On Fri, 8 Nov 2013, Frederic Weisbecker wrote:

> On Fri, Nov 08, 2013 at 03:23:07PM -0500, Vince Weaver wrote:
> > On Fri, 8 Nov 2013, Frederic Weisbecker wrote:
> > > 
> > > There seem to be a loop that takes too long in intel_pmu_handle_irq(). Your two
> > > previous reports seemed to suggest that lbr is involved, but not this one.
> > 
> > I may be wrong but I think everything between <NMI> and <EOE> is just
> > noise from the NMI perf-event watchdog timer kicking in.
> 
> Ah good point.
> 
> So the pattern seem to be that irq work/perf_event_wakeup is involved, may be
> interrupting a tracepoint event or so.

I managed to construct a reproducible test case, which is attached.  I 
sometimes have to run it 5-10 times before it triggers.

> It would be nice if you enable CONFIG_FRAME_POINTER in your future reports
> so that we get more precise traces. Or may be you actually enabled it and it
> doesn't work?

No, it's not enabled, I can try setting that in the future.

Vince