Re: [PATCH v3 2/2] Make hard lockup detection use timestamps

[Don Zickus <dzickus@redhat.com>]

On Fri, Jul 22, 2011 at 03:34:37PM -0700, ZAK Magnus wrote:
> On Fri, Jul 22, 2011 at 12:53 PM, Don Zickus <dzickus@redhat.com> wrote:
> 
> > So I played with the hardlockup case and I kinda like the timestamp thing.
> > It seems to give useful data.  In fact I feel like I can shrink the
> > hardlockup window, run some tests and see where the latencies are in a
> > system.  The patch itself I think is ok, I'll review on Monday or Tuesday
> > when I get some more free time.
> >
> > However, I ran the softlockup case and the output was a mess.  I think
> > rcu_sched stalls were being detected and as a result it was NMI dumping
> > stack traces for all cpus.  I can't tell if it was your patch or some
> > uncovered bug.
> >
> > I'll dig into on Monday.  Not sure if you were able to see that.
> >
> > Thanks,
> > Don
> >
> I'm not sure what you mean. One problem could be the wording I used.
> For the soft stalls I just called it LOCKUP, mostly to be very showy
> in order to cover that case where it's unclear what exactly is
> happening. This doesn't do much to distinguish soft and hard lockups,
> and I see LOCKUP otherwise seems to refer to hard lockup, so maybe
> that's misleading.

It had nothing to do with the wording.  It was spewing a ton of stack
traces.  Most of them related to rcu_sched stalls which requested stack
traces for each cpu (and the machine I as on had 16 cpus) repeatedly.

So from a user perspective, I just saw a flood of stack traces scroll
across the screen forever for a minute.  It was impossible to determine
what was going on without reviewing the logs once everything calmed down.
That is never a good thing.  It probably has nothing to do with your
patch, but it is something that should be looked at.

I'll try and poke today or tomorrow.

Cheers,
Don