ovn: Disable activation-strategy=rarp for DPDK ports

When vhosuser* is used, qemu won't send RARP; instead, it will request
guest's virtio to announce the guest. Which ultimately generates GARP
(for IPv4) and NA (for IPv6) addresses.

This results in port not being activated until it's too late (only when
nova updates neutron via api).

It's better to have activation-strategy disabled until ovn-controller
learns how to activate with GARPs and NAs.

Related-Bug: #2092407
Change-Id: I71e6ec0d87adec629262c5a488bc9739f78ad6f8

neutron/plugins/ml2/drivers/ovn/mech_driver/ovsdb/ovn_client.py

@@ -501,7 +501,11 @@ class OVNClient:
                         # Block traffic on destination host until libvirt sends
                         # a RARP packet from it to inform network about the new
                         # location of the port
-                        options['activation-strategy'] = 'rarp'
+                        # TODO(ihrachys) Remove this once OVN properly supports
+                        # activation of DPDK ports (bug 2092407)
+                        if (port[portbindings.VIF_TYPE] !=
+                                portbindings.VIF_TYPE_VHOST_USER):
+                            options['activation-strategy'] = 'rarp'
 
                 # Virtual ports can not be bound by using the requested-chassis
                 # mechanism, ovn-controller will create the Port_Binding entry

neutron/tests/unit/plugins/ml2/drivers/ovn/mech_driver/test_mech_driver.py

@@ -2069,7 +2069,8 @@ class TestOVNMechanismDriver(TestOVNMechanismDriverBase):
             portbindings.HOST_ID: 'fake-src',
             portbindings.PROFILE: {
                 ovn_const.MIGRATING_ATTR: 'fake-dest',
-            }
+            },
+            portbindings.VIF_TYPE: portbindings.VIF_TYPE_OVS,
         }
         with mock.patch.object(
                 self.mech_driver._ovn_client._sb_idl, 'is_col_present',
@@ -2079,6 +2080,27 @@ class TestOVNMechanismDriver(TestOVNMechanismDriverBase):
         self.assertEqual('fake-src,fake-dest',
                          options.options['requested-chassis'])
 
+    def test__get_port_options_migrating_vhostuser(self):
+        port = {
+            'id': 'virt-port',
+            'mac_address': '00:00:00:00:00:00',
+            'device_owner': 'device_owner',
+            'network_id': 'foo',
+            'fixed_ips': [],
+            portbindings.HOST_ID: 'fake-src',
+            portbindings.PROFILE: {
+                ovn_const.MIGRATING_ATTR: 'fake-dest',
+            },
+            portbindings.VIF_TYPE: portbindings.VIF_TYPE_VHOST_USER,
+        }
+        with mock.patch.object(
+                self.mech_driver._ovn_client._sb_idl, 'is_col_present',
+                return_value=True):
+            options = self.mech_driver._ovn_client._get_port_options(port)
+        self.assertNotIn('activation-strategy', options.options)
+        self.assertEqual('fake-src,fake-dest',
+                         options.options['requested-chassis'])
+
     def test__get_port_options_not_migrating_additional_chassis_present(self):
         port = {
             'id': 'virt-port',

releasenotes/notes/disable-activation-strategy-rarp-for-dpdk-ports-7f9164e57c578b95.yaml

@@ -0,0 +1,9 @@
+---
+fixes:
+  - |
+    For OVN DPDK ports, live migration activation strategy that expects a RARP
+    frame sent by QEMU is no longer used. This is because for DPDK ports, QEMU
+    does not send a RARP frame, which affects the time to recover network
+    connectivity for DPDK ports after live migration is complete. Note that
+    because of the change, some low number of duplicate packets from these
+    ports may be observed during live migration.