Re: [PATCH 2/3] clk: amlogic: drop clk_regmap tables

[Jerome Brunet <jbrunet@baylibre.com>]

On Fri 20 Dec 2024 at 16:12, Stephen Boyd <sboyd@kernel.org> wrote:

> Quoting Jerome Brunet (2024-12-20 09:17:43)
>> Remove the big clk_regmap tables that are used to populate the regmap
>> field of clk_regmap clocks at runtime. Instead of using tables, use devres
>> to allow the clocks to get the necessary regmap.
>> 
>> A simpler solution would have been to use dev_get_regmap() but this would
>> not work with syscon based controllers.
>
> Why not have two init functions, one that uses the syscon approach from
> the parent device?

That would duplicate all the ops and would not scale if anything else
comes along. It would also tie the controller quirks with
clock ops. I would like to keep to clock ops and controllers decoupled as
much as possible

> Is the typical path to not use a syscon anyway?
>

I sure hope there will be no new syscon based controller but, ATM, around
50% are syscon based in drivers/clk/meson. Those are here to stay and I
doubt we can do anything about it.

>> 
>> This rework save a bit memory and the result is less of a maintenance
>> burden.
>> 
>> Unfortunately meson8b is left out for now since it is an early clock
>> driver that does not have proper device support for now.
>
> We should add a clk_hw_get_of_node() function that returns
> hw->core->of_node. Then there can be a similar function that looks at
> the of_node of a clk registered with of_clk_hw_register() and tries to
> find the regmap either with
> syscon_node_to_regmap(clk_hw_get_of_node(hw)) or on the parent of the
> node for the clk.

That's the thing. It means encoding the controller quirk of how to get
regmap in the clock ops. I would be prefer to avoid that.

With what you are suggesting I could make an ops that
* Try dev_get_regmap() first
* Try the syscon/of_node way next

I can make this "trial an error" approach work but I think it is pretty
nasty and encode controller stuff inside the clock driver.

>
> TL;DR: Don't use devres.

Using it makes thing nice and tidy. clk_regmap does not care were regmap
comes from. It just picks it up where it has been prepared

That approach could be extended to support controller with multiple
regmaps, with a name that does not depend on regmap_config and is local
to the clock controller. This will be useful when the name if defined
somewhere else (syscon, auxiliary device, etc ...)

>
>> 
>> Signed-off-by: Jerome Brunet <jbrunet@baylibre.com>
>> ---
>>  drivers/clk/meson/a1-peripherals.c | 165 +-------------
>>  drivers/clk/meson/a1-pll.c         |  18 +-
>>  drivers/clk/meson/axg-aoclk.c      |  22 --
>>  drivers/clk/meson/axg-audio.c      | 435 +------------------------------------
>>  drivers/clk/meson/axg.c            | 131 -----------
>>  drivers/clk/meson/c3-peripherals.c | 212 +-----------------
>>  drivers/clk/meson/c3-pll.c         |  34 +--
>>  drivers/clk/meson/clk-cpu-dyndiv.c |   1 +
>>  drivers/clk/meson/clk-dualdiv.c    |   2 +
>>  drivers/clk/meson/clk-mpll.c       |   6 +
>>  drivers/clk/meson/clk-phase.c      |  11 +
>>  drivers/clk/meson/clk-pll.c        |   7 +
>>  drivers/clk/meson/clk-regmap.c     |  88 ++++++++
>>  drivers/clk/meson/clk-regmap.h     |   7 +
>>  drivers/clk/meson/g12a-aoclk.c     |  34 ---
>>  drivers/clk/meson/g12a.c           | 261 ----------------------
>>  drivers/clk/meson/gxbb-aoclk.c     |  19 --
>>  drivers/clk/meson/gxbb.c           | 393 ---------------------------------
>>  drivers/clk/meson/meson-aoclk.c    |   7 +-
>>  drivers/clk/meson/meson-aoclk.h    |   2 -
>>  drivers/clk/meson/meson-eeclk.c    |   6 +-
>>  drivers/clk/meson/meson-eeclk.h    |   2 -
>>  drivers/clk/meson/meson8-ddr.c     |  11 +-
>>  drivers/clk/meson/s4-peripherals.c | 231 +-------------------
>>  drivers/clk/meson/s4-pll.c         |  33 +--
>>  drivers/clk/meson/sclk-div.c       |   5 +
>>  drivers/clk/meson/vclk.c           |   2 +
>>  drivers/clk/meson/vid-pll-div.c    |   1 +
>>  28 files changed, 165 insertions(+), 1981 deletions(-)
>
> It would be more focused if you split the patch into two. One that
> installs the init clk_op and implements the logic to hook up the regmap
> and one that removes the arrays that are no longer used.

I should have done so indeed.


-- 
Jerome