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

[Thomas Renninger <trenn@suse.de>]

On Monday 25 October 2010 15:55:08 Arjan van de Ven wrote:
> On 10/25/2010 2:41 AM, Thomas Renninger wrote:
> > On Monday 25 October 2010 08:54:34 Arjan van de Ven wrote:
> >> On 10/19/2010 4:36 AM, Thomas Renninger wrote:
> >>>    static void poll_idle(void)
> >>>    {
> >>> -	trace_power_start(POWER_CSTATE, 0, smp_processor_id());
> >>>    	local_irq_enable();
> >>>    	while (!need_resched())
> >>>    		cpu_relax();
> >>> -	trace_power_end(0);
> >>>    }
> >> why did you remove the idle tracepoints from this one ???
> > Because no idle/sleep state is entered here.
> > State 0 does not exist or say, it means the machine is not idle.
> > The new event uses idle state 0 spec conform as "exit sleep state".
> >
> > If this should still be trackable some kind of dummy sleep state:
> > #define IDLE_BUSY_LOOP 0xFE
> > (or similar) must get defined and passed like this:
> > trace_processor_idle(IDLE_BUSY_LOOP, smp_processor_id());
> >      cpu_relax()
> > trace_processor_idle(0, smp_processor_id());
> >
> > I could imagine this is somewhat worth it to compare idle results
> > to "no idle state at all" is used.
> > But nobody should ever use idle=poll, comparing deep sleep states
> > with C1 with (idle=halt) should be sufficient?
> 
> this is not idle=poll on the command line only.
> this also gets used normally, in two cases
> 1) during real time operations, for some short periods of time
>      (think wallstreet trading)
> 2) by the menu governor when the next event is less than a few 
> microseconds away, so short that even C1 is too much
> 
> I know that your new API tries to use "0" as exit, but 0 is already 
> taken (in all power terminology at least on x86 it is) for this.
cpuidle indeed misuses C0 as "poll idle" state.
That's really bad/misleading, but nothing that can be changed easily.

I agree shifting C0 (cpuidle) <-> POLL_IDLE event
and              "not idle"   <-> real C0 (executing instructions)
or however this gets mapped makes things even worse.

Damn, it could be that easy and straight forward, but I agree that
this kills the approach to trigger state 0 event if C0 is entered
(C0 as defined as operational mode executing instructions).

     Thomas