On Fri, 29 April 2011 Thomas Gleixner wrote:
> On Thu, 28 Apr 2011, john stultz wrote:
> > On Thu, 2011-04-28 at 23:04 +0200, Thomas Gleixner wrote:
> > > /me suspects hrtimer changes to be the real culprit.
> >
> > I'm not seeing anything on right off, but it does smell like
> > e06383db9ec591696a06654257474b85bac1f8cb would be where such an issue
> > would crop up.
> >
> > Bruno, could you try checking out e06383db9ec, confirming it still
> > occurs (and then maybe seeing if it goes away at e06383db9ec^1)?
> >
> > I'll keep digging in the meantime.
>
> I found the bug already. The problem is that sched_init() calls
> init_rt_bandwidth() which calls hrtimer_init() _BEFORE_
> hrtimers_init() is called.
>
> That was unnoticed so far as the CLOCK id to hrtimer base conversion
> was hardcoded. Now we use a table which is set up at hrtimers_init(),
> so the bandwith hrtimer ends up on CLOCK_REALTIME because the table is
> in the bss.
>
> The patch below fixes this, by providing the table statically rather
> than runtime initialized. Though that whole ordering wants to be
> revisited.
Works here as well (applied alone), /proc/$(pidof rcu_kthread)/sched shows
total runtime continuing to increase beyond 950 and slubs continue being
released!
Thanks,
Bruno
> Thanks,
>
> tglx
>
> --- linux-2.6.orig/kernel/hrtimer.c
> +++ linux-2.6/kernel/hrtimer.c
> @@ -81,7 +81,11 @@ DEFINE_PER_CPU(struct hrtimer_cpu_base,
> }
> };
>
> -static int hrtimer_clock_to_base_table[MAX_CLOCKS];
> +static int hrtimer_clock_to_base_table[MAX_CLOCKS] = {
> + [CLOCK_REALTIME] = HRTIMER_BASE_REALTIME,
> + [CLOCK_MONOTONIC] = HRTIMER_BASE_MONOTONIC,
> + [CLOCK_BOOTTIME] = HRTIMER_BASE_BOOTTIME,
> +};
>
> static inline int hrtimer_clockid_to_base(clockid_t clock_id)
> {
> @@ -1722,10 +1726,6 @@ static struct notifier_block __cpuinitda
>
> void __init hrtimers_init(void)
> {
> - hrtimer_clock_to_base_table[CLOCK_REALTIME] = HRTIMER_BASE_REALTIME;
> - hrtimer_clock_to_base_table[CLOCK_MONOTONIC] = HRTIMER_BASE_MONOTONIC;
> - hrtimer_clock_to_base_table[CLOCK_BOOTTIME] = HRTIMER_BASE_BOOTTIME;
> -
> hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
> (void *)(long)smp_processor_id());
> register_cpu_notifier(&hrtimers_nb);
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>