Re: [PATCH] hwrng: nomadik: add ARM_AMBA dependency

[Arnd Bergmann]

On Mon, Oct 6, 2025, at 10:23, Geert Uytterhoeven wrote:
> On Tue, 29 Jul 2025 at 17:28, Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
>> From: Arnd Bergmann <arnd@xxxxxxxx>
>>
>> Compile-testing this driver is only possible when the AMBA bus driver is
>> available in the kernel:
>>
>> x86_64-linux-ld: drivers/char/hw_random/nomadik-rng.o: in function `nmk_rng_remove':
>> nomadik-rng.c:(.text+0x67): undefined reference to `amba_release_regions'
>> x86_64-linux-ld: drivers/char/hw_random/nomadik-rng.o: in function `nmk_rng_probe':
>> nomadik-rng.c:(.text+0xee): undefined reference to `amba_request_regions'
>> x86_64-linux-ld: nomadik-rng.c:(.text+0x18d): undefined reference to `amba_release_regions'
>>
>> The was previously implied by the 'depends on ARCH_NOMADIK', but needs to be
>> specified for the COMPILE_TEST case.
>>
>> Fixes: d5e93b3374e4 ("hwrng: Kconfig - Add helper dependency on COMPILE_TEST")
>> Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
>
> Thanks for your patch, which is now commit efaa2d815a0e4d1c ("hwrng:
> nomadik - add ARM_AMBA dependency") upstream.
>
>> --- a/drivers/char/hw_random/Kconfig
>> +++ b/drivers/char/hw_random/Kconfig
>> @@ -312,6 +312,7 @@ config HW_RANDOM_INGENIC_TRNG
>>  config HW_RANDOM_NOMADIK
>>         tristate "ST-Ericsson Nomadik Random Number Generator support"
>>         depends on ARCH_NOMADIK || COMPILE_TEST
>> +       depends on ARM_AMBA
>>         default HW_RANDOM
>>         help
>>           This driver provides kernel-side support for the Random Number
>
> After seeing CONFIG_HW_RANDOM_NOMADIK disappear from m68k
> all{mod,yes}config, I became intrigued, as it did build fine before?
> If CONFIG_ARM_AMBA is not enabled, both __amba_driver_register() and
> amba_driver_unregister() become static inline dummies, and the rest
> of the code and data is not referenced, thus optimized away by the
> compiler.  I verified this is the case on amd64 allmodconfig, too.
> How come this failed for you?

I believe this was the result of new build options getting adding
to the randconfig test matrix. If for any reason the
__amba_driver_register() function does not get inlined, the caller
will still have a permanent reference to the amba_driver structure,
and the rest of the file does not get removed by dead-code elimination.

Most likely this change happens because the __init attribute on
the module_init function puts it into a separate section from
the __amba_driver_register() function, causing the compiler to
refuse inlining them. Marking __amba_driver_register() as
__always_inline, or changing the amba_driver_register() macro
to not pass the pointer should also work here.

     Arnd