[PATCH v9 12/12] ARM: EXYNOS: refactor of mach-exynos to use chipid information

[pankaj.dubey]

On Friday 31 March 2017 01:52 PM, Arnd Bergmann wrote:
> On Fri, Mar 31, 2017 at 8:01 AM, pankaj.dubey <pankaj.dubey at samsung.com> wrote:
>> On Thursday 30 March 2017 07:31 PM, Arnd Bergmann wrote:
>>> On Thu, Mar 30, 2017 at 3:17 PM, Pankaj Dubey <pankaj.dubey at samsung.com> wrote:
>>>
>>> This seems really overcomplicated. From what I can tell, this is only needed
>>> to find the SCU address, but there are better ways to do that, either by
>>> looking up the SCU in the DT at the time you actually need it,
>>> or by calling scu_a9_get_base().
>>
>> Indeed it's complicated and ugly, for the same reason I attempted to
>> clean-up this part for Exynos SoC in patch series [1], which was
>> accepted by Samsung maintainer Krzysztof and later during pull request
>> rejected by ARM maintainers [2]
>>
>> [1]: https://www.spinics.net/lists/arm-kernel/msg540498.html
>> [2]: http://lkml.iu.edu/hypermail/linux/kernel/1612.0/01508.html
>>
>> Given the argument we can't adopt getting base address of SCU completely
>> via DT without breaking OLD dtbs at some point, difficult to simplify
>> these stuffs.
>>
>> Please guide which is the best way to handle such stuffs.
>>
>> In fact I made another attempt [3] to cleanup this SCU related stuffs
>> for other ARM based platforms, but we could not arrive on some consensus
>> at that time. I will try to attempt it again soon.
>>
>> [3]: http://www.spinics.net/lists/arm-kernel/msg542351.html
> 
> I very much appreciate your attempts, this is very helpful. I'm sorry
> for making it so hard for you. I had remembered your previous series,
> but did not realize that you were the author of both this patch and
> the SCU cleanup.
> 
> Let me try to understand the entire problem space better. These
> are my assumptions based on what  you write:
> 
> - A clean solution would be to map the SCU based on the DT as
>   you implemented in the earlier patch series, but that breaks
>   running new kernels with old DTBs, which we don't want.
> 
> - Another clean solution would be to map the SCU based on
>   scu_a9_get_base(), but that doesn't work on all variants of
>   exynos4, because there is at least one (which?) that has incorrect
>   data in the register

Sorry, I am also not aware which variant does not support this.
Currently for none of exynos4 we are using scu_a9_get_base(), and due to
unavailability of all exynos4 based board with me, I can't test it.

> 
> - The approach in this patch works, but is really ugly as it
>   requires parsing the fdt.
> 
> Assuming that we don't want any of the approaches above,
> here are some other ideas that might be a little better:
> 
> - Map the SCU later, just before we first need for calling
>   scu_enable(): exynos_smp_prepare_cpus(),
>   exynos_enter_aftr() or exynos_pm_resume(). All three
>   are late enough that we can simply call ioremap()
>   on the known physical address for the first caller,
>   and they are all guarded by a
>   ARM_CPU_PART_CORTEX_A9 check that should
>   be equivalent to the EXYNOS4 check you have.
> 

OK, let me try this, I think I can shift it at a little later point of boot.

> - Split the exynos4 machine descriptor from the one
>   for exynos5, and exynos3, the call the iotable_init
>   unconditionally for exynos4. The same method would
>   also let us get rid of some of the other
>   of_machine_is_compatible() checks.
> 

Once we move SCU mapping to later point of time, iotable_init can be
dropped. Currently only for SCU mapping it is getting called. So this
part will become more clean.


Thanks for suggestion. I will work on it and submit next version.

Pankaj Dubey

>        Arnd
> 
> 
>