On 04/10/17 12:30, Arnd Bergmann wrote:

quoted

+static int scmi_cpufreq_probe(struct platform_device *pdev)
+{
+       int ret;
+
+       handle = devm_scmi_handle_get(&pdev->dev);
+
+       if (IS_ERR_OR_NULL(handle) || !handle->perf_ops)
+               return -EPROBE_DEFER;

As mentioned before, never create an interface that needs to use
IS_ERR_OR_NULL(), make it return either NULL on error, or always
have an error code.

Agreed.

quoted

+
+static struct platform_driver scmi_cpufreq_platdrv = {
+       .driver = {
+               .name   = "scmi-cpufreq",
+       },
+       .probe          = scmi_cpufreq_probe,
+       .remove         = scmi_cpufreq_remove,
+};

You appear to have split this driver into the 'cpufreq' side and
the 'protocol' handler in a different file. I already commented that
the way the main scmi driver knows about all the protocols looks
bad, here we can see another aspect of the same problem.

Rather than manually register a platform_device for the purpose
of connecting it to the cpufreq driver, there should be a way
to dynamically register the protocol from the cpufreq driver
and then have both in the same file.

I agree that should be possible. I took this approach for 2 reasons:

1. to avoid all sorts of probe ordering issues
2. we may have system with multiple instances of SCMI. E.g. a system
   may have multiple remote processors, each controlling dvfs/power mgmt
   of a subset of CPUs/devices controller in OS.

I have to admit that I haven't thought too much in details yet. That's
the main idea behind scmi_handle and restricting access to that only to
sub-nodes in it. I am open to suggestions.

-- 
Regards,
Sudeep