Repository navigation
Conversation
Co-Authored-By: GPT-6 <noreply@openai.com>
|
Source fixes and focused validation are complete on both lines: #412 (0.6) and #413 (0.5). The end-to-end unread-stream and close-delivery regressions both fail on unmodified 0.5.2 and pass with the fixes. Suites pass 189 tests on 0.6 and 181 on 0.5, including doctests; focused Clippy passes on 0.5, and workspace native check/Clippy passed on 0.6 using MoQ's pinned Nix toolchain. Full checks with web-transport's own pinned Rust 1.94 toolchain and PR CI remain pending. The first broad run with MoQ's newer Clippy failed on the existing WASM harness MoQ progress is tracked in moq-dev/moq#4705. The remaining blocker is maintainer review/merge and publication of both patched release lines, then MoQ main/dev pin updates. No packages were published and no PRs were merged. Recommendation: finish upstream CI/review, merge/release both lines, then resume the pin steps. (written by GPT-6) |
Co-Authored-By: GPT-6 <noreply@openai.com>
|
Final sources are pushed: #412 at 82527d2 and #413 at 1adcb9c. Complete qmux tests and focused Clippy pass on these sources: 189 tests on 0.6 and 181 on 0.5, including doctests. The last guard preserves the legacy binding's lack of flow control. Both end-to-end regressions fail against unmodified qmux 0.5.2. The newly published qmux 0.6.2 contains the earlier receive-map lookup fix, not these credit/close fixes. MoQ must consume releases containing #412/#413 after review and merge. Broad checks with web-transport's pinned Rust 1.94 toolchain and latest-head CI remain pending. All PRs remain drafts; no packages were published or PRs merged by this work. Recommendation: complete upstream checks/review, merge/release both lines, then resume MoQ main/dev pin updates tracked in moq-dev/moq#4705. No unresolved API or wire choices. (written by GPT-6) |
|
Closing: QMux development has moved to moq-dev/moq. web-transport's copy only takes finished bug fixes until it is frozen (see moq's qmux-rs-retire). This draft is stale, and later merges like #437 and #451 likely overlap with it. Any remaining issue belongs in moq-dev/moq's qmux. (Written by Claude Opus 5.5 with high) |
Problem
Dropping unread receive streams leaks MAX_DATA credit. Removing STOP_SENDING state also loses later STREAM bytes and RESET final-size gaps. A local close during an in-flight write cancels that write before APPLICATION_CLOSE reaches the peer.
Approach
Close delivery before draining unread buffers, account stopped streams until peer FIN/RESET, and register receive state before exposing an accepted stream. Share stream-count ownership with the backend so retained accounting stays bounded. Drain and flush the first requested close before teardown, with a one-second deadline covering the in-flight write and transport close.
Impact
Validation
cargo test -p qmux --all-features: 152 unit tests, 34 integration tests, and 3 doctests passed using MoQ's pinned Nix toolchain.cargo clippy -p qmux --all-targets --all-features -- -D warnings) and workspace native check/Clippy passed. Fulljust checkwith MoQ's newer Clippy stopped on an existingchunks_exact_to_as_chunkslint in the WASM harness; validation with this repository's own pinned Rust 1.94 toolchain and CI are pending.Alternatives
Retiring receive accounting at STOP_SENDING cannot account the peer's final size. Immediate teardown cannot reliably deliver a close frame during a pending write.
Follow-ups
Backport #413 targets
qmux-v0.5.x. Merge both fixes, and publish patched 0.5.x and 0.6.x releases before updating MoQ's respective pins. This PR does not publish or merge either line.(written by GPT-6)