Re: [PATCH] platform/x86: int3472: Fix double free of GPIO device during unregister

[Dan Scally <dan.scally@ideasonboard.com>]

On 28/10/2025 10:54, Andy Shevchenko wrote:
> On Tue, Oct 28, 2025 at 11:38:00AM +0100, Hans de Goede wrote:
>> On 28-Oct-25 11:02 AM, Andy Shevchenko wrote:
>>> On Tue, Oct 28, 2025 at 08:55:07AM +0000, Dan Scally wrote:
>>>> On 24/10/2025 06:05, Qiu Wenbo wrote:
>>>>>
>>>>> regulator_unregister() already frees the associated GPIO device. On
>>>>> ThinkPad X9 (Lunar Lake), this causes a double free issue that leads to
>>>>> random failures when other drivers (typically Intel THC) attempt to
>>>>> allocate interrupts. The root cause is that the reference count of the
>>>>> pinctrl_intel_platform module unexpectedly drops to zero when this
>>>>> driver defers its probe.
>>>>>
>>>>> This behavior can also be reproduced by unloading the module directly.
>>>>>
>>>>> Fix the issue by removing the redundant release of the GPIO device
>>>>> during regulator unregistration.
>>>>>
>>>>> Fixes: 1e5d088a52c2 ("platform/x86: int3472: Stop using devm_gpiod_get()")
>>>
>>>> However the Fixes tag I wonder about; devm_gpiod_get() will also result in a
>>>> call to gpiod_put() when the module is unloaded; doesn't that mean that the
>>>> same issue  will occur before that commit?
>>>
>>> Actually a good question! To me sounds like it's a bug(?) in regulator code.
>>> It must not release resources it didn't acquire. This sounds like a clear
>>> layering violation.
>>
>> I think the problem is that when it comes from devicetree it is acquired
>> by the regulator core.
> 
> Hmm... I probably missed that, but I failed to see this. Any pointers?

They can come through the struct regulator_desc.of_parse_cb(), which is called in
regulator_of_init_data(), from regulator_register(). For example: 
https://elixir.bootlin.com/linux/v6.17.5/source/drivers/power/supply/mt6370-charger.c#L234>
>> Only when passed as platform-data as we do here does
>> this layering violation occur.
>>
>> I do believe that a transfer of ownership ad done here is ok for
>> the platform-data special case.
>