multiprocessing.Process cannot be closed if the process is reaped elsewhere

Here's a small reproducer where I start a multiprocessing.Process - then reap it with os.waitpid(

import os
import multiprocessing
import socket
import sys


ctx = multiprocessing.get_context("spawn")

def target(sock):
    with sock:
        sock.send(b"\x00")
        sock.recv(1)


def main():
    a, b = socket.socketpair()
    with a, b:
        with b:
            proc = ctx.Process(target=target, args=(b, ))
            proc.start()
        assert a.recv(1) == b"\x00"
        a.send(b"\x00")
        os.waitpid(proc.pid, 0)
        if sys.version_info >= (3, 7):
            proc.close()
        elif proc.is_alive():
            raise RuntimeError("proc={proc} is still running".format(proc=proc))


if __name__ == "__main__":
    sys.exit(main())

which results in:

 Traceback (most recent call last):
  File "/home/graingert/projects/distributed/demo.py", line 31, in <module>
    sys.exit(main())
  File "/home/graingert/projects/distributed/demo.py", line 25, in main
    proc.close()
  File "/home/graingert/miniconda3/lib/python3.9/multiprocessing/process.py", line 181, in close
    raise ValueError("Cannot close a process while it is still running. "
ValueError: Cannot close a process while it is still running. You should first call join() or terminate().

One issue is the call to self._popen.poll() suppresses the OSError of a reaped process:
https://github.com/python/cpython/blob/b6558d768f19584ad724be23030603280f9e6361/Lib/multiprocessing/popen_fork.py#L26-L31

I think this can be fixed by checking the state of the multiprocessing.Process.sentinel instead of using os.waitpid

从 issue 中的 reproducer 开始，检查 Lib/multiprocessing/popen_fork.py，尤其是 _popen.poll() 的处理，以及 multiprocessing.Process.close()。调查一个被外部 reaped 的进程如何由其 sentinel 表示，并将其与基于 waitpid 的状态检测进行比较。当 reproducer 可以从外部 reaped 子进程，然后关闭 Process，而不会错误地触发表示进程仍在运行的 ValueError 时，即视为完成。