[RFC/RFT 2/6] clk: qcom: Add runtime support to handle clocks using PM clocks

[Ulf Hansson]

On 30 April 2015 at 08:21, Ulf Hansson <ulf.hansson at linaro.org> wrote:
> On 29 April 2015 at 15:08, Geert Uytterhoeven <geert at linux-m68k.org> wrote:
>> Hi Ulf,
>>
>> On Wed, Apr 29, 2015 at 2:31 PM, Ulf Hansson <ulf.hansson at linaro.org> wrote:
>>>>> Geert, thanks, I was wondering how you handle the !CONFIG_PM case for
>>>>> rmobile. I mean who turns the clocks on for the devices when you build
>>>>> with CONFIG_PM disabled?
>>>>
>>>> We still use pm_clk_add_notifier() in drivers/sh/pm_runtime.c if
>>>> CONFIG_PM_GENERIC_DOMAINS_OF=n. Hence the second instance of
>>>> pm_clk_notify() will enable the clocks at driver binding time.
>>>> Hardware power domains are assumed enabled by reset state/the boot
>>>> loader.
>>>
>>> Yes, it a bit tricky - but I guess that's the current only viable
>>> solution if we have when using the pm_clk API.
>>>
>>>> Given the amount of PM infrastructure involved when hardware power and
>>>> clock domains are involved, I think at one point we'll have to start restricting
>>>> our builds to CONFIG_PM=y.
>>>
>>> I don't think that would solve the problem, since you may still have
>>> cross SoC drivers which may be used without CONFIG_PM.
>>
>> That's not as much of a problem as it sounds:
>>   - If the driver cares (needs to know) about the clock, it should already
>>     manage it itself.
>
> Agree!
>
>>   - If it doesn't care about the clock, it just needs Runtime PM support.
>>     That will do the right thing on platforms with (needs PM=y) and without
>>     (doesn't care about PM=x) clock domains.
>
> How about those drivers that cares about clocks and thus manages them,
> but also cares about runtime PM?
>
> I believe we will get a clock reference count issue in these cases,
> since both the PM domain and the driver will manage the clocks.
>
> I assume that's why we have had a few attempts to have "runtime PM
> clocks" specially marked, one was via DT, to have clear distinguish
> between who will be responsible to manage them.
>
> Those attempts did get nacked, so the problem is still there I assume.
>
>> So the bigger "problem" is making sure all drivers have at least minimal
>> Runtime PM support, else they can't be reused as-is on systems with PM
>> domains.
>>
>>> So all SoC that uses a driver which expects clocks to be managed using
>>> PM clocks from a PM domain, will need the above "trick". Right?
>>
>> One remaining issue with pm_clk_add_notifier() is that it applies to all
>> platform devices in the system, not just the on-SoC devices. Hence it may
>> inadvertently manage clocks for off-SoC devices it's not supposed to touch.
>
> Yes. That's really bad. :-)
>
> Additionally, it means devices that isn't part of the platform bus
> isn't able to use PM clk domains at all.

Correction: Of course they can register one PM clk notifier per bus
type. The API currently also provides this option.

>
> Within this context, I find it hard to advise people to use PM clk
> domains (via pm_clk_add_notifier()), since there just so many open
> issues. What works a _little_ better is to use PM clks via genpd.
>
> Kind regards
> Uffe