I can't make a reproducer for this.

I tried

async def socket_runner(port: int) -> None:
    loop = asyncio.get_running_loop()
    while True:
        sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
        sock.setblocking(False)
        task = asyncio.current_task()
        loop.call_later(0.1, task.cancel)
        try:
            await loop.sock_connect(sock, ("1.1.1.1", port))
        except asyncio.CancelledError:
            sock.close()

But the callback always fires in time.

I thinkthe problem is actually happening in _create_connection_transport

So if create_connection get cancelled, _create_connection_transport has already created the transport and protocol. The waiter gets cancelled at

I'll keep working on a reproducer so this issue is actually actionable.

I can't come up with a working reproducer for that either. I'm going to need more information for the original reporter.

I'll close this for now and reopen if I can get enough information to make a reproducer.

Sorry for the noise.

home-assistant/core#141855

related

Reproducer that works on 3.12 and 3.13

import asyncio
import socket
from contextlib import suppress


async def socket_runner() -> None:
    loop = asyncio.get_running_loop()
    while True:
        sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
        sock.setblocking(False)
        task = asyncio.create_task(
            loop.create_connection(asyncio.Protocol, None, None, sock=sock)
        )
        for _ in range(10):
            await asyncio.sleep(0)
        task.cancel()
        with suppress(asyncio.CancelledError):
            await task
        sock.close()


asyncio.run(socket_runner())

Will produce

RuntimeError: File descriptor <socket.socket fd=6, family=2, type=1, proto=0, laddr=('0.0.0.0', 0)> is used by transport <_SelectorSocketTransport fd=6 read=polling write=<idle, bufsize=0>>

After a few runs

If I take the close it looks like the socket leaks even after gc because something is holding a strong ref. Not that its a concern for cpython, but if I switch to uvloop, the exception doesn't happen and it just leaks instead.

Finished tracing it and its still not a good reproducer.

Can you see what refcycle says about why the socket is still alive? It should also be closed explicitly to avoid a ResourceWarning

Still looking into this one, if I come up with something useful, I will reopen