2.6.37-rc3 massive interactivity regression on ARM

Andi -

This bug appears to have crept into 2.6.35.12 via a3fe22ee824895aafdc1b788e19c081a2e6dd9da and is still present in 2.6.35.13.

Mainline patch fe44d62122829959e960bc699318d58966922a69 (sched: Fix the irqtime code to deal with u64 wraps) seems to fix my system although I recommend also applying 8e92c20183ed0579d94501311b81c42b65cb2129 (sched: Fix the irqtime code for 32bit). Any stable releases that have picked up 305e6835e05513406fa12820e40e4a8ecb63743c (sched: Do not account irq time to current task) should probably also apply these.

Before I could get these patches to apply to 2.6.35.13 I had to tweak sched.c:update_rq_clock() to match f26f9aff6aaf67e9a430d16c266f91b13a5bff64 (Sched: fix skip_clock_update optimization). That patch was applied to 2.6.35-longterm, except for the update_rq_clock() portion. I believe that was because that portion depended on a patch from Ventakesh that wasn't applied until later. It looks like the logic that ended up in 2.6.35.12 was slightly different than in mainline; I'm not familiar enough with the scheduler details to know what the effects would be.

------------------------------------------------------------------------

---

diff -uprN a/kernel/sched.c b/kernel/sched.c
@@ -654,8 +654,10 @@ inline void update_rq_clock(struct rq *r
 	int cpu = cpu_of(rq);
 	u64 irq_time;
 
-	if (!rq->skip_clock_update)
-		rq->clock = sched_clock_cpu(cpu_of(rq));
+	if (rq->skip_clock_update)
+		return;
+
+	rq->clock = sched_clock_cpu(cpu);
 	irq_time = irq_time_cpu(cpu);
 	if (rq->clock - irq_time > rq->clock_task)
 		rq->clock_task = rq->clock - irq_time;

------------------------------------------------------------------------
Steven J. Magnani "I claim this network for MARS!
www.digidescorp.com Earthling, return my space modulator!"

#include <standard.disclaimer>