On 19.11.2021 14:51, Takashi Iwai wrote:
> On Thu, 18 Nov 2021 23:13:50 +0100,
> Heiner Kallweit wrote:
>>
>> On 18.11.2021 22:28, Takashi Iwai wrote:
>>> On Thu, 18 Nov 2021 21:33:34 +0100,
>>> Heiner Kallweit wrote:
>>>>
>>>> I get the following warning caused by 4f66a9ef37d3 ("ALSA: hda: intel: More
>>>> comprehensive PM runtime setup for controller driver"):
>>>>
>>>> snd_hda_intel 0000:00:1f.3: Unbalanced pm_runtime_enable!
>>>>
>>>> Not sure how this patch was tested because the warning is obvious.
>>>> The patch doesn't consider what the PCI sub-system does with regard to
>>>> RPM. Have a look at pci_pm_init().
>>>>
>>>> I'd understand to add the call to pm_runtime_dont_use_autosuspend(),
>>>> but for all other added calls I see no justification.
>>>>
>>>> If being unsure about when to use which RPM call best involve
>>>> linux-pm@xxxxxxxxxxxxxxx.
>>>
>>> Thanks for the notice. It's been through Intel CI and tests on a few
>>> local machines, maybe we haven't checked carefully those errors but
>>> only concentrated on the other issues, as it seems.
>>>
>>> There were two problems: one was the runtime PM being kicked off even
>>> during the PCI driver remove call, and another was the proper runtime
>>> PM setup after re-binding.
>>>
>>
>> Having a look at the commit message of "ALSA: hda: fix general protection
>> fault in azx_runtime_idle" the following sounds weird:
>>
>> - pci-driver.c:pm_runtime_put_sync() leads to a call
>> to rpm_idle() which again calls azx_runtime_idle()
>>
>> rpm_idle() is only called if usage_count is 1 when entering
>> pm_runtime_put_sync. And this should not be the case.
>> pm_runtime_get_sync() increments the usage counter before remove()
>> is called, and remove() should also increment the usage counter.
>> This doesn't seem to happen. Maybe for whatever reason
>> pm_runtime_get_noresume() isn't called in azx_free(), or azx_free()
>> isn't called from remove().
>> I think you should trace the call chain from the PCI core calling
>> remove() to pm_runtime_get_noresume() getting called or not.
>
> Neither of them, supposedly. Now I took a deeper look at the code
> around it and dug into the git log, and found that the likely problem
> was the recent PCI core code refactoring (removal of pci->driver, etc)
> that have been already reverted; that was why linux-next-20211109 was
> broken and linux-next-20211110 worked. With the leftover pci->driver,
> the stale runtime PM callback was called at the pm_runtime_put_sync()
> call in pci_device_remove().
>
I also noticed that partially I was on the wrong path.
> In anyway, I'll drop the invalid calls of pm_runtime_enable() /
> disable() & co. Maybe keeping pm_runtime_forbid() and
> pm_runtime_dont_use_autosuspend() at remove still makes some sense as
> a counter-part for the probe calls, though.
>
The call to pm_runtime_forbid() in pci_pm_init() is a heritage from
early ACPI times when broken ACPI implementations had problems with RPM.
There's a discussion (w/o result yet) to enable RPM per default for
newer ACPI versions.
Calling pm_runtime_forbid() in the driver removal path isn't strictly
needed but it doesn't hurt.
>
> thanks,
>
> Takashi
>
