On 01/24/2011 01:04 PM, Peter Zijlstra wrote:

diff --git a/kernel/sched.c b/kernel/sched.c
index dc91a4d..e4e57ff 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -327,7 +327,7 @@ struct cfs_rq {
         * 'curr' points to currently running entity on this cfs_rq.
         * It is set to NULL otherwise (i.e when none are currently running).
         */
-       struct sched_entity *curr, *next, *last;
+       struct sched_entity *curr, *next, *last, *yield;

I'd prefer it be called: skip or somesuch..

I could do that.  Do any of the other scheduler people have
a preference?

+static struct sched_entity *__pick_second_entity(struct cfs_rq *cfs_rq)
+{
+       struct rb_node *left = cfs_rq->rb_leftmost;
+       struct rb_node *second;
+
+       if (!left)
+               return NULL;
+
+       second = rb_next(left);
+
+       if (!second)
+               second = left;
+
+       return rb_entry(second, struct sched_entity, run_node);
+}

So this works because you only ever skip the leftmost, should we perhaps
write this as something like the below?

Well, pick_next_entity only ever *picks* the leftmost entity,
so there's no reason to skip others.

@@ -813,6 +840,9 @@ static void clear_buddies(struct cfs_rq *cfs_rq, struct 
sched_entity *se)

        if (cfs_rq->next == se)
                __clear_buddies_next(se);
+
+       if (cfs_rq->yield == se)
+               __clear_buddies_yield(se);
  }

The 3rd hierarchy iteration.. :/

Except it won't actually walk up the tree above the level
where the buddy actually points at the se.  I suspect the
new code will do less tree walking than the old code.

+       /*
+        * Someone really wants this to run. If it's not unfair, run it.
+        */
+       if (cfs_rq->next&&  wakeup_preempt_entity(cfs_rq->next, left)<  1)
+               se = cfs_rq->next;
+
        clear_buddies(cfs_rq, se);

        return se;

This seems to assume ->yield cannot be ->next nor ->last, but I'm not
quite sure that will actually be true.

On the contrary, I specifically want ->next to be able to
override ->yield, for the reason that the _tasks_ that
have ->next and ->yield set could be inside the same _group_.

What I am assuming is that ->yield and ->last are not the
same task.  This is achieved by yield_task_fair calling
clear_buddies.

+/*
+ * sched_yield() is very simple
+ *
+ * The magic of dealing with the ->yield buddy is in pick_next_entity.
+ */
+static void yield_task_fair(struct rq *rq)
+{
+       struct task_struct *curr = rq->curr;
+       struct cfs_rq *cfs_rq = task_cfs_rq(curr);
+       struct sched_entity *se =&curr->se;
+
+       /*
+        * Are we the only task in the tree?
+        */
+       if (unlikely(rq->nr_running == 1))
+               return;
+
+       clear_buddies(cfs_rq, se);
+
+       if (curr->policy != SCHED_BATCH) {
+               update_rq_clock(rq);
+               /*
+                * Update run-time statistics of the 'current'.
+                */
+               update_curr(cfs_rq);
+       }
+
+       set_yield_buddy(se);
+}

You just lost sysctl_sched_compat_yield, someone might be upset (I
really can't be bothered much with people using sys_yield :-), but if
you're going down that road you want a hunk in kernel/sysctl.c as well I
think.

I lost sysctl_sched_compat_yield, because with my code
yield is no longer a noop.

I'd be glad to remove the sysctl.c bits if you want :)
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Reply via email to