Re: [PATCH 6/9] clk: Allow the common clk framework to be selectable

[Greg Ungerer <gerg@linux-m68k.org>]

Hi Arnd, Stephen

On 6/4/20 5:35 pm, Arnd Bergmann wrote:
> On Mon, Apr 6, 2020 at 5:01 AM Stephen Boyd <sboyd@kernel.org> wrote:
>> Quoting Arnd Bergmann (2020-04-05 05:45:20)
>>> On Sun, Apr 5, 2020 at 4:51 AM Stephen Boyd <sboyd@kernel.org> wrote:
>>>> There's one snag with doing this, and that's making sure that randconfig
>>>> builds don't select this option when some architecture or platform
>>>> implements 'struct clk' outside of the common clk framework. Introduce a
>>>> new config option 'HAVE_LEGACY_CLK' to indicate those platforms that
>>>> haven't migrated to the common clk framework and therefore shouldn't be
>>>> allowed to select this new config option. Also add a note that we hope
>>>> one day to remove this config entirely.
>>>
>>> Good idea!
>>>
>>> I've looked through the individual ones and commented a bit on
>>> what I think may or may not happen with them.
>>>
>>> ralink SOC_MT7621 is the only one that I think you got wrong,
>>> as it already has common-clk support.
>>
>> Ah I missed that it was inside a big if RALINK. Thanks. I suppose I
>> should just remove the select then for that config and not worry about
>> the duplication of clkdev and common clk configs.
> 
> Won't that cause build failures in those configurations that have
> both implementations?
> 
> According to the Makefile, the clk.c file is built whenever CONFIG_MIPS_GIC
> is unset, so I think we need
> 
>           select HAVE_LEGACY_CLK if !MIPS_GIC
> 
> or maybe move the select into the per-chip configs that need it:
> RT288X, RT305X, RT3883, and MT7620.
> 
>>>> diff --git a/arch/m68k/Kconfig.cpu b/arch/m68k/Kconfig.cpu
>>>> index 60ac1cd8b96f..bd2d29c22a10 100644
>>>> --- a/arch/m68k/Kconfig.cpu
>>>> +++ b/arch/m68k/Kconfig.cpu
>>>
>>>     text    data     bss     dec     hex filename
>>> 1934726 263616   83284 2281626 22d09a obj/vmlinux-before
>>> 1971989 266192   83308 2321489 236c51 obj/vmlinux-after
>>>
>>> The coldfire clock implementation looks rather simple compared
>>> to chips from the 2010s: most chips have only fixed clocks,
>>> and three of them have one of two registers of clock gates.
>>>
>>> It shouldn't be hard to convert, but enabling common-clk will
>>> cause a noticeable kernel size increase on the fairly limited
>>> hardware.
>>>
>>> Simply enabling COMMON_CLK in m5475evb_defconfig
>>> results in a 1.7% or 40KB growth in kernel size, plus there
>>> would be additional dynamic memory usage:
>> There could certainly be some work done to reduce the code size of the
>> CCF. I haven't looked but perhaps we could save some memory by making
>> the basic types selectable too and then push a bunch of kconfig updates
>> through for that.
> 
> Right, that might help. Another possibility would be to support both
> the common clk layer and the custom clk implementation on coldfire
> until we remove the other custom implementations, by which point
> even fewer people will care about coldfire.
> 
> Let's see what Geert and Greg think would be the best path for coldfire,
> maybe the added 40KB is less of a problem after all.

Losing another 40k is not ideal, but not the end of the world.
It would not stop me running it on any platforms I regularly
run on. For sure some of the really old hardware just doesn't
have the RAM to spare.

Any way, I say we have to move forward and and move to using
the common clock framework for ColdFire sooner than later.

Regards
Greg