Fork in Python

The os.fork() function causes many implementation issues. It is supposed on most platforms, but Windows and VxWorks.

posix_spawn() (fork+exec)

First of all, it’s important to know how to avoid fork, especially to spawn child processes which execute a new program :-)

On Windows, spawning a process can be done using CreateProcess() which doesn’t use fork. On Unix, posix_spawn() can avoid fork on some platforms, and handles the dirty work for us on other platforms which implement it in userland.

Python 3.8 provides the os.posix_spawn() function. Not only posix_spawn() is safer than calling manully fork() + exec(), it can also be way faster in some cases.

The subprocess module can use os.posix_spawn() under some conditions:

  • close_fds is false;
  • preexec_fn, pass_fds, cwd and start_new_session parameters are not set;
  • the executable path contains a directory.

fork() (fork without exec)

The POSIX standard says only calling exec() is after after calling fork(). Calling any function different than exec() after fork() is unsafe.

Forking a process creates a child process which only has 1 thread: all other threads are destroyed. The whole memory is duplicated. For performance, usually memory is copied using “copy-on-write”: physical memory pages are marked as read-only and shared between the parent and the child process. When one process modify a memory page, the page is really duplicated and is no longer read-only.

See os.fork() function documentation.

Reinitialize all locks after fork

When fork is called, all threads are destroyed immediately except of the thread which called fork(). Locks can be in an inconsistent state. Using a lock after a fork can lead to a hang or to a crash.

The meta-issue bpo-6721 “Locks in the standard library should be sanitized on fork” tracks this problem.

The bpo-40089 added _PyThread_at_fork_reinit() function to reinitialize a lock to the unlocked state. It can be called after a fork to prevent the bug.

PyOS_AfterFork_Child()

Python has multiple functions called before and after fork:

  • PyOS_BeforeFork()
  • PyOS_AfterFork_Parent()
  • PyOS_AfterFork_Child()

The importlib lock is acquired before fork and released after fork.

The PyOS_AfterFork_Child() function is the most important: update Python internal states after fork in the child process.

  • Reset the GIL state
  • Reset threads
  • Reset importlib lock
  • Clear pending signals to prevent to handle the same signal “twice” (once in the parent, once in the child)
  • Delete Python thread states of other threads which have been destroyed.
  • etc.

os.register_at_fork()

The os.register_at_fork() added to Python 3.7 allows to register functions which will be called at fork:

  • before fork (in the parent process)
  • after fork in the parent process
  • after fork in the child process

macOS 10.14 (Mojave)

The multiprocessing started to crash randomly on macOS 10.14 (Mojave). The multiprocessing default start method changed from fork to spawn in Python 3.8: see bpo-33725.

getaddrinfo() and gethostbyname() locks

The C library provides getaddrinfo() and gethostbyname() functions which are not thread-safe on some platforms. Python uses an internal lock on platforms where these functions are known to not be thread-safe.

There are numerous articles about bugs caused by threads or caused by the lock added to make the function thread-safe. For example, Python didn’t reinitialize the getaddrinfo() lock at fork in the child process (the lock has been removed).

A thread-safe gethostbyname_r() function was added to avoid this issue.

The bpo-25920 removed the getaddrinfo() lock.

ssl.RAND_bytes()

“Applications must change the PRNG state of the parent process if they use any SSL feature with os.fork().” https://docs.python.org/dev/library/ssl.html#multi-processing