ctdb/tcp/tcp_connect.c

   1 /*
   2    ctdb over TCP
   3
   4    Copyright (C) Andrew Tridgell  2006
   5    Copyright (C) Ronnie Sahlberg  2008
   6
   7    This program is free software; you can redistribute it and/or modify
   8    it under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3 of the License, or
  10    (at your option) any later version.
  11
  12    This program is distributed in the hope that it will be useful,
  13    but WITHOUT ANY WARRANTY; without even the implied warranty of
  14    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15    GNU General Public License for more details.
  16
  17    You should have received a copy of the GNU General Public License
  18    along with this program; if not, see <http://www.gnu.org/licenses/>.
  19 */
  20
  21 #include "replace.h"
  22 #include "system/network.h"
  23 #include "system/filesys.h"
  24
  25 #include <talloc.h>
  26 #include <tevent.h>
  27
  28 #include "lib/util/debug.h"
  29 #include "lib/util/time.h"
  30 #include "lib/util/blocking.h"
  31
  32 #include "ctdb_private.h"
  33
  34 #include "common/system.h"
  35 #include "common/common.h"
  36 #include "common/logging.h"
  37 #include "common/path.h"
  38
  39 #include "protocol/protocol_util.h"
  40
  41 #include "ctdb_tcp.h"
  42
  43 /*
  44   stop any outgoing connection (established or pending) to a node
  45  */
  46 void ctdb_tcp_stop_outgoing(struct ctdb_node *node)
  47 {
  48         struct ctdb_tcp_node *tnode = talloc_get_type(
  49                 node->transport_data, struct ctdb_tcp_node);
  50
  51         TALLOC_FREE(tnode->out_queue);
  52         TALLOC_FREE(tnode->connect_te);
  53         TALLOC_FREE(tnode->connect_fde);
  54         if (tnode->out_fd != -1) {
  55                 close(tnode->out_fd);
  56                 tnode->out_fd = -1;
  57         }
  58 }
  59
  60 /*
  61   stop incoming connection to a node
  62  */
  63 void ctdb_tcp_stop_incoming(struct ctdb_node *node)
  64 {
  65         struct ctdb_tcp_node *tnode = talloc_get_type(
  66                 node->transport_data, struct ctdb_tcp_node);
  67
  68         TALLOC_FREE(tnode->in_queue);
  69 }
  70
  71 /*
  72   called when a complete packet has come in - should not happen on this socket
  73   unless the other side closes the connection with RST or FIN
  74  */
  75 void ctdb_tcp_tnode_cb(uint8_t *data, size_t cnt, void *private_data)
  76 {
  77         struct ctdb_node *node = talloc_get_type(private_data, struct ctdb_node);
  78
  79         node->ctdb->upcalls->node_dead(node);
  80
  81         TALLOC_FREE(data);
  82 }
  83
  84 /*
  85   called when socket becomes writeable on connect
  86 */
  87 static void ctdb_node_connect_write(struct tevent_context *ev,
  88                                     struct tevent_fd *fde,
  89                                     uint16_t flags, void *private_data)
  90 {
  91         struct ctdb_node *node = talloc_get_type(private_data,
  92                                                  struct ctdb_node);
  93         struct ctdb_tcp_node *tnode = talloc_get_type(node->transport_data,
  94                                                       struct ctdb_tcp_node);
  95         struct ctdb_context *ctdb = node->ctdb;
  96         int error = 0;
  97         socklen_t len = sizeof(error);
  98         int one = 1;
  99         int ret;
 100
 101         TALLOC_FREE(tnode->connect_te);
 102
 103         ret = getsockopt(tnode->out_fd, SOL_SOCKET, SO_ERROR, &error, &len);
 104         if (ret != 0 || error != 0) {
 105                 ctdb_tcp_stop_outgoing(node);
 106                 tnode->connect_te = tevent_add_timer(ctdb->ev, tnode,
 107                                                     timeval_current_ofs(1, 0),
 108                                                     ctdb_tcp_node_connect, node);
 109                 return;
 110         }
 111
 112         TALLOC_FREE(tnode->connect_fde);
 113
 114         ret = setsockopt(tnode->out_fd,
 115                          IPPROTO_TCP,
 116                          TCP_NODELAY,
 117                          (char *)&one,
 118                          sizeof(one));
 119         if (ret == -1) {
 120                 DBG_WARNING("Failed to set TCP_NODELAY on fd - %s\n",
 121                           strerror(errno));
 122         }
 123         ret = setsockopt(tnode->out_fd,
 124                          SOL_SOCKET,
 125                          SO_KEEPALIVE,(char *)&one,
 126                          sizeof(one));
 127         if (ret == -1) {
 128                 DBG_WARNING("Failed to set KEEPALIVE on fd - %s\n",
 129                             strerror(errno));
 130         }
 131
 132         tnode->out_queue = ctdb_queue_setup(node->ctdb,
 133                                             tnode,
 134                                             tnode->out_fd,
 135                                             CTDB_TCP_ALIGNMENT,
 136                                             ctdb_tcp_tnode_cb,
 137                                             node,
 138                                             "to-node-%s",
 139                                             node->name);
 140         if (tnode->out_queue == NULL) {
 141                 DBG_ERR("Failed to set up outgoing queue\n");
 142                 ctdb_tcp_stop_outgoing(node);
 143                 tnode->connect_te = tevent_add_timer(ctdb->ev,
 144                                                      tnode,
 145                                                      timeval_current_ofs(1, 0),
 146                                                      ctdb_tcp_node_connect,
 147                                                      node);
 148                 return;
 149         }
 150
 151         /* the queue subsystem now owns this fd */
 152         tnode->out_fd = -1;
 153
 154         /*
 155          * Mark the node to which this connection has been established
 156          * as connected, but only if the corresponding listening
 157          * socket is also connected
 158          */
 159         if (tnode->in_queue != NULL) {
 160                 node->ctdb->upcalls->node_connected(node);
 161         }
 162 }
 163
 164
 165 static void ctdb_tcp_node_connect_timeout(struct tevent_context *ev,
 166                                           struct tevent_timer *te,
 167                                           struct timeval t,
 168                                           void *private_data);
 169
 170 /*
 171   called when we should try and establish a tcp connection to a node
 172 */
 173 static void ctdb_tcp_start_outgoing(struct ctdb_node *node)
 174 {
 175         struct ctdb_tcp_node *tnode = talloc_get_type(node->transport_data,
 176                                                       struct ctdb_tcp_node);
 177         struct ctdb_context *ctdb = node->ctdb;
 178         ctdb_sock_addr sock_in;
 179         int sockin_size;
 180         int sockout_size;
 181         ctdb_sock_addr sock_out;
 182         int ret;
 183
 184         sock_out = node->address;
 185
 186         tnode->out_fd = socket(sock_out.sa.sa_family, SOCK_STREAM, IPPROTO_TCP);
 187         if (tnode->out_fd == -1) {
 188                 DBG_ERR("Failed to create socket\n");
 189                 goto failed;
 190         }
 191
 192         ret = set_blocking(tnode->out_fd, false);
 193         if (ret != 0) {
 194                 DBG_ERR("Failed to set socket non-blocking (%s)\n",
 195                         strerror(errno));
 196                 goto failed;
 197         }
 198
 199         set_close_on_exec(tnode->out_fd);
 200
 201         DBG_DEBUG("Created TCP SOCKET FD:%d\n", tnode->out_fd);
 202
 203         /* Bind our side of the socketpair to the same address we use to listen
 204          * on incoming CTDB traffic.
 205          * We must specify this address to make sure that the address we expose to
 206          * the remote side is actually routable in case CTDB traffic will run on
 207          * a dedicated non-routeable network.
 208          */
 209         sock_in = *ctdb->address;
 210
 211         /* AIX libs check to see if the socket address and length
 212            arguments are consistent with each other on calls like
 213            connect().   Can not get by with just sizeof(sock_in),
 214            need sizeof(sock_in.ip).
 215         */
 216         switch (sock_in.sa.sa_family) {
 217         case AF_INET:
 218                 sock_in.ip.sin_port = 0 /* Any port */;
 219                 sockin_size = sizeof(sock_in.ip);
 220                 sockout_size = sizeof(sock_out.ip);
 221                 break;
 222         case AF_INET6:
 223                 sock_in.ip6.sin6_port = 0 /* Any port */;
 224                 sockin_size = sizeof(sock_in.ip6);
 225                 sockout_size = sizeof(sock_out.ip6);
 226                 break;
 227         default:
 228                 DBG_ERR("Unknown address family %u\n", sock_in.sa.sa_family);
 229                 /* Can't happen to due to address parsing restrictions */
 230                 goto failed;
 231         }
 232
 233         ret = bind(tnode->out_fd, (struct sockaddr *)&sock_in, sockin_size);
 234         if (ret == -1) {
 235                 DBG_ERR("Failed to bind socket (%s)\n", strerror(errno));
 236                 goto failed;
 237         }
 238
 239         ret = connect(tnode->out_fd,
 240                       (struct sockaddr *)&sock_out,
 241                       sockout_size);
 242         if (ret != 0 && errno != EINPROGRESS) {
 243                 goto failed;
 244         }
 245
 246         /* non-blocking connect - wait for write event */
 247         tnode->connect_fde = tevent_add_fd(node->ctdb->ev,
 248                                            tnode,
 249                                            tnode->out_fd,
 250                                            TEVENT_FD_WRITE|TEVENT_FD_READ,
 251                                            ctdb_node_connect_write,
 252                                            node);
 253
 254         /* don't give it long to connect - retry in one second. This ensures
 255            that we find a node is up quickly (tcp normally backs off a syn reply
 256            delay by quite a lot) */
 257         tnode->connect_te = tevent_add_timer(ctdb->ev,
 258                                              tnode,
 259                                              timeval_current_ofs(1, 0),
 260                                              ctdb_tcp_node_connect_timeout,
 261                                              node);
 262
 263         return;
 264
 265 failed:
 266         ctdb_tcp_stop_outgoing(node);
 267         tnode->connect_te = tevent_add_timer(ctdb->ev,
 268                                              tnode,
 269                                              timeval_current_ofs(1, 0),
 270                                              ctdb_tcp_node_connect,
 271                                              node);
 272 }
 273
 274 void ctdb_tcp_node_connect(struct tevent_context *ev,
 275                            struct tevent_timer *te,
 276                            struct timeval t,
 277                            void *private_data)
 278 {
 279         struct ctdb_node *node = talloc_get_type_abort(private_data,
 280                                                        struct ctdb_node);
 281
 282         ctdb_tcp_start_outgoing(node);
 283 }
 284
 285 static void ctdb_tcp_node_connect_timeout(struct tevent_context *ev,
 286                                           struct tevent_timer *te,
 287                                           struct timeval t,
 288                                           void *private_data)
 289 {
 290         struct ctdb_node *node = talloc_get_type_abort(private_data,
 291                                                        struct ctdb_node);
 292
 293         ctdb_tcp_stop_outgoing(node);
 294         ctdb_tcp_start_outgoing(node);
 295 }
 296
 297 /*
 298   called when we get contacted by another node
 299   currently makes no attempt to check if the connection is really from a ctdb
 300   node in our cluster
 301 */
 302 static void ctdb_listen_event(struct tevent_context *ev, struct tevent_fd *fde,
 303                               uint16_t flags, void *private_data)
 304 {
 305         struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
 306         struct ctdb_tcp *ctcp = talloc_get_type(ctdb->transport_data,
 307                                                 struct ctdb_tcp);
 308         ctdb_sock_addr addr;
 309         socklen_t len;
 310         int fd;
 311         struct ctdb_node *node;
 312         struct ctdb_tcp_node *tnode;
 313         int one = 1;
 314         int ret;
 315
 316         memset(&addr, 0, sizeof(addr));
 317         len = sizeof(addr);
 318         fd = accept(ctcp->listen_fd, (struct sockaddr *)&addr, &len);
 319         if (fd == -1) return;
 320         smb_set_close_on_exec(fd);
 321
 322         node = ctdb_ip_to_node(ctdb, &addr);
 323         if (node == NULL) {
 324                 char *t = ctdb_sock_addr_to_string(ctcp, &addr, true);
 325                 if (t == NULL) {
 326                         DBG_ERR("Refused connection from unparsable node\n");
 327                         goto failed;
 328                 }
 329
 330                 D_ERR("Refused connection from unknown node %s\n", t);
 331                 talloc_free(t);
 332                 goto failed;
 333         }
 334
 335         tnode = talloc_get_type_abort(node->transport_data,
 336                                       struct ctdb_tcp_node);
 337         if (tnode == NULL) {
 338                 /* This can't happen - see ctdb_tcp_initialise() */
 339                 DBG_ERR("INTERNAL ERROR setting up connection from node %s\n",
 340                         node->name);
 341                 goto failed;
 342         }
 343
 344         if (tnode->in_queue != NULL) {
 345                 DBG_ERR("Incoming queue active, rejecting connection from %s\n",
 346                         node->name);
 347                 goto failed;
 348         }
 349
 350         ret = set_blocking(fd, false);
 351         if (ret != 0) {
 352                 DBG_ERR("Failed to set socket non-blocking (%s)\n",
 353                         strerror(errno));
 354                 goto failed;
 355         }
 356
 357         set_close_on_exec(fd);
 358
 359         DBG_DEBUG("Created SOCKET FD:%d to incoming ctdb connection\n", fd);
 360
 361         ret = setsockopt(fd,
 362                          SOL_SOCKET,
 363                          SO_KEEPALIVE,
 364                          (char *)&one,
 365                          sizeof(one));
 366         if (ret == -1) {
 367                 DBG_WARNING("Failed to set KEEPALIVE on fd - %s\n",
 368                             strerror(errno));
 369         }
 370
 371         tnode->in_queue = ctdb_queue_setup(ctdb,
 372                                            tnode,
 373                                            fd,
 374                                            CTDB_TCP_ALIGNMENT,
 375                                            ctdb_tcp_read_cb,
 376                                            node,
 377                                            "ctdbd-%s",
 378                                            node->name);
 379         if (tnode->in_queue == NULL) {
 380                 DBG_ERR("Failed to set up incoming queue\n");
 381                 goto failed;
 382         }
 383
 384        /*
 385         * Mark the connecting node as connected, but only if the
 386         * corresponding outbound connected is also up
 387         */
 388         if (tnode->out_queue != NULL) {
 389                 node->ctdb->upcalls->node_connected(node);
 390         }
 391
 392         return;
 393
 394 failed:
 395         close(fd);
 396  }
 397
 398 static int ctdb_tcp_listen_addr(struct ctdb_context *ctdb,
 399                                 ctdb_sock_addr *addr,
 400                                 bool strict)
 401 {
 402         struct ctdb_tcp *ctcp = talloc_get_type_abort(
 403                 ctdb->transport_data, struct ctdb_tcp);
 404         ctdb_sock_addr sock;
 405         int sock_size;
 406         int one = 1;
 407         struct tevent_fd *fde = NULL;
 408         int ret;
 409
 410         sock = *addr;
 411         ctcp->listen_fd = -1;
 412
 413         switch (sock.sa.sa_family) {
 414         case AF_INET:
 415                 sock_size = sizeof(sock.ip);
 416                 break;
 417         case AF_INET6:
 418                 sock_size = sizeof(sock.ip6);
 419                 break;
 420         default:
 421                 DBG_ERR("Unknown family %u\n", sock.sa.sa_family);
 422                 goto failed;
 423         }
 424
 425         ctcp->listen_fd = socket(sock.sa.sa_family, SOCK_STREAM, IPPROTO_TCP);
 426         if (ctcp->listen_fd == -1) {
 427                 DBG_ERR("Socket failed - %s (%d)\n", strerror(errno), errno);
 428                 goto failed;
 429         }
 430
 431         set_close_on_exec(ctcp->listen_fd);
 432
 433         ret = setsockopt(ctcp->listen_fd,
 434                          SOL_SOCKET,
 435                          SO_REUSEADDR,
 436                          (char *)&one,
 437                          sizeof(one));
 438         if (ret == -1) {
 439                 DBG_WARNING("Failed to set REUSEADDR on fd - %s (%d)\n",
 440                             strerror(errno),
 441                             errno);
 442         }
 443
 444         ret =bind(ctcp->listen_fd, (struct sockaddr * )&sock, sock_size);
 445         if (ret != 0) {
 446                 if (strict || errno != EADDRNOTAVAIL) {
 447                         DBG_ERR("Failed to bind() to socket - %s (%d)\n",
 448                                 strerror(errno),
 449                                 errno);
 450                 } else {
 451                         DBG_DEBUG("Failed to bind() to socket - %s (%d)\n",
 452                                   strerror(errno),
 453                                   errno);
 454                 }
 455                 goto failed;
 456         }
 457
 458         ret = listen(ctcp->listen_fd, 10);
 459         if (ret == -1) {
 460                 DBG_ERR("Failed to listen() on socket - %s (%d)\n",
 461                         strerror(errno),
 462                         errno);
 463                 goto failed;
 464         }
 465
 466         fde = tevent_add_fd(ctdb->ev,
 467                             ctcp,
 468                             ctcp->listen_fd,
 469                             TEVENT_FD_READ,
 470                             ctdb_listen_event,
 471                             ctdb);
 472         tevent_fd_set_auto_close(fde);
 473
 474         return 0;
 475
 476 failed:
 477         if (ctcp->listen_fd != -1) {
 478                 close(ctcp->listen_fd);
 479                 ctcp->listen_fd = -1;
 480         }
 481         return -1;
 482 }
 483
 484 /*
 485   automatically find which address to listen on
 486 */
 487 static int ctdb_tcp_listen_automatic(struct ctdb_context *ctdb)
 488 {
 489         int lock_fd;
 490         unsigned int i;
 491         char *lock_path = NULL;
 492         struct flock lock;
 493         int ret;
 494
 495         /*
 496          * If there are no nodes, then it won't be possible to find
 497          * the first one.  Log a failure and short circuit the whole
 498          * process.
 499          */
 500         if (ctdb->num_nodes == 0) {
 501                 DEBUG(DEBUG_CRIT,("No nodes available to attempt bind to - is the nodes file empty?\n"));
 502                 return -1;
 503         }
 504
 505         /*
 506          * In order to ensure that we don't get two nodes with the
 507          * same address, we must make the bind() and listen() calls
 508          * atomic. The SO_REUSEADDR setsockopt only prevents double
 509          * binds if the first socket is in LISTEN state.
 510          */
 511         lock_path = path_rundir_append(ctdb, ".socket_lock");
 512         if (lock_path == NULL) {
 513                 DBG_ERR("Memory allocation error\n");
 514                 return -1;
 515         }
 516         lock_fd = open(lock_path, O_RDWR|O_CREAT, 0666);
 517         if (lock_fd == -1) {
 518                 DBG_ERR("Unable to open %s\n", lock_path);
 519                 talloc_free(lock_path);
 520                 return -1;
 521         }
 522
 523         lock.l_type = F_WRLCK;
 524         lock.l_whence = SEEK_SET;
 525         lock.l_start = 0;
 526         lock.l_len = 1;
 527         lock.l_pid = 0;
 528
 529         if (fcntl(lock_fd, F_SETLKW, &lock) != 0) {
 530                 DBG_ERR("Unable to lock %s\n", lock_path);
 531                 close(lock_fd);
 532                 talloc_free(lock_path);
 533                 return -1;
 534         }
 535         talloc_free(lock_path);
 536
 537         for (i=0; i < ctdb->num_nodes; i++) {
 538                 if (ctdb->nodes[i]->flags & NODE_FLAGS_DELETED) {
 539                         continue;
 540                 }
 541
 542                 ret = ctdb_tcp_listen_addr(ctdb,
 543                                            &ctdb->nodes[i]->address,
 544                                            false);
 545                 if (ret == 0) {
 546                         break;
 547                 }
 548         }
 549
 550         if (i == ctdb->num_nodes) {
 551                 D_ERR("Unable to bind to any node address - giving up\n");
 552                 return -1;
 553         }
 554
 555         ctdb->address = talloc_memdup(ctdb,
 556                                       &ctdb->nodes[i]->address,
 557                                       sizeof(ctdb_sock_addr));
 558         if (ctdb->address == NULL) {
 559                 DBG_ERR("Memory allocation error\n");
 560                 return -1;
 561         }
 562
 563         ctdb->name = talloc_strdup(ctdb, ctdb->nodes[i]->name);
 564         if (ctdb->name == NULL) {
 565                 DBG_ERR("Memory allocation error\n");
 566                 return -1;
 567         }
 568
 569         D_INFO("ctdb chose network address %s\n", ctdb->name);
 570         return 0;
 571 }
 572
 573
 574 /*
 575   listen on our own address
 576 */
 577 int ctdb_tcp_listen(struct ctdb_context *ctdb)
 578 {
 579         int ret;
 580
 581         /* we can either auto-bind to the first available address, or we can
 582            use a specified address */
 583         if (!ctdb->address) {
 584                 ret = ctdb_tcp_listen_automatic(ctdb);
 585                 return ret;
 586         }
 587
 588         ret = ctdb_tcp_listen_addr(ctdb, ctdb->address, true);
 589         return ret;
 590 }