Set max reanimation attempts on HA watchdog (#4784)

supervisor/misc/tasks.py

@@ -15,6 +15,8 @@ from ..utils.sentry import capture_exception
 _LOGGER: logging.Logger = logging.getLogger(__name__)
 
 HASS_WATCHDOG_API = "HASS_WATCHDOG_API"
+HASS_WATCHDOG_REANIMATE_FAILURES = "HASS_WATCHDOG_REANIMATE_FAILURES"
+HASS_WATCHDOG_MAX_REANIMATE_ATTEMPTS = 5
 
 RUN_UPDATE_SUPERVISOR = 29100
 RUN_UPDATE_ADDONS = 57600
@@ -154,6 +156,18 @@ class Tasks(CoreSysAttributes):
             return
         if await self.sys_homeassistant.api.check_api_state():
             # Home Assistant is running properly
+            self._cache[HASS_WATCHDOG_REANIMATE_FAILURES] = 0
+            return
+
+        # Give up after 5 reanimation failures in a row. Supervisor cannot fix this issue.
+        reanimate_fails = self._cache.get(HASS_WATCHDOG_REANIMATE_FAILURES, 0)
+        if reanimate_fails >= HASS_WATCHDOG_MAX_REANIMATE_ATTEMPTS:
+            if reanimate_fails == HASS_WATCHDOG_MAX_REANIMATE_ATTEMPTS:
+                _LOGGER.critical(
+                    "Watchdog cannot reanimate Home Assistant, failed all %s attempts.",
+                    reanimate_fails,
+                )
+                self._cache[HASS_WATCHDOG_REANIMATE_FAILURES] += 1
             return
 
         # Init cache data
@@ -171,7 +185,11 @@ class Tasks(CoreSysAttributes):
             await self.sys_homeassistant.core.restart()
         except HomeAssistantError as err:
             _LOGGER.error("Home Assistant watchdog reanimation failed!")
-            capture_exception(err)
+            if reanimate_fails == 0:
+                capture_exception(err)
+            self._cache[HASS_WATCHDOG_REANIMATE_FAILURES] = reanimate_fails + 1
+        else:
+            self._cache[HASS_WATCHDOG_REANIMATE_FAILURES] = 0
         finally:
             self._cache[HASS_WATCHDOG_API] = 0
 

tests/misc/test_tasks.py

@@ -0,0 +1,133 @@
+"""Test scheduled tasks."""
+
+from unittest.mock import MagicMock, Mock, patch
+
+from awesomeversion import AwesomeVersion
+import pytest
+
+from supervisor.coresys import CoreSys
+from supervisor.exceptions import HomeAssistantError
+from supervisor.homeassistant.api import HomeAssistantAPI
+from supervisor.homeassistant.const import LANDINGPAGE
+from supervisor.homeassistant.core import HomeAssistantCore
+from supervisor.misc.tasks import Tasks
+
+# pylint: disable=protected-access
+
+
+@pytest.fixture(name="tasks")
+async def fixture_tasks(coresys: CoreSys, container: MagicMock) -> Tasks:
+    """Return task manager."""
+    coresys.homeassistant.watchdog = True
+    coresys.homeassistant.version = AwesomeVersion("2023.12.0")
+    container.status = "running"
+    yield Tasks(coresys)
+
+
+async def test_watchdog_homeassistant_api(
+    tasks: Tasks, caplog: pytest.LogCaptureFixture
+):
+    """Test watchdog of homeassistant api."""
+    with patch.object(
+        HomeAssistantAPI, "check_api_state", return_value=False
+    ), patch.object(HomeAssistantCore, "restart") as restart:
+        await tasks._watchdog_homeassistant_api()
+
+        restart.assert_not_called()
+        assert "Watchdog miss API response from Home Assistant" in caplog.text
+        assert "Watchdog found a problem with Home Assistant API!" not in caplog.text
+
+        caplog.clear()
+        await tasks._watchdog_homeassistant_api()
+
+        restart.assert_called_once()
+        assert "Watchdog miss API response from Home Assistant" not in caplog.text
+        assert "Watchdog found a problem with Home Assistant API!" in caplog.text
+
+
+async def test_watchdog_homeassistant_api_off(tasks: Tasks, coresys: CoreSys):
+    """Test watchdog of homeassistant api does not run when disabled."""
+    coresys.homeassistant.watchdog = False
+
+    with patch.object(
+        HomeAssistantAPI, "check_api_state", return_value=False
+    ), patch.object(HomeAssistantCore, "restart") as restart:
+        await tasks._watchdog_homeassistant_api()
+        await tasks._watchdog_homeassistant_api()
+        restart.assert_not_called()
+
+
+async def test_watchdog_homeassistant_api_error_state(tasks: Tasks, coresys: CoreSys):
+    """Test watchdog of homeassistant api does not restart when in error state."""
+    coresys.homeassistant.core._error_state = True
+
+    with patch.object(
+        HomeAssistantAPI, "check_api_state", return_value=False
+    ), patch.object(HomeAssistantCore, "restart") as restart:
+        await tasks._watchdog_homeassistant_api()
+        await tasks._watchdog_homeassistant_api()
+        restart.assert_not_called()
+
+
+async def test_watchdog_homeassistant_api_landing_page(tasks: Tasks, coresys: CoreSys):
+    """Test watchdog of homeassistant api does not monitor landing page."""
+    coresys.homeassistant.version = LANDINGPAGE
+
+    with patch.object(
+        HomeAssistantAPI, "check_api_state", return_value=False
+    ), patch.object(HomeAssistantCore, "restart") as restart:
+        await tasks._watchdog_homeassistant_api()
+        await tasks._watchdog_homeassistant_api()
+        restart.assert_not_called()
+
+
+async def test_watchdog_homeassistant_api_not_running(
+    tasks: Tasks, container: MagicMock
+):
+    """Test watchdog of homeassistant api does not monitor when home assistant not running."""
+    container.status = "stopped"
+
+    with patch.object(
+        HomeAssistantAPI, "check_api_state", return_value=False
+    ), patch.object(HomeAssistantCore, "restart") as restart:
+        await tasks._watchdog_homeassistant_api()
+        await tasks._watchdog_homeassistant_api()
+        restart.assert_not_called()
+
+
+async def test_watchdog_homeassistant_api_reanimation_limit(
+    tasks: Tasks, caplog: pytest.LogCaptureFixture, capture_exception: Mock
+):
+    """Test watchdog of homeassistant api stops after max reanimation failures."""
+    with patch.object(
+        HomeAssistantAPI, "check_api_state", return_value=False
+    ), patch.object(
+        HomeAssistantCore, "restart", side_effect=(err := HomeAssistantError())
+    ) as restart:
+        for _ in range(5):
+            await tasks._watchdog_homeassistant_api()
+            restart.assert_not_called()
+
+            await tasks._watchdog_homeassistant_api()
+            restart.assert_called_once()
+            assert "Home Assistant watchdog reanimation failed!" in caplog.text
+
+            restart.reset_mock()
+
+        capture_exception.assert_called_once_with(err)
+
+        caplog.clear()
+        await tasks._watchdog_homeassistant_api()
+
+        restart.assert_not_called()
+        assert "Watchdog miss API response from Home Assistant" not in caplog.text
+        assert "Watchdog found a problem with Home Assistant API!" not in caplog.text
+        assert (
+            "Watchdog cannot reanimate Home Assistant, failed all 5 attempts."
+            in caplog.text
+        )
+
+        caplog.clear()
+        await tasks._watchdog_homeassistant_api()
+        restart.assert_not_called()
+        assert not caplog.text