Re: [PATCH v6 03/18] clk: qcom: gcc-ipq806x: add PXO_SRC in clk table

[Dmitry Baryshkov <dmitry.baryshkov@linaro.org>]

On 13/04/2022 20:54, Ansuel Smith wrote:
> On Wed, Apr 13, 2022 at 08:32:21PM +0300, Dmitry Baryshkov wrote:
>> On Wed, 13 Apr 2022 at 20:00, Ansuel Smith <ansuelsmth@gmail.com> wrote:
>>>
>>> On Thu, Mar 24, 2022 at 06:22:29PM -0700, Stephen Boyd wrote:
>>>> Quoting Ansuel Smith (2022-03-24 18:13:49)
>>>>> On Thu, Mar 24, 2022 at 06:10:35PM -0700, Stephen Boyd wrote:
>>>>>> Quoting Ansuel Smith (2022-03-21 16:15:33)
>>>>>>> PXO_SRC is currently defined in the gcc include and referenced in the
>>>>>>> ipq8064 DTSI. Correctly provide a clk after gcc probe to fix kernel
>>>>>>> panic if a driver starts to actually use it.
>>>>>>>
>>>>>>> Signed-off-by: Ansuel Smith <ansuelsmth@gmail.com>
>>>>>>> ---
>>>>>>
>>>>>> What is this patch about? clk providers shouldn't be calling clk_get().
>>>>>>
>>>>>
>>>>> If pxo is passed as a clock in dts and defined as a fixed clock, what
>>>>> should be used?
>>>>
>>>> clk_parent_data
>>>
>>> Sorry but I'm not following you. No idea if you missed the cover letter
>>> where i describe the problem with PXO_SRC.
>>>
>>> The problem here is that
>>> - In DTS we have node that reference <&gcc PXO_SRC>
>>> But
>>> - gcc driver NEVER defined PXO_SRC
>>> As
>>> - PXO_SRC is actually pxo_board that should be defined as a fixed-clock
>>>    in dts or is defined using qcom_cc_register_board_clk.
>>>
>>> So in theory we should just put in PXO_SRC the clk hw of the
>>> fixed-clock. That is why I'm using clk_get(). I can use __clk_lookup()
>>> as an alternative but I really can't find a way to get the clock defined
>>> from DTS or qcom_cc_register_board_clk.
>>>
>>> (I have the same exact problem with the cpu qsb clock where is defined
>>> using fixed-clock API but can also defined directly in DTS and I have to
>>> use clk_get())
>>>
>>> I'm totally missing something so I would love some hint on how to solve
>>> this.
>>
>> When we were doing such conversion for other  platforms, we pointed
>> clock consumers to the board clocks directly. There is no need to go
>> through the gcc to fetch pxo.
>> Instead you can use a <&pxo_board> in the dts directly. Typically the
>> sequence is the following:
>> - Minor cleanup of the clock-controller driver
>> (ARRAY_SIZE(parent_data), removal of unused clock sources, unused enum
>> entries, etc)
>> - update drivers to use both .name and  .fw_name in replacement of
>> parent_names. Use parent_hws where possible.
>> - update dtsi to reference clocks using clocks/clock-names properties.
>> Pass board/rpmh/rpm clocks directly to their consumers without
>> bandaids in the gcc driver.
>> - (optionally) after several major releases drop parent_data.name
>> completely. I think we mostly skipped this, since it provides no gain.
>>
>> This way you don't have to play around clk_get to return PXO_SRC from
>> gcc clock-controller.
>>
>> -- 
>> With best wishes
>> Dmitry
> 
> Thanks for the list of steps to do this kind of cleanup.
>  From what I'm reading this series is ""stuck"" in the sense that I first
> have to fix the wrong PXO_SRC reference and then I can continue the
> conversion work. A bit sad considering most of the time DTS proposal got
> ignored :(

Not really. You can leave "pxo" as is. Use { .fw_name = "pxo", .name = 
"pxo_board" } as parent_data. Then pass <&pxo_board> as the "pxo" clock 
to the consumers. Yes, you will still have the lingering "pxo" / "cxo" 
clocks at this step. It's okay, they might be used by other drivers.

After the whole conversion is finished, you can make "pxo"/"cxo" 
registration conditional on !of_find_property("clocks") rather than 
using clk_get.

As a rule of thumb, you don't have to complete the whole thing in a 
single commit. Having smaller commits might be better.

[And yes, I'm looking forward to testing your cpufreq changes on my 
apq8064 devices].

-- 
With best wishes
Dmitry