63 lines
6.2 KiB
Markdown
63 lines
6.2 KiB
Markdown
---
|
|
id: "bug-1917"
|
|
status: "fixed"
|
|
status_updated: "2026-10-04"
|
|
status_note: ""
|
|
component: "lib"
|
|
severity: "crash"
|
|
timestamp: "2026-10-03T22:55:30Z"
|
|
last_seen: "2026-10-03T22:55:30Z"
|
|
occurrences: 1
|
|
files: ["src/lib/xcectx.c", "src/lib/netio.c"]
|
|
tags: ["race", "failover", "client", "ssl"]
|
|
related: ["bug-1916"]
|
|
fix_commits: ["431a6d70", "f26a2d45"]
|
|
gitea_issue: 342
|
|
gitea_synced_status: "fixed"
|
|
gitea_synced_hash: "8e6a90f1f914"
|
|
---
|
|
|
|
# bug-1917: xce_ctx_reconnect re-inits the endpoint on the caller thread while the io_worker may still use it
|
|
|
|
## Description
|
|
**Observed:**
|
|
Not yet observed as a crash -- found by code review while fixing bug-1916. `xce_ctx_reconnect` (src/lib/xcectx.c ~366) calls `endpoint_fini` / `endpoint_init` on the client's caller thread for an endpoint embedded in the context, while the io_worker may still be dispatching the same endpoint (still monitored). Exposed operations: the io_worker's post-handler store into `ep->fsm_in.pkt` and `endpoint_unref` (harmless after the bug-1916 fix), a concurrent `SSL_read` / `handle_error` on the old fd/SSL against `SSL_free` / `close`, and `fsm_out.pkt` freed while `handle_output` is writing it. `vtask_entry` also reads `ep->owner` without a lock while reconnect rewrites it.
|
|
|
|
**Expected:** reconnect first unmonitors the old endpoint and waits for the io_worker to finish any in-flight dispatch on it (quiesce), then tears it down, as the controller side does with deferred per-endpoint teardown (bug-1839).
|
|
**Impact:** potential use-after-free / crash in OpenSSL or `handle_output` stacks under `xce_ctx_reconnect` during client failover (worker, scheduler, xce_test, apiserver backend), most likely under heavy CPU load, like bug-1916.
|
|
**Found by:** review (bug-1916 fix, 2026-10-03)
|
|
|
|
UPDATE (2026-10-03): Review of the bug-1916 fix (2026-10-03) adds a second hazard: a reply task from the OLD connection still queued when xce_ctx_reconnect re-arms ep->owner can run against the NEW session (ep->owner read unsynchronised in xcectx.c vtask_entry ~134-135).
|
|
|
|
## Environment
|
|
- OS: no /etc/os-release; Docker container; NixOS kernel; kernel Linux 6.18.33 x86_64
|
|
- libc: glibc 2.42-47 (interpreter /nix/store/wb6rhpznjfczwlwx23zmdrrw74bayxw4-glibc-2.42-47/lib/ld-linux-x86-64.so.2)
|
|
- Toolchain: gcc (GCC) 15.2.0, meson 1.9.1
|
|
- Build: asan (xce.build)
|
|
- Code: scheduler-framework @ 6d9ef3d0-dirty
|
|
- Deployment: any libxce client failing over during a controller failover (seen context: /xce-test-ha under 40 busy loops on 32 cores)
|
|
|
|
## Reproduction
|
|
1. Not reproduced yet. Candidate: loaded `/xce-test-ha` soaks (40 busy loops on 32 cores) with the bug-1916 fix in place; look for ASAN reports with `xce_ctx_reconnect` in the stack but not `free_packet` (e.g. SSL_read/SSL_free or handle_output frames).
|
|
- Frequency: 0/2 runs, 2026-10-03 (code-review finding; the 2 loaded soaks crashed on bug-1916 first)
|
|
- Evidence: code reading of src/lib/xcectx.c xce_ctx_reconnect and src/lib/netio.c endpoint_fini / handle_input / handle_output
|
|
|
|
## Root cause
|
|
`xce_ctx_reconnect` (src/lib/xcectx.c) ran `endpoint_fini`/`endpoint_init` on `ctx->ep` straight away, while work dispatched for the old connection could still be running: the io_worker inside its epoll batch (it holds an endpoint reference for the whole batch: `SSL_read`/`handle_output` on the fd and SSL being freed), and VTASK_IO reply tasks still queued on the context's pool (each holds a reference and reads `ep->owner` without a lock). A queued reply task of the old connection that ran after the reconnect found `ep->owner` re-armed to the context and was delivered to the new session.
|
|
|
|
## Fix
|
|
`xce_ctx_reconnect` first calls `xce_ctx_quiesce_ep`: unmonitor the old fd (no further io_worker events), clear `ep->owner` under `ep->lock` (stale reply tasks are dropped; `vtask_entry` frees their packet), then wait until the endpoint's reference count is back at its base value (0 for the embedded client endpoint, 1 for a heap one), polling every 1 ms. The wait is bounded at 2000 ms: if the caller is itself a task of this endpoint its own reference never drains, and the reconnect then proceeds as before with a WARN ("reconnecting anyway") rather than deadlocking. `vtask_entry` now reads `ep->owner` under `ep->lock`.
|
|
|
|
UPDATE (2026-10-08): Follow-up (2026-10-08, f26a2d45): a reconnect made from inside a task of the endpoint it re-establishes no longer waits out the 2 s bound (vtaskmgr tracks the endpoint whose VTASK_IO task the thread runs; xce_ctx_quiesce_ep returns at once in that case), and old-connection tasks queued behind it can no longer reach the new session: each VTASK_IO task records ep->reconnect_gen when queued and vtask_entry drops a task with a stale generation.
|
|
|
|
## Verification
|
|
- Test: `test_reconnect_drains_old_connection_tasks` (new src/lib/testsuite/test_ctx_reconnect_quiesce.c): reconnect to a local listener, queue an old-connection packet via INVOKE_MSG_HANDLER -> handle_msg with the pool stopped, reconnect again while a helper starts the pool 100 ms in. RED before the fix (the stale task reached the context's request handler: expected 0 calls, was 1), GREEN and LSan-clean after.
|
|
- Suites: `xce-build.sh all` asan, 127/127 meson suites OK, Go tests ok (2026-10-04).
|
|
- Live: loaded `/xce-test-ha` sequential 30 rounds (40 busy loops on 32 cores), 2026-10-04: PASS, zero job loss, all 3 nodes led, no ASAN report in any controller/worker/scheduler log; client failover reconnects in worker/scheduler/xce_test throughout, and the quiesce never needed its timeout (no "reconnecting anyway" WARN).
|
|
|
|
UPDATE (2026-10-05): Live observation (2026-10-05, bug-1736 E2E): the xce_test fake worker hit the bounded-wait path -- a vtask handler of its own connection timed out waiting for a reply queued behind it (-110), failed over through xce_ctx_reconnect, and the quiesce logged 'old connection still has 2 in-flight reference(s) after 2000 ms; reconnecting anyway' and proceeded, as designed; no crash or ASAN report.
|
|
|
|
UPDATE (2026-10-08): Follow-up: test_reconnect_from_own_task_does_not_wait_or_leak (test_ctx_reconnect_quiesce.c) -- a reconnect from the endpoint's own task with a stale message queued behind it: RED before (took the full 2 s bound), GREEN after (under 500 ms, the queued stale message never reaches the handler), LSan clean; xce-build all asan 127/127 (2026-10-07).
|
|
|
|
## References
|
|
- Related: none
|