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) {
357 if (type == OFPTYPE_ECHO_REQUEST) {
358 process_echo_request(sw, msg->data);
359 } else if (type == OFPTYPE_FEATURES_REPLY) {
360 if (sw->state == S_FEATURES_REPLY) {
361 if (!process_switch_features(sw, msg->data)) {
362 sw->state = S_SWITCHING;
364 rconn_disconnect(sw->rconn);
367 } else if (type == OFPTYPE_PACKET_IN) {
368 process_packet_in(sw, msg->data);
369 } else if (type == OFPTYPE_FLOW_REMOVED) {
371 } else if (VLOG_IS_DBG_ENABLED()) {
372 char *s = ofp_to_string(msg->data, msg->size, 2);
373 VLOG_DBG_RL(&rl, "%016llx: OpenFlow packet ignored: %s",
380 send_features_request(struct lswitch *sw)
383 int ofp_version = rconn_get_version(sw->rconn);
385 ovs_assert(ofp_version > 0 && ofp_version < 0xff);
387 /* Send OFPT_FEATURES_REQUEST. */
388 b = ofpraw_alloc(OFPRAW_OFPT_FEATURES_REQUEST, ofp_version, 0);
391 /* Send OFPT_SET_CONFIG. */
392 struct ofputil_switch_config config = {
393 .miss_send_len = OFP_DEFAULT_MISS_SEND_LEN
395 queue_tx(sw, ofputil_encode_set_config(&config, ofp_version));
399 queue_tx(struct lswitch *sw, struct ofpbuf *b)
401 int retval = rconn_send_with_limit(sw->rconn, b, sw->queued, 10);
402 if (retval && retval != ENOTCONN) {
403 if (retval == EAGAIN) {
404 VLOG_INFO_RL(&rl, "%016llx: %s: tx queue overflow",
405 sw->datapath_id, rconn_get_name(sw->rconn));
407 VLOG_WARN_RL(&rl, "%016llx: %s: send: %s",
408 sw->datapath_id, rconn_get_name(sw->rconn),
409 ovs_strerror(retval));
415 process_switch_features(struct lswitch *sw, struct ofp_header *oh)
417 struct ofputil_switch_features features;
418 struct ofputil_phy_port port;
420 struct ofpbuf b = ofpbuf_const_initializer(oh, ntohs(oh->length));
421 enum ofperr error = ofputil_pull_switch_features(&b, &features);
423 VLOG_ERR("received invalid switch feature reply (%s)",
424 ofperr_to_string(error));
428 sw->datapath_id = features.datapath_id;
430 while (!ofputil_pull_phy_port(oh->version, &b, &port)) {
431 struct lswitch_port *lp = shash_find_data(&sw->queue_names, port.name);
432 if (lp && hmap_node_is_null(&lp->hmap_node)) {
433 lp->port_no = port.port_no;
434 hmap_insert(&sw->queue_numbers, &lp->hmap_node,
435 hash_ofp_port(lp->port_no));
442 lswitch_choose_destination(struct lswitch *sw, const struct flow *flow)
446 /* Learn the source MAC. */
448 ovs_rwlock_wrlock(&sw->ml->rwlock);
449 if (mac_learning_may_learn(sw->ml, flow->dl_src, 0)) {
450 struct mac_entry *mac = mac_learning_insert(sw->ml, flow->dl_src,
452 if (get_mac_entry_ofp_port(sw->ml, mac)
453 != flow->in_port.ofp_port) {
454 VLOG_DBG_RL(&rl, "%016llx: learned that "ETH_ADDR_FMT" is on "
455 "port %"PRIu16, sw->datapath_id,
456 ETH_ADDR_ARGS(flow->dl_src),
457 flow->in_port.ofp_port);
459 set_mac_entry_ofp_port(sw->ml, mac, flow->in_port.ofp_port);
462 ovs_rwlock_unlock(&sw->ml->rwlock);
465 /* Drop frames for reserved multicast addresses. */
466 if (eth_addr_is_reserved(flow->dl_dst)) {
470 out_port = OFPP_FLOOD;
472 struct mac_entry *mac;
474 ovs_rwlock_rdlock(&sw->ml->rwlock);
475 mac = mac_learning_lookup(sw->ml, flow->dl_dst, 0);
477 out_port = get_mac_entry_ofp_port(sw->ml, mac);
478 if (out_port == flow->in_port.ofp_port) {
479 /* Don't send a packet back out its input port. */
480 ovs_rwlock_unlock(&sw->ml->rwlock);
484 ovs_rwlock_unlock(&sw->ml->rwlock);
487 /* Check if we need to use "NORMAL" action. */
488 if (sw->action_normal && out_port != OFPP_FLOOD) {
496 get_queue_id(const struct lswitch *sw, ofp_port_t in_port)
498 const struct lswitch_port *port;
500 HMAP_FOR_EACH_WITH_HASH (port, hmap_node, hash_ofp_port(in_port),
501 &sw->queue_numbers) {
502 if (port->port_no == in_port) {
503 return port->queue_id;
507 return sw->default_queue;
511 process_packet_in(struct lswitch *sw, const struct ofp_header *oh)
513 struct ofputil_packet_in pi;
518 uint64_t ofpacts_stub[64 / 8];
519 struct ofpbuf ofpacts;
521 struct ofputil_packet_out po;
524 struct dp_packet pkt;
527 error = ofputil_decode_packet_in(oh, true, &pi, NULL, &buffer_id, NULL);
529 VLOG_WARN_RL(&rl, "failed to decode packet-in: %s",
530 ofperr_to_string(error));
534 /* Ignore packets sent via output to OFPP_CONTROLLER. This library never
535 * uses such an action. You never know what experiments might be going on,
536 * though, and it seems best not to interfere with them. */
537 if (pi.reason != OFPR_NO_MATCH) {
541 /* Extract flow data from 'pi' into 'flow'. */
542 dp_packet_use_const(&pkt, pi.packet, pi.packet_len);
543 flow_extract(&pkt, &flow);
544 flow.in_port.ofp_port = pi.flow_metadata.flow.in_port.ofp_port;
545 flow.tunnel.tun_id = pi.flow_metadata.flow.tunnel.tun_id;
547 /* Choose output port. */
548 out_port = lswitch_choose_destination(sw, &flow);
551 queue_id = get_queue_id(sw, pi.flow_metadata.flow.in_port.ofp_port);
552 ofpbuf_use_stack(&ofpacts, ofpacts_stub, sizeof ofpacts_stub);
553 if (out_port == OFPP_NONE) {
555 } else if (queue_id == UINT32_MAX
556 || ofp_to_u16(out_port) >= ofp_to_u16(OFPP_MAX)) {
557 ofpact_put_OUTPUT(&ofpacts)->port = out_port;
559 struct ofpact_enqueue *enqueue = ofpact_put_ENQUEUE(&ofpacts);
560 enqueue->port = out_port;
561 enqueue->queue = queue_id;
564 /* Prepare packet_out in case we need one. */
565 po.buffer_id = buffer_id;
566 if (buffer_id == UINT32_MAX) {
567 po.packet = dp_packet_data(&pkt);
568 po.packet_len = dp_packet_size(&pkt);
573 po.in_port = pi.flow_metadata.flow.in_port.ofp_port;
574 po.ofpacts = ofpacts.data;
575 po.ofpacts_len = ofpacts.size;
577 /* Send the packet, and possibly the whole flow, to the output port. */
578 if (sw->max_idle >= 0 && (!sw->ml || out_port != OFPP_FLOOD)) {
579 /* The output port is known, or we always flood everything, so add a
581 struct ofputil_flow_mod fm = {
582 .priority = 1, /* Must be > 0 because of table-miss flow entry. */
584 .command = OFPFC_ADD,
585 .idle_timeout = sw->max_idle,
586 .buffer_id = buffer_id,
587 .out_port = OFPP_NONE,
588 .ofpacts = ofpacts.data,
589 .ofpacts_len = ofpacts.size,
591 match_init(&fm.match, &flow, &sw->wc);
592 ofputil_normalize_match_quiet(&fm.match);
594 struct ofpbuf *buffer = ofputil_encode_flow_mod(&fm, sw->protocol);
596 queue_tx(sw, buffer);
598 /* If the switch didn't buffer the packet, we need to send a copy. */
599 if (buffer_id == UINT32_MAX && out_port != OFPP_NONE) {
600 queue_tx(sw, ofputil_encode_packet_out(&po, sw->protocol));
603 /* We don't know that MAC, or we don't set up flows. Send along the
604 * packet without setting up a flow. */
605 if (buffer_id != UINT32_MAX || out_port != OFPP_NONE) {
606 queue_tx(sw, ofputil_encode_packet_out(&po, sw->protocol));
612 process_echo_request(struct lswitch *sw, const struct ofp_header *rq)
614 queue_tx(sw, make_echo_reply(rq));
618 get_mac_entry_ofp_port(const struct mac_learning *ml,
619 const struct mac_entry *e)
620 OVS_REQ_RDLOCK(ml->rwlock)
622 void *port = mac_entry_get_port(ml, e);
623 return (OVS_FORCE ofp_port_t) (uintptr_t) port;
627 set_mac_entry_ofp_port(struct mac_learning *ml,
628 struct mac_entry *e, ofp_port_t ofp_port)
629 OVS_REQ_WRLOCK(ml->rwlock)
631 mac_entry_set_port(ml, e, (void *) (OVS_FORCE uintptr_t) ofp_port);