>>>>>>     1 file changed, 8 insertions(+), 2 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/acpi/processor_idle.c b/drivers/acpi/processor_idle.c
>>>>>> index 5684925338b3..b0d6b51ee363 100644
>>>>>> --- a/drivers/acpi/processor_idle.c
>>>>>> +++ b/drivers/acpi/processor_idle.c
>>>>>> @@ -1264,7 +1264,7 @@ static int acpi_processor_setup_cpuidle_dev(struct acpi_processor *pr,
>>>>>>
>>>>>>            dev->cpu = pr->id;
>>>>>>            if (pr->flags.has_lpi)
>>>>>> -               return acpi_processor_ffh_lpi_probe(pr->id);
>>>>>> +               return 0;
>>>>>>
>>>>>>            return acpi_processor_setup_cpuidle_cx(pr, dev);
>>>>>>     }
>>>>>> @@ -1275,7 +1275,13 @@ static int acpi_processor_get_power_info(struct acpi_processor *pr)
>>>>>>
>>>>>>            ret = acpi_processor_get_lpi_info(pr);
>>>>>>            if (ret)
>>> So I think it would be better to check it here, that is
>>>
>>> if (!ret) {
>>>          ret = acpi_processor_ffh_lpi_probe(pr->id));
>>>          if (!ret)
>>>                  return 0;
>>>
>>>          pr_info("CPU%d: FFH LPI state is invalid\n", pr->id);
>>>          pr->flags.has_lpi = 0;
>>> }
>>>
>>> return acpi_processor_get_cstate_info(pr);
>>>
>>> And the default acpi_processor_ffh_lpi_probe() needs to be changed to return 0.
>> Sorry, I don't understand why pr->flags.has_lpi is true if
>> acpi_processor_ffh_lpi_probe() return failure.
> It is set by acpi_processor_get_lpi_info() on success and
> acpi_processor_ffh_lpi_probe() does not update it.
The acpi_processor_get_lpi_info() will return failure on X86 platform 
because this function first call acpi_processor_ffh_lpi_probe().
And acpi_processor_ffh_lpi_probe return EOPNOTSUPP because X86 platform 
doesn't implement it.
So I think pr->flags.has_lpi is false on X86 plaform.
>
>> In addition, X86 platform doesn't define acpi_processor_ffh_lpi_probe().
>> this function will return EOPNOTSUPP.
> Which is exactly why it is a problem.  x86 has no reason to implement
> it because FFH always works there.
Sorry, I still don't understand why X86 has no reason to implement it.
I simply think that X86 doesn't need it.
AFAICS, the platform doesn't need to get LPI info if this platform 
doesn't implement acpi_processor_ffh_lpi_probe().
>
>>>>>> -               ret = acpi_processor_get_cstate_info(pr);
>>>>>> +               return acpi_processor_get_cstate_info(pr);
>>>>>> +
>>>>>> +       if (pr->flags.has_lpi) {
>>>>>> +               ret = acpi_processor_ffh_lpi_probe(pr->id);
>>>>>> +               if (ret)
>>>>>> +                       pr_err("Processor FFH LPI state is invalid.\n");
>>>>>> +       }
>>>>>>
>>>>>>            return ret;
>>>>>>     }
>>>>>> --