macOS: Workers crash with NSCharacterSet fork-safety error after upgrading to 25.1.0

[Preston-Landers]

Type

Bug Report

Description

After upgrading from 25.0.3 to 25.1.0, forked workers crash immediately on macOS (Tahoe 26.3, M4 Pro) with the following error:

objc[19365]: +[NSCharacterSet initialize] may have been in progress in another thread when fork() was called.
objc[19365]: +[NSCharacterSet initialize] may have been in progress in another thread when fork() was called. We cannot safely call it or ignore it in the fork() child process. Crashing instead.

The crash occurs when the application has established a PostgreSQL connection (via psycopg2-binary) before workers are forked. The same application works fine under 25.0.3 with no other changes.

It does seem like setting --no-control-socket makes the problem go away.

Note that I'm using Python 3.14.3 from python-build-standalone: cpython-3.14.3+20260203-aarch64-apple-darwin-install_only_stripped.tar.gz

Steps to Reproduce (for bugs)

Save as test_gu.py:

import psycopg2

conn = psycopg2.connect(
    host="localhost", port=5432,
    dbname="postgres", user="postgres", password="postgres",
)
cur = conn.cursor()
cur.execute("SELECT 1")
print(f"DB connected: {cur.fetchone()}")
cur.close()

def app(environ, start_response):
    start_response("200 OK", [("Content-Type", "text/plain")])
    return [b"Hello, World!"]

Run with:

python -m gunicorn --bind 0.0.0.0:8080 --workers 2 --worker-class sync test_gu:app

Configuration

python -m gunicorn --bind 0.0.0.0:8080 --workers 2 --worker-class sync test_gu:app

Logs / Error Output

[2026-02-21 11:07:45 -0600] [21593] [INFO] Booting worker with pid: 21593
[2026-02-21 11:07:45 -0600] [20867] [ERROR] Worker (pid:21471) was sent SIGKILL! Perhaps out of memory?
[2026-02-21 11:07:45 -0600] [21599] [INFO] Booting worker with pid: 21599
objc[21593]: +[NSCharacterSet initialize] may have been in progress in another thread when fork() was called.
objc[21593]: +[NSCharacterSet initialize] may have been in progress in another thread when fork() was called. We cannot safely call it or ignore it in the fork() child process. Crashing instead. Set a breakpoint on objc_initializeAfterForkError to debug.

Gunicorn Version

gunicorn 25.1.0

Python Version

Python 3.14.3

Worker Class

sync (default)

Operating System

macOS Tahoe 26.3

Additional Context

Analysis / speculation:

The macOS ObjC runtime aborts forked child processes when it detects that an ObjC class was being initialized on another thread at the time of fork(). The error specifically mentions "in another thread," which suggests that new thread activity in the master process in 25.1.0 is overlapping with fork().

Based on testing, --no-control-socket appears to resolve the issue, suggesting the new control interface thread is the trigger. The control socket listener thread likely touches system frameworks that initialize ObjC classes, and when fork() occurs while that thread is active, the runtime kills the child.

psycopg2-binary directly links libssl and libcrypto, which on macOS trigger ObjC framework initialization (Security.framework / Foundation.framework). Interestingly, psycopg3-binary (psycopg[binary]) does NOT reproduce the issue - its C extension delegates SSL to libpq rather than linking libssl/libcrypto directly, which appears to change the initialization timing enough to avoid the conflict.

Setting OBJC_DISABLE_INITIALIZE_FORK_SAFETY=YES also suppresses the crash, but this masks the fork-safety violation rather than resolving it. It's worth noting that while macOS makes this violation fatal and visible, the same underlying condition (forking while threads are active) could potentially cause subtle issues on Linux as well - deadlocks from stale mutexes, etc. - that would be harder to detect.

Workarounds:

1. --no-control-socket
2. OBJC_DISABLE_INITIALIZE_FORK_SAFETY=YES (masks the symptom)
3. Pin to gunicorn==25.0.3

Checklist

• I have searched existing discussions and issues for duplicates
• I have checked the documentation and FAQ