Re: 2.6.27.8 scheduler bug - threads not being scheduled for long periods

[Dimitri Sivanich <sivanich@sgi.com>]

On Mon, Jan 05, 2009 at 03:54:29PM -0600, Dimitri Sivanich wrote:
> On Mon, Jan 05, 2009 at 09:55:58PM +0100, Peter Zijlstra wrote:
> > On Mon, 2009-01-05 at 11:56 -0600, Dimitri Sivanich wrote:
> > > One place we've found this happens is in update_curr(), which calculates a
> > > delta_exec value as follows:
> > >         delta_exec = (unsigned long)(now - curr->exec_start);
> > > 
> > > Sometimes this value will be very large, as 'now' (the rq clock time) will
> > > be less than 'exec_start'.  When this happens, __update_curr() will
> > > calculate a delta_exec_weighted based on this large value and add it to the
> > > thread's vruntime:
> > >         curr->vruntime += delta_exec_weighted;
> > 
> > So you're saying your rq->clock = sched_clock_cpu(cpu) = sched_clock()
> > [on ia64] goes backwards?
> > 
> > If so, then that's an architecture bug, sched_clock() must never be seen
> > to go backwards!
> 
> Actually, sched_clock() should not go backwards on any one cpu, but the readings will be different between cpus.
> 
> Also, we noticed the following code is being used for sched_clock_cpu():
> u64 sched_clock_cpu(int cpu)
> {
>         if (unlikely(!sched_clock_running))
>                 return 0;
>         return sched_clock();
> }
> 
> and is called when smp_processor_id() != cpu.

And sure enough, the rq->clock is sometimes going backwards.

The comment for sched_clock() in arch/ia64/kernel/head.S:
 * Return a CPU-local timestamp in nano-seconds.  This timestamp is
 * NOT synchronized across CPUs its return value must never be
 * compared against the values returned on another CPU.  The usage in
 * kernel/sched.c ensures that.

We will try this with CONFIG_HAVE_UNSTABLE_SCHED_CLOCKS.

Thanks for the heads up!