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-parse.h"
30 #include "ofp-print.h"
32 #include "openflow/openflow.h"
33 #include "poll-loop.h"
41 VLOG_DEFINE_THIS_MODULE(learning_switch)
44 /* If nonnegative, the switch sets up flows that expire after the given
45 * number of seconds (or never expire, if the value is OFP_FLOW_PERMANENT).
46 * Otherwise, the switch processes every packet. */
49 unsigned long long int datapath_id;
50 time_t last_features_request;
51 struct mac_learning *ml; /* NULL to act as hub instead of switch. */
52 uint32_t wildcards; /* Wildcards to apply to flows. */
53 bool action_normal; /* Use OFPP_NORMAL? */
54 uint32_t queue; /* OpenFlow queue to use, or UINT32_MAX. */
56 /* Number of outgoing queued packets on the rconn. */
57 struct rconn_packet_counter *queued;
60 /* The log messages here could actually be useful in debugging, so keep the
61 * rate limit relatively high. */
62 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(30, 300);
64 static void queue_tx(struct lswitch *, struct rconn *, struct ofpbuf *);
65 static void send_features_request(struct lswitch *, struct rconn *);
66 static void send_default_flows(struct lswitch *sw, struct rconn *rconn,
69 typedef void packet_handler_func(struct lswitch *, struct rconn *, void *);
70 static packet_handler_func process_switch_features;
71 static packet_handler_func process_packet_in;
72 static packet_handler_func process_echo_request;
74 /* Creates and returns a new learning switch.
76 * If 'learn_macs' is true, the new switch will learn the ports on which MAC
77 * addresses appear. Otherwise, the new switch will flood all packets.
79 * If 'max_idle' is nonnegative, the new switch will set up flows that expire
80 * after the given number of seconds (or never expire, if 'max_idle' is
81 * OFP_FLOW_PERMANENT). Otherwise, the new switch will process every packet.
83 * The caller may provide the file stream 'default_flows' that defines
84 * default flows that should be pushed when a switch connects. Each
85 * line is a flow entry in the format described for "add-flows" command
86 * in the Flow Syntax section of the ovs-ofct(8) man page. The caller
87 * is responsible for closing the stream.
89 * 'rconn' is used to send out an OpenFlow features request. */
91 lswitch_create(struct rconn *rconn, bool learn_macs,
92 bool exact_flows, int max_idle, bool action_normal,
97 sw = xzalloc(sizeof *sw);
98 sw->max_idle = max_idle;
100 sw->last_features_request = time_now() - 1;
101 sw->ml = learn_macs ? mac_learning_create() : NULL;
102 sw->action_normal = action_normal;
107 /* We cannot wildcard all fields.
108 * We need in_port to detect moves.
109 * We need both SA and DA to do learning. */
110 sw->wildcards = (OFPFW_DL_TYPE | OFPFW_NW_SRC_MASK | OFPFW_NW_DST_MASK
111 | OFPFW_NW_PROTO | OFPFW_TP_SRC | OFPFW_TP_DST);
113 sw->queue = UINT32_MAX;
114 sw->queued = rconn_packet_counter_create();
115 send_features_request(sw, rconn);
117 send_default_flows(sw, rconn, default_flows);
124 lswitch_destroy(struct lswitch *sw)
127 mac_learning_destroy(sw->ml);
128 rconn_packet_counter_destroy(sw->queued);
133 /* Sets 'queue' as the OpenFlow queue used by packets and flows set up by 'sw'.
134 * Specify UINT32_MAX to avoid specifying a particular queue, which is also the
135 * default if this function is never called for 'sw'. */
137 lswitch_set_queue(struct lswitch *sw, uint32_t queue)
142 /* Takes care of necessary 'sw' activity, except for receiving packets (which
143 * the caller must do). */
145 lswitch_run(struct lswitch *sw)
148 mac_learning_run(sw->ml, NULL);
153 lswitch_wait(struct lswitch *sw)
156 mac_learning_wait(sw->ml);
160 /* Processes 'msg', which should be an OpenFlow received on 'rconn', according
161 * to the learning switch state in 'sw'. The most likely result of processing
162 * is that flow-setup and packet-out OpenFlow messages will be sent out on
165 lswitch_process_packet(struct lswitch *sw, struct rconn *rconn,
166 const struct ofpbuf *msg)
171 packet_handler_func *handler;
173 static const struct processor processors[] = {
176 sizeof(struct ofp_header),
181 sizeof(struct ofp_switch_features),
182 process_switch_features
186 offsetof(struct ofp_packet_in, data),
191 sizeof(struct ofp_flow_removed),
195 const size_t n_processors = ARRAY_SIZE(processors);
196 const struct processor *p;
197 struct ofp_header *oh;
200 if (sw->datapath_id == 0
201 && oh->type != OFPT_ECHO_REQUEST
202 && oh->type != OFPT_FEATURES_REPLY) {
203 send_features_request(sw, rconn);
207 for (p = processors; p < &processors[n_processors]; p++) {
208 if (oh->type == p->type) {
209 if (msg->size < p->min_size) {
210 VLOG_WARN_RL(&rl, "%016llx: %s: too short (%zu bytes) for "
211 "type %"PRIu8" (min %zu)", sw->datapath_id,
212 rconn_get_name(rconn), msg->size, oh->type,
217 (p->handler)(sw, rconn, msg->data);
222 if (VLOG_IS_DBG_ENABLED()) {
223 char *p = ofp_to_string(msg->data, msg->size, 2);
224 VLOG_DBG_RL(&rl, "%016llx: OpenFlow packet ignored: %s",
231 send_features_request(struct lswitch *sw, struct rconn *rconn)
233 time_t now = time_now();
234 if (now >= sw->last_features_request + 1) {
236 struct ofp_switch_config *osc;
238 /* Send OFPT_FEATURES_REQUEST. */
239 make_openflow(sizeof(struct ofp_header), OFPT_FEATURES_REQUEST, &b);
240 queue_tx(sw, rconn, b);
242 /* Send OFPT_SET_CONFIG. */
243 osc = make_openflow(sizeof *osc, OFPT_SET_CONFIG, &b);
244 osc->miss_send_len = htons(OFP_DEFAULT_MISS_SEND_LEN);
245 queue_tx(sw, rconn, b);
247 sw->last_features_request = now;
252 send_default_flows(struct lswitch *sw, struct rconn *rconn,
257 while (fgets(line, sizeof line, default_flows)) {
259 struct ofp_flow_mod *ofm;
260 uint16_t priority, idle_timeout, hard_timeout;
262 struct ofp_match match;
266 /* Delete comments. */
267 comment = strchr(line, '#');
272 /* Drop empty lines. */
273 if (line[strspn(line, " \t\n")] == '\0') {
277 /* Parse and send. str_to_flow() will expand and reallocate the data
278 * in 'buffer', so we can't keep pointers to across the str_to_flow()
280 make_openflow(sizeof *ofm, OFPT_FLOW_MOD, &b);
281 parse_ofp_str(line, &match, b,
282 NULL, NULL, &priority, &idle_timeout, &hard_timeout,
286 ofm->command = htons(OFPFC_ADD);
287 ofm->cookie = htonll(cookie);
288 ofm->idle_timeout = htons(idle_timeout);
289 ofm->hard_timeout = htons(hard_timeout);
290 ofm->buffer_id = htonl(UINT32_MAX);
291 ofm->priority = htons(priority);
293 update_openflow_length(b);
294 queue_tx(sw, rconn, b);
299 queue_tx(struct lswitch *sw, struct rconn *rconn, struct ofpbuf *b)
301 int retval = rconn_send_with_limit(rconn, b, sw->queued, 10);
302 if (retval && retval != ENOTCONN) {
303 if (retval == EAGAIN) {
304 VLOG_INFO_RL(&rl, "%016llx: %s: tx queue overflow",
305 sw->datapath_id, rconn_get_name(rconn));
307 VLOG_WARN_RL(&rl, "%016llx: %s: send: %s",
308 sw->datapath_id, rconn_get_name(rconn),
315 process_switch_features(struct lswitch *sw, struct rconn *rconn OVS_UNUSED,
318 struct ofp_switch_features *osf = osf_;
320 sw->datapath_id = ntohll(osf->datapath_id);
324 lswitch_choose_destination(struct lswitch *sw, const flow_t *flow)
328 /* Learn the source MAC. */
330 if (mac_learning_learn(sw->ml, flow->dl_src, 0, flow->in_port,
331 GRAT_ARP_LOCK_NONE)) {
332 VLOG_DBG_RL(&rl, "%016llx: learned that "ETH_ADDR_FMT" is on "
333 "port %"PRIu16, sw->datapath_id,
334 ETH_ADDR_ARGS(flow->dl_src), flow->in_port);
338 /* Drop frames for reserved multicast addresses. */
339 if (eth_addr_is_reserved(flow->dl_dst)) {
343 out_port = OFPP_FLOOD;
345 int learned_port = mac_learning_lookup(sw->ml, flow->dl_dst, 0, NULL);
346 if (learned_port >= 0) {
347 out_port = learned_port;
348 if (out_port == flow->in_port) {
349 /* Don't send a packet back out its input port. */
355 /* Check if we need to use "NORMAL" action. */
356 if (sw->action_normal && out_port != OFPP_FLOOD) {
364 process_packet_in(struct lswitch *sw, struct rconn *rconn, void *opi_)
366 struct ofp_packet_in *opi = opi_;
367 uint16_t in_port = ntohs(opi->in_port);
370 struct ofp_action_header actions[2];
373 size_t pkt_ofs, pkt_len;
377 /* Ignore packets sent via output to OFPP_CONTROLLER. This library never
378 * uses such an action. You never know what experiments might be going on,
379 * though, and it seems best not to interfere with them. */
380 if (opi->reason != OFPR_NO_MATCH) {
384 /* Extract flow data from 'opi' into 'flow'. */
385 pkt_ofs = offsetof(struct ofp_packet_in, data);
386 pkt_len = ntohs(opi->header.length) - pkt_ofs;
387 pkt.data = opi->data;
389 flow_extract(&pkt, 0, in_port, &flow);
391 /* Choose output port. */
392 out_port = lswitch_choose_destination(sw, &flow);
395 if (out_port == OFPP_NONE) {
397 } else if (sw->queue == UINT32_MAX || out_port >= OFPP_MAX) {
398 struct ofp_action_output oao;
400 memset(&oao, 0, sizeof oao);
401 oao.type = htons(OFPAT_OUTPUT);
402 oao.len = htons(sizeof oao);
403 oao.port = htons(out_port);
405 memcpy(actions, &oao, sizeof oao);
406 actions_len = sizeof oao;
408 struct ofp_action_enqueue oae;
410 memset(&oae, 0, sizeof oae);
411 oae.type = htons(OFPAT_ENQUEUE);
412 oae.len = htons(sizeof oae);
413 oae.port = htons(out_port);
414 oae.queue_id = htonl(sw->queue);
416 memcpy(actions, &oae, sizeof oae);
417 actions_len = sizeof oae;
419 assert(actions_len <= sizeof actions);
421 /* Send the packet, and possibly the whole flow, to the output port. */
422 if (sw->max_idle >= 0 && (!sw->ml || out_port != OFPP_FLOOD)) {
423 struct ofpbuf *buffer;
424 struct ofp_flow_mod *ofm;
426 /* The output port is known, or we always flood everything, so add a
428 buffer = make_add_flow(&flow, ntohl(opi->buffer_id),
429 sw->max_idle, actions_len);
430 ofpbuf_put(buffer, actions, actions_len);
432 ofm->match.wildcards = htonl(sw->wildcards);
433 queue_tx(sw, rconn, buffer);
435 /* If the switch didn't buffer the packet, we need to send a copy. */
436 if (ntohl(opi->buffer_id) == UINT32_MAX && actions_len > 0) {
438 make_packet_out(&pkt, UINT32_MAX, in_port,
439 actions, actions_len / sizeof *actions));
442 /* We don't know that MAC, or we don't set up flows. Send along the
443 * packet without setting up a flow. */
444 if (ntohl(opi->buffer_id) != UINT32_MAX || actions_len > 0) {
446 make_packet_out(&pkt, ntohl(opi->buffer_id), in_port,
447 actions, actions_len / sizeof *actions));
453 process_echo_request(struct lswitch *sw, struct rconn *rconn, void *rq_)
455 struct ofp_header *rq = rq_;
456 queue_tx(sw, rconn, make_echo_reply(rq));