L2 Extension flow management support


The l2-agent-extension-api that merged in Mitaka allows extensions of the Open vSwitch agent to manipulate the Open vSwitch OpenFlow tables.

This introduces a challenge of interoperability between extensions creating flows.

Problem Description

As the Open vSwitch flow table matches traffic to flows based on packet characteristics in the order of priority assigned to the flow, it will inevitably result in two extensions using flows that will block the extension that uses the lower priority flows.

Therefore a flow manager or flow pipeline is necessary, that would reconcile flows for multiple extensions in a consistent and mutually compatible way.

Proposed Change

In order to make extensions optional, pluggable and interoperable in the context of OpenFlow rules we need to define a pipeline model, and a l2 extension API interface, with the main purpose of avoiding cases where two extensions manipulate the same exact table, inserting eventually incompatible flows, and to let extensions:

  • declare OpenFlowFunctions, where an OpenFlowFunction is a set of tables that implement the extension behaviour and can be plugged in either the egress or ingress pipeline.

  • declare the need to reference other OpenFlowFunctions within the same extension (for example an extension could need to learn flows into another of its own functions on a different stage of the pipeline ingress/egress, a typical example of this is mac-learning ingress function learns to an egress function).

  • declare OpenFlowRegisters, which will be used inside the function, and sometimes can be used by other functions as a medium to exchange metadata.

  • declare the need for a specific OpenFlowRegister, generally the needed registers are filled or received up by the fixed blocks, but eventually extensions, or different extension functions may need to interoperate by exchanging metadata in registers.

    Public registers or functions can be accessed by other extensions, but non public registers can’t. This sets a minimum base for extension interoperation, where that makes sense.

Three classes will be provided for the extensions and other OpenFlow mechanisms (like the OpenFlow firewall driver, or the proposed fixed blocks -see below-) to declare their OpenFlow functions:

  • OpenFlowFunction

  • OpenFlowTable

  • OpenFlowRegister

The code could roughly look like:

class EgressTrafficClassifier(openflow_pipeline.OpenFlowFunction):

    extension = "FIXED_BLOCKS"
    bridge = openflow_pipeline.INTEGRATION_BRIDGE

    declares_registers = [OpenFlowRegister(bits=16, name='CLASSIFIER_MARK',
    declares_tables = [OpenFlowTable(name='CLASSIFIER', default=NEXT)]
    needs_registers = [OpenFlowRegister(name='OUTPUT_PORT')]
    needs_function = [OpenFlowfunction(name='INGRESS_TRAFFIC_X'),

    # The openflow manager will register the following attributes on the
    # class once it has resolved all dependencies and creations
    #          <name>_table
    #          <name>_reg
    #          <name>_function
    # for all the declared and needed registers, so in this example
    # we would have the following class attributes:
    #     - classifier_mark_reg
    #     - classifier_table
    #     - output_port_reg
    #     - ingress_traffic_x_function
    #     - input_tagging_function

A top level OpenFlow manager, used by the l2 extension manager will be responsible of taking all the OpenFlowFunctions declared, resolving order, and assigning table numbers and registers to every function.

OpenFlowFunctions can choose the bridge or where they want to live (provider network bridges, integration bridge, tunnel bridge), taking baby steps we will start by supporting functions that want to live in the integration bridge, but the bridge intention should be declared in the function to allow future extensibility.

The OpenFlowManager will inspect the declared tables default action, which can be:

  • NEXT, the lowest priority rule inserted in table jumping to next OpenFlow function.

  • DROP, the lowest priority rule inserted in the table drops the packet.

Description of the integration bridge pipeline

The pipeline is divided in two paths, from the neutron port of VM instance point of view:

  • Egress path (traffic that leaves the VM/port)

  • Ingress path (traffic that goes to a VM/port)

The pipeline looks like this:

           +<--+ vm-tap-a <---------------+
           |   +----------+               |
           |                              |
           |   +----------+               |
           |<--+ vm-tap-b <-----+         |
           |   +--+-------+     |         |
           |                    |         |
     0+----v------------+ 253+--+---------+--+
      | initial_tagging |  |  input_stage    |
      +-----------+-----+  +--------^--------+
                  |                 |
      +-----------v-----+  +--------+--------+
 E    |  [ function A ] |  | [ function .. ] |
 G    +-----------+-----+  +--------^--------+
 R                |                 |
 E    +-----------v-----+  +--------+--------+
 S    |  [ function ..] |  | [ function F ]  |
 S    +-----------+-----+  +--------^--------+
                  |                 |
 |    +-----------v-----+  +--------+--------+   .
 |    | egress filter   |  |  ingress filter |  /|\
\|/   +-----------+-----+  +--------^--------+   |
 °                |                 |            |
      +-----------v-----+  +--------+--------+
      |  [ function C ] |  | [ function .. ] |   I
      +-----------+-----+  +--------^--------+   N
                  |                 |            G
      +-----------v-----+  +--------+--------+   R
      |  [ function ..] |  ^  [ function E ] |   E
      +-----------+-----+ /+--------^--------+   S
                  |      /          |            S
      +-----------v-----+ 0+--------+--------+
      |ingress deflector|  | initial_tagging |
      +----------+------+  +---------^-------+
                 |                   |
      +----------v------+            |
      |  output_stage   +------+     |
      +-----------+-----+      |     |
                               |     |
                             | patch-br-xxx  |


  • vm-tap-a and vm-tap-b: are VMs or service ports attached to a bridge.

  • patch-br-xxx: is the patch port going to bridge br-xxx

  • Functions can be made of several tables, it’s the responsibility of the function to jump around its own tables or to the next function once its processing has finished.

OpenFlowFunction capabilities

OpenFlow functions would have the right to add_flows/del_flows to normally manipulate packets as usual in the next cases:

  • Drop a packet

  • Alter packet details

  • Set a queue

  • push/pop data to the stack (stack size must not change after ending function)

  • learn actions in other functions belonging to the same extension

  • etc

Special handling is required for:

  • NORMAL action: normal action should not be executed immediately via action, but a register will be marked so the specific packet will be handled with “NORMAL” at output stage.

  • Modify destination on the ingress path: the input tagging details must be reset, and based on the locality/externality of the packet destination, update of input tagging fields (using a helper).

  • Output a clone of the packet: “output:X,goto_table:<next-function-first-table>” may be used.

  • Redirect packet to a new destination and stop pipeline processing (for extensions that want to move the packet to an ancillary bridge and are completely sure that further processing on the bridge should not happen) will use output:X with no resubmit to next function.

Helpers will be provided, for cases like the “NORMAL” one, or packet cloning, and we may want to consider some filter to enforce sanity of the actions.


Inserted flows will have to be re-inserted by extensions by calling their openflow functions as necessary when, for example, openvswitch has been reset and the flows need to be recreated.

We could have a DSL to define how flows are created for each function, but that would increase the complexity of this, and we prefer to take baby steps toward a better integration between extensions. In a future we could consider such option we have the foundations of a first openflow pipeline.

OpenFlowFunction fixed blocks (integration bridge)

Some common OpenFlowFunctions are provided by the Open vSwitch agent, the agent will declare and register those:

  • INITIAL_TAGGING: Takes care of identifying the source or destination of the packet, marks the network in a register, and sends the packet to the egress or ingress path (with priority for egress path when both the source and destination of the packet is local)

    Some extensions may want to change the destination, or duplicate packets with new destinations. Helpers should be provided for that kind of manipulation, so the registers will resemble the same settings done by INITIAL_TAGGING function.

    The egress path has priority to cover the case when a packet is both egress and ingress in the same bridge (source and destination ports are local to the hypervisor)

  • INGRESS_DEFLECTOR: For packets that are both egress and ingress on the local hypervisor, this stage is in charge of moving packets from the end of the egress path, to the start of the ingress path, so any ingress filtering or packet processing happens before the packet is finally sent to the output port.

  • INGRESS_FILTER, EGRESS_FILTER: Is provided by the security groups firewall driver (openvswitch firewall), if no OpenFlow filter is provided such stage of the pipeline is reserved for the L2 filtering of the hybrid firewall used for L2 MAC filtering to overcome the iptables/ebtables limitations.

  • INPUT_STAGE: Goes after the ingress path, the packet is directed to its destination port, or gets the NORMAL rule executed based on the packet details. In a future revision, we could get rid of all NORMAL rules, with that, transparent vlans and port trunking could be implemented as simple OpenFlow functions, with no need for separate bridges.

  • OUTPUT_STAGE: Goes after the egress path, the packet is directed to its destination patch port or with NORMAL, when going to another bridge.

Each l2-agent extension for the Open vSwitch agent can declare one or several functional blocks to be plugged.

Every functional block can be composed of one or more OpenFlow tables.

Every functional block has an entry and at least one exit point (next function). The manager will insert a default jump to next function as the lowest priority rule for every function table if table “default” is NEXT.

Ordering of functions

Ordering of functional blocks is defined in the manager, and maintained in the core neutron. If somebody wants to declare a new function in the pipeline, they can come to the community and discuss about its insertion point, alternatively we will include common extension points that can be used right away.

COMMON prefixed points can be used by several extensions, and the ordering between extension functions won’t be guaranteed.

Extensions that require several plug points should declare all of them as separate OpenFlow functions, and make use of them as necessary (note the <>_<DIR>_PREFILTER / <>_<DIR>_POSTFILTER examples below)

An example of the functional blocks per pipeline could be:



Please note that the fixed blocks (INITIAL_TAGGING, INPUT_STAGE, OUTPUT_STAGE, INGRESS_DEFLECTOR) are not included in the configuration definition, as are neither ingress or egress pipeline, but blocks that will steer traffic to the right pipelines.


[1] https://www.opennetworking.org/images/stories/downloads/sdn-resources/onf-specifications/openflow/openflow-switch-v1.4.1.pdf