Possible regression due to "hrtimer: Get rid of hrtimer softirq"

[Thomas Gleixner]

On Thu, 7 May 2015, Simon Horman wrote:
> ------------[ cut here ]------------
> kernel BUG at kernel/irq_work.c:135!

  BUG_ON(!irqs_disabled());

So something enables interrupts in the periodic tick handling
machinery. Seems you have high resolution timers disabled, but nohz
enabled. And that code path has a local_irq_disable/enable pair which
causes havoc. Patch below.

Thanks,

	tglx

diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
index 753c211f6195..812f7a3b9898 100644
--- a/kernel/time/tick-sched.c
+++ b/kernel/time/tick-sched.c
@@ -967,11 +967,9 @@ static void tick_nohz_switch_to_nohz(void)
 	if (!tick_nohz_enabled)
 		return;
 
-	local_irq_disable();
-	if (tick_switch_to_oneshot(tick_nohz_handler)) {
-		local_irq_enable();
+	if (tick_switch_to_oneshot(tick_nohz_handler))
 		return;
-	}
+
 	tick_nohz_active = 1;
 	ts->nohz_mode = NOHZ_MODE_LOWRES;
 
@@ -986,7 +984,6 @@ static void tick_nohz_switch_to_nohz(void)
 	hrtimer_forward_now(&ts->sched_timer, tick_period);
 	hrtimer_set_expires(&ts->sched_timer, next);
 	tick_program_event(next, 1);
-	local_irq_enable();
 }
 
 /*
@@ -1171,7 +1168,7 @@ void tick_oneshot_notify(void)
  * Called cyclic from the hrtimer softirq (driven by the timer
  * softirq) allow_nohz signals, that we can switch into low-res nohz
  * mode, because high resolution timers are disabled (either compile
- * or runtime).
+ * or runtime). Called with interrupts disabled.
  */
 int tick_check_oneshot_change(int allow_nohz)
 {