openvswitch.txt 10.9 KB
Open vSwitch datapath developer documentation
=============================================

The Open vSwitch kernel module allows flexible userspace control over
flow-level packet processing on selected network devices.  It can be
used to implement a plain Ethernet switch, network device bonding,
VLAN processing, network access control, flow-based network control,
and so on.

The kernel module implements multiple "datapaths" (analogous to
bridges), each of which can have multiple "vports" (analogous to ports
within a bridge).  Each datapath also has associated with it a "flow
table" that userspace populates with "flows" that map from keys based
on packet headers and metadata to sets of actions.  The most common
action forwards the packet to another vport; other actions are also
implemented.

When a packet arrives on a vport, the kernel module processes it by
extracting its flow key and looking it up in the flow table.  If there
is a matching flow, it executes the associated actions.  If there is
no match, it queues the packet to userspace for processing (as part of
its processing, userspace will likely set up a flow to handle further
packets of the same type entirely in-kernel).


Flow key compatibility
----------------------

Network protocols evolve over time.  New protocols become important
and existing protocols lose their prominence.  For the Open vSwitch
kernel module to remain relevant, it must be possible for newer
versions to parse additional protocols as part of the flow key.  It
might even be desirable, someday, to drop support for parsing
protocols that have become obsolete.  Therefore, the Netlink interface
to Open vSwitch is designed to allow carefully written userspace
applications to work with any version of the flow key, past or future.

To support this forward and backward compatibility, whenever the
kernel module passes a packet to userspace, it also passes along the
flow key that it parsed from the packet.  Userspace then extracts its
own notion of a flow key from the packet and compares it against the
kernel-provided version:

    - If userspace's notion of the flow key for the packet matches the
      kernel's, then nothing special is necessary.

    - If the kernel's flow key includes more fields than the userspace
      version of the flow key, for example if the kernel decoded IPv6
      headers but userspace stopped at the Ethernet type (because it
      does not understand IPv6), then again nothing special is
      necessary.  Userspace can still set up a flow in the usual way,
      as long as it uses the kernel-provided flow key to do it.

    - If the userspace flow key includes more fields than the
      kernel's, for example if userspace decoded an IPv6 header but
      the kernel stopped at the Ethernet type, then userspace can
      forward the packet manually, without setting up a flow in the
      kernel.  This case is bad for performance because every packet
      that the kernel considers part of the flow must go to userspace,
      but the forwarding behavior is correct.  (If userspace can
      determine that the values of the extra fields would not affect
      forwarding behavior, then it could set up a flow anyway.)

How flow keys evolve over time is important to making this work, so
the following sections go into detail.


Flow key format
---------------

A flow key is passed over a Netlink socket as a sequence of Netlink
attributes.  Some attributes represent packet metadata, defined as any
information about a packet that cannot be extracted from the packet
itself, e.g. the vport on which the packet was received.  Most
attributes, however, are extracted from headers within the packet,
e.g. source and destination addresses from Ethernet, IP, or TCP
headers.

The <linux/openvswitch.h> header file defines the exact format of the
flow key attributes.  For informal explanatory purposes here, we write
them as comma-separated strings, with parentheses indicating arguments
and nesting.  For example, the following could represent a flow key
corresponding to a TCP packet that arrived on vport 1:

    in_port(1), eth(src=e0:91:f5:21:d0:b2, dst=00:02:e3:0f:80:a4),
    eth_type(0x0800), ipv4(src=172.16.0.20, dst=172.18.0.52, proto=17, tos=0,
    frag=no), tcp(src=49163, dst=80)

Often we ellipsize arguments not important to the discussion, e.g.:

    in_port(1), eth(...), eth_type(0x0800), ipv4(...), tcp(...)


Wildcarded flow key format
--------------------------

A wildcarded flow is described with two sequences of Netlink attributes
passed over the Netlink socket. A flow key, exactly as described above, and an
optional corresponding flow mask.

A wildcarded flow can represent a group of exact match flows. Each '1' bit
in the mask specifies a exact match with the corresponding bit in the flow key.
A '0' bit specifies a don't care bit, which will match either a '1' or '0' bit
of a incoming packet. Using wildcarded flow can improve the flow set up rate
by reduce the number of new flows need to be processed by the user space program.

Support for the mask Netlink attribute is optional for both the kernel and user
space program. The kernel can ignore the mask attribute, installing an exact
match flow, or reduce the number of don't care bits in the kernel to less than
what was specified by the user space program. In this case, variations in bits
that the kernel does not implement will simply result in additional flow setups.
The kernel module will also work with user space programs that neither support
nor supply flow mask attributes.

Since the kernel may ignore or modify wildcard bits, it can be difficult for
the userspace program to know exactly what matches are installed. There are
two possible approaches: reactively install flows as they miss the kernel
flow table (and therefore not attempt to determine wildcard changes at all)
or use the kernel's response messages to determine the installed wildcards.

When interacting with userspace, the kernel should maintain the match portion
of the key exactly as originally installed. This will provides a handle to
identify the flow for all future operations. However, when reporting the
mask of an installed flow, the mask should include any restrictions imposed
by the kernel.

The behavior when using overlapping wildcarded flows is undefined. It is the
responsibility of the user space program to ensure that any incoming packet
can match at most one flow, wildcarded or not. The current implementation
performs best-effort detection of overlapping wildcarded flows and may reject
some but not all of them. However, this behavior may change in future versions.


Basic rule for evolving flow keys
---------------------------------

Some care is needed to really maintain forward and backward
compatibility for applications that follow the rules listed under
"Flow key compatibility" above.

The basic rule is obvious:

    ------------------------------------------------------------------
    New network protocol support must only supplement existing flow
    key attributes.  It must not change the meaning of already defined
    flow key attributes.
    ------------------------------------------------------------------

This rule does have less-obvious consequences so it is worth working
through a few examples.  Suppose, for example, that the kernel module
did not already implement VLAN parsing.  Instead, it just interpreted
the 802.1Q TPID (0x8100) as the Ethertype then stopped parsing the
packet.  The flow key for any packet with an 802.1Q header would look
essentially like this, ignoring metadata:

    eth(...), eth_type(0x8100)

Naively, to add VLAN support, it makes sense to add a new "vlan" flow
key attribute to contain the VLAN tag, then continue to decode the
encapsulated headers beyond the VLAN tag using the existing field
definitions.  With this change, a TCP packet in VLAN 10 would have a
flow key much like this:

    eth(...), vlan(vid=10, pcp=0), eth_type(0x0800), ip(proto=6, ...), tcp(...)

But this change would negatively affect a userspace application that
has not been updated to understand the new "vlan" flow key attribute.
The application could, following the flow compatibility rules above,
ignore the "vlan" attribute that it does not understand and therefore
assume that the flow contained IP packets.  This is a bad assumption
(the flow only contains IP packets if one parses and skips over the
802.1Q header) and it could cause the application's behavior to change
across kernel versions even though it follows the compatibility rules.

The solution is to use a set of nested attributes.  This is, for
example, why 802.1Q support uses nested attributes.  A TCP packet in
VLAN 10 is actually expressed as:

    eth(...), eth_type(0x8100), vlan(vid=10, pcp=0), encap(eth_type(0x0800),
    ip(proto=6, ...), tcp(...)))

Notice how the "eth_type", "ip", and "tcp" flow key attributes are
nested inside the "encap" attribute.  Thus, an application that does
not understand the "vlan" key will not see either of those attributes
and therefore will not misinterpret them.  (Also, the outer eth_type
is still 0x8100, not changed to 0x0800.)

Handling malformed packets
--------------------------

Don't drop packets in the kernel for malformed protocol headers, bad
checksums, etc.  This would prevent userspace from implementing a
simple Ethernet switch that forwards every packet.

Instead, in such a case, include an attribute with "empty" content.
It doesn't matter if the empty content could be valid protocol values,
as long as those values are rarely seen in practice, because userspace
can always forward all packets with those values to userspace and
handle them individually.

For example, consider a packet that contains an IP header that
indicates protocol 6 for TCP, but which is truncated just after the IP
header, so that the TCP header is missing.  The flow key for this
packet would include a tcp attribute with all-zero src and dst, like
this:

    eth(...), eth_type(0x0800), ip(proto=6, ...), tcp(src=0, dst=0)

As another example, consider a packet with an Ethernet type of 0x8100,
indicating that a VLAN TCI should follow, but which is truncated just
after the Ethernet type.  The flow key for this packet would include
an all-zero-bits vlan and an empty encap attribute, like this:

    eth(...), eth_type(0x8100), vlan(0), encap()

Unlike a TCP packet with source and destination ports 0, an
all-zero-bits VLAN TCI is not that rare, so the CFI bit (aka
VLAN_TAG_PRESENT inside the kernel) is ordinarily set in a vlan
attribute expressly to allow this situation to be distinguished.
Thus, the flow key in this second example unambiguously indicates a
missing or malformed VLAN TCI.

Other rules
-----------

The other rules for flow keys are much less subtle:

    - Duplicate attributes are not allowed at a given nesting level.

    - Ordering of attributes is not significant.

    - When the kernel sends a given flow key to userspace, it always
      composes it the same way.  This allows userspace to hash and
      compare entire flow keys that it may not be able to fully
      interpret.