You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
fix(cuda.core): make IPC imports wait for a pending release of the same descriptor - #3450
deviceptr_import_ipc() shares one handle per descriptor, because the driver's first free of an imported pointer unmaps every import of it. The deleter took ipc_import_mutex only after the handle's last reference was gone. In that gap, an import on another thread could:
find the cache entry expired,
re-import the still-mapped pointer, and
register a new entry.
The deleter then erased the new entry and freed the pointer, which unmapped it under the new Buffer.
Fix (memory.cpp):
The cache entry outlives the last reference. Only its own deleter erases it, under the mutex, together with the free.
An import that finds an expired entry waits on a condition variable until that release is done, then imports fresh.
cleanup_in_context always calls the after_cleanup hook, and the hook notifies the waiters. Every exit path therefore wakes them, including activation and driver failures.
Validation
Test
Without the fix
With the fix
Deterministic window test
Fails 50/50
Passes 50/50
8-thread stress test
Re-imports a mapped pointer every run
0 failures, plain and TSan builds, no TSan reports
New testtests/memory_ipc/test_ipc_import_release_race.py: threads in one child import, use and close one descriptor in a loop. It catches a hang (missed wakeup) and a lost mapping (failed copy). On real hardware the race did not trigger on the unpatched build either (about 6M cycles on an RTX 5880 Ada, driver 595.99), so this test guards the new waiting path rather than reproducing the original bug.
Performance: no measurable regression. Each cell is the median ns per import, n=20, before → after, measured on the final commit. The GPU runs alternate between builds.
…me descriptor
deviceptr_import_ipc() shares one handle per descriptor because the
driver's first free of an imported pointer unmaps every import of it
(nvbug 5570902). The handle's deleter took ipc_import_mutex only after
the last reference was gone. In that gap an import found the cache entry
expired, re-imported the still-mapped pointer and registered a new entry.
The deleter then erased that new entry and freed the pointer, which
unmapped it under the new Buffer.
The cache entry now outlives the last reference and only its own deleter
erases it, under the mutex, together with the free. An import that finds
an expired entry waits on a condition variable until that release is
done, then imports fresh.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reviewing files that changed from the base of the PR and between 43f6ea4 and 9711882.
You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.
Use the checkbox below for a quick retry:
Trigger review
Summary
Summary by CodeRabbit
Bug Fixes
Improved reliability when repeated imports of the same shared GPU memory descriptor overlap with the final handle release. Imports now share a handle, and any import arriving during release waits for it to finish, helping prevent hangs and invalid memory mappings.
Tests
Added stress coverage for concurrent imports, data copies, and releases, including checks for stalled threads and processes.
The IPC import cache now coordinates imports with release of the last handle. A new multiprocessing stress test exercises concurrent imports and releases.
The comments document shared handles and waiting for an in-progress free. The cache tracks weak handles and their owning boxes. Importers wait on expired entries until the deleter frees the mapping, erases the entry, and notifies waiters.
Concurrent import and release stress test cuda_core/tests/memory_ipc/test_ipc_import_release_race.py
A spawned child runs four workers that repeatedly import the descriptor, copy data, and close buffers. The parent checks worker completion, child exit status, and join timeout.
Change: Bug fix
Merge Risk: Low · previous review 43f6e
The new IPC import/release synchronization has no established production defect. A new stress test could leave a child process running when the test exits abnormally. This is a small test-hygiene fix and is safe to follow up on.
Security architecture risk: Low · up to 43f6e
The change closes a mapping-lifetime race without expanding access permissions. A narrow reentrant-warning path may still stall final release and subsequent imports of the same descriptor.
Retained concerns
Low · reliability · inferred: The new wait protocol can self-deadlock if final release on a different host thread emits the per-thread-default-stream warning and a synchronous warning handler imports the same descriptor. Context resolution invokes the warning before the deleter removes its expired cache entry; the nested import waits for that removal, while the deleter cannot proceed until the handler returns. This can strand mapping cleanup and other importers of that descriptor. It requires application-side reentrancy and has not been reproduced at runtime.
Security review details
Security Blast Radius
inferred — The directly affected ownership domain is callers sharing a descriptor in one importing process. Cache coordination does not add privileges or create a new cross-service entrypoint. The inferred reentrancy stall affects the releasing call and imports waiting on that descriptor; no broader tenant or environment exposure is established.
Trust Boundaries and Controls
observed — The new test's descriptor and allocation size originate from its own parent allocation and travel over a queue to its own child. Its process-kill helper receives that child process object. These routed test entrypoints do not introduce an external descriptor source or general process-control interface.
Resilience and Maintainability Implications
observed — Context activation and driver failures still reach notification and unlock, preventing those completed failure paths from leaving waiters asleep. Cleanup can nevertheless leak the mapping on failure, as before; notification is a progress mechanism rather than a successful-cleanup guarantee.
- [ ] Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts
Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.
MSVC 14.44 cannot deduce std::unique_lock's template argument for a lock
declared inside the deleter lambda and used from the nested after_cleanup
lambda (C2955, C2662). The import path keeps CTAD.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Wrap the child-process checks in try/finally so a failed assertion no
longer leaks the exported buffer (CodeRabbit review note).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.
The reason will be displayed to describe this comment to others. Learn more.
Minor · Stability & Availability · Quick win
important: Stop the child before closing the exported buffer.
If queue population raises or process.join() receives KeyboardInterrupt, execution skips kill_subprocesses. This finally then closes the exported buffer while the child can remain blocked or continue using its imports. Terminate and join any started child in finally before closing buffer.
As per coding guidelines, use “track_child_processes for fixture cleanup so a stuck child cannot block IPC resource teardown.”
The reason will be displayed to describe this comment to others. Learn more.
Approving. I went through the fix and found no problem with it: an import can't get past an expired entry until that entry's own deleter has erased it and freed the pointer, and every exit path wakes the waiters. Thanks for finding this.
…me descriptor (NVIDIA#3450)
* fix(cuda.core): make IPC imports wait for a pending release of the same descriptor
deviceptr_import_ipc() shares one handle per descriptor because the
driver's first free of an imported pointer unmaps every import of it
(nvbug 5570902). The handle's deleter took ipc_import_mutex only after
the last reference was gone. In that gap an import found the cache entry
expired, re-imported the still-mapped pointer and registered a new entry.
The deleter then erased that new entry and freed the pointer, which
unmapped it under the new Buffer.
The cache entry now outlives the last reference and only its own deleter
erases it, under the mutex, together with the free. An import that finds
an expired entry waits on a condition variable until that release is
done, then imports fresh.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Use CTAD for the ipc_import_mutex locks
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 2f1185e)
…ce fix (#3479)
* docs(cuda.core): add 1.3.0 release note for the IPC import/release race fix
Follow-up to #3450, which merged without a release note.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(cuda.core): scope the IPC race release note to memory pool buffers
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Andy-Jost
pushed a commit
to Andy-Jost/cuda-python
that referenced
this pull request
Nov 17, 2026
…ce fix (NVIDIA#3479)
* docs(cuda.core): add 1.3.0 release note for the IPC import/release race fix
Follow-up to NVIDIA#3450, which merged without a release note.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(cuda.core): scope the IPC race release note to memory pool buffers
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit de4c60d)
* [doc-only] docs(cuda.core): fix the floor-move wording; advise against pinning cuda-bindings (#3441)
* docs(cuda.core): drop the Breaking Changes section name from the floor policy
The release notes no longer have a Breaking Changes section for a floor
move; 1.3.0 lists it under Highlights. Say only that the release notes
list every move.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(cuda.core): advise against pinning cuda-bindings next to cuda-core
Users who pin cuda-bindings themselves can conflict with the floor of a
later cuda-core release. Tell them to depend on cuda-core with the cu12
or cu13 extra and let cuda-core choose the cuda-bindings version.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 83a3422)
* Mark cuda-gdb source-line test thread_unsafe (#3433)
Signed-off-by: Ralf Juengling <rjuengling@nvidia.com>
(cherry picked from commit 1a22eec)
* fix(cuda.core): serialize pinned-memory kernel-access test on Windows (#3436)
* fix(cuda.core): serialize pinned-memory kernel-access test on Windows
closes#3426
Add @thread_unsafe_on_windows to the pinned-memory kernel-access test.
The test accesses host memory while parallel workers may execute GPU
work; synchronizing its own stream does not coordinate those workers.
This follows the guard used by similar tests on Windows and WSL.
Windows CI is needed to confirm it resolves the reported crashes.
The separate OOM failures remain under investigation.
Validation: Ruff and git diff --check pass.
* fix(cuda.core): report CUDA synchronization errors during test teardown
Check the cuCtxSynchronize result in init_cuda so asynchronous errors
surface during the triggering test's teardown instead of going unnoticed
until a later test. Always attempt context cleanup via finally.
Related to #3426.
* fix(cuda.core): synchronize the bound context during test teardown
Use device.sync() because the test wrapper may pop the current context
before init_cuda teardown. This avoids spurious INVALID_CONTEXT errors
while preserving synchronization error reporting and context cleanup.
Add regression coverage for a missing current context and cleanup after
synchronization failures.
Related to #3426.
* fix(cuda.core): serialize memory-pool watermark reset tests
Mark test_mempool_watermark_reset thread_unsafe, matching test_mempool_trim.
Both exercise the pool teardown and synchronization sequence associated
with the suspected driver race in #3350.
* fix(cuda.core): bypass access revocation when destroying host pools
Skip peer-access revocation for locally created HOST and HOST_NUMA pools
to investigate pinned-pool teardown OOMs and crashes. Continue destroying
the pools and preserve cleanup for device, managed, and imported pools.
This is a diagnostic workaround pending Windows CI validation.
Related to #3426.
* fix(cuda.core): preserve the first exception during fixture cleanup
Keep setup or synchronization failures primary and attach subsequent
cleanup failures as exception notes. Always attempt to pop the context.
Add regression coverage for combined failures, including interrupts.
* fix(cuda.core): limit host-pool cleanup workaround to pinned allocations
HOST and HOST_NUMA also describe managed-pool preferences. Check the allocation type so the pinned-memory workaround preserves managed-pool peer-access revocation.
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
* test(cuda.core): expose pinned allocation OOM and retain its traceback
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
* ci: diagnose PR3436 pinned pool teardown on Windows
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
* ci: move watermark investigation out of PR3436
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
* test(cuda.core): restore expected pinned allocation OOM handling
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
* revert(cuda.core): defer pinned host pool cleanup workaround
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
---------
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
Co-authored-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
(cherry picked from commit b715078)
* [no-ci] ci: exclude unavailable historical GitHub references (#3471)
* ci: exclude unavailable historical GitHub references
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
* ci: require manual copy-pr-bot approval
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
---------
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
(cherry picked from commit 3991a69)
* cuda.core: map ml_dtypes bfloat16 to kDLBfloat in DLPack export (#3359)
_smv_dtype_numpy_to_dlpack dispatched on numpy dtype kind, but
ml_dtypes.bfloat16 has kind "V", so bf16 views fell through to the
unsupported-dtype BufferError and could not be re-exported, even though
the DLPack import path already accepts kDLBfloat and _tensor_map.pyx
maps bfloat16. Detect bfloat16 up front and export it as kDLBfloat
(16 bits).
The new round-trip test goes through StridedMemoryView.from_dlpack so
it cannot skip on stacks where NumPy's own DLPack support for extended
dtypes is unavailable, unlike the existing parametrized round-trip
test's NumPy-consumer probe.
Signed-off-by: Sainath-Reddy7 <180950590+Sainath-Reddy7@users.noreply.github.com>
Co-authored-by: Sainath-Reddy7 <180950590+Sainath-Reddy7@users.noreply.github.com>
Co-authored-by: Sebastian Berg <sebastianb@nvidia.com>
(cherry picked from commit d1d1457)
* fix(cuda.core): make IPC imports wait for a pending release of the same descriptor (#3450)
* fix(cuda.core): make IPC imports wait for a pending release of the same descriptor
deviceptr_import_ipc() shares one handle per descriptor because the
driver's first free of an imported pointer unmaps every import of it
(nvbug 5570902). The handle's deleter took ipc_import_mutex only after
the last reference was gone. In that gap an import found the cache entry
expired, re-imported the still-mapped pointer and registered a new entry.
The deleter then erased that new entry and freed the pointer, which
unmapped it under the new Buffer.
The cache entry now outlives the last reference and only its own deleter
erases it, under the mutex, together with the free. An import that finds
an expired entry waits on a condition variable until that release is
done, then imports fresh.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Use CTAD for the ipc_import_mutex locks
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 2f1185e)
* test(cuda.core): skip HOST graph allocation before CUDA 13 (#3446)
(cherry picked from commit 11cc666)
* test(cuda.core): query green stream contexts with v2 (#3448)
Signed-off-by: Rui Luo <ruluo@nvidia.com>
(cherry picked from commit 7076a67)
* test(cuda.core): skip the VMM RDMA policy test on WSL (#3478)
GPUDirect RDMA is not supported on WSL, but some drivers report the
capability there, and the RDMA-enabled VMM allocation in
test_vmm_allocator_policy_configuration then fails. Skip the test on
WSL; the capability checks stay as they are.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 36bc530)
* docs(cuda.core): add 1.3.0 release note for the IPC import/release race fix (#3479)
* docs(cuda.core): add 1.3.0 release note for the IPC import/release race fix
Follow-up to #3450, which merged without a release note.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(cuda.core): scope the IPC race release note to memory pool buffers
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit de4c60d)
* Fix conditional graph lifetime tests to use supported nodes (#3469)
* Fix conditional graph lifetime tests to use supported nodes
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
* ci: require manual copy-pr-bot approval
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
---------
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
Co-authored-by: Andy Jost <ajost@nvidia.com>
(cherry picked from commit b80866b)
* fix(cuda.core): run cuMemPoolSetAccess in relaxed mode in PinnedMemoryResource (#3475)
Since #3405, PinnedMemoryResource.__init__ without options calls
cuMemPoolSetAccess to grant the current device read/write access to
the driver's default pool. The driver refuses that call, and
invalidates the capture, while a global-mode capture is active on any
thread. Users of PyTorch's torch.cuda.graph (which defaults to global
mode) therefore see CUDA_ERROR_STREAM_CAPTURE_UNSUPPORTED on
construction and CUDA_ERROR_STREAM_CAPTURE_INVALIDATED on end_building.
Fix: introduce MP_set_access in _memory_pool.pyx, wrapping
cuMemPoolSetAccess in the existing _enter_relaxed_capture_mode /
_exit_relaxed_capture_mode helpers, following the identical pattern
used by pool attribute reads and writes and trim() since #2834.
Call it from _PMR_init instead of the bare with-nogil block.
The with-options path (cuMemPoolCreate) is unaffected; that call is
still refused under global capture as documented in the 1.3.0 notes.
Add tests for the three capture combinations (global+same_thread,
global+other_thread, thread_local+same_thread) and a capture-mode
restore check modelled on test_dmr_init_restores_capture_mode.
Closes#3451
Co-authored-by: Andy Jost <ajost@nvidia.com>
(cherry picked from commit d94f19c)
* cuda.core: carry the cuda-bindings floors through minimal-sysctk12/13 extras (#3460)
* cuda.core: carry the cuda-bindings floors through minimal-sysctk extras
The floors lived only in the cu12/cu13 extras, so a package that depends
on plain cuda-core never passed them to pip or uv, and an older
cuda-bindings failed only at import (#3459).
Declare the floors in new minimal-sysctk12/13 extras, include them from
cu12/cu13, and depend on cuda-bindings without a version. The floor
parser, the CI wheel-floor reader, the pre-commit hook and the docs read
the new extras.
* cuda.core: read cu<N> floors through the minimal-sysctk self-reference
Drop the unversioned cuda-bindings from the base dependencies, so a
plain cuda-core install keeps asking for an extra.
Keep the floor readers on the cu<N> extras, as before, and teach them
to follow a cuda-core[...] self-reference and to skip an unversioned
cuda-bindings[all]. The cu<N> layout then yields the floor written in
minimal-sysctk<N>, and the existing tests of the cu<N> form stay as
they are. New tests cover the self-reference.
---------
Co-authored-by: Andy Jost <ajost@nvidia.com>
(cherry picked from commit ad38141)
---------
Signed-off-by: Ralf Juengling <rjuengling@nvidia.com>
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
Signed-off-by: Sainath-Reddy7 <180950590+Sainath-Reddy7@users.noreply.github.com>
Signed-off-by: Rui Luo <ruluo@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Ralf Juengling <rjuengling@nvidia.com>
Co-authored-by: jpascucci-nv <jpascucci@nvidia.com>
Co-authored-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
Co-authored-by: Ralf W. Grosse-Kunstleve <rwgkio@gmail.com>
Co-authored-by: B Sainath Reddy <bsainathreddy4@gmail.com>
Co-authored-by: Sainath-Reddy7 <180950590+Sainath-Reddy7@users.noreply.github.com>
Co-authored-by: Sebastian Berg <sebastianb@nvidia.com>
Co-authored-by: Rob Parolin <rparolin@users.noreply.github.com>
Co-authored-by: Rui Luo <ruluo@nvidia.com>
Co-authored-by: Leo Fang <leof@nvidia.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
bugSomething isn't workingcuda.coreEverything related to the cuda.core module
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
deviceptr_import_ipc()shares one handle per descriptor, because the driver's first free of an imported pointer unmaps every import of it. The deleter tookipc_import_mutexonly after the handle's last reference was gone. In that gap, an import on another thread could:The deleter then erased the new entry and freed the pointer, which unmapped it under the new
Buffer.Fix (
memory.cpp):cleanup_in_contextalways calls theafter_cleanuphook, and the hook notifies the waiters. Every exit path therefore wakes them, including activation and driver failures.Validation
tests/memory_ipc/test_ipc_import_release_race.py: threads in one child import, use and close one descriptor in a loop. It catches a hang (missed wakeup) and a lost mapping (failed copy). On real hardware the race did not trigger on the unpatched build either (about 6M cycles on an RTX 5880 Ada, driver 595.99), so this test guards the new waiting path rather than reproducing the original bug.Performance: no measurable regression. Each cell is the median ns per import, n=20, before → after, measured on the final commit. The GPU runs alternate between builds.
🤖 Generated with Claude Code