On Fri, 2 May 2014, Thomas Gleixner wrote:

On Fri, 2 May 2014, Sebastian Andrzej Siewior wrote:

quoted

* Steven Rostedt | 2014-04-22 14:16:50 [-0400]:

quoted

/*
 * Called by the local, per-CPU timer interrupt on SMP.
 */

@@ -1467,7 +1473,7 @@ void run_local_timers(void)

		return;
	}

-	if (!spin_do_trylock(&base->lock)) {
+	if (timer_should_raise_softirq(&base->lock)) {
		raise_softirq(TIMER_SOFTIRQ);
		return;
	}

Okay. So Peter said that it is okay to apply this since FULL_NO_HZ users
wouldn't complain on UP. I still wouldn't say it is broken but that is a
different story.
We have two users of this trylock. run_local_timers() which pops up
quite often (and you patched here) and the other is
get_next_timer_interrupt(). What do you suggest we do here? It is
basically the same thing.

It's different as it CANNOT fail on UP. That's called from the idle
code and there is no way that anything holds that lock on UP when idle
runs.

So yeah, you are right, that it's called from irq_exit() so it needs
an annotation at least. Maybe it's really cleaner to make it #if SMP
as well just for clarity raisins.

Thanks,

	tglx