Re: [patch] voluntary-preempt-2.6.9-rc3-mm1-S9

[Peter Williams]

Rui Nuno Capela wrote:
> Ingo wrote:
>
> > i've released the -S9 VP patch:
> > http://redhat.com/~mingo/voluntary-preempt/voluntary-preempt-2.6.9-rc3-mm2-S9
>
>
> Me again, we bad humor :(
>
> My SMP/HT box is (again) terribly in that uglyness of being quite
> unfriendly to -mm1, -mm2, and indirectly to -S8 and -S9 labeled kernels.
>
> It works quite well with vanilla 2.6.9-rc3, though.
>
> But very, very bad with those -mm1 or -mm2 patches. To get it straight,
> almost all the time it hangs, randomly, but not as completely as to a
> dramatic cold-reboot. It stalls on the the most administrative tasks. Name
> one, and it stalls! I can hardly feel lucky if it sometimes reaches the
> login prompt, while boot/initing.
>
> I know you remember this story. Yeah. This seems quite similar to some of
> earlier problems, but (un/fortunately) it doesn't seem related to VP at
> all. Just having -mm1 or -mm2 is enough to make this machine go astray.
>
> However, as usual, this seems to be ix86 SMP/HT specific. On my laptop, I
> get to run full 2.6.9-rc3-mm2-S9 UP very happily.
>
> Sorry if I can't get any real or useful debug data for now. The bad
> behavior I'm referring to, is terribly non-deterministic, so I couldn't
> get a pattern yet.
>
> I just wanted to let you know ;)

This may be my fault.  I made changes in the SMH code as part of the 
ZAPHOD patch but was unable to test them on a hyperthreaded machine due 
to a lack thereof (I have one on order).  I've attached a patch that 
reverts the changes that I made.  Can you give them a try please and let 
me know if they fix your problem?

Thanks
Peter
--
Peter Williams                                   pwil3058@xxxxxxxxxxxxxx

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce
--- linux-2.6.9-rc3-mm2/kernel/sched.c	2004-10-05 15:17:09.157119322 +1000
+++ linux-2.6.9-rc3-mm2-mod/kernel/sched.c	2004-10-05 15:27:05.355561830 +1000
@@ -3240,18 +3240,21 @@ need_resched:
 		put_task_in_sinbin(prev);
 
 	cpu = smp_processor_id();
-	if (unlikely(needs_idle_balance(rq))) {
+	if (unlikely(!rq->nr_running)) {
 go_idle:
 		idle_balance(cpu, rq);
-		/* This code should get optimised away when CONFIG_SCHED_SMT
-		 * is not defined
-		 */
-		if (dependent_idle(rq))
+		if (!rq->nr_running) {
+			next = rq->idle;
 			wake_sleeping_dependent(cpu, rq);
+			/*
+			 * wake_sleeping_dependent() might have released
+			 * the runqueue, so break out if we got new
+			 * tasks meanwhile:
+			 */
+			if (!rq->nr_running)
+				goto switch_tasks;
+		}
 	} else {
-		/* This code should all get optimised away when CONFIG_SCHED_SMT
-		 * is not defined
-		 */
 		if (dependent_sleeper(cpu, rq)) {
 			schedstat_inc(rq, sched_goidle);
 			next = rq->idle;
@@ -3262,7 +3265,7 @@ go_idle:
 		 * lock, hence go into the idle loop if the rq went
 		 * empty meanwhile:
 		 */
-		if (unlikely(recheck_needs_idle_balance(rq)))
+		if (unlikely(!rq->nr_running))
 			goto go_idle;
 	}