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 "includes.h"
  22 #include "lib/tevent/tevent.h"
  23 #include "lib/tdb/include/tdb.h"
  24 #include "system/network.h"
  25 #include "system/filesys.h"
  26 #include "../include/ctdb_private.h"
  27 #include "ctdb_tcp.h"
  28
  29 /*
  30   stop any connecting (established or pending) to a node
  31  */
  32 void ctdb_tcp_stop_connection(struct ctdb_node *node)
  33 {
  34         struct ctdb_tcp_node *tnode = talloc_get_type(
  35                 node->private_data, struct ctdb_tcp_node);
  36
  37         ctdb_queue_set_fd(tnode->out_queue, -1);
  38         talloc_free(tnode->connect_te);
  39         talloc_free(tnode->connect_fde);
  40         tnode->connect_fde = NULL;
  41         tnode->connect_te = NULL;
  42         if (tnode->fd != -1) {
  43                 close(tnode->fd);
  44                 tnode->fd = -1;
  45         }
  46 }
  47
  48
  49 /*
  50   called when a complete packet has come in - should not happen on this socket
  51   unless the other side closes the connection with RST or FIN
  52  */
  53 void ctdb_tcp_tnode_cb(uint8_t *data, size_t cnt, void *private_data)
  54 {
  55         struct ctdb_node *node = talloc_get_type(private_data, struct ctdb_node);
  56         struct ctdb_tcp_node *tnode = talloc_get_type(
  57                 node->private_data, struct ctdb_tcp_node);
  58
  59         if (data == NULL) {
  60                 node->ctdb->upcalls->node_dead(node);
  61         }
  62
  63         ctdb_tcp_stop_connection(node);
  64         tnode->connect_te = event_add_timed(node->ctdb->ev, tnode,
  65                                             timeval_current_ofs(3, 0),
  66                                             ctdb_tcp_node_connect, node);
  67 }
  68
  69 /*
  70   called when socket becomes writeable on connect
  71 */
  72 static void ctdb_node_connect_write(struct event_context *ev, struct fd_event *fde,
  73                                     uint16_t flags, void *private_data)
  74 {
  75         struct ctdb_node *node = talloc_get_type(private_data,
  76                                                  struct ctdb_node);
  77         struct ctdb_tcp_node *tnode = talloc_get_type(node->private_data,
  78                                                       struct ctdb_tcp_node);
  79         struct ctdb_context *ctdb = node->ctdb;
  80         int error = 0;
  81         socklen_t len = sizeof(error);
  82         int one = 1;
  83
  84         talloc_free(tnode->connect_te);
  85         tnode->connect_te = NULL;
  86
  87         if (getsockopt(tnode->fd, SOL_SOCKET, SO_ERROR, &error, &len) != 0 ||
  88             error != 0) {
  89                 ctdb_tcp_stop_connection(node);
  90                 tnode->connect_te = event_add_timed(ctdb->ev, tnode,
  91                                                     timeval_current_ofs(1, 0),
  92                                                     ctdb_tcp_node_connect, node);
  93                 return;
  94         }
  95
  96         talloc_free(tnode->connect_fde);
  97         tnode->connect_fde = NULL;
  98
  99         setsockopt(tnode->fd,IPPROTO_TCP,TCP_NODELAY,(char *)&one,sizeof(one));
 100         setsockopt(tnode->fd,SOL_SOCKET,SO_KEEPALIVE,(char *)&one,sizeof(one));
 101
 102         ctdb_queue_set_fd(tnode->out_queue, tnode->fd);
 103
 104         /* the queue subsystem now owns this fd */
 105         tnode->fd = -1;
 106 }
 107
 108
 109 static int ctdb_tcp_get_address(struct ctdb_context *ctdb,
 110                                 const char *address, ctdb_sock_addr *addr)
 111 {
 112         if (parse_ip(address, NULL, 0, addr) == 0) {
 113                 DEBUG(DEBUG_CRIT, (__location__ " Unparsable address : %s.\n", address));
 114                 return -1;
 115         }
 116         return 0;
 117 }
 118
 119 /*
 120   called when we should try and establish a tcp connection to a node
 121 */
 122 void ctdb_tcp_node_connect(struct event_context *ev, struct timed_event *te,
 123                            struct timeval t, void *private_data)
 124 {
 125         struct ctdb_node *node = talloc_get_type(private_data,
 126                                                  struct ctdb_node);
 127         struct ctdb_tcp_node *tnode = talloc_get_type(node->private_data,
 128                                                       struct ctdb_tcp_node);
 129         struct ctdb_context *ctdb = node->ctdb;
 130         ctdb_sock_addr sock_in;
 131         int sockin_size;
 132         int sockout_size;
 133         ctdb_sock_addr sock_out;
 134
 135         ctdb_tcp_stop_connection(node);
 136
 137         ZERO_STRUCT(sock_out);
 138 #ifdef HAVE_SOCK_SIN_LEN
 139         sock_out.ip.sin_len = sizeof(sock_out);
 140 #endif
 141         if (ctdb_tcp_get_address(ctdb, node->address.address, &sock_out) != 0) {
 142                 return;
 143         }
 144         switch (sock_out.sa.sa_family) {
 145         case AF_INET:
 146                 sock_out.ip.sin_port = htons(node->address.port);
 147                 break;
 148         case AF_INET6:
 149                 sock_out.ip6.sin6_port = htons(node->address.port);
 150                 break;
 151         default:
 152                 DEBUG(DEBUG_ERR, (__location__ " unknown family %u\n",
 153                         sock_out.sa.sa_family));
 154                 return;
 155         }
 156
 157         tnode->fd = socket(sock_out.sa.sa_family, SOCK_STREAM, IPPROTO_TCP);
 158         set_nonblocking(tnode->fd);
 159         set_close_on_exec(tnode->fd);
 160
 161         DEBUG(DEBUG_DEBUG, (__location__ " Created TCP SOCKET FD:%d\n", tnode->fd));
 162
 163         /* Bind our side of the socketpair to the same address we use to listen
 164          * on incoming CTDB traffic.
 165          * We must specify this address to make sure that the address we expose to
 166          * the remote side is actually routable in case CTDB traffic will run on
 167          * a dedicated non-routeable network.
 168          */
 169         ZERO_STRUCT(sock_in);
 170         if (ctdb_tcp_get_address(ctdb, ctdb->address.address, &sock_in) != 0) {
 171                 DEBUG(DEBUG_ERR, (__location__ " Failed to find our address. Failing bind.\n"));
 172                 close(tnode->fd);
 173                 return;
 174         }
 175
 176         /* AIX libs check to see if the socket address and length
 177            arguments are consistent with each other on calls like
 178            connect().   Can not get by with just sizeof(sock_in),
 179            need sizeof(sock_in.ip).
 180         */
 181         switch (sock_in.sa.sa_family) {
 182         case AF_INET:
 183                 sockin_size = sizeof(sock_in.ip);
 184                 sockout_size = sizeof(sock_out.ip);
 185                 break;
 186         case AF_INET6:
 187                 sockin_size = sizeof(sock_in.ip6);
 188                 sockout_size = sizeof(sock_out.ip6);
 189                 break;
 190         default:
 191                 DEBUG(DEBUG_ERR, (__location__ " unknown family %u\n",
 192                         sock_in.sa.sa_family));
 193                 close(tnode->fd);
 194                 return;
 195         }
 196 #ifdef HAVE_SOCK_SIN_LEN
 197         sock_in.ip.sin_len = sockin_size;
 198         sock_out.ip.sin_len = sockout_size;
 199 #endif
 200         bind(tnode->fd, (struct sockaddr *)&sock_in, sockin_size);
 201
 202         if (connect(tnode->fd, (struct sockaddr *)&sock_out, sockout_size) != 0 &&
 203             errno != EINPROGRESS) {
 204                 ctdb_tcp_stop_connection(node);
 205                 tnode->connect_te = event_add_timed(ctdb->ev, tnode,
 206                                                     timeval_current_ofs(1, 0),
 207                                                     ctdb_tcp_node_connect, node);
 208                 return;
 209         }
 210
 211         /* non-blocking connect - wait for write event */
 212         tnode->connect_fde = event_add_fd(node->ctdb->ev, tnode, tnode->fd,
 213                                           EVENT_FD_WRITE|EVENT_FD_READ,
 214                                           ctdb_node_connect_write, node);
 215
 216         /* don't give it long to connect - retry in one second. This ensures
 217            that we find a node is up quickly (tcp normally backs off a syn reply
 218            delay by quite a lot) */
 219         tnode->connect_te = event_add_timed(ctdb->ev, tnode, timeval_current_ofs(1, 0),
 220                                             ctdb_tcp_node_connect, node);
 221 }
 222
 223 /*
 224   called when we get contacted by another node
 225   currently makes no attempt to check if the connection is really from a ctdb
 226   node in our cluster
 227 */
 228 static void ctdb_listen_event(struct event_context *ev, struct fd_event *fde,
 229                               uint16_t flags, void *private_data)
 230 {
 231         struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
 232         struct ctdb_tcp *ctcp = talloc_get_type(ctdb->private_data, struct ctdb_tcp);
 233         ctdb_sock_addr addr;
 234         socklen_t len;
 235         int fd, nodeid;
 236         struct ctdb_incoming *in;
 237         int one = 1;
 238         const char *incoming_node;
 239
 240         memset(&addr, 0, sizeof(addr));
 241         len = sizeof(addr);
 242         fd = accept(ctcp->listen_fd, (struct sockaddr *)&addr, &len);
 243         if (fd == -1) return;
 244
 245         incoming_node = ctdb_addr_to_str(&addr);
 246         nodeid = ctdb_ip_to_nodeid(ctdb, incoming_node);
 247
 248         if (nodeid == -1) {
 249                 DEBUG(DEBUG_ERR, ("Refused connection from unknown node %s\n", incoming_node));
 250                 close(fd);
 251                 return;
 252         }
 253
 254         in = talloc_zero(ctcp, struct ctdb_incoming);
 255         in->fd = fd;
 256         in->ctdb = ctdb;
 257
 258         set_nonblocking(in->fd);
 259         set_close_on_exec(in->fd);
 260
 261         DEBUG(DEBUG_DEBUG, (__location__ " Created SOCKET FD:%d to incoming ctdb connection\n", fd));
 262
 263         setsockopt(in->fd,SOL_SOCKET,SO_KEEPALIVE,(char *)&one,sizeof(one));
 264
 265         in->queue = ctdb_queue_setup(ctdb, in, in->fd, CTDB_TCP_ALIGNMENT,
 266                                      ctdb_tcp_read_cb, in, "ctdbd-%s", incoming_node);
 267 }
 268
 269
 270 /*
 271   automatically find which address to listen on
 272 */
 273 static int ctdb_tcp_listen_automatic(struct ctdb_context *ctdb)
 274 {
 275         struct ctdb_tcp *ctcp = talloc_get_type(ctdb->private_data,
 276                                                 struct ctdb_tcp);
 277         ctdb_sock_addr sock;
 278         int lock_fd, i;
 279         const char *lock_path = "/tmp/.ctdb_socket_lock";
 280         struct flock lock;
 281         int one = 1;
 282         int sock_size;
 283         struct tevent_fd *fde;
 284
 285         /* in order to ensure that we don't get two nodes with the
 286            same adddress, we must make the bind() and listen() calls
 287            atomic. The SO_REUSEADDR setsockopt only prevents double
 288            binds if the first socket is in LISTEN state  */
 289         lock_fd = open(lock_path, O_RDWR|O_CREAT, 0666);
 290         if (lock_fd == -1) {
 291                 DEBUG(DEBUG_CRIT,("Unable to open %s\n", lock_path));
 292                 return -1;
 293         }
 294
 295         lock.l_type = F_WRLCK;
 296         lock.l_whence = SEEK_SET;
 297         lock.l_start = 0;
 298         lock.l_len = 1;
 299         lock.l_pid = 0;
 300
 301         if (fcntl(lock_fd, F_SETLKW, &lock) != 0) {
 302                 DEBUG(DEBUG_CRIT,("Unable to lock %s\n", lock_path));
 303                 close(lock_fd);
 304                 return -1;
 305         }
 306
 307         for (i=0; i < ctdb->num_nodes; i++) {
 308                 if (ctdb->nodes[i]->flags & NODE_FLAGS_DELETED) {
 309                         continue;
 310                 }
 311
 312                 /* if node_ip is specified we will only try to bind to that
 313                    ip.
 314                 */
 315                 if (ctdb->node_ip != NULL) {
 316                         if (strcmp(ctdb->node_ip, ctdb->nodes[i]->address.address)) {
 317                                 continue;
 318                         }
 319                 }
 320
 321                 ZERO_STRUCT(sock);
 322                 if (ctdb_tcp_get_address(ctdb,
 323                                 ctdb->nodes[i]->address.address,
 324                                 &sock) != 0) {
 325                         continue;
 326                 }
 327
 328                 switch (sock.sa.sa_family) {
 329                 case AF_INET:
 330                         sock.ip.sin_port = htons(ctdb->nodes[i]->address.port);
 331                         sock_size = sizeof(sock.ip);
 332                         break;
 333                 case AF_INET6:
 334                         sock.ip6.sin6_port = htons(ctdb->nodes[i]->address.port);
 335                         sock_size = sizeof(sock.ip6);
 336                         break;
 337                 default:
 338                         DEBUG(DEBUG_ERR, (__location__ " unknown family %u\n",
 339                                 sock.sa.sa_family));
 340                         continue;
 341                 }
 342 #ifdef HAVE_SOCK_SIN_LEN
 343                 sock.ip.sin_len = sock_size;
 344 #endif
 345
 346                 ctcp->listen_fd = socket(sock.sa.sa_family, SOCK_STREAM, IPPROTO_TCP);
 347                 if (ctcp->listen_fd == -1) {
 348                         ctdb_set_error(ctdb, "socket failed\n");
 349                         continue;
 350                 }
 351
 352                 set_close_on_exec(ctcp->listen_fd);
 353
 354                 setsockopt(ctcp->listen_fd,SOL_SOCKET,SO_REUSEADDR,(char *)&one,sizeof(one));
 355
 356                 if (bind(ctcp->listen_fd, (struct sockaddr * )&sock, sock_size) == 0) {
 357                         break;
 358                 }
 359
 360                 if (errno == EADDRNOTAVAIL) {
 361                         DEBUG(DEBUG_DEBUG,(__location__ " Failed to bind() to socket. %s(%d)\n",
 362                                         strerror(errno), errno));
 363                 } else {
 364                         DEBUG(DEBUG_ERR,(__location__ " Failed to bind() to socket. %s(%d)\n",
 365                                         strerror(errno), errno));
 366                 }
 367         }
 368
 369         if (i == ctdb->num_nodes) {
 370                 DEBUG(DEBUG_CRIT,("Unable to bind to any of the node addresses - giving up\n"));
 371                 goto failed;
 372         }
 373         ctdb->address.address = talloc_strdup(ctdb, ctdb->nodes[i]->address.address);
 374         ctdb->address.port    = ctdb->nodes[i]->address.port;
 375         ctdb->name = talloc_asprintf(ctdb, "%s:%u",
 376                                      ctdb->address.address,
 377                                      ctdb->address.port);
 378         ctdb->pnn = ctdb->nodes[i]->pnn;
 379         ctdb->nodes[i]->flags &= ~NODE_FLAGS_DISCONNECTED;
 380         DEBUG(DEBUG_INFO,("ctdb chose network address %s:%u pnn %u\n",
 381                  ctdb->address.address,
 382                  ctdb->address.port,
 383                  ctdb->pnn));
 384         /* do we start out in DISABLED mode? */
 385         if (ctdb->start_as_disabled != 0) {
 386                 DEBUG(DEBUG_INFO, ("This node is configured to start in DISABLED state\n"));
 387                 ctdb->nodes[i]->flags |= NODE_FLAGS_DISABLED;
 388         }
 389         /* do we start out in STOPPED mode? */
 390         if (ctdb->start_as_stopped != 0) {
 391                 DEBUG(DEBUG_INFO, ("This node is configured to start in STOPPED state\n"));
 392                 ctdb->nodes[i]->flags |= NODE_FLAGS_STOPPED;
 393         }
 394
 395         if (listen(ctcp->listen_fd, 10) == -1) {
 396                 goto failed;
 397         }
 398
 399         fde = event_add_fd(ctdb->ev, ctcp, ctcp->listen_fd, EVENT_FD_READ,
 400                            ctdb_listen_event, ctdb);
 401         tevent_fd_set_auto_close(fde);
 402
 403         close(lock_fd);
 404         return 0;
 405
 406 failed:
 407         close(lock_fd);
 408         close(ctcp->listen_fd);
 409         ctcp->listen_fd = -1;
 410         return -1;
 411 }
 412
 413
 414 /*
 415   listen on our own address
 416 */
 417 int ctdb_tcp_listen(struct ctdb_context *ctdb)
 418 {
 419         struct ctdb_tcp *ctcp = talloc_get_type(ctdb->private_data,
 420                                                 struct ctdb_tcp);
 421         ctdb_sock_addr sock;
 422         int sock_size;
 423         int one = 1;
 424         struct tevent_fd *fde;
 425
 426         /* we can either auto-bind to the first available address, or we can
 427            use a specified address */
 428         if (!ctdb->address.address) {
 429                 return ctdb_tcp_listen_automatic(ctdb);
 430         }
 431
 432         ZERO_STRUCT(sock);
 433         if (ctdb_tcp_get_address(ctdb, ctdb->address.address,
 434                                  &sock) != 0) {
 435                 goto failed;
 436         }
 437
 438         switch (sock.sa.sa_family) {
 439         case AF_INET:
 440                 sock.ip.sin_port = htons(ctdb->address.port);
 441                 sock_size = sizeof(sock.ip);
 442                 break;
 443         case AF_INET6:
 444                 sock.ip6.sin6_port = htons(ctdb->address.port);
 445                 sock_size = sizeof(sock.ip6);
 446                 break;
 447         default:
 448                 DEBUG(DEBUG_ERR, (__location__ " unknown family %u\n",
 449                         sock.sa.sa_family));
 450                 goto failed;
 451         }
 452 #ifdef HAVE_SOCK_SIN_LEN
 453         sock.ip.sin_len = sock_size;
 454 #endif
 455
 456         ctcp->listen_fd = socket(sock.sa.sa_family, SOCK_STREAM, IPPROTO_TCP);
 457         if (ctcp->listen_fd == -1) {
 458                 ctdb_set_error(ctdb, "socket failed\n");
 459                 return -1;
 460         }
 461
 462         set_close_on_exec(ctcp->listen_fd);
 463
 464         setsockopt(ctcp->listen_fd,SOL_SOCKET,SO_REUSEADDR,(char *)&one,sizeof(one));
 465
 466         if (bind(ctcp->listen_fd, (struct sockaddr * )&sock, sock_size) != 0) {
 467                 DEBUG(DEBUG_ERR,(__location__ " Failed to bind() to socket. %s(%d)\n", strerror(errno), errno));
 468                 goto failed;
 469         }
 470
 471         if (listen(ctcp->listen_fd, 10) == -1) {
 472                 goto failed;
 473         }
 474
 475         fde = event_add_fd(ctdb->ev, ctcp, ctcp->listen_fd, EVENT_FD_READ,
 476                      ctdb_listen_event, ctdb);
 477         tevent_fd_set_auto_close(fde);
 478
 479         return 0;
 480
 481 failed:
 482         if (ctcp->listen_fd != -1) {
 483                 close(ctcp->listen_fd);
 484         }
 485         ctcp->listen_fd = -1;
 486         return -1;
 487 }
 488