Subject: hrtimer: Don't call the timer handler from hrtimer_start
From: Peter Zijlstra <a.p.zijlstra@chello.nl>
Date: Fri Aug 12 17:39:54 CEST 2011

 [<ffffffff812de4a9>] __delay+0xf/0x11
 [<ffffffff812e36e9>] do_raw_spin_lock+0xd2/0x13c
 [<ffffffff815028ee>] _raw_spin_lock+0x60/0x73              rt_b->rt_runtime_lock
 [<ffffffff81068f68>] ? sched_rt_period_timer+0xad/0x281
 [<ffffffff81068f68>] sched_rt_period_timer+0xad/0x281
 [<ffffffff8109e5e1>] __run_hrtimer+0x1e4/0x347
 [<ffffffff81068ebb>] ? enqueue_rt_entity+0x36/0x36
 [<ffffffff8109f2b1>] __hrtimer_start_range_ns+0x2b5/0x40a  base->cpu_base->lock  (lock_hrtimer_base)
 [<ffffffff81068b6f>] __enqueue_rt_entity+0x26f/0x2aa       rt_b->rt_runtime_lock (start_rt_bandwidth)
 [<ffffffff81068ead>] enqueue_rt_entity+0x28/0x36
 [<ffffffff81069355>] enqueue_task_rt+0x3d/0xb0
 [<ffffffff810679d6>] enqueue_task+0x5d/0x64
 [<ffffffff810714fc>] task_setprio+0x210/0x29c              rq->lock
 [<ffffffff810b56cb>] __rt_mutex_adjust_prio+0x25/0x2a      p->pi_lock
 [<ffffffff810b5d2c>] task_blocks_on_rt_mutex+0x196/0x20f

Instead make __hrtimer_start_range_ns() return -ETIME when the timer
is in the past. Since body actually uses the hrtimer_start*() return
value its pretty safe to wreck it.

Also, it will only ever return -ETIME for timer->irqsafe || !wakeup
timers.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 kernel/hrtimer.c |   48 +++++++++++++++++++++++-------------------------
 1 file changed, 23 insertions(+), 25 deletions(-)

Index: linux-stable/kernel/hrtimer.c
===================================================================
--- linux-stable.orig/kernel/hrtimer.c
+++ linux-stable/kernel/hrtimer.c
@@ -646,37 +646,24 @@ static inline int hrtimer_enqueue_reprog
 					    struct hrtimer_clock_base *base,
 					    int wakeup)
 {
-#ifdef CONFIG_PREEMPT_RT_BASE
-again:
 	if (base->cpu_base->hres_active && hrtimer_reprogram(timer, base)) {
+		if (!wakeup)
+			return -ETIME;
+
+#ifdef CONFIG_PREEMPT_RT_BASE
 		/*
 		 * Move softirq based timers away from the rbtree in
 		 * case it expired already. Otherwise we would have a
 		 * stale base->first entry until the softirq runs.
 		 */
-		if (!hrtimer_rt_defer(timer)) {
-			ktime_t now = ktime_get();
-
-			__run_hrtimer(timer, &now);
-			/*
-			 * __run_hrtimer might have requeued timer and
-			 * it could be base->first again.
-			 */
-			if (&timer->node == base->active.next)
-				goto again;
-			return 1;
-		}
-#else
-	if (base->cpu_base->hres_active && hrtimer_reprogram(timer, base)) {
+		if (!hrtimer_rt_defer(timer))
+			return -ETIME;
 #endif
-		if (wakeup) {
-			raw_spin_unlock(&base->cpu_base->lock);
-			raise_softirq_irqoff(HRTIMER_SOFTIRQ);
-			raw_spin_lock(&base->cpu_base->lock);
-		} else
-			__raise_softirq_irqoff(HRTIMER_SOFTIRQ);
+		raw_spin_unlock(&base->cpu_base->lock);
+		raise_softirq_irqoff(HRTIMER_SOFTIRQ);
+		raw_spin_lock(&base->cpu_base->lock);
 
-		return 1;
+		return 0;
 	}
 
 	return 0;
@@ -1067,8 +1054,19 @@ int __hrtimer_start_range_ns(struct hrti
 	 *
 	 * XXX send_remote_softirq() ?
 	 */
-	if (leftmost && new_base->cpu_base == &__get_cpu_var(hrtimer_bases))
-		hrtimer_enqueue_reprogram(timer, new_base, wakeup);
+	if (leftmost && new_base->cpu_base == &__get_cpu_var(hrtimer_bases)) {
+		ret = hrtimer_enqueue_reprogram(timer, new_base, wakeup);
+		if (ret) {
+			/*
+			 * In case we failed to reprogram the timer (mostly
+			 * because out current timer is already elapsed),
+			 * remove it again and report a failure. This avoids
+			 * stale base->first entries.
+			 */
+			__remove_hrtimer(timer, new_base,
+					timer->state & HRTIMER_STATE_CALLBACK, 0);
+		}
+	}
 
 	unlock_hrtimer_base(timer, &flags);