Asyncio: fork() in coroutine causes bad file descriptor

[wjmelements]

Bug report

Bug description:

Here is a simple repro:

import asyncio
import os

async def main():
    pid = os.fork()
    if pid:
        os.waitpid(pid, 0)

asyncio.run(main())

The traceback looks like:

Traceback (most recent call last):
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/runners.py", line 190, in run
    return runner.run(main)
           ^^^^^^^^^^^^^^^^
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/runners.py", line 118, in run
    return self._loop.run_until_complete(task)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/base_events.py", line 641, in run_until_complete
    self.run_forever()
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/base_events.py", line 608, in run_forever
    self._run_once()
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/base_events.py", line 1898, in _run_once
    event_list = self._selector.select(timeout)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/selectors.py", line 566, in select
    kev_list = self._selector.control(None, max_ev, timeout)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
OSError: [Errno 9] Bad file descriptor

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "~/forkasyncio.py", line 9, in <module>
    asyncio.run(main())
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/runners.py", line 189, in run
    with Runner(debug=debug) as runner:
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/runners.py", line 63, in __exit__
    self.close()
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/runners.py", line 72, in close
    loop.run_until_complete(loop.shutdown_asyncgens())
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/base_events.py", line 641, in run_until_complete
    self.run_forever()
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/base_events.py", line 608, in run_forever
    self._run_once()
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/asyncio/base_events.py", line 1898, in _run_once
    event_list = self._selector.select(timeout)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/Cellar/python@3.11/3.11.10/Frameworks/Python.framework/Versions/3.11/lib/python3.11/selectors.py", line 566, in select
    kev_list = self._selector.control(None, max_ev, timeout)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
OSError: [Errno 9] Bad file descriptor
sys:1: RuntimeWarning: coroutine 'BaseEventLoop.shutdown_asyncgens' was never awaited

I believe the traceback is produced by the child process.

I would not expect bad file descriptor in either the parent or the child because fork is supposed to copy the open file descriptors. I have even tried a custom fork module to ensure os.fork isn't closing the file descriptor.

The issue also happens if the child does sys.exit(). I find the best workaround is to have the child do os._exit(0):

import asyncio
import os
import sys

async def main():
    pid = os.fork()
    if pid:
        os.waitpid(pid, 0)
    else:
        sys.stdout.flush()
        sys.stderr.flush()
        os._exit(0)

asyncio.run(main())

This produces no traceback. I also avoid doing anything asyncio in the child.

I'm mainly curious about what file descriptor is bad; it doesn't seem possible.

It sounds like this issue could be fixed by #99539, but I still reproduce the issue in 3.12.7 and 3.13.0, though with a slightly different exception:

Traceback (most recent call last):
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/runners.py", line 194, in run
    return runner.run(main)
           ^^^^^^^^^^^^^^^^
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/runners.py", line 118, in run
    return self._loop.run_until_complete(task)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/base_events.py", line 674, in run_until_complete
    self.run_forever()
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/base_events.py", line 641, in run_forever
    self._run_once()
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/base_events.py", line 1948, in _run_once
    event_list = self._selector.select(timeout)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/selectors.py", line 566, in select
    kev_list = self._selector.control(None, max_ev, timeout)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: I/O operation on closed kqueue object

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/runners.py", line 71, in close
    loop.run_until_complete(loop.shutdown_asyncgens())
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/base_events.py", line 674, in run_until_complete
    self.run_forever()
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/base_events.py", line 641, in run_forever
    self._run_once()
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/base_events.py", line 1948, in _run_once
    event_list = self._selector.select(timeout)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/selectors.py", line 566, in select
    kev_list = self._selector.control(None, max_ev, timeout)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: I/O operation on closed kqueue object

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "~/forkasyncio.py", line 10, in <module>
    asyncio.run(main())
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/runners.py", line 193, in run
    with Runner(debug=debug, loop_factory=loop_factory) as runner:
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/runners.py", line 62, in __exit__
    self.close()
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/runners.py", line 77, in close
    loop.close()
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/unix_events.py", line 68, in close
    super().close()
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/selector_events.py", line 104, in close
    self._close_self_pipe()
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/selector_events.py", line 111, in _close_self_pipe
    self._remove_reader(self._ssock.fileno())
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/selector_events.py", line 305, in _remove_reader
    self._selector.unregister(fd)
  File "/Users/will/.pyenv/versions/3.12.7/lib/python3.12/selectors.py", line 542, in unregister
    self._selector.control([kev], 0, 0)
ValueError: I/O operation on closed kqueue object
/Users/will/.pyenv/versions/3.12.7/lib/python3.12/asyncio/base_events.py:712: RuntimeWarning: coroutine 'BaseEventLoop.shutdown_asyncgens' was never awaited

CPython versions tested on:

3.11

Operating systems tested on:

macOS

[gvanrossum]

For various reasons, it's a very bad idea to fork while an event loop is running and then continue running the loop in both parent and child process, so in the child process the event loop is intentionally sabotaged (I forget by which mechanism). However, it would be nice to get a better error for this. You can't just warn about forking when an event loop is running: if you don't use the event loop in the child, you can do something else, notably exec another binary. (If the child exits normally, the end-of-process GC might wake up daemons.)

So I give this a low importance and low priority, but in principle it is a bug and worthy of being fixed -- after most other bugs have been fixed, though. :-)

[kumaraditya303]

It sounds like this issue could be fixed by #99539, but I still reproduce the issue in 3.12.7 and 3.13.0, though with a slightly different exception:

That fix is meant for not sharing the loop with the forked child process from the parent process, if you do something else in child it will work.

Mixing a running loop with forked process is not supported if you intend to continue using old loop and asyncio in forked child process.

[justineakehurst]

I'm experiencing this issue on Python 3.11.12, when trying to run Salt's 'salt-master' daemon. Applying the fix from #99539 seems to address the issue.