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