1 Design Decisions In Open vSwitch
2 ================================
4 This document describes design decisions that went into implementing
5 Open vSwitch. While we believe these to be reasonable decisions, it is
6 impossible to predict how Open vSwitch will be used in all environments.
7 Understanding assumptions made by Open vSwitch is critical to a
8 successful deployment. The end of this document contains contact
9 information that can be used to let us know how we can make Open vSwitch
10 more generally useful.
15 Over time, Open vSwitch has added many knobs that control whether a
16 given controller receives OpenFlow asynchronous messages. This
17 section describes how all of these features interact.
19 First, a service controller never receives any asynchronous messages
20 unless it changes its miss_send_len from the service controller
21 default of zero in one of the following ways:
23 - Sending an OFPT_SET_CONFIG message with nonzero miss_send_len.
25 - Sending any NXT_SET_ASYNC_CONFIG message: as a side effect, this
26 message changes the miss_send_len to
27 OFP_DEFAULT_MISS_SEND_LEN (128) for service controllers.
29 Second, OFPT_FLOW_REMOVED and NXT_FLOW_REMOVED messages are generated
30 only if the flow that was removed had the OFPFF_SEND_FLOW_REM flag
33 Third, OFPT_PACKET_IN and NXT_PACKET_IN messages are sent only to
34 OpenFlow controller connections that have the correct connection ID
35 (see "struct nx_controller_id" and "struct nx_action_controller"):
37 - For packet-in messages generated by a NXAST_CONTROLLER action,
38 the controller ID specified in the action.
40 - For other packet-in messages, controller ID zero. (This is the
41 default ID when an OpenFlow controller does not configure one.)
43 Finally, Open vSwitch consults a per-connection table indexed by the
44 message type, reason code, and current role. The following table
45 shows how this table is initialized by default when an OpenFlow
46 connection is made. An entry labeled "yes" means that the message is
47 sent, an entry labeled "---" means that the message is suppressed.
51 message and reason code other slave
52 ---------------------------------------- ------- -----
53 OFPT_PACKET_IN / NXT_PACKET_IN
56 OFPR_INVALID_TTL --- ---
58 OFPT_FLOW_REMOVED / NXT_FLOW_REMOVED
59 OFPRR_IDLE_TIMEOUT yes ---
60 OFPRR_HARD_TIMEOUT yes ---
69 The NXT_SET_ASYNC_CONFIG message directly sets all of the values in
70 this table for the current connection. The
71 OFPC_INVALID_TTL_TO_CONTROLLER bit in the OFPT_SET_CONFIG message
72 controls the setting for OFPR_INVALID_TTL for the "master" role.
78 The OpenFlow 1.0 specification requires the output port of the OFPAT_ENQUEUE
79 action to "refer to a valid physical port (i.e. < OFPP_MAX) or OFPP_IN_PORT".
80 Although OFPP_LOCAL is not less than OFPP_MAX, it is an 'internal' port which
81 can have QoS applied to it in Linux. Since we allow the OFPAT_ENQUEUE to apply
82 to 'internal' ports whose port numbers are less than OFPP_MAX, we interpret
83 OFPP_LOCAL as a physical port and support OFPAT_ENQUEUE on it as well.
89 The OpenFlow specification for the behavior of OFPT_FLOW_MOD is
90 confusing. The following tables summarize the Open vSwitch
91 implementation of its behavior in the following categories:
93 - "match on priority": Whether the flow_mod acts only on flows
94 whose priority matches that included in the flow_mod message.
96 - "match on out_port": Whether the flow_mod acts only on flows
97 that output to the out_port included in the flow_mod message (if
98 out_port is not OFPP_NONE). OpenFlow 1.1 and later have a
99 similar feature (not listed separately here) for out_group.
101 - "match on flow_cookie": Whether the flow_mod acts only on flows
102 whose flow_cookie matches an optional controller-specified value
105 - "updates flow_cookie": Whether the flow_mod changes the
106 flow_cookie of the flow or flows that it matches to the
107 flow_cookie included in the flow_mod message.
109 - "updates OFPFF_ flags": Whether the flow_mod changes the
110 OFPFF_SEND_FLOW_REM flag of the flow or flows that it matches to
111 the setting included in the flags of the flow_mod message.
113 - "honors OFPFF_CHECK_OVERLAP": Whether the OFPFF_CHECK_OVERLAP
114 flag in the flow_mod is significant.
116 - "updates idle_timeout" and "updates hard_timeout": Whether the
117 idle_timeout and hard_timeout in the flow_mod, respectively,
118 have an effect on the flow or flows matched by the flow_mod.
120 - "updates idle timer": Whether the flow_mod resets the per-flow
121 timer that measures how long a flow has been idle.
123 - "updates hard timer": Whether the flow_mod resets the per-flow
124 timer that measures how long it has been since a flow was
127 - "zeros counters": Whether the flow_mod resets per-flow packet
128 and byte counters to zero.
130 - "may add a new flow": Whether the flow_mod may add a new flow to
131 the flow table. (Obviously this is always true for "add"
132 commands but in some OpenFlow versions "modify" and
133 "modify-strict" can also add new flows.)
135 - "sends flow_removed message": Whether the flow_mod generates a
136 flow_removed message for the flow or flows that it affects.
138 An entry labeled "yes" means that the flow mod type does have the
139 indicated behavior, "---" means that it does not, an empty cell means
140 that the property is not applicable, and other values are explained
148 ADD MODIFY STRICT DELETE STRICT
149 === ====== ====== ====== ======
150 match on priority yes --- yes --- yes
151 match on out_port --- --- --- yes yes
152 match on flow_cookie --- --- --- --- ---
153 match on table_id --- --- --- --- ---
154 controller chooses table_id --- --- ---
155 updates flow_cookie yes yes yes
156 updates OFPFF_SEND_FLOW_REM yes + +
157 honors OFPFF_CHECK_OVERLAP yes + +
158 updates idle_timeout yes + +
159 updates hard_timeout yes + +
160 resets idle timer yes + +
161 resets hard timer yes yes yes
162 zeros counters yes + +
163 may add a new flow yes yes yes
164 sends flow_removed message --- --- --- % %
166 (+) "modify" and "modify-strict" only take these actions when they
167 create a new flow, not when they update an existing flow.
169 (%) "delete" and "delete_strict" generates a flow_removed message if
170 the deleted flow or flows have the OFPFF_SEND_FLOW_REM flag set.
171 (Each controller can separately control whether it wants to
172 receive the generated messages.)
178 OpenFlow 1.1 makes these changes:
180 - The controller now must specify the table_id of the flow match
181 searched and into which a flow may be inserted. Behavior for a
182 table_id of 255 is undefined.
184 - A flow_mod, except an "add", can now match on the flow_cookie.
186 - When a flow_mod matches on the flow_cookie, "modify" and
187 "modify-strict" never insert a new flow.
191 ADD MODIFY STRICT DELETE STRICT
192 === ====== ====== ====== ======
193 match on priority yes --- yes --- yes
194 match on out_port --- --- --- yes yes
195 match on flow_cookie --- yes yes yes yes
196 match on table_id yes yes yes yes yes
197 controller chooses table_id yes yes yes
198 updates flow_cookie yes --- ---
199 updates OFPFF_SEND_FLOW_REM yes + +
200 honors OFPFF_CHECK_OVERLAP yes + +
201 updates idle_timeout yes + +
202 updates hard_timeout yes + +
203 resets idle timer yes + +
204 resets hard timer yes yes yes
205 zeros counters yes + +
206 may add a new flow yes # #
207 sends flow_removed message --- --- --- % %
209 (+) "modify" and "modify-strict" only take these actions when they
210 create a new flow, not when they update an existing flow.
212 (%) "delete" and "delete_strict" generates a flow_removed message if
213 the deleted flow or flows have the OFPFF_SEND_FLOW_REM flag set.
214 (Each controller can separately control whether it wants to
215 receive the generated messages.)
217 (#) "modify" and "modify-strict" only add a new flow if the flow_mod
218 does not match on any bits of the flow cookie
224 OpenFlow 1.2 makes these changes:
226 - Only "add" commands ever add flows, "modify" and "modify-strict"
229 - A new flag OFPFF_RESET_COUNTS now controls whether "modify" and
230 "modify-strict" reset counters, whereas previously they never
231 reset counters (except when they inserted a new flow).
235 ADD MODIFY STRICT DELETE STRICT
236 === ====== ====== ====== ======
237 match on priority yes --- yes --- yes
238 match on out_port --- --- --- yes yes
239 match on flow_cookie --- yes yes yes yes
240 match on table_id yes yes yes yes yes
241 controller chooses table_id yes yes yes
242 updates flow_cookie yes --- ---
243 updates OFPFF_SEND_FLOW_REM yes --- ---
244 honors OFPFF_CHECK_OVERLAP yes --- ---
245 updates idle_timeout yes --- ---
246 updates hard_timeout yes --- ---
247 resets idle timer yes --- ---
248 resets hard timer yes yes yes
249 zeros counters yes & &
250 may add a new flow yes --- ---
251 sends flow_removed message --- --- --- % %
253 (%) "delete" and "delete_strict" generates a flow_removed message if
254 the deleted flow or flows have the OFPFF_SEND_FLOW_REM flag set.
255 (Each controller can separately control whether it wants to
256 receive the generated messages.)
258 (&) "modify" and "modify-strict" reset counters if the
259 OFPFF_RESET_COUNTS flag is specified.
265 OpenFlow 1.3 makes these changes:
267 - Behavior for a table_id of 255 is now defined, for "delete" and
268 "delete-strict" commands, as meaning to delete from all tables.
269 A table_id of 255 is now explicitly invalid for other commands.
271 - New flags OFPFF_NO_PKT_COUNTS and OFPFF_NO_BYT_COUNTS for "add"
274 The table for 1.3 is the same as the one shown above for 1.2.
280 OpenFlow 1.4 does not change flow_mod semantics.
286 The OpenFlow 1.1 specification for OFPT_PACKET_IN is confusing. The
287 definition in OF1.1 openflow.h is[*]:
290 /* Packet received on port (datapath -> controller). */
291 struct ofp_packet_in {
292 struct ofp_header header;
293 uint32_t buffer_id; /* ID assigned by datapath. */
294 uint32_t in_port; /* Port on which frame was received. */
295 uint32_t in_phy_port; /* Physical Port on which frame was received. */
296 uint16_t total_len; /* Full length of frame. */
297 uint8_t reason; /* Reason packet is being sent (one of OFPR_*) */
298 uint8_t table_id; /* ID of the table that was looked up */
299 uint8_t data[0]; /* Ethernet frame, halfway through 32-bit word,
300 so the IP header is 32-bit aligned. The
301 amount of data is inferred from the length
302 field in the header. Because of padding,
303 offsetof(struct ofp_packet_in, data) ==
304 sizeof(struct ofp_packet_in) - 2. */
306 OFP_ASSERT(sizeof(struct ofp_packet_in) == 24);
309 The confusing part is the comment on the data[] member. This comment
310 is a leftover from OF1.0 openflow.h, in which the comment was correct:
311 sizeof(struct ofp_packet_in) is 20 in OF1.0 and offsetof(struct
312 ofp_packet_in, data) is 18. When OF1.1 was written, the structure
313 members were changed but the comment was carelessly not updated, and
314 the comment became wrong: sizeof(struct ofp_packet_in) and
315 offsetof(struct ofp_packet_in, data) are both 24 in OF1.1.
317 That leaves the question of how to implement ofp_packet_in in OF1.1.
318 The OpenFlow reference implementation for OF1.1 does not include any
319 padding, that is, the first byte of the encapsulated frame immediately
320 follows the 'table_id' member without a gap. Open vSwitch therefore
321 implements it the same way for compatibility.
323 For an earlier discussion, please see the thread archived at:
324 https://mailman.stanford.edu/pipermail/openflow-discuss/2011-August/002604.html
326 [*] The quoted definition is directly from OF1.1. Definitions used
327 inside OVS omit the 8-byte ofp_header members, so the sizes in
328 this discussion are 8 bytes larger than those declared in OVS
335 The 802.1Q VLAN header causes more trouble than any other 4 bytes in
336 networking. More specifically, three versions of OpenFlow and Open
337 vSwitch have among them four different ways to match the contents and
338 presence of the VLAN header. The following table describes how each
341 Match NXM OF1.0 OF1.1 OF1.2
342 ----- --------- ----------- ----------- ------------
343 [1] 0000/0000 ????/1,??/? ????/1,??/? 0000/0000,--
344 [2] 0000/ffff ffff/0,??/? ffff/0,??/? 0000/ffff,--
345 [3] 1xxx/1fff 0xxx/0,??/1 0xxx/0,??/1 1xxx/ffff,--
346 [4] z000/f000 ????/1,0y/0 fffe/0,0y/0 1000/1000,0y
347 [5] zxxx/ffff 0xxx/0,0y/0 0xxx/0,0y/0 1xxx/ffff,0y
348 [6] 0000/0fff <none> <none> <none>
349 [7] 0000/f000 <none> <none> <none>
350 [8] 0000/efff <none> <none> <none>
351 [9] 1001/1001 <none> <none> 1001/1001,--
352 [10] 3000/3000 <none> <none> <none>
354 Each column is interpreted as follows.
356 - Match: See the list below.
358 - NXM: xxxx/yyyy means NXM_OF_VLAN_TCI_W with value xxxx and mask
359 yyyy. A mask of 0000 is equivalent to omitting
360 NXM_OF_VLAN_TCI(_W), a mask of ffff is equivalent to
363 - OF1.0 and OF1.1: wwww/x,yy/z means dl_vlan wwww, OFPFW_DL_VLAN
364 x, dl_vlan_pcp yy, and OFPFW_DL_VLAN_PCP z. ? means that the
365 given nibble is ignored (and conventionally 0 for wwww or yy,
366 conventionally 1 for x or z). <none> means that the given match
369 - OF1.2: xxxx/yyyy,zz means OXM_OF_VLAN_VID_W with value xxxx and
370 mask yyyy, and OXM_OF_VLAN_PCP (which is not maskable) with
371 value zz. A mask of 0000 is equivalent to omitting
372 OXM_OF_VLAN_VID(_W), a mask of ffff is equivalent to
373 OXM_OF_VLAN_VID. -- means that OXM_OF_VLAN_PCP is omitted.
374 <none> means that the given match is not supported.
378 [1] Matches any packet, that is, one without an 802.1Q header or with
379 an 802.1Q header with any TCI value.
381 [2] Matches only packets without an 802.1Q header.
383 NXM: Any match with (vlan_tci == 0) and (vlan_tci_mask & 0x1000)
384 != 0 is equivalent to the one listed in the table.
386 OF1.0: The spec doesn't define behavior if dl_vlan is set to
387 0xffff and OFPFW_DL_VLAN_PCP is not set.
389 OF1.1: The spec says explicitly to ignore dl_vlan_pcp when
390 dl_vlan is set to 0xffff.
392 OF1.2: The spec doesn't say what should happen if (vlan_vid == 0)
393 and (vlan_vid_mask & 0x1000) != 0 but (vlan_vid_mask != 0x1000),
394 but it would be straightforward to also interpret as [2].
396 [3] Matches only packets that have an 802.1Q header with VID xxx (and
399 [4] Matches only packets that have an 802.1Q header with PCP y (and
402 NXM: z is ((y << 1) | 1).
404 OF1.0: The spec isn't very clear, but OVS implements it this way.
406 OF1.2: Presumably other masks such that (vlan_vid_mask & 0x1fff)
407 == 0x1000 would also work, but the spec doesn't define their
410 [5] Matches only packets that have an 802.1Q header with VID xxx and
413 NXM: z is ((y << 1) | 1).
415 OF1.2: Presumably other masks such that (vlan_vid_mask & 0x1fff)
416 == 0x1fff would also work.
418 [6] Matches packets with no 802.1Q header or with an 802.1Q header
419 with a VID of 0. Only possible with NXM.
421 [7] Matches packets with no 802.1Q header or with an 802.1Q header
422 with a PCP of 0. Only possible with NXM.
424 [8] Matches packets with no 802.1Q header or with an 802.1Q header
425 with both VID and PCP of 0. Only possible with NXM.
427 [9] Matches only packets that have an 802.1Q header with an
428 odd-numbered VID (and any PCP). Only possible with NXM and
429 OF1.2. (This is just an example; one can match on any desired
432 [10] Matches only packets that have an 802.1Q header with an
433 odd-numbered PCP (and any VID). Only possible with NXM. (This
434 is just an example; one can match on any desired VID bit
439 - OF1.2: The top three bits of OXM_OF_VLAN_VID are fixed to zero,
440 so bits 13, 14, and 15 in the masks listed in the table may be
441 set to arbitrary values, as long as the corresponding value bits
442 are also zero. The suggested ffff mask for [2], [3], and [5]
443 allows a shorter OXM representation (the mask is omitted) than
444 the minimal 1fff mask.
450 OpenFlow 1.0 and later versions have the concept of a "flow cookie",
451 which is a 64-bit integer value attached to each flow. The treatment
452 of the flow cookie has varied greatly across OpenFlow versions,
457 - OFPFC_ADD set the cookie in the flow that it added.
459 - OFPFC_MODIFY and OFPFC_MODIFY_STRICT updated the cookie for
460 the flow or flows that it modified.
462 - OFPST_FLOW messages included the flow cookie.
464 - OFPT_FLOW_REMOVED messages reported the cookie of the flow
467 OpenFlow 1.1 made the following changes:
469 - Flow mod operations OFPFC_MODIFY, OFPFC_MODIFY_STRICT,
470 OFPFC_DELETE, and OFPFC_DELETE_STRICT, plus flow stats
471 requests and aggregate stats requests, gained the ability to
472 match on flow cookies with an arbitrary mask.
474 - OFPFC_MODIFY and OFPFC_MODIFY_STRICT were changed to add a
475 new flow, in the case of no match, only if the flow table
476 modification operation did not match on the cookie field.
477 (In OpenFlow 1.0, modify operations always added a new flow
478 when there was no match.)
480 - OFPFC_MODIFY and OFPFC_MODIFY_STRICT no longer updated flow
483 OpenFlow 1.2 made the following changes:
485 - OFPC_MODIFY and OFPFC_MODIFY_STRICT were changed to never
486 add a new flow, regardless of whether the flow cookie was
489 Open vSwitch support for OpenFlow 1.0 implements the OpenFlow 1.0
490 behavior with the following extensions:
492 - An NXM extension field NXM_NX_COOKIE(_W) allows the NXM
493 versions of OFPFC_MODIFY, OFPFC_MODIFY_STRICT, OFPFC_DELETE,
494 and OFPFC_DELETE_STRICT flow_mods, plus flow stats requests
495 and aggregate stats requests, to match on flow cookies with
496 arbitrary masks. This is much like the equivalent OpenFlow
499 - Like OpenFlow 1.1, OFPC_MODIFY and OFPFC_MODIFY_STRICT add a
500 new flow if there is no match and the mask is zero (or not
503 - The "cookie" field in OFPT_FLOW_MOD and NXT_FLOW_MOD messages
504 is used as the cookie value for OFPFC_ADD commands, as
505 described in OpenFlow 1.0. For OFPFC_MODIFY and
506 OFPFC_MODIFY_STRICT commands, the "cookie" field is used as a
507 new cookie for flows that match unless it is UINT64_MAX, in
508 which case the flow's cookie is not updated.
510 - NXT_PACKET_IN (the Nicira extended version of
511 OFPT_PACKET_IN) reports the cookie of the rule that
512 generated the packet, or all-1-bits if no rule generated the
513 packet. (Older versions of OVS used all-0-bits instead of
516 The following table shows the handling of different protocols when
517 receiving OFPFC_MODIFY and OFPFC_MODIFY_STRICT messages. A mask of 0
518 indicates either an explicit mask of zero or an implicit one by not
519 specifying the NXM_NX_COOKIE(_W) field.
522 Match Update Add on miss Add on miss
523 cookie cookie mask!=0 mask==0
524 ====== ====== =========== ===========
525 OpenFlow 1.0 no yes <always add on miss>
526 OpenFlow 1.1 yes no no yes
527 OpenFlow 1.2 yes no no no
530 * Updates the flow's cookie unless the "cookie" field is UINT64_MAX.
533 Multiple Table Support
534 ======================
536 OpenFlow 1.0 has only rudimentary support for multiple flow tables.
537 Notably, OpenFlow 1.0 does not allow the controller to specify the
538 flow table to which a flow is to be added. Open vSwitch adds an
539 extension for this purpose, which is enabled on a per-OpenFlow
540 connection basis using the NXT_FLOW_MOD_TABLE_ID message. When the
541 extension is enabled, the upper 8 bits of the 'command' member in an
542 OFPT_FLOW_MOD or NXT_FLOW_MOD message designates the table to which a
545 The Open vSwitch software switch implementation offers 255 flow
546 tables. On packet ingress, only the first flow table (table 0) is
547 searched, and the contents of the remaining tables are not considered
548 in any way. Tables other than table 0 only come into play when an
549 NXAST_RESUBMIT_TABLE action specifies another table to search.
551 Tables 128 and above are reserved for use by the switch itself.
552 Controllers should use only tables 0 through 127.
558 Open vSwitch supports stateless handling of IPv6 packets. Flows can be
559 written to support matching TCP, UDP, and ICMPv6 headers within an IPv6
560 packet. Deeper matching of some Neighbor Discovery messages is also
563 IPv6 was not designed to interact well with middle-boxes. This,
564 combined with Open vSwitch's stateless nature, have affected the
565 processing of IPv6 traffic, which is detailed below.
570 The base IPv6 header is incredibly simple with the intention of only
571 containing information relevant for routing packets between two
572 endpoints. IPv6 relies heavily on the use of extension headers to
573 provide any other functionality. Unfortunately, the extension headers
574 were designed in such a way that it is impossible to move to the next
575 header (including the layer-4 payload) unless the current header is
578 Open vSwitch will process the following extension headers and continue
581 * Fragment (see the next section)
582 * AH (Authentication Header)
585 * Destination Options
587 When a header is encountered that is not in that list, it is considered
588 "terminal". A terminal header's IPv6 protocol value is stored in
589 "nw_proto" for matching purposes. If a terminal header is TCP, UDP, or
590 ICMPv6, the packet will be further processed in an attempt to extract
596 IPv6 requires that every link in the internet have an MTU of 1280 octets
597 or greater (RFC 2460). As such, a terminal header (as described above in
598 "Extension Headers") in the first fragment should generally be
599 reachable. In this case, the terminal header's IPv6 protocol type is
600 stored in the "nw_proto" field for matching purposes. If a terminal
601 header cannot be found in the first fragment (one with a fragment offset
602 of zero), the "nw_proto" field is set to 0. Subsequent fragments (those
603 with a non-zero fragment offset) have the "nw_proto" field set to the
604 IPv6 protocol type for fragments (44).
609 An IPv6 jumbogram (RFC 2675) is a packet containing a payload longer
610 than 65,535 octets. A jumbogram is only relevant in subnets with a link
611 MTU greater than 65,575 octets, and are not required to be supported on
612 nodes that do not connect to link with such large MTUs. Currently, Open
613 vSwitch doesn't process jumbograms.
622 An OpenFlow switch must establish and maintain a TCP network
623 connection to its controller. There are two basic ways to categorize
624 the network that this connection traverses: either it is completely
625 separate from the one that the switch is otherwise controlling, or its
626 path may overlap the network that the switch controls. We call the
627 former case "out-of-band control", the latter case "in-band control".
629 Out-of-band control has the following benefits:
631 - Simplicity: Out-of-band control slightly simplifies the switch
634 - Reliability: Excessive switch traffic volume cannot interfere
635 with control traffic.
637 - Integrity: Machines not on the control network cannot
638 impersonate a switch or a controller.
640 - Confidentiality: Machines not on the control network cannot
641 snoop on control traffic.
643 In-band control, on the other hand, has the following advantages:
645 - No dedicated port: There is no need to dedicate a physical
646 switch port to control, which is important on switches that have
647 few ports (e.g. wireless routers, low-end embedded platforms).
649 - No dedicated network: There is no need to build and maintain a
650 separate control network. This is important in many
651 environments because it reduces proliferation of switches and
654 Open vSwitch supports both out-of-band and in-band control. This
655 section describes the principles behind in-band control. See the
656 description of the Controller table in ovs-vswitchd.conf.db(5) to
657 configure OVS for in-band control.
662 The fundamental principle of in-band control is that an OpenFlow
663 switch must recognize and switch control traffic without involving the
664 OpenFlow controller. All the details of implementing in-band control
665 are special cases of this principle.
667 The rationale for this principle is simple. If the switch does not
668 handle in-band control traffic itself, then it will be caught in a
669 contradiction: it must contact the controller, but it cannot, because
670 only the controller can set up the flows that are needed to contact
673 The following points describe important special cases of this
676 - In-band control must be implemented regardless of whether the
679 It is tempting to implement the in-band control rules only when
680 the switch is not connected to the controller, using the
681 reasoning that the controller should have complete control once
682 it has established a connection with the switch.
684 This does not work in practice. Consider the case where the
685 switch is connected to the controller. Occasionally it can
686 happen that the controller forgets or otherwise needs to obtain
687 the MAC address of the switch. To do so, the controller sends a
688 broadcast ARP request. A switch that implements the in-band
689 control rules only when it is disconnected will then send an
690 OFPT_PACKET_IN message up to the controller. The controller will
691 be unable to respond, because it does not know the MAC address of
692 the switch. This is a deadlock situation that can only be
693 resolved by the switch noticing that its connection to the
694 controller has hung and reconnecting.
696 - In-band control must override flows set up by the controller.
698 It is reasonable to assume that flows set up by the OpenFlow
699 controller should take precedence over in-band control, on the
700 basis that the controller should be in charge of the switch.
702 Again, this does not work in practice. Reasonable controller
703 implementations may set up a "last resort" fallback rule that
704 wildcards every field and, e.g., sends it up to the controller or
705 discards it. If a controller does that, then it will isolate
706 itself from the switch.
708 - The switch must recognize all control traffic.
710 The fundamental principle of in-band control states, in part,
711 that a switch must recognize control traffic without involving
712 the OpenFlow controller. More specifically, the switch must
713 recognize *all* control traffic. "False negatives", that is,
714 packets that constitute control traffic but that the switch does
715 not recognize as control traffic, lead to control traffic storms.
717 Consider an OpenFlow switch that only recognizes control packets
718 sent to or from that switch. Now suppose that two switches of
719 this type, named A and B, are connected to ports on an Ethernet
720 hub (not a switch) and that an OpenFlow controller is connected
721 to a third hub port. In this setup, control traffic sent by
722 switch A will be seen by switch B, which will send it to the
723 controller as part of an OFPT_PACKET_IN message. Switch A will
724 then see the OFPT_PACKET_IN message's packet, re-encapsulate it
725 in another OFPT_PACKET_IN, and send it to the controller. Switch
726 B will then see that OFPT_PACKET_IN, and so on in an infinite
729 Incidentally, the consequences of "false positives", where
730 packets that are not control traffic are nevertheless recognized
731 as control traffic, are much less severe. The controller will
732 not be able to control their behavior, but the network will
733 remain in working order. False positives do constitute a
736 - The switch should use echo-requests to detect disconnection.
738 TCP will notice that a connection has hung, but this can take a
739 considerable amount of time. For example, with default settings
740 the Linux kernel TCP implementation will retransmit for between
741 13 and 30 minutes, depending on the connection's retransmission
742 timeout, according to kernel documentation. This is far too long
743 for a switch to be disconnected, so an OpenFlow switch should
744 implement its own connection timeout. OpenFlow OFPT_ECHO_REQUEST
745 messages are the best way to do this, since they test the
746 OpenFlow connection itself.
751 This section describes how Open vSwitch implements in-band control.
752 Correctly implementing in-band control has proven difficult due to its
753 many subtleties, and has thus gone through many iterations. Please
754 read through and understand the reasoning behind the chosen rules
755 before making modifications.
757 Open vSwitch implements in-band control as "hidden" flows, that is,
758 flows that are not visible through OpenFlow, and at a higher priority
759 than wildcarded flows can be set up through OpenFlow. This is done so
760 that the OpenFlow controller cannot interfere with them and possibly
761 break connectivity with its switches. It is possible to see all
762 flows, including in-band ones, with the ovs-appctl "bridge/dump-flows"
765 The Open vSwitch implementation of in-band control can hide traffic to
766 arbitrary "remotes", where each remote is one TCP port on one IP address.
767 Currently the remotes are automatically configured as the in-band OpenFlow
768 controllers plus the OVSDB managers, if any. (The latter is a requirement
769 because OVSDB managers are responsible for configuring OpenFlow controllers,
770 so if the manager cannot be reached then OpenFlow cannot be reconfigured.)
772 The following rules (with the OFPP_NORMAL action) are set up on any bridge
773 that has any remotes:
775 (a) DHCP requests sent from the local port.
776 (b) ARP replies to the local port's MAC address.
777 (c) ARP requests from the local port's MAC address.
779 In-band also sets up the following rules for each unique next-hop MAC
780 address for the remotes' IPs (the "next hop" is either the remote
781 itself, if it is on a local subnet, or the gateway to reach the remote):
783 (d) ARP replies to the next hop's MAC address.
784 (e) ARP requests from the next hop's MAC address.
786 In-band also sets up the following rules for each unique remote IP address:
788 (f) ARP replies containing the remote's IP address as a target.
789 (g) ARP requests containing the remote's IP address as a source.
791 In-band also sets up the following rules for each unique remote (IP,port)
794 (h) TCP traffic to the remote's IP and port.
795 (i) TCP traffic from the remote's IP and port.
797 The goal of these rules is to be as narrow as possible to allow a
798 switch to join a network and be able to communicate with the
799 remotes. As mentioned earlier, these rules have higher priority
800 than the controller's rules, so if they are too broad, they may
801 prevent the controller from implementing its policy. As such,
802 in-band actively monitors some aspects of flow and packet processing
803 so that the rules can be made more precise.
805 In-band control monitors attempts to add flows into the datapath that
806 could interfere with its duties. The datapath only allows exact
807 match entries, so in-band control is able to be very precise about
808 the flows it prevents. Flows that miss in the datapath are sent to
809 userspace to be processed, so preventing these flows from being
810 cached in the "fast path" does not affect correctness. The only type
811 of flow that is currently prevented is one that would prevent DHCP
812 replies from being seen by the local port. For example, a rule that
813 forwarded all DHCP traffic to the controller would not be allowed,
814 but one that forwarded to all ports (including the local port) would.
816 As mentioned earlier, packets that miss in the datapath are sent to
817 the userspace for processing. The userspace has its own flow table,
818 the "classifier", so in-band checks whether any special processing
819 is needed before the classifier is consulted. If a packet is a DHCP
820 response to a request from the local port, the packet is forwarded to
821 the local port, regardless of the flow table. Note that this requires
822 L7 processing of DHCP replies to determine whether the 'chaddr' field
823 matches the MAC address of the local port.
825 It is interesting to note that for an L3-based in-band control
826 mechanism, the majority of rules are devoted to ARP traffic. At first
827 glance, some of these rules appear redundant. However, each serves an
828 important role. First, in order to determine the MAC address of the
829 remote side (controller or gateway) for other ARP rules, we must allow
830 ARP traffic for our local port with rules (b) and (c). If we are
831 between a switch and its connection to the remote, we have to
832 allow the other switch's ARP traffic to through. This is done with
833 rules (d) and (e), since we do not know the addresses of the other
834 switches a priori, but do know the remote's or gateway's. Finally,
835 if the remote is running in a local guest VM that is not reached
836 through the local port, the switch that is connected to the VM must
837 allow ARP traffic based on the remote's IP address, since it will
838 not know the MAC address of the local port that is sending the traffic
839 or the MAC address of the remote in the guest VM.
841 With a few notable exceptions below, in-band should work in most
842 network setups. The following are considered "supported' in the
843 current implementation:
845 - Locally Connected. The switch and remote are on the same
846 subnet. This uses rules (a), (b), (c), (h), and (i).
848 - Reached through Gateway. The switch and remote are on
849 different subnets and must go through a gateway. This uses
850 rules (a), (b), (c), (h), and (i).
852 - Between Switch and Remote. This switch is between another
853 switch and the remote, and we want to allow the other
854 switch's traffic through. This uses rules (d), (e), (h), and
855 (i). It uses (b) and (c) indirectly in order to know the MAC
856 address for rules (d) and (e). Note that DHCP for the other
857 switch will not work unless an OpenFlow controller explicitly lets this
858 switch pass the traffic.
860 - Between Switch and Gateway. This switch is between another
861 switch and the gateway, and we want to allow the other switch's
862 traffic through. This uses the same rules and logic as the
863 "Between Switch and Remote" configuration described earlier.
865 - Remote on Local VM. The remote is a guest VM on the
866 system running in-band control. This uses rules (a), (b), (c),
869 - Remote on Local VM with Different Networks. The remote
870 is a guest VM on the system running in-band control, but the
871 local port is not used to connect to the remote. For
872 example, an IP address is configured on eth0 of the switch. The
873 remote's VM is connected through eth1 of the switch, but an
874 IP address has not been configured for that port on the switch.
875 As such, the switch will use eth0 to connect to the remote,
876 and eth1's rules about the local port will not work. In the
877 example, the switch attached to eth0 would use rules (a), (b),
878 (c), (h), and (i) on eth0. The switch attached to eth1 would use
879 rules (f), (g), (h), and (i).
881 The following are explicitly *not* supported by in-band control:
883 - Specify Remote by Name. Currently, the remote must be
884 identified by IP address. A naive approach would be to permit
885 all DNS traffic. Unfortunately, this would prevent the
886 controller from defining any policy over DNS. Since switches
887 that are located behind us need to connect to the remote,
888 in-band cannot simply add a rule that allows DNS traffic from
889 the local port. The "correct" way to support this is to parse
890 DNS requests to allow all traffic related to a request for the
891 remote's name through. Due to the potential security
892 problems and amount of processing, we decided to hold off for
895 - Differing Remotes for Switches. All switches must know
896 the L3 addresses for all the remotes that other switches
897 may use, since rules need to be set up to allow traffic related
898 to those remotes through. See rules (f), (g), (h), and (i).
900 - Differing Routes for Switches. In order for the switch to
901 allow other switches to connect to a remote through a
902 gateway, it allows the gateway's traffic through with rules (d)
903 and (e). If the routes to the remote differ for the two
904 switches, we will not know the MAC address of the alternate
911 It seems likely that many controllers, at least at startup, use the
912 OpenFlow "flow statistics" request to obtain existing flows, then
913 compare the flows' actions against the actions that they expect to
914 find. Before version 1.8.0, Open vSwitch always returned exact,
915 byte-for-byte copies of the actions that had been added to the flow
916 table. The current version of Open vSwitch does not always do this in
917 some exceptional cases. This section lists the exceptions that
918 controller authors must keep in mind if they compare actual actions
919 against desired actions in a bytewise fashion:
921 - Open vSwitch zeros padding bytes in action structures,
922 regardless of their values when the flows were added.
924 - Open vSwitch "normalizes" the instructions in OpenFlow 1.1
925 (and later) in the following way:
927 * OVS sorts the instructions into the following order:
928 Apply-Actions, Clear-Actions, Write-Actions,
929 Write-Metadata, Goto-Table.
931 * OVS drops Apply-Actions instructions that have empty
934 * OVS drops Write-Actions instructions that have empty
937 Please report other discrepancies, if you notice any, so that we can
938 fix or document them.
944 Suggestions to improve Open vSwitch are welcome at discuss@openvswitch.org.
projects / ovs.git / blob