[L3][DVR][vlan] east-west traffic flooding on physical bridges

[LIU Yulong]

For L3 DVR of VLAN networks, the east west traffic between router subnets will be flooding on the physical bridge.

Assuming we have resources like this:

1. subnet-1 10.10.10.0/24 with gateway-1 10.10.10.1, mac-address-[01]
2. subnet-2 20.20.20.0/24 with gateway-2 20.20.20.1, mac-address-[02]
3. subnet-1 and subnet-2 are all connected to router-1
4. host-1 with dvr_host_mac-[03]
5. host-2 with dvr_host_mac-[04]
6. vm-1 on host-1 with IP 10.10.10.100 and mac-address-[05]
7. vm-2 on host-2 with IP 20.20.20.100 and mac-address-[06]

The packet procedure looks like this:
1. VM-1 is trying to ping VM-2
2. VM-1 gets the gateway-1's MAC address, and send ping request out.
    packet tuple (src_mac=[05], des_mac=[01], src_ip=10.10.10.100, dst_ip=20.20.20.100)
3. packet goes to dvr local router namespace, and routing
4. packet source mac address is changed to mac-address-[02] and destination to mac-addres-[06].
    packet tuple (src_mac=[02], des_mac=[06], src_ip=10.10.10.100, dst_ip=20.20.20.100)
5. the source MAC address will be changed to dvr_host_mac-[03] on physical bridge (br-vlan), and hit NORMAL
   packet tuple (src_mac=[03], des_mac=[06], src_ip=10.10.10.100, dst_ip=20.20.20.100)

* flooding happens, because the br-vlan had never learnt the destination mac address of VM-2 mac-address-[06].

  Tracing the packets, you will see:
    NORMAL
     -> no learned MAC for destination, flooding

  Typically, the flooding action of fast flows are:
  actions:set(eth(src=fa:16:3f:xx:xx:03)),push_vlan(vid=XXXXX,pcp=0),br-vlan,NIC_PORT

6. On host-2, it hits a dvr_host_mac-[03] flows in table=3 and output to patch-to-int-br, which has a side effect is that the source mac address will not be learnt.

packet tuple (src_mac=[03], des_mac=[06], src_ip=10.10.10.100, dst_ip=20.20.20.100)

flows: table=3, priority=2,dl_src=fa:16:3f:xx:xx:03 actions=output:"phy-br-vlan"

7. On host-2, VM-2 send reply back, and finally go to br-vlan, it flooding again

packet tuple (src_mac=[04], des_mac=[05], src_ip=20.20.20.100, dst_ip=10.10.10.100)

8. ping reply back to host-2, on br-vlan, it hits the output flows which match the dvr_host_mac-[04], then the VM-2's mac address will never be learnt.

So, ping request is flooding on br-vlan forever.

[Rodolfo Alonso (rodolfo-alonso-hernandez)]

Hi Liu:

Are you working on this bug?

Regards.

[LIU Yulong (dragon889)]

yes, I will upload the patch today.

[OpenStack Infra (hudson-openstack)]

Fix proposed to branch: master
Review: https://review.opendev.org/c/openstack/neutron/+/868150