bpo-37410: subprocess closes the process handle when done

[vstinner]

On Windows, subprocess.Popen now closes the process handle when the
process completes. Previously, it was only closed when the Popen
object was destroyed.

https://bugs.python.org/issue37410

[vstinner]

cc @efiop @eryksun @zooba

[vstinner]

See also https://bugs.python.org/issue37380

[vstinner]

cc @gpshead

[vstinner]

Oh, a test now fails on Windows:

ERROR: test_threadsafe_wait (test.test_subprocess.ProcessTestCase)
Issue21291: Popen.wait() needs to be threadsafe for returncode.
----------------------------------------------------------------------
Traceback (most recent call last):
  File "C:\projects\cpython\lib\test\test_subprocess.py", line 1284, in test_threadsafe_wait
    proc.wait(timeout=20)
  File "C:\projects\cpython\lib\subprocess.py", line 1050, in wait
    return self._wait(timeout=timeout)
  File "C:\projects\cpython\lib\subprocess.py", line 1340, in _wait
    exitcode = _winapi.GetExitCodeProcess(self._handle)
TypeError: GetExitCodeProcess() argument must be int, not None

It seems like _wait() is not longer thread-safe with my change... but I'm surprised that it was thread-safe before my change.

The Unix implementation uses a lock to make _wait() thread-safe. Extract:

                    with self._waitpid_lock:
                        if self.returncode is not None:
                            break  # Another thread waited.
                        (pid, sts) = self._try_wait(0)
                        # Check the pid and loop as waitpid has been known to
                        # return 0 even without WNOHANG in odd situations.
                        # http://bugs.python.org/issue14396.
                        if pid == self.pid:
                            self._set_returncode(sts)

Maybe the Windows implementation should also use a lock?

The failing test comes from https://bugs.python.org/issue21291

[eryksun]

It seems like _wait() is not longer thread-safe with my change... but I'm surprised that it was thread-safe before my change.

Before we could rely on _handle being valid. Now I think we need to acquire a lock before checking and setting the return code. The following is an untested version of _wait that acquires a lock:

    def _wait(self, timeout):
        """Internal implementation of wait() on Windows."""
        if self.returncode is not None:
            return self.returncode

        if timeout is None:
            self._status_lock.acquire()
        else:
            if timeout < 0:
                # Disallow the -1 default timeout value of acquire.
                raise ValueError('timeout value must be positive')
            endtime = _time() + timeout
            if not self._status_lock.acquire(timeout=timeout):
                raise TimeoutExpired(self.args, timeout)
        try:
            if self.returncode is None:
                if timeout is None:
                    remaining_ms = _winapi.INFINITE
                else:
                    remaining = endtime - _time() if timeout != 0 else 0
                    if remaining < 0:
                        raise TimeoutExpired(self.args, timeout)
                    remaining_ms = int(remaining * 1000)
                # API note: returns immediately if remaining_ms is 0.
                result = _winapi.WaitForSingleObject(self._handle,
                    remaining_ms)
                if result == _winapi.WAIT_TIMEOUT:
                    raise TimeoutExpired(self.args, timeout)
                exitcode = _winapi.GetExitCodeProcess(self._handle)
                self._set_returncode(exitcode)
        finally:
            self._status_lock.release()
        return self.returncode

[efiop]

@eryksun It seems like we could re-use _waitpid_lock(it was requested for removal in #14360 (comment)) instead of introducing new _status_lock, right? They seem to have pretty similar meaning to me, but maybe I'm not quite getting the difference.

[vstinner]

The Unix code path has the same issue and it was already been fixed. I suggest to factorize the code between Windows and Unix: reuse the same lock and the same locking code, but add sub-methods for Windows or Unix specific code.

I plan to write such PR.

[eryksun]

The Unix code path has the same issue and it was already been fixed. I suggest to factorize the code between Windows and Unix: reuse the same lock and the same locking code, but add sub-methods for Windows or Unix specific code.

The POSIX code uses a busy loop. I see no reason for that in Windows.

[vstinner]

My PR doesn't work: more locking is needed. Code should be shared with Unix, or copied from the Unix implementation, to add locking. I will not be available next weeks, so I prefer to close the issue. I may rewrite it properly later.