Re: BUG: perf top enters loop synthesizing events for existing threads

[Arnaldo Carvalho de Melo <arnaldo.melo@gmail.com>]

Em Wed, Apr 08, 2015 at 11:34:37AM -0300, Arnaldo Carvalho de Melo escreveu:
> Em Wed, Apr 08, 2015 at 11:14:27AM -0300, Arnaldo Carvalho de Melo escreveu:
> > Em Wed, Apr 08, 2015 at 07:48:23AM -0600, David Ahern escreveu:
> > > On 4/8/15 7:45 AM, Arnaldo Carvalho de Melo wrote:
> > > >>I also noticed that if I set /proc/sys/kernel/perf_event_paranoid to -1
> > > >>to run it as !root, then the problem "goes away", which I think probably
> > > >>is explained by, as !root, not being able to parse some of the /proc
> > > >>files for existing threads and thus not triggering the bug, still
> > > >>investigating...
> > > 
> > > odd. because this has nothing to do with perf_events; it is just walking
> > > /proc and for the ppid adds a strstr and atoi(str). Something else is at
> > > play. I'll take a look.
> > 
> > You are correcly setting the pid values, that will have effects when
> > using findnew, i.e. threads will be added to the rbtree, which causes
> > allocations, etc I.e. it is probably triggering a dormant bug
> 
> Sitting there, will stop chrome to see if the problem is triggered by it...

Yeah, stopping chrome gets the system to a state that doesn't make it
trigger the bug, why it works when your patch is reverted still a
puzzle...

- Arnaldo