| 1353 | } |
| 1354 | |
| 1355 | static void runtime_worker_disconnect(cbm_daemon_runtime_worker_t *worker) { |
| 1356 | cbm_daemon_runtime_service_t *service = worker->service; |
| 1357 | cbm_daemon_client_id_t client_id = CBM_DAEMON_CLIENT_ID_INVALID; |
| 1358 | uint64_t shutdown_deadline = runtime_deadline_after(service->shutdown_timeout_ms); |
| 1359 | bool last_committed_left = false; |
| 1360 | bool linger_armed = false; |
| 1361 | atomic_store_explicit(&worker->disconnecting, true, memory_order_release); |
| 1362 | cbm_mutex_lock(&service->mutex); |
| 1363 | if (worker->admitted) { |
| 1364 | client_id = worker->client_id; |
| 1365 | worker->admitted = false; |
| 1366 | } |
| 1367 | if (worker->admission_committed) { |
| 1368 | worker->admission_committed = false; |
| 1369 | if (service->committed_clients > 0) { |
| 1370 | service->committed_clients--; |
| 1371 | } |
| 1372 | if (service->committed_clients == 0 && !service->permanent) { |
| 1373 | /* A HELLO whose application session is still opening is only a |
| 1374 | * provisional coordinator client. It cannot keep the generation |
| 1375 | * alive after the final fully committed frontend disconnects. |
| 1376 | * A permanent generation (`daemon start`) deliberately survives |
| 1377 | * this: only the stop/drain ops or a process kill end it. */ |
| 1378 | last_committed_left = true; |
| 1379 | /* Cold-storm race (2026-09): if another connection is already |
| 1380 | * accepted and still mid-HELLO (not yet admitted) when the last |
| 1381 | * committed client leaves, it is a one-shot peer racing to commit — |
| 1382 | * retiring now strands it on a STOPPING daemon and forces a full |
| 1383 | * cold respawn ("cold-storm client failed (racing daemon spawn)"). |
| 1384 | * Linger a bounded window for it; the accept loop retires the |
| 1385 | * generation once that racer drains without committing or the window |
| 1386 | * elapses, and a racer that does commit clears the linger |
| 1387 | * (runtime_worker_commit_admission). This deliberately lingers ONLY |
| 1388 | * for a pre-HELLO racer, never for an already-admitted provisional |
| 1389 | * session (HELLO done, app session opening): the contract that a |
| 1390 | * provisional coordinator client cannot keep a retiring generation |
| 1391 | * alive is preserved (the reject-blocked-provisional-session guard). |
| 1392 | * With no such racer — the common single-client case — retire |
| 1393 | * immediately, unchanged. (A cross-process cohort peer count is |
| 1394 | * unavailable: advisory locks expose presence, not a holder count, |
| 1395 | * and Windows LockFileEx has no non-owning probe.) */ |
| 1396 | if (runtime_has_pending_hello_peer_locked(service, worker)) { |
| 1397 | service->ephemeral_linger_deadline_ms = |
| 1398 | runtime_deadline_after(runtime_ephemeral_linger_timeout_ms()); |
| 1399 | linger_armed = true; |
| 1400 | } else { |
| 1401 | service->ephemeral_linger_deadline_ms = 0; |
| 1402 | runtime_service_begin_stopping_locked(service, shutdown_deadline, false, |
| 1403 | "last_committed_client_disconnected"); |
| 1404 | } |
| 1405 | } |
| 1406 | } |
| 1407 | cbm_mutex_unlock(&service->mutex); |
| 1408 | if (client_id == CBM_DAEMON_CLIENT_ID_INVALID) { |
| 1409 | return; |
| 1410 | } |
| 1411 | /* Set the coordinator hold BEFORE releasing this client so its last-client |
| 1412 | * self-transition to STOPPING is suppressed while a racing peer is still |
no test coverage detected