| blob | b3b9ac61d29d8751baff13446f800a2c9bc9e9fb |
1 Open vSwitch datapath developer documentation
2 =============================================
4 The Open vSwitch kernel module allows flexible userspace control over
5 flow-level packet processing on selected network devices. It can be
6 used to implement a plain Ethernet switch, network device bonding,
7 VLAN processing, network access control, flow-based network control,
8 and so on.
10 The kernel module implements multiple "datapaths" (analogous to
11 bridges), each of which can have multiple "vports" (analogous to ports
12 within a bridge). Each datapath also has associated with it a "flow
13 table" that userspace populates with "flows" that map from keys based
14 on packet headers and metadata to sets of actions. The most common
15 action forwards the packet to another vport; other actions are also
16 implemented.
18 When a packet arrives on a vport, the kernel module processes it by
19 extracting its flow key and looking it up in the flow table. If there
20 is a matching flow, it executes the associated actions. If there is
21 no match, it queues the packet to userspace for processing (as part of
22 its processing, userspace will likely set up a flow to handle further
23 packets of the same type entirely in-kernel).
26 Flow key compatibility
27 ----------------------
29 Network protocols evolve over time. New protocols become important
30 and existing protocols lose their prominence. For the Open vSwitch
31 kernel module to remain relevant, it must be possible for newer
32 versions to parse additional protocols as part of the flow key. It
33 might even be desirable, someday, to drop support for parsing
34 protocols that have become obsolete. Therefore, the Netlink interface
35 to Open vSwitch is designed to allow carefully written userspace
36 applications to work with any version of the flow key, past or future.
38 To support this forward and backward compatibility, whenever the
39 kernel module passes a packet to userspace, it also passes along the
40 flow key that it parsed from the packet. Userspace then extracts its
41 own notion of a flow key from the packet and compares it against the
42 kernel-provided version:
44 - If userspace's notion of the flow key for the packet matches the
45 kernel's, then nothing special is necessary.
47 - If the kernel's flow key includes more fields than the userspace
48 version of the flow key, for example if the kernel decoded IPv6
49 headers but userspace stopped at the Ethernet type (because it
50 does not understand IPv6), then again nothing special is
51 necessary. Userspace can still set up a flow in the usual way,
52 as long as it uses the kernel-provided flow key to do it.
54 - If the userspace flow key includes more fields than the
55 kernel's, for example if userspace decoded an IPv6 header but
56 the kernel stopped at the Ethernet type, then userspace can
57 forward the packet manually, without setting up a flow in the
58 kernel. This case is bad for performance because every packet
59 that the kernel considers part of the flow must go to userspace,
60 but the forwarding behavior is correct. (If userspace can
61 determine that the values of the extra fields would not affect
62 forwarding behavior, then it could set up a flow anyway.)
64 How flow keys evolve over time is important to making this work, so
65 the following sections go into detail.
68 Flow key format
69 ---------------
71 A flow key is passed over a Netlink socket as a sequence of Netlink
72 attributes. Some attributes represent packet metadata, defined as any
73 information about a packet that cannot be extracted from the packet
74 itself, e.g. the vport on which the packet was received. Most
75 attributes, however, are extracted from headers within the packet,
76 e.g. source and destination addresses from Ethernet, IP, or TCP
77 headers.
79 The <linux/openvswitch.h> header file defines the exact format of the
80 flow key attributes. For informal explanatory purposes here, we write
81 them as comma-separated strings, with parentheses indicating arguments
82 and nesting. For example, the following could represent a flow key
83 corresponding to a TCP packet that arrived on vport 1:
85 in_port(1), eth(src=e0:91:f5:21:d0:b2, dst=00:02:e3:0f:80:a4),
86 eth_type(0x0800), ipv4(src=172.16.0.20, dst=172.18.0.52, proto=17, tos=0,
87 frag=no), tcp(src=49163, dst=80)
89 Often we ellipsize arguments not important to the discussion, e.g.:
91 in_port(1), eth(...), eth_type(0x0800), ipv4(...), tcp(...)
94 Wildcarded flow key format
95 --------------------------
97 A wildcarded flow is described with two sequences of Netlink attributes
98 passed over the Netlink socket. A flow key, exactly as described above, and an
99 optional corresponding flow mask.
101 A wildcarded flow can represent a group of exact match flows. Each '1' bit
102 in the mask specifies a exact match with the corresponding bit in the flow key.
103 A '0' bit specifies a don't care bit, which will match either a '1' or '0' bit
104 of a incoming packet. Using wildcarded flow can improve the flow set up rate
105 by reduce the number of new flows need to be processed by the user space program.
107 Support for the mask Netlink attribute is optional for both the kernel and user
108 space program. The kernel can ignore the mask attribute, installing an exact
109 match flow, or reduce the number of don't care bits in the kernel to less than
110 what was specified by the user space program. In this case, variations in bits
111 that the kernel does not implement will simply result in additional flow setups.
112 The kernel module will also work with user space programs that neither support
113 nor supply flow mask attributes.
115 Since the kernel may ignore or modify wildcard bits, it can be difficult for
116 the userspace program to know exactly what matches are installed. There are
117 two possible approaches: reactively install flows as they miss the kernel
118 flow table (and therefore not attempt to determine wildcard changes at all)
119 or use the kernel's response messages to determine the installed wildcards.
121 When interacting with userspace, the kernel should maintain the match portion
122 of the key exactly as originally installed. This will provides a handle to
123 identify the flow for all future operations. However, when reporting the
124 mask of an installed flow, the mask should include any restrictions imposed
125 by the kernel.
127 The behavior when using overlapping wildcarded flows is undefined. It is the
128 responsibility of the user space program to ensure that any incoming packet
129 can match at most one flow, wildcarded or not. The current implementation
130 performs best-effort detection of overlapping wildcarded flows and may reject
131 some but not all of them. However, this behavior may change in future versions.
134 Unique flow identifiers
135 -----------------------
137 An alternative to using the original match portion of a key as the handle for
138 flow identification is a unique flow identifier, or "UFID". UFIDs are optional
139 for both the kernel and user space program.
141 User space programs that support UFID are expected to provide it during flow
142 setup in addition to the flow, then refer to the flow using the UFID for all
143 future operations. The kernel is not required to index flows by the original
144 flow key if a UFID is specified.
147 Basic rule for evolving flow keys
148 ---------------------------------
150 Some care is needed to really maintain forward and backward
151 compatibility for applications that follow the rules listed under
152 "Flow key compatibility" above.
154 The basic rule is obvious:
156 ------------------------------------------------------------------
157 New network protocol support must only supplement existing flow
158 key attributes. It must not change the meaning of already defined
159 flow key attributes.
160 ------------------------------------------------------------------
162 This rule does have less-obvious consequences so it is worth working
163 through a few examples. Suppose, for example, that the kernel module
164 did not already implement VLAN parsing. Instead, it just interpreted
165 the 802.1Q TPID (0x8100) as the Ethertype then stopped parsing the
166 packet. The flow key for any packet with an 802.1Q header would look
167 essentially like this, ignoring metadata:
169 eth(...), eth_type(0x8100)
171 Naively, to add VLAN support, it makes sense to add a new "vlan" flow
172 key attribute to contain the VLAN tag, then continue to decode the
173 encapsulated headers beyond the VLAN tag using the existing field
174 definitions. With this change, a TCP packet in VLAN 10 would have a
175 flow key much like this:
177 eth(...), vlan(vid=10, pcp=0), eth_type(0x0800), ip(proto=6, ...), tcp(...)
179 But this change would negatively affect a userspace application that
180 has not been updated to understand the new "vlan" flow key attribute.
181 The application could, following the flow compatibility rules above,
182 ignore the "vlan" attribute that it does not understand and therefore
183 assume that the flow contained IP packets. This is a bad assumption
184 (the flow only contains IP packets if one parses and skips over the
185 802.1Q header) and it could cause the application's behavior to change
186 across kernel versions even though it follows the compatibility rules.
188 The solution is to use a set of nested attributes. This is, for
189 example, why 802.1Q support uses nested attributes. A TCP packet in
190 VLAN 10 is actually expressed as:
192 eth(...), eth_type(0x8100), vlan(vid=10, pcp=0), encap(eth_type(0x0800),
193 ip(proto=6, ...), tcp(...)))
195 Notice how the "eth_type", "ip", and "tcp" flow key attributes are
196 nested inside the "encap" attribute. Thus, an application that does
197 not understand the "vlan" key will not see either of those attributes
198 and therefore will not misinterpret them. (Also, the outer eth_type
199 is still 0x8100, not changed to 0x0800.)
201 Handling malformed packets
202 --------------------------
204 Don't drop packets in the kernel for malformed protocol headers, bad
205 checksums, etc. This would prevent userspace from implementing a
206 simple Ethernet switch that forwards every packet.
208 Instead, in such a case, include an attribute with "empty" content.
209 It doesn't matter if the empty content could be valid protocol values,
210 as long as those values are rarely seen in practice, because userspace
211 can always forward all packets with those values to userspace and
212 handle them individually.
214 For example, consider a packet that contains an IP header that
215 indicates protocol 6 for TCP, but which is truncated just after the IP
216 header, so that the TCP header is missing. The flow key for this
217 packet would include a tcp attribute with all-zero src and dst, like
218 this:
220 eth(...), eth_type(0x0800), ip(proto=6, ...), tcp(src=0, dst=0)
222 As another example, consider a packet with an Ethernet type of 0x8100,
223 indicating that a VLAN TCI should follow, but which is truncated just
224 after the Ethernet type. The flow key for this packet would include
225 an all-zero-bits vlan and an empty encap attribute, like this:
227 eth(...), eth_type(0x8100), vlan(0), encap()
229 Unlike a TCP packet with source and destination ports 0, an
230 all-zero-bits VLAN TCI is not that rare, so the CFI bit (aka
231 VLAN_TAG_PRESENT inside the kernel) is ordinarily set in a vlan
232 attribute expressly to allow this situation to be distinguished.
233 Thus, the flow key in this second example unambiguously indicates a
234 missing or malformed VLAN TCI.
236 Other rules
237 -----------
239 The other rules for flow keys are much less subtle:
241 - Duplicate attributes are not allowed at a given nesting level.
243 - Ordering of attributes is not significant.
245 - When the kernel sends a given flow key to userspace, it always
246 composes it the same way. This allows userspace to hash and
247 compare entire flow keys that it may not be able to fully
248 interpret.