On Tue, Mar 08, 2016 at 06:52:06PM +0100, Sebastian Andrzej Siewior wrote:

* Daniel Wagner | 2016-03-08 16:59:13 [+0100]:

quoted

Hi,

Hi,

quoted

As Peter correctly pointed out in [1] a simple conversion from
wait to swait in completion.c wont work. I played a bit around and
came up with this rather ugly idea.

besides all the things I mentioned privatly, here is what I have
currently in -RT:

+void swake_up_all_locked(struct swait_queue_head *q)
+{
+       struct swait_queue *curr;
+       int wakes = 0;
+
+       while (!list_empty(&q->task_list)) {
+
+               curr = list_first_entry(&q->task_list, typeof(*curr),
+                                       task_list);
+               wake_up_process(curr->task);
+               list_del_init(&curr->task_list);
+               wakes++;
+       }
+       WARN_ON(wakes > 2);
+}
+EXPORT_SYMBOL(swake_up_all_locked);

the remaining part is what you have. The only user so far is complete()
and currently I see ony complete_all() with zero or one waiter.
If none of my boxes die over the night, I intend to release this
tomorrow in -RT and see if someone else triggers the limit.

However I don't think if your DEFER flag solution is all that bad.  I
have also the block-mq in -RT using swait and they perform wakes with
irqs-off. Not in -RT but mainline. So me might need something to make
it work properly. But if we defer the wakeup they might come at us and
complain about the latency???

Is it really just about latency?  Does this deferral not lead to an
inversion in the case where the single woken task isn't the highest
priority waiter on the completion (and doesn't run due to a
middle-priority thing spinning)?

In order for this to work, it seems like the chosen waiter would need to
inherit the highest priority of all waiters (which AFAICT isn't
happening).

  Josh