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

[Trinabh Gupta <trinabh@linux.vnet.ibm.com>]

On 10/20/2010 12:19 AM, Venkatesh Pallipadi wrote:
> On Tue, Oct 19, 2010 at 11:38 AM, Arjan van de Ven
> <arjan@linux.intel.com>  wrote:
>>   On 10/19/2010 11:36 AM, Trinabh Gupta wrote:
>>>
>>> The core of the kernel's idle routine on x86 presently depends on an
>>> exported pm_idle function pointer that is unmanaged and causing
>>> hazard to various subsystems when they save and restore it.
>>> The first problem is that this exported pointer can be modified/flipped
>>> by any subsystem. There is no tracking or notification mechanism.
>>> Secondly and more importantly, various subsystems save the value of
>>> this pointer, flip it and later restore to the saved value. There is
>>> no guarantee that the saved value is still valid. The problem has
>>> been discussed in [2,3] and Peter Zijlstra suggested removing pm_idle
>>> and implementing a list based registration [1].
>>>
>>> This patch is an initial RFC implementation for x86 architecture
>>> only. This framework can be generalised for other archs and also
>>> include the current cpuidle framework for managing multiple idle
>>> routines.
>>>
>>> Tests done with the patch:
>>> ------------------------
>>> 1. Build (on 2.6.36-rc7) and booted on x86 with C1E as deepest idle
>>>     state and current_idle was selected to be mwait_idle.
>>>
>>> 2. Build (on 2.5.36-rc8) and booted on x86 (Nehalem) with ACPI C3 as
>>>     deepest sleep state.  The current_idle was selected to be
>>>     cpuidle_idle_call which is the cpuidle subsystem that will further
>>>     select idle routines from {C1,C2,C3}.
>>>
>>> Future implementation will try to eliminate this hirearchy and have
>>> a single registration and menu/idle cpuidle governor for selection
>>> of idle routine.
>>
>>
>> looks like you're duplicating the cpuidle subsystem
>>
>> how about biting the bullet and just always and only use the cpuidle
>> subsystem for all idle on x86 ?
>>
>
> I agree with Arjan.
> If we have a default_cpuidle driver which parses idle= params, handles
> various mwait quirks in x86 process*.c and registers with cpuidle, we
> can then always call cpuidle idle routine on x86.

This wouldn't duplicate code. It would move parts/functionality of 
cpuidle into the kernel, keeping governors alone as modules.

If we directly call cpuidle_idle_call() then this may be too much
overhead for architectures that have single idle routine i.e cases where 
cpuidle is not used and will be seen as a bloat. (c.f goals 4a,b).

>
> Thanks,
> Venki

Regards,
-Trinabh