Re: System lockup with 2.6.26.8-rt16 on ARM9 [Solved]

[Daniel Walker <dwalker@fifo99.com>]

On Sat, 2009-08-29 at 09:47 +0200, Remy Bohmer wrote:

> Well, we found the root cause of this problem.
> It turned out to be caused by sched_clock() that made disjunct time jumps.
> This caused this check to become true in kernel/sched_rt.c:370:
>          if (rt_rq->rt_time > runtime) {
>                 rt_rq->rt_throttled = 1;
>                 if (rt_rq_throttled(rt_rq)) {
>                         sched_rt_rq_dequeue(rt_rq);
>                         return 1;
>                 }
>         }
> 
> The end results is that all realtime tasks got throttled for a long
> time, and that time got extended every time sched_clock() made such a
> jump. I would never have expected the scheduler would show this kind
> of behaviour while CONFIG_RT_GROUP_SCHED is _not_ set...
> 
> The root-cause of the sched_clock being faulty was a synchronisation
> issue between 2 clock domains. The CPU clock and the clock domain of
> the peripheral (GPT) on which the sched_clock() implementation was
> based. The GPT made jumps backwards which triggered a false wraparound
> detection in the conversion of 32->64 bit timestamps, causing the time
> to jump about 356 seconds in the future...
> 

Can you tell us more about what type of board this was? I've never heard
of a ARM board having an unstable clocksource before ..

Daniel