Re: [PATCH 11/18] platform/x86: int3472: fix object shared between several modules

From: Hans de Goede
Date: Mon Nov 21 2022 - 03:13:57 EST


Hi,

On 11/21/22 00:45, Masahiro Yamada wrote:
> On Mon, Nov 21, 2022 at 5:55 AM Hans de Goede <hdegoede@xxxxxxxxxx> wrote:
>>
>> Hi,
>>
>> On 11/20/22 14:55, Andy Shevchenko wrote:
>>> On Sat, Nov 19, 2022 at 11:08:17PM +0000, Alexander Lobakin wrote:
>>>> common.o is linked to both intel_skl_int3472_{discrete,tps68470}:
>>>>
>>>>> scripts/Makefile.build:252: ./drivers/platform/x86/intel/int3472/Makefile:
>>>>> common.o is added to multiple modules: intel_skl_int3472_discrete
>>>>> intel_skl_int3472_tps68470
>>>>
>>>> Although both drivers share one Kconfig option
>>>> (CONFIG_INTEL_SKL_INT3472), it's better to not link one object file
>>>> into several modules (and/or vmlinux).
>>>> Under certain circumstances, such can lead to the situation fixed by
>>>> commit 637a642f5ca5 ("zstd: Fixing mixed module-builtin objects").
>>>>
>>>> Introduce the new module, intel_skl_int3472_common, to provide the
>>>> functions from common.o to both discrete and tps68470 drivers. This
>>>> adds only 3 exports and doesn't provide any changes to the actual
>>>> code.
>>
>> Replying to Andy's reply here since I never saw the original submission
>> which was not Cc-ed to platform-driver-x86@xxxxxxxxxxxxxxx .
>>
>> As you mention already in the commit msg, the issue from:
>>
>> commit 637a642f5ca5 ("zstd: Fixing mixed module-builtin objects")
>>
>> is not an issue here since both modules sharing the .o file are
>> behind the same Kconfig option.
>>
>> So there is not really an issue here and common.o is tiny, so
>> small chances are it does not ever increase the .ko size
>> when looking a the .ko size rounded up to a multiple of
>> the filesystem size.
>>
>> At the same time adding an extra module does come with significant
>> costs, it will eat up at least 1 possibly more then 1 fs blocks
>> (I don't know what the module header size overhead is).
>>
>> And it needs to be loaded separately and module loading is slow;
>> and it will grow the /lib/modules/<kver>/modules.* sizes.
>>
>> So nack from me for this patch, since I really don't see
>> it adding any value.
>
>
>
>
> This does have a value.
>
> This clarifies the ownership of the common.o,
> in other words, makes KBUILD_MODNAME deterministic.
>
>
> If an object belongs to a module,
> KBUILD_MODNAME is defined as the module name.
>
> If an object is always built-in,
> KBUILD_MODNAME is defined as the basename of the object.
>
>
>
> Here is a question:
> if common.o is shared by two modules intel_skl_int3472_discrete
> and intel_skl_int3472_tps68470, what should KBUILD_MODNAME be?
>
>
> I see some patch submissions relying on the assumption that
> KBUILD_MODNAME is unique.
> We cannot determine KBUILD_MODNAME correctly if an object is shared
> by multiple modules.
>
>
>
>
>
>
> BTW, this patch is not the way I suggested.
> The Suggested-by should not have been there
> (or at least Reported-by)
>
>
> You argued "common.o is tiny", so I would vote for
> making them inline functions, like
>
>
> https://lore.kernel.org/linux-kbuild/20221119225650.1044591-2-alobakin@xxxxx/T/#u

Yes just moving the contents of common.c to static inline helpers in common.h
would be much better.

If someone creates such a patch, please do not forget to Cc
platform-driver-x86@xxxxxxxxxxxxxxx

Regards,

Hans



>
>
>
>
>
>
>
>
>> Regards,
>>
>> Hans
>>
>>
>>
>>
>>
>>>
>>> ...
>>>
>>>> +MODULE_IMPORT_NS(INTEL_SKL_INT3472);
>>>> +
>>>
>>> Redundant blank line. You may put it to be last MODULE_*() in the file, if you
>>> think it would be more visible.
>>>
>>>> MODULE_DESCRIPTION("Intel SkyLake INT3472 ACPI Discrete Device Driver");
>>>> MODULE_AUTHOR("Daniel Scally <djrscally@xxxxxxxxx>");
>>>> MODULE_LICENSE("GPL v2");
>>>
>>> ...
>>>
>>>> +MODULE_IMPORT_NS(INTEL_SKL_INT3472);
>>>> +
>>>> MODULE_DESCRIPTION("Intel SkyLake INT3472 ACPI TPS68470 Device Driver");
>>>> MODULE_AUTHOR("Daniel Scally <djrscally@xxxxxxxxx>");
>>>> MODULE_LICENSE("GPL v2");
>>>
>>> Ditto. And the same to all your patches.
>>>
>>
>
>