Fix std::terminate in ibverbs destructors on systems without RDMA hardware#500
Merged
Conversation
d4l3k
force-pushed
the
d4l3k/fix_ibv
branch
2 times, most recently
from
7de27ba to
7b0254f
Compare
…vice On CI runners without real RDMA hardware, rdma-core software providers let ibv_open_device/ibv_alloc_pd/ibv_create_comp_channel succeed but ibv_create_qp fails with EINVAL. Creating a gloo Device starts a background thread; after fork() in TransportMultiProcTest the thread handle is invalid, causing SIGSEGV (exit 139) in Device::~Device. Fix: probe ibverbs capability using raw APIs (through ibv_create_qp) in the test's createDevice() before constructing a gloo Device. If QP creation fails, mark IBVERBS as unavailable and return nullptr. Also moves GTEST_SKIP() out of worker threads to avoid concurrent calls racing on GTest internals (exit 134), adds a SIGSEGV backtrace handler for test debugging, and builds with RelWithDebInfo.
dolpm
approved these changes
Mar 20, 2026
kapilsh
approved these changes
Mar 20, 2026
pytorchmergebot
pushed a commit
to pytorch/pytorch
that referenced
this pull request
Jun 11, 2026
…87077) ## Summary Bumps the `gloo` submodule from `3135b0b` to `74cc005` (15 commits, 2026-01-09 -> 2026-06-10; +615/-144 across 29 files). The headline change is a **security fix**: a `size_t` overflow in the TCP transport's recv bounds check that allowed a relative write-what-where. `roffset` and `length` are read directly off the wire, and the old check `roffset + length <= size_` could wrap around 2^64, letting an out-of-bounds pair pass and yield an arbitrary write at `ptr_ + roffset`. The fix validates each term independently before forming the pointer. The rest is a mix of portability/bug fixes (ibverbs destructor crash on no-RDMA hosts, TCP address-family mismatch, ROCm `[[nodiscard]]` and `-Wnontrivial-memcall` warnings), toolchain/CI work (move to C++20, native CMake HIP support for ROCm, new CUDA/ROCm/arm64 CI runners, clang-format 21), and named gloo threads for observability. Note: the SHM allreduce optimization ([#458](pytorch/gloo#458)) was added then reverted ([#490](pytorch/gloo#490)) within this range, so it is **not** present in the final tree. ## Included commits (newest -> oldest) | Commit | Author | Date | Description | |--------|--------|------|-------------| | `74cc005` | Tristan Rice | 2026-06-10 | **gloo/tcp: fix size_t overflow in recv bounds check (relative write-what-where)** ([#509](pytorch/gloo#509)) | | `70dc360` | Tristan Rice | 2026-04-20 | Fix address family mismatch by reusing bound socket ([#503](pytorch/gloo#503)) | | `6f4c667` | Tristan Rice | 2026-03-19 | Fix `std::terminate` in ibverbs destructors on systems without RDMA hardware ([#500](pytorch/gloo#500)) | | `2ba34a6` | Karl Gyllstrom | 2026-03-11 | Fix `[[nodiscard]]` `cudaDeviceEnablePeerAccess` warning | | `845824e` | Richard Barnes | 2026-03-11 | Move gloo onto C++20 | | `bcd1672` | Gavin Zhao | 2026-02-12 | ROCm: Migrate to native CMake HIP support ([#478](pytorch/gloo#478)) | | `9322e67` | Tristan Rice | 2026-02-09 | Revert "Intra-node shared memory (SHM) optimizations" ([#490](pytorch/gloo#490)) | | `f834c75` | Tristan Rice | 2026-02-06 | gloo: improve error message on connection closed | | `8789be7` | Tristan Rice | 2026-02-05 | ci: add CUDA and rocm builds | | `d8d0f77` | Nathan Brown | 2026-02-05 | ci: add arm64 runner for github actions ([#487](pytorch/gloo#487)) | | `8d0b9a4` | Lydia Kim | 2026-01-14 | Fix `-Wnontrivial-memcall` | | `980c925` | Lucian Adrian Grijincu | 2026-01-13 | Fix `-Wnontrivial-memcall` error in AllreduceLocal | | `b9cac96` | Tristan Rice | 2026-01-13 | Add `setThreadName` helper and name all gloo threads | | `7ec708d` | Tristan Rice | 2026-01-12 | ci: bump linter to clang-format 21.1.2 | | `5994546` | gaopengff | 2026-01-09 | Intra-node shared memory (SHM) optimizations for CPU primitives ([#458](pytorch/gloo#458)) -- *later reverted by `9322e67`* | ## Test Plan CI. Submodule bump only; the gloo changes carry their own unit tests (TCP pair bounds-check coverage added in #509). Pull Request resolved: #187077 Approved by: https://github.com/dolpm, https://github.com/kapilsh, https://github.com/Regina8023, https://github.com/malfet
jemitche1
pushed a commit
to jemitche1/pytorch
that referenced
this pull request
Jun 13, 2026
…torch#187077) ## Summary Bumps the `gloo` submodule from `3135b0b` to `74cc005` (15 commits, 2026-01-09 -> 2026-06-10; +615/-144 across 29 files). The headline change is a **security fix**: a `size_t` overflow in the TCP transport's recv bounds check that allowed a relative write-what-where. `roffset` and `length` are read directly off the wire, and the old check `roffset + length <= size_` could wrap around 2^64, letting an out-of-bounds pair pass and yield an arbitrary write at `ptr_ + roffset`. The fix validates each term independently before forming the pointer. The rest is a mix of portability/bug fixes (ibverbs destructor crash on no-RDMA hosts, TCP address-family mismatch, ROCm `[[nodiscard]]` and `-Wnontrivial-memcall` warnings), toolchain/CI work (move to C++20, native CMake HIP support for ROCm, new CUDA/ROCm/arm64 CI runners, clang-format 21), and named gloo threads for observability. Note: the SHM allreduce optimization ([pytorch#458](pytorch/gloo#458)) was added then reverted ([pytorch#490](pytorch/gloo#490)) within this range, so it is **not** present in the final tree. ## Included commits (newest -> oldest) | Commit | Author | Date | Description | |--------|--------|------|-------------| | `74cc005` | Tristan Rice | 2026-06-10 | **gloo/tcp: fix size_t overflow in recv bounds check (relative write-what-where)** ([pytorch#509](pytorch/gloo#509)) | | `70dc360` | Tristan Rice | 2026-04-20 | Fix address family mismatch by reusing bound socket ([pytorch#503](pytorch/gloo#503)) | | `6f4c667` | Tristan Rice | 2026-03-19 | Fix `std::terminate` in ibverbs destructors on systems without RDMA hardware ([pytorch#500](pytorch/gloo#500)) | | `2ba34a6` | Karl Gyllstrom | 2026-03-11 | Fix `[[nodiscard]]` `cudaDeviceEnablePeerAccess` warning | | `845824e` | Richard Barnes | 2026-03-11 | Move gloo onto C++20 | | `bcd1672` | Gavin Zhao | 2026-02-12 | ROCm: Migrate to native CMake HIP support ([pytorch#478](pytorch/gloo#478)) | | `9322e67` | Tristan Rice | 2026-02-09 | Revert "Intra-node shared memory (SHM) optimizations" ([pytorch#490](pytorch/gloo#490)) | | `f834c75` | Tristan Rice | 2026-02-06 | gloo: improve error message on connection closed | | `8789be7` | Tristan Rice | 2026-02-05 | ci: add CUDA and rocm builds | | `d8d0f77` | Nathan Brown | 2026-02-05 | ci: add arm64 runner for github actions ([pytorch#487](pytorch/gloo#487)) | | `8d0b9a4` | Lydia Kim | 2026-01-14 | Fix `-Wnontrivial-memcall` | | `980c925` | Lucian Adrian Grijincu | 2026-01-13 | Fix `-Wnontrivial-memcall` error in AllreduceLocal | | `b9cac96` | Tristan Rice | 2026-01-13 | Add `setThreadName` helper and name all gloo threads | | `7ec708d` | Tristan Rice | 2026-01-12 | ci: bump linter to clang-format 21.1.2 | | `5994546` | gaopengff | 2026-01-09 | Intra-node shared memory (SHM) optimizations for CPU primitives ([pytorch#458](pytorch/gloo#458)) -- *later reverted by `9322e67`* | ## Test Plan CI. Submodule bump only; the gloo changes carry their own unit tests (TCP pair bounds-check coverage added in pytorch#509). Pull Request resolved: pytorch#187077 Approved by: https://github.com/dolpm, https://github.com/kapilsh, https://github.com/Regina8023, https://github.com/malfet
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
Summary
-DUSE_IBVERBS=ON), rdma-core userspace providers makeibv_get_device_list()return devices even without real RDMA hardware. The ibverbsDeviceconstructor succeeds, butibv_create_qp()fails in thePairconstructor, throwingEnforceNotMet. During stack unwinding,~Pair()and~Device()callGLOO_ENFORCEwhich throws from implicitlynoexceptdestructors (C++11+), causingstd::terminate().Fixes the crash at
AllgatherRing/AllgatherTest.VarNumPointer/360seen in every CI run: https://github.com/pytorch/gloo/actions/runs/22975489184/job/66702898253See also #497 which identified the same crash but addressed it differently.
Test plan
-DUSE_IBVERBS=ON -DUSE_LIBUV=ON -DUSE_TCP_OPENSSL_LINK=ON