Tunnel ports are not cleaned up with several ovs agents restart

[Ann Taraday]

If we update local_ip config value on several openvswitch agents and restart them we may get tunnel ports with old local_ip(remote_ip) setting are not cleaned up.

Before

        Port "vxlan-c0a80006"
            Interface "vxlan-c0a80006"
                type: vxlan
                options: {df_default="true", egress_pkt_mark="0", in_key=flow, local_ip="192.168.0.7", out_key=flow, remote_ip="192.168.0.6"}
        Port "vxlan-c0a89b04"
            Interface "vxlan-c0a89b04"
                type: vxlan
                options: {df_default="true", egress_pkt_mark="0", in_key=flow, local_ip="192.168.0.7", out_key=flow, remote_ip="192.168.155.4"}

After restart

        Port "vxlan-c0a80006"
            Interface "vxlan-c0a80006"
                type: vxlan
                options: {df_default="true", egress_pkt_mark="0", in_key=flow, local_ip="192.168.0.7", out_key=flow, remote_ip="192.168.0.6"}
        Port "vxlan-c0a89b04"
            Interface "vxlan-c0a89b04"
                type: vxlan
                options: {df_default="true", egress_pkt_mark="0", in_key=flow, local_ip="192.168.155.10", out_key=flow, remote_ip="192.168.155.4"}
        Port "vxlan-c0a89b05"
            Interface "vxlan-c0a89b05"
                type: vxlan
                options: {df_default="true", egress_pkt_mark="0", in_key=flow, local_ip="192.168.155.10", out_key=flow, remote_ip="192.168.155.5"}

If agents restart one by one this issue does not appear.

Debug shows that during tunnel_sync https://github.com/openstack/neutron/blob/master/neutron/plugins/ml2/drivers/openvswitch/agent/ovs_neutron_agent.py#L2240 rpc call https://github.com/openstack/neutron/blob/master/neutron/plugins/ml2/drivers/type_tunnel.py#L538 is getting interrupted http://paste.openstack.org/show/792876/ and some tunnels are not proceeded properly.

Originally found on Queens version, but is reproduced with master code and multinode devstack.

[Miguel Lavalle (minsel)]

In the failed case, what is the sequence you followed to update the config file and re-start the agents?

[Ann Taraday (akamyshnikova)]

In failed case for devstack: change config files on all nodes, then restart openvswitch agents almost at once with minimum delay.