2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014, 2015, 2016 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "learning-switch.h"
22 #include <netinet/in.h>
26 #include "byte-order.h"
27 #include "classifier.h"
28 #include "dp-packet.h"
31 #include "mac-learning.h"
33 #include "ofp-actions.h"
34 #include "ofp-errors.h"
36 #include "ofp-parse.h"
37 #include "ofp-print.h"
39 #include "openflow/openflow.h"
40 #include "poll-loop.h"
45 #include "openvswitch/vconn.h"
46 #include "openvswitch/vlog.h"
48 VLOG_DEFINE_THIS_MODULE(learning_switch);
51 struct hmap_node hmap_node; /* Hash node for port number. */
52 ofp_port_t port_no; /* OpenFlow port number. */
53 uint32_t queue_id; /* OpenFlow queue number. */
57 S_CONNECTING, /* Waiting for connection to complete. */
58 S_FEATURES_REPLY, /* Waiting for features reply. */
59 S_SWITCHING, /* Switching flows. */
64 enum lswitch_state state;
66 /* If nonnegative, the switch sets up flows that expire after the given
67 * number of seconds (or never expire, if the value is OFP_FLOW_PERMANENT).
68 * Otherwise, the switch processes every packet. */
71 enum ofputil_protocol protocol;
72 unsigned long long int datapath_id;
73 struct mac_learning *ml; /* NULL to act as hub instead of switch. */
74 struct flow_wildcards wc; /* Wildcards to apply to flows. */
75 bool action_normal; /* Use OFPP_NORMAL? */
77 /* Queue distribution. */
78 uint32_t default_queue; /* Default OpenFlow queue, or UINT32_MAX. */
79 struct hmap queue_numbers; /* Map from port number to lswitch_port. */
80 struct shash queue_names; /* Map from port name to lswitch_port. */
82 /* Number of outgoing queued packets on the rconn. */
83 struct rconn_packet_counter *queued;
85 /* If true, do not reply to any messages from the switch (for debugging
89 /* Optional "flow mod" requests to send to the switch at connection time,
90 * to set up the flow table. */
91 const struct ofputil_flow_mod *default_flows;
92 size_t n_default_flows;
93 enum ofputil_protocol usable_protocols;
96 /* The log messages here could actually be useful in debugging, so keep the
97 * rate limit relatively high. */
98 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(30, 300);
100 static void queue_tx(struct lswitch *, struct ofpbuf *);
101 static void send_features_request(struct lswitch *);
103 static void lswitch_process_packet(struct lswitch *, const struct ofpbuf *);
104 static enum ofperr process_switch_features(struct lswitch *,
105 struct ofp_header *);
106 static void process_packet_in(struct lswitch *, const struct ofp_header *);
107 static void process_echo_request(struct lswitch *, const struct ofp_header *);
109 static ofp_port_t get_mac_entry_ofp_port(const struct mac_learning *ml,
110 const struct mac_entry *)
111 OVS_REQ_RDLOCK(ml->rwlock);
112 static void set_mac_entry_ofp_port(struct mac_learning *ml,
113 struct mac_entry *, ofp_port_t)
114 OVS_REQ_WRLOCK(ml->rwlock);
116 /* Creates and returns a new learning switch whose configuration is given by
119 * 'rconn' is used to send out an OpenFlow features request. */
121 lswitch_create(struct rconn *rconn, const struct lswitch_config *cfg)
126 sw = xzalloc(sizeof *sw);
128 sw->state = S_CONNECTING;
129 sw->max_idle = cfg->max_idle;
131 sw->ml = (cfg->mode == LSW_LEARN
132 ? mac_learning_create(MAC_ENTRY_DEFAULT_IDLE_TIME)
134 sw->action_normal = cfg->mode == LSW_NORMAL;
136 switch (cfg->wildcards) {
142 /* Try to wildcard as many fields as possible, but we cannot
143 * wildcard all fields. We need in_port to detect moves. We need
144 * Ethernet source and dest and VLAN VID to do L2 learning. */
145 ofpfw = (OFPFW10_DL_TYPE | OFPFW10_DL_VLAN_PCP
146 | OFPFW10_NW_SRC_ALL | OFPFW10_NW_DST_ALL
147 | OFPFW10_NW_TOS | OFPFW10_NW_PROTO
148 | OFPFW10_TP_SRC | OFPFW10_TP_DST);
152 ofpfw = cfg->wildcards;
155 ofputil_wildcard_from_ofpfw10(ofpfw, &sw->wc);
157 sw->default_queue = cfg->default_queue;
158 hmap_init(&sw->queue_numbers);
159 shash_init(&sw->queue_names);
160 if (cfg->port_queues) {
161 struct simap_node *node;
163 SIMAP_FOR_EACH (node, cfg->port_queues) {
164 struct lswitch_port *port = xmalloc(sizeof *port);
165 hmap_node_nullify(&port->hmap_node);
166 port->queue_id = node->data;
167 shash_add(&sw->queue_names, node->name, port);
171 sw->default_flows = cfg->default_flows;
172 sw->n_default_flows = cfg->n_default_flows;
173 sw->usable_protocols = cfg->usable_protocols;
175 sw->queued = rconn_packet_counter_create();
181 lswitch_handshake(struct lswitch *sw)
183 enum ofputil_protocol protocol;
184 enum ofp_version version;
186 send_features_request(sw);
188 version = rconn_get_version(sw->rconn);
189 protocol = ofputil_protocol_from_ofp_version(version);
190 if (version >= OFP13_VERSION) {
191 /* OpenFlow 1.3 and later by default drop packets that miss in the flow
192 * table. Set up a flow to send packets to the controller by
194 struct ofpact_output output;
198 ofpact_init_OUTPUT(&output);
199 output.port = OFPP_CONTROLLER;
200 output.max_len = OFP_DEFAULT_MISS_SEND_LEN;
202 struct ofputil_flow_mod fm = {
203 .match = MATCH_CATCHALL_INITIALIZER,
206 .command = OFPFC_ADD,
207 .buffer_id = UINT32_MAX,
208 .out_port = OFPP_NONE,
209 .out_group = OFPG_ANY,
210 .ofpacts = &output.ofpact,
211 .ofpacts_len = sizeof output,
214 msg = ofputil_encode_flow_mod(&fm, protocol);
215 error = rconn_send(sw->rconn, msg, NULL);
217 VLOG_INFO_RL(&rl, "%s: failed to add default flow (%s)",
218 rconn_get_name(sw->rconn), ovs_strerror(error));
221 if (sw->default_flows) {
222 struct ofpbuf *msg = NULL;
226 /* If the initial protocol isn't good enough for default_flows, then
227 * pick one that will work and encode messages to set up that
230 * This could be improved by actually negotiating a mutually acceptable
231 * flow format with the switch, but that would require an asynchronous
232 * state machine. This version ought to work fine in practice. */
233 if (!(protocol & sw->usable_protocols)) {
234 enum ofputil_protocol want = rightmost_1bit(sw->usable_protocols);
236 msg = ofputil_encode_set_protocol(protocol, want, &protocol);
240 error = rconn_send(sw->rconn, msg, NULL);
243 if (protocol & sw->usable_protocols) {
244 for (i = 0; !error && i < sw->n_default_flows; i++) {
245 msg = ofputil_encode_flow_mod(&sw->default_flows[i], protocol);
246 error = rconn_send(sw->rconn, msg, NULL);
250 VLOG_INFO_RL(&rl, "%s: failed to queue default flows (%s)",
251 rconn_get_name(sw->rconn), ovs_strerror(error));
254 VLOG_INFO_RL(&rl, "%s: failed to set usable protocol",
255 rconn_get_name(sw->rconn));
258 sw->protocol = protocol;
262 lswitch_is_alive(const struct lswitch *sw)
264 return rconn_is_alive(sw->rconn);
269 lswitch_destroy(struct lswitch *sw)
272 struct lswitch_port *node, *next;
274 rconn_destroy(sw->rconn);
275 HMAP_FOR_EACH_SAFE (node, next, hmap_node, &sw->queue_numbers) {
276 hmap_remove(&sw->queue_numbers, &node->hmap_node);
279 shash_destroy(&sw->queue_names);
280 mac_learning_unref(sw->ml);
281 rconn_packet_counter_destroy(sw->queued);
286 /* Takes care of necessary 'sw' activity, except for receiving packets (which
287 * the caller must do). */
289 lswitch_run(struct lswitch *sw)
294 ovs_rwlock_wrlock(&sw->ml->rwlock);
295 mac_learning_run(sw->ml);
296 ovs_rwlock_unlock(&sw->ml->rwlock);
299 rconn_run(sw->rconn);
301 if (sw->state == S_CONNECTING) {
302 if (rconn_get_version(sw->rconn) != -1) {
303 lswitch_handshake(sw);
304 sw->state = S_FEATURES_REPLY;
309 for (i = 0; i < 50; i++) {
312 msg = rconn_recv(sw->rconn);
318 lswitch_process_packet(sw, msg);
325 lswitch_wait(struct lswitch *sw)
328 ovs_rwlock_rdlock(&sw->ml->rwlock);
329 mac_learning_wait(sw->ml);
330 ovs_rwlock_unlock(&sw->ml->rwlock);
332 rconn_run_wait(sw->rconn);
333 rconn_recv_wait(sw->rconn);
336 /* Processes 'msg', which should be an OpenFlow received on 'rconn', according
337 * to the learning switch state in 'sw'. The most likely result of processing
338 * is that flow-setup and packet-out OpenFlow messages will be sent out on
341 lswitch_process_packet(struct lswitch *sw, const struct ofpbuf *msg)
347 if (ofptype_pull(&type, &b)) {
351 if (sw->state == S_FEATURES_REPLY
352 && type != OFPTYPE_ECHO_REQUEST
353 && type != OFPTYPE_FEATURES_REPLY) {
358 case OFPTYPE_ECHO_REQUEST:
359 process_echo_request(sw, msg->data);
362 case OFPTYPE_FEATURES_REPLY:
363 if (sw->state == S_FEATURES_REPLY) {
364 if (!process_switch_features(sw, msg->data)) {
365 sw->state = S_SWITCHING;
367 rconn_disconnect(sw->rconn);
372 case OFPTYPE_PACKET_IN:
373 process_packet_in(sw, msg->data);
376 case OFPTYPE_FLOW_REMOVED:
382 case OFPTYPE_ECHO_REPLY:
383 case OFPTYPE_FEATURES_REQUEST:
384 case OFPTYPE_GET_CONFIG_REQUEST:
385 case OFPTYPE_GET_CONFIG_REPLY:
386 case OFPTYPE_SET_CONFIG:
387 case OFPTYPE_PORT_STATUS:
388 case OFPTYPE_PACKET_OUT:
389 case OFPTYPE_FLOW_MOD:
390 case OFPTYPE_GROUP_MOD:
391 case OFPTYPE_PORT_MOD:
392 case OFPTYPE_TABLE_MOD:
393 case OFPTYPE_BARRIER_REQUEST:
394 case OFPTYPE_BARRIER_REPLY:
395 case OFPTYPE_QUEUE_GET_CONFIG_REQUEST:
396 case OFPTYPE_QUEUE_GET_CONFIG_REPLY:
397 case OFPTYPE_DESC_STATS_REQUEST:
398 case OFPTYPE_DESC_STATS_REPLY:
399 case OFPTYPE_FLOW_STATS_REQUEST:
400 case OFPTYPE_FLOW_STATS_REPLY:
401 case OFPTYPE_AGGREGATE_STATS_REQUEST:
402 case OFPTYPE_AGGREGATE_STATS_REPLY:
403 case OFPTYPE_TABLE_STATS_REQUEST:
404 case OFPTYPE_TABLE_STATS_REPLY:
405 case OFPTYPE_PORT_STATS_REQUEST:
406 case OFPTYPE_PORT_STATS_REPLY:
407 case OFPTYPE_QUEUE_STATS_REQUEST:
408 case OFPTYPE_QUEUE_STATS_REPLY:
409 case OFPTYPE_PORT_DESC_STATS_REQUEST:
410 case OFPTYPE_PORT_DESC_STATS_REPLY:
411 case OFPTYPE_ROLE_REQUEST:
412 case OFPTYPE_ROLE_REPLY:
413 case OFPTYPE_ROLE_STATUS:
414 case OFPTYPE_REQUESTFORWARD:
415 case OFPTYPE_SET_FLOW_FORMAT:
416 case OFPTYPE_FLOW_MOD_TABLE_ID:
417 case OFPTYPE_SET_PACKET_IN_FORMAT:
418 case OFPTYPE_FLOW_AGE:
419 case OFPTYPE_SET_CONTROLLER_ID:
420 case OFPTYPE_FLOW_MONITOR_STATS_REQUEST:
421 case OFPTYPE_FLOW_MONITOR_STATS_REPLY:
422 case OFPTYPE_FLOW_MONITOR_CANCEL:
423 case OFPTYPE_FLOW_MONITOR_PAUSED:
424 case OFPTYPE_FLOW_MONITOR_RESUMED:
425 case OFPTYPE_GET_ASYNC_REQUEST:
426 case OFPTYPE_GET_ASYNC_REPLY:
427 case OFPTYPE_SET_ASYNC_CONFIG:
428 case OFPTYPE_METER_MOD:
429 case OFPTYPE_GROUP_STATS_REQUEST:
430 case OFPTYPE_GROUP_STATS_REPLY:
431 case OFPTYPE_GROUP_DESC_STATS_REQUEST:
432 case OFPTYPE_GROUP_DESC_STATS_REPLY:
433 case OFPTYPE_GROUP_FEATURES_STATS_REQUEST:
434 case OFPTYPE_GROUP_FEATURES_STATS_REPLY:
435 case OFPTYPE_METER_STATS_REQUEST:
436 case OFPTYPE_METER_STATS_REPLY:
437 case OFPTYPE_METER_CONFIG_STATS_REQUEST:
438 case OFPTYPE_METER_CONFIG_STATS_REPLY:
439 case OFPTYPE_METER_FEATURES_STATS_REQUEST:
440 case OFPTYPE_METER_FEATURES_STATS_REPLY:
441 case OFPTYPE_TABLE_FEATURES_STATS_REQUEST:
442 case OFPTYPE_TABLE_FEATURES_STATS_REPLY:
443 case OFPTYPE_TABLE_DESC_REQUEST:
444 case OFPTYPE_TABLE_DESC_REPLY:
445 case OFPTYPE_BUNDLE_CONTROL:
446 case OFPTYPE_BUNDLE_ADD_MESSAGE:
447 case OFPTYPE_NXT_TLV_TABLE_MOD:
448 case OFPTYPE_NXT_TLV_TABLE_REQUEST:
449 case OFPTYPE_NXT_TLV_TABLE_REPLY:
451 if (VLOG_IS_DBG_ENABLED()) {
452 char *s = ofp_to_string(msg->data, msg->size, 2);
453 VLOG_DBG_RL(&rl, "%016llx: OpenFlow packet ignored: %s",
461 send_features_request(struct lswitch *sw)
464 struct ofp_switch_config *osc;
465 int ofp_version = rconn_get_version(sw->rconn);
467 ovs_assert(ofp_version > 0 && ofp_version < 0xff);
469 /* Send OFPT_FEATURES_REQUEST. */
470 b = ofpraw_alloc(OFPRAW_OFPT_FEATURES_REQUEST, ofp_version, 0);
473 /* Send OFPT_SET_CONFIG. */
474 b = ofpraw_alloc(OFPRAW_OFPT_SET_CONFIG, ofp_version, sizeof *osc);
475 osc = ofpbuf_put_zeros(b, sizeof *osc);
476 osc->miss_send_len = htons(OFP_DEFAULT_MISS_SEND_LEN);
481 queue_tx(struct lswitch *sw, struct ofpbuf *b)
483 int retval = rconn_send_with_limit(sw->rconn, b, sw->queued, 10);
484 if (retval && retval != ENOTCONN) {
485 if (retval == EAGAIN) {
486 VLOG_INFO_RL(&rl, "%016llx: %s: tx queue overflow",
487 sw->datapath_id, rconn_get_name(sw->rconn));
489 VLOG_WARN_RL(&rl, "%016llx: %s: send: %s",
490 sw->datapath_id, rconn_get_name(sw->rconn),
491 ovs_strerror(retval));
497 process_switch_features(struct lswitch *sw, struct ofp_header *oh)
499 struct ofputil_switch_features features;
500 struct ofputil_phy_port port;
504 error = ofputil_decode_switch_features(oh, &features, &b);
506 VLOG_ERR("received invalid switch feature reply (%s)",
507 ofperr_to_string(error));
511 sw->datapath_id = features.datapath_id;
513 while (!ofputil_pull_phy_port(oh->version, &b, &port)) {
514 struct lswitch_port *lp = shash_find_data(&sw->queue_names, port.name);
515 if (lp && hmap_node_is_null(&lp->hmap_node)) {
516 lp->port_no = port.port_no;
517 hmap_insert(&sw->queue_numbers, &lp->hmap_node,
518 hash_ofp_port(lp->port_no));
525 lswitch_choose_destination(struct lswitch *sw, const struct flow *flow)
529 /* Learn the source MAC. */
531 ovs_rwlock_wrlock(&sw->ml->rwlock);
532 if (mac_learning_may_learn(sw->ml, flow->dl_src, 0)) {
533 struct mac_entry *mac = mac_learning_insert(sw->ml, flow->dl_src,
535 if (get_mac_entry_ofp_port(sw->ml, mac)
536 != flow->in_port.ofp_port) {
537 VLOG_DBG_RL(&rl, "%016llx: learned that "ETH_ADDR_FMT" is on "
538 "port %"PRIu16, sw->datapath_id,
539 ETH_ADDR_ARGS(flow->dl_src),
540 flow->in_port.ofp_port);
542 set_mac_entry_ofp_port(sw->ml, mac, flow->in_port.ofp_port);
545 ovs_rwlock_unlock(&sw->ml->rwlock);
548 /* Drop frames for reserved multicast addresses. */
549 if (eth_addr_is_reserved(flow->dl_dst)) {
553 out_port = OFPP_FLOOD;
555 struct mac_entry *mac;
557 ovs_rwlock_rdlock(&sw->ml->rwlock);
558 mac = mac_learning_lookup(sw->ml, flow->dl_dst, 0);
560 out_port = get_mac_entry_ofp_port(sw->ml, mac);
561 if (out_port == flow->in_port.ofp_port) {
562 /* Don't send a packet back out its input port. */
563 ovs_rwlock_unlock(&sw->ml->rwlock);
567 ovs_rwlock_unlock(&sw->ml->rwlock);
570 /* Check if we need to use "NORMAL" action. */
571 if (sw->action_normal && out_port != OFPP_FLOOD) {
579 get_queue_id(const struct lswitch *sw, ofp_port_t in_port)
581 const struct lswitch_port *port;
583 HMAP_FOR_EACH_WITH_HASH (port, hmap_node, hash_ofp_port(in_port),
584 &sw->queue_numbers) {
585 if (port->port_no == in_port) {
586 return port->queue_id;
590 return sw->default_queue;
594 process_packet_in(struct lswitch *sw, const struct ofp_header *oh)
596 struct ofputil_packet_in pi;
600 uint64_t ofpacts_stub[64 / 8];
601 struct ofpbuf ofpacts;
603 struct ofputil_packet_out po;
606 struct dp_packet pkt;
609 error = ofputil_decode_packet_in(&pi, oh);
611 VLOG_WARN_RL(&rl, "failed to decode packet-in: %s",
612 ofperr_to_string(error));
616 /* Ignore packets sent via output to OFPP_CONTROLLER. This library never
617 * uses such an action. You never know what experiments might be going on,
618 * though, and it seems best not to interfere with them. */
619 if (pi.reason != OFPR_NO_MATCH) {
623 /* Extract flow data from 'opi' into 'flow'. */
624 dp_packet_use_const(&pkt, pi.packet, pi.packet_len);
625 flow_extract(&pkt, &flow);
626 flow.in_port.ofp_port = pi.flow_metadata.flow.in_port.ofp_port;
627 flow.tunnel.tun_id = pi.flow_metadata.flow.tunnel.tun_id;
629 /* Choose output port. */
630 out_port = lswitch_choose_destination(sw, &flow);
633 queue_id = get_queue_id(sw, pi.flow_metadata.flow.in_port.ofp_port);
634 ofpbuf_use_stack(&ofpacts, ofpacts_stub, sizeof ofpacts_stub);
635 if (out_port == OFPP_NONE) {
637 } else if (queue_id == UINT32_MAX
638 || ofp_to_u16(out_port) >= ofp_to_u16(OFPP_MAX)) {
639 ofpact_put_OUTPUT(&ofpacts)->port = out_port;
641 struct ofpact_enqueue *enqueue = ofpact_put_ENQUEUE(&ofpacts);
642 enqueue->port = out_port;
643 enqueue->queue = queue_id;
646 /* Prepare packet_out in case we need one. */
647 po.buffer_id = pi.buffer_id;
648 if (po.buffer_id == UINT32_MAX) {
649 po.packet = dp_packet_data(&pkt);
650 po.packet_len = dp_packet_size(&pkt);
655 po.in_port = pi.flow_metadata.flow.in_port.ofp_port;
656 po.ofpacts = ofpacts.data;
657 po.ofpacts_len = ofpacts.size;
659 /* Send the packet, and possibly the whole flow, to the output port. */
660 if (sw->max_idle >= 0 && (!sw->ml || out_port != OFPP_FLOOD)) {
661 /* The output port is known, or we always flood everything, so add a
663 struct ofputil_flow_mod fm = {
664 .priority = 1, /* Must be > 0 because of table-miss flow entry. */
666 .command = OFPFC_ADD,
667 .idle_timeout = sw->max_idle,
668 .buffer_id = pi.buffer_id,
669 .out_port = OFPP_NONE,
670 .ofpacts = ofpacts.data,
671 .ofpacts_len = ofpacts.size,
673 match_init(&fm.match, &flow, &sw->wc);
674 ofputil_normalize_match_quiet(&fm.match);
676 struct ofpbuf *buffer = ofputil_encode_flow_mod(&fm, sw->protocol);
678 queue_tx(sw, buffer);
680 /* If the switch didn't buffer the packet, we need to send a copy. */
681 if (pi.buffer_id == UINT32_MAX && out_port != OFPP_NONE) {
682 queue_tx(sw, ofputil_encode_packet_out(&po, sw->protocol));
685 /* We don't know that MAC, or we don't set up flows. Send along the
686 * packet without setting up a flow. */
687 if (pi.buffer_id != UINT32_MAX || out_port != OFPP_NONE) {
688 queue_tx(sw, ofputil_encode_packet_out(&po, sw->protocol));
694 process_echo_request(struct lswitch *sw, const struct ofp_header *rq)
696 queue_tx(sw, make_echo_reply(rq));
700 get_mac_entry_ofp_port(const struct mac_learning *ml,
701 const struct mac_entry *e)
702 OVS_REQ_RDLOCK(ml->rwlock)
704 void *port = mac_entry_get_port(ml, e);
705 return (OVS_FORCE ofp_port_t) (uintptr_t) port;
709 set_mac_entry_ofp_port(struct mac_learning *ml,
710 struct mac_entry *e, ofp_port_t ofp_port)
711 OVS_REQ_WRLOCK(ml->rwlock)
713 mac_entry_set_port(ml, e, (void *) (OVS_FORCE uintptr_t) ofp_port);