Skip to content

Fix crash when registering/unregistering callbacks concurrently under free-threaded Python #3308

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
lijinf2 wants to merge 2 commits into
base: main
Choose a base branch
from
+79 −11
Open

Fix crash when registering/unregistering callbacks concurrently under free-threaded Python #3308

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8e3dd5b
Lock the callback bookkeeping map in cuda.bindings
lijinf2 Oct 30, 2026
1ceee45
add a test case to increase probability triggering contention
lijinf2 Oct 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 18 additions & 7 deletions cuda_bindings_13/cuda/bindings/driver.pyx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,7 @@
# SPDX-License-Identifier: Apache-2.0

# This code was automatically generated with version 13.4.1. Do not modify it directly.
# CYTHON-BINDINGS-GENERATED-DO-NOT-MODIFY-THIS-FILE: format=1; content-sha256=348b02f0a76a3a271fbb11535f02d660d811d3c21a877e46bfc070bbabf43e76
# CYTHON-BINDINGS-GENERATED-DO-NOT-MODIFY-THIS-FILE: format=1; content-sha256=8932ee959b6a2f1e69316f813cb8e8492c8e112094812dc1310c49970f00779a
from typing import Any, Optional
import cython
import ctypes
Expand Down Expand Up @@ -30899,7 +30899,7 @@ def cuDeviceRegisterAsyncNotification(device, callbackFunc, userData):
if err != cydriver.CUDA_SUCCESS:
free(cbData)
else:
m_global._allocated[int(callback)] = cbData
m_global._insert(int(callback), cbData)
_helper_input_void_ptr_free(&cyuserDataHelper)
if err != cydriver.CUDA_SUCCESS:
return (_CUresult(err), None)
Expand Down Expand Up @@ -30948,8 +30948,7 @@ def cuDeviceUnregisterAsyncNotification(device, callback):
with nogil:
err = cydriver.cuDeviceUnregisterAsyncNotification(cydevice, cycallback)
if err == cydriver.CUDA_SUCCESS:
free(m_global._allocated[pcallback])
m_global._allocated.erase(<void_ptr>pcallback)
free(m_global._take(pcallback))
return (_CUresult(err),)

@cython.embedsignature(True)
Expand Down Expand Up @@ -55743,7 +55742,7 @@ def cuLogsRegisterCallback(callbackFunc, userData):
if err != cydriver.CUDA_SUCCESS:
free(cbData)
else:
m_global._allocated[int(callback_out)] = cbData
m_global._insert(int(callback_out), cbData)
_helper_input_void_ptr_free(&cyuserDataHelper)
if err != cydriver.CUDA_SUCCESS:
return (_CUresult(err), None)
Expand Down Expand Up @@ -55774,8 +55773,7 @@ def cuLogsUnregisterCallback(callback):
with nogil:
err = cydriver.cuLogsUnregisterCallback(cycallback)
if err == cydriver.CUDA_SUCCESS:
free(m_global._allocated[pcallback])
m_global._allocated.erase(<void_ptr>pcallback)
free(m_global._take(pcallback))
return (_CUresult(err),)

@cython.embedsignature(True)
Expand Down Expand Up @@ -57375,6 +57373,19 @@ def cuGraphicsVDPAURegisterOutputSurface(vdpSurface, unsigned int flags):

cdef class cudaBindingsDriverGlobal:
cdef map[void_ptr, void*] _allocated
cdef cython.pymutex _lock

cdef void _insert(self, void_ptr key, void* value) noexcept:
with self._lock:
self._allocated[key] = value

cdef void* _take(self, void_ptr key) noexcept:
cdef void* value = NULL
with self._lock:
if self._allocated.count(key) != 0:
value = self._allocated[key]
self._allocated.erase(key)
return value

def __dealloc__(self):
for item in self._allocated:
Expand Down
20 changes: 16 additions & 4 deletions cuda_bindings_13/cuda/bindings/runtime.pyx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,7 @@
# SPDX-License-Identifier: Apache-2.0

# This code was automatically generated with version 13.4.1. Do not modify it directly.
# CYTHON-BINDINGS-GENERATED-DO-NOT-MODIFY-THIS-FILE: format=1; content-sha256=ea95eb12334cbd4c1bf3fc8e890fb6f9493063b12faea853bb3d8c6288934182
# CYTHON-BINDINGS-GENERATED-DO-NOT-MODIFY-THIS-FILE: format=1; content-sha256=4cc5d08de16bc7cac781ba6e469ffe16b0d8e30beea5ec6f5e579f0fae24e834
from typing import Any, Optional
import cython
import ctypes
Expand Down Expand Up @@ -21228,7 +21228,7 @@ def cudaDeviceRegisterAsyncNotification(int device, callbackFunc, userData):
if err != cyruntime.cudaSuccess:
free(cbData)
else:
m_global._allocated[int(callback)] = cbData
m_global._insert(int(callback), cbData)
_helper_input_void_ptr_free(&cyuserDataHelper)

if err != cyruntime.cudaSuccess:
Expand Down Expand Up @@ -21270,8 +21270,7 @@ def cudaDeviceUnregisterAsyncNotification(int device, callback):
with nogil:
err = cyruntime.cudaDeviceUnregisterAsyncNotification(device, cycallback)
if err == cyruntime.cudaSuccess:
free(m_global._allocated[pcallback])
m_global._allocated.erase(<void_ptr>pcallback)
free(m_global._take(pcallback))
return (_cudaError_t(err),)

@cython.embedsignature(True)
Expand Down Expand Up @@ -40262,6 +40261,19 @@ def getLocalRuntimeVersion():

cdef class cudaBindingsRuntimeGlobal:
cdef map[void_ptr, void*] _allocated
cdef cython.pymutex _lock

cdef void _insert(self, void_ptr key, void* value) noexcept:
with self._lock:
self._allocated[key] = value

cdef void* _take(self, void_ptr key) noexcept:
cdef void* value = NULL
with self._lock:
if self._allocated.count(key) != 0:
value = self._allocated[key]
self._allocated.erase(key)
return value

def __dealloc__(self):
for item in self._allocated:
Expand Down
45 changes: 45 additions & 0 deletions cuda_bindings_13/tests/test_callback_threads.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,45 @@
# 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
from cuda_python_test_helpers import driver_version_less_than

NUM_THREADS = 16
NUM_ITERATIONS = 500


@pytest.mark.agent_authored(model="claude-sonnet-5-5")
@pytest.mark.skipif(
driver_version_less_than(12090),
reason="cuLogs* functions were introduced in CUDA 12.9",
)
@pytest.mark.thread_unsafe(reason="starts its own threads and needs them to overlap")
def test_culogs_register_unregister_from_many_threads():
(err,) = driver.cuInit(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.cuLogsRegisterCallback(0, None)

@coderabbitai coderabbitai Bot Oct 31, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Major · Stability & Availability · Quick win

important: Register a valid callback. If CUDA emits a log while a callback is registered, driver.pyx Line 51114–51117 calls this null pointer and crashes the pytest process. (raw.githubusercontent.com)

Register a no-op callback with the correct calling convention and keep it alive until the workers exit.

assert err == driver.CUresult.CUDA_SUCCESS
(err,) = driver.cuLogsUnregisterCallback(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
Loading