Re: [PATCH 2/3] PERF(kernel): Cleanup power events

[Thomas Renninger <trenn@suse.de>]

On Monday 25 October 2010 16:11:10 Arjan van de Ven wrote:
> On 10/25/2010 5:55 AM, Thomas Renninger wrote:
> 
> 
> >> But the actual code does not actually deal with any 'state 0', does it?
> > It does. Not being idle is tracked by cpuidle driver as state 0
> > (arch independent):
> > /sys/devices/system/cpu/cpu0/cpuidle/state0/
> > halt/C1 on X86 is:
> > /sys/devices/system/cpu/cpu0/cpuidle/state1/
> > ...
> state0 is still OS idle!
Yes, I just realized that.
Which is very unfortunate.
The whole cpuidle stuff is based on ACPI C-states and
cat /sys/devices/system/cpu/cpu0/cpuidle/state0/name
C0
is plain wrong if it's used as "poll idle" time.
C0 is defined as (in the ACPI spec):
----------
2.5 Processor Power State Definitions
C0 Processor Power State
While the processor is in this state, it executes instructions.
----------

> the API is just weird for this, from a userspace perspective
> 
> if the kernel picks this state 0 for the idle handler, the userspace app 
> gets
> two events
> 
> one for going to state 0 to enter the idle state
> one for going to state 0 to exit idle
> 
> but they're the exact same event in your API.
> 
> rather unpleasant from a userspace program perspective....
Yeah. But the re-definition of C0 being "Linux poll idle"
will confuse users as well. Not sure whether this should get
touched, though.

Thanks for clarification, I wasn't aware of that...

    Thomas