1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
|
OpenFlow 1.1+ support in Open vSwitch
=====================================
Open vSwitch support for OpenFlow 1.1, 1.2, and 1.3 is a work in
progress. This file describes the work still to be done.
The Plan
--------
OpenFlow version support is not a build-time option. A single build
of Open vSwitch must be able to handle all supported versions of
OpenFlow. Ideally, even at runtime it should be able to support all
protocol versions at the same time on different OpenFlow bridges (and
perhaps even on the same bridge).
At the same time, it would be a shame to litter the core of the OVS
code with lots of ugly code concerned with the details of various
OpenFlow protocol versions.
The primary approach to compatibility is to abstract most of the
details of the differences from the core code, by adding a protocol
layer that translates between OF1.x and a slightly higher-level
abstract representation. The core of this approach is the many struct
ofputil_* structures in lib/ofp-util.h.
As a consequence of this approach, OVS cannot use OpenFlow protocol
definitions that closely resemble those in the OpenFlow specification,
because openflow.h in different versions of the OpenFlow specification
defines the same identifier with different values. Instead,
openflow-common.h contains definitions that are common to all the
specifications and separate protocol version-specific headers contain
protocol-specific definitions renamed so as not to conflict,
e.g. OFPAT10_ENQUEUE and OFPAT11_ENQUEUE for the OpenFlow 1.0 and 1.1
values for OFPAT_ENQUEUE. Generally, in cases of conflict, the
protocol layer will define a more abstract OFPUTIL_* or struct
ofputil_*.
Here are the current approaches in a few tricky areas:
* Port numbering. OpenFlow 1.0 has 16-bit port numbers and later
OpenFlow versions have 32-bit port numbers. For now, OVS
support for later protocol versions requires all port numbers to
fall into the 16-bit range, translating the reserved OFPP_* port
numbers.
* Actions. OpenFlow 1.0 and later versions have very different
ideas of actions. OVS reconciles by translating all the
versions' actions (and instructions) to and from a common
internal representation.
OpenFlow 1.1
------------
The list of remaining work items for OpenFlow 1.1 is below. It is
probably incomplete.
* Implement Write-Actions instruction.
* The new in_phy_port field in OFPT_PACKET_IN needs some kind of
implementation. It has a sensible interpretation for tunnels
but in general the physical port is not in the datapath for OVS
so the value is not necessarily meaningful. We might have to
just fix it as the same as in_port.
* On OF1.1+ flow_mods, updates by MODIFY are now much better
specified. Check that OVS implements the new behavior, fix it
if not.
* On OF1.1+ flow_mods, DELETE now ignores buffer_id.
* OFPST_PORT and OFPST_QUEUE stats. These differ little from
OF1.0 to OF1.1 (just the size of the port number field) but do
require abstraction (Done?).
* OFPT_TABLE_MOD stats. This is new in OF1.1, so we need to
implement it. It should be implemented so that the default OVS
behavior does not change.
* Document how OVS does packet buffering.
* MPLS. Simon Horman maintains a patch series that adds this
feature. It needs review and possible revision before it is
merged.
* SCTP. Joe Stringer maintains a patch series that adds this
feature. It has received review comments that need to be
addressed before it is merged.
* Match and set double-tagged VLANs (QinQ). This requires kernel
work for reasonable performance.
* VLANs tagged with 88a8 Ethertype. This requires kernel work for
reasonable performance.
* Groups.
OpenFlow 1.2
------------
OpenFlow 1.2 support requires OpenFlow 1.1 as a prerequisite, plus the
following additional work. (This is based on the change log at the
end of the OF1.2 spec. I didn’t compare the specs carefully yet.)
* Use new OpenFlow extensible error infrastructure, on OF1.2+
only, instead of the OVS-specific extension used until now.
* OFPT_FLOW_MOD:
* MODIFY and MODIFY_STRICT commands now never insert new flows
in the table. We will still need variations that do,
though, both to support older OpenFlow protocols and to get
sensible behavior for the internal implementation of the
NXAST_LEARN action.
* New flag OFPFF_RESET_COUNTS.
* New cookie field behavior.
* Add ability to delete flow in all tables.
* Update DESIGN to describe OF1.2 behavior also.
* Implement OFPT_ROLE_REQUEST. Patch submitted by Jarno
Rajahalme, currently under revision.
* Add ability to turn off packet buffering with OFPCML_NO_BUFFER.
OpenFlow 1.3
------------
OpenFlow 1.3 support requires OpenFlow 1.2 as a prerequisite, plus the
following additional work. (This is based on the change log at the
end of the OF1.3 spec, reusing most of the section titles directly. I
didn’t compare the specs carefully yet.)
* Add support for multipart requests.
* Add OFPMP_TABLE_FEATURES statistics.
* More flexible table miss support.
* IPv6 extension header handling support. Fully implementing this
requires kernel support. This likely will take some careful and
probably time-consuming design work. The actual coding, once
that is all done, is probably 2 or 3 days work.
* Per-flow meters. Similar to IPv6 extension headers in kernel
and design requirements. Might be politically difficult to add
directly to the kernel module, since its functionality overlaps
with tc. Ideally, therefore, we could implement these somehow
with tc, but I haven’t investigated whether that makes sense.
* Per-connection event filtering. OF1.3 adopted Open vSwitch’s
existing design for this feature so implementation should be
easy.
* Auxiliary connections. These are optional, so a minimal
implementation would not need them. An implementation in
generic code might be a week’s worth of work. The value of an
implementation in generic code is questionable, though, since
much of the benefit of axuiliary connections is supposed to be
to take advantage of hardware support. (We could make the
kernel module somehow send packets across the auxiliary
connections directly, for some kind of “hardware” support, if we
judged it useful enough.)
* MPLS BoS matching. (Included in Simon's MPLS series?)
* Provider Backbone Bridge tagging. I don’t plan to implement
this (but we’d accept an implementation).
* Rework tag order. I’m not sure whether we need to do anything
for this.
* Duration for stats.
* On-demand flow counters. I think this might be a real
optimization in some cases for the software switch.
How to contribute
-----------------
If you plan to contribute code for a feature, please let everyone know
on ovs-dev before you start work. This will help avoid duplicating
work.
Please consider the following:
* Testing. Please test your code.
* Unit tests. Please consider writing some. The tests directory
has many examples that you can use as a starting point.
* ovs-ofctl. If you add a feature that is useful for some
ovs-ofctl command then you should add support for it there.
* Documentation. If you add a user-visible feature, then you
should document it in the appropriate manpage and mention it in
NEWS as well.
* Coding style (see the CodingStyle file at the top of the source
tree).
* The patch submission guidelines (see SubmittingPatches). I
recommend using “git send-email”, which automatically follows a
lot of those guidelines.
Bug Reporting
-------------
Please report problems to bugs@openvswitch.org.
|
