Re: [GIT PULL 0/8] Annotation weekly ponies delivery

[Namhyung Kim]

2012-05-14 (ì), 13:55 +0200, Ingo Molnar:
> * Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> > On Sat, 2012-05-12 at 13:40 -0700, Linus Torvalds wrote:
> > > The whole "-e cycles:pp" doesn't work any more. I don't get any nice
> > > PEBS information, I get the totally useless irq-based profiling.
> > > 
> > > The difference for a "make -j" profile is quite stunning:
> > > 
> > > Doing "perf record -f -e cycles:pp -F 20000 make -j"
> > > 
> > >  - my current git:
> > > 
> > >     [ perf record: Woken up 47 times to write data ]
> > >     [ perf record: Captured and wrote 11.890 MB perf.data (~519462 samples) ]
> > > 
> > >  - with the above tree pulled into my current git tree (but compiling
> > > the same old tree):
> > > 
> > >     [ perf record: Woken up 1 times to write data ]
> > >     [ perf record: Captured and wrote 0.031 MB perf.data (~1375 samples) ] 
> > 
> > The output simply suggests we're not getting enough samples 
> > not that PEBS isn't working, in fact I can could reproduce 
> > without using PEBS.
> > 
> > This bisected to the below commit, the code has since been 
> > changed again and all that evlist stuff gives me a head-ache. 
> > Acme, Namhyung ?
> > 
> > 55261f46702cec96911a81aacfb3cba13434d304 is the first bad commit
> > commit 55261f46702cec96911a81aacfb3cba13434d304
> > Author: Namhyung Kim <namhyung.kim@xxxxxxx>
> > Date:   Mon May 7 14:08:59 2012 +0900
> > 
> >     perf evlist: Fix creation of cpu map
> 
> Another detail seems to be that the bug takes per-task-inherited 
> profiling. Doing:
> 
>    perf record -a -e cycles:pp make -j64 bzImage
> 
> produces the expected number of events. Without the '-a' the bug 
> Linus found triggers.
> 
> Similarly, 'perf top' does not trigger the bug - because it's 
> using per CPU, not per task (inherited) profiling.
> 

perf_evlist__config_attrs() has this:

	if (evlist->cpus->map[0] < 0)
		opts->no_inherit = true;

meaning that per task profiling won't enable event inheritance. I don't
know why it's needed though.


-- 
Regards,
Namhyung Kim


--
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/