Re: [PATCH] hrtimer: Replace cpu_base->active_bases with a direct check of the active list

[Ingo Molnar <mingo@kernel.org>]

* Ingo Molnar <mingo@kernel.org> wrote:

> This would speed up various hrtimer primitives like 
> hrtimer_remove()/add and simplify the code. It would be a net code 
> shrink as well.
> 
> Totally untested patch below. It gives:
> 
>    text    data     bss     dec     hex filename
>    7502     427       0    7929    1ef9 hrtimer.o.before
>    7422     427       0    7849    1ea9 hrtimer.o.after
> 
> and half of that code removal is from hot paths.
> 
> This would simplify the followup step of skipping over inactive bases 
> as well.

The followup step is attached below (untested as well).

Note that all other iterations already had a check for active.next, so 
the patch doesn't even bloat anything:

   text    data     bss     dec     hex filename
   7422     427       0    7849    1ea9 hrtimer.o.before
   7422     427       0    7849    1ea9 hrtimer.o.after

(I did a rename within migrate_hrtimers() because it used 'cpu_base' 
vs 'clock_base' inconsistently in a confusing (to me) manner.)

I'd also suggest the removal of the timerqueue_getnext() obfuscation: 
it 'sounds' complex but in reality it's a simple dereference to 
active.next. I think this is what triggered this rather pointless 
maintenance of active_bases.

Thanks,

	Ingo

---
 kernel/time/hrtimer.c |   24 +++++++++++++++---------
 1 file changed, 15 insertions(+), 9 deletions(-)

Index: tip/kernel/time/hrtimer.c
===================================================================
--- tip.orig/kernel/time/hrtimer.c
+++ tip/kernel/time/hrtimer.c
@@ -1660,29 +1660,35 @@ static void migrate_hrtimer_list(struct
 
 static void migrate_hrtimers(int scpu)
 {
-	struct hrtimer_cpu_base *old_base, *new_base;
+	struct hrtimer_cpu_base *old_cpu_base, *new_cpu_base;
 	int i;
 
 	BUG_ON(cpu_online(scpu));
 	tick_cancel_sched_timer(scpu);
 
 	local_irq_disable();
-	old_base = &per_cpu(hrtimer_bases, scpu);
-	new_base = this_cpu_ptr(&hrtimer_bases);
+	old_cpu_base = &per_cpu(hrtimer_bases, scpu);
+	new_cpu_base = this_cpu_ptr(&hrtimer_bases);
 	/*
 	 * The caller is globally serialized and nobody else
 	 * takes two locks at once, deadlock is not possible.
 	 */
-	raw_spin_lock(&new_base->lock);
-	raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
+	raw_spin_lock(&new_cpu_base->lock);
+	raw_spin_lock_nested(&old_cpu_base->lock, SINGLE_DEPTH_NESTING);
 
 	for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
-		migrate_hrtimer_list(&old_base->clock_base[i],
-				     &new_base->clock_base[i]);
+		struct hrtimer_clock_base *old_base = old_cpu_base->clock_base + i;
+		struct hrtimer_clock_base *new_base;
+
+		if (!old_base->active.next)
+			continue;
+
+		new_base = new_cpu_base->clock_base + i;
+		migrate_hrtimer_list(old_base, new_base);
 	}
 
-	raw_spin_unlock(&old_base->lock);
-	raw_spin_unlock(&new_base->lock);
+	raw_spin_unlock(&old_cpu_base->lock);
+	raw_spin_unlock(&new_cpu_base->lock);
 
 	/* Check, if we got expired work to do */
 	__hrtimer_peek_ahead_timers();