Regression in multiprocessing example using venv on Windows in 3.11rc2

[edge-python]

I'm programming in PyCharm in Virtual Environment on Win11 on a 12900k

I use 2 modules:
• mp_tst_all.py
• mp_tst_a.py

mp_tst_all.py

import multiprocessing
import platform

import mp_tst_a


def run_my_multi():
    p_xyz = multiprocessing.Process(
        target=mp_tst_a.run_main,
        args=())

    # p_... = ...

    p_xyz.start()
    # p_... .start()

    p_xyz.join()
    # p_... .join()

    return


if __name__ == '__main__':

    print('py:', platform.python_version())

    run_my_multi()

mp_tst_a.py

import multiprocessing
import time
import platform

main_queue = multiprocessing.Queue()


def load_directory(def_queue):

    content_def = def_queue.get()
    content_def['load_directory_is_alive'] = True
    def_queue.put(content_def)

    # ---- do something ... --------
    print('\n' + 'LOAD DIRECTORY:')
    for x in range(3):
        time.sleep(.6)
        print('do something ...')
    print('\n' + 'LOAD ok' + '\n')


    content_def = def_queue.get()
    content_def['load_directory_is_alive'] = False
    content_def['run_load_file'] = True
    def_queue.put(content_def)

    return


def load_files(def_queue):

    content_def = def_queue.get()
    run_load_file_i = content_def['run_load_file']
    def_queue.put(content_def)

    # ---- do something ... --------
    print('LOAD FILE:')
    if run_load_file_i:
        for x in range(3):
            time.sleep(.6)
            print('FILE ' + str(x))
    print('\n' + 'LOAD ok' + '\n')

    return


def run_main():

    global main_queue

    content = {
        'load_directory_is_alive': None,
        'run_load_file': None,
        'last_download_date': None
        }

    main_queue.put(content)

    rest = 0.2  # for test, try different values
    time.sleep(rest)

    jobs = []

    content = main_queue.get()
    content['last_download_date'] = 'xx.xx.xxxx'
    main_queue.put(content)
    time.sleep(rest)


    # ---- process_1.1: load_directory -----------------------------------
    process_1_1 = multiprocessing.Process(
        target=load_directory, name='load_directory',
        args=(main_queue, ))

    jobs.append(process_1_1)
    process_1_1.start()
    time.sleep(rest)


    # ---- do not begin process_2 before process_1.1: load_directory exits
    while process_1_1.is_alive():
        pass
        time.sleep(rest)


    # ---- process_2: load_files -----------------------------------------
    process_2 = multiprocessing.Process(
        target=load_files, name='load_files',
        args=(main_queue, ))

    jobs.append(process_2)
    process_2.start()


    process_1_1.join()
    process_2.join()

    return


if __name__ == "__main__":

    print('py:', platform.python_version())
    run_main()

Bug

first: with PyCharm

mp_tst_a.py works in py 3.11.0_rc2 and in py 3.10(.7)

but

when mp_tst_a.py is called from mp_tst_all.py there is an error in py 3.11.0_rc2,
however, in py 3.10(.7) it works

the error message varies depending on the 'rest' variable (mp_tst_a.py -> def run_main)

Process load_directory:
Traceback (most recent call last):
  File "C:\...\Python311\Lib\multiprocessing\process.py", line 314, in _bootstrap
    self.run()
  File "C:\...Python311\Lib\multiprocessing\process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File "C:\...\mp_test_single_a.py", line 10, in load_directory
    content_def = def_queue.get()
                  ^^^^^^^^^^^^^^^
  File "C:\...\Python311\Lib\multiprocessing\queues.py", line 102, in get
    with self._rlock:
  File "C:\...\Python311\Lib\multiprocessing\synchronize.py", line 95, in __enter__
    return self._semlock.__enter__()
           ^^^^^^^^^^^^^^^^^^^^^^^^^
**PermissionError: [WinError 5]** Zugriff verweigert

Process load_directory:
Traceback (most recent call last):
  File "C:\...\Python311\Lib\multiprocessing\process.py", line 314, in _bootstrap
    self.run()
  File "C:\...\Python311\Lib\multiprocessing\process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File "C:\...\mp_test_single_a.py", line 10, in load_directory
    content_def = def_queue.get()
                  ^^^^^^^^^^^^^^^
  File "C:\...\Python311\Lib\multiprocessing\queues.py", line 102, in get
    with self._rlock:
  File "C:\...\Python311\Lib\multiprocessing\synchronize.py", line 98, in __exit__
    return self._semlock.__exit__(*args)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
**OSError: [WinError 6]** Das Handle ist ungültig

secondly: in Command Prompt Window (cmd) it runs:

C:\...\Python311\Lib\site-packages>mp_tst_all.py
py: 3.11.0rc2

LOAD DIRECTORY:
do something ...
do something ...
do something ...

LOAD ok

LOAD FILE:
FILE 0
FILE 1
FILE 2

LOAD ok


C:\...\Python311\Lib\site-packages>

is it a bug in 3.11.0_rc2 or do i need to change something in the code ?
Thank you

[eryksun]

There is a bug with using multiprocessing in a virtual environment, which I presume applies to development under PyCharm.

Under a virtual environment, multiprocessing bypasses the venv launcher to instead directly run sys._base_executable, and it sets the "__PYVENV_LAUNCHER__" environment variable in order to propagate the virtual environment. It bypasses the launcher because the addition of a process in between the parent and the worker breaks the basic design of muliprocessing, which relies on manual duplication of handles from parent to child and vice versa.

However, multiprocessing leaves the path of the venv launcher in the command line, which hasn't been a problem until 3.11. With the new implementation of path initialization in 3.11, the value of sys._base_executable is now based on the command line if the parsed argv[0] path contains one or more backslashes ¹. Of course this breaks spawning the next generation of worker processes.

A high-level fix in Popen.__init__(), when running in a virtual environment, would be to modify the result from spawn.get_command_line() to replace the first item with sys._base_executable. (The venv launcher itself does this, in terms of modifying the command line of the launched process.) Current source:

cpython/Lib/multiprocessing/popen_spawn_win32.py

Lines 55 to 68 in 5fe0431

	cmd = spawn.get_command_line(parent_pid=os.getpid(),
	pipe_handle=rhandle)
	cmd = ' '.join('"%s"' % x for x in cmd)
	python_exe = spawn.get_executable()
	# bpo-35797: When running in a venv, we bypass the redirect
	# executor and launch our base Python.
	if WINENV and _path_eq(python_exe, sys.executable):
	python_exe = sys._base_executable
	env = os.environ.copy()
	env["__PYVENV_LAUNCHER__"] = sys.executable
	else:
	env = None

Footnotes

1. In previous versions, sys._base_executable is always based on the process image path, from GetModuleFileNameW(NULL, ...). The new behavior allows the command line to override this if the command is a qualified path that contains at least one backslash (not forward slash). Refer to the source in Modules/getpath.py. ↩

[edge-python]

Many Thanks,

I can confirm that using the System Interpreter in PyCharm (without Virtual Environment) py 3.11.0_rc2 works correctly.
(but not in a virtual environment)

[gvanrossum]

@zooba Can you have a look at this release blocker?

[zooba]

A high-level fix in Popen.__init__(), when running in a virtual environment, would be to modify the result from spawn.get_command_line() to replace the first item with sys._base_executable. (The venv launcher itself does this, in terms of modifying the command line of the launched process.)

This sounds like the right fix here. It should be easy to test (in case someone else gets to it before I can):

if WINENV and _path_eq(python_exe, sys.executable): 
    python_exe = sys._base_executable 
    env = os.environ.copy() 
    env["__PYVENV_LAUNCHER__"] = sys.executable 
    cmd[0] = python_exe
else:
    env = None

cmd = ' '.join('"%s"' % x for x in cmd)   # moved down from earlier