Re: [PATCH v3 03/15] soc: tegra: Add Tegra PMC clock registrations into PMC driver

[Dmitry Osipenko]

16.12.2019 18:24, Peter De Schrijver ÐÐÑÐÑ:
> On Mon, Dec 16, 2019 at 05:11:32PM +0200, Peter De Schrijver wrote:
>> External email: Use caution opening links or attachments
>>
>>
>> On Mon, Dec 16, 2019 at 05:23:23PM +0300, Dmitry Osipenko wrote:
>>>>> Could you please clarify what do you mean by the "existing users"?
>>>>> AFAIK, nothing in kernel uses mux clocks.
>>>>
>>>> The DT clk bindings allow for parent initialization, so it's certainly
>>>> possible there are some DTs which rely on this. We promised to never
>>>> break the bindings, which changing to 1 clock would do.
>>>
>>> What about this variant:
>>>
>>>   1. Keep the old CaR code in place.
>>>
>>>   2. Make CaR driver to scan whole device-tree for the legacy PMC clocks.
>>>
>>>   3. If legacy clock is found, then register PMC clocks from CaR.
>>>
>>>   4. If legacy clocks are not found, then don't register PMC clocks from
>>> CaR.
>>>
>>>   5. Add clocks support to the PMC driver and only register them if
>>> legacy clocks are not registered by CaR.
>>>
>>> Now both old and new DTBs can co-exist and work, everyone happy.
>>
>> That seems even more work.. Today the only upstream user is audio.
>> Currently these clocks are setup by the CAR clock driver. However
>> as they will move to the PMC driver, this mechanism cannot be used.
>> Hence the plan is to modify the audio driver to check for the PMC clocks
>> in DT and if they are not available use the CAR clocks as fallback.
>> The whole reason the clocks move to the PMC driver, is that when PMC
>> becomes secure, all accesses have to go via an SMC. Given that we don't
>> want SMCs all over the Tegra drivers, it's a good opportunity to move
>> the PMC clock handling into the PMC driver. PMC can be secure with both
>> 'new' and old DTs, so just registering the PMC clocks in the CAR driver
>> if legacy clocks are found in the DT, won't always work.
>>
> 
> Given the audio driver needs to change anyway, we can indeed use 1 clock
> per PMC clk_out_ rather than 2 as we have today. As this models the hw
> slightly better, I think we should do that as you suggested.
> 
> Peter.
> 

Okay, let's try and see how it will go.