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

[deadlovelll]

Fix data race on the recursion count of threading.RLock

For more details see gh-154928

[deadlovelll]

I want to give a little disclaimer about my solution, maybe it would answer the possible questions during the review.

At the first attempt i tried to change type of _PyRecursiveMutex's level from size_t to Py_ssize_t. At first look that was nice solution, Py_ssize_t have atomics macro, but! size_t is unsigned type, Py_ssize_t is signed.

where it shoots into the knee:

import_threadr=_thread.RLock()
r._acquire_restore((0, _thread.get_ident()))
r.release()

When we execute this code with Py_ssize_t version we will get the assertion error:

Assertionfailed: (m->level==0), function_PyRecursiveMutex_TryUnlock,
filelock.c, line466.

Because in _thread_RLock__acquire_restore_implcount - 1 stores at the self->lock.level, our count is 0, so we got -1, size_t cant be -1 and it transforms to 18446744073709551615. In Py_ssize_t solution it will be -1

Then in _PyRecursiveMutex_TryUnlock in size_t solution we will end a the check if (m->level > 0), in Py_ssize_t we will go to the assert and then fail

int_PyRecursiveMutex_TryUnlock(_PyRecursiveMutex*m)
{
PyThread_ident_tthread=PyThread_get_thread_ident_ex();
if (!recursive_mutex_is_owned_by(m, thread)) {
return-1;
}
if (m->level>0) {
FT_ATOMIC_STORE_SIZE_RELAXED(m->level, m->level-1);
return0;
}
assert(m->level==0);
_Py_atomic_store_ullong_relaxed(&m->thread, 0);
PyMutex_Unlock(&m->mutex);
return0;
}

So in this fix I preferred to be conservative and keep the current semantics, despite the bigger diff.

My first version - deadlovelll@c2f316e