Re: [PATCH] ACPI / bus: ignore rather than fail bus driver registrations on non-ACPI boot

[Ard Biesheuvel]

On 23 April 2018 at 09:39, Rafael J. Wysocki <rjw@xxxxxxxxxxxxx> wrote:
> On Sunday, April 22, 2018 2:34:17 PM CEST Ard Biesheuvel wrote:
>> On 22 April 2018 at 11:57, Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote:
>> > On 22 April 2018 at 11:27, Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
>> >> On Thu, Apr 19, 2018 at 6:58 PM, Ard Biesheuvel
>> >> <ard.biesheuvel@xxxxxxxxxx> wrote:
>> >>> When building ACPI bus drivers such as button.ko into the core kernel,
>> >>> other drivers that depend on its symbols are loadable even when booting
>> >>> with ACPI disabled. For instance, nouveau.ko has a link time dependency
>> >>> on acpi_lid_open() on ACPI capable kernels, and calls it regardless of
>> >>> whether the system booted via ACPI.
>> >>>
>> >>> However, when building button.ko as a module, it will refuse to load if
>> >>> the system did not boot in ACPI mode, which subsequently prevents the
>> >>> nouveau driver from loading as well, resulting in broken graphics.
>> >>>
>> >>> Given that returning an error from an initcall() is ignored for drivers
>> >>> that are built into the kernel,
>> >>
>> >> Which makes sense, because they are present in the kernel anyway.
>> >>
>> >>> let's align the module case with this,
>> >>> and not return an error when registering an ACPI bus driver on a system
>> >>> that did not boot via ACPI.
>> >>
>> >> But why is loading a module that's never going to be used actually OK?
>> >>
>> >> Isn't this a problem with the assumptions made by the nouveau driver
>> >> that need not be met depending on what configuration the kernel is run
>> >> in?
>> >>
>> >> Honestly, it doesn't appear quite right to try to change the rest of
>> >> the kernel to follow the nouveau's expectations.
>> >>
>> >
>> > I don't disagree here, I am just unsure whether other options are any better.
>> >
>> > I think the alternative is to make acpi_lid_open() a non-modular
>> > function of the ACPI core that invokes the button ACPI bus driver if
>> > it was loaded, and always returns false otherwise. Would that work for
>> > you?
>>
>> BTW not only nouveau invokes acpi_lid_open(), i915 does it as well.
>
> Clearly, the design is somewhat ad-hoc here.
>
> It looks like using module_acpi_driver() in button.c is a mistake given the
> dependencies.  The module initialization should ignore the
> acpi_bus_register_driver() failure in there, but there's no reason for the
> other ACPI driver modules to be affected by that.
>
> And if you change that, please add a comment referring to the dependencies
> in question.
>

OK, will do.