Re: 2.6.14-rc3-rt2

[Steven Rostedt <rostedt@goodmis.org>]

On Wed, 5 Oct 2005, Ingo Molnar wrote:

>
> * Thomas Gleixner <tglx@linutronix.de> wrote:
>
> > On Wed, 2005-10-05 at 10:29 -0400, Steven Rostedt wrote:
> > > Hmm, Ingo,
> > >
> > > Do you know why time goes backwards when I run hackbench as a realtime
> > > process?  I added the output of start and stop and it does seem to go
> > > backwards.
> > >
> > > Thomas?
> >
> > Yes. Thats happening. I moved the priority of softirq-timer above
> > hackbench priority and the problem goes away. I look into this
> > further.
>
> wouldnt hackbench then permanently starve run_timer_softirq(), and
> update_times() in particular?
>

It seems that the problem comes down to the call to getnstimeofday in
do_gettimeofday.

void getnstimeofday(struct timespec *ts)
{
	nsec_t delta;
	unsigned long seq;

	/* atomically read __monotonic_clock() */
	do {
		seq = read_seqbegin(&system_time_lock);

		delta = __monotonic_clock() - xtime_last_update;
		*ts = xtime_last_update_ts;

	} while (read_seqretry(&system_time_lock, seq));

	set_normalized_timespec(ts,
				ts->tv_sec,
				ts->tv_nsec + (long) delta);

}


I found that xtime_last_update is not updated while hackbench is running,
and the delta ends up being 16 billion and some change. So the call to
set_normalized_timespec overflows with the (long) delta.

-- Steve