Re: problem with "sched: revert back to per-rq vruntime"?

From: Jayson King
Date: Wed Dec 31 2008 - 15:15:36 EST


Mike Galbraith wrote:
On Tue, 2008-12-30 at 13:37 -0600, Jayson King wrote:
Here's some times using bash, for example. Let me know if there are some other tests I can do.

[jayson@atlas tmp]$ time bash -i -c 'exit'
exit

real 0m0.753s
user 0m0.011s
sys 0m0.011s

Yeah, it's b0rked. The problem is that we call sched_slice() before the
new task's load weight has been added to the runqueue.

Normally this doesn't cause horrible problems, but when there is a +nice
hog on the runqueue, the lighter the hog, and/or heavier the victim...

[ 160.489402] bash:5858 vslice: 3840000000 weight: 1024 rq_weight: 15 inv_weight: 268435456

...the closer the victim's vruntime gets to the middle of next week.

Not-signed-off diag hack.

diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
index 5ad4440..3254213 100644
--- a/kernel/sched_fair.c
+++ b/kernel/sched_fair.c
@@ -449,13 +449,34 @@ static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
}
/*
+ * A freshly forked task's load weight has not yet been added to the runqueue,
+ * so we can't call calc_delta_weight() and get sane results. Therefore, we roll
+ * our own so place_entity() can place this task in the right spot. The runqueue
+ * weight and nr_running will be updated when the task is enqueued.
+ */
+static u64 sched_initial_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
+{
+ u64 slice = __sched_period(cfs_rq->nr_running + 1);
+
+ for_each_sched_entity(se) {
+ struct load_weight lw = cfs_rq->load;
+
+ lw.weight += se->load.weight;
+ lw.inv_weight = 0;
+ slice = calc_delta_mine(slice, se->load.weight, &lw);
+ }
+
+ return slice;
+}
+
+/*
* We calculate the vruntime slice of a to be inserted task
*
* vs = s/w
*/
static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
- return calc_delta_fair(sched_slice(cfs_rq, se), se);
+ return calc_delta_fair(sched_initial_slice(cfs_rq, se), se);
}
/*

I tested it. Now the times are back to normal.

# 2.6.28 under 100% nice CPU and Mike Galbraith patch
[jayson@atlas tmp]$ time bash -i -c 'exit'
exit

real 0m0.042s
user 0m0.006s
sys 0m0.003s



Jayson

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/