OpenFlow is a switching standard and open protocol enabling distributed control of the flow tables contained within Ethernet switches in a network. Each OpenFlow switch has three parts:
- A datapath, containing a flow table, associating set of actions with each flow entry;
- A secure channel, connecting to a controller; and
- The OpenFlow protocol, used by the controller to talk to switches.
Following this standard model, the implementation comprises three parts:
switch.ml, containing a skeleton OpenFlow switch;
controller.ml, containing a skeleton OpenFlow controller; and
Bitstringparsers/writers for the OpenFlow protocol.
N.B. There are two versions of the OpenFlow protocol: v1.0.0 (
the wire) and v1.1.0 (
0x02 on the wire). The implementation supports wire
0x01 as this is what is implemented in Open vSwitch,
used for debugging.
The file begins with some utility functions, operators, types. The
bulk of the code is organised following the v1.0.0
protocol specification, as implemented by
Open vSwitch v1.2. Each set of messages is contained
within its own module, most of which contain a type
the entity named by the module, plus relevant parsers to convert a
bitstring to a type (
parse_*) and pretty printers for the type
string_of_*). At the end of the file, in the root
module scope, are definitions for interacting with the protocol as a
whole, e.g., error codes, OpenFlow message types and standard header,
root OpenFlow parser, OpenFlow packet builders.
Queue module is really a placeholder currently. OpenFlow
defines limited quality-of-service support via a simple queueing
mechanism. Flows are mapped to queues attached to ports, and each
queue is then configured as desired. The specification currently
defines just a minimum rate, although specific implementations may
Port module wraps several port related elements:
- t, where that is either simply the index of the port in the switch, or the special indexes (> 0xff00) representing the controller, flooding, etc.
- config, a specific port's configuration (up/down, STP supported, etc).
- features, a port's feature set (rate, fiber/copper, etc).
- state, a port's current state (up/down, STP learning mode, etc).
- phy, a port's physical details (index, address, name, etc).
- stats, current statistics of the port (packet and byte counters, collisions, etc).
reason and status, for reporting changes to a port's
configuration; reason is one of
Switch wraps elements pertaining to a whole switch, that is
a collection of ports, tables (including the group table), and the
connection to the controller.
- capabilities, the switch's capabilities in terms of supporting IP fragment reassembly, various statistics, etc.
- action, the types of action the switch's ports support (setting various fields, etc).
- features, the switch's id, number of buffers, tables, port list etc.
- config, for masking against handling of IP fragments: no special handling, drop, reassemble.
Match modules both simply wrap types
respectively representing the fields to wildcard in a flow match, and
the flow match specification itself.
Flow module then contains structures representing:
- t, the flow itself (its age, activity, priority, etc); and
stats, extended statistics association with a flow identified by a
These represent messages associated with receipt or transmission of a packet in response to a controller initiated action.
Packet_in is used where a packet arrives at the switch and is
forwarded to the controller, either due to lack of matching entry, or
an explicit action.
Packet_out contains the structure used by the controller to indicate
to the switch that a packet it has been buffering must now have some
actions performed on it, typically culminating in it being forward out
of one or more ports.
These represent modification messages to existing flow and port state in the switch.
Stats module contains structures representing the
different statistics messages available through OpenFlow, as well as
the request and response messages that transport them.
Initially modelled after NOX, this is a skeleton controller that provides a simple event based wrapper around the OpenFlow protocol. It currently provides the minimal set of events corresponding to basic switch operation:
DATAPATH_JOIN, representing the connection of a datapath to the controller, i.e., notification of the existence of a switch.
DATAPATH_LEAVE, representing the disconnection of a datapath from the controller, i.e., notification of the destruction of a switch.
PACKET_IN, representing the forwarding of a packet to the controller, whether through an explicit action corresponding to a flow match, or simply as the default when flow match is found.
The controller state is mutable and modelled as:
- A list of callbacks per event, each taking the current state, the originating datapath, and the event;
- Mappings from switch (
datapath_id) to a Mirage communications channel (
- Mappings from channel (
endhostcomprising an IPv4 address and port) tp datapath (
The main work of the controller is carried out in
which processes each received packet within the context given by the
current state of the switch: this is where the OpenFlow state machine
The controller entry point is via the
listen function which
effectively creates a receiving channel to parse OpenFlow packets, and
pass them to
process_of_packet which handles a range of standard
protocol-level interactions, e.g.,
generating Mirage events as appropriate. Specifically,
is passed as callback to
Channel.listen, and recursively evaluates
echo to read the incoming packet and pass it to
N.B. This is unwritten as yet, awaiting the new device model being applied to the network stack.
An OpenFlow switch or datapath consists of one or more flow tables, a group table (in later versions, not supported in v1.0.0), and a secure channel back to the controller. Communication over the channel is via the OpenFlow protocol, and is how the controller manages the switch.
In short, each table contains flow entries consisting of match fields, counters, and instructions to apply to packets. Starting with the first flow table, if an incoming packet matches an entry, the counters are updated and the instructions carried out. If no entry in the first table matches, (part of) the packet is forwarded to the controller, or it is dropped, or it proceeds to the next flow table.
Skeleton code is as follows:
Represents a single flow table entry. Each entry consists of:
fields, against which to match (
counters, to keep statistics per-table, -flow, -port, -queue
Entry.queue_counter list); and
actions, to perform on packets matching the fields (
A simple module representing a table of flow entries. Currently just an id
tid) and a list of entries (
Encapsulating the switch (or datapath) itself. Currently defines a port as:
details, a physical port configuration (
device, some handle to the physical device (mocked out as a
The switch is then modelled as:
ports, a list of physical ports (
- table, the table of flow entries for this switch;
stats, a set of per-switch counters (
- p_sflow, the probability in use when sFlow sampling.
Note that the vocabulary of a number of these changes with v1.1.0, in addition to the table structure becoming more complex (support for chains of tables, forwarding to tables, and the group table).
What's the best way to structure the controller so that application code can introduce generation and consumption of new events? NOX permits this within a single event-handling framework -- is this simply out-of-scope here, or should we have a separate event-based programming framework available, or is there a straightforward Ocaml-ish way to incorporate this into the OpenFlow Controller?
What's the best way to expose parsing as a separate activity to reading data
off the wire? Specifically, I'd really like to reuse functions from
Net.Ipv4, etc to recover structure from the bitstring without
need to have
OfPacket.Match.parse_from_raw_packet. Previously I have found
having parsers that return structured data and then wrapping up the packet
structure as a nested type, e.g.,
PCAP(pcaph, ETH(ethh, IPv4(iph, payload)))
...TCP(tcph, payload)))) worked well, permitting fairly natural pattern
matching. The depth to which the packet was deumltiplexed was controlled by a
parameter to the entry-point parser.
Switch design is almost certainly very inefficient, and needs working
on. This is waiting on implementation -- although sketched out, waiting on
network driver model to actually be able to get hold of physical devices and
frames. When we can, also need to consider how to control packet parsing, and
demultiplexing of frames for switching from frames comprising the TCP stream
carrying the controller channel. Ideally, it would be transparent to have
Channel for the controller's OpenFlow messages and a per-device frame
handler for everything else. That is, Mirage would do the necessary
demultiplexing -- but only what's necessary -- passing non-OpenFlow frames to
the switch to be matched, but reassembling the TCP flow carrying the
controller's OpenFlow traffic.