JDK-8150134 changed the call to DCQS::apply_closure_to_completed_buffer, passing the thread's deactivation threshold as the stop_at argument. The intent was that this would result in the apply_xxx doing nothing and returning false if there aren't at least that many buffers available, and the false return value is used to trigger deactivation.

Unfortunately, that was based on a mis-understanding of the current meaning of the stop_at parameter.  The problem is in DCQS::get_completed_buffer, which contains

  if (_n_completed_buffers <= stop_at) {
    _process_completed = false;
    return NULL;
  }

With the old green_zone value for stop_at, any concurrent refinement thread would only stop here when the the number of completed buffers fell to green_zone, at which point _process_completed would be set false, which is used by the primary refinement thread to determine whether it should be active.