tick: Make oneshot broadcast robust vs. CPU offlining
authorThomas Gleixner <tglx@linutronix.de>
Wed, 26 Jun 2013 10:17:32 +0000 (12:17 +0200)
committerThomas Gleixner <tglx@linutronix.de>
Tue, 2 Jul 2013 12:26:44 +0000 (14:26 +0200)
In periodic mode we remove offline cpus from the broadcast propagation
mask. In oneshot mode we fail to do so. This was not a problem so far,
but the recent changes to the broadcast propagation introduced a
constellation which can result in a NULL pointer dereference.

What happens is:

CPU0 CPU1
idle()
  arch_idle()
    tick_broadcast_oneshot_control(OFF);
      set cpu1 in tick_broadcast_force_mask
  if (cpu_offline())
     arch_cpu_dead()

cpu_dead_cleanup(cpu1)
 cpu1 tickdevice pointer = NULL

broadcast interrupt
  dereference cpu1 tickdevice pointer -> OOPS

We dereference the pointer because cpu1 is still set in
tick_broadcast_force_mask and tick_do_broadcast() expects a valid
cpumask and therefor lacks any further checks.

Remove the cpu from the tick_broadcast_force_mask before we set the
tick device pointer to NULL. Also add a sanity check to the oneshot
broadcast function, so we can detect such issues w/o crashing the
machine.

Reported-by: Prarit Bhargava <prarit@redhat.com>
Cc: athorlton@sgi.com
Cc: CAI Qian <caiqian@redhat.com>
Link: http://lkml.kernel.org/r/alpine.DEB.2.02.1306261303260.4013@ionos.tec.linutronix.de
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
kernel/time/tick-broadcast.c

index d067c01586f56fca68ad0e030b5b400555f1ac73..4790037163f6e001c6aa4cd1d06f9b64c9188970 100644 (file)
@@ -532,6 +532,13 @@ again:
        cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
        cpumask_clear(tick_broadcast_force_mask);
 
+       /*
+        * Sanity check. Catch the case where we try to broadcast to
+        * offline cpus.
+        */
+       if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
+               cpumask_and(tmpmask, tmpmask, cpu_online_mask);
+
        /*
         * Wakeup the cpus which have an expired event.
         */
@@ -773,10 +780,12 @@ void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
        raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 
        /*
-        * Clear the broadcast mask flag for the dead cpu, but do not
-        * stop the broadcast device!
+        * Clear the broadcast masks for the dead cpu, but do not stop
+        * the broadcast device!
         */
        cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
+       cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
+       cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
 
        raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 }