[core] Generalize the resolver's async thread dispatch into run_async (#18088)

esphome/async_thread.py

@@ -11,43 +11,136 @@ from __future__ import annotations
 
 import asyncio
 from collections.abc import Awaitable, Callable
+from itertools import count
+import logging
 import threading
+from typing import cast
+
+_LOGGER = logging.getLogger(__name__)
+
+# How long the orphan watcher waits for an abandoned coroutine before giving
+# up, so a hung operation does not park a watcher thread forever.
+ORPHAN_WAIT_TIMEOUT = 300.0
+
+
+_runner_ids = count(1)
+
+
+class AsyncDispatchTimeout(TimeoutError):
+    """The caller stopped waiting; the coroutine was abandoned.
+
+    A subclass so callers can tell the dispatcher's own expiry apart from a
+    ``TimeoutError`` raised inside the coroutine, while existing
+    ``except TimeoutError`` handlers keep working.
+    """
 
 
 class AsyncThreadRunner[T](threading.Thread):
     """Run an async coroutine in a daemon thread and expose its result.
 
-    The runner catches all exceptions from the coroutine and stores them in
-    ``exception`` so ``event`` is always set — this prevents callers waiting
-    on ``event`` from hanging forever when the coroutine crashes.
-
-    Typical usage::
-
-        runner = AsyncThreadRunner(lambda: my_coro(arg))
-        runner.start()
-        if not runner.event.wait(timeout=5.0):
-            ...  # timed out
-        if runner.exception is not None:
-            raise runner.exception
-        result = runner.result
+    ``event`` is always set, even when the coroutine crashes, so waiters
+    never hang; ``completed`` distinguishes a delivered result (even a
+    legitimate ``None``) from a coroutine that never finished. Prefer
+    :func:`run_async`; use this class directly only when a failure should
+    degrade to a default value instead of raising.
     """
 
     def __init__(self, coro_factory: Callable[[], Awaitable[T]]) -> None:
-        super().__init__(daemon=True)
+        super().__init__(daemon=True, name=f"async-thread-runner-{next(_runner_ids)}")
         self._coro_factory = coro_factory
         self.result: T | None = None
         self.exception: BaseException | None = None
+        self.completed = False
         self.event = threading.Event()
 
     async def _runner(self) -> None:
         try:
             self.result = await self._coro_factory()
-        except Exception as exc:  # noqa: BLE001  # pylint: disable=broad-except
-            # Capture all exceptions so ``event`` is always set — otherwise a
-            # crash would hang the waiter forever.
+            # Distinguishes a delivered result from "never ran", since None
+            # is a valid result value.
+            self.completed = True
+        except BaseException as exc:  # noqa: BLE001  # pylint: disable=broad-except
+            # Capture everything, including BaseException — otherwise a
+            # cancellation or SystemExit would leave ``exception`` unset and
+            # waiters would mistake the empty ``result`` for success.
             self.exception = exc
         finally:
             self.event.set()
 
     def run(self) -> None:
-        asyncio.run(self._runner())
+        try:
+            asyncio.run(self._runner())
+        except BaseException as exc:  # noqa: BLE001  # pylint: disable=broad-except
+            # asyncio.run itself can fail before _runner executes (e.g. loop
+            # creation under fd exhaustion); record it so waiters never hang.
+            # A failure during loop cleanup after the coroutine completed
+            # must not clobber the delivered result, hence the guard.
+            if self.exception is None and not self.completed:
+                self.exception = exc
+            else:
+                _LOGGER.debug(
+                    "Event loop teardown failed after outcome recorded",
+                    exc_info=True,
+                )
+        finally:
+            self.event.set()
+
+
+def run_async[T](
+    coro_factory: Callable[[], Awaitable[T]],
+    timeout: float | None = None,
+    on_orphan: Callable[[T], None] | None = None,
+) -> T:
+    """Run a coroutine in a daemon-thread event loop and return its result.
+
+    Raises :class:`AsyncDispatchTimeout` if the coroutine does not finish
+    within ``timeout`` seconds; the thread is abandoned and exits with the
+    interpreter. If the abandoned coroutine later produces a result,
+    ``on_orphan`` (if given) is called with it so resources such as a
+    connected socket can be released; delivery is best effort and bounded
+    by ``ORPHAN_WAIT_TIMEOUT``.
+    """
+    runner: AsyncThreadRunner[T] = AsyncThreadRunner(coro_factory)
+    runner.start()
+    if not runner.event.wait(timeout):
+
+        def _cleanup() -> None:
+            if not runner.event.wait(ORPHAN_WAIT_TIMEOUT):
+                # The one state where a resource can genuinely leak; leave
+                # a trace so a recurring hang is attributable.
+                _LOGGER.info(
+                    "Orphan watcher gave up after %.0fs; a late result may leak",
+                    ORPHAN_WAIT_TIMEOUT,
+                )
+                return
+            if not runner.completed:
+                # The only place an abandoned thread's real error surfaces;
+                # without it a late failure hides behind the TimeoutError.
+                # INFO, not DEBUG: it fires at most once per abandoned
+                # operation and the cause may not reproduce on a rerun.
+                _LOGGER.info(
+                    "Abandoned async operation failed",
+                    exc_info=runner.exception,
+                )
+                return
+            if (result := runner.result) is None:
+                return
+            if on_orphan is None:
+                _LOGGER.debug("Discarding late result; no on_orphan handler")
+                return
+            try:
+                on_orphan(result)
+            except Exception:  # pylint: disable=broad-except
+                # INFO, not DEBUG: a failed release means a real leak, and
+                # it fires at most once per abandoned operation.
+                _LOGGER.info("Error releasing orphaned result", exc_info=True)
+
+        threading.Thread(
+            target=_cleanup, daemon=True, name="async-orphan-cleanup"
+        ).start()
+        raise AsyncDispatchTimeout("Timed out waiting for async operation")
+    if (exc := runner.exception) is not None:
+        raise exc
+    if not runner.completed:
+        raise RuntimeError("Async operation finished without a result or an exception")
+    return cast("T", runner.result)