Soichi Shigeta <shigeta.soi...@jp.fujitsu.com> wrote:
Hi, We find a problem that neutron ovs-agent deletes taas flows. o) Problem description: Background: At Liberty, a bug fix to drop only old flows was merged to Neutron. When ovs-agent is restarted, the cleanup logic drops flow entries unless they are stamped by agent_uuid (recorded as a cookie). bug: #1383674 "Restarting neutron openvswitch agent causes network hiccup by throwing away all flows" https://bugs.launchpad.net/neutron/+bug/1383674 commit: 73673beacd75a2d9f51f15b284f1b458d32e992e (patch) https://git.openstack.org/cgit/openstack/neutron/commit/?id=73673beacd75a2d9f51f15b284f1b458d32e992e Problem: Cleanup will be done only once, but it seems not to work until port configuration is changed. Therefore, taas flows will be deleted as follows: 1. Start a new compute node or restart an existing compute node. 2. Start taas agent on the compute node. --> taas agent creates flows (these flows are not stamped by using ovs-agent's uuid) 3. Deploy a vm on the compute node. --> 1. neutron changes port configuration 2. subsequently, the cleanup logic is invoked 3. ovs-agent drops taas flows Specifically, following taas flows in br_tun are dropped: ----- table=35, priority=2,reg0=0x0 actions=resubmit(,36) table=35, priority=1,reg0=0x1 actions=resubmit(,36) table=35, priority=1,reg0=0x2 actions=resubmit(,37) ----- log in q-agt.log ----- neutron.plugins.ml2.drivers.openvswitch.agent.openflow.ovs_ofctl.ofswitch req-e5739280-7116-4802-b5ba-d6964b4c5557 Deleting flow cookie=0x0, duration=434.59s, table=35, n_packets=0, n_bytes=0, idle_age=434, priority=2,reg0=0x0 actions=resubmit(,36) neutron.plugins.ml2.drivers.openvswitch.agent.openflow.ovs_ofctl.ofswitch req-e5739280-7116-4802-b5ba-d6964b4c5557 Deleting flow cookie=0x0, duration=434.587s, table=35, n_packets=0, n_bytes=0, idle_age=434, priority=1,reg0=0x1 actions=resubmit(,36) neutron.plugins.ml2.drivers.openvswitch.agent.openflow.ovs_ofctl.ofswitch req-e5739280-7116-4802-b5ba-d6964b4c5557 Deleting flow cookie=0x0, duration=434.583s, table=35, n_packets=0, n_bytes=0, idle_age=434, priority=1,reg0=0x2 actions=resubmit(,37) ----- o) Impact for TaaS: Because flows in br_tun are dropped by the cleanup logic, mirrored packets will not send to a monitoring vm running on another host. Note: Mirrored packets are sent in case of both source vm and monitoring vm are running on the same host. (not affected by flows in br_tun) o) How to reproduce: 1. Start a new compute node or restart an existing compute node. (Actually, restarting ovs-agent is enough.) 2. Start (or restart) taas agent on the compute node. 3. Deploy a vm on the compute node. --> The cleanup logic drops taas flows. o) Workaround: After a vm is deployed on a (re)started compute node, restart taas agent before creating a tap-service or tap-flow. That is, create taas flows after cleanup has been done. Note that cleanup will be done only once during an ovs-agent is running. o) An idea to fix: 1. Set "taas" stamp(*) to taas flows. 2. Modify the cleanup logic in ovs-agent not to delete entries stamped as "taas". * Maybe a static string. If we need to use a string which generated dynamically (e.g. uuid), API to interact with ovs-agent is required.
API proposal with some consideration for flow cleanup not dropping flows for external code is covered in the following email thread: http://lists.openstack.org/pipermail/openstack-dev/2015-December/081264.html
I believe you would need to adopt the extensions API once it’s in, moving from setup with a separate agent for your feature to l2 agent extension for taas that will run inside OVS agent.
Ihar __________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev