At Enthought we have been tracking a deadlock in some code that turned out to be due to the following scenario:

  0) There is some cyclic garbage that requires collection; an object in the garbage is referred to by a weakref with a callback
  1) You have a lock that is acquired
  2) While the lock is held, the garbage collector runs (on the same thread)
  3) The weakref callback in (0) is called via the garbage collecter and the callback tries to acquire the same lock that was acquired in (1)

Attached is a simple script that exercises the issue in Python 3.3. The script monkey-patches Queue.get to force garbage collection after acquiring the Queue.not_empty lock.

Adding Brian Quinlan to the nosy list.

I don't see an easy way around this other than documenting it and recommending that people do explicit executor shutdown where necessary.

Weakref callbacks are dangerous things.  Weakref callbacks that acquire a lock at any point doubly so.

I've encounter this issue. To easily avoid this issue, I've change `queue.put(None)` to `queue.put(None, block=False)` to work around this.

Sounds related to issue14976.

Per the decision of issue14976.

Davin: I don't see the link with #14976. There are no external signals involved in this issue. There's just a piece of simple pure Python code, which can deadlock if gc happens to kick in at the wrong moment.

I don't think this issue should be closed.

I don't really understand this issue.  How is concurrent.futures supposed to be the culprit here?  Can you elaborate which weakref callback you are talking about?

In the script attached to the original issue, the weakref callback that causes the hang is the callback defined in ThreadPoolExecutor._adjust_thread_count

Attached is a faulthandler stack captured from Python 3.6.1. The script submitted here uses a patched queue.get() method to force garbage collection to reliably trigger the issue, but this deadlock was observed in a real application using threadpool executors.

there's a PR on the other issue i opened for this without noticing this one so i'm marking this as the dup.

There's a proposed fix in https://github.com/python/cpython/pull/5216
It would be nice if someone able to reproduce this issue could test it (the reproducer script here wouldn't work with the new SimpleQueue implementation).