Re: Patch for tracing c states (power_end) on x86

[Robert Schöne]

Hi,

I tried to pass 0 in "my" sleep routine "static void mwait_idle(void)"
Which led to the following behaviour:
The event was reported on /sys/kernel/debug/tracing, but still not
for sys_perf_open.
As 1 had been the argument which led to a working tracing, I assumed,
that the argument should be the same as the 2nd arg of the last
power_start event.
Since this argument had been 1 in my case, it worked for me. However, 0
did not.

Bye Robert



Am Mittwoch, den 24.02.2010, 16:36 +0800 schrieb Li Zefan:
> Robert SchÃne wrote:
> > Hello,
> > 
> > Since noone replied to my last mail (Febr. 15th, 11:42), describing the
> > way to fix the missing c-state tracing, here's a patch.
> > Maybe its easier that way.
> > 
> > (I used the perf-fixes-for-linus git tree to obtain a
> > more-then-up-to-date version)
> > 
> > Bye Robert
> > 
> > 
> > diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
> > index 02d6780..b1cfb88 100644
> > --- a/arch/x86/kernel/process.c
> > +++ b/arch/x86/kernel/process.c
> > @@ -384,6 +384,7 @@ void default_idle(void)
> >                 else
> >                         local_irq_enable();
> >                 current_thread_info()->status |= TS_POLLING;
> > +               trace_power_end(1);
> >         } else {
> >                 local_irq_enable();
> >                 /* loop is done by the caller */
> > @@ -451,6 +452,7 @@ void mwait_idle_with_hints(unsigned long ax,
> > unsigned long cx)
> >                 if (!need_resched())
> >                         __mwait(ax, cx);
> >         }
> > +       trace_power_end((ax>>4)+1);
> 
> The only argument of trace_power_end() is a dummy, so you can just
> pass 0 or 1 to the trace hook, actually better pass 0 to be
> consistent with other parts.
> 
> The dummy argument can't be eliminated, because the macros that
> automatically generates racing code have some limitations, and
> seems it's not so easy to get over.
> 
> >  }
> >  
> >  /* Default MONITOR/MWAIT with no hints, used for default C1 state */
> > @@ -467,6 +469,7 @@ static void mwait_idle(void)
> >                         __sti_mwait(0, 0);
> >                 else
> >                         local_irq_enable();
> > +               trace_power_end(1);
> >         } else
> >                 local_irq_enable();
> >  }
> > 
> > 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
> 
------------------------- Mail from Febr. 15th --------------------
Hi,
I have a question regarding the event "power/power_end".
For the standard linux kernel (2.6.32.8), it's just not reported -
neither for the /sys/kernel/debug/tracing nor for the sys_perf_open
approach.

System:
        Intel Core 2 Quad,
        Kernel 2.6.32.8,
        for sys_perf_open:always using sampling counters,
        (Kernel 2.6.33-rcX should show the same behavior)

After finding "my" c-state procedure in arch/x86/kernel/process.c
(which was "static void mwait_idle(void)" btw), I added a
trace_power_end call on the correct line:
...
                else
                        local_irq_enable();
        } else
...
->
...
                else
                        local_irq_enable();
                trace_power_end(0);
        } else
...
Now the event was reported on /sys/kernel/debug/tracing, but still not
for sys_perf_open.

Then I had the idea, that trace_power_end's argument should be the same
as the 2nd argument of the previous power_start.

That worked.

However, things to be done are: add trace_power_end's to some
process.c's procedures.


Bye Robert

-------------------------End of Mail of Febr. 15th --------------------


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/