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