2 * Copyright (c) 2008, 2009, 2010 Nicira Networks.
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>
27 #include "mac-learning.h"
29 #include "ofp-print.h"
31 #include "openflow/openflow.h"
32 #include "poll-loop.h"
40 #define THIS_MODULE VLM_learning_switch
47 P_FORWARDING = 1 << 3,
52 /* If nonnegative, the switch sets up flows that expire after the given
53 * number of seconds (or never expire, if the value is OFP_FLOW_PERMANENT).
54 * Otherwise, the switch processes every packet. */
57 unsigned long long int datapath_id;
58 uint32_t capabilities;
59 time_t last_features_request;
60 struct mac_learning *ml; /* NULL to act as hub instead of switch. */
61 uint32_t wildcards; /* Wildcards to apply to flows. */
62 bool action_normal; /* Use OFPP_NORMAL? */
63 uint32_t queue; /* OpenFlow queue to use, or UINT32_MAX. */
65 /* Number of outgoing queued packets on the rconn. */
66 struct rconn_packet_counter *queued;
68 /* Spanning tree protocol implementation.
70 * We implement STP states by, whenever a port's STP state changes,
71 * querying all the flows on the switch and then deleting any of them that
72 * are inappropriate for a port's STP state. */
73 long long int next_query; /* Next time at which to query all flows. */
74 long long int last_query; /* Last time we sent a query. */
75 long long int last_reply; /* Last time we received a query reply. */
76 unsigned int port_states[STP_MAX_PORTS];
77 uint32_t query_xid; /* XID used for query. */
78 int n_flows, n_no_recv, n_no_send;
81 /* The log messages here could actually be useful in debugging, so keep the
82 * rate limit relatively high. */
83 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(30, 300);
85 static void queue_tx(struct lswitch *, struct rconn *, struct ofpbuf *);
86 static void send_features_request(struct lswitch *, struct rconn *);
87 static void schedule_query(struct lswitch *, long long int delay);
88 static bool may_learn(const struct lswitch *, uint16_t port_no);
89 static bool may_recv(const struct lswitch *, uint16_t port_no,
91 static bool may_send(const struct lswitch *, uint16_t port_no);
93 typedef void packet_handler_func(struct lswitch *, struct rconn *, void *);
94 static packet_handler_func process_switch_features;
95 static packet_handler_func process_packet_in;
96 static packet_handler_func process_echo_request;
97 static packet_handler_func process_port_status;
98 static packet_handler_func process_phy_port;
99 static packet_handler_func process_stats_reply;
101 /* Creates and returns a new learning switch.
103 * If 'learn_macs' is true, the new switch will learn the ports on which MAC
104 * addresses appear. Otherwise, the new switch will flood all packets.
106 * If 'max_idle' is nonnegative, the new switch will set up flows that expire
107 * after the given number of seconds (or never expire, if 'max_idle' is
108 * OFP_FLOW_PERMANENT). Otherwise, the new switch will process every packet.
110 * 'rconn' is used to send out an OpenFlow features request. */
112 lswitch_create(struct rconn *rconn, bool learn_macs,
113 bool exact_flows, int max_idle, bool action_normal)
118 sw = xzalloc(sizeof *sw);
119 sw->max_idle = max_idle;
121 sw->last_features_request = time_now() - 1;
122 sw->ml = learn_macs ? mac_learning_create() : NULL;
123 sw->action_normal = action_normal;
128 /* We cannot wildcard all fields.
129 * We need in_port to detect moves.
130 * We need both SA and DA to do learning. */
131 sw->wildcards = (OFPFW_DL_TYPE | OFPFW_NW_SRC_MASK | OFPFW_NW_DST_MASK
132 | OFPFW_NW_PROTO | OFPFW_TP_SRC | OFPFW_TP_DST);
134 sw->queue = UINT32_MAX;
135 sw->queued = rconn_packet_counter_create();
136 sw->next_query = LLONG_MIN;
137 sw->last_query = LLONG_MIN;
138 sw->last_reply = LLONG_MIN;
139 for (i = 0; i < STP_MAX_PORTS; i++) {
140 sw->port_states[i] = P_DISABLED;
142 send_features_request(sw, rconn);
148 lswitch_destroy(struct lswitch *sw)
151 mac_learning_destroy(sw->ml);
152 rconn_packet_counter_destroy(sw->queued);
157 /* Sets 'queue' as the OpenFlow queue used by packets and flows set up by 'sw'.
158 * Specify UINT32_MAX to avoid specifying a particular queue, which is also the
159 * default if this function is never called for 'sw'. */
161 lswitch_set_queue(struct lswitch *sw, uint32_t queue)
166 /* Takes care of necessary 'sw' activity, except for receiving packets (which
167 * the caller must do). */
169 lswitch_run(struct lswitch *sw, struct rconn *rconn)
171 long long int now = time_msec();
174 mac_learning_run(sw->ml, NULL);
177 /* If we're waiting for more replies, keeping waiting for up to 10 s. */
178 if (sw->last_reply != LLONG_MIN) {
179 if (now - sw->last_reply > 10000) {
180 VLOG_ERR_RL(&rl, "%016llx: No more flow stat replies last 10 s",
182 sw->last_reply = LLONG_MIN;
183 sw->last_query = LLONG_MIN;
184 schedule_query(sw, 0);
190 /* If we're waiting for any reply at all, keep waiting for up to 10 s. */
191 if (sw->last_query != LLONG_MIN) {
192 if (now - sw->last_query > 10000) {
193 VLOG_ERR_RL(&rl, "%016llx: No flow stat replies in last 10 s",
195 sw->last_query = LLONG_MIN;
196 schedule_query(sw, 0);
202 /* If it's time to send another query, do so. */
203 if (sw->next_query != LLONG_MIN && now >= sw->next_query) {
204 sw->next_query = LLONG_MIN;
205 if (!rconn_is_connected(rconn)) {
206 schedule_query(sw, 1000);
208 struct ofp_stats_request *osr;
209 struct ofp_flow_stats_request *ofsr;
213 VLOG_DBG("%016llx: Sending flow stats request to implement STP",
216 sw->last_query = now;
217 sw->query_xid = random_uint32();
221 osr = make_openflow_xid(sizeof *osr + sizeof *ofsr,
222 OFPT_STATS_REQUEST, sw->query_xid, &b);
223 osr->type = htons(OFPST_FLOW);
224 osr->flags = htons(0);
225 ofsr = (struct ofp_flow_stats_request *) osr->body;
226 ofsr->match.wildcards = htonl(OFPFW_ALL);
227 ofsr->table_id = 0xff;
228 ofsr->out_port = htons(OFPP_NONE);
230 error = rconn_send(rconn, b, NULL);
232 VLOG_WARN_RL(&rl, "%016llx: sending flow stats request "
233 "failed: %s", sw->datapath_id, strerror(error));
235 schedule_query(sw, 1000);
242 wait_timeout(long long int started)
244 poll_timer_wait_until(started + 10000);
248 lswitch_wait(struct lswitch *sw)
251 mac_learning_wait(sw->ml);
254 if (sw->last_reply != LLONG_MIN) {
255 wait_timeout(sw->last_reply);
256 } else if (sw->last_query != LLONG_MIN) {
257 wait_timeout(sw->last_query);
261 /* Processes 'msg', which should be an OpenFlow received on 'rconn', according
262 * to the learning switch state in 'sw'. The most likely result of processing
263 * is that flow-setup and packet-out OpenFlow messages will be sent out on
266 lswitch_process_packet(struct lswitch *sw, struct rconn *rconn,
267 const struct ofpbuf *msg)
272 packet_handler_func *handler;
274 static const struct processor processors[] = {
277 sizeof(struct ofp_header),
282 sizeof(struct ofp_switch_features),
283 process_switch_features
287 offsetof(struct ofp_packet_in, data),
292 sizeof(struct ofp_port_status),
297 offsetof(struct ofp_stats_reply, body),
302 sizeof(struct ofp_flow_removed),
306 const size_t n_processors = ARRAY_SIZE(processors);
307 const struct processor *p;
308 struct ofp_header *oh;
311 if (sw->datapath_id == 0
312 && oh->type != OFPT_ECHO_REQUEST
313 && oh->type != OFPT_FEATURES_REPLY) {
314 send_features_request(sw, rconn);
318 for (p = processors; p < &processors[n_processors]; p++) {
319 if (oh->type == p->type) {
320 if (msg->size < p->min_size) {
321 VLOG_WARN_RL(&rl, "%016llx: %s: too short (%zu bytes) for "
322 "type %"PRIu8" (min %zu)", sw->datapath_id,
323 rconn_get_name(rconn), msg->size, oh->type,
328 (p->handler)(sw, rconn, msg->data);
333 if (VLOG_IS_DBG_ENABLED()) {
334 char *p = ofp_to_string(msg->data, msg->size, 2);
335 VLOG_DBG_RL(&rl, "%016llx: OpenFlow packet ignored: %s",
342 send_features_request(struct lswitch *sw, struct rconn *rconn)
344 time_t now = time_now();
345 if (now >= sw->last_features_request + 1) {
347 struct ofp_switch_config *osc;
349 /* Send OFPT_FEATURES_REQUEST. */
350 make_openflow(sizeof(struct ofp_header), OFPT_FEATURES_REQUEST, &b);
351 queue_tx(sw, rconn, b);
353 /* Send OFPT_SET_CONFIG. */
354 osc = make_openflow(sizeof *osc, OFPT_SET_CONFIG, &b);
355 osc->miss_send_len = htons(OFP_DEFAULT_MISS_SEND_LEN);
356 queue_tx(sw, rconn, b);
358 sw->last_features_request = now;
363 queue_tx(struct lswitch *sw, struct rconn *rconn, struct ofpbuf *b)
365 int retval = rconn_send_with_limit(rconn, b, sw->queued, 10);
366 if (retval && retval != ENOTCONN) {
367 if (retval == EAGAIN) {
368 VLOG_INFO_RL(&rl, "%016llx: %s: tx queue overflow",
369 sw->datapath_id, rconn_get_name(rconn));
371 VLOG_WARN_RL(&rl, "%016llx: %s: send: %s",
372 sw->datapath_id, rconn_get_name(rconn),
379 schedule_query(struct lswitch *sw, long long int delay)
381 long long int now = time_msec();
382 if (sw->next_query == LLONG_MIN || sw->next_query > now + delay) {
383 sw->next_query = now + delay;
388 process_switch_features(struct lswitch *sw, struct rconn *rconn, void *osf_)
390 struct ofp_switch_features *osf = osf_;
391 size_t n_ports = ((ntohs(osf->header.length)
392 - offsetof(struct ofp_switch_features, ports))
393 / sizeof *osf->ports);
396 sw->datapath_id = ntohll(osf->datapath_id);
397 sw->capabilities = ntohl(osf->capabilities);
398 for (i = 0; i < n_ports; i++) {
399 process_phy_port(sw, rconn, &osf->ports[i]);
401 if (sw->capabilities & OFPC_STP) {
402 schedule_query(sw, 1000);
407 lswitch_choose_destination(struct lswitch *sw, const flow_t *flow)
411 /* Learn the source MAC. */
412 if (may_learn(sw, flow->in_port) && sw->ml) {
413 if (mac_learning_learn(sw->ml, flow->dl_src, 0, flow->in_port,
414 GRAT_ARP_LOCK_NONE)) {
415 VLOG_DBG_RL(&rl, "%016llx: learned that "ETH_ADDR_FMT" is on "
416 "port %"PRIu16, sw->datapath_id,
417 ETH_ADDR_ARGS(flow->dl_src), flow->in_port);
421 /* Drop frames for reserved multicast addresses. */
422 if (eth_addr_is_reserved(flow->dl_dst)) {
426 if (!may_recv(sw, flow->in_port, false)) {
427 /* STP prevents receiving anything on this port. */
431 out_port = OFPP_FLOOD;
433 int learned_port = mac_learning_lookup(sw->ml, flow->dl_dst, 0, NULL);
434 if (learned_port >= 0 && may_send(sw, learned_port)) {
435 out_port = learned_port;
436 if (out_port == flow->in_port) {
437 /* Don't send a packet back out its input port. */
443 /* Check if we need to use "NORMAL" action. */
444 if (sw->action_normal && out_port != OFPP_FLOOD) {
452 process_packet_in(struct lswitch *sw, struct rconn *rconn, void *opi_)
454 struct ofp_packet_in *opi = opi_;
455 uint16_t in_port = ntohs(opi->in_port);
458 struct ofp_action_header actions[2];
461 size_t pkt_ofs, pkt_len;
465 /* Extract flow data from 'opi' into 'flow'. */
466 pkt_ofs = offsetof(struct ofp_packet_in, data);
467 pkt_len = ntohs(opi->header.length) - pkt_ofs;
468 pkt.data = opi->data;
470 flow_extract(&pkt, 0, in_port, &flow);
472 /* Choose output port. */
473 out_port = lswitch_choose_destination(sw, &flow);
476 memset(actions, 0, sizeof actions);
477 if (out_port == OFPP_NONE) {
479 } else if (sw->queue == UINT32_MAX || out_port >= OFPP_MAX) {
480 struct ofp_action_output *oao = (struct ofp_action_output *) actions;
481 oao->type = htons(OFPAT_OUTPUT);
482 oao->len = htons(sizeof *oao);
483 oao->port = htons(out_port);
484 actions_len = sizeof *oao;
486 struct ofp_action_enqueue *oae = (struct ofp_action_enqueue *) actions;
487 oae->type = htons(OFPAT_ENQUEUE);
488 oae->len = htons(sizeof *oae);
489 oae->port = htons(out_port);
490 oae->queue_id = htonl(sw->queue);
491 actions_len = sizeof *oae;
493 assert(actions_len <= sizeof actions);
495 /* Send the packet, and possibly the whole flow, to the output port. */
496 if (sw->max_idle >= 0 && (!sw->ml || out_port != OFPP_FLOOD)) {
497 struct ofpbuf *buffer;
498 struct ofp_flow_mod *ofm;
500 /* The output port is known, or we always flood everything, so add a
502 buffer = make_add_flow(&flow, ntohl(opi->buffer_id),
503 sw->max_idle, actions_len);
504 ofpbuf_put(buffer, actions, actions_len);
506 ofm->match.wildcards = htonl(sw->wildcards);
507 queue_tx(sw, rconn, buffer);
509 /* If the switch didn't buffer the packet, we need to send a copy. */
510 if (ntohl(opi->buffer_id) == UINT32_MAX && actions_len > 0) {
512 make_packet_out(&pkt, UINT32_MAX, in_port,
513 actions, actions_len / sizeof *actions));
516 /* We don't know that MAC, or we don't set up flows. Send along the
517 * packet without setting up a flow. */
518 if (ntohl(opi->buffer_id) != UINT32_MAX || actions_len > 0) {
520 make_packet_out(&pkt, ntohl(opi->buffer_id), in_port,
521 actions, actions_len / sizeof *actions));
527 process_echo_request(struct lswitch *sw, struct rconn *rconn, void *rq_)
529 struct ofp_header *rq = rq_;
530 queue_tx(sw, rconn, make_echo_reply(rq));
534 process_port_status(struct lswitch *sw, struct rconn *rconn, void *ops_)
536 struct ofp_port_status *ops = ops_;
537 process_phy_port(sw, rconn, &ops->desc);
541 process_phy_port(struct lswitch *sw, struct rconn *rconn OVS_UNUSED,
544 const struct ofp_phy_port *opp = opp_;
545 uint16_t port_no = ntohs(opp->port_no);
546 if (sw->capabilities & OFPC_STP && port_no < STP_MAX_PORTS) {
547 uint32_t config = ntohl(opp->config);
548 uint32_t state = ntohl(opp->state);
549 unsigned int *port_state = &sw->port_states[port_no];
550 unsigned int new_port_state;
552 if (!(config & (OFPPC_NO_STP | OFPPC_PORT_DOWN))
553 && !(state & OFPPS_LINK_DOWN))
555 switch (state & OFPPS_STP_MASK) {
556 case OFPPS_STP_LISTEN:
557 new_port_state = P_LISTENING;
559 case OFPPS_STP_LEARN:
560 new_port_state = P_LEARNING;
562 case OFPPS_STP_FORWARD:
563 new_port_state = P_FORWARDING;
565 case OFPPS_STP_BLOCK:
566 new_port_state = P_BLOCKING;
569 new_port_state = P_DISABLED;
573 new_port_state = P_FORWARDING;
575 if (*port_state != new_port_state) {
576 *port_state = new_port_state;
577 schedule_query(sw, 1000);
583 get_port_state(const struct lswitch *sw, uint16_t port_no)
585 return (port_no >= STP_MAX_PORTS || !(sw->capabilities & OFPC_STP)
587 : sw->port_states[port_no]);
591 may_learn(const struct lswitch *sw, uint16_t port_no)
593 return get_port_state(sw, port_no) & (P_LEARNING | P_FORWARDING);
597 may_recv(const struct lswitch *sw, uint16_t port_no, bool any_actions)
599 unsigned int state = get_port_state(sw, port_no);
601 ? state & (P_DISABLED | P_LISTENING | P_BLOCKING)
602 : state & (P_DISABLED | P_LISTENING | P_BLOCKING | P_LEARNING));
606 may_send(const struct lswitch *sw, uint16_t port_no)
608 return get_port_state(sw, port_no) & P_FORWARDING;
612 process_flow_stats(struct lswitch *sw, struct rconn *rconn,
613 const struct ofp_flow_stats *ofs)
615 const char *end = (char *) ofs + ntohs(ofs->length);
618 /* Decide to delete the flow if it matches on an STP-disabled physical
619 * port. But don't delete it if the flow just drops all received packets,
620 * because that's a perfectly reasonable thing to do for disabled physical
622 if (!(ofs->match.wildcards & htonl(OFPFW_IN_PORT))) {
623 if (!may_recv(sw, ntohs(ofs->match.in_port),
624 end > (char *) ofs->actions)) {
630 /* Decide to delete the flow if it forwards to an STP-disabled physical
633 const struct ofp_action_header *a;
636 for (a = ofs->actions; (char *) a < end; a += len / 8) {
638 if (len > end - (char *) a) {
639 VLOG_DBG_RL(&rl, "%016llx: action exceeds available space "
641 sw->datapath_id, len, end - (char *) a);
643 } else if (len % 8) {
644 VLOG_DBG_RL(&rl, "%016llx: action length (%zu) not multiple "
645 "of 8 bytes", sw->datapath_id, len);
649 if (a->type == htons(OFPAT_OUTPUT)) {
650 struct ofp_action_output *oao = (struct ofp_action_output *) a;
651 if (!may_send(sw, ntohs(oao->port))) {
660 /* Delete the flow. */
662 struct ofp_flow_mod *ofm;
665 ofm = make_openflow(offsetof(struct ofp_flow_mod, actions),
667 ofm->match = ofs->match;
668 ofm->command = OFPFC_DELETE_STRICT;
669 rconn_send(rconn, b, NULL);
674 process_stats_reply(struct lswitch *sw, struct rconn *rconn, void *osr_)
676 struct ofp_stats_reply *osr = osr_;
677 struct flow_stats_iterator i;
678 const struct ofp_flow_stats *fs;
680 if (sw->last_query == LLONG_MIN
681 || osr->type != htons(OFPST_FLOW)
682 || osr->header.xid != sw->query_xid) {
685 for (fs = flow_stats_first(&i, osr); fs; fs = flow_stats_next(&i)) {
687 process_flow_stats(sw, rconn, fs);
689 if (!(osr->flags & htons(OFPSF_REPLY_MORE))) {
690 VLOG_DBG("%016llx: Deleted %d of %d received flows to "
691 "implement STP, %d because of no-recv, %d because of "
692 "no-send", sw->datapath_id,
693 sw->n_no_recv + sw->n_no_send, sw->n_flows,
694 sw->n_no_recv, sw->n_no_send);
695 sw->last_query = LLONG_MIN;
696 sw->last_reply = LLONG_MIN;
698 sw->last_reply = time_msec();