Re: 2.6.39-rc4+: Kernel leaking memory during FS scanning, regression?

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Apr 29, 2011 at 01:35:44AM +0200, Sedat Dilek wrote:
> On Fri, Apr 29, 2011 at 1:06 AM, Sedat Dilek <sedat.dilek@googlemail.com> wrote:
> > On Fri, Apr 29, 2011 at 12:02 AM, Thomas Gleixner <tglx@linutronix.de> 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.
> >>
> >> 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);
> >>
> >>
> >>
> >
> > Looks good so far, no stalls or call-traces.
> >
> > Really stressing with 20+ open tabs in firefox with flash-movie
> > running in one of them , tar-job, IRC-client etc.
> > I will run some more tests and collect data and send them later.
> >
> > - Sedat -
> >
> > P.S.: Patchset against linux-2.6-rcu.git#sedat.2011.04.23a where 0003
> > is from [2]
> >
> > [1] http://git.us.kernel.org/?p=linux/kernel/git/paulmck/linux-2.6-rcu.git;a=shortlog;h=refs/heads/sedat.2011.04.23a
> > [2] https://patchwork.kernel.org/patch/739782/
> >
> > $ l ../RCU-HOORAY/
> > insgesamt 40
> > drwxr-xr-x  2 sd sd  4096 29. Apr 01:02 .
> > drwxr-xr-x 35 sd sd 20480 29. Apr 01:01 ..
> > -rw-r--r--  1 sd sd   726 29. Apr 01:01
> > 0001-Revert-rcu-restrict-TREE_RCU-to-SMP-builds-with-PREE.patch
> > -rw-r--r--  1 sd sd   735 29. Apr 01:01
> > 0002-sched-Add-warning-when-RT-throttling-is-activated.patch
> > -rw-r--r--  1 sd sd  2376 29. Apr 01:01
> > 0003-2.6.39-rc4-Kernel-leaking-memory-during-FS-scanning-.patch
> >
> 
> As promised the tarball (at the end of the log I made some XZ compressing).
> 
> Wow!
> $ uptime
>  01:35:17 up 45 min,  3 users,  load average: 0.45, 0.57, 1.27
> 
> Thanks to all involved people helping to kill that bug (Come on Paul, smile!).

Woo-hoo!!!!

Many thanks to Thomas for tracking this down -- it is fair to say that
I never would have thought to look at timer initialization!  ;-)

							Thanx, Paul