summaryrefslogtreecommitdiffstats
path: root/kernel
diff options
context:
space:
mode:
authorThomas Gleixner <tglx@linutronix.de>2013-04-06 10:10:27 +0200
committerThomas Gleixner <tglx@linutronix.de>2013-04-08 11:50:44 +0200
commita1cbcaa9ea87b87a96b9fc465951dcf36e459ca2 (patch)
treea58fb45d03bb4c55d7f32a33c6bcbc59414a3682 /kernel
parent383efcd00053ec40023010ce5034bd702e7ab373 (diff)
downloadtalos-obmc-linux-a1cbcaa9ea87b87a96b9fc465951dcf36e459ca2.tar.gz
talos-obmc-linux-a1cbcaa9ea87b87a96b9fc465951dcf36e459ca2.zip
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. #1 can be excluded because sched_clock would continue to increase for one jiffy and then go stale. #2 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
Diffstat (limited to 'kernel')
-rw-r--r--kernel/sched/clock.c26
1 files changed, 26 insertions, 0 deletions
diff --git a/kernel/sched/clock.c b/kernel/sched/clock.c
index c685e31492df..c3ae1446461c 100644
--- a/kernel/sched/clock.c
+++ b/kernel/sched/clock.c
@@ -176,10 +176,36 @@ static u64 sched_clock_remote(struct sched_clock_data *scd)
u64 this_clock, remote_clock;
u64 *ptr, old_val, val;
+#if BITS_PER_LONG != 64
+again:
+ /*
+ * Careful here: The local and the remote clock values need to
+ * be read out atomic as we need to compare the values and
+ * then update either the local or the remote side. So the
+ * cmpxchg64 below only protects one readout.
+ *
+ * We must reread via sched_clock_local() in the retry case on
+ * 32bit as an NMI could use sched_clock_local() via the
+ * tracer and hit between the readout of
+ * the low32bit and the high 32bit portion.
+ */
+ this_clock = sched_clock_local(my_scd);
+ /*
+ * We must enforce atomic readout on 32bit, otherwise the
+ * update on the remote cpu can hit inbetween the readout of
+ * the low32bit and the high 32bit portion.
+ */
+ remote_clock = cmpxchg64(&scd->clock, 0, 0);
+#else
+ /*
+ * On 64bit the read of [my]scd->clock is atomic versus the
+ * update, so we can avoid the above 32bit dance.
+ */
sched_clock_local(my_scd);
again:
this_clock = my_scd->clock;
remote_clock = scd->clock;
+#endif
/*
* Use the opportunity that we have both locks
OpenPOWER on IntegriCloud