Re: [PATCH] firmware: arm_scmi: fix i.MX build dependency

["Arnd Bergmann" <arnd@arndb.de>]

On Sun, Nov 17, 2024, at 11:03, Cristian Marussi wrote:
> On Sat, Nov 16, 2024 at 12:05:18AM +0100, Arnd Bergmann wrote:
>> 
>> arm-linux-gnueabi-ld: sound/soc/fsl/fsl_mqs.o: in function `fsl_mqs_sm_write':
>> fsl_mqs.c:(.text+0x1aa): undefined reference to `scmi_imx_misc_ctrl_set'
>> arm-linux-gnueabi-ld: sound/soc/fsl/fsl_mqs.o: in function `fsl_mqs_sm_read':
>> fsl_mqs.c:(.text+0x1ee): undefined reference to `scmi_imx_misc_ctrl_get'
>> 
>
> The SCMI drivers, like the newly added IMX_SCMI_MISC_DRV, generally make
> ue of the related vendor protocol like IMX_SCMI_MISC_EXT, BUT the SCMI
> stack is designed in a way that NO symbols are needed to be exported by
> the protocol layer (to avoid a huge and growing number of symbols
> exports)...so usually the current DRV-->PROTO dependency is fine.
>
> In this case, AFAIU, it is the SCMI driver that in turn exports a few
> helpers that are used by another driver fsl_mqs, which in turn could be
> compiled and work with or without the SCMI stack, so with this patch we
> are artificially reversing the DRV<--PROTO dependency to solve this
> scenario in all the compillation scenarios...
>
> ....BUT given that the IMX_SCMI_MISC_DRV is the one that should export
> the missing symbols could NOT this solved in a cleaner way, without
> adding the fake reverse dependency, by instead modifying the header of
> the driver with something like the classic:

> --->8-----
> diff --git a/include/linux/firmware/imx/sm.h b/include/linux/firmware/imx/sm.h
> index 9b85a3f028d1..3a7a3ec367c5 100644
> --- a/include/linux/firmware/imx/sm.h
> +++ b/include/linux/firmware/imx/sm.h
> @@ -17,7 +17,19 @@
>  #define SCMI_IMX_CTRL_SAI4_MCLK                4       /* WAKE SAI4 MCLK */
>  #define SCMI_IMX_CTRL_SAI5_MCLK                5       /* WAKE SAI5 MCLK */
> 
> +#ifdef IMX_SCMI_MISC_DRV
>  int scmi_imx_misc_ctrl_get(u32 id, u32 *num, u32 *val);
>  int scmi_imx_misc_ctrl_set(u32 id, u32 val);
> +#else
> +static inline int scmi_imx_misc_ctrl_get(u32 id, u32 *num, u32 *val)
> +{
> +       return 0;
> +}
> +
> +static inline int scmi_imx_misc_ctrl_set(u32 id, u32 val)
> +{
> +       return 0;
> +}
> +#endif

This usually doesn't work if the provider of these interfaces
can be in a loadable module. The #ifdef above means this won't
be usable when CONFIG_IMX_SCMI_MISC_DRV=m, while changing it to
IS_ENABLED(CONFIG_IMX_SCMI_MISC_DRV) still produces a link error
when the consumer is built-in. Changing it to IS_REACHABLE()
in turn is even worse because it avoids the link failure but
makes it silently do the wrong thing in some configurations.

>  #endif
> ----->8-----------
>
> ....to just support compilation in all the scenarios.
>
>> This however only works after changing the dependency in the SND_SOC_FSL_MQS
>> driver as well, which uses 'select IMX_SCMI_MISC_DRV' to turn on a
>> driver it depends on. This is generally a bad idea, so the best solution
>> is to change that into a dependency.
>> 
>> To allow the ASoC driver to keep building with the SCMI support, this
>> needs to be an optional dependency that enforces the link-time
>> dependency if IMX_SCMI_MISC_DRV is a loadable module but not
>> depend on it if that is disabled.
>> 
>
> ...and maybe with the above additions you could avoid also these other
> dep changes...
>
> ...not sure if I am missing something and I have definitely not tested
> any of my babbling above...

In my experience, there is no way to avoid reflecting the dependencies
correctly in Kconfig: if one driver has an EXPORT_SYMBOL that
gets picked up by another driver, you need a matching 'depends on'.

       Arnd