Re: RFC: futex_wait() can DoS the tick

[Thomas Gleixner <tglx@linutronix.de>]

On Thu, 11 Jun 2015, Mike Galbraith wrote:
> On Thu, 2015-06-11 at 10:34 +0200, Thomas Gleixner wrote:
> 
> > Can you try the following, please?
> > 
> > Enable function tracer and hrtimer events manually. Then watch the irq
> > count on cpu3. If it stalls or becomes slow, then stop the trace with
> > 
> >       echo 0 >/sys/kernel/debug/tracing/tracing_on
> > 
> > If the overhead of the function tracer hides the problem, then try just
> > with hrtimer, sched_switch and irq events.
> 
> Yeah, function tracer hides it.  After some fiddling with buffer size to
> be able to ever see a tick, I finally have a trace trimmed with a tick
> on each end, 175ms gap between them.  It's still 1.5 after xz, sending
> off list.

What you are looking at is the actual timer interrupt vector, but
that's not the the tick.

# grep 'expire_entry.*tick_sched_timer' trace2 | wc -l
45

175 / 45  =~ 4 . So I you have CPNFIG_HZ=250

The kernel does not care whether the actual interrupt happens or the
timer is expired by other means.

If Xen relies on the vector, it's none of my problems ....

Thanks,

	tglx