2012/5/4 Frederic Weisbecker [off-list ref]:

On Thu, May 03, 2012 at 05:55:57PM +0300, Gilad Ben-Yossef wrote:

quoted

@@ -1317,9 +1322,15 @@ unsigned long get_next_timer_interrupt(unsigned long now)

      if (cpu_is_offline(smp_processor_id()))
              return now + NEXT_TIMER_MAX_DELTA;
      spin_lock(&base->lock);
-     if (time_before_eq(base->next_timer, base->timer_jiffies))
-             base->next_timer = __next_timer_interrupt(base);
-     expires = base->next_timer;
+     if (time_before_eq(base->next_timer, base->timer_jiffies)) {
+
+             if (__next_timer_interrupt(base, &expires))
+                     base->next_timer = expires;
+             else
+                     expires = now + NEXT_TIMER_MAX_DELTA;

I believe you can update base->next_timer to now + NEXT_TIMER_MAX_DELTA,
so on any further idle interrupt exit that call tick_nohz_stop_sched_tick(),
we won't get again the overhead of __next_timer_interrupt().

Ah forget that, I was confused. If we do that we actually get the useless timer
at now + NEXT_TIMER_MAX_DELTA.

So I think the patch is fine.

Acked-by: Frederic Weisbecker <redacted>