Data race: `sys.addaudithook()` lazily creates `interp->audit_hooks` with no lock

[sobolevn]

Bug report

Full report: https://gist.github.com/devdanzin/0b13838fd6089e73a3f063ed8f68e733

Repro:

import sys, threading
assert not sys._is_gil_enabled(), "run free-threaded: PYTHON_GIL=0"

NADD = 24      # threads slamming the first-time lazy-init store (write @540) at once
NAUD = 8       # threads spinning audit events (should_audit read @239)
barrier = threading.Barrier(NADD + NAUD)

def _hook(*a):
    return None

def adder():
    barrier.wait()
    for _ in range(200):
        sys.addaudithook(_hook)          # write interp->audit_hooks (first time) @540

def auditor():
    barrier.wait()
    for _ in range(200000):
        sys.audit("fusil.tsan.test")     # should_audit read of interp->audit_hooks @239

ts = [threading.Thread(target=adder) for _ in range(NADD)]
ts += [threading.Thread(target=auditor) for _ in range(NAUD)]
for t in ts: t.start()
for t in ts: t.join()
print("done, no crash")

The fix is incoming.

Linked PRs

• gh-154431: Fix data race in sys.audithook #154462
• [3.15] gh-154431: Fix data race in sys.audithook (GH-154462) #154477
• [3.14] gh-154431: Fix data race in sys.audithook (GH-154462) #154478
• gh-154431: Use existing runtime->audit_hooks.mutex instead of a new one #154494
• gh-154431: Eagerly initialize the audit hooks list #154554
• gh-154431: Revert "gh-154431: Fix data race in sys.audithook (#154462)" #154769
• [3.15] gh-154431: Revert "gh-154431: Fix data race in sys.audithook (GH-154462)" (GH-154769) #154774