The code to do this is in the CPython header this is cribbed from. It would be dead code if I included it.

Move the release before the error check, or we will dead-lock on error!

Good catch! Also, the obj == NULL check doesn't make sense -- it would need to be *obj == NULL to check the result.

But I think it would be better not to hold the lock around the npy_import:

• Imports in general can run arbitrary code, which may reentrantly trigger npy_cache_import_runtime.
• Imports (i.e., PyImport_ImportModule) are already thread-safe internally

So I'd suggest writing this as:

if (!npy_atomic_load_ptr(obj)) {
    PyObject *value = npy_import(module, attr);
    if (value == NULL) {
        return -1;
    }
    PyThread_acquire_lock(npy_runtime_imports.import_mutex, WAIT_LOCK);
    if (!npy_atomic_load_ptr(obj)) {
        npy_atomic_store_ptr(obj, Py_NewRef(value));
    }
    PyThread_release_lock(npy_runtime_imports.import_mutex);
    Py_DECREF(value);
}

Or you can get rid of the lock if you implement compare-exchange:

if (!npy_atomic_load_ptr(obj)) {
    PyObject *value = npy_import(module, attr);
    if (value == NULL) {
        return -1;
    }
    PyObject *exepected = NULL;
    if (!npy_atomic_compare_exchange_ptr(obj, &expected, value)) {
        Py_DECREF(value);
    }
}

Thanks! I kept the lock since the next iteration will use PyMutex and I'd like to make sure this gets in for NumPy 2.1. Thankfully deadsnakes was updated over the weekend so we can test against a version with a public PyMutex now.