Repository navigation
Make concurrent type metadata retirement stack-bounded - #7
Conversation
Replace snapshot-based shared-pointer teardown with the existing intrusive retirement mechanism. Detach all descriptor edges into the last-owner queue before deleting a plan, while preserving immutable sharing and source-preserving descriptor moves. Add deferred-disposal and concurrent ownership regressions, retaining deep destruction and unwinding coverage. Clients must rebuild because the private descriptor layout changes. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Suppress the Ubuntu analyzer's unhandled-self-assignment false positive on the existing copy-and-swap implementation. The temporary retains the source before swapping and releasing prior ownership; the ownership algorithm is unchanged. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
sean-parent
left a comment
There was a problem hiding this comment.
Talos review
Verdict: No accepted findings.
Reviewed commit: 8d7dfc0a6442e7c1ca67a0960d896d5510978549, full PR diff against main (merge-base cacd8cf6c43831470769ba2917088ac21ce14c77). The review was rerun for this final pushed snapshot after the CI lint annotation.
| Files | Reviewers | Candidates -> validated -> accepted | Shown | Tests |
|---|---|---|---|---|
| 6 of 6; none skipped | 6 personas plus challenge | 0 -> 0 -> 0 | 0 of 0 | Local normal, clang-tidy, and ASan/UBSan suites: 20/20 each; hosted CI: 20/20 in all five test jobs |
Personas completed: security, correctness, standards, architecture, devops, and documentation. Challenge accepted 0 and rejected 0. Accepted findings: 0; inline comments posted: 0; overflow: 0; unanchored findings: 0; duplicate/recurring findings: 0. Existing reviews/comments were checked before publication. This is a comment review, not an approval.
Verification
Fresh configure/build/test validation also passed release installation, generated package artifact checks, strict documentation, clang-format, and whitespace checks. CI run 43943911436 completed successfully at the reviewed SHA: Ubuntu GCC, Ubuntu Clang, macOS, Windows, and clang-tidy. Logs confirm each test job ran all 20 tests; all four platform jobs checked installed package artifacts. Release-only Pages deployment was skipped as expected; documentation was built locally.
The initial Ubuntu clang-tidy diagnostic on the existing self-copy-safe copy-and-swap implementation was addressed with a narrowly scoped, justified annotation in 8d7dfc0; the ownership algorithm is unchanged, and the final Ubuntu clang-tidy job passed.
No ThreadSanitizer or macOS LeakSanitizer coverage is claimed. The historical scheduling-dependent race was not deterministically reproduced. Controlled mutation during implementation verified that omitting metadata element-edge detachment causes the deep-release regression to crash; restored code passes. The private descriptor layout changes, so clients must rebuild.
Persona routing: static roster (DCAP routing off; rerun with --dcap to enable).
Org knowledge: off (rerun with --knowledge).
Scout
Used: risk CRITICAL, evidence LOW, status INCOMPLETE (scout 0.9.135, <1s).
Cross-file leads given to reviewers: 3 changed symbol(s) with uses outside the diff, 0 new value(s) not handled everywhere, 0 possibly unused symbol(s).
Scout rated its evidence LOW for this change; reviewers used it only as leads.
Scout evidence is incomplete (status INCOMPLETE; change analysis bounded; rows cut by the packet budget); reviewers treated missing evidence as unknown, not as proof.
Tests to run (Scout: 19 Must Run test(s)):
Scout also lists 4 Should Run test(s).
Scout left 23 affected-test entries out of its packet; for the full list run scout affected-tests -r '/Users/sean-parent/repos/github.com/stlab/cel-runtime/.claude/worktrees/metadata-retirement' -s compare -b origin/main.
Summary
shared_ptr::use_count()snapshots with the existing atomic last-owner retirement queue.Compatibility: Public source signatures are unchanged, but the private descriptor layout changes. Rebuild clients.
Test plan
Fresh local validation on macOS arm64 with AppleClang 21 and C++20. Configure commands used the existing CPM source cache through
-DCPM_SOURCE_CACHE; the presets were otherwise unchanged.cmake --preset test,cmake --build --preset test,ctest --preset test --no-tests=error: 20/20 passed, including source and installed consumers.cmake --preset clang-tidy,cmake --build --preset clang-tidy,ctest --preset clang-tidy --no-tests=error: full configured static analysis and 20/20 tests passed.cmake --preset default,cmake --build --preset default,cmake --install build/default --prefix <temporary-prefix>,cmake -DPACKAGE_TEST_PREFIX=<temporary-prefix> -P build/default/cpp-library-test-package.cmake: release build, installation, and artifact checks passed.cmake --preset docs,cmake --build --preset docs: documentation passed.build/sanitizedwith-fsanitize=address,undefined -fno-omit-frame-pointerand matching linker flags;cmake --build build/sanitizedandASAN_OPTIONS=detect_leaks=0 UBSAN_OPTIONS=halt_on_error=1 ctest --test-dir build/sanitized --output-on-failure --no-tests=error: 20/20 passed.clang-format --dry-run --Werroron all five changed C++ files andgit diff --check: passed.8d7dfc0: Ubuntu GCC, Ubuntu Clang, macOS, Windows, and clang-tidy each passed all 20 tests. All four platform jobs also passed release installation and package checks.The concurrency tests exercise real ownership and exact-once disposal without sleeps or timing assertions. They do not deterministically reproduce the historical scheduling-dependent race. No ThreadSanitizer or macOS LeakSanitizer coverage is claimed. Pages deployment is release-only and is not part of this PR.
Generated with GitHub Copilot CLI