Re: [PATCH] PERF(kernel): Cleanup power events V2

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Tuesday, October 26, 2010, Arjan van de Ven wrote:
> On 10/26/2010 1:38 PM, Rafael J. Wysocki wrote:
> > On Tuesday, October 26, 2010, Pierre Tardy wrote:
> >> On Tue, Oct 26, 2010 at 2:08 PM, Rafael J. Wysocki<rjw@sisk.pl>  wrote:
> >>> On Tuesday, October 26, 2010, Pierre Tardy wrote:
> >>>> On Tue, Oct 26, 2010 at 12:58 PM, Peter Zijlstra<peterz@infradead.org>  wrote:
> >>>>> On Tue, 2010-10-26 at 11:56 -0500, Pierre Tardy wrote:
> >>>>>> +       trace_runtime_pm_usage(dev, atomic_read(&dev->power.usage_count)+1);
> >>>>>>          atomic_inc(&dev->power.usage_count);
> >>>>> That's terribly racy..
> >>>>>
> >>>> I know. I'm not proud of this.. As I said, this is preliminary patch.
> >>>> We dont really need to have this prev_usage. This is just for debug.
> >>>> It mayprobably endup with something like:
> >>>>
> >>>>           atomic_inc(&dev->power.usage_count);
> >>>> +       trace_power_device_usage(dev);
> >>> Well, please tell me what you're trying to achieve.
> >> Please see attached the kind of pytimechart output I'm trying to
> >> achieve (yes, this chart is not coherent, seems I'm still missing some
> >> traces)
> >>
> >> We basically want to have a trace point eachtime the usage_counter
> >> changes, so that I can display nice timecharts, and Arjan can have the
> >> comm of the process that eventually generated the rpm_get, in order to
> >> pinpoint it in powertop.
> >>
> >> What you dont see in the above two lines is that
> >> trace_power_device_usage(dev); actually reads the usage_count, as well
> >> as the driver and device name.
> > I'm afraid that for this to really work you'd need to put usage_count under a
> > spinlock along with your trace point, which I'm not really sure I like.
> >
> > Besides, I'm not really sure the manipulations of usage_count are worth
> > tracing.
> 
> what's most interesting is the 0->1  and 1->0 transitions.

But they are only meaningful in specific situations.  For example, if someone
does pm_runtime_get_noresume() when the device is active, there may be
a device suspend already under way at the same time.  So IMO what really
is interesting is when rpm_resume() is called with usage_count > 0 and then
perhaps when rpm_idle() or rpm_suspend() is called after usage_count drops
back to 0.

There are some other interesting cases, but they all need to be checked under
->power.lock and you need to do that cleverly, so that the _functionality_ is
not harmed.

Overall, I think that adding tracepoints to the runtime PM core code is really
premature at this point, given that we've just reworked it quite a bit recently.

Thanks,
Rafael