Hyprland module-host timeout leaves socket work alive and executes fallback after failure #46
Breadway
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Found during independent review of Breadway/bread PR44, exact head
44a5a0b740. Confidence high: real daemon and sandboxed host with private fake compositor reproduce. Evidence /tmp/opencode/review-pr44/timeout-probe.py and timeout-probe.log on hermes. Repair on existing PR44, not a duplicate PR. breadd/src/ipc/module_host_bridge.rs:587-590 times out a spawn_blocking JoinHandle without cancelling the blocking work. hyprland_compat.rs:63-71 has no socket read/write deadline or reply-size cap. Fake compositor receives dispatch at 1.6s, stalls 9.5s, then rejects it. Module already receives hyprland request timed out, but daemon still sends eval hl.dispatch(...) at 11.1s. Thus commands can execute after reported failure and caller retry, blocked workers can accumulate, and shutdown can wait indefinitely. Bound the actual transport/operation lifetime across fallback and read-back, prevent any further request after expiry, bound request/reply bytes and in-flight work, and add stalled/late/oversized response regressions.Exact-head re-review at
5ed063907a: original late-after-timeout defect is repaired (independent 8.006s timeout, only initial dispatch, no fallback). Remaining R1 lifecycle defect reproduced against real daemon + authenticated sandboxed fake host: initial dispatch at0.083s, exact fake_module_host PID SIGKILL at0.083s, /proc/PID/exe absent afterwards, compositor rejects at2.083s and daemon still sends eval hl.dispatch(...) at2.083s. High confidence, probe /tmp/opencode/review-pr44/disconnect-probe.py and .log. module_host_bridge.rs:128-162 awaits dispatch without monitoring connection EOF; HyprlandSession.closed is set only on Drop at178, after operation completes. Unit dropping_the_host_session_stops_an_in_flight_operation manually drops session and misses real lifecycle. Repair same PR44: observe actual disconnect concurrently and signal cancellation promptly, no fallback/read-back after disconnect, release permits. Add real host-close/death regression. Related test defect: deadlines.rs:191-203 rewrites script and go on an already exited one-shot fake host (examples/fake_module_host.rs reads script once and exits), so its second in-flight shutdown scenario never starts. Reviewer separate shutdown probe observed initial request then SIGTERM: exit0 in0.007s and no fallback. Keep regression but make it observe a genuinely live second stalled call before termination. No source changes or merge.Coordinator verification on the same integrated merge
d5748d6093(PR44). Bound transport lifetime, no post-expiry request, request/reply byte caps and in-flight limits are covered by the deadlines suite 6/6 plus wire 3/3 and host 4/4 on the merge tree. Independent pass atb05a6f6(comment 5360, criterion R1) retained. Fake compositor and private sockets only; no host compositor.