I've decided to make the test more clean, so I added `time.sleep` and now we can be sure that here is enough time for queue's underlying process to synchronize all variables:
https://github.com/szobov/cpython/compare/master...regression-test-on-multiprocessing-queue-feeder-ignore-exception

It shows that sometimes `Queue._sem` will never be decreased.
https://github.com/szobov/cpython/blob/master/Lib/multiprocessing/queues.py#L266
If it's undesirable behavior, would it be better to ignore this lines: https://github.com/szobov/cpython/blob/master/Lib/multiprocessing/queues.py#L257-L259 ?

But in my thoughts the problem is in the functions `util._exit_function` and `util.is_exiting`. It's look like that `Queue._thread` share the same variable `util._exiting` with the process that started this thread. If you'll try to print it you see that when function `util._exit_function` executed for process it changes the value of `util._exiting` flag to `True` and after, from `Queue._thread`, the value for this flag stay `True`. But it's the daemon process and it should still work, isn't it? Or I'm missing something important here.