lib/learning-switch.c

   1 /*
   2  * Copyright (c) 2008, 2009, 2010 Nicira Networks.
   3  *
   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:
   7  *
   8  *     http://www.apache.org/licenses/LICENSE-2.0
   9  *
  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.
  15  */
  16
  17 #include <config.h>
  18 #include "learning-switch.h"
  19
  20 #include <errno.h>
  21 #include <inttypes.h>
  22 #include <netinet/in.h>
  23 #include <stdlib.h>
  24 #include <time.h>
  25
  26 #include "byte-order.h"
  27 #include "flow.h"
  28 #include "hmap.h"
  29 #include "mac-learning.h"
  30 #include "ofpbuf.h"
  31 #include "ofp-parse.h"
  32 #include "ofp-print.h"
  33 #include "ofp-util.h"
  34 #include "openflow/openflow.h"
  35 #include "poll-loop.h"
  36 #include "queue.h"
  37 #include "rconn.h"
  38 #include "shash.h"
  39 #include "timeval.h"
  40 #include "vconn.h"
  41 #include "vlog.h"
  42
  43 VLOG_DEFINE_THIS_MODULE(learning_switch)
  44
  45 struct lswitch_port {
  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. */
  49 };
  50
  51 struct lswitch {
  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. */
  55     int max_idle;
  56
  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     uint32_t wildcards;         /* Wildcards to apply to flows. */
  61     bool action_normal;         /* Use OFPP_NORMAL? */
  62
  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. */
  67
  68     /* Number of outgoing queued packets on the rconn. */
  69     struct rconn_packet_counter *queued;
  70 };
  71
  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);
  75
  76 static void queue_tx(struct lswitch *, struct rconn *, struct ofpbuf *);
  77 static void send_features_request(struct lswitch *, struct rconn *);
  78
  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;
  83
  84 /* Creates and returns a new learning switch whose configuration is given by
  85  * 'cfg'.
  86  *
  87  * 'rconn' is used to send out an OpenFlow features request. */
  88 struct lswitch *
  89 lswitch_create(struct rconn *rconn, const struct lswitch_config *cfg)
  90 {
  91     const struct ofpbuf *b;
  92     struct lswitch *sw;
  93
  94     sw = xzalloc(sizeof *sw);
  95     sw->max_idle = cfg->max_idle;
  96     sw->datapath_id = 0;
  97     sw->last_features_request = time_now() - 1;
  98     sw->ml = cfg->mode == LSW_LEARN ? mac_learning_create() : NULL;
  99     sw->action_normal = cfg->mode == LSW_NORMAL;
 100     if (cfg->exact_flows) {
 101         /* Exact match. */
 102         sw->wildcards = 0;
 103     } else {
 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->wildcards = (OFPFW_DL_TYPE | OFPFW_NW_SRC_MASK | OFPFW_NW_DST_MASK
 108                          | OFPFW_NW_PROTO | OFPFW_TP_SRC | OFPFW_TP_DST);
 109     }
 110
 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;
 116
 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);
 122         }
 123     }
 124
 125     sw->queued = rconn_packet_counter_create();
 126     send_features_request(sw, rconn);
 127
 128     for (b = cfg->default_flows; b; b = b->next) {
 129         queue_tx(sw, rconn, ofpbuf_clone(b));
 130     }
 131
 132     return sw;
 133 }
 134
 135 /* Destroys 'sw'. */
 136 void
 137 lswitch_destroy(struct lswitch *sw)
 138 {
 139     if (sw) {
 140         struct lswitch_port *node, *next;
 141
 142         HMAP_FOR_EACH_SAFE (node, next, hmap_node, &sw->queue_numbers) {
 143             hmap_remove(&sw->queue_numbers, &node->hmap_node);
 144             free(node);
 145         }
 146         shash_destroy(&sw->queue_names);
 147         mac_learning_destroy(sw->ml);
 148         rconn_packet_counter_destroy(sw->queued);
 149         free(sw);
 150     }
 151 }
 152
 153 /* Takes care of necessary 'sw' activity, except for receiving packets (which
 154  * the caller must do). */
 155 void
 156 lswitch_run(struct lswitch *sw)
 157 {
 158     if (sw->ml) {
 159         mac_learning_run(sw->ml, NULL);
 160     }
 161 }
 162
 163 void
 164 lswitch_wait(struct lswitch *sw)
 165 {
 166     if (sw->ml) {
 167         mac_learning_wait(sw->ml);
 168     }
 169 }
 170
 171 /* Processes 'msg', which should be an OpenFlow received on 'rconn', according
 172  * to the learning switch state in 'sw'.  The most likely result of processing
 173  * is that flow-setup and packet-out OpenFlow messages will be sent out on
 174  * 'rconn'.  */
 175 void
 176 lswitch_process_packet(struct lswitch *sw, struct rconn *rconn,
 177                        const struct ofpbuf *msg)
 178 {
 179     struct processor {
 180         uint8_t type;
 181         size_t min_size;
 182         packet_handler_func *handler;
 183     };
 184     static const struct processor processors[] = {
 185         {
 186             OFPT_ECHO_REQUEST,
 187             sizeof(struct ofp_header),
 188             process_echo_request
 189         },
 190         {
 191             OFPT_FEATURES_REPLY,
 192             sizeof(struct ofp_switch_features),
 193             process_switch_features
 194         },
 195         {
 196             OFPT_PACKET_IN,
 197             offsetof(struct ofp_packet_in, data),
 198             process_packet_in
 199         },
 200         {
 201             OFPT_FLOW_REMOVED,
 202             sizeof(struct ofp_flow_removed),
 203             NULL
 204         },
 205     };
 206     const size_t n_processors = ARRAY_SIZE(processors);
 207     const struct processor *p;
 208     struct ofp_header *oh;
 209
 210     oh = msg->data;
 211     if (sw->datapath_id == 0
 212         && oh->type != OFPT_ECHO_REQUEST
 213         && oh->type != OFPT_FEATURES_REPLY) {
 214         send_features_request(sw, rconn);
 215         return;
 216     }
 217
 218     for (p = processors; p < &processors[n_processors]; p++) {
 219         if (oh->type == p->type) {
 220             if (msg->size < p->min_size) {
 221                 VLOG_WARN_RL(&rl, "%016llx: %s: too short (%zu bytes) for "
 222                              "type %"PRIu8" (min %zu)", sw->datapath_id,
 223                              rconn_get_name(rconn), msg->size, oh->type,
 224                              p->min_size);
 225                 return;
 226             }
 227             if (p->handler) {
 228                 (p->handler)(sw, rconn, msg->data);
 229             }
 230             return;
 231         }
 232     }
 233     if (VLOG_IS_DBG_ENABLED()) {
 234         char *s = ofp_to_string(msg->data, msg->size, 2);
 235         VLOG_DBG_RL(&rl, "%016llx: OpenFlow packet ignored: %s",
 236                     sw->datapath_id, s);
 237         free(s);
 238     }
 239 }
 240 \f
 241 static void
 242 send_features_request(struct lswitch *sw, struct rconn *rconn)
 243 {
 244     time_t now = time_now();
 245     if (now >= sw->last_features_request + 1) {
 246         struct ofpbuf *b;
 247         struct ofp_switch_config *osc;
 248
 249         /* Send OFPT_FEATURES_REQUEST. */
 250         make_openflow(sizeof(struct ofp_header), OFPT_FEATURES_REQUEST, &b);
 251         queue_tx(sw, rconn, b);
 252
 253         /* Send OFPT_SET_CONFIG. */
 254         osc = make_openflow(sizeof *osc, OFPT_SET_CONFIG, &b);
 255         osc->miss_send_len = htons(OFP_DEFAULT_MISS_SEND_LEN);
 256         queue_tx(sw, rconn, b);
 257
 258         sw->last_features_request = now;
 259     }
 260 }
 261
 262 static void
 263 queue_tx(struct lswitch *sw, struct rconn *rconn, struct ofpbuf *b)
 264 {
 265     int retval = rconn_send_with_limit(rconn, b, sw->queued, 10);
 266     if (retval && retval != ENOTCONN) {
 267         if (retval == EAGAIN) {
 268             VLOG_INFO_RL(&rl, "%016llx: %s: tx queue overflow",
 269                          sw->datapath_id, rconn_get_name(rconn));
 270         } else {
 271             VLOG_WARN_RL(&rl, "%016llx: %s: send: %s",
 272                          sw->datapath_id, rconn_get_name(rconn),
 273                          strerror(retval));
 274         }
 275     }
 276 }
 277
 278 static void
 279 process_switch_features(struct lswitch *sw, struct rconn *rconn OVS_UNUSED,
 280                         void *osf_)
 281 {
 282     struct ofp_switch_features *osf = osf_;
 283     size_t n_ports;
 284     size_t i;
 285
 286     if (check_ofp_message_array(&osf->header, OFPT_FEATURES_REPLY,
 287                                 sizeof *osf, sizeof *osf->ports, &n_ports)) {
 288         return;
 289     }
 290
 291     sw->datapath_id = ntohll(osf->datapath_id);
 292
 293     for (i = 0; i < n_ports; i++) {
 294         struct ofp_phy_port *opp = &osf->ports[i];
 295         struct lswitch_port *lp;
 296
 297         opp->name[OFP_MAX_PORT_NAME_LEN - 1] = '\0';
 298         lp = shash_find_data(&sw->queue_names, (char *) opp->name);
 299         if (lp && hmap_node_is_null(&lp->hmap_node)) {
 300             lp->port_no = ntohs(opp->port_no);
 301             hmap_insert(&sw->queue_numbers, &lp->hmap_node,
 302                         hash_int(lp->port_no, 0));
 303         }
 304     }
 305 }
 306
 307 static uint16_t
 308 lswitch_choose_destination(struct lswitch *sw, const struct flow *flow)
 309 {
 310     uint16_t out_port;
 311
 312     /* Learn the source MAC. */
 313     if (sw->ml) {
 314         if (mac_learning_learn(sw->ml, flow->dl_src, 0, flow->in_port,
 315                                GRAT_ARP_LOCK_NONE)) {
 316             VLOG_DBG_RL(&rl, "%016llx: learned that "ETH_ADDR_FMT" is on "
 317                         "port %"PRIu16, sw->datapath_id,
 318                         ETH_ADDR_ARGS(flow->dl_src), flow->in_port);
 319         }
 320     }
 321
 322     /* Drop frames for reserved multicast addresses. */
 323     if (eth_addr_is_reserved(flow->dl_dst)) {
 324         return OFPP_NONE;
 325     }
 326
 327     out_port = OFPP_FLOOD;
 328     if (sw->ml) {
 329         int learned_port = mac_learning_lookup(sw->ml, flow->dl_dst, 0, NULL);
 330         if (learned_port >= 0) {
 331             out_port = learned_port;
 332             if (out_port == flow->in_port) {
 333                 /* Don't send a packet back out its input port. */
 334                 return OFPP_NONE;
 335             }
 336         }
 337     }
 338
 339     /* Check if we need to use "NORMAL" action. */
 340     if (sw->action_normal && out_port != OFPP_FLOOD) {
 341         return OFPP_NORMAL;
 342     }
 343
 344     return out_port;
 345 }
 346
 347 static uint32_t
 348 get_queue_id(const struct lswitch *sw, uint16_t in_port)
 349 {
 350     const struct lswitch_port *port;
 351
 352     HMAP_FOR_EACH_WITH_HASH (port, hmap_node, hash_int(in_port, 0),
 353                              &sw->queue_numbers) {
 354         if (port->port_no == in_port) {
 355             return port->queue_id;
 356         }
 357     }
 358
 359     return sw->default_queue;
 360 }
 361
 362 static void
 363 process_packet_in(struct lswitch *sw, struct rconn *rconn, void *opi_)
 364 {
 365     struct ofp_packet_in *opi = opi_;
 366     uint16_t in_port = ntohs(opi->in_port);
 367     uint32_t queue_id;
 368     uint16_t out_port;
 369
 370     struct ofp_action_header actions[2];
 371     size_t actions_len;
 372
 373     size_t pkt_ofs, pkt_len;
 374     struct ofpbuf pkt;
 375     struct flow flow;
 376
 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) {
 381         return;
 382     }
 383
 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;
 388     pkt.size = pkt_len;
 389     flow_extract(&pkt, 0, in_port, &flow);
 390
 391     /* Choose output port. */
 392     out_port = lswitch_choose_destination(sw, &flow);
 393
 394     /* Make actions. */
 395     queue_id = get_queue_id(sw, in_port);
 396     if (out_port == OFPP_NONE) {
 397         actions_len = 0;
 398     } else if (queue_id == UINT32_MAX || out_port >= OFPP_MAX) {
 399         struct ofp_action_output oao;
 400
 401         memset(&oao, 0, sizeof oao);
 402         oao.type = htons(OFPAT_OUTPUT);
 403         oao.len = htons(sizeof oao);
 404         oao.port = htons(out_port);
 405
 406         memcpy(actions, &oao, sizeof oao);
 407         actions_len = sizeof oao;
 408     } else {
 409         struct ofp_action_enqueue oae;
 410
 411         memset(&oae, 0, sizeof oae);
 412         oae.type = htons(OFPAT_ENQUEUE);
 413         oae.len = htons(sizeof oae);
 414         oae.port = htons(out_port);
 415         oae.queue_id = htonl(queue_id);
 416
 417         memcpy(actions, &oae, sizeof oae);
 418         actions_len = sizeof oae;
 419     }
 420     assert(actions_len <= sizeof actions);
 421
 422     /* Send the packet, and possibly the whole flow, to the output port. */
 423     if (sw->max_idle >= 0 && (!sw->ml || out_port != OFPP_FLOOD)) {
 424         struct ofpbuf *buffer;
 425         struct ofp_flow_mod *ofm;
 426
 427         /* The output port is known, or we always flood everything, so add a
 428          * new flow. */
 429         buffer = make_add_flow(&flow, ntohl(opi->buffer_id),
 430                                sw->max_idle, actions_len);
 431         ofpbuf_put(buffer, actions, actions_len);
 432         ofm = buffer->data;
 433         ofm->match.wildcards = htonl(sw->wildcards);
 434         queue_tx(sw, rconn, buffer);
 435
 436         /* If the switch didn't buffer the packet, we need to send a copy. */
 437         if (ntohl(opi->buffer_id) == UINT32_MAX && actions_len > 0) {
 438             queue_tx(sw, rconn,
 439                      make_packet_out(&pkt, UINT32_MAX, in_port,
 440                                      actions, actions_len / sizeof *actions));
 441         }
 442     } else {
 443         /* We don't know that MAC, or we don't set up flows.  Send along the
 444          * packet without setting up a flow. */
 445         if (ntohl(opi->buffer_id) != UINT32_MAX || actions_len > 0) {
 446             queue_tx(sw, rconn,
 447                      make_packet_out(&pkt, ntohl(opi->buffer_id), in_port,
 448                                      actions, actions_len / sizeof *actions));
 449         }
 450     }
 451 }
 452
 453 static void
 454 process_echo_request(struct lswitch *sw, struct rconn *rconn, void *rq_)
 455 {
 456     struct ofp_header *rq = rq_;
 457     queue_tx(sw, rconn, make_echo_reply(rq));
 458 }