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"
30 #include "openflow/openflow.h"
31 #include "poll-loop.h"
39 #define THIS_MODULE VLM_learning_switch
46 P_FORWARDING = 1 << 3,
51 /* If nonnegative, the switch sets up flows that expire after the given
52 * number of seconds (or never expire, if the value is OFP_FLOW_PERMANENT).
53 * Otherwise, the switch processes every packet. */
56 unsigned long long int datapath_id;
57 uint32_t capabilities;
58 time_t last_features_request;
59 struct mac_learning *ml; /* NULL to act as hub instead of switch. */
60 bool exact_flows; /* Use exact-match flows? */
61 bool action_normal; /* Use OFPP_NORMAL? */
63 /* Number of outgoing queued packets on the rconn. */
64 struct rconn_packet_counter *queued;
66 /* Spanning tree protocol implementation.
68 * We implement STP states by, whenever a port's STP state changes,
69 * querying all the flows on the switch and then deleting any of them that
70 * are inappropriate for a port's STP state. */
71 long long int next_query; /* Next time at which to query all flows. */
72 long long int last_query; /* Last time we sent a query. */
73 long long int last_reply; /* Last time we received a query reply. */
74 unsigned int port_states[STP_MAX_PORTS];
75 uint32_t query_xid; /* XID used for query. */
76 int n_flows, n_no_recv, n_no_send;
79 /* The log messages here could actually be useful in debugging, so keep the
80 * rate limit relatively high. */
81 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(30, 300);
83 static void queue_tx(struct lswitch *, struct rconn *, struct ofpbuf *);
84 static void send_features_request(struct lswitch *, struct rconn *);
85 static void schedule_query(struct lswitch *, long long int delay);
86 static bool may_learn(const struct lswitch *, uint16_t port_no);
87 static bool may_recv(const struct lswitch *, uint16_t port_no,
89 static bool may_send(const struct lswitch *, uint16_t port_no);
91 typedef void packet_handler_func(struct lswitch *, struct rconn *, void *);
92 static packet_handler_func process_switch_features;
93 static packet_handler_func process_packet_in;
94 static packet_handler_func process_echo_request;
95 static packet_handler_func process_port_status;
96 static packet_handler_func process_phy_port;
97 static packet_handler_func process_stats_reply;
99 /* Creates and returns a new learning switch.
101 * If 'learn_macs' is true, the new switch will learn the ports on which MAC
102 * addresses appear. Otherwise, the new switch will flood all packets.
104 * If 'max_idle' is nonnegative, the new switch will set up flows that expire
105 * after the given number of seconds (or never expire, if 'max_idle' is
106 * OFP_FLOW_PERMANENT). Otherwise, the new switch will process every packet.
108 * 'rconn' is used to send out an OpenFlow features request. */
110 lswitch_create(struct rconn *rconn, bool learn_macs,
111 bool exact_flows, int max_idle, bool action_normal)
116 sw = xzalloc(sizeof *sw);
117 sw->max_idle = max_idle;
119 sw->last_features_request = time_now() - 1;
120 sw->ml = learn_macs ? mac_learning_create() : NULL;
121 sw->action_normal = action_normal;
122 sw->exact_flows = exact_flows;
123 sw->queued = rconn_packet_counter_create();
124 sw->next_query = LLONG_MIN;
125 sw->last_query = LLONG_MIN;
126 sw->last_reply = LLONG_MIN;
127 for (i = 0; i < STP_MAX_PORTS; i++) {
128 sw->port_states[i] = P_DISABLED;
130 send_features_request(sw, rconn);
136 lswitch_destroy(struct lswitch *sw)
139 mac_learning_destroy(sw->ml);
140 rconn_packet_counter_destroy(sw->queued);
145 /* Takes care of necessary 'sw' activity, except for receiving packets (which
146 * the caller must do). */
148 lswitch_run(struct lswitch *sw, struct rconn *rconn)
150 long long int now = time_msec();
153 mac_learning_run(sw->ml, NULL);
156 /* If we're waiting for more replies, keeping waiting for up to 10 s. */
157 if (sw->last_reply != LLONG_MIN) {
158 if (now - sw->last_reply > 10000) {
159 VLOG_ERR_RL(&rl, "%016llx: No more flow stat replies last 10 s",
161 sw->last_reply = LLONG_MIN;
162 sw->last_query = LLONG_MIN;
163 schedule_query(sw, 0);
169 /* If we're waiting for any reply at all, keep waiting for up to 10 s. */
170 if (sw->last_query != LLONG_MIN) {
171 if (now - sw->last_query > 10000) {
172 VLOG_ERR_RL(&rl, "%016llx: No flow stat replies in last 10 s",
174 sw->last_query = LLONG_MIN;
175 schedule_query(sw, 0);
181 /* If it's time to send another query, do so. */
182 if (sw->next_query != LLONG_MIN && now >= sw->next_query) {
183 sw->next_query = LLONG_MIN;
184 if (!rconn_is_connected(rconn)) {
185 schedule_query(sw, 1000);
187 struct ofp_stats_request *osr;
188 struct ofp_flow_stats_request *ofsr;
192 VLOG_DBG("%016llx: Sending flow stats request to implement STP",
195 sw->last_query = now;
196 sw->query_xid = random_uint32();
200 osr = make_openflow_xid(sizeof *osr + sizeof *ofsr,
201 OFPT_STATS_REQUEST, sw->query_xid, &b);
202 osr->type = htons(OFPST_FLOW);
203 osr->flags = htons(0);
204 ofsr = (struct ofp_flow_stats_request *) osr->body;
205 ofsr->match.wildcards = htonl(OFPFW_ALL);
206 ofsr->table_id = 0xff;
207 ofsr->out_port = htons(OFPP_NONE);
209 error = rconn_send(rconn, b, NULL);
211 VLOG_WARN_RL(&rl, "%016llx: sending flow stats request "
212 "failed: %s", sw->datapath_id, strerror(error));
214 schedule_query(sw, 1000);
221 wait_timeout(long long int started)
223 poll_timer_wait_until(started + 10000);
227 lswitch_wait(struct lswitch *sw)
230 mac_learning_wait(sw->ml);
233 if (sw->last_reply != LLONG_MIN) {
234 wait_timeout(sw->last_reply);
235 } else if (sw->last_query != LLONG_MIN) {
236 wait_timeout(sw->last_query);
240 /* Processes 'msg', which should be an OpenFlow received on 'rconn', according
241 * to the learning switch state in 'sw'. The most likely result of processing
242 * is that flow-setup and packet-out OpenFlow messages will be sent out on
245 lswitch_process_packet(struct lswitch *sw, struct rconn *rconn,
246 const struct ofpbuf *msg)
251 packet_handler_func *handler;
253 static const struct processor processors[] = {
256 sizeof(struct ofp_header),
261 sizeof(struct ofp_switch_features),
262 process_switch_features
266 offsetof(struct ofp_packet_in, data),
271 sizeof(struct ofp_port_status),
276 offsetof(struct ofp_stats_reply, body),
281 sizeof(struct ofp_flow_removed),
285 const size_t n_processors = ARRAY_SIZE(processors);
286 const struct processor *p;
287 struct ofp_header *oh;
290 if (sw->datapath_id == 0
291 && oh->type != OFPT_ECHO_REQUEST
292 && oh->type != OFPT_FEATURES_REPLY) {
293 send_features_request(sw, rconn);
297 for (p = processors; p < &processors[n_processors]; p++) {
298 if (oh->type == p->type) {
299 if (msg->size < p->min_size) {
300 VLOG_WARN_RL(&rl, "%016llx: %s: too short (%zu bytes) for "
301 "type %"PRIu8" (min %zu)", sw->datapath_id,
302 rconn_get_name(rconn), msg->size, oh->type,
307 (p->handler)(sw, rconn, msg->data);
312 if (VLOG_IS_DBG_ENABLED()) {
313 char *p = ofp_to_string(msg->data, msg->size, 2);
314 VLOG_DBG_RL(&rl, "%016llx: OpenFlow packet ignored: %s",
321 send_features_request(struct lswitch *sw, struct rconn *rconn)
323 time_t now = time_now();
324 if (now >= sw->last_features_request + 1) {
326 struct ofp_switch_config *osc;
328 /* Send OFPT_FEATURES_REQUEST. */
329 make_openflow(sizeof(struct ofp_header), OFPT_FEATURES_REQUEST, &b);
330 queue_tx(sw, rconn, b);
332 /* Send OFPT_SET_CONFIG. */
333 osc = make_openflow(sizeof *osc, OFPT_SET_CONFIG, &b);
334 osc->miss_send_len = htons(OFP_DEFAULT_MISS_SEND_LEN);
335 queue_tx(sw, rconn, b);
337 sw->last_features_request = now;
342 queue_tx(struct lswitch *sw, struct rconn *rconn, struct ofpbuf *b)
344 int retval = rconn_send_with_limit(rconn, b, sw->queued, 10);
345 if (retval && retval != ENOTCONN) {
346 if (retval == EAGAIN) {
347 VLOG_INFO_RL(&rl, "%016llx: %s: tx queue overflow",
348 sw->datapath_id, rconn_get_name(rconn));
350 VLOG_WARN_RL(&rl, "%016llx: %s: send: %s",
351 sw->datapath_id, rconn_get_name(rconn),
358 schedule_query(struct lswitch *sw, long long int delay)
360 long long int now = time_msec();
361 if (sw->next_query == LLONG_MIN || sw->next_query > now + delay) {
362 sw->next_query = now + delay;
367 process_switch_features(struct lswitch *sw, struct rconn *rconn, void *osf_)
369 struct ofp_switch_features *osf = osf_;
370 size_t n_ports = ((ntohs(osf->header.length)
371 - offsetof(struct ofp_switch_features, ports))
372 / sizeof *osf->ports);
375 sw->datapath_id = ntohll(osf->datapath_id);
376 sw->capabilities = ntohl(osf->capabilities);
377 for (i = 0; i < n_ports; i++) {
378 process_phy_port(sw, rconn, &osf->ports[i]);
380 if (sw->capabilities & OFPC_STP) {
381 schedule_query(sw, 1000);
386 process_packet_in(struct lswitch *sw, struct rconn *rconn, void *opi_)
388 struct ofp_packet_in *opi = opi_;
389 uint16_t in_port = ntohs(opi->in_port);
390 uint16_t out_port = OFPP_FLOOD;
392 size_t pkt_ofs, pkt_len;
396 /* Extract flow data from 'opi' into 'flow'. */
397 pkt_ofs = offsetof(struct ofp_packet_in, data);
398 pkt_len = ntohs(opi->header.length) - pkt_ofs;
399 pkt.data = opi->data;
401 flow_extract(&pkt, 0, in_port, &flow);
403 if (may_learn(sw, in_port) && sw->ml) {
404 if (mac_learning_learn(sw->ml, flow.dl_src, 0, in_port,
405 GRAT_ARP_LOCK_NONE)) {
406 VLOG_DBG_RL(&rl, "%016llx: learned that "ETH_ADDR_FMT" is on "
407 "port %"PRIu16, sw->datapath_id,
408 ETH_ADDR_ARGS(flow.dl_src), in_port);
412 if (eth_addr_is_reserved(flow.dl_src)) {
416 if (!may_recv(sw, in_port, false)) {
417 /* STP prevents receiving anything on this port. */
422 int learned_port = mac_learning_lookup(sw->ml, flow.dl_dst, 0, NULL);
423 if (learned_port >= 0 && may_send(sw, learned_port)) {
424 out_port = learned_port;
428 if (in_port == out_port) {
429 /* Don't send out packets on their input ports. */
431 } else if (sw->max_idle >= 0 && (!sw->ml || out_port != OFPP_FLOOD)) {
432 struct ofpbuf *buffer;
433 struct ofp_flow_mod *ofm;
436 /* Check if we need to wildcard the flows. */
437 if (!sw->exact_flows) {
438 /* We can not wildcard all fields.
439 * We need in_port to detect moves.
440 * We need both SA and DA to do learning. */
441 wildcards = (OFPFW_DL_TYPE | OFPFW_NW_SRC_MASK | OFPFW_NW_DST_MASK
442 | OFPFW_NW_PROTO | OFPFW_TP_SRC | OFPFW_TP_DST);
448 /* Check if we need to use "NORMAL" action. */
449 if (sw->action_normal && out_port != OFPP_FLOOD) {
450 out_port = OFPP_NORMAL;
453 /* The output port is known, or we always flood everything, so add a
455 buffer = make_add_simple_flow(&flow, ntohl(opi->buffer_id),
456 out_port, sw->max_idle);
458 ofm->match.wildcards = htonl(wildcards);
459 queue_tx(sw, rconn, buffer);
461 /* If the switch didn't buffer the packet, we need to send a copy. */
462 if (ntohl(opi->buffer_id) == UINT32_MAX) {
464 make_unbuffered_packet_out(&pkt, in_port, out_port));
469 /* Check if we need to use "NORMAL" action. */
470 if (sw->action_normal && out_port != OFPP_FLOOD) {
471 out_port = OFPP_NORMAL;
474 /* We don't know that MAC, or we don't set up flows. Send along the
475 * packet without setting up a flow. */
476 if (ntohl(opi->buffer_id) == UINT32_MAX) {
477 b = make_unbuffered_packet_out(&pkt, in_port, out_port);
479 b = make_buffered_packet_out(ntohl(opi->buffer_id),
482 queue_tx(sw, rconn, b);
487 if (sw->max_idle >= 0) {
488 /* Set up a flow to drop packets. */
489 queue_tx(sw, rconn, make_add_flow(&flow, ntohl(opi->buffer_id),
492 /* Just drop the packet, since we don't set up flows at all.
493 * XXX we should send a packet_out with no actions if buffer_id !=
494 * UINT32_MAX, to avoid clogging the kernel buffers. */
500 process_echo_request(struct lswitch *sw, struct rconn *rconn, void *rq_)
502 struct ofp_header *rq = rq_;
503 queue_tx(sw, rconn, make_echo_reply(rq));
507 process_port_status(struct lswitch *sw, struct rconn *rconn, void *ops_)
509 struct ofp_port_status *ops = ops_;
510 process_phy_port(sw, rconn, &ops->desc);
514 process_phy_port(struct lswitch *sw, struct rconn *rconn OVS_UNUSED,
517 const struct ofp_phy_port *opp = opp_;
518 uint16_t port_no = ntohs(opp->port_no);
519 if (sw->capabilities & OFPC_STP && port_no < STP_MAX_PORTS) {
520 uint32_t config = ntohl(opp->config);
521 uint32_t state = ntohl(opp->state);
522 unsigned int *port_state = &sw->port_states[port_no];
523 unsigned int new_port_state;
525 if (!(config & (OFPPC_NO_STP | OFPPC_PORT_DOWN))
526 && !(state & OFPPS_LINK_DOWN))
528 switch (state & OFPPS_STP_MASK) {
529 case OFPPS_STP_LISTEN:
530 new_port_state = P_LISTENING;
532 case OFPPS_STP_LEARN:
533 new_port_state = P_LEARNING;
535 case OFPPS_STP_FORWARD:
536 new_port_state = P_FORWARDING;
538 case OFPPS_STP_BLOCK:
539 new_port_state = P_BLOCKING;
542 new_port_state = P_DISABLED;
546 new_port_state = P_FORWARDING;
548 if (*port_state != new_port_state) {
549 *port_state = new_port_state;
550 schedule_query(sw, 1000);
556 get_port_state(const struct lswitch *sw, uint16_t port_no)
558 return (port_no >= STP_MAX_PORTS || !(sw->capabilities & OFPC_STP)
560 : sw->port_states[port_no]);
564 may_learn(const struct lswitch *sw, uint16_t port_no)
566 return get_port_state(sw, port_no) & (P_LEARNING | P_FORWARDING);
570 may_recv(const struct lswitch *sw, uint16_t port_no, bool any_actions)
572 unsigned int state = get_port_state(sw, port_no);
574 ? state & (P_DISABLED | P_LISTENING | P_BLOCKING)
575 : state & (P_DISABLED | P_LISTENING | P_BLOCKING | P_LEARNING));
579 may_send(const struct lswitch *sw, uint16_t port_no)
581 return get_port_state(sw, port_no) & P_FORWARDING;
585 process_flow_stats(struct lswitch *sw, struct rconn *rconn,
586 const struct ofp_flow_stats *ofs)
588 const char *end = (char *) ofs + ntohs(ofs->length);
591 /* Decide to delete the flow if it matches on an STP-disabled physical
592 * port. But don't delete it if the flow just drops all received packets,
593 * because that's a perfectly reasonable thing to do for disabled physical
595 if (!(ofs->match.wildcards & htonl(OFPFW_IN_PORT))) {
596 if (!may_recv(sw, ntohs(ofs->match.in_port),
597 end > (char *) ofs->actions)) {
603 /* Decide to delete the flow if it forwards to an STP-disabled physical
606 const struct ofp_action_header *a;
609 for (a = ofs->actions; (char *) a < end; a += len / 8) {
611 if (len > end - (char *) a) {
612 VLOG_DBG_RL(&rl, "%016llx: action exceeds available space "
614 sw->datapath_id, len, end - (char *) a);
616 } else if (len % 8) {
617 VLOG_DBG_RL(&rl, "%016llx: action length (%zu) not multiple "
618 "of 8 bytes", sw->datapath_id, len);
622 if (a->type == htons(OFPAT_OUTPUT)) {
623 struct ofp_action_output *oao = (struct ofp_action_output *) a;
624 if (!may_send(sw, ntohs(oao->port))) {
633 /* Delete the flow. */
635 struct ofp_flow_mod *ofm;
638 ofm = make_openflow(offsetof(struct ofp_flow_mod, actions),
640 ofm->match = ofs->match;
641 ofm->command = OFPFC_DELETE_STRICT;
642 rconn_send(rconn, b, NULL);
647 process_stats_reply(struct lswitch *sw, struct rconn *rconn, void *osr_)
649 struct ofp_stats_reply *osr = osr_;
650 struct flow_stats_iterator i;
651 const struct ofp_flow_stats *fs;
653 if (sw->last_query == LLONG_MIN
654 || osr->type != htons(OFPST_FLOW)
655 || osr->header.xid != sw->query_xid) {
658 for (fs = flow_stats_first(&i, osr); fs; fs = flow_stats_next(&i)) {
660 process_flow_stats(sw, rconn, fs);
662 if (!(osr->flags & htons(OFPSF_REPLY_MORE))) {
663 VLOG_DBG("%016llx: Deleted %d of %d received flows to "
664 "implement STP, %d because of no-recv, %d because of "
665 "no-send", sw->datapath_id,
666 sw->n_no_recv + sw->n_no_send, sw->n_flows,
667 sw->n_no_recv, sw->n_no_send);
668 sw->last_query = LLONG_MIN;
669 sw->last_reply = LLONG_MIN;
671 sw->last_reply = time_msec();