Data race reading `_thread.RLock` recursion count in `repr()` under free-threading

[Naserume]

Bug report

Bug description:

This is a follow-up to #153292 (data race in repr() of _thread.RLock), which was fixed by making rlock_repr read the lock's owner with an atomic load. The fix covered the owner (self->lock.thread) field, but rlock_repr still reads self->lock.level with a plain (non-atomic) load to compute the recursion count:

cpython/Modules/_threadmodule.c

Lines 1287 to 1305 in 22a6c51

	static PyObject *
	rlock_repr(PyObject *op)
	{
	rlockobject *self = rlockobject_CAST(op);
	PyThread_ident_t owner = FT_ATOMIC_LOAD_ULLONG_RELAXED(self->lock.thread);
	int locked = rlock_locked_impl(self);
	size_t count;
	if (locked) {
	count = self->lock.level + 1;
	}
	else {
	count = 0;
	}
	return PyUnicode_FromFormat(
	"<%s %s object owner=%" PY_FORMAT_THREAD_IDENT_T " count=%zu at %p>",
	locked ? "locked" : "unlocked",
	Py_TYPE(self)->tp_name, owner,
	count, self);
	}

self->lock.level is written by acquire / release / _acquire_restore, e.g.:

cpython/Modules/_threadmodule.c

Lines 1198 to 1210 in 22a6c51

	_thread_RLock__acquire_restore_impl(rlockobject *self, PyObject *state)
	/*[clinic end generated code: output=beb8f2713a35e775 input=c8f2094fde059447]*/
	{
	PyThread_ident_t owner;
	Py_ssize_t count;
	if (!PyArg_Parse(state, "(n" Py_PARSE_THREAD_IDENT_T "):_acquire_restore",
	&count, &owner))
	return NULL;
	_PyRecursiveMutex_Lock(&self->lock);
	_Py_atomic_store_ullong_relaxed(&self->lock.thread, owner);
	self->lock.level = (size_t)count - 1;

So on a free-threaded build, repr(rlock) concurrent with acquire() or release() is still a data race, now on the level field rather than the owner field the earlier fix addressed.

Reproducer:

import _thread
from threading import Thread, Barrier

shared_rlock = _thread.RLock()
state = (5, _thread.get_ident())

def chain1_thread():
 for _ in range(20000):
 try:
 shared_rlock._acquire_restore(state)
 except Exception:
 pass

def chain2_thread():
 for _ in range(20000):
 try:
 repr(shared_rlock)
 except Exception:
 pass

N_C1 = 2
N_C2 = 4
barrier = Barrier(N_C1 + N_C2)

def _c1():
 barrier.wait()
 chain1_thread()

def _c2():
 barrier.wait()
 chain2_thread()

threads = [Thread(target=_c1) for _ in range(N_C1)]
threads += [Thread(target=_c2) for _ in range(N_C2)]
for t in threads: t.start()
for t in threads: t.join()

TSAN Report :

==================
WARNING: ThreadSanitizer: data race (pid=655652)
 Read of size 8 at 0x7fffb6610540 by thread T6:
 #0 rlock_repr /cpython/./Modules/_threadmodule.c:1295:28 
 #1 PyObject_Repr /cpython/Objects/object.c:784:11 
 #2 builtin_repr /cpython/Python/bltinmodule.c:2677:12 
 #3 _PyEval_EvalFrameDefault /cpython/Python/generated_cases.c.h:2712:35 

 Previous write of size 8 at 0x7fffb6610540 by thread T1:
 #0 _thread_RLock__acquire_restore_impl /cpython/./Modules/_threadmodule.c:1210:22 
 #1 _thread_RLock__acquire_restore /cpython/./Modules/clinic/_threadmodule.c.h:537:20 
 #2 method_vectorcall_O /cpython/Objects/descrobject.c:476:24 
 #3 _PyObject_VectorcallTstate /cpython/./Include/internal/pycore_call.h:144:11 
 #4 PyObject_Vectorcall /cpython/Objects/call.c:327:12 
 #5 _Py_VectorCallInstrumentation_StackRefSteal /cpython/Python/ceval.c:768:11 
 #6 _PyEval_EvalFrameDefault /cpython/Python/generated_cases.c.h:1906:35

SUMMARY: ThreadSanitizer: data race /cpython/./Modules/_threadmodule.c:1295:28 in rlock_repr
==================

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Linked PRs

• gh-154928: Fix data race on the recursion count of threading.RLock #155381

[deadlovelll]

hey @Naserume! i can try work on this