bgp.c

   1 /*
   2  * BGPv4
   3  * Used to advertise routes for upstream (l2tp port, rather than gratiutious
   4  * arp) and downstream--allowing routers to load-balance both.
   5  *
   6  * Implementation limitations:
   7  * - We never listen for incoming connections (session always initiated by us).
   8  * - Any routes advertised by the peer are accepted, but ignored.
   9  * - No password support; neither RFC1771 (which no-one seems to do anyway)
  10  *   nor RFC2385 (which requires a kernel patch on 2.4 kernels).
  11  */
  12
  13 char const *cvs_id_bgp = "$Id: bgp.c,v 1.12 2005/09/02 23:39:36 bodea Exp $";
  14
  15 #include <stdlib.h>
  16 #include <unistd.h>
  17 #include <string.h>
  18 #include <time.h>
  19 #include <errno.h>
  20 #include <sys/socket.h>
  21 #include <netinet/in.h>
  22 #include <arpa/inet.h>
  23 #include <netdb.h>
  24 #include <fcntl.h>
  25
  26 #include "l2tpns.h"
  27 #include "bgp.h"
  28 #include "util.h"
  29
  30 static void bgp_clear(struct bgp_peer *peer);
  31 static void bgp_set_retry(struct bgp_peer *peer);
  32 static void bgp_cidr(in_addr_t ip, in_addr_t mask, struct bgp_ip_prefix *pfx);
  33 static struct bgp_route_list *bgp_insert_route(struct bgp_route_list *head,
  34     struct bgp_route_list *new);
  35
  36 static void bgp_free_routes(struct bgp_route_list *routes);
  37 static char const *bgp_msg_type_str(uint8_t type);
  38 static int bgp_connect(struct bgp_peer *peer);
  39 static int bgp_handle_connect(struct bgp_peer *peer);
  40 static int bgp_write(struct bgp_peer *peer);
  41 static int bgp_read(struct bgp_peer *peer);
  42 static int bgp_handle_input(struct bgp_peer *peer);
  43 static int bgp_send_open(struct bgp_peer *peer);
  44 static int bgp_send_keepalive(struct bgp_peer *peer);
  45 static int bgp_send_update(struct bgp_peer *peer);
  46 static int bgp_send_notification(struct bgp_peer *peer, uint8_t code,
  47     uint8_t subcode);
  48
  49 static uint16_t our_as;
  50 static struct bgp_route_list *bgp_routes = 0;
  51
  52 int bgp_configured = 0;
  53 struct bgp_peer *bgp_peers = 0;
  54
  55 /* prepare peer structure, globals */
  56 int bgp_setup(int as)
  57 {
  58     int i;
  59     struct bgp_peer *peer;
  60
  61     for (i = 0; i < BGP_NUM_PEERS; i++)
  62     {
  63         peer = &bgp_peers[i];
  64         memset(peer, 0, sizeof(*peer));
  65
  66         peer->addr = INADDR_NONE;
  67         peer->sock = -1;
  68         peer->state = peer->next_state = Disabled;
  69
  70         if (!((peer->outbuf = malloc(sizeof(*peer->outbuf)))
  71             && (peer->inbuf = malloc(sizeof(*peer->inbuf)))))
  72         {
  73             LOG(0, 0, 0, "Can't allocate buffers for bgp peer (%s)\n",
  74                 strerror(errno));
  75
  76             return 0;
  77         }
  78
  79         peer->edata.type = FD_TYPE_BGP;
  80         peer->edata.index = i;
  81         peer->events = 0;
  82     }
  83
  84     if (as < 1)
  85         as = 0;
  86
  87     if ((our_as = as))
  88         return 0;
  89
  90     bgp_routes = 0;
  91     bgp_configured = 0; /* set by bgp_start */
  92
  93     return 1;
  94 }
  95
  96 /* start connection with a peer */
  97 int bgp_start(struct bgp_peer *peer, char *name, int as, int keepalive,
  98     int hold, struct in_addr update_source, int enable)
  99 {
 100     struct hostent *h;
 101     int ibgp;
 102     int i;
 103     struct bgp_path_attr a;
 104     char path_attrs[64];
 105     char *p = path_attrs;
 106     in_addr_t ip;
 107     uint32_t metric = htonl(BGP_METRIC);
 108     uint32_t no_export = htonl(BGP_COMMUNITY_NO_EXPORT);
 109
 110     if (!our_as)
 111         return 0;
 112
 113     if (peer->state != Disabled)
 114         bgp_halt(peer);
 115
 116     snprintf(peer->name, sizeof(peer->name), "%s", name);
 117
 118     if (!(h = gethostbyname(name)) || h->h_addrtype != AF_INET)
 119     {
 120         LOG(0, 0, 0, "Can't get address for BGP peer %s (%s)\n",
 121             name, h ? "no address" : hstrerror(h_errno));
 122
 123         return 0;
 124     }
 125
 126     memcpy(&peer->addr, h->h_addr, sizeof(peer->addr));
 127     peer->source_addr = update_source.s_addr;
 128     peer->as = as > 0 ? as : our_as;
 129     ibgp = peer->as == our_as;
 130
 131     /* set initial timer values */
 132     peer->init_keepalive = keepalive == -1 ? BGP_KEEPALIVE_TIME : keepalive;
 133     peer->init_hold = hold == -1 ? BGP_HOLD_TIME : hold;
 134
 135     if (peer->init_hold < 3)
 136         peer->init_hold = 3;
 137
 138     if (peer->init_keepalive * 3 > peer->init_hold)
 139         peer->init_keepalive = peer->init_hold / 3;
 140
 141     /* clear buffers, go to Idle state */
 142     peer->next_state = Idle;
 143     bgp_clear(peer);
 144
 145     /* set initial routing state */
 146     peer->routing = enable;
 147
 148     /* all our routes use the same attributes, so prepare it in advance */
 149     if (peer->path_attrs)
 150         free(peer->path_attrs);
 151
 152     peer->path_attr_len = 0;
 153
 154     /* ORIGIN */
 155     a.flags = BGP_PATH_ATTR_FLAG_TRANS;
 156     a.code = BGP_PATH_ATTR_CODE_ORIGIN;
 157     a.data.s.len = 1;
 158     a.data.s.value[0] = BGP_PATH_ATTR_CODE_ORIGIN_IGP;
 159
 160 #define ADD_ATTRIBUTE()         do { \
 161     i = BGP_PATH_ATTR_SIZE(a);  \
 162     memcpy(p, &a, i);           \
 163     p += i;                     \
 164     peer->path_attr_len += i;   } while (0)
 165
 166     ADD_ATTRIBUTE();
 167
 168     /* AS_PATH */
 169     a.flags = BGP_PATH_ATTR_FLAG_TRANS;
 170     a.code = BGP_PATH_ATTR_CODE_AS_PATH;
 171     if (ibgp)
 172     {
 173         /* empty path */
 174         a.data.s.len = 0;
 175     }
 176     else
 177     {
 178         /* just our AS */
 179         struct {
 180             uint8_t type;
 181             uint8_t len;
 182             uint16_t value;
 183         } as_path = {
 184             BGP_PATH_ATTR_CODE_AS_PATH_AS_SEQUENCE,
 185             1,
 186             htons(our_as),
 187         };
 188
 189         a.data.s.len = sizeof(as_path);
 190         memcpy(&a.data.s.value, &as_path, sizeof(as_path));
 191     }
 192
 193     ADD_ATTRIBUTE();
 194
 195     /* NEXT_HOP */
 196     a.flags = BGP_PATH_ATTR_FLAG_TRANS;
 197     a.code = BGP_PATH_ATTR_CODE_NEXT_HOP;
 198     if (config->nexthop_address)
 199     {
 200         ip = config->nexthop_address;
 201     }
 202     else
 203     {
 204         ip = my_address; /* we're it */
 205     }
 206     a.data.s.len = sizeof(ip);
 207     memcpy(a.data.s.value, &ip, sizeof(ip));
 208
 209     ADD_ATTRIBUTE();
 210
 211     /* MULTI_EXIT_DISC */
 212     a.flags = BGP_PATH_ATTR_FLAG_OPTIONAL;
 213     a.code = BGP_PATH_ATTR_CODE_MULTI_EXIT_DISC;
 214     a.data.s.len = sizeof(metric);
 215     memcpy(a.data.s.value, &metric, sizeof(metric));
 216
 217     ADD_ATTRIBUTE();
 218
 219     if (ibgp)
 220     {
 221         uint32_t local_pref = htonl(BGP_LOCAL_PREF);
 222
 223         /* LOCAL_PREF */
 224         a.flags = BGP_PATH_ATTR_FLAG_TRANS;
 225         a.code = BGP_PATH_ATTR_CODE_LOCAL_PREF;
 226         a.data.s.len = sizeof(local_pref);
 227         memcpy(a.data.s.value, &local_pref, sizeof(local_pref));
 228
 229         ADD_ATTRIBUTE();
 230     }
 231
 232     /* COMMUNITIES */
 233     a.flags = BGP_PATH_ATTR_FLAG_OPTIONAL | BGP_PATH_ATTR_FLAG_TRANS;
 234     a.code = BGP_PATH_ATTR_CODE_COMMUNITIES;
 235     a.data.s.len = sizeof(no_export);
 236     memcpy(a.data.s.value, &no_export, sizeof(no_export));
 237
 238     ADD_ATTRIBUTE();
 239
 240     if (!(peer->path_attrs = malloc(peer->path_attr_len)))
 241     {
 242         LOG(0, 0, 0, "Can't allocate path_attrs for %s (%s)\n",
 243             name, strerror(errno));
 244
 245         return 0;
 246     }
 247
 248     memcpy(peer->path_attrs, path_attrs, peer->path_attr_len);
 249
 250     LOG(4, 0, 0, "Initiating BGP connection to %s (routing %s)\n",
 251         name, enable ? "enabled" : "suspended");
 252
 253     /* we have at least one peer configured */
 254     bgp_configured = 1;
 255
 256     /* connect */
 257     return bgp_connect(peer);
 258 }
 259
 260 /* clear counters, timers, routes and buffers; close socket; move to
 261    next_state, which may be Disabled or Idle */
 262 static void bgp_clear(struct bgp_peer *peer)
 263 {
 264     if (peer->sock != -1)
 265     {
 266         close(peer->sock);
 267         peer->sock = -1;
 268     }
 269
 270     peer->keepalive_time = 0;
 271     peer->expire_time = 0;
 272
 273     peer->keepalive = peer->init_keepalive;
 274     peer->hold = peer->init_hold;
 275
 276     bgp_free_routes(peer->routes);
 277     peer->routes = 0;
 278
 279     peer->outbuf->packet.header.len = 0;
 280     peer->outbuf->done = 0;
 281     peer->inbuf->packet.header.len = 0;
 282     peer->inbuf->done = 0;
 283
 284     peer->cli_flag = 0;
 285     peer->events = 0;
 286
 287     if (peer->state != peer->next_state)
 288     {
 289         peer->state = peer->next_state;
 290         peer->state_time = time_now;
 291
 292         LOG(4, 0, 0, "BGP peer %s: state %s\n", peer->name,
 293             bgp_state_str(peer->next_state));
 294     }
 295 }
 296
 297 /* initiate a clean shutdown */
 298 void bgp_stop(struct bgp_peer *peer)
 299 {
 300     LOG(4, 0, 0, "Terminating BGP connection to %s\n", peer->name);
 301     bgp_send_notification(peer, BGP_ERR_CEASE, 0);
 302 }
 303
 304 /* drop connection (if any) and set state to Disabled */
 305 void bgp_halt(struct bgp_peer *peer)
 306 {
 307     LOG(4, 0, 0, "Aborting BGP connection to %s\n", peer->name);
 308     peer->next_state = Disabled;
 309     bgp_clear(peer);
 310 }
 311
 312 /* drop connection (if any) and set to Idle for connection retry */
 313 int bgp_restart(struct bgp_peer *peer)
 314 {
 315     peer->next_state = Idle;
 316     bgp_clear(peer);
 317
 318     /* restart now */
 319     peer->retry_time = time_now;
 320     peer->retry_count = 0;
 321
 322     /* connect */
 323     return bgp_connect(peer);
 324 }
 325
 326 static void bgp_set_retry(struct bgp_peer *peer)
 327 {
 328     if (peer->retry_count++ < BGP_MAX_RETRY)
 329     {
 330         peer->retry_time = time_now + (BGP_RETRY_BACKOFF * peer->retry_count);
 331         peer->next_state = Idle;
 332         bgp_clear(peer);
 333     }
 334     else
 335         bgp_halt(peer); /* give up */
 336 }
 337
 338 /* convert ip/mask to CIDR notation */
 339 static void bgp_cidr(in_addr_t ip, in_addr_t mask, struct bgp_ip_prefix *pfx)
 340 {
 341     int i;
 342     uint32_t b;
 343
 344     /* convert to prefix notation */
 345     pfx->len = 32;
 346     pfx->prefix = ip;
 347
 348     if (!mask) /* bogus */
 349         mask = 0xffffffff;
 350
 351     for (i = 0; i < 32 && ((b = ntohl(1 << i)), !(mask & b)); i++)
 352     {
 353         pfx->len--;
 354         pfx->prefix &= ~b;
 355     }
 356 }
 357
 358 /* insert route into list; sorted */
 359 static struct bgp_route_list *bgp_insert_route(struct bgp_route_list *head,
 360     struct bgp_route_list *new)
 361 {
 362     struct bgp_route_list *p = head;
 363     struct bgp_route_list *e = 0;
 364
 365     while (p && memcmp(&p->dest, &new->dest, sizeof(p->dest)) < 0)
 366     {
 367         e = p;
 368         p = p->next;
 369     }
 370
 371     if (e)
 372     {
 373         new->next = e->next;
 374         e->next = new;
 375     }
 376     else
 377     {
 378         new->next = head;
 379         head = new;
 380     }
 381
 382     return head;
 383 }
 384
 385 /* add route to list for peers */
 386 /*
 387  * Note:  this doesn't do route aggregation, nor drop routes if a less
 388  * specific match already exists (partly because I'm lazy, but also so
 389  * that if that route is later deleted we don't have to be concerned
 390  * about adding back the more specific one).
 391  */
 392 int bgp_add_route(in_addr_t ip, in_addr_t mask)
 393 {
 394     struct bgp_route_list *r = bgp_routes;
 395     struct bgp_route_list add;
 396     int i;
 397
 398     bgp_cidr(ip, mask, &add.dest);
 399     add.next = 0;
 400
 401     /* check for duplicate */
 402     while (r)
 403     {
 404         i = memcmp(&r->dest, &add.dest, sizeof(r->dest));
 405         if (!i)
 406             return 1; /* already covered */
 407
 408         if (i > 0)
 409             break;
 410
 411         r = r->next;
 412     }
 413
 414     /* insert into route list; sorted */
 415     if (!(r = malloc(sizeof(*r))))
 416     {
 417         LOG(0, 0, 0, "Can't allocate route for %s/%d (%s)\n",
 418             fmtaddr(add.dest.prefix, 0), add.dest.len, strerror(errno));
 419
 420         return 0;
 421     }
 422
 423     memcpy(r, &add, sizeof(*r));
 424     bgp_routes = bgp_insert_route(bgp_routes, r);
 425
 426     /* flag established peers for update */
 427     for (i = 0; i < BGP_NUM_PEERS; i++)
 428         if (bgp_peers[i].state == Established)
 429             bgp_peers[i].update_routes = 1;
 430
 431     LOG(4, 0, 0, "Registered BGP route %s/%d\n",
 432         fmtaddr(add.dest.prefix, 0), add.dest.len);
 433
 434     return 1;
 435 }
 436
 437 /* remove route from list for peers */
 438 int bgp_del_route(in_addr_t ip, in_addr_t mask)
 439 {
 440     struct bgp_route_list *r = bgp_routes;
 441     struct bgp_route_list *e = 0;
 442     struct bgp_route_list del;
 443     int i;
 444
 445     bgp_cidr(ip, mask, &del.dest);
 446     del.next = 0;
 447
 448     /* find entry in routes list and remove */
 449     while (r)
 450     {
 451         i = memcmp(&r->dest, &del.dest, sizeof(r->dest));
 452         if (!i)
 453         {
 454             if (e)
 455                 e->next = r->next;
 456             else
 457                 bgp_routes = r->next;
 458
 459             free(r);
 460             break;
 461         }
 462
 463         e = r;
 464
 465         if (i > 0)
 466             r = 0; /* stop */
 467         else
 468             r = r->next;
 469     }
 470
 471     /* not found */
 472     if (!r)
 473         return 1;
 474
 475     /* flag established peers for update */
 476     for (i = 0; i < BGP_NUM_PEERS; i++)
 477         if (bgp_peers[i].state == Established)
 478             bgp_peers[i].update_routes = 1;
 479
 480     LOG(4, 0, 0, "Removed BGP route %s/%d\n",
 481         fmtaddr(del.dest.prefix, 0), del.dest.len);
 482
 483     return 1;
 484 }
 485
 486 /* enable or disable routing */
 487 void bgp_enable_routing(int enable)
 488 {
 489     int i;
 490
 491     for (i = 0; i < BGP_NUM_PEERS; i++)
 492     {
 493         bgp_peers[i].routing = enable;
 494
 495         /* flag established peers for update */
 496         if (bgp_peers[i].state == Established)
 497             bgp_peers[i].update_routes = 1;
 498     }
 499
 500     LOG(4, 0, 0, "%s BGP routing\n", enable ? "Enabled" : "Suspended");
 501 }
 502
 503 #ifdef HAVE_EPOLL
 504 # include <sys/epoll.h>
 505 #else
 506 # include "fake_epoll.h"
 507 #endif
 508
 509 /* return a bitmask of the events required to poll this peer's fd */
 510 int bgp_set_poll()
 511 {
 512     int i;
 513
 514     if (!bgp_configured)
 515         return 0;
 516
 517     for (i = 0; i < BGP_NUM_PEERS; i++)
 518     {
 519         struct bgp_peer *peer = &bgp_peers[i];
 520         int events = 0;
 521
 522         if (peer->state == Disabled || peer->state == Idle)
 523             continue;
 524
 525         if (peer->inbuf->done < BGP_MAX_PACKET_SIZE)
 526             events |= EPOLLIN;
 527
 528         if (peer->state == Connect ||           /* connection in progress */
 529             peer->update_routes ||              /* routing updates */
 530             peer->outbuf->packet.header.len)    /* pending output */
 531             events |= EPOLLOUT;
 532
 533         if (peer->events != events)
 534         {
 535             struct epoll_event ev;
 536
 537             ev.events = peer->events = events;
 538             ev.data.ptr = &peer->edata;
 539             epoll_ctl(epollfd, EPOLL_CTL_MOD, peer->sock, &ev);
 540         }
 541     }
 542
 543     return 1;
 544 }
 545
 546 /* process bgp events/timers */
 547 int bgp_process(uint32_t events[])
 548 {
 549     int i;
 550
 551     if (!bgp_configured)
 552         return 0;
 553
 554     for (i = 0; i < BGP_NUM_PEERS; i++)
 555     {
 556         struct bgp_peer *peer = &bgp_peers[i];
 557
 558         if (*peer->name && peer->cli_flag == BGP_CLI_RESTART)
 559         {
 560             bgp_restart(peer);
 561             continue;
 562         }
 563
 564         if (peer->state == Disabled)
 565             continue;
 566
 567         if (peer->cli_flag)
 568         {
 569             switch (peer->cli_flag)
 570             {
 571             case BGP_CLI_SUSPEND:
 572                 if (peer->routing)
 573                 {
 574                     peer->routing = 0;
 575                     if (peer->state == Established)
 576                         peer->update_routes = 1;
 577                 }
 578
 579                 break;
 580
 581             case BGP_CLI_ENABLE:
 582                 if (!peer->routing)
 583                 {
 584                     peer->routing = 1;
 585                     if (peer->state == Established)
 586                         peer->update_routes = 1;
 587                 }
 588
 589                 break;
 590             }
 591
 592             peer->cli_flag = 0;
 593         }
 594
 595         /* handle empty/fill of buffers */
 596         if (events[i] & EPOLLOUT)
 597         {
 598             int r = 1;
 599             if (peer->state == Connect)
 600                 r = bgp_handle_connect(peer);
 601             else if (peer->outbuf->packet.header.len)
 602                 r = bgp_write(peer);
 603
 604             if (!r)
 605                 continue;
 606         }
 607
 608         if (events[i] & (EPOLLIN|EPOLLHUP))
 609         {
 610             if (!bgp_read(peer))
 611                 continue;
 612         }
 613
 614         /* process input buffer contents */
 615         while (peer->inbuf->done >= sizeof(peer->inbuf->packet.header)
 616             && !peer->outbuf->packet.header.len) /* may need to queue a response */
 617         {
 618             if (bgp_handle_input(peer) < 0)
 619                 continue;
 620         }
 621
 622         /* process pending updates */
 623         if (peer->update_routes
 624             && !peer->outbuf->packet.header.len) /* ditto */
 625         {
 626             if (!bgp_send_update(peer))
 627                 continue;
 628         }
 629
 630         /* process timers */
 631         if (peer->state == Established)
 632         {
 633             if (time_now > peer->expire_time)
 634             {
 635                 LOG(1, 0, 0, "No message from BGP peer %s in %ds\n",
 636                     peer->name, peer->hold);
 637
 638                 bgp_send_notification(peer, BGP_ERR_HOLD_TIMER_EXP, 0);
 639                 continue;
 640             }
 641
 642             if (time_now > peer->keepalive_time && !peer->outbuf->packet.header.len)
 643                 bgp_send_keepalive(peer);
 644         }
 645         else if (peer->state == Idle)
 646         {
 647             if (time_now > peer->retry_time)
 648                 bgp_connect(peer);
 649         }
 650         else if (time_now > peer->state_time + BGP_STATE_TIME)
 651         {
 652             LOG(1, 0, 0, "%s timer expired for BGP peer %s\n",
 653                 bgp_state_str(peer->state), peer->name);
 654
 655             bgp_restart(peer);
 656         }
 657     }
 658
 659     return 1;
 660 }
 661
 662 static void bgp_free_routes(struct bgp_route_list *routes)
 663 {
 664     struct bgp_route_list *tmp;
 665
 666     while ((tmp = routes))
 667     {
 668         routes = tmp->next;
 669         free(tmp);
 670     }
 671 }
 672
 673 char const *bgp_state_str(enum bgp_state state)
 674 {
 675     switch (state)
 676     {
 677     case Disabled:      return "Disabled";
 678     case Idle:          return "Idle";
 679     case Connect:       return "Connect";
 680     case Active:        return "Active";
 681     case OpenSent:      return "OpenSent";
 682     case OpenConfirm:   return "OpenConfirm";
 683     case Established:   return "Established";
 684     }
 685
 686     return "?";
 687 }
 688
 689 static char const *bgp_msg_type_str(uint8_t type)
 690 {
 691     switch (type)
 692     {
 693     case BGP_MSG_OPEN:          return "OPEN";
 694     case BGP_MSG_UPDATE:        return "UPDATE";
 695     case BGP_MSG_NOTIFICATION:  return "NOTIFICATION";
 696     case BGP_MSG_KEEPALIVE:     return "KEEPALIVE";
 697     }
 698
 699     return "?";
 700 }
 701
 702 /* attempt to connect to peer */
 703 static int bgp_connect(struct bgp_peer *peer)
 704 {
 705     static int bgp_port = 0;
 706     struct sockaddr_in addr;
 707     struct sockaddr_in source_addr;
 708     struct epoll_event ev;
 709
 710     if (!bgp_port)
 711     {
 712         struct servent *serv;
 713         if (!(serv = getservbyname("bgp", "tcp")))
 714         {
 715             LOG(0, 0, 0, "Can't get bgp service (%s)\n", strerror(errno));
 716             return 0;
 717         }
 718
 719         bgp_port = serv->s_port;
 720     }
 721
 722     if ((peer->sock = socket(PF_INET, SOCK_STREAM, IPPROTO_TCP)) < 0)
 723     {
 724         LOG(0, 0, 0, "Can't create a socket for BGP peer %s (%s)\n",
 725             peer->name, strerror(errno));
 726
 727         peer->state = peer->next_state = Disabled;
 728         return 0;
 729     }
 730
 731     /* add to poll set */
 732     ev.events = peer->events = EPOLLOUT;
 733     ev.data.ptr = &peer->edata;
 734     epoll_ctl(epollfd, EPOLL_CTL_ADD, peer->sock, &ev);
 735
 736     /* set to non-blocking */
 737     fcntl(peer->sock, F_SETFL, fcntl(peer->sock, F_GETFL, 0) | O_NONBLOCK);
 738
 739     /* set source address */
 740     memset(&source_addr, 0, sizeof(source_addr));
 741     source_addr.sin_family = AF_INET;
 742     source_addr.sin_addr.s_addr = peer->source_addr; /* defaults to INADDR_ANY */
 743     if (bind(peer->sock, (struct sockaddr *) &source_addr, sizeof(source_addr)) < 0)
 744     {
 745         LOG(1, 0, 0, "Can't set source address to %s: %s\n",
 746             inet_ntoa(source_addr.sin_addr), strerror(errno));
 747
 748         bgp_set_retry(peer);
 749         return 0;
 750     }
 751
 752     /* try connect */
 753     memset(&addr, 0, sizeof(addr));
 754     addr.sin_family = AF_INET;
 755     addr.sin_port = bgp_port;
 756     addr.sin_addr.s_addr = peer->addr;
 757
 758     while (connect(peer->sock, (struct sockaddr *) &addr, sizeof(addr)) == -1)
 759     {
 760         if (errno == EINTR) /* SIGALARM handler */
 761             continue;
 762
 763         if (errno != EINPROGRESS)
 764         {
 765             LOG(1, 0, 0, "Can't connect to BGP peer %s (%s)\n",
 766                 inet_ntoa(addr.sin_addr), strerror(errno));
 767
 768             bgp_set_retry(peer);
 769             return 0;
 770         }
 771
 772         peer->state = Connect;
 773         peer->state_time = time_now;
 774
 775         LOG(4, 0, 0, "BGP peer %s: state Connect\n", peer->name);
 776         return 1;
 777     }
 778
 779     peer->state = Active;
 780     peer->state_time = time_now;
 781     peer->retry_time = peer->retry_count = 0;
 782
 783     LOG(4, 0, 0, "BGP peer %s: state Active\n", inet_ntoa(addr.sin_addr));
 784
 785     return bgp_send_open(peer);
 786 }
 787
 788 /* complete partial connection (state = Connect) */
 789 static int bgp_handle_connect(struct bgp_peer *peer)
 790 {
 791     int err = 0;
 792     socklen_t len = sizeof(int);
 793     getsockopt(peer->sock, SOL_SOCKET, SO_ERROR, &err, &len);
 794     if (err)
 795     {
 796         LOG(1, 0, 0, "Can't connect to BGP peer %s (%s)\n", peer->name,
 797             strerror(err));
 798
 799         bgp_set_retry(peer);
 800         return 0;
 801     }
 802
 803     peer->state = Active;
 804     peer->state_time = time_now;
 805
 806     LOG(4, 0, 0, "BGP peer %s: state Active\n", peer->name);
 807
 808     return bgp_send_open(peer);
 809 }
 810
 811 /* initiate a write */
 812 static int bgp_write(struct bgp_peer *peer)
 813 {
 814     int len = htons(peer->outbuf->packet.header.len);
 815     int r;
 816
 817     while ((r = write(peer->sock, &peer->outbuf->packet + peer->outbuf->done,
 818         len - peer->outbuf->done)) == -1)
 819     {
 820         if (errno == EINTR)
 821             continue;
 822
 823         if (errno == EAGAIN)
 824             return 1;
 825
 826         if (errno == EPIPE)
 827             LOG(1, 0, 0, "Connection to BGP peer %s closed\n", peer->name);
 828         else
 829             LOG(1, 0, 0, "Can't write to BGP peer %s (%s)\n", peer->name,
 830                 strerror(errno));
 831
 832         bgp_set_retry(peer);
 833         return 0;
 834     }
 835
 836     if (r < len)
 837     {
 838         peer->outbuf->done += r;
 839         return 1;
 840     }
 841
 842     LOG(4, 0, 0, "Sent %s to BGP peer %s\n",
 843         bgp_msg_type_str(peer->outbuf->packet.header.type), peer->name);
 844
 845     peer->outbuf->packet.header.len = 0;
 846     peer->outbuf->done = 0;
 847
 848     if (peer->state == Established)
 849         peer->keepalive_time = time_now + peer->keepalive;
 850
 851     if (peer->state != peer->next_state)
 852     {
 853         if (peer->next_state == Disabled || peer->next_state == Idle)
 854         {
 855             bgp_clear(peer);
 856             return 0;
 857         }
 858
 859         peer->state = peer->next_state;
 860         peer->state_time = time_now;
 861
 862         LOG(4, 0, 0, "BGP peer %s: state %s\n", peer->name,
 863             bgp_state_str(peer->state));
 864     }
 865
 866     return 1;
 867 }
 868
 869 /* initiate a read */
 870 static int bgp_read(struct bgp_peer *peer)
 871 {
 872     int r;
 873
 874     while ((r = read(peer->sock, &peer->inbuf->packet + peer->inbuf->done,
 875         BGP_MAX_PACKET_SIZE - peer->inbuf->done)) < 1)
 876     {
 877         if (!r)
 878         {
 879             LOG(1, 0, 0, "Connection to BGP peer %s closed\n", peer->name);
 880         }
 881         else
 882         {
 883             if (errno == EINTR)
 884                 continue;
 885
 886             if (errno == EAGAIN)
 887                 return 1;
 888
 889             LOG(1, 0, 0, "Can't read from BGP peer %s (%s)\n", peer->name,
 890                 strerror(errno));
 891         }
 892
 893         bgp_set_retry(peer);
 894         return 0;
 895     }
 896
 897     peer->inbuf->done += r;
 898     return 1;
 899 }
 900
 901 /* process buffered packets */
 902 static int bgp_handle_input(struct bgp_peer *peer)
 903 {
 904     struct bgp_packet *p = &peer->inbuf->packet;
 905     int len = ntohs(p->header.len);
 906
 907     if (len > BGP_MAX_PACKET_SIZE)
 908     {
 909         LOG(1, 0, 0, "Bad header length from BGP %s\n", peer->name);
 910         bgp_send_notification(peer, BGP_ERR_HEADER, BGP_ERR_HDR_BAD_LEN);
 911         return 0;
 912     }
 913
 914     if (peer->inbuf->done < len)
 915         return 0;
 916
 917     LOG(4, 0, 0, "Received %s from BGP peer %s\n",
 918         bgp_msg_type_str(p->header.type), peer->name);
 919
 920     switch (p->header.type)
 921     {
 922     case BGP_MSG_OPEN:
 923         {
 924             struct bgp_data_open data;
 925             int hold;
 926             int i;
 927
 928             for (i = 0; i < sizeof(p->header.marker); i++)
 929             {
 930                 if ((unsigned char) p->header.marker[i] != 0xff)
 931                 {
 932                     LOG(1, 0, 0, "Invalid marker from BGP peer %s\n",
 933                         peer->name);
 934
 935                     bgp_send_notification(peer, BGP_ERR_HEADER,
 936                         BGP_ERR_HDR_NOT_SYNC);
 937
 938                     return 0;
 939                 }
 940             }
 941
 942             if (peer->state != OpenSent)
 943             {
 944                 LOG(1, 0, 0, "OPEN from BGP peer %s in %s state\n",
 945                     peer->name, bgp_state_str(peer->state));
 946
 947                 bgp_send_notification(peer, BGP_ERR_FSM, 0);
 948                 return 0;
 949             }
 950
 951             memcpy(&data, p->data, len - sizeof(p->header));
 952
 953             if (data.version != BGP_VERSION)
 954             {
 955                 LOG(1, 0, 0, "Bad version (%d) sent by BGP peer %s\n",
 956                     (int) data.version, peer->name);
 957
 958                 bgp_send_notification(peer, BGP_ERR_OPEN, BGP_ERR_OPN_VERSION);
 959                 return 0;
 960             }
 961
 962             if (ntohs(data.as) != peer->as)
 963             {
 964                 LOG(1, 0, 0, "Bad AS sent by BGP peer %s (got %d, "
 965                     "expected %d)\n", peer->name, (int) htons(data.as),
 966                     (int) peer->as);
 967
 968                 bgp_send_notification(peer, BGP_ERR_OPEN, BGP_ERR_OPN_BAD_AS);
 969                 return 0;
 970             }
 971
 972             if ((hold = ntohs(data.hold_time)) < 3)
 973             {
 974                 LOG(1, 0, 0, "Bad hold time (%d) from BGP peer %s\n",
 975                     hold, peer->name);
 976
 977                 bgp_send_notification(peer, BGP_ERR_OPEN, BGP_ERR_OPN_HOLD_TIME);
 978                 return 0;
 979             }
 980
 981             /* pick lowest hold time */
 982             if (hold < peer->hold)
 983                 peer->hold = hold;
 984
 985             /* adjust our keepalive based on negotiated hold value */
 986             if (peer->keepalive * 3 > peer->hold)
 987                 peer->keepalive = peer->hold / 3;
 988
 989             /* next transition requires an exchange of keepalives */
 990             bgp_send_keepalive(peer);
 991
 992             /* FIXME: may need to check for optional params */
 993         }
 994
 995         break;
 996
 997     case BGP_MSG_KEEPALIVE:
 998         if (peer->state == OpenConfirm)
 999         {
1000             peer->state = peer->next_state = Established;
1001             peer->state_time = time_now;
1002             peer->keepalive_time = time_now + peer->keepalive;
1003             peer->update_routes = 1;
1004             peer->retry_count = 0;
1005             peer->retry_time = 0;
1006
1007             LOG(4, 0, 0, "BGP peer %s: state Established\n", peer->name);
1008         }
1009
1010         break;
1011
1012     case BGP_MSG_NOTIFICATION:
1013         if (len > sizeof(p->header))
1014         {
1015             struct bgp_data_notification *notification =
1016                 (struct bgp_data_notification *) p->data;
1017
1018             if (notification->error_code == BGP_ERR_CEASE)
1019             {
1020                 LOG(4, 0, 0, "BGP peer %s sent CEASE\n", peer->name);
1021                 bgp_set_retry(peer);
1022                 return 0;
1023             }
1024
1025             /* FIXME: should handle more notifications */
1026             LOG(4, 0, 0, "BGP peer %s sent unhandled NOTIFICATION %d\n",
1027                 peer->name, (int) notification->error_code);
1028         }
1029
1030         break;
1031     }
1032
1033     /* reset timer */
1034     peer->expire_time = time_now + peer->hold;
1035
1036     /* see if there's another message in the same packet/buffer */
1037     if (peer->inbuf->done > len)
1038     {
1039         peer->inbuf->done -= len;
1040         memmove(p, (char *) p + len, peer->inbuf->done);
1041     }
1042     else
1043     {
1044         peer->inbuf->packet.header.len = 0;
1045         peer->inbuf->done = 0;
1046     }
1047
1048     return peer->inbuf->done;
1049 }
1050
1051 /* send/buffer OPEN message */
1052 static int bgp_send_open(struct bgp_peer *peer)
1053 {
1054     struct bgp_data_open data;
1055     uint16_t len = sizeof(peer->outbuf->packet.header);
1056
1057     memset(peer->outbuf->packet.header.marker, 0xff,
1058         sizeof(peer->outbuf->packet.header.marker));
1059
1060     peer->outbuf->packet.header.type = BGP_MSG_OPEN;
1061
1062     data.version = BGP_VERSION;
1063     data.as = htons(our_as);
1064     data.hold_time = htons(peer->hold);
1065     data.identifier = my_address;
1066     data.opt_len = 0;
1067
1068     memcpy(peer->outbuf->packet.data, &data, BGP_DATA_OPEN_SIZE);
1069     len += BGP_DATA_OPEN_SIZE;
1070
1071     peer->outbuf->packet.header.len = htons(len);
1072     peer->outbuf->done = 0;
1073     peer->next_state = OpenSent;
1074
1075     return bgp_write(peer);
1076 }
1077
1078 /* send/buffer KEEPALIVE message */
1079 static int bgp_send_keepalive(struct bgp_peer *peer)
1080 {
1081     memset(peer->outbuf->packet.header.marker, 0xff,
1082         sizeof(peer->outbuf->packet.header.marker));
1083
1084     peer->outbuf->packet.header.type = BGP_MSG_KEEPALIVE;
1085     peer->outbuf->packet.header.len =
1086         htons(sizeof(peer->outbuf->packet.header));
1087
1088     peer->outbuf->done = 0;
1089     peer->next_state = (peer->state == OpenSent) ? OpenConfirm : peer->state;
1090
1091     return bgp_write(peer);
1092 }
1093
1094 /* send/buffer UPDATE message */
1095 static int bgp_send_update(struct bgp_peer *peer)
1096 {
1097     uint16_t unf_len = 0;
1098     uint16_t attr_len;
1099     uint16_t len = sizeof(peer->outbuf->packet.header);
1100     struct bgp_route_list *have = peer->routes;
1101     struct bgp_route_list *want = peer->routing ? bgp_routes : 0;
1102     struct bgp_route_list *e = 0;
1103     struct bgp_route_list *add = 0;
1104     int s;
1105
1106     char *data = (char *) &peer->outbuf->packet.data;
1107
1108     /* need leave room for attr_len, bgp_path_attrs and one prefix */
1109     char *max = (char *) &peer->outbuf->packet.data
1110         + sizeof(peer->outbuf->packet.data)
1111         - sizeof(attr_len) - peer->path_attr_len - sizeof(struct bgp_ip_prefix);
1112
1113     /* skip over unf_len */
1114     data += sizeof(unf_len);
1115     len += sizeof(unf_len);
1116
1117     memset(peer->outbuf->packet.header.marker, 0xff,
1118         sizeof(peer->outbuf->packet.header.marker));
1119
1120     peer->outbuf->packet.header.type = BGP_MSG_UPDATE;
1121
1122     peer->update_routes = 0; /* tentatively clear */
1123
1124     /* find differences */
1125     while ((have || want) && data < (max - sizeof(struct bgp_ip_prefix)))
1126     {
1127         if (have)
1128             s = want
1129                 ? memcmp(&have->dest, &want->dest, sizeof(have->dest))
1130                 : -1;
1131         else
1132             s = 1;
1133
1134         if (s < 0) /* found one to delete */
1135         {
1136             struct bgp_route_list *tmp = have;
1137             have = have->next;
1138
1139             s = BGP_IP_PREFIX_SIZE(tmp->dest);
1140             memcpy(data, &tmp->dest, s);
1141             data += s;
1142             unf_len += s;
1143             len += s;
1144
1145             LOG(5, 0, 0, "Withdrawing route %s/%d from BGP peer %s\n",
1146                 fmtaddr(tmp->dest.prefix, 0), tmp->dest.len, peer->name);
1147
1148             free(tmp);
1149
1150             if (e)
1151                 e->next = have;
1152             else
1153                 peer->routes = have;
1154         }
1155         else
1156         {
1157             if (!s) /* same */
1158             {
1159                 e = have; /* stash the last found to relink above */
1160                 have = have->next;
1161                 want = want->next;
1162             }
1163             else if (s > 0) /* addition reqd. */
1164             {
1165                 if (add)
1166                 {
1167                     peer->update_routes = 1; /* only one add per packet */
1168                     if (!have)
1169                         break;
1170                 }
1171                 else
1172                     add = want;
1173
1174                 if (want)
1175                     want = want->next;
1176             }
1177         }
1178     }
1179
1180     if (have || want)
1181         peer->update_routes = 1; /* more to do */
1182
1183     /* anything changed? */
1184     if (!(unf_len || add))
1185         return 1;
1186
1187     /* go back and insert unf_len */
1188     unf_len = htons(unf_len);
1189     memcpy(&peer->outbuf->packet.data, &unf_len, sizeof(unf_len));
1190
1191     if (add)
1192     {
1193         if (!(e = malloc(sizeof(*e))))
1194         {
1195             LOG(0, 0, 0, "Can't allocate route for %s/%d (%s)\n",
1196                 fmtaddr(add->dest.prefix, 0), add->dest.len, strerror(errno));
1197
1198             return 0;
1199         }
1200
1201         memcpy(e, add, sizeof(*e));
1202         e->next = 0;
1203         peer->routes = bgp_insert_route(peer->routes, e);
1204
1205         attr_len = htons(peer->path_attr_len);
1206         memcpy(data, &attr_len, sizeof(attr_len));
1207         data += sizeof(attr_len);
1208         len += sizeof(attr_len);
1209
1210         memcpy(data, peer->path_attrs, peer->path_attr_len);
1211         data += peer->path_attr_len;
1212         len += peer->path_attr_len;
1213
1214         s = BGP_IP_PREFIX_SIZE(add->dest);
1215         memcpy(data, &add->dest, s);
1216         data += s;
1217         len += s;
1218
1219         LOG(5, 0, 0, "Advertising route %s/%d to BGP peer %s\n",
1220             fmtaddr(add->dest.prefix, 0), add->dest.len, peer->name);
1221     }
1222     else
1223     {
1224         attr_len = 0;
1225         memcpy(data, &attr_len, sizeof(attr_len));
1226         data += sizeof(attr_len);
1227         len += sizeof(attr_len);
1228     }
1229
1230     peer->outbuf->packet.header.len = htons(len);
1231     peer->outbuf->done = 0;
1232
1233     return bgp_write(peer);
1234 }
1235
1236 /* send/buffer NOTIFICATION message */
1237 static int bgp_send_notification(struct bgp_peer *peer, uint8_t code,
1238     uint8_t subcode)
1239 {
1240     struct bgp_data_notification data;
1241     uint16_t len = 0;
1242
1243     data.error_code = code;
1244     len += sizeof(data.error_code);
1245
1246     data.error_subcode = subcode;
1247     len += sizeof(data.error_code);
1248
1249     memset(peer->outbuf->packet.header.marker, 0xff,
1250         sizeof(peer->outbuf->packet.header.marker));
1251
1252     peer->outbuf->packet.header.type = BGP_MSG_NOTIFICATION;
1253     peer->outbuf->packet.header.len =
1254         htons(sizeof(peer->outbuf->packet.header) + len);
1255
1256     memcpy(peer->outbuf->packet.data, &data, len);
1257
1258     peer->outbuf->done = 0;
1259     peer->next_state = code == BGP_ERR_CEASE ? Disabled : Idle;
1260
1261     /* we're dying; ignore any pending input */
1262     peer->inbuf->packet.header.len = 0;
1263     peer->inbuf->done = 0;
1264
1265     return bgp_write(peer);
1266 }