yahoo-eng-team team mailing list archive

[Dmitrii Shcherbakov <1783654@xxxxxxxxxxxxxxxxxx>]

** Also affects: cloud-archive
   Importance: Undecided
       Status: New

** Also affects: neutron (Ubuntu)
   Importance: Undecided
       Status: New

-- 
You received this bug notification because you are a member of Yahoo!
Engineering Team, which is subscribed to neutron.
https://bugs.launchpad.net/bugs/1783654

Title:
  DVR process flow not installed on physical bridge for shared tenant
  network

Status in Ubuntu Cloud Archive:
  New
Status in neutron:
  Fix Released
Status in neutron package in Ubuntu:
  New

Bug description:
  Seems like collateral from
  https://bugs.launchpad.net/neutron/+bug/1751396

  In DVR, the distributed gateway port's IP and MAC are shared in the
  qrouter across all hosts.

  The dvr_process_flow on the physical bridge (which replaces the shared
  router_distributed MAC address with the unique per-host MAC when its
  the source), is missing, and so is the drop rule which instructs the
  bridge to drop all traffic destined for the shared distributed MAC.

  Because of this, we are seeing the router MAC on the network
  infrastructure, causing it on flap on br-int on every compute host:

  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
     11     4  fa:16:3e:42:a2:ec    1
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
     11     4  fa:16:3e:42:a2:ec    2
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
      1     4  fa:16:3e:42:a2:ec    0
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
     11     4  fa:16:3e:42:a2:ec    0
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
     11     4  fa:16:3e:42:a2:ec    0
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
      1     4  fa:16:3e:42:a2:ec    0
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
      1     4  fa:16:3e:42:a2:ec    0
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
      1     4  fa:16:3e:42:a2:ec    0
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
      1     4  fa:16:3e:42:a2:ec    1
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
     11     4  fa:16:3e:42:a2:ec    0
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
     11     4  fa:16:3e:42:a2:ec    0
  root@milhouse:~# ovs-appctl fdb/show br-int | grep fa:16:3e:42:a2:ec
     11     4  fa:16:3e:42:a2:ec    0

  
  Where port 1 is phy-br-vlan, connecting to the physical bridge, and port 11 is the correct local qr-interface. Because these dvr flows are missing on br-vlan, pkts w/ source mac ingress into the host and br-int learns it upstream.

  
  The symptom is when pinging a VM's floating IP, we see occasional packet loss (10-30%), and sometimes the responses are sent upstream by br-int instead of the qrouter, so the ICMP replies come with fixed IP of the replier since no NAT'ing took place, and on the tenant network rather than external network.

  When I force net_shared_only to False here, the problem goes away:
  https://github.com/openstack/neutron/blob/stable/pike/neutron/plugins/ml2/drivers/openvswitch/agent/ovs_dvr_neutron_agent.py#L436

  It should we noted we *ONLY* need to do this on our dvr_snat host. The
  dvr process's are missing on every compute host. But if we shut
  qrouter on the snat host, FIP functionality works and DVR mac stops
  flapping on others. Or if we apply fix only to snat host, it works.
  Perhaps there is something on SNAT node that is unique

To manage notifications about this bug go to:
https://bugs.launchpad.net/cloud-archive/+bug/1783654/+subscriptions