Please describe the problem:
"test_thread_pools" does an unatomic add of "leftover_task_counter", while "test_thread_pools_entry_func" does an atomic decrement.  The race occurs when the main thread reads the value from memory, a worker thread decrements leftover_task_counter, then the main thread commits back to memory thus losing the atomic decrement done by the worker.

Steps to reproduce:
1. Bump "runs" in "test_thread_pools" up to something large (I used 30000) so that worker threads are running while the main thread is still pushing jobs into the thread pools


Actual results:
** ERROR:(threadpool-test.c:196):test_thread_pools: assertion failed: (runs * 3 == abs_thread_counter + leftover_task_counter)

Expected results:
Non-deterministic fails, about 3 in 20 runs; tested on POWER6 

Does this happen every time?
no

Other information:
Can be fixed by grabbing the "thread_counter_pools" lock when we do "leftover_task_counter += 3;".  Seems to happen more reliably on 2.12 for both POWER and i686, not sure why.