`test_lock_two_threads` flaky in `test_capi`

[colesbury]

Bug report

The lock may have the _Py_HAS_PARKED bit set so that m->_bits == 3 instead of 1. This can happen due to collisions in the parking lot hash table -- that's rare so we don't see it frequently. I think the possible collision here is if the PyEvent that the main thread is waiting on uses the same bucket as the PyMutex.

We should weaken the assertion in the test case.

See in https://buildbot.python.org/all/#/builders/1609/builds/2878/steps/6/logs/stdio:

test_lock_two_threads (test.test_capi.test_misc.Test_PyLock.test_lock_two_threads) ... python: ./Modules/_testinternalcapi/test_lock.c:60: lock_thread: Assertion `m->_bits == 1' failed.
Fatal Python error: Aborted


Current thread's C stack trace (most recent call first):
  Binary file "/home/buildbot/buildarea/3.x.itamaro-centos-aws.nogil/build/python", at _Py_DumpStack+0x31 [0x75b0f6]
  Binary file "/home/buildbot/buildarea/3.x.itamaro-centos-aws.nogil/build/python" [0x770ae8]
  Binary file "/home/buildbot/buildarea/3.x.itamaro-centos-aws.nogil/build/python" [0x770bd5]
  Binary file "/lib64/libc.so.6", at +0x3ebf0 [0x7f3a97a3ebf0]
  Binary file "/lib64/libc.so.6", at +0x8c1ec [0x7f3a97a8c1ec]
  Binary file "/lib64/libc.so.6", at raise+0x16 [0x7f3a97a3eb46]
  Binary file "/lib64/libc.so.6", at abort+0xd3 [0x7f3a97a28833]
  Binary file "/lib64/libc.so.6", at +0x2875b [0x7f3a97a2875b]
  Binary file "/lib64/libc.so.6", at +0x37886 [0x7f3a97a37886]
  Binary file "/home/buildbot/buildarea/3.x.itamaro-centos-aws.nogil/build/build/lib.linux-x86_64-3.15/_testinternalcapi.cpython-315td-x86_64-linux-gnu.so", at +0xe2cf [0x7f3a970612cf]
  Binary file "/home/buildbot/buildarea/3.x.itamaro-centos-aws.nogil/build/python" [0x756d51]
  Binary file "/lib64/libc.so.6", at +0x8a4aa [0x7f3a97a8a4aa]
  Binary file "/lib64/libc.so.6", at +0x10f510 [0x7f3a97b0f510]

cpython/Modules/_testinternalcapi/test_lock.c

Line 60 in cb39410

assert(m->_bits == 1);

cpython/Modules/_testinternalcapi/test_lock.c

Lines 52 to 98 in cb39410

	static void
	lock_thread(void *arg)
	{
	struct test_lock2_data *test_data = arg;
	PyMutex *m = &test_data->m;
	_Py_atomic_store_int(&test_data->started, 1);
	PyMutex_Lock(m);
	assert(m->_bits == 1);
	PyMutex_Unlock(m);
	assert(m->_bits == 0);
	_PyEvent_Notify(&test_data->done);
	}
	static PyObject *
	test_lock_two_threads(PyObject *self, PyObject *obj)
	{
	// lock attempt by two threads
	struct test_lock2_data test_data;
	memset(&test_data, 0, sizeof(test_data));
	PyMutex_Lock(&test_data.m);
	assert(test_data.m._bits == 1);
	PyThread_start_new_thread(lock_thread, &test_data);
	// wait up to two seconds for the lock_thread to attempt to lock "m"
	int iters = 0;
	uint8_t v;
	do {
	pysleep(10); // allow some time for the other thread to try to lock
	v = _Py_atomic_load_uint8_relaxed(&test_data.m._bits);
	assert(v == 1 || v == 3);
	iters++;
	} while (v != 3 && iters < 200);
	// both the "locked" and the "has parked" bits should be set
	assert(test_data.m._bits == 3);
	PyMutex_Unlock(&test_data.m);
	PyEvent_Wait(&test_data.done);
	assert(test_data.m._bits == 0);
	Py_RETURN_NONE;
	}

Linked PRs

• gh-135641: Fix flaky test_capi.test_lock_two_threads test case #135642