On 2017-06-22 19:30:07 [+0200], Mike Galbraith wrote:

On Thu, 2017-06-22 at 18:34 +0200, Sebastian Andrzej Siewior wrote:

quoted

On 2017-06-20 09:45:06 [+0200], Mike Galbraith wrote:

quoted

See ! and ?

See see.
What about this:

I'll give it a go, likely during the weekend.

It survived >1d on my AMD-A10 box. I am adding this description:

|If a task is queued as a sleeper for a wakeup and never goes to
|schedule() (because it just obtained the lock) then it will receive a
|spurious wake up which is not "bad", it is considered. Until that wake
|up happens this task can no be enqueued for any wake ups handled by the
|WAKE_Q infrastructure (because a task can only be enqueued once). This
|wouldn't be bad if we would use the same wakeup mechanism for the wake
|up of sleepers as we do for "normal" wake ups. But we don't…
|
|So.
|   T1			T2		T3
|   spin_lock(x)				spin_unlock(x);
|   					wake_q_add(q1, T1)
|   spin_unlock(x)
|   set_state(TASK_INTERRUPTIBLE)
|   if (!condition)
|	schedule()
|			condition = true
|			wake_q_add(q2, T1)
|			// T1 not added, still enqueued
|			wake_up_q(q2)
|					wake_up_q_sleeper(q1)
|					// T1 not woken up, wrong task state
|
|In order to solve this race this patch adds a wake_q_node for the
|sleeper case.

and consider this closed unless I hear from you different things :)

Sebastian