threading + multiprocessing + logging = 死锁 ?

前段时间有个程序突然出现了子进程不工作的情况。

后来通过调查发现是因为程序中同时使用了多线程，多进程以及 logging 模块，导致子进程中出现了死锁的情况。

当创建子进程的时候，后台线程中的 logging 模块正好获取了一个锁(threading.RLock)在记录日志信息。由于在 unix/linux 平台下 Python 是通过 fork 来创建子进程的，因此创建子进程的时候会把 logging 中的锁也复制了一份，当子进程中需要记录日志的时候发现 logging 的锁一直处于被占用的状态，从而出现了死锁（复制的这个锁永远也不会被释放，因为它的所有者是父进程的某个线程，但是这个线程释放锁的时候又不会影响子进程里的这个锁）。

复现问题的代码如下:

import os
import sys
import threading
import time


class ThreadWorker(threading.Thread):
    def __init__(self):
        print('ThreadWorker: init')
        super().__init__()

    def run(self):
        print('ThreadWorker: running (rlock = {0})'.format(global_rlock))

        global_rlock.acquire()
        print('ThreadWorker: i got lock {0}'.format(global_rlock))
        time.sleep(5)
        global_rlock.release()
        print('ThreadWorker: release lock {0} and '
              'sleeping forever'.format(global_rlock))

        time.sleep(600000)

global_rlock = threading.RLock(verbose=True)
worker = ThreadWorker()
worker.start()

time.sleep(1)
print('forking')
pid = os.fork()
if pid != 0:    # pid != 0 当前处于父进程
    print('parent: running (rlock = {0})'.format(global_rlock))
else:      # pid = 0 当前处于子进程
    print('child: running (rlock = {0}), '
          'getting the lock...'.format(global_rlock))
    global_rlock.acquire()
    print('child: got the lock {0}'.format(global_rlock))
    sys.exit(0)

time.sleep(10)

上面代码的执行结果如下:

$ python fork.py
ThreadWorker: init
ThreadWorker: running (rlock = <unlocked _thread.RLock object owner=0 count=0 at 0x10116cb40>)
ThreadWorker: i got lock <locked _thread.RLock object owner=123145307557888 count=1 at 0x10116cb40>
forking
parent: running (rlock = <locked _thread.RLock object owner=123145307557888 count=1 at 0x10116cb40>)
child: running (rlock = <locked _thread.RLock object owner=123145307557888 count=1 at 0x10116cb40>), getting the lock...
ThreadWorker: release lock <unlocked _thread.RLock object owner=0 count=0 at 0x10116cb40> and sleeping forever

从上面的结果中可以看出来：虽然线程随后释放了获得的锁，但是子进程却永远的卡在了获取锁的地方。

那么, 应该如何解决这个问题呢？至少有三种解决办法:

• 先创建子进程，然后再创建线程:

import os
import sys
import threading
import time


class ThreadWorker(threading.Thread):
    def __init__(self):
        print('ThreadWorker: init')
        super().__init__()

    def run(self):
        print('ThreadWorker: running (rlock = {0})'.format(global_rlock))

        global_rlock.acquire()
        print('ThreadWorker: i got lock {0}'.format(global_rlock))
        time.sleep(5)
        global_rlock.release()
        print('ThreadWorker: release lock {0} and '
              'sleeping forever'.format(global_rlock))

        time.sleep(600000)

global_rlock = threading.RLock(verbose=True)
worker = ThreadWorker()

print('forking')
pid = os.fork()
if pid != 0:    # pid != 0 当前处于父进程
    print('parent: running (rlock = {0})'.format(global_rlock))
    worker.start()
else:      # pid = 0 当前处于子进程
    time.sleep(1)
    print('child: running (rlock = {0}), '
          'getting the lock...'.format(global_rlock))
    global_rlock.acquire()
    print('child: got the lock {0}'.format(global_rlock))
    global_rlock.release()
    print('child: release the lock {0}'.format(global_rlock))
    sys.exit(0)

time.sleep(10)

结果:

$ python fork2.py
ThreadWorker: init
forking
parent: running (rlock = <unlocked _thread.RLock object owner=0 count=0 at 0x10f24cb70>)
ThreadWorker: running (rlock = <unlocked _thread.RLock object owner=0 count=0 at 0x10f24cb70>)
ThreadWorker: i got lock <locked _thread.RLock object owner=123145307557888 count=1 at 0x10f24cb70>
child: running (rlock = <unlocked _thread.RLock object owner=0 count=0 at 0x10f24cb70>), getting the lock...
child: got the lock <locked _thread.RLock object owner=140735162044416 count=1 at 0x10f24cb70>
child: release the lock <unlocked _thread.RLock object owner=0 count=0 at 0x10f24cb70>
ThreadWorker: release lock <unlocked _thread.RLock object owner=0 count=0 at 0x10f24cb70> and sleeping forever

可以看到子进程和线程都能够正常获取锁。

• 不要混合使用 threading, multiprocessing, logging/其他使用了线程锁的模块。 要么都是多线程，要么都是多进程。

• 另一个办法就是配置 logging 使用无锁的 handler 来记录日志信息。

参考资料

• PythonLoggingThreadingMultiprocessingIntermixedStudy(Using modules Python logging, threading and multiprocessing in a single application.)

• Issue 6721: Locks in the standard library should be sanitized on fork - Python tracker

• multithreading - Deadlock with logging multiprocess/multithread python script - Stack Overflow

• python - 使用multiprocessing.Process调用start方法后，有较小的几率子进程中run方法未执行 - SegmentFault

• python multiprocessing hanging, potential queue memory error? - Stack Overflow

• Threads and fork(): think twice before mixing them. | Linux Programming Blog

原文地址: https://mozillazg.com/2016/09...