Fix cuda.bindings: lock lazy init of _cyb_func_ptrs in driver _inspect_function_pointer by lijinf2 · Pull Request #3406 · NVIDIA/cuda-python

lijinf2 · 2026-11-07T00:43:54Z

reference cybind MR#: 619

Summary

_inspect_function_pointer lazily builds a module-level _cyb_func_ptrs dict with an unsynchronized "check None, then assign". Under free-threaded Python several threads can each build and publish their own dict, and the replaced dict can be freed while another thread reads it. This is the SystemError: ... dictobject.c: bad argument to internal function seen in test_private_function_pointer_inspector on Python 3.14t.

The fix uses double-checked locking with a dedicated _cyb_func_ptrs_lock: only the first initialization takes the lock, and later lookups are lock-free reads of the published dict.

Notes

The _internal/driver_{linux,windows}.pyx files are regenerated from the cybind change (generator-side fix), not hand-edited.

🤖 Generated with Claude Code

copy-pr-bot · 2026-11-07T00:43:59Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

coderabbitai · 2026-11-07T00:48:24Z

No actionable comments were generated in the recent review.

Summary

Summary by CodeRabbit

Bug Fixes
- Improved the reliability of CUDA driver bindings on Linux and Windows when multiple threads access function pointers at the same time. Function pointers are now initialized once and consistently retrieved from the shared cache, reducing the chance of inconsistent behavior in multithreaded applications.

The Linux and Windows drivers now use a lock to initialize the cached function-pointer dictionary once. Both generated-file headers add a prerelease API warning and update the content hash.

Changes: Driver cache synchronization

Layer / File(s)	Summary
Lock function-pointer cache initialization `cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx`, `cuda_bindings_13/cuda/bindings/_internal/driver_windows.pyx`	Both drivers separate dictionary construction from cache access. They use a lock and a second cache check before storing the dictionary. The generated-file headers add a prerelease API warning and update the content hash.

Assessment against linked issues

Objective	Addressed	Explanation
Prevent `test_private_function_pointer_inspector` from failing under free-threaded Python on Windows [`#3373`]	✅

Suggested reviewers: mdboom

Change: Bug fix · Severity of issue fixed: Low

Merge Risk: Minimal · up to 45c27

The reported concurrent cache-initialization risk is addressed in both drivers. No actionable merge-blocking issue remains after normal checks.

Architecture risk: Low · up to 45c27

The change affects 1 system.

Changed systems: cuda_bindings_13

Architecture concerns
No architecture-level concerns identified.

Review details

Systems and components

observed — cuda_bindings_13 (service) was modified; 2 changed files map to changed impact.

Before / after behavior

observed — Modified behavior in cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx: Added a prerelease-API warning and updated the generated-file content hash.
observed — Modified behavior in cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx: Added _cyb_func_ptrs_lock to synchronize initialization of the cached function-pointer dictionary.
observed — Modified behavior in cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx: _build_function_pointers returns the constructed pointer dictionary without assigning it to the cache. _inspect_function_pointers now initializes that cache once using a lock and a second None check; _inspect_function_pointer delegates lookup to the cached accessor instead of initializing the cache itself.
observed — Modified behavior in cuda_bindings_13/cuda/bindings/_internal/driver_windows.pyx: Adds the prerelease-API warning and updates the generated-file content hash.

- [ ] Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Tip

🎯 Customize your risk assessments

Guide how CodeRabbit assesses merge risk and security architecture risk. Highlight sensitive changes and explain your security boundaries using Merge Risk Instructions and Security Risk Instructions in your review settings. Security architecture risk assessments require the Advanced plan or above.

✅ Automatic approvals, with your rules

Let CodeRabbit approve pull requests once required threads are resolved, pre-merge checks pass, and your risk limits are met. Enable Automatic approvals in review settings, with overrides for teams or file paths.

Recent review info

Run configuration

Source mode: Auto

Repository: NVIDIA/cuda-python

Branch: main

File: .coderabbit.yaml

Commit: db463bf

Layers: Repository YAML

Review profile: CHILL

Plan: Enterprise

Run ID: `e66e19cd-ff57-482f-a984-2a55dd2d7a9b`

Commits

Reviewing files that changed from the base of the PR and between `ab8ac53` and `45c27f2`.

Files selected for processing: 2

`cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx`

`cuda_bindings_13/cuda/bindings/_internal/driver_windows.pyx`

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

coderabbitai

Actionable comments posted: 1

Review info

Run configuration

Source mode: Auto

Repository: NVIDIA/cuda-python

Branch: main

File: .coderabbit.yaml

Commit: db463bf

Layers: Repository YAML

Review profile: CHILL

Plan: Enterprise

Run ID: `49bb1a46-3897-41f5-8de0-5feb3550b0a8`

Commits

Reviewing files that changed from the base of the PR and between `db463bf` and `ab8ac53`.

Files selected for processing: 2

`cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx`

`cuda_bindings_13/cuda/bindings/_internal/driver_windows.pyx`

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

coderabbitai · 2026-11-07T00:48:27Z

-        _cyb_func_ptrs = _inspect_function_pointers()
+        with _cyb_func_ptrs_lock:
+            if _cyb_func_ptrs is None:
+                _cyb_func_ptrs = _inspect_function_pointers()


Major · Stability & Availability · Heavy lift

Supported by static analysis

Script executed:

git diff 960be69abeb9b12ac91ce7604ee07cca803a401e ab8ac534c72cf441adb529951230336c0ea51f38 -- cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx rg -n -F -e '_inspect_function_pointers' -e '_inspect_function_pointer(' -e '_cyb_func_ptrs' -- cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx cuda_bindings_13/cuda/bindings/_internal/driver_windows.pyx cuda_bindings_13/tests cybind 2>/dev/null sed -n '1,90p;3740,3830p' cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx | nl -ba

Repository: NVIDIA/cuda-python

Length of output: 13432

Script executed:

printf '%s\n' '--- inspector start/end ---' nl -ba cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx | sed -n '2180,2235p;3775,3805p' printf '%s\n' '--- all plural-inspector references ---' rg -n -F --glob '!*.pyx' --glob '!*.pxd' -- '_inspect_function_pointers' . || test "$?" -eq 1 printf '%s\n' '--- generator and free-threading configuration ---' rg -n -F -e 'freethreading_compatible' -e 'free-threading' -e '_cyb_func_ptrs_lock' -e 'def _inspect_function_pointers' -- cuda_bindings_13 cybind pyproject.toml 2>/dev/null || test "$?" -eq 1 printf '%s\n' '--- test binding ---' nl -ba cuda_bindings_13/tests/test_cuda.py | sed -n '940,970p'

Repository: NVIDIA/cuda-python

Length of output: 8014

Script executed:

printf '%s\n' '--- cuda.core caller ---' nl -ba cuda_core/cuda/core/_cpp/rt/py_driver_fns.cpp | sed -n '170,250p' nl -ba cuda_core/cuda/core/_cpp/rt/driver_api.hpp | sed -n '165,205p' printf '%s\n' '--- generator locations ---' rg -n -F -e '_cyb_func_ptrs' -e '_inspect_function_pointers' -- cython_bindings cybind cuda_bindings_13 2>/dev/null | grep -v -E '/_internal/[^:]+\.pyx:' | sed -n '1,120p' printf '%s\n' '--- build directives ---' nl -ba cuda_bindings_13/build_hooks.py | sed -n '325,350p'

Repository: NVIDIA/cuda-python

Length of output: 7611

Script executed:

rg -n -i -e 'cython[>=<~ ]' -e 'cython_version' -- cuda_bindings_13/pyproject.toml cuda_bindings_13/build_hooks.py pyproject.toml 2>/dev/null || test "$?" -eq 1 rg --files | rg -i '(cybind|cython.binding|generator)' | sed -n '1,100p'

Repository: NVIDIA/cuda-python

Length of output: 452

Web query:

Cython free threading cdef global variables concurrent read write thread safety documentation

Result:

Sources consulted: - cython/docs/src/userguide/source_files_and_compilation.rst at master · cython/cython · GitHub — https://github.com/cython/cython/blob/master/docs/src/userguide/source_files_and_compilation.rst - Free threading — Cython 3.4.0a0 documentation — https://docs.cython.org/en/latest/src/userguide/freethreading.html - cython/docs/src/userguide/freethreading.rst at master · cython/cython · GitHub — https://github.com/cython/cython/blob/master/docs/src/userguide/freethreading.rst - …and 22 more

Web query:

Cython 3.2 free threading cdef global variable thread safety read write reference counting documentation

Result:

Sources consulted: - cython/docs/src/userguide/freethreading.rst at master · cython/cython · GitHub — https://github.com/cython/cython/blob/master/docs/src/userguide/freethreading.rst - Source Files and Compilation — Cython 3.3.0 documentation — https://cython.readthedocs.io/en/latest/src/userguide/source_files_and_compilation.html?highlight=cythonize - Free threading — Cython 3.3.0 documentation — https://cython.readthedocs.io/en/latest/src/userguide/freethreading.html - cython/docs/src/userguide/source_files_and_compilation.rst at master · cython/cython · GitHub — https://github.com/cython/cython/blob/master/docs/src/userguide/source_files_and_compilation.rst - …and 29 more

Script executed:

python3 - <<'PY' import importlib.metadata as metadata import importlib.util try: print("Cython version:", metadata.version("Cython")) spec = importlib.util.find_spec("Cython") print("Cython source root:", list(spec.submodule_search_locations) if spec else None) except metadata.PackageNotFoundError: print("Cython distribution unavailable") PY

Repository: NVIDIA/cuda-python

Length of output: 188

important: Synchronize both cache writers.

On free-threaded Python, a direct call to _inspect_function_pointers() can overlap the first _inspect_function_pointer() call and replace _cyb_func_ptrs without the lock. This can cause unsafe reference access or a crash.

Coordinate initialization in both entry points. Avoid acquiring the non-reentrant lock twice. Update the generator and regenerate both platform variants.

Good catch. This has been fixed. please review the fix.

Analysis chain

Script executed:

#!/bin/bash set -u echo '--- checked-out revision ---' git rev-parse HEAD echo '--- changes since reviewed commit ---' git diff --stat ab8ac534c72cf441adb529951230336c0ea51f38 HEAD echo '--- cache references in both variants ---' rg -n -C 4 '(^cdef (dict|object) _cyb_func_ptrs|^cdef dict _build_function_pointers|^cpdef (dict )?_inspect_function_pointers|^cpdef _inspect_function_pointer|_cyb_func_ptrs =)' cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx cuda_bindings_13/cuda/bindings/_internal/driver_windows.pyx echo '--- generated-source diff since reviewed commit ---' git diff ab8ac534c72cf441adb529951230336c0ea51f38 HEAD -- cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx cuda_bindings_13/cuda/bindings/_internal/driver_windows.pyx | head -180 echo '--- inspector tests and generator references ---' rg -n -C 3 'test_private_function_pointer_inspector|_inspect_function_pointers|_cyb_func_ptrs_lock' cuda_bindings_13/tests cybind -g '!*.pyx' -g '!*.pxd' 2>/dev/null | head -160 echo '--- relevant policy files ---' for f in AGENTS.md CLAUDE.md cuda_bindings_13/AGENTS.md cuda_bindings_13/CLAUDE.md; do if [ -f "$f" ]; then echo "--- $f ---" rg -n -i -C 3 'review|resolv|thread|generated|platform|cybind' "$f" || true fi done echo '--- cython dependency ---' rg -n -i 'cython[[:space:]]*[><=~]|freethreading_compatible' cuda_bindings_13/pyproject.toml cuda_bindings_13/build_hooks.py pyproject.toml 2>/dev/null || true

Length of output: 24893

web_search executed:

For Cython 3 free-threaded Python, are cdef module-level Python object variables automatically protected for concurrent reads and writes? Does a lock held by writers alone make an unlocked reader safe? Find relevant Cython and CPython documentation.

Length of output: 30297

Script executed:

#!/bin/bash set -u echo '--- nested repository instructions ---' fd -H -t f '^(AGENTS|CLAUDE)\.md$' cuda_bindings_13/cuda cuda_bindings_13/tests 2>/dev/null | while IFS= read -r f; do echo "--- $f ---" rg -n -i -C 3 'review|resolv|thread|generated|platform|cybind' "$f" || true done echo '--- installed Cython metadata ---' (cd /tmp && python3 - <<'PY' import importlib.metadata import importlib.util try: print("version:", importlib.metadata.version("Cython")) spec = importlib.util.find_spec("Cython") print("source:", list(spec.submodule_search_locations) if spec else None) except importlib.metadata.PackageNotFoundError: print("Cython not installed") PY )

Length of output: 946

web_search executed:

In Cython 3.2.5 free-threaded Python, are reads and assignments to cdef module-level Python object variables automatically atomic or synchronized? Find documentation or Cython source evidence for concurrent read/write safety and memory publication.

Length of output: 30296

Script executed:

#!/bin/bash set -euo pipefail tmp=$(mktemp -d) trap 'rm -rf "$tmp"' EXIT python3 -m pip install --disable-pip-version-check --quiet --target "$tmp/packages" 'Cython==3.2.5' cat > "$tmp/cache_example.pyx" <<'EOF' # cython: freethreading_compatible=True import threading cdef dict cache = None cdef object lock = threading.Lock() cdef dict build(): return {"key": 1} cpdef dict inspect_cache(): global cache if cache is None: with lock: if cache is None: cache = build() return cache EOF PYTHONPATH="$tmp/packages" python3 -m cython -3 -o "$tmp/cache_example.c" "$tmp/cache_example.pyx" echo '--- generated cache reads and writes ---' rg -n -C 3 '__pyx_v_13cache_example_cache|__pyx_f_13cache_example_inspect_cache' "$tmp/cache_example.c" | head -170

Length of output: 3882

@lijinf2 important: Both entry points now share one initializer, but the first read of _cyb_func_ptrs remains outside the lock. It can race with the assignment during initialization on free-threaded Python.

A standalone Cython 3.2.5 translation of this pattern emitted ordinary C pointer accesses. Please protect the read and assignment with the same lock, or use a safe acquire/release publication scheme.

Both platform variants have the same implementation. I could not inspect the generator source or run the repository tests. Please verify the generator change and run a concurrent first-call test for both entry points on Windows Python 3.14t.

Learnings used

Learnt from: CR Repo: NVIDIA/cuda-python PR: 0 File: cuda_bindings_13/AGENTS.md:0-0 Timestamp: 2026-10-20T22:59:48.070Z Learning: Applies to cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx : Source excerpt: ## Generated-source workflow - **Platform split files**: keep `_linux.pyx` and `_windows.pyx` variants aligned when behavior should be equivalent.

Learnt from: CR Repo: NVIDIA/cuda-python PR: 0 File: cuda_bindings_13/CLAUDE.md:0-0 Timestamp: 2026-10-31T04:59:49.354Z Learning: Applies to cuda_bindings_13/cuda/bindings/_internal/driver_linux.pyx : Source excerpt: ## Generated-source workflow - **Platform split files**: keep `_linux.pyx` and `_windows.pyx` variants aligned when behavior should be equivalent.

_{You are interacting with an AI system.}

lijinf2 · 2026-11-07T00:48:44Z

/ok to test ab8ac53

github-actions · 2026-11-07T01:03:30Z

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-3406/
https://nvidia.github.io/cuda-python/pr-preview/pr-3406/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-3406/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-3406/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

…_function_pointers() to a single build function having the fix

lijinf2 · 2026-11-07T02:07:19Z

/ok to test 45c27f2

lijinf2 added 3 commits November 6, 2026 22:23

add locking for _cyb_func_ptrs

8672ee4

use cybind gen

3f2a5c8

let read do not use lock

ab8ac53

github-actions Bot added the cuda.bindings Everything related to the cuda.bindings module label Nov 7, 2026

lijinf2 self-assigned this Nov 7, 2026

lijinf2 changed the title ~~Lock lazy init of _cyb_func_ptrs in driver _inspect_function_pointer~~ Fix cuda.bindings bug: lock lazy init of _cyb_func_ptrs in driver _inspect_function_pointer Nov 7, 2026

lijinf2 added the bug Something isn't working label Nov 7, 2026

lijinf2 added this to the cuda.bindings next milestone Nov 7, 2026

lijinf2 changed the title ~~Fix cuda.bindings bug: lock lazy init of _cyb_func_ptrs in driver _inspect_function_pointer~~ Fix cuda.bindings: lock lazy init of _cyb_func_ptrs in driver _inspect_function_pointer Nov 7, 2026

coderabbitai Bot reviewed Nov 7, 2026

View reviewed changes

reroute both entry point _inspect_function_pointer(name) and _inspect…

45c27f2

…_function_pointers() to a single build function having the fix

Navigation Menu

Repository navigation

Fix cuda.bindings: lock lazy init of _cyb_func_ptrs in driver _inspect_function_pointer - #3406

Fix cuda.bindings: lock lazy init of _cyb_func_ptrs in driver _inspect_function_pointer#3406
lijinf2 wants to merge 4 commits into
NVIDIA:mainfrom
lijinf2:fix-3373-func-ptrs-lock

lijinf2 commented Nov 7, 2026 •

edited

Loading

Uh oh!

copy-pr-bot Bot commented Nov 7, 2026

Uh oh!

coderabbitai Bot commented Nov 7, 2026 •

edited

Loading

Summary by CodeRabbit

Assessment against linked issues

Uh oh!

coderabbitai Bot left a comment

Uh oh!

coderabbitai Bot Nov 7, 2026 •

edited

Loading

Uh oh!

lijinf2 Nov 7, 2026

Uh oh!

coderabbitai Bot Nov 7, 2026

Uh oh!

lijinf2 commented Nov 7, 2026

Uh oh!

github-actions Bot commented Nov 7, 2026

Preview will be ready when the GitHub Pages deployment is complete.

Uh oh!

lijinf2 commented Nov 7, 2026

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

lijinf2 commented Nov 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Notes

Uh oh!

copy-pr-bot Bot commented Nov 7, 2026

Uh oh!

coderabbitai Bot commented Nov 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Assessment against linked issues

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

coderabbitai Bot Nov 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lijinf2 Nov 7, 2026

Choose a reason for hiding this comment

Uh oh!

coderabbitai Bot Nov 7, 2026

Choose a reason for hiding this comment

Uh oh!

lijinf2 commented Nov 7, 2026

Uh oh!

github-actions Bot commented Nov 7, 2026

Preview will be ready when the GitHub Pages deployment is complete.

Uh oh!

lijinf2 commented Nov 7, 2026

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

lijinf2 commented Nov 7, 2026 •

edited

Loading

coderabbitai Bot commented Nov 7, 2026 •

edited

Loading

coderabbitai Bot Nov 7, 2026 •

edited

Loading