Merge "libvirt: handle code=38 + sigkill (ebusy) in destroy()" into stable/kilo

nova/tests/unit/virt/libvirt/test_driver.py

@@ -8366,6 +8366,54 @@ class LibvirtConnTestCase(test.NoDBTestCase):
             self.assertRaises(fakelibvirt.libvirtError, conn._destroy,
                               instance)
 
+    def test_private_destroy_ebusy_timeout(self):
+        # Tests that _destroy will retry 3 times to destroy the guest when an
+        # EBUSY is raised, but eventually times out and raises the libvirtError
+        ex = fakelibvirt.make_libvirtError(
+                fakelibvirt.libvirtError,
+                ("Failed to terminate process 26425 with SIGKILL: "
+                 "Device or resource busy"),
+                error_code=fakelibvirt.VIR_ERR_SYSTEM_ERROR,
+                int1=errno.EBUSY)
+
+        instance = objects.Instance(**self.test_instance)
+        drvr = libvirt_driver.LibvirtDriver(fake.FakeVirtAPI(), False)
+
+        with mock.patch.object(drvr._host, 'get_domain') as mock_get_domain:
+            mock_domain = mock.MagicMock()
+            mock_domain.ID.return_value = 1
+            mock_get_domain.return_value = mock_domain
+            mock_domain.destroy.side_effect = ex
+
+            self.assertRaises(fakelibvirt.libvirtError, drvr._destroy,
+                              instance)
+
+        self.assertEqual(3, mock_domain.destroy.call_count)
+
+    def test_private_destroy_ebusy_multiple_attempt_ok(self):
+        # Tests that the _destroy attempt loop is broken when EBUSY is no
+        # longer raised.
+        ex = fakelibvirt.make_libvirtError(
+                fakelibvirt.libvirtError,
+                ("Failed to terminate process 26425 with SIGKILL: "
+                 "Device or resource busy"),
+                error_code=fakelibvirt.VIR_ERR_SYSTEM_ERROR,
+                int1=errno.EBUSY)
+
+        inst_info = hardware.InstanceInfo(power_state.SHUTDOWN, id=1)
+        instance = objects.Instance(**self.test_instance)
+        drvr = libvirt_driver.LibvirtDriver(fake.FakeVirtAPI(), False)
+
+        with mock.patch.object(drvr._host, 'get_domain') as mock_get_domain, \
+             mock.patch.object(drvr, 'get_info', return_value=inst_info):
+                mock_domain = mock.MagicMock()
+                mock_domain.ID.return_value = 1
+                mock_get_domain.return_value = mock_domain
+                mock_domain.destroy.side_effect = ex, None
+                drvr._destroy(instance)
+
+        self.assertEqual(2, mock_domain.destroy.call_count)
+
     def test_undefine_domain_with_not_found_instance(self):
         def fake_get_domain(self, instance):
             raise exception.InstanceNotFound(instance_id=instance.name)

nova/virt/libvirt/driver.py

@@ -617,7 +617,7 @@ class LibvirtDriver(driver.ComputeDriver):
         rootfs_dev = instance.system_metadata.get('rootfs_device_name')
         disk.teardown_container(container_dir, rootfs_dev)
 
-    def _destroy(self, instance):
+    def _destroy(self, instance, attempt=1):
         try:
             virt_dom = self._host.get_domain(instance)
         except exception.InstanceNotFound:
@@ -660,6 +660,34 @@ class LibvirtDriver(driver.ComputeDriver):
                             instance=instance)
                     reason = _("operation time out")
                     raise exception.InstancePowerOffFailure(reason=reason)
+                elif errcode == libvirt.VIR_ERR_SYSTEM_ERROR:
+                    if e.get_int1() == errno.EBUSY:
+                        # NOTE(danpb): When libvirt kills a process it sends it
+                        # SIGTERM first and waits 10 seconds. If it hasn't gone
+                        # it sends SIGKILL and waits another 5 seconds. If it
+                        # still hasn't gone then you get this EBUSY error.
+                        # Usually when a QEMU process fails to go away upon
+                        # SIGKILL it is because it is stuck in an
+                        # uninterruptable kernel sleep waiting on I/O from
+                        # some non-responsive server.
+                        # Given the CPU load of the gate tests though, it is
+                        # conceivable that the 15 second timeout is too short,
+                        # particularly if the VM running tempest has a high
+                        # steal time from the cloud host. ie 15 wallclock
+                        # seconds may have passed, but the VM might have only
+                        # have a few seconds of scheduled run time.
+                        LOG.warn(_LW('Error from libvirt during destroy. '
+                                     'Code=%(errcode)s Error=%(e)s; '
+                                     'attempt %(attempt)d of 3'),
+                                 {'errcode': errcode, 'e': e,
+                                  'attempt': attempt},
+                                 instance=instance)
+                        with excutils.save_and_reraise_exception() as ctxt:
+                            # Try up to 3 times before giving up.
+                            if attempt < 3:
+                                ctxt.reraise = False
+                                self._destroy(instance, attempt + 1)
+                                return
 
                 if not is_okay:
                     with excutils.save_and_reraise_exception():