Li, Aubrey wrote:
> On Saturday, December 01, 2007 4:01 AM, Mark.Haywood at Sun.COM wrote:
> 
>> Mark Haywood wrote:
>>> Eric Saxe wrote:
>>>> Eric Saxe wrote:
>>>>> Li, Aubrey wrote:
>>>>>> C-state driver will let the processor go into deeper c-state.
>>>>>>
>>>>> I was thinking would use the existing mwait code path to bring the
>>>>> processor into the deeper c-states?
>>>>>   ...or are you suggesting that the c-state driver would contain
>>>>> code that is invoked out of that path?
>>>>>
>>>> Sorry. In case you can't parse the above, let me try again. ;)
>>>>
>>>> I was thinking we would use the existing mwait code path to bring
>>>> the processor into the deeper c-states.
>>>> Are you suggesting that the c-state driver would contain the code
>>>> that could be invoked out of that path?...
>>>> or something else?
>>> I was wondering what we were going to do here too. Do the deeper
>>> c-states tend to be vendor specific mechanisms? If so, then I think
>>> we might want this support in a separate driver that can be accessed
>>> via something like layered ioctls? This is what I was thinking of
>>> doing with the reworked p-state support.
>> BTW, I don't know how expensive this kind of mechanism might be.
> 
> I think the deeper c-states shouldn't be vendor specific.
> As long as:
> 1) the processor supports different power saving mode when it's in idle
> 2) the platform supports ACPI spec.

Do we know if 2) is true. What I mean is that ACPI defines how to do 
P-state transitions, but SpeedStep and PowerNow! (particularly the 
Opteron type) are pretty different to implement. Is it unlikely that 
there will be Vendor specific requirements for C-state support from 
different vendors?

> As for how expensive, I think
> 1)the policy is cheap. 
> 2) And the transition, C2 is no more than C1. C3 and deeper will be
> responsible for ensuring that the caches maintain coherency.
> 3) And the timer/interrupts needs to be fixed.
> Others needs to be further investigated, ;-)
> 
> Thanks,
> -Aubrey
> 
> 
> 
> 
> 
> 


Reply via email to