"""Out-of-loop shutdown and event-loop liveness backstops. A frozen asyncio loop takes every asyncio-based recovery path down with it, and launchd/systemd KeepAlive only restarts a *dead* process. Hence: (1) an OS-thread shutdown watchdog that dumps stacks and ``os._exit``s past ``restart_drain_timeout + grace``; (2) a heartbeat file at ``/state/gateway.heartbeat`` so supervisors can tell "process alive" from "loop frozen"; (3) a lifetime thread watchdog that hard-exits when the loop is too frozen to run its own callbacks; (4) a self-rescheduling floor timer that keeps the selector timeout finite.""" from __future__ import annotations import asyncio import contextlib import faulthandler import json import logging import os import sys import threading import time import traceback from datetime import datetime, timezone, UTC from pathlib import Path from typing import Any, Callable, Dict, Optional from gateway.restart import GATEWAY_SERVICE_RESTART_EXIT_CODE from hermes_constants import get_hermes_home, get_process_hermes_home from utils import atomic_json_write logger = logging.getLogger(__name__) # Extra leash beyond ``agent.restart_drain_timeout`` so a slow-but-progressing drain survives. # Matches the issue #66892 suggested hardening. DEFAULT_SHUTDOWN_WATCHDOG_GRACE_S = 60.0 DEFAULT_HEARTBEAT_INTERVAL_S = 30.0 DEFAULT_LOOP_FLOOR_TIMER_INTERVAL_S = 5.0 DEFAULT_LOOP_WATCHDOG_INTERVAL_S = 30.0 DEFAULT_LOOP_WATCHDOG_TIMEOUT_S = 10.0 # 3 sustained misses (~90-120s of loop block) escalate; stays tight because the heartbeat write is # off-loop. Slow loops tune gateway.loop_watchdog_* in config.yaml. # The false-positive class that motivated raising this (the watchdog's own on-loop heartbeat fsync stalling # the loop it monitors) is fixed at the root by the off-loop heartbeat write + two-witness probe (#90502), # so the default stays tight for genuine wedges. DEFAULT_LOOP_WATCHDOG_MAX_STRIKES = 3 _HEARTBEAT_RELATIVE = ("state", "gateway.heartbeat") _WATCHDOG_DUMP_RELATIVE = ("logs", "gateway-shutdown-watchdog.log") def _coerce_float(value: Any, default: float, floor: float = 0.0) -> float: """``max(float(value), floor)``, or ``default`` when not coercible.""" try: return max(float(value), floor) except (TypeError, ValueError): return default class _LoopFloorTimerHandle: """Cancelable owner for the currently scheduled selector floor timer.""" def __init__(self, loop: asyncio.AbstractEventLoop, interval: float): self._loop, self._interval, self._cancelled = loop, interval, False self._timer: Optional[asyncio.TimerHandle] = None self._tick() def _tick(self) -> None: if not self._cancelled: self._timer = self._loop.call_later(self._interval, self._tick) def cancel(self) -> None: self._cancelled = True if self._timer is not None: self._timer.cancel() class _LoopLivenessWatchdogHandle: def __init__(self, stop_event: threading.Event, thread: threading.Thread): self._stop_event = stop_event self.stop, self.join, self.is_alive = stop_event.set, thread.join, thread.is_alive def _arm_loop_floor_timer( loop: asyncio.AbstractEventLoop, interval: float = DEFAULT_LOOP_FLOOR_TIMER_INTERVAL_S ) -> _LoopFloorTimerHandle: """Keep at least one timer pending so selector waits remain bounded.""" iv = _coerce_float(interval, 0.0) return _LoopFloorTimerHandle(loop, iv if iv > 0 else DEFAULT_LOOP_FLOOR_TIMER_INTERVAL_S) def start_loop_liveness_watchdog( loop: asyncio.AbstractEventLoop, *, probe_interval: float = DEFAULT_LOOP_WATCHDOG_INTERVAL_S, probe_timeout: float = DEFAULT_LOOP_WATCHDOG_TIMEOUT_S, max_strikes: int = DEFAULT_LOOP_WATCHDOG_MAX_STRIKES, exit_code: int = GATEWAY_SERVICE_RESTART_EXIT_CODE, ) -> Optional[_LoopLivenessWatchdogHandle]: """Start an out-of-loop watchdog that hard-exits after missed probes. The caller (``GatewayRunner._start_loop_liveness_guards``) enforces the ``gateway.loop_watchdog: false`` opt-out.""" stop_event = threading.Event() # The thread a post-mortem has to read first is the loop's own. ``BaseEventLoop`` records it # while the loop runs, and this starter runs on the loop thread, so its own ident covers the # mock and not-yet-running cases. loop_thread_id = getattr(loop, "_thread_id", None) if not isinstance(loop_thread_id, int): loop_thread_id = threading.get_ident() def _watchdog() -> None: strikes = 0 while not stop_event.wait(timeout=probe_interval): probe_event = threading.Event() try: loop.call_soon_threadsafe(probe_event.set) except RuntimeError: # normally closed loop: nothing left to backstop return except Exception: logger.debug("Failed to schedule gateway loop liveness probe", exc_info=True) return deadline = time.monotonic() + probe_timeout while not stop_event.is_set(): # poll so a stop() mid-wait is honoured within ~50ms remaining = deadline - time.monotonic() if remaining <= 0 or probe_event.wait(timeout=min(remaining, 0.05)): break else: return if probe_event.is_set(): strikes = 0 continue if stop_event.is_set(): # re-checked before each irreversible step: a late stop() wins return strikes += 1 if strikes < max_strikes: continue if stop_event.is_set(): return with contextlib.suppress(Exception): logger.critical( "Gateway event loop missed %d consecutive liveness probes; dumping all thread " "stacks and exiting with code %d so the service supervisor can restart it.", strikes, exit_code) # File + stderr: a console-less launcher (Windows .vbs login item) discards stderr, so a # liveness death used to leave no stack anywhere (#120356, #91097). _write_watchdog_dump( get_shutdown_watchdog_dump_path(), delay_s=probe_timeout * strikes, snapshot={"strikes": strikes, "probe_interval_s": probe_interval}, event="loop_liveness_watchdog_fired", loop_thread_id=loop_thread_id) if stop_event.is_set(): return _mark_exited_quietly(exit_code, "loop_liveness_watchdog") _hard_exit(exit_code) thread = threading.Thread(target=_watchdog, daemon=True, name="gateway-loop-liveness-watchdog") try: thread.start() except Exception: logger.debug("Failed to start gateway loop liveness watchdog", exc_info=True) return None return _LoopLivenessWatchdogHandle(stop_event, thread) def _hard_exit(exit_code: int) -> None: """``os._exit`` skips every cleanup: SIGKILL in-flight foreground commands first, they run in their own process group and would outlive the gateway, reparented to init.""" with contextlib.suppress(Exception): from tools.environments.base import kill_live_foreground_processes kill_live_foreground_processes(now=True) os._exit(exit_code) def _mark_exited_quietly(exit_code: int, reason: str) -> None: """Best-effort terminal stamp on BOTH lifecycle records before ``os._exit`` skips teardown: the lifecycle ledger (so the next boot names the watchdog, not SIGKILL/OOM) and ``gateway_state.json`` (so ``hermes gateway status`` and every other reader of that file stop seeing ``running`` for a process the watchdog killed — #113372). The runtime-status write goes LAST: it is the record housekeeping refreshes, so nothing may overwrite it after we stamp it.""" with contextlib.suppress(Exception): from gateway.lifecycle_ledger import mark_exited mark_exited(exit_code, reason=reason) with contextlib.suppress(Exception): # os._exit skips atexit: stamp the exit-metrics marker now from hermes_cli.observability.shared_metrics_process import stamp_exit stamp_exit("watchdog") with contextlib.suppress(Exception): from gateway.status import write_runtime_status # Only the supervisor-restart code asserts a restart; other codes leave the recorded # operator intent (a restart-drain that wedged is still a requested restart) untouched. restart = {"restart_requested": True} if exit_code == GATEWAY_SERVICE_RESTART_EXIT_CODE else {} write_runtime_status( gateway_state="degraded", exit_reason=reason, wait_timeout=0.25, **restart) def _process_hermes_home() -> Path: """HERMES_HOME for process-level identity files (ignore profile overrides).""" return get_process_hermes_home() if os.environ.get("HERMES_HOME", "").strip() else get_hermes_home() def _home(home: Optional[Path]) -> Path: return home if home is not None else _process_hermes_home() def get_loop_heartbeat_path(home: Optional[Path] = None) -> Path: return _home(home).joinpath(*_HEARTBEAT_RELATIVE) def get_loop_tick_socket_path(home: Optional[Path] = None, pid: Optional[int] = None) -> Path: """``/state/gateway.loop-tick..sock`` — PID-suffixed so a stale node from a dead process is never mistaken for this gateway's witness. Served by the loop itself (``_tick_socket_handler``), so an answer proves the loop dispatches; the heartbeat cannot. Served by the gateway loop itself (see ``_tick_socket_handler``): an answer is direct proof that the loop is dispatching, which is exactly the property the heartbeat file lost when its write moved off-loop (#90502). """ pid = int(pid if pid is not None else os.getpid()) return _home(home) / "state" / f"gateway.loop-tick.{pid}.sock" def get_shutdown_watchdog_dump_path(home: Optional[Path] = None) -> Path: return _home(home).joinpath(*_WATCHDOG_DUMP_RELATIVE) def _thread_name(ident: Optional[int]) -> Optional[str]: """Name of the live thread with this ident, or None once it has exited.""" return next((thread.name for thread in threading.enumerate() if thread.ident == ident), None) def write_loop_heartbeat( *, pid: Optional[int] = None, start_time: Optional[float] = None, home: Optional[Path] = None, extra: Optional[dict[str, Any]] = None) -> Path: """Atomically rewrite the loop-liveness heartbeat file; never raises. ``start_time`` (process start, epoch seconds) lets supervisors detect PID reuse.""" path = get_loop_heartbeat_path(home) payload: dict[str, Any] = {"pid": int(pid if pid is not None else os.getpid()), "updated_at": datetime.now(UTC).isoformat(), "monotonic": time.monotonic()} if start_time is not None: payload["start_time"] = float(start_time) with contextlib.suppress(Exception): # after an unclean death this is the last memory record from gateway.lifecycle_ledger import sample_memory if mem := sample_memory(): payload["mem"] = mem if extra: payload.update(extra) try: atomic_json_write(path, payload, indent=None) except Exception: logger.debug("Failed to write gateway loop heartbeat", exc_info=True) return path def resolve_shutdown_watchdog_delay( drain_timeout: float, *, grace_s: float = DEFAULT_SHUTDOWN_WATCHDOG_GRACE_S) -> float: """Return the wall-clock leash for the shutdown watchdog thread.""" grace = _coerce_float(grace_s, DEFAULT_SHUTDOWN_WATCHDOG_GRACE_S) return _coerce_float(drain_timeout, 0.0) + grace def _write_watchdog_dump(dump_path: Path, *, delay_s: float, snapshot: Optional[dict[str, Any]], event: str = "shutdown_watchdog_fired", loop_thread_id: Optional[int] = None) -> None: """Best-effort faulthandler + metadata dump before hard-exit (file, then stderr). ``loop_thread_id`` names the monitored event loop's thread, whose own stack is written before the all-threads dump: that dump does not say which thread runs the loop, which is the first thing a post-mortem of a liveness death needs. """ header: dict[str, Any] = {"event": event, "pid": os.getpid(), "delay_s": delay_s, "fired_at": datetime.now(UTC).isoformat(), "snapshot": snapshot or {}} loop_section = "" if loop_thread_id is not None: header["monitored_loop_thread_id"] = loop_thread_id header["monitored_loop_thread_name"] = _thread_name(loop_thread_id) loop_section = _format_loop_thread_section(loop_thread_id) with contextlib.suppress(OSError): # unwritable dir: skip only the file copy, stderr still runs dump_path.parent.mkdir(parents=True, exist_ok=True) with contextlib.suppress(Exception), open(dump_path, "a", encoding="utf-8") as fh: fh.write(json.dumps(header, default=str) + "\n" + loop_section) fh.flush() fh.write("--- faulthandler dump (all threads) ---\n") fh.flush() try: faulthandler.dump_traceback(file=fh, all_threads=True) except Exception: fh.write("(faulthandler.dump_traceback failed)\n") fh.write("--- end dump ---\n") fh.flush() with contextlib.suppress(Exception): # stderr too: journald/launchd get it if disk is wedged # The loop section goes to stderr as well: with the disk wedged, journald/launchd keep only # this copy, and its faulthandler dump is capped at 100 threads, newest first. sys.stderr.write(f"Gateway {event.replace('_', ' ')} after {delay_s:.0f}s " f"(pid={os.getpid()}); dumping all thread stacks.\n" + loop_section) sys.stderr.flush() faulthandler.dump_traceback(all_threads=True) def _format_loop_thread_section(loop_thread_id: int) -> str: """The monitored loop thread's stack, formatted once for both dump copies. Built from ``sys._current_frames()`` + ``traceback.format_stack``: file, line, function and source line only, never frame locals, so no token or message text a frame holds can reach the dump. ``faulthandler`` cannot be asked for one thread, and its all-threads list stops after 100 threads, newest first: the loop thread, started before every worker, is the one it drops.""" try: frame = sys._current_frames().get(loop_thread_id) stack = ("".join(traceback.format_stack(frame)) if frame is not None else "(monitored loop thread stack unavailable: thread may have exited)\n") except Exception as exc: # health: allow BLE001 -- a hard-exit diagnostic path must never raise stack = f"(monitored loop thread stack unavailable: {type(exc).__name__})\n" return "--- monitored gateway loop thread stack ---\n" + stack def arm_shutdown_watchdog( delay_s: float, *, done_event: Optional[threading.Event] = None, snapshot_fn: Optional[Callable[[], dict[str, Any]]] = None, exit_code: int = 1, dump_path: Optional[Path] = None, name: str = "gateway-shutdown-watchdog") -> threading.Event: """Arm a daemon-thread hard-exit backstop for a wedged shutdown path: exits quietly if ``done_event`` is set within ``delay_s``, else dumps diagnostics and ``os._exit(exit_code)``. Never raises; returns ``done_event`` for disarming.""" done = done_event if done_event is not None else threading.Event() delay = _coerce_float(delay_s, DEFAULT_SHUTDOWN_WATCHDOG_GRACE_S) if delay <= 0: return done def _watchdog() -> None: deadline = time.monotonic() + delay # chunked wait so a late disarm is observed within ~1s while time.monotonic() < deadline: if done.wait(timeout=min(deadline - time.monotonic(), 1.0)): return if done.is_set(): return try: snapshot = snapshot_fn() if snapshot_fn is not None else None except Exception as exc: snapshot = {"snapshot_error": repr(exc)} target = dump_path if dump_path is not None else get_shutdown_watchdog_dump_path() _write_watchdog_dump(target, delay_s=delay, snapshot=snapshot) with contextlib.suppress(Exception): logger.critical("Shutdown watchdog fired after %.0fs — forcing process exit " "(asyncio drain path appears wedged; see %s)", delay, target) for stream in (sys.stdout, sys.stderr): with contextlib.suppress(Exception): stream.flush() # Mirror _exit_after_graceful_shutdown: release PID file + runtime lock BEFORE the log drain # (never strand locks), then drain the log queue so logger.critical lands before os._exit. with contextlib.suppress(Exception): # Mirror _exit_after_graceful_shutdown: release PID file + runtime lock BEFORE the log drain # (locks must never be stranded), then drain the async log queue so the logger.critical above # actually reaches the file before os._exit bypasses atexit. (#66892) from gateway.status import remove_pid_file, release_gateway_runtime_lock remove_pid_file() release_gateway_runtime_lock() with contextlib.suppress(Exception): from hermes_logging import drain_log_queue drain_log_queue(timeout=1.0) _mark_exited_quietly(exit_code, "shutdown_watchdog") _hard_exit(exit_code) try: threading.Thread(target=_watchdog, daemon=True, name=name).start() except Exception: logger.debug("Failed to arm shutdown watchdog", exc_info=True) return done async def _tick_socket_handler(reader: asyncio.StreamReader, writer: asyncio.StreamWriter) -> None: """Answer a liveness ping with one byte; never raises. Runs on the gateway loop, so a reply witnesses loop schedulability; the write is a socket-buffer copy (no fsync), immune to the stalls that age the heartbeat.""" try: writer.write(b"1") await writer.drain() except Exception: pass finally: # close even on CancelledError (BaseException), as on BASE with contextlib.suppress(Exception): writer.close() def _sweep_stale_tick_sockets(own_path: Path) -> None: """Unlink loop-tick socket nodes left by dead PIDs (POSIX only; never raises). create_unix_server removes a leftover node at OUR path (os._exit / SIGKILL skip the finally-unlink) but not SIBLING nodes from other dead PIDs. os.kill(pid, 0) is a liveness probe only on POSIX (Windows would TerminateProcess).""" try: for stale in (p for p in own_path.parent.glob("gateway.loop-tick.*.sock") if p != own_path): try: os.kill(int(stale.name.split(".")[-2]), 0) # windows-footgun: ok — POSIX-only except (ValueError, IndexError, OSError): stale.unlink(missing_ok=True) except Exception: logger.debug("stale loop-tick socket sweep failed", exc_info=True) async def loop_heartbeat_forever( *, interval_s: float = DEFAULT_HEARTBEAT_INTERVAL_S, start_time: Optional[float] = None, home: Optional[Path] = None, should_continue: Optional[Callable[[], bool]] = None) -> None: """Rewrite the loop heartbeat file on a cadence until cancelled / gated off. Runs on the gateway loop so a frozen loop lets the file age for monitors. The fsync write goes to a thread (inline, a stalled filesystem blocked the loop inside its own heartbeat and the liveness watchdog killed it: WSL2 fsync stalls p99 31s / max 112s vs a ~90-120s budget) but is *awaited*, never fire-and-forget: an unawaited task would keep the file fresh while the loop was wedged, and one in-flight write at a time keeps a long stall from queuing a thread per interval. Off-loop, file freshness no longer proves loop schedulability, so this task also arms a loop-scheduling witness (``_tick_socket_handler``, flagged as ``loop_tick_socket``); probes must require the witness to agree before classifying WEDGED.""" interval = _coerce_float(interval_s, DEFAULT_HEARTBEAT_INTERVAL_S, floor=1.0) # The witness exposes only a scheduling ping, not control/auth operations. Reuse its # loopback transport when AF_UNIX is unavailable or the encoded address will not fit; # consumers already discover this port in the heartbeat, independent of their TMPDIR. from gateway.control_socket import _fits_sun_path tick_server = tick_socket_path = tick_tcp_port = None try: unix_path = get_loop_tick_socket_path(home) if os.name == "posix" else None if unix_path is not None and _fits_sun_path(unix_path): tick_socket_path = unix_path tick_socket_path.parent.mkdir(parents=True, exist_ok=True) _sweep_stale_tick_sockets(tick_socket_path) tick_server = await asyncio.start_unix_server(_tick_socket_handler, path=str(tick_socket_path)) else: tick_server = await asyncio.start_server(_tick_socket_handler, host="127.0.0.1", port=0) for _s in tick_server.sockets or []: with contextlib.suppress(Exception): if isinstance(_sname := _s.getsockname(), tuple) and len(_sname) >= 2: tick_tcp_port = int(_sname[1]) break except Exception: tick_server = tick_tcp_port = None logger.warning("Loop tick socket unavailable — liveness probes will have no " "loop-scheduling witness and will not escalate on a stale heartbeat", exc_info=True) extra = {"loop_tick_socket": tick_server is not None, "loop_tick_tcp_port": tick_tcp_port} try: while True: # first write is immediate so monitors see a fresh file at once try: await asyncio.to_thread(write_loop_heartbeat, start_time=start_time, home=home, extra=extra) except Exception: # write_loop_heartbeat never raises: executor problem, keep the task logger.debug("Loop heartbeat write failed off-loop", exc_info=True) if should_continue is not None and not should_continue(): return await asyncio.sleep(interval) if should_continue is not None and not should_continue(): return finally: if tick_server is not None: tick_server.close() with contextlib.suppress(Exception): await tick_server.wait_closed() if tick_socket_path is not None: with contextlib.suppress(Exception): tick_socket_path.unlink(missing_ok=True)