Re: [PATCH 2/3] perf: add OMAP support for the new power events

[Kevin Hilman <khilman@ti.com>]

Thomas Renninger <trenn@suse.de> writes:

> On Wednesday 05 January 2011 12:05:18 Jean Pihet wrote:
>> Hi Paul,
>> 
>> On Tue, Jan 4, 2011 at 7:48 PM, Paul Walmsley <paul@pwsan.com> wrote:
>> > Hello Jean,
>> >
>> > On Tue, 4 Jan 2011, jean.pihet@newoldbits.com wrote:
>> >
>> >> From: Jean Pihet <j-pihet@ti.com>
>> >>
>> >> The patch adds the new power management trace points for
>> >> the OMAP architecture.
>> >>
>> >> The trace points are for:
>> >> - default idle handler. Since the cpuidle framework is
>> >>   instrumented in the generic way there is no need to
>> >>   add trace points in the OMAP specific cpuidle handler;
>> >> - cpufreq (DVFS),
>> >> - clocks changes (enable, disable, set_rate),
>> >
>> > A question about these.  Are these only meant to track calls to these
>> > functions from outside the clock code?  Or meant to track actual hardware
>> > clock changes?
>> The former: this is used to track the clock requests from outside the
>> clock framework.
>> 
>> > If the latter, then it might make sense to put these
>> > trace points into the functions that actually change the hardware
>> > registers, e.g., omap2_dflt_clk_{enable,disable}(), etc., since a
>> > clk_enable() on a leaf clock may result in many internal system clocks
>> > being enabled up the clock tree.
>> I agree with you it is better to track the actual clock changes instead.
>> I propose to move the tracepoints to omap2_clk_{enable...} which
>> enables all the clocks irrespectively of the installed handler.
>> Note about the clock handlers: omap2_dflt_clk_enable happens to be the
>> handler for all controllable clocks but could that change in the
>> future?

> Looks like there is cpuidle34xx using cpuidle framework on specific
> boards only.  And pm34xx.c and others override pm_idle and use the
> same low level functions to reduce power consumption as cpuidle34xx.
> Ideally pm34xx.c (and others) would not override pm_idle, but register as
> a cpuidle driver. Then the idle events would be tracked by the
> cpuidle subsystem automatically (with my latest patches).
> But this would be a more intrusive change (are there efforts to do that?).

Whenever CPUidle is enabled though, the cpuidle34xx code is used so the
pm_idle in pm34xx is not used. This allows us to have a way to do PM
with and without CPUidle, so without CPUidle, we can still get some
basic PM in idle.

> Even if it should happen at some point, adding some additional events for
> people to better debug/monitor the stuff now does not hurt.

I agree, as the two both very useful.

Tracking CPUidle transitions gives us some high-level information on PM
transitions, but what is most interesting for real PM analysis on OMAP
is exactly which powerdomains, clockdomains and clocks transitions (or
didn't transition) for a given state.

Kevin