The origin of the issue I've seen is related to
a missing memory barrier between check for task->state and
the check for task->on_rq.

The task being woken up is already awake from a schedule()
and is doing the following:

	do {
		schedule()
		set_current_state(TASK_(UN)INTERRUPTIBLE);
	} while (!cond);

The waker, actually gets stuck doing the following in
try_to_wake_up():

	while (p->on_cpu)
		cpu_relax();

Analysis:

The instance I've seen involves the following race:

 CPU1					CPU2

 while () {
   if (cond)
     break;
   do {
     schedule();
     set_current_state(TASK_UN..)
   } while (!cond);
					wakeup_routine()
					  spin_lock_irqsave(wait_lock)
   raw_spin_lock_irqsave(wait_lock)	  wake_up_process()
 }					  try_to_wake_up()
 set_current_state(TASK_RUNNING);	  ..
 list_del(&waiter.list);

CPU2 wakes up CPU1, but before it can get the wait_lock and set
current state to TASK_RUNNING the following occurs:

 CPU3
 wakeup_routine()
 raw_spin_lock_irqsave(wait_lock)
 if (!list_empty)
   wake_up_process()
   try_to_wake_up()
   raw_spin_lock_irqsave(p->pi_lock)
   ..
   if (p->on_rq && ttwu_wakeup())
   ..
   while (p->on_cpu)
     cpu_relax()
   ..

CPU3 tries to wake up the task on CPU1 again since it finds
it on the wait_queue, CPU1 is spinning on wait_lock, but immediately
after CPU2, CPU3 got it.

CPU3 checks the state of p on CPU1, it is TASK_UNINTERRUPTIBLE and
the task is spinning on the wait_lock. Interestingly since p->on_rq
is checked under pi_lock, I've noticed that try_to_wake_up() finds
p->on_rq to be 0. This was the most confusing bit of the analysis,
but p->on_rq is changed under runqueue lock, rq_lock, the p->on_rq
check is not reliable without this fix IMHO. The race is visible
(based on the analysis) only when ttwu_queue() does a remote wakeup
via ttwu_queue_remote. In which case the p->on_rq change is not
done uder the pi_lock.

The result is that after a while the entire system locks up on
the raw_spin_irqlock_save(wait_lock) and the holder spins infintely

Reproduction of the issue:

The issue can be reproduced after a long run on my system with 80
threads and having to tweak available memory to very low and running
memory stress-ng mmapfork test. It usually takes a long time to
reproduce. I am trying to work on a test case that can reproduce
the issue faster, but thats work in progress. I am still testing the
changes on my still in a loop and the tests seem OK thus far.

Big thanks to Benjamin and Nick for helping debug this as well.
Ben helped catch the missing barrier, Nick caught every missing
bit in my theory.
Signed-off-by: Balbir Singh <bsingharora@gmail.com>
[ Updated comment to clarify matching barriers. Many
  architectures do not have a full barrier in switch_to()
  so that cannot be relied upon. ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Alexey Kardashevskiy <aik@ozlabs.ru>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Nicholas Piggin <nicholas.piggin@gmail.com>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: <stable@vger.kernel.org>
Link: http://lkml.kernel.org/r/e02cce7b-d9ca-1ad0-7a61-ea97c7582b37@gmail.comSigned-off-by: Ingo Molnar <mingo@kernel.org>