linux/kernel/sched
Thomas Gleixner a1cbcaa9ea sched_clock: Prevent 64bit inatomicity on 32bit systems
The sched_clock_remote() implementation has the following inatomicity
problem on 32bit systems when accessing the remote scd->clock, which
is a 64bit value.

CPU0			CPU1

sched_clock_local()	sched_clock_remote(CPU0)
...
			remote_clock = scd[CPU0]->clock
			    read_low32bit(scd[CPU0]->clock)
cmpxchg64(scd->clock,...)
			    read_high32bit(scd[CPU0]->clock)

While the update of scd->clock is using an atomic64 mechanism, the
readout on the remote cpu is not, which can cause completely bogus
readouts.

It is a quite rare problem, because it requires the update to hit the
narrow race window between the low/high readout and the update must go
across the 32bit boundary.

The resulting misbehaviour is, that CPU1 will see the sched_clock on
CPU1 ~4 seconds ahead of it's own and update CPU1s sched_clock value
to this bogus timestamp. This stays that way due to the clamping
implementation for about 4 seconds until the synchronization with
CLOCK_MONOTONIC undoes the problem.

The issue is hard to observe, because it might only result in a less
accurate SCHED_OTHER timeslicing behaviour. To create observable
damage on realtime scheduling classes, it is necessary that the bogus
update of CPU1 sched_clock happens in the context of an realtime
thread, which then gets charged 4 seconds of RT runtime, which results
in the RT throttler mechanism to trigger and prevent scheduling of RT
tasks for a little less than 4 seconds. So this is quite unlikely as
well.

The issue was quite hard to decode as the reproduction time is between
2 days and 3 weeks and intrusive tracing makes it less likely, but the
following trace recorded with trace_clock=global, which uses
sched_clock_local(), gave the final hint:

  <idle>-0   0d..30 400269.477150: hrtimer_cancel: hrtimer=0xf7061e80
  <idle>-0   0d..30 400269.477151: hrtimer_start:  hrtimer=0xf7061e80 ...
irq/20-S-587 1d..32 400273.772118: sched_wakeup:   comm= ... target_cpu=0
  <idle>-0   0dN.30 400273.772118: hrtimer_cancel: hrtimer=0xf7061e80

What happens is that CPU0 goes idle and invokes
sched_clock_idle_sleep_event() which invokes sched_clock_local() and
CPU1 runs a remote wakeup for CPU0 at the same time, which invokes
sched_remote_clock(). The time jump gets propagated to CPU0 via
sched_remote_clock() and stays stale on both cores for ~4 seconds.

There are only two other possibilities, which could cause a stale
sched clock:

1) ktime_get() which reads out CLOCK_MONOTONIC returns a sporadic
   wrong value.

2) sched_clock() which reads the TSC returns a sporadic wrong value.

 can be excluded because sched_clock would continue to increase for
   one jiffy and then go stale.

 can be excluded because it would not make the clock jump
   forward. It would just result in a stale sched_clock for one jiffy.

After quite some brain twisting and finding the same pattern on other
traces, sched_clock_remote() remained the only place which could cause
such a problem and as explained above it's indeed racy on 32bit
systems.

So while on 64bit systems the readout is atomic, we need to verify the
remote readout on 32bit machines. We need to protect the local->clock
readout in sched_clock_remote() on 32bit as well because an NMI could
hit between the low and the high readout, call sched_clock_local() and
modify local->clock.

Thanks to Siegfried Wulsch for bearing with my debug requests and
going through the tedious tasks of running a bunch of reproducer
systems to generate the debug information which let me decode the
issue.

Reported-by: Siegfried Wulsch <Siegfried.Wulsch@rovema.de>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Link: http://lkml.kernel.org/r/alpine.LFD.2.02.1304051544160.21884@ionos
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@vger.kernel.org
2013-04-08 11:50:44 +02:00
..
auto_group.c Revert "sched/autogroup: Fix crash on reboot when autogroup is disabled" 2012-12-11 10:23:45 +01:00
auto_group.h Revert "sched/autogroup: Fix crash on reboot when autogroup is disabled" 2012-12-11 10:23:45 +01:00
clock.c sched_clock: Prevent 64bit inatomicity on 32bit systems 2013-04-08 11:50:44 +02:00
core.c sched: Convert BUG_ON()s in try_to_wake_up_local() to WARN_ON_ONCE()s 2013-03-21 11:48:20 +01:00
cpupri.c sched/rt: Move rt specific bits into new header file 2013-02-07 20:51:08 +01:00
cpupri.h
cputime.c cputime: Use local_clock() for full dynticks cputime accounting 2013-02-24 12:57:16 +01:00
debug.c sched: Fix /proc/sched_debug failure on very very large systems 2013-02-22 10:27:25 +01:00
fair.c Merge branch 'sched-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip 2013-02-19 18:19:48 -08:00
features.h Automatic NUMA Balancing V11 2012-12-16 15:18:08 -08:00
idle_task.c sched/nohz: Rewrite and fix load-avg computation -- again 2012-07-05 20:58:13 +02:00
Makefile sched: Move cputime code to its own file 2012-08-20 13:05:17 +02:00
rt.c Merge branch 'sched-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip 2013-02-19 18:19:48 -08:00
sched.h sched/rt: Move rt specific bits into new header file 2013-02-07 20:51:08 +01:00
stats.c sched: Fix /proc/sched_stat failure on very very large systems 2013-02-22 10:27:24 +01:00
stats.h Merge branch 'sched/core' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip into cputime-tip 2011-12-19 19:23:15 +01:00
stop_task.c sched: Fix migration thread runtime bogosity 2012-08-13 18:41:55 +02:00