Fix concurrent port binding activate

Fix an issue with concurrent requests to activate a port binding.
If there are two activate requests in parallel, one might set the
binding on the new host to active and the other request may
not find the previously INACTIVE row anymore in
_commit_port_binding and initializing the driver_context.PortContext
crashed.

Closes-Bug: #1986003
Change-Id: I047e33062bc38f36848e0149c6e670cb5828c8e3
(cherry picked from commit 5b4ed5b117)

neutron/plugins/ml2/plugin.py

@@ -680,6 +680,14 @@ class Ml2Plugin(db_base_plugin_v2.NeutronDbPluginV2,
                 # transaction that completed before the deletion.
                 LOG.debug("Port %s has been deleted concurrently", port_id)
                 return orig_context, False, False
+
+            if (new_binding.status == const.INACTIVE and
+                    new_binding.host == cur_binding.host):
+                # The binding is already active on the target host,
+                # probably because of a concurrent activate request.
+                raise exc.PortBindingAlreadyActive(port_id=port_id,
+                                                   host=new_binding.host)
+
             # Since the mechanism driver bind_port() calls must be made
             # outside a DB transaction locking the port state, it is
             # possible (but unlikely) that the port's state could change

neutron/tests/unit/plugins/ml2/test_port_binding.py

@@ -13,6 +13,7 @@
 #    License for the specific language governing permissions and limitations
 #    under the License.
 
+from concurrent import futures
 from unittest import mock
 
 from neutron_lib.api.definitions import portbindings
@@ -558,6 +559,23 @@ class ExtendedPortBindingTestCase(test_plugin.NeutronDbPluginV2TestCase):
             retrieved_bindings, const.INACTIVE)
         self._assert_unbound_port_binding(retrieved_inactive_binding)
 
+    def test_activate_port_binding_concurrency(self):
+        port, _ = self._create_port_and_binding()
+        with mock.patch.object(mechanism_test.TestMechanismDriver,
+                               '_check_port_context'):
+            with futures.ThreadPoolExecutor() as executor:
+                f1 = executor.submit(
+                    self._activate_port_binding, port['id'], self.host)
+                f2 = executor.submit(
+                    self._activate_port_binding, port['id'], self.host)
+                result_1 = f1.result()
+                result_2 = f2.result()
+
+        # One request should be successful and the other should receive a
+        # HTTPConflict. The order is arbitrary.
+        self.assertEqual({webob.exc.HTTPConflict.code, webob.exc.HTTPOk.code},
+                         {result_1.status_int, result_2.status_int})
+
     def test_activate_port_binding_for_non_compute_owner(self):
         port, new_binding = self._create_port_and_binding()
         data = {'port': {'device_owner': ''}}

releasenotes/notes/bug-1986003-9bf5ca04f9304336.yaml

@@ -0,0 +1,10 @@
+---
+fixes:
+  - |
+    `1986003 <https://bugs.launchpad.net/neutron/+bug/1986003>`_
+    Fixed an issue with concurrent requests to activate the same port binding
+    where one of the requests returned a 500 Internal Server Error.
+    With the fix one request will return successfully and the other will
+    return a 409 Conflict (Binding already active).
+    This fixes errors in nova live-migrations where those concurrent requests
+    might be sent. Nova handles the 409/Conflict response gracefully.