[PATCH] dynticks: fix hrtimer rounding error in next_timer_interrupt

The rework of next_timer_interrupt() fixed the timer wheel bugs, but
invented a rounding error versus the next hrtimer event. This is caused
by the conversion of the hrtimer internal representation to relative
jiffies.

This causes bug #8100:
http://bugzilla.kernel.org/show_bug.cgi?id=8100

next_timer_interrupt() returns "now" in such a case and causes the code
in tick_nohz_stop_sched_tick() to trigger the timer softirq, which is
bogus as no timer is due for expiry. This results in an endless context
switching between idle and ksoftirqd until a timer is due for expiry.

Modify the hrtimer evaluation so that, it returns now + 1, when the
conversion results in a delta < 1 jiffie.

It's confirmed to resolve bug #8100

Reported-by: Emil Karlson <jkarlson@cc.hut.fi>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 16 insertions, 3 deletions

diff --git a/kernel/timer.c b/kernel/timer.c
index 797cccb86431..440048acaea1 100644
--- a/kernel/timer.c
+++ b/kernel/timer.c
@@ -695,15 +695,28 @@ static unsigned long cmp_next_hrtimer_event(unsigned long now,
 {
 	ktime_t hr_delta = hrtimer_get_next_event();
 	struct timespec tsdelta;
+	unsigned long delta;
 
 	if (hr_delta.tv64 == KTIME_MAX)
 		return expires;
 
-	if (hr_delta.tv64 <= TICK_NSEC)
-		return now;
+	/*
+	 * Expired timer available, let it expire in the next tick
+	 */
+	if (hr_delta.tv64 <= 0)
+		return now + 1;
 
 	tsdelta = ktime_to_timespec(hr_delta);
-	now += timespec_to_jiffies(&tsdelta);
+	delta = timespec_to_jiffies(&tsdelta);
+	/*
+	 * Take rounding errors in to account and make sure, that it
+	 * expires in the next tick. Otherwise we go into an endless
+	 * ping pong due to tick_nohz_stop_sched_tick() retriggering
+	 * the timer softirq
+	 */
+	if (delta < 1)
+		delta = 1;
+	now += delta;
 	if (time_before(now, expires))
 		return now;
 	return expires;