queue: clear next before re-enqueue by dacheng-gao · Pull Request #170 · mitchellh/libxev

dacheng-gao · 2025-06-20T16:13:07Z

Probablly fixes #169

mitchellh · 2025-06-20T16:16:09Z

Thanks, we're definitely going to need a test to cover this. If you don't get to it I can try eventually

dacheng-gao · 2025-09-30T06:39:12Z

…xt in submit Bug mitchellh#209: kqueue tick(0) EV_DELETE flush - Changed early exit condition from 'if (wait == 0) break' to 'if (wait == 0 and changes == 0) break' so pending EV_DELETE kevent changes from the event-processing loop are flushed before tick(0) returns. Bug mitchellh#170: rearm state correction - When a non-kqueue completion (close, shutdown, threadpool fallback) returns .rearm, set flags.state to .adding before pushing to the submissions queue. Previously .dead state caused submit() to route through stop_completion (cancel) instead of start() (resubmit). Issue mitchellh#169: defensive next-pointer clear in submit() - Explicitly clear c.next before push to completions queue when processing kevent-returned events in submit(). This prevents an assertion failure if a completion is present in multiple queues concurrently (race condition).

Add tests for two bugs: Bug mitchellh#209: tick(0) EV_DELETE flush - kqueue tick(0) may leave stale kernel filters when disarming. Six tests added covering multiple disarms, write/read disarm on same fd, rapid cycles, no-op tick, stale filter detection, and the kevent-event-processing race path (fires without fix, passes with fix). Bug mitchellh#170: rearm from non-kqueue completions - when a completion that completes without kqueue registration returns .rearm, it is pushed to submissions with .dead state which gets routed to stop_completion instead of start. Integration test added (fails without fix, passes with fix). Queue unit tests added verifying re-enqueue across queues, drain-and-refill, and rapid cycle invariants. (cherry picked from commit efdf2cb) (cherry picked from commit 6995a3a)

…xt in submit Bug mitchellh#209: kqueue tick(0) EV_DELETE flush - Changed early exit condition from 'if (wait == 0) break' to 'if (wait == 0 and changes == 0) break' so pending EV_DELETE kevent changes from the event-processing loop are flushed before tick(0) returns. Bug mitchellh#170: rearm state correction - When a non-kqueue completion (close, shutdown, threadpool fallback) returns .rearm, set flags.state to .adding before pushing to the submissions queue. Previously .dead state caused submit() to route through stop_completion (cancel) instead of start() (resubmit). Issue mitchellh#169: defensive next-pointer clear in submit() - Explicitly clear c.next before push to completions queue when processing kevent-returned events in submit(). This prevents an assertion failure if a completion is present in multiple queues concurrently (race condition). (cherry picked from mitchellh#224; the .rearm arm is merged with this fork's keep-active fix: a completion still registered with kqueue stays .active, one never registered is resubmitted as .adding) (cherry picked from commit a7502d6) (cherry picked from commit 570e3b0)