Re: [PATCH v4 1/2] ARM: keystone: pm: switch to use generic pm domains

From: Grygorii Strashko
Date: Wed Nov 19 2014 - 06:33:38 EST


Hi Arnd,

On 11/18/2014 09:32 PM, Arnd Bergmann wrote:
> On Tuesday 18 November 2014 20:54:36 Grygorii Strashko wrote:
>> Hi All,
>>
>> Thank you for your comments.
>>
>> On 11/17/2014 11:50 PM, Kevin Hilman wrote:
>>> Arnd Bergmann <arnd@xxxxxxxx> writes:
>>>
>>>> On Monday 17 November 2014 11:14:16 Kevin Hilman wrote:
>>>>>>>
>>>>>>> So, The Keystone 2 Generic PM Controller is just a proxy PM layer here between
>>>>>>> device and Generic clock manipulation PM callbacks.
>>>>>>> It fills per-device clock list when device is attached to GPD and
>>>>>>> ensures that all clocks from that list enabled/disabled when device is
>>>>>>> started/stopped.
>>>>>>
>>>>>> The idea of such a generic power domain implementation sounds useful, but
>>>>>> it has absolutely no business in platform specific code.
>>>>>
>>>>> Yes it does. This isn't a generic power domain implementation, but
>>>>> rather just the platform-specific glue that hooks up the clocks to the
>>>>> right devices and power-domains so that the generic power-domain and
>>>>> generic pm_clocks code does the right thing.
>>>>
>>>> How would you do this on an arm64 version of keystone then? With
>>>> the current approach, you'd need to add a machine specific directory,
>>>> and that seems completely pointless since this is not even about
>>>> a hardware requirement.
>>>
>>> Yeah, you're right. I misunderstood you're original comment.
>>>
>>>>>> I suggest you either remove the power domain proxy from your drivers
>>>>>> and use the clocks directly,
>>
>> Hm. I've been thinking about this, but the problem is that Keystone 2
>> reuses a lot of IPs from Davinci and PM for Davinci is based on Generic clock
>> manipulation PM callbacks framework, but for non-DT case. So, I can't simply
>> use clocks directly.
>
> I think you could get that to work without too much trouble, but as Kevin
> comments, the generic pmdomain code is helpful here, and we should find
> a way to make it work better.
>
>>>>> No. That's a step in the wrong direction. This change isn't affecting
>>>>> drivers directly. It's the runtime PM and generic power domain layers
>>>>> that handle this, and runtime PM adapted drivers don't need any changes.
>>>>>
>>>>>> or come up with an implementation that can be used across other
>>>>>> platforms and CPU architectures.
>>>>>
>>>>> We already have those in the generic power domain and the pm_clock
>>>>> layers. This series is just hooking those up for Keystone.
>>>>
>>>> Then why not add the missing piece to the generic power domain
>>>> code to avoid having to add infrastructure to the platform
>>>> for it?
>>>
>>> Yes, good point. There is nothing keystone-specific in this glue.
>>>
>>> Grygorii, what about adding a feature to the generic domain parsing so
>>> that it can get clocks from device nodes that are part of the domain,
>>> and so it sets up pm_clk accordingly.
>>
>> I'd like to mention few points here:
>> 1) not all platforms may need this
>>
>> 2) not all platforms may allow to add ALL clocks from "clocks" property
>> to pm_clk as some of them can be optional or have to be controlled by drivers only
>> (for example, initially, it was the case for SH-mobile https://lkml.org/lkml/2014/4/24/197
>> also now, last implementation for shmobile add only first clock from "clocks" property
>> to pm_clk https://lkml.org/lkml/2014/11/17/272).
>>
>> 3) such functionality have to be enabled for devices selectively, for example
>> now we are going to enable it for devices which a ready for runtime PM.
>>
>> Current implementation cover 1 & 3, but also it allows to cover 2 too, because
>> it's platform specific implementation and .attach_dev() can be updated to skip some
>> clocks or devices if needed.
>
> Well, not all drivers and not all platforms have to use it, I think it would
> just be good to make the case you are interested in really easy, and definitely
> work without platform specific code.
>
>>> I've recently seen other SoCs doing very similar, so this really should
>>> be generalized.
>>>
>>> I've been looking at this primarily as a right incremental improvement
>>> from what is there for Keystone today, but Arnd is right. This should
>>> be moved out of platform code.
>>
>> I'm ready to do what ever you want, but I don't fully understand what exactly to do :(
>> Should I create some generic_pm_clk_domain.c?
>> - or - Do you mean to integrate it in domain.c (see no way to do it:()?
>> - or - smth. else
>>
>> What about introduced DT bindings? For example, How will devices be
>> selected for attachment to Generic pm_clk domain if I'll introduce
>> generic_pm_clk_domain.c?
>
> I am not really familiar with the pmdomain code at all, but here is
> what I had thought the simplest generic pmdomain code would do:
>
> Have one pmdomain driver in the generic code that knows about clocks,
> possibly also regulators and pins and just turns them on when needed.
> You can have a "simple-pmdomain" or "generic-pmdomain" compatible
> string.
>
> I'm a bit surprised that your pmdomain code looks up the clocks from the
> respective device, rather than know about the clocks itself. There is
> probably a good reason for this, but I don't see it yet.

The keystone 2 uses simple PM schema based on clocks only:
- clocks enabled -> dev is active
- clocks disabled -> dev is suspended

To achieve explained above the Generic clock manipulation PM callbacks framework (pm_clk) is used.
- list of managed clocks is filled for each device (for non-DT case the con_id list
is specified by platform code like:
.con_ids = { "fck", "master", "slave", NULL },
- or -
.con_ids = { }, <-- in this case only first clock will be added to pm_clk
)
- dev.pm_domain is assigned to pm_clk_domain
- pm_clk_domain has .runtime_resume/runtime_suspend callbacks set and implemented like:
static int keystone_pm_runtime_suspend(struct device *dev)
{
ret = pm_generic_runtime_suspend(dev);
if (ret)
return ret;

ret = pm_clk_suspend(dev);
if (ret) {
pm_generic_runtime_resume(dev);
return ret;
}

return 0;
}
static int keystone_pm_runtime_resume(struct device *dev)
{
pm_clk_resume(dev);

return pm_generic_runtime_resume(dev);
}
Now this configuration can be done from Platform Bus notifier (clock_ops.c) only and It'll
be done for ALL Platform devices and, once above steps have been done, the Runtime PM
starts working for device.

As was described early in this thread, Keystone 2 PM domain is glue layer which
allows:
- configure pm_clk for devices from DT and get rid of .con_ids
- configure only selected devices
- get rid of using Platform Bus notifier
- enable suspend/resume for devices as side effect :)
and which manages state of its devices through dev_ops->start()/stop() callbacks.

>
> Another option would be to have a special case for an empty
> "power-domains" property if this is the most common case: if that
> property exists but is empty, the pmdomain core could interpret
> it as an indication to control all the clocks/regulators/pins
> of a device.

I can try to do the following:
- move Keystone PM domain code in drivers/base/power/simple_domain.c
- add DT bindings like:
+ clk_pmdomain: simple-clk-pmdomain {
+ compatible = "simple-clk-pmdomain", "simple-pmdomain";
+ #power-domain-cells = <0>;
+ };
- in case if compatible == "simple-clk-pmdomain" it will do the same sings as in this patch

In the future, additional support for regulators/clocks/gpios can be added and these resources
can be managed in .power_on/power_off.

Is it ok for you?

Regards,
- grygorii


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/