Re: [PATCH] clk: Warn (and therefore taint the kernel) on clk_ignore_unused

[Konrad Dybcio <konrad.dybcio@oss.qualcomm.com>]

On 3/4/25 8:34 PM, Stephen Boyd wrote:
> Quoting Dmitry Baryshkov (2025-03-03 15:17:21)
>> On Tue, 4 Mar 2025 at 00:16, Florian Fainelli
>> <florian.fainelli@broadcom.com> wrote:
>>>
>>> On 3/3/25 14:48, Stephen Boyd wrote:
>>>> Quoting Konrad Dybcio (2025-02-01 08:52:30)
> [...]
>>>>>
>>>>> The clock subsystem plays a crucial part in this quest, as even if
>>>>> the clock controllers themselves don't draw a lot of power when on
>>>>> (comparatively), improper description of clock requirements has been
>>>>> the #1 cause of incomplete/incorrect devicetree bindings in my
>>>>> experience.
>>>>
>>>> What is a user supposed to do about this warning stack? We already print
>>>> a warning. I don't see us dumping the stack when a driver is unfinished
>>>> and doesn't implement runtime PM to save power.
>>>>
>>>
>>> Agreed, I don't think this is tremendously helpful given that it does
>>> not even tell you what part is incomplete, it's just a broad warning for
>>> the entire system.
>>>
>>> Assuming you have a clock provided that can be used to turn clocks off,
>>> and you did not boot with 'clk_ignore_unused' set on the kernel command
>>> line, then you should discover pretty quickly which driver is not
>>> managing the clocks as it should no?
>>
>> Unfortunately it's sometimes not that easy. And some developers
>> pretend that 'clk_ignore_unused' is a viable way to run the system.
>>
> 
> Maybe we would be better off with a config option that removes the clk
> ignore unused ability entirely. Then you can have a kernel config check
> somewhere in the build process that verifies that a user can't even set
> the kernel commandline to change the behavior.

I used WARN specifically to taint the kernel (which would in turn throw off
any reasonable CI checks). Perhaps we could add a Kconfig entry (unless
there already is one) that would do the same, and clk_ignore_unused could
be gated behind it.

But then, it would make it harder to debug production kernels with that
parameter, which could potentially come in handy too

Konrad