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 1.8 2004-11-29 02:17:17 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(u8 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, u8 code, u8 subcode);
  47
  48 static u16 our_as;
  49 static struct bgp_route_list *bgp_routes = 0;
  50
  51 int bgp_configured = 0;
  52 struct bgp_peer *bgp_peers = 0;
  53
  54 /* prepare peer structure, globals */
  55 int bgp_setup(int as)
  56 {
  57     int i;
  58     struct bgp_peer *peer;
  59
  60     for (i = 0; i < BGP_NUM_PEERS; i++)
  61     {
  62         peer = &bgp_peers[i];
  63         memset(peer, 0, sizeof(*peer));
  64
  65         peer->addr = INADDR_NONE;
  66         peer->sock = -1;
  67         peer->state = peer->next_state = Disabled;
  68
  69         if (!((peer->outbuf = malloc(sizeof(*peer->outbuf)))
  70             && (peer->inbuf = malloc(sizeof(*peer->inbuf)))))
  71         {
  72             LOG(0, 0, 0, "Can't allocate buffers for bgp peer (%s)\n",
  73                 strerror(errno));
  74
  75             return 0;
  76         }
  77     }
  78
  79     if (as < 1)
  80         as = 0;
  81
  82     if ((our_as = as))
  83         return 0;
  84
  85     bgp_routes = 0;
  86     bgp_configured = 0; /* set by bgp_start */
  87
  88     return 1;
  89 }
  90
  91 /* start connection with a peer */
  92 int bgp_start(struct bgp_peer *peer, char *name, int as, int keepalive, int hold, int enable)
  93 {
  94     struct hostent *h;
  95     int ibgp;
  96     int i;
  97     struct bgp_path_attr a;
  98     char path_attrs[64];
  99     char *p = path_attrs;
 100     in_addr_t ip;
 101     u32 metric = htonl(BGP_METRIC);
 102     u32 no_export = htonl(BGP_COMMUNITY_NO_EXPORT);
 103
 104     if (!our_as)
 105         return 0;
 106
 107     if (peer->state != Disabled)
 108         bgp_halt(peer);
 109
 110     snprintf(peer->name, sizeof(peer->name), "%s", name);
 111
 112     if (!(h = gethostbyname(name)) || h->h_addrtype != AF_INET)
 113     {
 114         LOG(0, 0, 0, "Can't get address for BGP peer %s (%s)\n",
 115             name, h ? "no address" : hstrerror(h_errno));
 116
 117         return 0;
 118     }
 119
 120     memcpy(&peer->addr, h->h_addr, sizeof(peer->addr));
 121     peer->as = as > 0 ? as : our_as;
 122     ibgp = peer->as == our_as;
 123
 124     /* set initial timer values */
 125     peer->init_keepalive = keepalive == -1 ? BGP_KEEPALIVE_TIME : keepalive;
 126     peer->init_hold = hold == -1 ? BGP_HOLD_TIME : hold;
 127
 128     if (peer->init_hold < 3)
 129         peer->init_hold = 3;
 130
 131     if (peer->init_keepalive * 3 > peer->init_hold)
 132         peer->init_keepalive = peer->init_hold / 3;
 133
 134     /* clear buffers, go to Idle state */
 135     peer->next_state = Idle;
 136     bgp_clear(peer);
 137
 138     /* set initial routing state */
 139     peer->routing = enable;
 140
 141     /* all our routes use the same attributes, so prepare it in advance */
 142     if (peer->path_attrs)
 143         free(peer->path_attrs);
 144
 145     peer->path_attr_len = 0;
 146
 147     /* ORIGIN */
 148     a.flags = BGP_PATH_ATTR_FLAG_TRANS;
 149     a.code = BGP_PATH_ATTR_CODE_ORIGIN;
 150     a.data.s.len = 1;
 151     a.data.s.value[0] = BGP_PATH_ATTR_CODE_ORIGIN_IGP;
 152
 153 #define ADD_ATTRIBUTE()         do { \
 154     i = BGP_PATH_ATTR_SIZE(a);  \
 155     memcpy(p, &a, i);           \
 156     p += i;                     \
 157     peer->path_attr_len += i;   } while (0)
 158
 159     ADD_ATTRIBUTE();
 160
 161     /* AS_PATH */
 162     a.flags = BGP_PATH_ATTR_FLAG_TRANS;
 163     a.code = BGP_PATH_ATTR_CODE_AS_PATH;
 164     if (ibgp)
 165     {
 166         /* empty path */
 167         a.data.s.len = 0;
 168     }
 169     else
 170     {
 171         /* just our AS */
 172         struct {
 173             u8 type;
 174             u8 len;
 175             u16 value;
 176         } as_path = {
 177             BGP_PATH_ATTR_CODE_AS_PATH_AS_SEQUENCE,
 178             1,
 179             htons(our_as),
 180         };
 181
 182         a.data.s.len = sizeof(as_path);
 183         memcpy(&a.data.s.value, &as_path, sizeof(as_path));
 184     }
 185
 186     ADD_ATTRIBUTE();
 187
 188     /* NEXT_HOP */
 189     a.flags = BGP_PATH_ATTR_FLAG_TRANS;
 190     a.code = BGP_PATH_ATTR_CODE_NEXT_HOP;
 191     ip = my_address; /* we're it */
 192     a.data.s.len = sizeof(ip);
 193     memcpy(a.data.s.value, &ip, sizeof(ip));
 194
 195     ADD_ATTRIBUTE();
 196
 197     /* MULTI_EXIT_DISC */
 198     a.flags = BGP_PATH_ATTR_FLAG_OPTIONAL;
 199     a.code = BGP_PATH_ATTR_CODE_MULTI_EXIT_DISC;
 200     a.data.s.len = sizeof(metric);
 201     memcpy(a.data.s.value, &metric, sizeof(metric));
 202
 203     ADD_ATTRIBUTE();
 204
 205     if (ibgp)
 206     {
 207         u32 local_pref = htonl(BGP_LOCAL_PREF);
 208
 209         /* LOCAL_PREF */
 210         a.flags = BGP_PATH_ATTR_FLAG_TRANS;
 211         a.code = BGP_PATH_ATTR_CODE_LOCAL_PREF;
 212         a.data.s.len = sizeof(local_pref);
 213         memcpy(a.data.s.value, &local_pref, sizeof(local_pref));
 214
 215         ADD_ATTRIBUTE();
 216     }
 217
 218     /* COMMUNITIES */
 219     a.flags = BGP_PATH_ATTR_FLAG_OPTIONAL | BGP_PATH_ATTR_FLAG_TRANS;
 220     a.code = BGP_PATH_ATTR_CODE_COMMUNITIES;
 221     a.data.s.len = sizeof(no_export);
 222     memcpy(a.data.s.value, &no_export, sizeof(no_export));
 223
 224     ADD_ATTRIBUTE();
 225
 226     if (!(peer->path_attrs = malloc(peer->path_attr_len)))
 227     {
 228         LOG(0, 0, 0, "Can't allocate path_attrs for %s (%s)\n",
 229             name, strerror(errno));
 230
 231         return 0;
 232     }
 233
 234     memcpy(peer->path_attrs, path_attrs, peer->path_attr_len);
 235
 236     LOG(4, 0, 0, "Initiating BGP connection to %s (routing %s)\n",
 237         name, enable ? "enabled" : "suspended");
 238
 239     /* we have at least one peer configured */
 240     bgp_configured = 1;
 241
 242     /* connect */
 243     return bgp_connect(peer);
 244 }
 245
 246 /* clear counters, timers, routes and buffers; close socket; move to
 247    next_state, which may be Disabled or Idle */
 248 static void bgp_clear(struct bgp_peer *peer)
 249 {
 250     if (peer->sock != -1)
 251     {
 252         close(peer->sock);
 253         peer->sock = -1;
 254     }
 255
 256     peer->keepalive_time = 0;
 257     peer->expire_time = 0;
 258
 259     peer->keepalive = peer->init_keepalive;
 260     peer->hold = peer->init_hold;
 261
 262     bgp_free_routes(peer->routes);
 263     peer->routes = 0;
 264
 265     peer->outbuf->packet.header.len = 0;
 266     peer->outbuf->done = 0;
 267     peer->inbuf->packet.header.len = 0;
 268     peer->inbuf->done = 0;
 269
 270     peer->cli_flag = 0;
 271
 272     if (peer->state != peer->next_state)
 273     {
 274         peer->state = peer->next_state;
 275         peer->state_time = time_now;
 276
 277         LOG(4, 0, 0, "BGP peer %s: state %s\n", peer->name,
 278             bgp_state_str(peer->next_state));
 279     }
 280 }
 281
 282 /* initiate a clean shutdown */
 283 void bgp_stop(struct bgp_peer *peer)
 284 {
 285     LOG(4, 0, 0, "Terminating BGP connection to %s\n", peer->name);
 286     bgp_send_notification(peer, BGP_ERR_CEASE, 0);
 287 }
 288
 289 /* drop connection (if any) and set state to Disabled */
 290 void bgp_halt(struct bgp_peer *peer)
 291 {
 292     LOG(4, 0, 0, "Aborting BGP connection to %s\n", peer->name);
 293     peer->next_state = Disabled;
 294     bgp_clear(peer);
 295 }
 296
 297 /* drop connection (if any) and set to Idle for connection retry */
 298 int bgp_restart(struct bgp_peer *peer)
 299 {
 300     peer->next_state = Idle;
 301     bgp_clear(peer);
 302
 303     /* restart now */
 304     peer->retry_time = time_now;
 305     peer->retry_count = 0;
 306
 307     /* connect */
 308     return bgp_connect(peer);
 309 }
 310
 311 static void bgp_set_retry(struct bgp_peer *peer)
 312 {
 313     if (peer->retry_count++ < BGP_MAX_RETRY)
 314     {
 315         peer->retry_time = time_now + (BGP_RETRY_BACKOFF * peer->retry_count);
 316         peer->next_state = Idle;
 317         bgp_clear(peer);
 318     }
 319     else
 320         bgp_halt(peer); /* give up */
 321 }
 322
 323 /* convert ip/mask to CIDR notation */
 324 static void bgp_cidr(in_addr_t ip, in_addr_t mask, struct bgp_ip_prefix *pfx)
 325 {
 326     int i;
 327     u32 b;
 328
 329     /* convert to prefix notation */
 330     pfx->len = 32;
 331     pfx->prefix = ip;
 332
 333     if (!mask) /* bogus */
 334         mask = 0xffffffff;
 335
 336     for (i = 0; i < 32 && ((b = ntohl(1 << i)), !(mask & b)); i++)
 337     {
 338         pfx->len--;
 339         pfx->prefix &= ~b;
 340     }
 341 }
 342
 343 /* insert route into list; sorted */
 344 static struct bgp_route_list *bgp_insert_route(struct bgp_route_list *head,
 345     struct bgp_route_list *new)
 346 {
 347     struct bgp_route_list *p = head;
 348     struct bgp_route_list *e = 0;
 349
 350     while (p && memcmp(&p->dest, &new->dest, sizeof(p->dest)) < 0)
 351     {
 352         e = p;
 353         p = p->next;
 354     }
 355
 356     if (e)
 357     {
 358         new->next = e->next;
 359         e->next = new;
 360     }
 361     else
 362     {
 363         new->next = head;
 364         head = new;
 365     }
 366
 367     return head;
 368 }
 369
 370 /* add route to list for peers */
 371 /*
 372  * Note:  this doesn't do route aggregation, nor drop routes if a less
 373  * specific match already exists (partly because I'm lazy, but also so
 374  * that if that route is later deleted we don't have to be concerned
 375  * about adding back the more specific one).
 376  */
 377 int bgp_add_route(in_addr_t ip, in_addr_t mask)
 378 {
 379     struct bgp_route_list *r = bgp_routes;
 380     struct bgp_route_list add;
 381     int i;
 382
 383     bgp_cidr(ip, mask, &add.dest);
 384     add.next = 0;
 385
 386     /* check for duplicate */
 387     while (r)
 388     {
 389         i = memcmp(&r->dest, &add.dest, sizeof(r->dest));
 390         if (!i)
 391             return 1; /* already covered */
 392
 393         if (i > 0)
 394             break;
 395
 396         r = r->next;
 397     }
 398
 399     /* insert into route list; sorted */
 400     if (!(r = malloc(sizeof(*r))))
 401     {
 402         LOG(0, 0, 0, "Can't allocate route for %s/%d (%s)\n",
 403             fmtaddr(add.dest.prefix, 0), add.dest.len, strerror(errno));
 404
 405         return 0;
 406     }
 407
 408     memcpy(r, &add, sizeof(*r));
 409     bgp_routes = bgp_insert_route(bgp_routes, r);
 410
 411     /* flag established peers for update */
 412     for (i = 0; i < BGP_NUM_PEERS; i++)
 413         if (bgp_peers[i].state == Established)
 414             bgp_peers[i].update_routes = 1;
 415
 416     LOG(4, 0, 0, "Registered BGP route %s/%d\n",
 417         fmtaddr(add.dest.prefix, 0), add.dest.len);
 418
 419     return 1;
 420 }
 421
 422 /* remove route from list for peers */
 423 int bgp_del_route(in_addr_t ip, in_addr_t mask)
 424 {
 425     struct bgp_route_list *r = bgp_routes;
 426     struct bgp_route_list *e = 0;
 427     struct bgp_route_list del;
 428     int i;
 429
 430     bgp_cidr(ip, mask, &del.dest);
 431     del.next = 0;
 432
 433     /* find entry in routes list and remove */
 434     while (r)
 435     {
 436         i = memcmp(&r->dest, &del.dest, sizeof(r->dest));
 437         if (!i)
 438         {
 439             if (e)
 440                 e->next = r->next;
 441             else
 442                 bgp_routes = r->next;
 443
 444             free(r);
 445             break;
 446         }
 447
 448         e = r;
 449
 450         if (i > 0)
 451             r = 0; /* stop */
 452         else
 453             r = r->next;
 454     }
 455
 456     /* not found */
 457     if (!r)
 458         return 1;
 459
 460     /* flag established peers for update */
 461     for (i = 0; i < BGP_NUM_PEERS; i++)
 462         if (bgp_peers[i].state == Established)
 463             bgp_peers[i].update_routes = 1;
 464
 465     LOG(4, 0, 0, "Removed BGP route %s/%d\n",
 466         fmtaddr(del.dest.prefix, 0), del.dest.len);
 467
 468     return 1;
 469 }
 470
 471 /* enable or disable routing */
 472 void bgp_enable_routing(int enable)
 473 {
 474     int i;
 475
 476     for (i = 0; i < BGP_NUM_PEERS; i++)
 477     {
 478         bgp_peers[i].routing = enable;
 479
 480         /* flag established peers for update */
 481         if (bgp_peers[i].state == Established)
 482             bgp_peers[i].update_routes = 1;
 483     }
 484
 485     LOG(4, 0, 0, "%s BGP routing\n", enable ? "Enabled" : "Suspended");
 486 }
 487
 488 /* return a bitmask indicating if the socket should be added to the
 489    read set (1) and or write set (2) for select */
 490 int bgp_select_state(struct bgp_peer *peer)
 491 {
 492     int flags = 0;
 493
 494     if (!bgp_configured)
 495         return 0;
 496
 497     if (peer->state == Disabled || peer->state == Idle)
 498         return 0;
 499
 500     if (peer->inbuf->done < BGP_MAX_PACKET_SIZE)
 501         flags |= 1;
 502
 503     if (peer->state == Connect ||               /* connection in progress */
 504         peer->update_routes ||                  /* routing updates */
 505         peer->outbuf->packet.header.len)        /* pending output */
 506         flags |= 2;
 507
 508     return flags;
 509 }
 510
 511 /* process bgp peer */
 512 int bgp_process(struct bgp_peer *peer, int readable, int writable)
 513 {
 514     if (!bgp_configured)
 515         return 0;
 516
 517     if (*peer->name && peer->cli_flag == BGP_CLI_RESTART)
 518         return bgp_restart(peer);
 519
 520     if (peer->state == Disabled)
 521         return 1;
 522
 523     if (peer->cli_flag)
 524     {
 525         switch (peer->cli_flag)
 526         {
 527         case BGP_CLI_SUSPEND:
 528             if (peer->routing)
 529             {
 530                 peer->routing = 0;
 531                 if (peer->state == Established)
 532                     peer->update_routes = 1;
 533             }
 534
 535             break;
 536
 537         case BGP_CLI_ENABLE:
 538             if (!peer->routing)
 539             {
 540                 peer->routing = 1;
 541                 if (peer->state == Established)
 542                     peer->update_routes = 1;
 543             }
 544
 545             break;
 546         }
 547
 548         peer->cli_flag = 0;
 549     }
 550
 551     /* handle empty/fill of buffers */
 552     if (writable)
 553     {
 554         int r = 1;
 555         if (peer->state == Connect)
 556             r = bgp_handle_connect(peer);
 557         else if (peer->outbuf->packet.header.len)
 558             r = bgp_write(peer);
 559
 560         if (!r)
 561             return 0;
 562     }
 563
 564     if (readable)
 565     {
 566         if (!bgp_read(peer))
 567             return 0;
 568     }
 569
 570     /* process input buffer contents */
 571     while (peer->inbuf->done >= sizeof(peer->inbuf->packet.header)
 572         && !peer->outbuf->packet.header.len) /* may need to queue a response */
 573     {
 574         if (bgp_handle_input(peer) < 0)
 575             return 0;
 576     }
 577
 578     /* process pending updates */
 579     if (peer->update_routes
 580         && !peer->outbuf->packet.header.len) /* ditto */
 581     {
 582         if (!bgp_send_update(peer))
 583             return 0;
 584     }
 585
 586     /* process timers */
 587     if (peer->state == Established)
 588     {
 589         if (time_now > peer->expire_time)
 590         {
 591             LOG(1, 0, 0, "No message from BGP peer %s in %ds\n",
 592                 peer->name, peer->hold);
 593
 594             bgp_send_notification(peer, BGP_ERR_HOLD_TIMER_EXP, 0);
 595             return 0;
 596         }
 597
 598         if (time_now > peer->keepalive_time && !peer->outbuf->packet.header.len)
 599             bgp_send_keepalive(peer);
 600     }
 601     else if (peer->state == Idle)
 602     {
 603         if (time_now > peer->retry_time)
 604             return bgp_connect(peer);
 605     }
 606     else if (time_now > peer->state_time + BGP_STATE_TIME)
 607     {
 608         LOG(1, 0, 0, "%s timer expired for BGP peer %s\n",
 609             bgp_state_str(peer->state), peer->name);
 610
 611         return bgp_restart(peer);
 612     }
 613
 614     return 1;
 615 }
 616
 617 static void bgp_free_routes(struct bgp_route_list *routes)
 618 {
 619     struct bgp_route_list *tmp;
 620
 621     while ((tmp = routes))
 622     {
 623         routes = tmp->next;
 624         free(tmp);
 625     }
 626 }
 627
 628 char const *bgp_state_str(enum bgp_state state)
 629 {
 630     switch (state)
 631     {
 632     case Disabled:      return "Disabled";
 633     case Idle:          return "Idle";
 634     case Connect:       return "Connect";
 635     case Active:        return "Active";
 636     case OpenSent:      return "OpenSent";
 637     case OpenConfirm:   return "OpenConfirm";
 638     case Established:   return "Established";
 639     }
 640
 641     return "?";
 642 }
 643
 644 static char const *bgp_msg_type_str(u8 type)
 645 {
 646     switch (type)
 647     {
 648     case BGP_MSG_OPEN:          return "OPEN";
 649     case BGP_MSG_UPDATE:        return "UPDATE";
 650     case BGP_MSG_NOTIFICATION:  return "NOTIFICATION";
 651     case BGP_MSG_KEEPALIVE:     return "KEEPALIVE";
 652     }
 653
 654     return "?";
 655 }
 656
 657 /* attempt to connect to peer */
 658 static int bgp_connect(struct bgp_peer *peer)
 659 {
 660     static int bgp_port = 0;
 661     struct sockaddr_in addr;
 662
 663     if (!bgp_port)
 664     {
 665         struct servent *serv;
 666         if (!(serv = getservbyname("bgp", "tcp")))
 667         {
 668             LOG(0, 0, 0, "Can't get bgp service (%s)\n", strerror(errno));
 669             return 0;
 670         }
 671
 672         bgp_port = serv->s_port;
 673     }
 674
 675     if ((peer->sock = socket(PF_INET, SOCK_STREAM, IPPROTO_TCP)) < 0)
 676     {
 677         LOG(0, 0, 0, "Can't create a socket for BGP peer %s (%s)\n",
 678             peer->name, strerror(errno));
 679
 680         peer->state = peer->next_state = Disabled;
 681         return 0;
 682     }
 683
 684     /* set to non-blocking */
 685     fcntl(peer->sock, F_SETFL, fcntl(peer->sock, F_GETFL, 0) | O_NONBLOCK);
 686
 687     /* try connect */
 688     memset(&addr, 0, sizeof(addr));
 689     addr.sin_family = AF_INET;
 690     addr.sin_port = bgp_port;
 691     addr.sin_addr.s_addr = peer->addr;
 692
 693     while (connect(peer->sock, (struct sockaddr *) &addr, sizeof(addr)) == -1)
 694     {
 695         if (errno == EINTR) /* SIGALARM handler */
 696             continue;
 697
 698         if (errno != EINPROGRESS)
 699         {
 700             LOG(1, 0, 0, "Can't connect to BGP peer %s (%s)\n",
 701                 inet_ntoa(addr.sin_addr), strerror(errno));
 702
 703             bgp_set_retry(peer);
 704             return 0;
 705         }
 706
 707         peer->state = Connect;
 708         peer->state_time = time_now;
 709
 710         LOG(4, 0, 0, "BGP peer %s: state Connect\n", peer->name);
 711         return 1;
 712     }
 713
 714     peer->state = Active;
 715     peer->state_time = time_now;
 716     peer->retry_time = peer->retry_count = 0;
 717
 718     LOG(4, 0, 0, "BGP peer %s: state Active\n", inet_ntoa(addr.sin_addr));
 719
 720     return bgp_send_open(peer);
 721 }
 722
 723 /* complete partial connection (state = Connect) */
 724 static int bgp_handle_connect(struct bgp_peer *peer)
 725 {
 726     int err = 0;
 727     int len = sizeof(int);
 728     getsockopt(peer->sock, SOL_SOCKET, SO_ERROR, &err, &len);
 729     if (err)
 730     {
 731         LOG(1, 0, 0, "Can't connect to BGP peer %s (%s)\n", peer->name,
 732             strerror(err));
 733
 734         bgp_set_retry(peer);
 735         return 0;
 736     }
 737
 738     peer->state = Active;
 739     peer->state_time = time_now;
 740
 741     LOG(4, 0, 0, "BGP peer %s: state Active\n", peer->name);
 742
 743     return bgp_send_open(peer);
 744 }
 745
 746 /* initiate a write */
 747 static int bgp_write(struct bgp_peer *peer)
 748 {
 749     int len = htons(peer->outbuf->packet.header.len);
 750     int r;
 751
 752     while ((r = write(peer->sock, &peer->outbuf->packet + peer->outbuf->done,
 753         len - peer->outbuf->done)) == -1)
 754     {
 755         if (errno == EINTR)
 756             continue;
 757
 758         if (errno == EAGAIN)
 759             return 1;
 760
 761         if (errno == EPIPE)
 762             LOG(1, 0, 0, "Connection to BGP peer %s closed\n", peer->name);
 763         else
 764             LOG(1, 0, 0, "Can't write to BGP peer %s (%s)\n", peer->name,
 765                 strerror(errno));
 766
 767         bgp_set_retry(peer);
 768         return 0;
 769     }
 770
 771     if (r < len)
 772     {
 773         peer->outbuf->done += r;
 774         return 1;
 775     }
 776
 777     LOG(4, 0, 0, "Sent %s to BGP peer %s\n",
 778         bgp_msg_type_str(peer->outbuf->packet.header.type), peer->name);
 779
 780     peer->outbuf->packet.header.len = 0;
 781     peer->outbuf->done = 0;
 782
 783     if (peer->state == Established)
 784         peer->keepalive_time = time_now + peer->keepalive;
 785
 786     if (peer->state != peer->next_state)
 787     {
 788         if (peer->next_state == Disabled || peer->next_state == Idle)
 789         {
 790             bgp_clear(peer);
 791             return 0;
 792         }
 793
 794         peer->state = peer->next_state;
 795         peer->state_time = time_now;
 796
 797         LOG(4, 0, 0, "BGP peer %s: state %s\n", peer->name,
 798             bgp_state_str(peer->state));
 799     }
 800
 801     return 1;
 802 }
 803
 804 /* initiate a read */
 805 static int bgp_read(struct bgp_peer *peer)
 806 {
 807     int r;
 808
 809     while ((r = read(peer->sock, &peer->inbuf->packet + peer->inbuf->done,
 810         BGP_MAX_PACKET_SIZE - peer->inbuf->done)) < 1)
 811     {
 812         if (!r)
 813         {
 814             LOG(1, 0, 0, "Connection to BGP peer %s closed\n", peer->name);
 815         }
 816         else
 817         {
 818             if (errno == EINTR)
 819                 continue;
 820
 821             if (errno == EAGAIN)
 822                 return 1;
 823
 824             LOG(1, 0, 0, "Can't read from BGP peer %s (%s)\n", peer->name,
 825                 strerror(errno));
 826         }
 827
 828         bgp_set_retry(peer);
 829         return 0;
 830     }
 831
 832     peer->inbuf->done += r;
 833     return 1;
 834 }
 835
 836 /* process buffered packets */
 837 static int bgp_handle_input(struct bgp_peer *peer)
 838 {
 839     struct bgp_packet *p = &peer->inbuf->packet;
 840     int len = ntohs(p->header.len);
 841
 842     if (len > BGP_MAX_PACKET_SIZE)
 843     {
 844         LOG(1, 0, 0, "Bad header length from BGP %s\n", peer->name);
 845         bgp_send_notification(peer, BGP_ERR_HEADER, BGP_ERR_HDR_BAD_LEN);
 846         return 0;
 847     }
 848
 849     if (peer->inbuf->done < len)
 850         return 0;
 851
 852     LOG(4, 0, 0, "Received %s from BGP peer %s\n",
 853         bgp_msg_type_str(p->header.type), peer->name);
 854
 855     switch (p->header.type)
 856     {
 857     case BGP_MSG_OPEN:
 858         {
 859             struct bgp_data_open data;
 860             int hold;
 861             int i;
 862
 863             for (i = 0; i < sizeof(p->header.marker); i++)
 864             {
 865                 if ((unsigned char) p->header.marker[i] != 0xff)
 866                 {
 867                     LOG(1, 0, 0, "Invalid marker from BGP peer %s\n",
 868                         peer->name);
 869
 870                     bgp_send_notification(peer, BGP_ERR_HEADER,
 871                         BGP_ERR_HDR_NOT_SYNC);
 872
 873                     return 0;
 874                 }
 875             }
 876
 877             if (peer->state != OpenSent)
 878             {
 879                 LOG(1, 0, 0, "OPEN from BGP peer %s in %s state\n",
 880                     peer->name, bgp_state_str(peer->state));
 881
 882                 bgp_send_notification(peer, BGP_ERR_FSM, 0);
 883                 return 0;
 884             }
 885
 886             memcpy(&data, p->data, len - sizeof(p->header));
 887
 888             if (data.version != BGP_VERSION)
 889             {
 890                 LOG(1, 0, 0, "Bad version (%d) sent by BGP peer %s\n",
 891                     (int) data.version, peer->name);
 892
 893                 bgp_send_notification(peer, BGP_ERR_OPEN, BGP_ERR_OPN_VERSION);
 894                 return 0;
 895             }
 896
 897             if (ntohs(data.as) != peer->as)
 898             {
 899                 LOG(1, 0, 0, "Bad AS sent by BGP peer %s (got %d, "
 900                     "expected %d)\n", peer->name, (int) htons(data.as),
 901                     (int) peer->as);
 902
 903                 bgp_send_notification(peer, BGP_ERR_OPEN, BGP_ERR_OPN_BAD_AS);
 904                 return 0;
 905             }
 906
 907             if ((hold = ntohs(data.hold_time)) < 3)
 908             {
 909                 LOG(1, 0, 0, "Bad hold time (%d) from BGP peer %s\n",
 910                     hold, peer->name);
 911
 912                 bgp_send_notification(peer, BGP_ERR_OPEN, BGP_ERR_OPN_HOLD_TIME);
 913                 return 0;
 914             }
 915
 916             /* pick lowest hold time */
 917             if (hold < peer->hold)
 918                 peer->hold = hold;
 919
 920             /* adjust our keepalive based on negotiated hold value */
 921             if (peer->keepalive * 3 > peer->hold)
 922                 peer->keepalive = peer->hold / 3;
 923
 924             /* next transition requires an exchange of keepalives */
 925             bgp_send_keepalive(peer);
 926
 927             /* FIXME: may need to check for optional params */
 928         }
 929
 930         break;
 931
 932     case BGP_MSG_KEEPALIVE:
 933         if (peer->state == OpenConfirm)
 934         {
 935             peer->state = peer->next_state = Established;
 936             peer->state_time = time_now;
 937             peer->keepalive_time = time_now + peer->keepalive;
 938             peer->update_routes = 1;
 939             peer->retry_count = 0;
 940             peer->retry_time = 0;
 941
 942             LOG(4, 0, 0, "BGP peer %s: state Established\n", peer->name);
 943         }
 944
 945         break;
 946
 947     case BGP_MSG_NOTIFICATION:
 948         if (len > sizeof(p->header))
 949         {
 950             struct bgp_data_notification *notification =
 951                 (struct bgp_data_notification *) p->data;
 952
 953             if (notification->error_code == BGP_ERR_CEASE)
 954             {
 955                 LOG(4, 0, 0, "BGP peer %s sent CEASE\n", peer->name);
 956                 bgp_halt(peer);
 957                 return 0;
 958             }
 959
 960             /* FIXME: should handle more notifications */
 961             LOG(4, 0, 0, "BGP peer %s sent unhandled NOTIFICATION %d\n",
 962                 peer->name, (int) notification->error_code);
 963         }
 964
 965         break;
 966     }
 967
 968     /* reset timer */
 969     peer->expire_time = time_now + peer->hold;
 970
 971     /* see if there's another message in the same packet/buffer */
 972     if (peer->inbuf->done > len)
 973     {
 974         peer->inbuf->done -= len;
 975         memmove(p, (char *) p + len, peer->inbuf->done);
 976     }
 977     else
 978     {
 979         peer->inbuf->packet.header.len = 0;
 980         peer->inbuf->done = 0;
 981     }
 982
 983     return peer->inbuf->done;
 984 }
 985
 986 /* send/buffer OPEN message */
 987 static int bgp_send_open(struct bgp_peer *peer)
 988 {
 989     struct bgp_data_open data;
 990     u16 len = sizeof(peer->outbuf->packet.header);
 991
 992     memset(peer->outbuf->packet.header.marker, 0xff,
 993         sizeof(peer->outbuf->packet.header.marker));
 994
 995     peer->outbuf->packet.header.type = BGP_MSG_OPEN;
 996
 997     data.version = BGP_VERSION;
 998     data.as = htons(our_as);
 999     data.hold_time = htons(peer->hold);
1000     data.identifier = my_address;
1001     data.opt_len = 0;
1002
1003     memcpy(peer->outbuf->packet.data, &data, BGP_DATA_OPEN_SIZE);
1004     len += BGP_DATA_OPEN_SIZE;
1005
1006     peer->outbuf->packet.header.len = htons(len);
1007     peer->outbuf->done = 0;
1008     peer->next_state = OpenSent;
1009
1010     return bgp_write(peer);
1011 }
1012
1013 /* send/buffer KEEPALIVE message */
1014 static int bgp_send_keepalive(struct bgp_peer *peer)
1015 {
1016     memset(peer->outbuf->packet.header.marker, 0xff,
1017         sizeof(peer->outbuf->packet.header.marker));
1018
1019     peer->outbuf->packet.header.type = BGP_MSG_KEEPALIVE;
1020     peer->outbuf->packet.header.len =
1021         htons(sizeof(peer->outbuf->packet.header));
1022
1023     peer->outbuf->done = 0;
1024     peer->next_state = (peer->state == OpenSent) ? OpenConfirm : peer->state;
1025
1026     return bgp_write(peer);
1027 }
1028
1029 /* send/buffer UPDATE message */
1030 static int bgp_send_update(struct bgp_peer *peer)
1031 {
1032     u16 unf_len = 0;
1033     u16 attr_len;
1034     u16 len = sizeof(peer->outbuf->packet.header);
1035     struct bgp_route_list *have = peer->routes;
1036     struct bgp_route_list *want = peer->routing ? bgp_routes : 0;
1037     struct bgp_route_list *e = 0;
1038     struct bgp_route_list *add = 0;
1039     int s;
1040
1041     char *data = (char *) &peer->outbuf->packet.data;
1042
1043     /* need leave room for attr_len, bgp_path_attrs and one prefix */
1044     char *max = (char *) &peer->outbuf->packet.data
1045         + sizeof(peer->outbuf->packet.data)
1046         - sizeof(attr_len) - peer->path_attr_len - sizeof(struct bgp_ip_prefix);
1047
1048     /* skip over unf_len */
1049     data += sizeof(unf_len);
1050     len += sizeof(unf_len);
1051
1052     memset(peer->outbuf->packet.header.marker, 0xff,
1053         sizeof(peer->outbuf->packet.header.marker));
1054
1055     peer->outbuf->packet.header.type = BGP_MSG_UPDATE;
1056
1057     peer->update_routes = 0; /* tentatively clear */
1058
1059     /* find differences */
1060     while ((have || want) && data < (max - sizeof(struct bgp_ip_prefix)))
1061     {
1062         if (have)
1063             s = want
1064                 ? memcmp(&have->dest, &want->dest, sizeof(have->dest))
1065                 : -1;
1066         else
1067             s = 1;
1068
1069         if (s < 0) /* found one to delete */
1070         {
1071             struct bgp_route_list *tmp = have;
1072             have = have->next;
1073
1074             s = BGP_IP_PREFIX_SIZE(tmp->dest);
1075             memcpy(data, &tmp->dest, s);
1076             data += s;
1077             unf_len += s;
1078             len += s;
1079
1080             LOG(5, 0, 0, "Withdrawing route %s/%d from BGP peer %s\n",
1081                 fmtaddr(tmp->dest.prefix, 0), tmp->dest.len, peer->name);
1082
1083             free(tmp);
1084
1085             if (e)
1086                 e->next = have;
1087             else
1088                 peer->routes = have;
1089         }
1090         else
1091         {
1092             if (!s) /* same */
1093             {
1094                 e = have; /* stash the last found to relink above */
1095                 have = have->next;
1096                 want = want->next;
1097             }
1098             else if (s > 0) /* addition reqd. */
1099             {
1100                 if (add)
1101                 {
1102                     peer->update_routes = 1; /* only one add per packet */
1103                     if (!have)
1104                         break;
1105                 }
1106                 else
1107                     add = want;
1108
1109                 if (want)
1110                     want = want->next;
1111             }
1112         }
1113     }
1114
1115     if (have || want)
1116         peer->update_routes = 1; /* more to do */
1117
1118     /* anything changed? */
1119     if (!(unf_len || add))
1120         return 1;
1121
1122     /* go back and insert unf_len */
1123     unf_len = htons(unf_len);
1124     memcpy(&peer->outbuf->packet.data, &unf_len, sizeof(unf_len));
1125
1126     if (add)
1127     {
1128         if (!(e = malloc(sizeof(*e))))
1129         {
1130             LOG(0, 0, 0, "Can't allocate route for %s/%d (%s)\n",
1131                 fmtaddr(add->dest.prefix, 0), add->dest.len, strerror(errno));
1132
1133             return 0;
1134         }
1135
1136         memcpy(e, add, sizeof(*e));
1137         e->next = 0;
1138         peer->routes = bgp_insert_route(peer->routes, e);
1139
1140         attr_len = htons(peer->path_attr_len);
1141         memcpy(data, &attr_len, sizeof(attr_len));
1142         data += sizeof(attr_len);
1143         len += sizeof(attr_len);
1144
1145         memcpy(data, peer->path_attrs, peer->path_attr_len);
1146         data += peer->path_attr_len;
1147         len += peer->path_attr_len;
1148
1149         s = BGP_IP_PREFIX_SIZE(add->dest);
1150         memcpy(data, &add->dest, s);
1151         data += s;
1152         len += s;
1153
1154         LOG(5, 0, 0, "Advertising route %s/%d to BGP peer %s\n",
1155             fmtaddr(add->dest.prefix, 0), add->dest.len, peer->name);
1156     }
1157     else
1158     {
1159         attr_len = 0;
1160         memcpy(data, &attr_len, sizeof(attr_len));
1161         data += sizeof(attr_len);
1162         len += sizeof(attr_len);
1163     }
1164
1165     peer->outbuf->packet.header.len = htons(len);
1166     peer->outbuf->done = 0;
1167
1168     return bgp_write(peer);
1169 }
1170
1171 /* send/buffer NOTIFICATION message */
1172 static int bgp_send_notification(struct bgp_peer *peer, u8 code, u8 subcode)
1173 {
1174     struct bgp_data_notification data;
1175     u16 len = 0;
1176
1177     data.error_code = code;
1178     len += sizeof(data.error_code);
1179
1180     data.error_subcode = subcode;
1181     len += sizeof(data.error_code);
1182
1183     memset(peer->outbuf->packet.header.marker, 0xff,
1184         sizeof(peer->outbuf->packet.header.marker));
1185
1186     peer->outbuf->packet.header.type = BGP_MSG_NOTIFICATION;
1187     peer->outbuf->packet.header.len =
1188         htons(sizeof(peer->outbuf->packet.header) + len);
1189
1190     memcpy(peer->outbuf->packet.data, &data, len);
1191
1192     peer->outbuf->done = 0;
1193     peer->next_state = code == BGP_ERR_CEASE ? Disabled : Idle;
1194
1195     /* we're dying; ignore any pending input */
1196     peer->inbuf->packet.header.len = 0;
1197     peer->inbuf->done = 0;
1198
1199     return bgp_write(peer);
1200 }