[RFC V2] OPP: Redefine bindings to overcome shortcomings

[Rob Herring]

On Wed, Jan 28, 2015 at 7:39 PM, Viresh Kumar <viresh.kumar at linaro.org> wrote:
> On 29 January 2015 at 01:36, Mark Brown <broonie at kernel.org> wrote:
>> On Fri, Jan 23, 2015 at 04:14:50PM +0530, Viresh Kumar wrote:

[...]

>>> +  Required properties:
>>> +  - compatible: allow OPPs to express their compatibility.
>>> +  - opp-list: phandle to opp-list defined above.
>>
>> I don't understand what that compatible property is intended to mean and
>> I expect other readers might be similarly confused - is it a standard
>> compatbile property meaning this noe corresponds to some sort of device?
>
> The goal is to choose the driver which we want to probe for a platform. There
> can be multiple DT enabled cpufreq drivers present in a build and the platform
> needs some way to choose one of them.

That was not my original thought, but rather it is versioning the
binding and a way to find OPP nodes. That doesn't mean we can't do
things to make it easier to match to drivers.

> For example, if both cpufreq-dt, exynos-cpufreq and arm-big-little drivers are
> present in a kernel build, the how do we specify which one a platform wants to
> use.

What does the exynos driver do that the other 2 don't?

> The generic drivers (cpufreq-dt and big LITTLE) should match with some generic
> strings, as these aren't specific to any platform. But exynos one can be matched
> with the machine or Soc-family's existing compatible names.

I will likely nak any such strings that are solely there to match to
these 2 drivers. The information is there in DT already. It's just not
a single property to match on. You can tell it is a big.LITTLE system
by the fact your have 2 different cpu types. You can use the top level
compatible strings when you need a machine specific driver. I would
not be trying to get cpufreq platform devices to be created by
of_platform_populate. The other issue I have is you may not always
have 2 drivers for bL and non-bL. Really symmetric multi-cluster or
single cluster should be a subset of functionality of bL.

I think you just need to have an initcall which does this:

if (of_find_matching_node_and_match(machine_specific_driver_match list))
  platform_device_create(match->data);
else if (is_big_little() && has_opp)
  platform_device_create(big_little_device);
else if (has_opp)
  platform_device_create(cpufreq_dt_device);

Or perhaps you handle the generic ones late in boot only if no machine
specific driver has been probed. There is no generic solution ATM for
selecting the best driver when there are multiple matches.

Rob