[12.9.x] Manual backport of #3308: lock the global callback map

cuda_bindings/docs/source/release/12.X.Y-notes.rst

@@ -33,6 +33,8 @@ Highlights
 Bug fixes
 ---------
 
+* Fixed a crash on free-threaded Python when callbacks (for example ``cuLogsRegisterCallback``) are registered or
+  unregistered from multiple threads at the same time. Access to the internal callback registry is now synchronized.
 
 Miscellaneous
 -------------

cuda_bindings/tests/test_callback_threads.py

@@ -0,0 +1,41 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import threading
+
+import pytest
+
+from cuda.bindings import driver
+
+NUM_THREADS = 16
+NUM_ITERATIONS = 500
+
+
+@pytest.mark.agent_authored(model="claude-sonnet-5-5")
+def test_async_notification_register_unregister_from_many_threads():
+    (err,) = driver.cuInit(0)
+    assert err == driver.CUresult.CUDA_SUCCESS
+    err, device = driver.cuDeviceGet(0)
+    assert err == driver.CUresult.CUDA_SUCCESS
+
+    barrier = threading.Barrier(NUM_THREADS)
+    failures = []
+
+    def worker():
+        try:
+            barrier.wait()
+            for _ in range(NUM_ITERATIONS):
+                err, handle = driver.cuDeviceRegisterAsyncNotification(device, 0, 0)
+                assert err == driver.CUresult.CUDA_SUCCESS
+                (err,) = driver.cuDeviceUnregisterAsyncNotification(device, handle)
+                assert err == driver.CUresult.CUDA_SUCCESS
+        except BaseException as e:
+            failures.append(e)
+
+    threads = [threading.Thread(target=worker) for _ in range(NUM_THREADS)]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join()
+
+    assert not failures, failures