On Thu, 2011-09-22 at 16:52 +0200, Oleg Nesterov wrote:

On 09/22, Peter Zijlstra wrote:

quoted

+static void wait_task_inactive_sched_in(struct preempt_notifier *n, int cpu)
+{
+	struct task_struct *p;
+	struct wait_task_inactive_blocked *blocked =
+		container_of(n, struct wait_task_inactive_blocked, notifier);
+
+	hlist_del(&n->link);
+
+	p = ACCESS_ONCE(blocked->waiter);
+	blocked->waiter = NULL;
+	wake_up_process(p);
+}
...
+static void
+wait_task_inactive_sched_out(struct preempt_notifier *n, struct task_struct *next)
+{
+	if (current->on_rq) /* we're not inactive yet */
+		return;
+
+	hlist_del(&n->link);
+	n->ops = &wait_task_inactive_ops_post;
+	hlist_add_head(&n->link, &next->preempt_notifiers);
+}

Tricky ;) Yes, the first ->sched_out() is not enough.

Not enough isn't the problem, its ran with rq->lock held and irqs
disabled, you simply cannot do ttwu() from there.

If we could, the subsequent task_rq_lock() in wait_task_inactive() would
be enough to serialize against the still in-flight context switch.

One of the problems with doing it from the next sched_in notifier, is
that next can be idle, and then we do a A -> idle -> B switch, which is
of course sub-optimal.

quoted

 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
 {
...
+	rq = task_rq_lock(p, &flags);
+	trace_sched_wait_task(p);
+	if (!p->on_rq) /* we're already blocked */
+		goto done;

This doesn't look right. schedule() clears ->on_rq a long before
__switch_to/etc.

Oh, bugger, yes its before we can drop the rq for idle balance and
nonsense like that. (!p->on_rq && !p->on_cpu) should suffice I think.

And it seems that we check ->on_cpu above, this is not UP friendly.

True, but its what the old code did.. and I was seeing performance
suckage compared to the unpatched kernel (not that the p->on_cpu busy
wait fixed it)...

quoted

-			set_current_state(TASK_UNINTERRUPTIBLE);
-			schedule_hrtimeout(&to, HRTIMER_MODE_REL);
-			continue;
-		}
+	hlist_add_head(&blocked.notifier.link, &p->preempt_notifiers);
+	task_rq_unlock(rq, p, &flags);

I thought about reimplementing wait_task_inactive() too, but afaics there
is a problem: why we can't race with p doing register_preempt_notifier() ?
I guess register_ needs rq->lock too.

We can actually, now you mention it.. ->pi_lock would be sufficient and
less expensive to acquire.