Application processes hung in direct reclaim under congestion_wait() / shrink_inactive_list()

Issue

• The problem is that several processes may get stuck pretty much forever
  in shrink_inactive_list()/too_many_isolated() loop in the direct reclaim path.
  These would than block other processes via rw sema and the hang would cascade.

• Some of symptoms can be seen :

  • Lots of tasks were waiting on semaphore.
  • The semaphore holder was "numad". It was in page reclaim code, calling congestion_wait() and waiting.
  • Another tasks were waiting in the same congestion wait queue at that time.
  • The system was running out of memory and a little bit swapping at that time, the number of free pages did not
    fall below the minimum watermark on all zones though.