---
id: "bug-646"
status: "fixed"
status_updated: "2026-09-24"
status_note: ""
component: "lib"
severity: "hang"
timestamp: "2026-06-17T08:21:32.099Z"
last_seen: "2026-09-19T09:00:00Z"
occurrences: 2
files: ["src/lib/netio.c"]
tags: ["wrong-reference", "c", "libxce", "failover", "netio", "epoll", "EPOLL_CTL_MOD", "ENOENT", "silent-failure", "hang", "cbuf_get", "reconnect"]
related: ["bug-610", "bug-609"]
legacy: true
---

# bug-646: xce_test jobsub hangs forever after a failover; main thread blocked in cbuf_get (xceapi.c:304)…

## Description
xce_test jobsub hangs forever after a failover; main thread blocked in cbuf_get (xceapi.c:304) while io_worker sits idle in epoll_wait. Logs: 'rotate failed (rc=-111)' then 'endpoint_update_epoll rc1=-1 < 0'.

## Environment
_Not recorded (migrated from buglog.json)._

## Reproduction
_Not recorded (migrated from buglog.json)._

## Root cause
endpoint_update_epoll() swallowed iomgr_update_poll() failures: it WARN'd 'rc1=-1' but returned 0 and left ep->flags unchanged, so the poll interest was never armed. iomgr_update_poll uses EPOLL_CTL_MOD, which fails with ENOENT when the fd is no longer registered (e.g. after a failed reconnect: xce_ctx_reconnect endpoint_fini's the old fd but only iomgr_monitor/EPOLL_CTL_ADDs on a SUCCESSFUL connect). The next endpoint_enq() send then MOD-failed silently, request bytes were never flushed, no reply arrived, and the synchronous client blocked forever in cbuf_get (no timeout) with no socket-drop event to wake it (handle_disconnect only fires on a broken socket). Server side was a red herring: DEFAULT_SYNC_TIMEOUT_MS=500 bounds the commit barrier, and ha_dispatch_gate already replies STANDBY/NOT_READY for non-serving roles.

## Fix
endpoint_update_epoll() captures errno and RETURNS -errno (not 0) on iomgr_update_poll failure. endpoint_enq(): on negative arm result, run handle_error(ep) (iomgr_unmonitor + handle_disconnect -> state=ERR + inject -ECONNRESET into wbuf, waking the blocked cbuf_get) and return the error so the failover loop rotates. handle_output(): propagates the arm failure so io_handler runs the error path. TDD: test_netio_arm_failure.c drives EPOLL_CTL_MOD ENOENT via endpoint_enq on an unregistered fd; asserts loud rc<0 + err_handler fired (verified red without fix, green with).

## Verification
_Not recorded (migrated from buglog.json)._

## References
- Related: bug-610, bug-609