Connecting a socket in timeout mode to a server with full listen backlog raises BlockingIOError

[wferi]

Bug report

Bug description:

I'm working on a Debian bookworm system. Please consider this executable script ms:

#!/usr/bin/python3 -i

from pathlib import Path
import socket

SOCK = 'stream.sock'
s = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
Path(SOCK).unlink(missing_ok=True)
s.bind(SOCK)
s.listen(0)

and this one mc:

#!/usr/bin/python3

import socket

SOCK = 'stream.sock'
s = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
s.settimeout(3)
s.connect(SOCK)

Let's start ./ms, and once it reaches the Python prompt, run ./mc twice in another terminal:

$ ./mc
$ ./mc
Traceback (most recent call last):
  File "/home/wferi/block/./mc", line 8, in <module>
    s.connect(SOCK)
BlockingIOError: [Errno 11] Resource temporarily unavailable

The first¹ invocation runs to completion all right, but all succeeding ones immediately fail with BlockingIOError, as if the connecting socket was in non-blocking mode, not in timeout mode. I understand that the documentation² notes

at the operating system level, sockets in timeout mode are internally set in non-blocking mode

so that must be where this comes from, but such an implementation detail shouldn't leak out of the abstraction, as the documentation also states that

A socket object can be in one of three modes: blocking, non-blocking, or timeout.

and

exception BlockingIOError: Raised when an operation would block on an object (e.g. socket) set for non-blocking operation. ³

CPython versions tested on:

3.11

Operating systems tested on:

Linux

Linked PRs

• gh-117208: handle EAGAIN in non-blocking connect #134688

Footnotes

1. In general the first backlog+1, but I set the listen backlog to 0 in the code of ms above. ↩

2. https://docs.python.org/3/library/socket.html#socket-timeouts ↩

3. https://docs.python.org/3/library/exceptions.html ↩

[duaneg]

I found this unreliable to reproduce initially, but after some tweaking managed to come up with a self-contained script that works on my system:

repro

from pathlib import Path
import socket
import sys
import threading

SOCK = 'stream.sock'

def open_socket():
    return socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)

def server(listening):
    s = open_socket()
    Path(SOCK).unlink(missing_ok=True)
    s.bind(SOCK)
    s.listen(0)
    listening.set()

def client(listening):
    s = open_socket()
    s.settimeout(3)
    listening.wait()
    s.connect(SOCK)

def run():
    listening = threading.Event()

    threads = (threading.Thread(target=server, args=(listening, )),
               threading.Thread(target=client, args=(listening, )),
               threading.Thread(target=client, args=(listening, )))

    for t in threads:
        t.start()

    for t in threads:
        t.join()

stopping = threading.Event()

def check_exception(args):
    if args.exc_type != ConnectionRefusedError:
        stopping.set()
        sys.excepthook(args.exc_type, args.exc_value, args.exc_traceback)

threading.excepthook = check_exception
while not stopping.is_set():
    run()

The exception is coming from internal_connect, where it calls connect. If there is a timeout and connect fails with EINPROGRESS, we wait. However, man 2 connect on Linux tells us:

EINPROGRESS: The socket is nonblocking and the connection cannot be completed immediately. (UNIX domain sockets failed with EAGAIN instead.)

That is what is happening in this case. Since it is a UNIX domain socket errno is set to EAGAIN not EINPROGRESS, so it doesn't try to wait, and immediately raises the error.