Re: [RFC PATCH v2 0/2] clk: Support for DT assigned clock parents and rates

[Sylwester Nawrocki <s.nawrocki@samsung.com>]

On 21/03/14 12:20, Maxime Coquelin wrote:
> On 03/20/2014 01:42 PM, Sylwester Nawrocki wrote:
>> On 06/03/14 14:45, Maxime Coquelin wrote:
>>> Hi Sylwester,
>>>
>>> 	I like the principle of your implementation, but I have two questions:
>>> 	1 - How can we manage PM with this solution, as the parent/rate will be
>>> set only once at probe time?
>>> 	2 - How to set the parent of a parent clock (which can be shared with
>>> other devices)? Same question about the parent rates.
>>
>> Thanks for your feedback and apologies for late reply.
>
> No problem! Apologies accepted ;)

:)
>> IIUC your first concern is about a situation when clocks need to be
>> reconfigured upon each resume from system sleep or runtime PM resume ?
> 
> Yes. This is the case of the STi SoCs.
> When resuming from system sleep, the clocks-related registers are 
> restored at their boot state.

Couldn't this be handled at the clocks controller driver suspend/resume
operations ?

>> As I mentioned in v1 of the RFC I was considering having individual
>> drivers calling explicitly the clocks set up routine. Presumably this
>> would allow to resolve the power management related issue.
>
>  From a functional point of view, that would indeed resolve the PM 
> related issue.
> But I'm not sure that on a performance point of view, parsing the DT at 
> each driver's resume call is an efficient way.

Right, it might not indeed be a good idea to do the parsing repeatedly.

>> One example I'm aware the approach as in this RFC wouldn't work is
>> when a device in a SoC belongs to a power domain, which needs to be
>> first switched on before we can start setting up and the clocks'
>> configuration get lost after the power domain switch off.
>
> Yes, that's another case to handle.
> I don't know which platforms are in that case, but not STi SoCs for your 
> information.

OK, this could still hopefully be resolved in a clock controller driver,
making it aware it belongs to same power domain as the clock consumer
devices.

>> OTOH I suspect devices for which one-time clocks setup is sufficient
>> will be quite common. And for these there would need to be a single
>> call to the setup routine in probe() I guess, since it wouldn't be
>> possible to figure out just from the DT data when the actual call
>> should be made.
>>
>> For a global clocks configuration, I thought about specifying that
>> in the clocks controller node, and then to have the setup routine
>> called e.g. from of_clk_init(). I think that could work well enough,
>> together with the patch [1], adding clock dependencies handling.
>> But then the clock frequency set up function would need to be
>> modified to respect the clock parent relationships, similarly as
>> in patch series [2]. A just noticed [2] recently, after posting
>> this RFC (adding Tero at Cc).
>
> OK, I agree with the approach.
> There is still the PM issue remaining with these clocks, but I think 
> that is not related to your series as we already have the issue currently.

>> [1] http://www.spinics.net/lists/arm-kernel/msg310507.html
>> [2] http://www.spinics.net/lists/linux-omap/msg103069.html

--
Regards,
Sylwester