[Question] vendor-specific cpu enable-method

[Question] vendor-specific cpu enable-method

[Masahiro Yamada]

Hello.


Sorry if I am asking a stupid question.


For arm64, there are only 2 cpu methods, psci and spin-table.

Why do we still allow vendor-specific methods upstreamed
for arm 32bit ports?

To me, it looks like SoC vendors continue inventing
different (but similar) ways to do the same thing.

It is a historical reason for old platforms.

However, if I look at Documentation/devicetree/bindings/arm/cpus.txt
enable-method properties are still increasing.


psci is available in arch/arm/kernel/psci_smp.c,
but not all SoCs support the security extension.
Is there a simpler one like spin-table available for arm32?

If we force generic methods like psci or spin-table
for new platforms, we can stop proliferated smp code.
(Of course, we are just shifting the complexity
from the kernel to firmware.)

Am I missing something?


-- 
Best Regards
Masahiro Yamada

Re: [Question] vendor-specific cpu enable-method

[Jisheng Zhang]

On Thu, 13 Sep 2018 10:23:35 +0900 Masahiro Yamada wrote:

> Hello.
> 
> 
> Sorry if I am asking a stupid question.
> 
> 
> For arm64, there are only 2 cpu methods, psci and spin-table.
> 
> Why do we still allow vendor-specific methods upstreamed
> for arm 32bit ports?
> 
> To me, it looks like SoC vendors continue inventing
> different (but similar) ways to do the same thing.
> 
> It is a historical reason for old platforms.
> 
> However, if I look at Documentation/devicetree/bindings/arm/cpus.txt
> enable-method properties are still increasing.
> 
> 
> psci is available in arch/arm/kernel/psci_smp.c,
> but not all SoCs support the security extension.
> Is there a simpler one like spin-table available for arm32?

Per my understanding, spin-table is similar as the "pen" based
solution in arm32, both can't reliably support kexec, suspend etc...

> 
> If we force generic methods like psci or spin-table
> for new platforms, we can stop proliferated smp code.
> (Of course, we are just shifting the complexity
> from the kernel to firmware.)

psci is good but not all SoCs support secure extensions. spin-table
can't support kexec, suspend. Except prefer psci for news SoCs
with secure extensions, no better solutions AFAIK.

Re: [Question] vendor-specific cpu enable-method

[Masahiro Yamada]

Hi.


2018-09-13 11:29 GMT+09:00 Jisheng Zhang <Jisheng.Zhang@synaptics.com>:
> On Thu, 13 Sep 2018 10:23:35 +0900 Masahiro Yamada wrote:
>
>> Hello.
>>
>>
>> Sorry if I am asking a stupid question.
>>
>>
>> For arm64, there are only 2 cpu methods, psci and spin-table.
>>
>> Why do we still allow vendor-specific methods upstreamed
>> for arm 32bit ports?
>>
>> To me, it looks like SoC vendors continue inventing
>> different (but similar) ways to do the same thing.
>>
>> It is a historical reason for old platforms.
>>
>> However, if I look at Documentation/devicetree/bindings/arm/cpus.txt
>> enable-method properties are still increasing.
>>
>>
>> psci is available in arch/arm/kernel/psci_smp.c,
>> but not all SoCs support the security extension.
>> Is there a simpler one like spin-table available for arm32?
>
> Per my understanding, spin-table is similar as the "pen" based
> solution in arm32, both can't reliably support kexec, suspend etc...

Right.  spin-table is based on pen-based implementation,
and just a back-up plan in case psci is not available for some reasons.


>>
>> If we force generic methods like psci or spin-table
>> for new platforms, we can stop proliferated smp code.
>> (Of course, we are just shifting the complexity
>> from the kernel to firmware.)
>
> psci is good but not all SoCs support secure extensions. spin-table
> can't support kexec, suspend. Except prefer psci for news SoCs
> with secure extensions, no better solutions AFAIK.

OK, psci is preferred if it is available.

Otherwise, ... vendor specific code.


Thanks.


-- 
Best Regards
Masahiro Yamada