Re: [PATCH 4/4] PERF: fix power:cpu_idle double end events

[Thomas Renninger <trenn@suse.de>]

On Wednesday 27 October 2010 01:43:25 Thomas Renninger wrote:
> cpu_idle events (transition into sleep state and exiting) are
> both fired in pm_idle().
> 
> Entering sleep state and exiting should always get fired inside
> pm_idle() already.
> 
> This is a revert of commit c882e0feb937af4e5b991cbd1c
Robert: I expect you tested this on a machine with no cpuidle
driver registered?

I should have had a deeper look at this at once, done so now:

Current cpu_idle, power_start/end (same before my changes) behavior
on X86 is rather weird (without this patch):

if pm_idle is:
   poll_idle -> should throw double end events

   default_idle -> only throws power_start, your patch
                   fixed that, but in the generic cpu_idle
                   thread function which always gets executed
                   also if pm_idle != default_idle
                   So Robert fixed this case, but at the wrong place.

   cpuidle_idle_call -> depends whether intel_idle or acpi_idle
                        driver registered:

        intel_idle -> throws a cpu idle state event, still double end
                      events (one from cpuidle, one from
                      process_{32,64}.c, due to Robert's patch

        acpi_idle  -> whether a power_start event is thrown at all
                      depends on (cmp with acpi_idle_do_entry()):
                      cx->entry_method == ACPI_CSTATE_FFH
                      will end up in a power_start event via:
                      acpi_processor_ffh_cstate_enter(cx)
                           mwait_idle_with_hints()
                      but it will not in case of:
                      cx->entry_method == ACPI_CSTATE_HALT
                      or IO based switching (the else path there):
                      /* IO port based C-state */
                      

Again without this(my) patch you get:
   poll_idle                  -> double end events
   default_idle               -> all is fine (with your patch)
   cpuidle_idle_call
        intel_idle registered -> double end events
        acpi_idle registered  -> double end events
                                 start events may not be thrown
                                 at all.

perf timechart can handle double end events, this may be
the reason this was overseen.

Argh, I tried to come up with patches, but run out of
time. I will send something soon.
I also found a bug in my userspace stuff: I forgot to convert
u64 to u32..., sorry about that.
I try to resend everything tomorrow.

     Thomas