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>
26 #include "byte-order.h"
27 #include "classifier.h"
30 #include "mac-learning.h"
32 #include "ofp-parse.h"
33 #include "ofp-print.h"
35 #include "openflow/openflow.h"
36 #include "poll-loop.h"
44 VLOG_DEFINE_THIS_MODULE(learning_switch);
47 struct hmap_node hmap_node; /* Hash node for port number. */
48 uint16_t port_no; /* OpenFlow port number, in host byte order. */
49 uint32_t queue_id; /* OpenFlow queue number. */
53 /* If nonnegative, the switch sets up flows that expire after the given
54 * number of seconds (or never expire, if the value is OFP_FLOW_PERMANENT).
55 * Otherwise, the switch processes every packet. */
58 unsigned long long int datapath_id;
59 time_t last_features_request;
60 struct mac_learning *ml; /* NULL to act as hub instead of switch. */
61 struct flow_wildcards wc; /* Wildcards to apply to flows. */
62 bool action_normal; /* Use OFPP_NORMAL? */
64 /* Queue distribution. */
65 uint32_t default_queue; /* Default OpenFlow queue, or UINT32_MAX. */
66 struct hmap queue_numbers; /* Map from port number to lswitch_port. */
67 struct shash queue_names; /* Map from port name to lswitch_port. */
69 /* Number of outgoing queued packets on the rconn. */
70 struct rconn_packet_counter *queued;
73 /* The log messages here could actually be useful in debugging, so keep the
74 * rate limit relatively high. */
75 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(30, 300);
77 static void queue_tx(struct lswitch *, struct rconn *, struct ofpbuf *);
78 static void send_features_request(struct lswitch *, struct rconn *);
80 typedef void packet_handler_func(struct lswitch *, struct rconn *, void *);
81 static packet_handler_func process_switch_features;
82 static packet_handler_func process_packet_in;
83 static packet_handler_func process_echo_request;
85 /* Creates and returns a new learning switch whose configuration is given by
88 * 'rconn' is used to send out an OpenFlow features request. */
90 lswitch_create(struct rconn *rconn, const struct lswitch_config *cfg)
92 const struct ofpbuf *b;
95 sw = xzalloc(sizeof *sw);
96 sw->max_idle = cfg->max_idle;
98 sw->last_features_request = time_now() - 1;
99 sw->ml = cfg->mode == LSW_LEARN ? mac_learning_create() : NULL;
100 sw->action_normal = cfg->mode == LSW_NORMAL;
102 flow_wildcards_init_exact(&sw->wc);
103 if (!cfg->exact_flows) {
104 /* We cannot wildcard all fields.
105 * We need in_port to detect moves.
106 * We need both SA and DA to do learning. */
107 sw->wc.wildcards = (FWW_DL_TYPE | FWW_NW_PROTO
108 | FWW_TP_SRC | FWW_TP_DST);
109 sw->wc.nw_src_mask = htonl(0);
110 sw->wc.nw_dst_mask = htonl(0);
113 sw->default_queue = cfg->default_queue;
114 hmap_init(&sw->queue_numbers);
115 shash_init(&sw->queue_names);
116 if (cfg->port_queues) {
117 struct shash_node *node;
119 SHASH_FOR_EACH (node, cfg->port_queues) {
120 struct lswitch_port *port = xmalloc(sizeof *port);
121 hmap_node_nullify(&port->hmap_node);
122 port->queue_id = (uintptr_t) node->data;
123 shash_add(&sw->queue_names, node->name, port);
127 sw->queued = rconn_packet_counter_create();
128 send_features_request(sw, rconn);
130 for (b = cfg->default_flows; b; b = b->next) {
131 queue_tx(sw, rconn, ofpbuf_clone(b));
139 lswitch_destroy(struct lswitch *sw)
142 struct lswitch_port *node, *next;
144 HMAP_FOR_EACH_SAFE (node, next, hmap_node, &sw->queue_numbers) {
145 hmap_remove(&sw->queue_numbers, &node->hmap_node);
148 shash_destroy(&sw->queue_names);
149 mac_learning_destroy(sw->ml);
150 rconn_packet_counter_destroy(sw->queued);
155 /* Takes care of necessary 'sw' activity, except for receiving packets (which
156 * the caller must do). */
158 lswitch_run(struct lswitch *sw)
161 mac_learning_run(sw->ml, NULL);
166 lswitch_wait(struct lswitch *sw)
169 mac_learning_wait(sw->ml);
173 /* Processes 'msg', which should be an OpenFlow received on 'rconn', according
174 * to the learning switch state in 'sw'. The most likely result of processing
175 * is that flow-setup and packet-out OpenFlow messages will be sent out on
178 lswitch_process_packet(struct lswitch *sw, struct rconn *rconn,
179 const struct ofpbuf *msg)
184 packet_handler_func *handler;
186 static const struct processor processors[] = {
189 sizeof(struct ofp_header),
194 sizeof(struct ofp_switch_features),
195 process_switch_features
199 offsetof(struct ofp_packet_in, data),
204 sizeof(struct ofp_flow_removed),
208 const size_t n_processors = ARRAY_SIZE(processors);
209 const struct processor *p;
210 struct ofp_header *oh;
213 if (sw->datapath_id == 0
214 && oh->type != OFPT_ECHO_REQUEST
215 && oh->type != OFPT_FEATURES_REPLY) {
216 send_features_request(sw, rconn);
220 for (p = processors; p < &processors[n_processors]; p++) {
221 if (oh->type == p->type) {
222 if (msg->size < p->min_size) {
223 VLOG_WARN_RL(&rl, "%016llx: %s: too short (%zu bytes) for "
224 "type %"PRIu8" (min %zu)", sw->datapath_id,
225 rconn_get_name(rconn), msg->size, oh->type,
230 (p->handler)(sw, rconn, msg->data);
235 if (VLOG_IS_DBG_ENABLED()) {
236 char *s = ofp_to_string(msg->data, msg->size, 2);
237 VLOG_DBG_RL(&rl, "%016llx: OpenFlow packet ignored: %s",
244 send_features_request(struct lswitch *sw, struct rconn *rconn)
246 time_t now = time_now();
247 if (now >= sw->last_features_request + 1) {
249 struct ofp_switch_config *osc;
251 /* Send OFPT_FEATURES_REQUEST. */
252 make_openflow(sizeof(struct ofp_header), OFPT_FEATURES_REQUEST, &b);
253 queue_tx(sw, rconn, b);
255 /* Send OFPT_SET_CONFIG. */
256 osc = make_openflow(sizeof *osc, OFPT_SET_CONFIG, &b);
257 osc->miss_send_len = htons(OFP_DEFAULT_MISS_SEND_LEN);
258 queue_tx(sw, rconn, b);
260 sw->last_features_request = now;
265 queue_tx(struct lswitch *sw, struct rconn *rconn, struct ofpbuf *b)
267 int retval = rconn_send_with_limit(rconn, b, sw->queued, 10);
268 if (retval && retval != ENOTCONN) {
269 if (retval == EAGAIN) {
270 VLOG_INFO_RL(&rl, "%016llx: %s: tx queue overflow",
271 sw->datapath_id, rconn_get_name(rconn));
273 VLOG_WARN_RL(&rl, "%016llx: %s: send: %s",
274 sw->datapath_id, rconn_get_name(rconn),
281 process_switch_features(struct lswitch *sw, struct rconn *rconn OVS_UNUSED,
284 struct ofp_switch_features *osf = osf_;
288 if (check_ofp_message_array(&osf->header, OFPT_FEATURES_REPLY,
289 sizeof *osf, sizeof *osf->ports, &n_ports)) {
293 sw->datapath_id = ntohll(osf->datapath_id);
295 for (i = 0; i < n_ports; i++) {
296 struct ofp_phy_port *opp = &osf->ports[i];
297 struct lswitch_port *lp;
299 opp->name[OFP_MAX_PORT_NAME_LEN - 1] = '\0';
300 lp = shash_find_data(&sw->queue_names, (char *) opp->name);
301 if (lp && hmap_node_is_null(&lp->hmap_node)) {
302 lp->port_no = ntohs(opp->port_no);
303 hmap_insert(&sw->queue_numbers, &lp->hmap_node,
304 hash_int(lp->port_no, 0));
310 lswitch_choose_destination(struct lswitch *sw, const struct flow *flow)
314 /* Learn the source MAC. */
316 if (mac_learning_learn(sw->ml, flow->dl_src, 0, flow->in_port,
317 GRAT_ARP_LOCK_NONE)) {
318 VLOG_DBG_RL(&rl, "%016llx: learned that "ETH_ADDR_FMT" is on "
319 "port %"PRIu16, sw->datapath_id,
320 ETH_ADDR_ARGS(flow->dl_src), flow->in_port);
324 /* Drop frames for reserved multicast addresses. */
325 if (eth_addr_is_reserved(flow->dl_dst)) {
329 out_port = OFPP_FLOOD;
331 int learned_port = mac_learning_lookup(sw->ml, flow->dl_dst, 0, NULL);
332 if (learned_port >= 0) {
333 out_port = learned_port;
334 if (out_port == flow->in_port) {
335 /* Don't send a packet back out its input port. */
341 /* Check if we need to use "NORMAL" action. */
342 if (sw->action_normal && out_port != OFPP_FLOOD) {
350 get_queue_id(const struct lswitch *sw, uint16_t in_port)
352 const struct lswitch_port *port;
354 HMAP_FOR_EACH_WITH_HASH (port, hmap_node, hash_int(in_port, 0),
355 &sw->queue_numbers) {
356 if (port->port_no == in_port) {
357 return port->queue_id;
361 return sw->default_queue;
365 process_packet_in(struct lswitch *sw, struct rconn *rconn, void *opi_)
367 struct ofp_packet_in *opi = opi_;
368 uint16_t in_port = ntohs(opi->in_port);
372 struct ofp_action_header actions[2];
375 size_t pkt_ofs, pkt_len;
379 /* Ignore packets sent via output to OFPP_CONTROLLER. This library never
380 * uses such an action. You never know what experiments might be going on,
381 * though, and it seems best not to interfere with them. */
382 if (opi->reason != OFPR_NO_MATCH) {
386 /* Extract flow data from 'opi' into 'flow'. */
387 pkt_ofs = offsetof(struct ofp_packet_in, data);
388 pkt_len = ntohs(opi->header.length) - pkt_ofs;
389 pkt.data = opi->data;
391 flow_extract(&pkt, 0, in_port, &flow);
393 /* Choose output port. */
394 out_port = lswitch_choose_destination(sw, &flow);
397 queue_id = get_queue_id(sw, in_port);
398 if (out_port == OFPP_NONE) {
400 } else if (queue_id == UINT32_MAX || out_port >= OFPP_MAX) {
401 struct ofp_action_output oao;
403 memset(&oao, 0, sizeof oao);
404 oao.type = htons(OFPAT_OUTPUT);
405 oao.len = htons(sizeof oao);
406 oao.port = htons(out_port);
408 memcpy(actions, &oao, sizeof oao);
409 actions_len = sizeof oao;
411 struct ofp_action_enqueue oae;
413 memset(&oae, 0, sizeof oae);
414 oae.type = htons(OFPAT_ENQUEUE);
415 oae.len = htons(sizeof oae);
416 oae.port = htons(out_port);
417 oae.queue_id = htonl(queue_id);
419 memcpy(actions, &oae, sizeof oae);
420 actions_len = sizeof oae;
422 assert(actions_len <= sizeof actions);
424 /* Send the packet, and possibly the whole flow, to the output port. */
425 if (sw->max_idle >= 0 && (!sw->ml || out_port != OFPP_FLOOD)) {
426 struct ofpbuf *buffer;
427 struct ofp_flow_mod *ofm;
428 struct cls_rule rule;
430 /* The output port is known, or we always flood everything, so add a
432 cls_rule_init(&flow, &sw->wc, 0, &rule);
433 buffer = make_add_flow(&rule, ntohl(opi->buffer_id),
434 sw->max_idle, actions_len);
435 ofpbuf_put(buffer, actions, actions_len);
437 queue_tx(sw, rconn, buffer);
439 /* If the switch didn't buffer the packet, we need to send a copy. */
440 if (ntohl(opi->buffer_id) == UINT32_MAX && actions_len > 0) {
442 make_packet_out(&pkt, UINT32_MAX, in_port,
443 actions, actions_len / sizeof *actions));
446 /* We don't know that MAC, or we don't set up flows. Send along the
447 * packet without setting up a flow. */
448 if (ntohl(opi->buffer_id) != UINT32_MAX || actions_len > 0) {
450 make_packet_out(&pkt, ntohl(opi->buffer_id), in_port,
451 actions, actions_len / sizeof *actions));
457 process_echo_request(struct lswitch *sw, struct rconn *rconn, void *rq_)
459 struct ofp_header *rq = rq_;
460 queue_tx(sw, rconn, make_echo_reply(rq));