Comment 0 for bug 1973276

Revision history for this message
Gregory Thiemonge (gthiemonge) wrote :

Bug found in Octavia (master)

Octavia creates at least 2 ports for each load balancer:
- the VIP port, it is down, it keeps/stores the IP address of the LB
- the VRRP port, plugged into a VM, it has the VIP address in the allowed-address list (and the VIP address is configured on the interface in the VM)

When sending an ARP request for the VIP address, the VRRP port should reply with its mac-address.

In OVN the VIP port is marked as "type: virtual".

But when the VIP port is updated, it loses its "port: virtual" status and that breaks the ARP resolution (OVN replies to the ARP request by sending the mac-address of the VIP port - which is not used/down).

Quick reproducer that simulates the Octavia behavior:

===========================

import subprocess
import time

import openstack

conn = openstack.connect(cloud="devstack-admin-demo")

network = conn.network.find_network("public")

sg = conn.network.find_security_group('sg')
if not sg:
    sg = conn.network.create_security_group(name='sg')

vip_port = conn.network.create_port(
    name="lb-vip",
    network_id=network.id,
    device_id="lb-1",
    device_owner="me",
    is_admin_state_up=False)

vip_address = [
    fixed_ip['ip_address']
    for fixed_ip in vip_port.fixed_ips
    if '.' in fixed_ip['ip_address']][0]

vrrp_port = conn.network.create_port(
    name="lb-vrrp",
    device_id="vrrp",
    device_owner="vm",
    network_id=network.id)
vrrp_port = conn.network.update_port(
    vrrp_port,
    allowed_address_pairs=[
        {"ip_address": vip_address,
         "mac_address": vrrp_port.mac_address}])

time.sleep(1)

output = subprocess.check_output(
    f"sudo ovn-nbctl show | grep -A2 'port {vip_port.id}'",
    shell=True)
output = output.decode('utf-8')

if 'type: virtual' in output:
    print("Port is virtual, this is ok.")
print(output)

conn.network.update_port(
    vip_port,
    security_group_ids=[sg.id])

time.sleep(1)

output = subprocess.check_output(
    f"sudo ovn-nbctl show | grep -A2 'port {vip_port.id}'",
    shell=True)
output = output.decode('utf-8')

if 'type: virtual' not in output:
    print("Port is not virtual, this is an issue.")
print(output)

===========================

In my env (devstack master on c9s):
$ python3 /mnt/host/virtual_port_issue.py
Port is virtual, this is ok.
    port e0fe2894-e306-42d9-8c5e-6e77b77659e2 (aka lb-vip)
        type: virtual
        addresses: ["fa:16:3e:93:00:8f 172.24.4.111 2001:db8::178"]

Port is not virtual, this is an issue.
    port e0fe2894-e306-42d9-8c5e-6e77b77659e2 (aka lb-vip)
        addresses: ["fa:16:3e:93:00:8f 172.24.4.111 2001:db8::178"]
    port 8ec36278-82b1-436b-bc5e-ea03ef22192f

In Octavia, the "port: virtual" is _sometimes_ back after other updates of the ports, but in some cases the LB is unreachable.

(and "ovn-nbctl lsp-set-type <vip-port-id> virtual" fixes the LB)