Race condition TypeError under py3.14t

[kevmo314]

Hi there, I'm getting a TypeError within five calls with the GIL disabled under free-threaded python on master. Here's a repro script:

# Run on free-threaded Python, e.g. PYTHON_GIL=0 python3.14t repro.py
import threading, time
import httpcore

conn = httpcore.HTTP11Connection(
    origin=httpcore.Origin(b"http", b"example.com", 80),
    stream=httpcore.MockStream([]),
    keepalive_expiry=5.0,
)

def reuse():
    # What handle_request() and _response_closed() do as the connection is reused.
    while True:
        conn._expire_at = None
        conn._expire_at = time.monotonic() + 5.0

threading.Thread(target=reuse, daemon=True).start()
while True:
    conn.has_expired()
# TypeError: '>' not supported between instances of 'float' and 'NoneType'

I created a PR to fix the issue here: kevmo314#1 however PRs and issues are restricted in this repo, let me know how I can contribute this fix appropriately.

[YCistak]

Nice find. The race is in has_expired() reading self._expire_at twice:

keepalive_expired = self._expire_at is not None and now > self._expire_at

Another thread can set it to None between the is not None check and the comparison. Without the GIL nothing prevents that interleaving. The minimal fix is to read the attribute once into a local:

expire_at = self._expire_at
keepalive_expired = expire_at is not None and now > expire_at

A few things to include when you turn this into a PR:

• The same pattern exists in http2.py (return self._expire_at is not None and now > self._expire_at), so fix both.
• Don't edit httpcore/_sync/* by hand. Those files are generated from httpcore/_async/* by scripts/unasync.py, and CI checks they match. Change _async/http11.py and _async/http2.py, then run scripts/unasync (or scripts/check) to regenerate the sync versions.

On the process: encode routes everything through Discussions first (the issue template chooser points here, and blank issues are disabled), and a maintainer converts it to an issue or invites a PR once they agree it's a bug. So this discussion is the right place. Posting the two-line diff above here, plus a link to your fork's branch, makes it easy for them to say "go ahead".