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"
43 VLOG_DEFINE_THIS_MODULE(learning_switch);
46 struct hmap_node hmap_node; /* Hash node for port number. */
47 uint16_t port_no; /* OpenFlow port number, in host byte order. */
48 uint32_t queue_id; /* OpenFlow queue number. */
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 time_t last_features_request;
59 struct mac_learning *ml; /* NULL to act as hub instead of switch. */
60 struct flow_wildcards wc; /* Wildcards to apply to flows. */
61 bool action_normal; /* Use OFPP_NORMAL? */
63 /* Queue distribution. */
64 uint32_t default_queue; /* Default OpenFlow queue, or UINT32_MAX. */
65 struct hmap queue_numbers; /* Map from port number to lswitch_port. */
66 struct shash queue_names; /* Map from port name to lswitch_port. */
68 /* Number of outgoing queued packets on the rconn. */
69 struct rconn_packet_counter *queued;
72 /* The log messages here could actually be useful in debugging, so keep the
73 * rate limit relatively high. */
74 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(30, 300);
76 static void queue_tx(struct lswitch *, struct rconn *, struct ofpbuf *);
77 static void send_features_request(struct lswitch *, struct rconn *);
79 typedef void packet_handler_func(struct lswitch *, struct rconn *, void *);
80 static packet_handler_func process_switch_features;
81 static packet_handler_func process_packet_in;
82 static packet_handler_func process_echo_request;
84 /* Creates and returns a new learning switch whose configuration is given by
87 * 'rconn' is used to send out an OpenFlow features request. */
89 lswitch_create(struct rconn *rconn, const struct lswitch_config *cfg)
93 sw = xzalloc(sizeof *sw);
94 sw->max_idle = cfg->max_idle;
96 sw->last_features_request = time_now() - 1;
97 sw->ml = cfg->mode == LSW_LEARN ? mac_learning_create() : NULL;
98 sw->action_normal = cfg->mode == LSW_NORMAL;
100 flow_wildcards_init_exact(&sw->wc);
101 if (!cfg->exact_flows) {
102 /* We cannot wildcard all fields.
103 * We need in_port to detect moves.
104 * We need both SA and DA to do learning. */
105 sw->wc.wildcards = (FWW_DL_TYPE | FWW_NW_PROTO
106 | FWW_TP_SRC | FWW_TP_DST);
107 sw->wc.nw_src_mask = htonl(0);
108 sw->wc.nw_dst_mask = htonl(0);
111 sw->default_queue = cfg->default_queue;
112 hmap_init(&sw->queue_numbers);
113 shash_init(&sw->queue_names);
114 if (cfg->port_queues) {
115 struct shash_node *node;
117 SHASH_FOR_EACH (node, cfg->port_queues) {
118 struct lswitch_port *port = xmalloc(sizeof *port);
119 hmap_node_nullify(&port->hmap_node);
120 port->queue_id = (uintptr_t) node->data;
121 shash_add(&sw->queue_names, node->name, port);
125 sw->queued = rconn_packet_counter_create();
126 send_features_request(sw, rconn);
128 if (cfg->default_flows) {
129 const struct ofpbuf *b;
131 LIST_FOR_EACH (b, list_node, cfg->default_flows) {
132 queue_tx(sw, rconn, ofpbuf_clone(b));
141 lswitch_destroy(struct lswitch *sw)
144 struct lswitch_port *node, *next;
146 HMAP_FOR_EACH_SAFE (node, next, hmap_node, &sw->queue_numbers) {
147 hmap_remove(&sw->queue_numbers, &node->hmap_node);
150 shash_destroy(&sw->queue_names);
151 mac_learning_destroy(sw->ml);
152 rconn_packet_counter_destroy(sw->queued);
157 /* Takes care of necessary 'sw' activity, except for receiving packets (which
158 * the caller must do). */
160 lswitch_run(struct lswitch *sw)
163 mac_learning_run(sw->ml, NULL);
168 lswitch_wait(struct lswitch *sw)
171 mac_learning_wait(sw->ml);
175 /* Processes 'msg', which should be an OpenFlow received on 'rconn', according
176 * to the learning switch state in 'sw'. The most likely result of processing
177 * is that flow-setup and packet-out OpenFlow messages will be sent out on
180 lswitch_process_packet(struct lswitch *sw, struct rconn *rconn,
181 const struct ofpbuf *msg)
186 packet_handler_func *handler;
188 static const struct processor processors[] = {
191 sizeof(struct ofp_header),
196 sizeof(struct ofp_switch_features),
197 process_switch_features
201 offsetof(struct ofp_packet_in, data),
206 sizeof(struct ofp_flow_removed),
210 const size_t n_processors = ARRAY_SIZE(processors);
211 const struct processor *p;
212 struct ofp_header *oh;
215 if (sw->datapath_id == 0
216 && oh->type != OFPT_ECHO_REQUEST
217 && oh->type != OFPT_FEATURES_REPLY) {
218 send_features_request(sw, rconn);
222 for (p = processors; p < &processors[n_processors]; p++) {
223 if (oh->type == p->type) {
224 if (msg->size < p->min_size) {
225 VLOG_WARN_RL(&rl, "%016llx: %s: too short (%zu bytes) for "
226 "type %"PRIu8" (min %zu)", sw->datapath_id,
227 rconn_get_name(rconn), msg->size, oh->type,
232 (p->handler)(sw, rconn, msg->data);
237 if (VLOG_IS_DBG_ENABLED()) {
238 char *s = ofp_to_string(msg->data, msg->size, 2);
239 VLOG_DBG_RL(&rl, "%016llx: OpenFlow packet ignored: %s",
246 send_features_request(struct lswitch *sw, struct rconn *rconn)
248 time_t now = time_now();
249 if (now >= sw->last_features_request + 1) {
251 struct ofp_switch_config *osc;
253 /* Send OFPT_FEATURES_REQUEST. */
254 make_openflow(sizeof(struct ofp_header), OFPT_FEATURES_REQUEST, &b);
255 queue_tx(sw, rconn, b);
257 /* Send OFPT_SET_CONFIG. */
258 osc = make_openflow(sizeof *osc, OFPT_SET_CONFIG, &b);
259 osc->miss_send_len = htons(OFP_DEFAULT_MISS_SEND_LEN);
260 queue_tx(sw, rconn, b);
262 sw->last_features_request = now;
267 queue_tx(struct lswitch *sw, struct rconn *rconn, struct ofpbuf *b)
269 int retval = rconn_send_with_limit(rconn, b, sw->queued, 10);
270 if (retval && retval != ENOTCONN) {
271 if (retval == EAGAIN) {
272 VLOG_INFO_RL(&rl, "%016llx: %s: tx queue overflow",
273 sw->datapath_id, rconn_get_name(rconn));
275 VLOG_WARN_RL(&rl, "%016llx: %s: send: %s",
276 sw->datapath_id, rconn_get_name(rconn),
283 process_switch_features(struct lswitch *sw, struct rconn *rconn OVS_UNUSED,
286 struct ofp_switch_features *osf = osf_;
290 if (check_ofp_message_array(&osf->header, OFPT_FEATURES_REPLY,
291 sizeof *osf, sizeof *osf->ports, &n_ports)) {
295 sw->datapath_id = ntohll(osf->datapath_id);
297 for (i = 0; i < n_ports; i++) {
298 struct ofp_phy_port *opp = &osf->ports[i];
299 struct lswitch_port *lp;
301 opp->name[OFP_MAX_PORT_NAME_LEN - 1] = '\0';
302 lp = shash_find_data(&sw->queue_names, opp->name);
303 if (lp && hmap_node_is_null(&lp->hmap_node)) {
304 lp->port_no = ntohs(opp->port_no);
305 hmap_insert(&sw->queue_numbers, &lp->hmap_node,
306 hash_int(lp->port_no, 0));
312 lswitch_choose_destination(struct lswitch *sw, const struct flow *flow)
316 /* Learn the source MAC. */
318 if (mac_learning_learn(sw->ml, flow->dl_src, 0, flow->in_port,
319 GRAT_ARP_LOCK_NONE)) {
320 VLOG_DBG_RL(&rl, "%016llx: learned that "ETH_ADDR_FMT" is on "
321 "port %"PRIu16, sw->datapath_id,
322 ETH_ADDR_ARGS(flow->dl_src), flow->in_port);
326 /* Drop frames for reserved multicast addresses. */
327 if (eth_addr_is_reserved(flow->dl_dst)) {
331 out_port = OFPP_FLOOD;
333 int learned_port = mac_learning_lookup(sw->ml, flow->dl_dst, 0, NULL);
334 if (learned_port >= 0) {
335 out_port = learned_port;
336 if (out_port == flow->in_port) {
337 /* Don't send a packet back out its input port. */
343 /* Check if we need to use "NORMAL" action. */
344 if (sw->action_normal && out_port != OFPP_FLOOD) {
352 get_queue_id(const struct lswitch *sw, uint16_t in_port)
354 const struct lswitch_port *port;
356 HMAP_FOR_EACH_WITH_HASH (port, hmap_node, hash_int(in_port, 0),
357 &sw->queue_numbers) {
358 if (port->port_no == in_port) {
359 return port->queue_id;
363 return sw->default_queue;
367 process_packet_in(struct lswitch *sw, struct rconn *rconn, void *opi_)
369 struct ofp_packet_in *opi = opi_;
370 uint16_t in_port = ntohs(opi->in_port);
374 struct ofp_action_header actions[2];
377 size_t pkt_ofs, pkt_len;
381 /* Ignore packets sent via output to OFPP_CONTROLLER. This library never
382 * uses such an action. You never know what experiments might be going on,
383 * though, and it seems best not to interfere with them. */
384 if (opi->reason != OFPR_NO_MATCH) {
388 /* Extract flow data from 'opi' into 'flow'. */
389 pkt_ofs = offsetof(struct ofp_packet_in, data);
390 pkt_len = ntohs(opi->header.length) - pkt_ofs;
391 pkt.data = opi->data;
393 flow_extract(&pkt, 0, in_port, &flow);
395 /* Choose output port. */
396 out_port = lswitch_choose_destination(sw, &flow);
399 queue_id = get_queue_id(sw, in_port);
400 if (out_port == OFPP_NONE) {
402 } else if (queue_id == UINT32_MAX || out_port >= OFPP_MAX) {
403 struct ofp_action_output oao;
405 memset(&oao, 0, sizeof oao);
406 oao.type = htons(OFPAT_OUTPUT);
407 oao.len = htons(sizeof oao);
408 oao.port = htons(out_port);
410 memcpy(actions, &oao, sizeof oao);
411 actions_len = sizeof oao;
413 struct ofp_action_enqueue oae;
415 memset(&oae, 0, sizeof oae);
416 oae.type = htons(OFPAT_ENQUEUE);
417 oae.len = htons(sizeof oae);
418 oae.port = htons(out_port);
419 oae.queue_id = htonl(queue_id);
421 memcpy(actions, &oae, sizeof oae);
422 actions_len = sizeof oae;
424 assert(actions_len <= sizeof actions);
426 /* Send the packet, and possibly the whole flow, to the output port. */
427 if (sw->max_idle >= 0 && (!sw->ml || out_port != OFPP_FLOOD)) {
428 struct ofpbuf *buffer;
429 struct ofp_flow_mod *ofm;
430 struct cls_rule rule;
432 /* The output port is known, or we always flood everything, so add a
434 cls_rule_init(&flow, &sw->wc, 0, &rule);
435 buffer = make_add_flow(&rule, ntohl(opi->buffer_id),
436 sw->max_idle, actions_len);
437 ofpbuf_put(buffer, actions, actions_len);
439 queue_tx(sw, rconn, buffer);
441 /* If the switch didn't buffer the packet, we need to send a copy. */
442 if (ntohl(opi->buffer_id) == UINT32_MAX && actions_len > 0) {
444 make_packet_out(&pkt, UINT32_MAX, in_port,
445 actions, actions_len / sizeof *actions));
448 /* We don't know that MAC, or we don't set up flows. Send along the
449 * packet without setting up a flow. */
450 if (ntohl(opi->buffer_id) != UINT32_MAX || actions_len > 0) {
452 make_packet_out(&pkt, ntohl(opi->buffer_id), in_port,
453 actions, actions_len / sizeof *actions));
459 process_echo_request(struct lswitch *sw, struct rconn *rconn, void *rq_)
461 struct ofp_header *rq = rq_;
462 queue_tx(sw, rconn, make_echo_reply(rq));