On Thu, Aug 21, 2025, at 11:56, Peng Fan wrote:

On Wed, Aug 20, 2025 at 03:55:20PM +0200, Arnd Bergmann wrote:

quoted

On Thu, Aug 7, 2025, at 03:47, Peng Fan wrote:


When a caller of this function is in a built-in driver but the
IMX_SCMI_MISC_DRV code is in a loadable module, you still
get a link failure, see 514b2262ade4 ("firmware: arm_scmi:
Fix i.MX build dependency") for an example.

As you still need the correct Kconfig dependencies, I
think your patch here is not helpful.

The consumer driver still needs Kconfig dependcies, such as
  depends on IMX_SCMI_MISC_DRV || !IMX_SCMI_MISC_DRV

So when IMX_SCMI_MISC_DRV is module built, the consumer driver will
also be module built.

But if IMX_SCMI_MISC_DRV is n, the consumer driver is y, there will be
link error.

The consumer driver is to support platform A and platform B.

Platform A does not require the real API in IMX_SCMI_MISC_DRV.
Platform B requires the real API in IMX_SCMI_MISC_DRV.

So when producing an image for platform A, IMX_SCMI_MISC_DRV could set
to n to make Image smaller. Introducing the stub API is mainly for this
case.

Hope this is clear

I see. In this case the stub helpers are not wrong, but I
still find them more error-prone than not having them and
using IS_ENABLED() checks as in commit 101c9023594a
("ASoC: fsl_mqs: Support accessing registers by scmi interface"):

+static int fsl_mqs_sm_read(void *context, unsigned int reg, unsigned int *val)
+{
+       struct fsl_mqs *mqs_priv = context;
+       int num = 1;
+
+       if (IS_ENABLED(CONFIG_IMX_SCMI_MISC_DRV) &&
+           mqs_priv->soc->ctrl_off == reg)
+               return scmi_imx_misc_ctrl_get(SCMI_IMX_CTRL_MQS1_SETTINGS, &num, val);
+
+       return -EINVAL;
+};

The logic is the same here in the end, but the link failure
is easier to trigger and repair if someone gets it wrong.

Also, for drivers that actually need the exported interface,
the dependency becomes the simpler 'depends on IMX_SCMI_MISC_DRV'.

Which driver using this symbol are you actually looking
at? I see you have three similar patches for a couple of
interfaces, and want to make sure the added complexity is
really needed here. I do a lot of randconfig build tests,
so quite often I end up being the one that runs into
the subtle link failures from these.

      Arnd