Repository navigation
Fix double-close of _serverFd in SocketServer teardown - #585
Merged
Merged
Conversation
Collaborator
|
Do you think you can merge master in your branch, I think I fixed the build issue. with uwp, then we'll reckick the PR checks. |
SocketServer::stop() closed _serverFd but never cleared it, and stop()
runs again from both ~WebSocketServer() and ~SocketServer(), so every
teardown closed the same fd number two or three times. If another
thread opens a descriptor between those closes, the kernel reuses the
number and the stale close destroys it — observed as intermittent
process aborts ("Unexpected error 9 on netlink descriptor", glibc's
getaddrinfo netlink probe) when a DNS lookup raced a server teardown.
Close the fd once and set it to -1, in stop() and in the listen()
error paths.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Contributor
Author
|
Done! |
bsergean
added a commit
that referenced
this pull request
Jun 25, 2026
Release of accumulated fixes since v12.0.0: DNS lookup crash/offline resolution (#589), OpenSSL IP-address cert validation (#588), case-insensitive Upgrade header (#584), server fd double-close (#585), fragment buffering optimization (#562), and the install include-dir fix (#582). Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
kahrendt
added a commit
to Sendspin/sendspin-cpp
that referenced
this pull request
Oct 6, 2026
…173) ## What Three lifecycle tests asserted a state another thread could still change. Each now waits on what it asserts, or reads a value only the test thread writes: - `TwoPeersAtOnceSettleOnThePreferredOne` and `TwoServerRaceResolvedByPreference` asserted `is_connected()` as soon as the winner's server information appeared. `refresh_published_state()` lowers the connected flag as soon as no operational connection remains, and publishes the winner's server information on install. Only the end of the tick raises the flag again. The two-peers test now waits for both together. The race test waits for `is_connected()` after the incumbent's close, so a stale flag from the incumbent cannot satisfy it. - `DestroyingARunningClientWithAReturnedLocalItemIsClean` asserted the player quota was nonzero right after handing a chunk over. The idle sync task discards that chunk and empties the quota on its own thread, so the assertion could run too late. The hand-over is now proven from two things the sync thread cannot change: the host ring's `items_waiting_` shows the chunk was copied into the ring, and the consumer's drop log shows it was not dropped after the copy. `tests/CMakeLists.txt` names the new private reach. ## Why Under full CPU load, 300 repeats each on `main` (failed runs, TSan / ASan): two-peers 13 / 6, race 17 / 20, destroy 163 / 43. CI runs serially and rarely hits these, but a loaded runner will. No ThreadSanitizer reports in any run: the races are in the tests, not the library. ## Review notes - Same load, this branch: 0 failures in 300 repeats of all three under TSan and ASan. Full suite 693/693 under both, run serially. - Mutants: - Never publishing the connected flag hangs both connection tests into the suite watchdog, the intended structural failure for a wait. - The destroy test fails, each time at the assertion for that path, when the hand-over is never attempted, dropped before the copy (too long), or dropped after it (over quota). - Checked and left alone: every other `is_connected()` assertion. Most follow a connection that was already up. `PostRehandshakeInadmissibleActivateDrops` and `TogglingUnpairedAccessBringsTheUnpairedSessionInLine` assert it false after a close; they passed 1000 and 567 repeats under the same load. - Not in this PR: a separate host-only abort under load, glibc's "Unexpected error 9 on netlink descriptor". IXWebSocket v11.4.5 closes the server's listen fd up to three times on stop. Fixed upstream in machinezone/IXWebSocket#585, which is not yet in a release. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01TVRGCEo9dwfwss5LhHu4cf --- _Generated by [Claude Code](https://claude.ai/code/session_01TVRGCEo9dwfwss5LhHu4cf)_ --- _Generated by [Claude Code](https://claude.ai/code)_ --------- Co-authored-by: Claude <noreply@anthropic.com>
kahrendt
added a commit
to Sendspin/sendspin-cpp
that referenced
this pull request
Oct 6, 2026
## What Bumps the host build's IXWebSocket from v11.4.5 to v12.0.1. ESP builds don't use IXWebSocket and are unaffected. ## Why v11.4.5 can close the server's listen fd up to three times on stop. Under load this caused glibc's "Unexpected error 9 on netlink descriptor" abort, noted in #173. v12.0.1 includes the fix (machinezone/IXWebSocket#585), and also #589, which avoids the `AI_ADDRCONFIG` netlink path during DNS lookups. ## Review notes - No API changes affect us. The new `WebSocketServer` arguments for ping interval and send timeout default to `-1` (off), which matches the old behavior. - Upstream's `IX_WEBSOCKET_VERSION` header still reads 11.4.6. The fetched tree is tag v12.0.1.
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
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.
Hi, I'm using IXWebSocket in https://github.com/openglad/openglad, and ran into an issue when using this in a multithreaded setting.
It seems that
SocketServer::stop()closes_serverFdbut never clears it, and stop() runs again from both~WebSocketServer()and~SocketServer(), so every teardown closes the same fd number two or three times. If another thread opens a descriptor between those closes, the kernel reuses the number and the stale close destroys it. I was hitting intermittent process aborts ("Unexpected error 9 on netlink descriptor", glibc's getaddrinfo netlink probe) whenever a DNS lookup in another thread raced a server teardown.This PR sets
_serverFdto -1 after closing it (including in error paths), like what seems to happen to_sockfdcurrently elsewhere in the code. This seems to have fixed it for me.