Re: [RFC V1] cpuidle: add idle routine registration and cleanup pm_idle pointer

[Arjan van de Ven <arjan@linux.intel.com>]

  On 10/20/2010 12:47 PM, Venkatesh Pallipadi wrote:
> On Wed, Oct 20, 2010 at 12:44 PM, Arjan van de Ven
> <arjan@linux.intel.com>  wrote:
>>   On 10/20/2010 12:40 PM, Vaid
>>>> you ALWAYS have at least 2 idle handling states. The platform idle
>>>> one and the generic busy waiting one.
>>>> the later is needed for "I want absolutely 0 latency" cases.
>>> Some special overrides like idle=poll should handle this case even if
>>> cpuidle and related registration mechanism is compiled out. The point
>>> is that we need some flexibility even if the full framework is not
>>> included.
>> this is not idle=poll
>>
>> this is an (privileged) app or driver, at runtime, requesting a 0 usec max
>> latency for a short or long period of time.
>>
>>
>>>>> Making current cpuidle as default in kernel
>>>> not "in the kernel" but "for x86".
>>>> You're solving an x86 problem here, right?
>>>> (the pm_idle is an x86 only problem. other architectures should be
>>>> able to keep doing what they are doing)
>>>> For x86, lets solve it by going to cpuidle period... and if Andi can
>>>> find some bloat in cpuidle, lets see if the fat can be trimmed.
>>> Ok, you are suggesting that for x86 lets move cpuidle in kernel
>>> always, while it can be an optional module for other archs as it
>>> stands today.  We can slim down the cpuidle from current 7K or atleast
>>> split some parts like governors as modules if needed.
>> governors as modules is a total pain. modules don't solve the problem.
>> really. it's still code you need.
>> we have two governors today, menu and ladder
>> menu is best on anything that is tickless
>> ladder is useless on any tickless kernel, and likely not better than menu on
>> non-tickless.
>> that's it.
>> It will be good to have other archs also follow the same cpuidle
> I don't think they have to be modules. There can be a CPUIDLE_LITE
> which deCONFIG entire governor.c and individual governors (and
> probably sysfs stuff as well) for archs that can only use one state at
> any time, but still want to do config or runtime detection of one
> state and register that state.

there is no such case though; you always have at least 2 options!
(the hardware equivalent of "hlt" and the polling option)