src/socket.c

   1 /*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
   2
   3 /***
   4   This file is part of systemd.
   5
   6   Copyright 2010 Lennart Poettering
   7
   8   systemd is free software; you can redistribute it and/or modify it
   9   under the terms of the GNU General Public License as published by
  10   the Free Software Foundation; either version 2 of the License, or
  11   (at your option) any later version.
  12
  13   systemd is distributed in the hope that it will be useful, but
  14   WITHOUT ANY WARRANTY; without even the implied warranty of
  15   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  16   General Public License for more details.
  17
  18   You should have received a copy of the GNU General Public License
  19   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  20 ***/
  21
  22 #include <sys/types.h>
  23 #include <sys/stat.h>
  24 #include <unistd.h>
  25 #include <errno.h>
  26 #include <fcntl.h>
  27 #include <sys/epoll.h>
  28 #include <signal.h>
  29 #include <arpa/inet.h>
  30 #include <mqueue.h>
  31
  32 #include "unit.h"
  33 #include "socket.h"
  34 #include "netinet/tcp.h"
  35 #include "log.h"
  36 #include "load-dropin.h"
  37 #include "load-fragment.h"
  38 #include "strv.h"
  39 #include "unit-name.h"
  40 #include "dbus-socket.h"
  41 #include "missing.h"
  42 #include "special.h"
  43 #include "bus-errors.h"
  44 #include "label.h"
  45 #include "exit-status.h"
  46 #include "def.h"
  47
  48 static const UnitActiveState state_translation_table[_SOCKET_STATE_MAX] = {
  49         [SOCKET_DEAD] = UNIT_INACTIVE,
  50         [SOCKET_START_PRE] = UNIT_ACTIVATING,
  51         [SOCKET_START_POST] = UNIT_ACTIVATING,
  52         [SOCKET_LISTENING] = UNIT_ACTIVE,
  53         [SOCKET_RUNNING] = UNIT_ACTIVE,
  54         [SOCKET_STOP_PRE] = UNIT_DEACTIVATING,
  55         [SOCKET_STOP_PRE_SIGTERM] = UNIT_DEACTIVATING,
  56         [SOCKET_STOP_PRE_SIGKILL] = UNIT_DEACTIVATING,
  57         [SOCKET_STOP_POST] = UNIT_DEACTIVATING,
  58         [SOCKET_FINAL_SIGTERM] = UNIT_DEACTIVATING,
  59         [SOCKET_FINAL_SIGKILL] = UNIT_DEACTIVATING,
  60         [SOCKET_FAILED] = UNIT_FAILED
  61 };
  62
  63 static void socket_init(Unit *u) {
  64         Socket *s = SOCKET(u);
  65
  66         assert(u);
  67         assert(u->meta.load_state == UNIT_STUB);
  68
  69         s->backlog = SOMAXCONN;
  70         s->timeout_usec = DEFAULT_TIMEOUT_USEC;
  71         s->directory_mode = 0755;
  72         s->socket_mode = 0666;
  73
  74         s->max_connections = 64;
  75
  76         s->priority = -1;
  77         s->ip_tos = -1;
  78         s->ip_ttl = -1;
  79         s->mark = -1;
  80
  81         exec_context_init(&s->exec_context);
  82         s->exec_context.std_output = u->meta.manager->default_std_output;
  83         s->exec_context.std_error = u->meta.manager->default_std_error;
  84
  85         s->control_command_id = _SOCKET_EXEC_COMMAND_INVALID;
  86 }
  87
  88 static void socket_unwatch_control_pid(Socket *s) {
  89         assert(s);
  90
  91         if (s->control_pid <= 0)
  92                 return;
  93
  94         unit_unwatch_pid(UNIT(s), s->control_pid);
  95         s->control_pid = 0;
  96 }
  97
  98 static void socket_done(Unit *u) {
  99         Socket *s = SOCKET(u);
 100         SocketPort *p;
 101         Meta *i;
 102
 103         assert(s);
 104
 105         while ((p = s->ports)) {
 106                 LIST_REMOVE(SocketPort, port, s->ports, p);
 107
 108                 if (p->fd >= 0) {
 109                         unit_unwatch_fd(UNIT(s), &p->fd_watch);
 110                         close_nointr_nofail(p->fd);
 111                 }
 112
 113                 free(p->path);
 114                 free(p);
 115         }
 116
 117         exec_context_done(&s->exec_context);
 118         exec_command_free_array(s->exec_command, _SOCKET_EXEC_COMMAND_MAX);
 119         s->control_command = NULL;
 120
 121         socket_unwatch_control_pid(s);
 122
 123         s->service = NULL;
 124
 125         free(s->tcp_congestion);
 126         s->tcp_congestion = NULL;
 127
 128         free(s->bind_to_device);
 129         s->bind_to_device = NULL;
 130
 131         unit_unwatch_timer(u, &s->timer_watch);
 132
 133         /* Make sure no service instance refers to us anymore. */
 134         LIST_FOREACH(units_per_type, i, u->meta.manager->units_per_type[UNIT_SERVICE]) {
 135                 Service *service = (Service *) i;
 136
 137                 if (service->accept_socket == s)
 138                         service->accept_socket = NULL;
 139
 140                 set_remove(service->configured_sockets, s);
 141         }
 142 }
 143
 144 static int socket_instantiate_service(Socket *s) {
 145         char *prefix, *name;
 146         int r;
 147         Unit *u;
 148
 149         assert(s);
 150
 151         /* This fills in s->service if it isn't filled in yet. For
 152          * Accept=yes sockets we create the next connection service
 153          * here. For Accept=no this is mostly a NOP since the service
 154          * is figured out at load time anyway. */
 155
 156         if (s->service)
 157                 return 0;
 158
 159         assert(s->accept);
 160
 161         if (!(prefix = unit_name_to_prefix(s->meta.id)))
 162                 return -ENOMEM;
 163
 164         r = asprintf(&name, "%s@%u.service", prefix, s->n_accepted);
 165         free(prefix);
 166
 167         if (r < 0)
 168                 return -ENOMEM;
 169
 170         r = manager_load_unit(s->meta.manager, name, NULL, NULL, &u);
 171         free(name);
 172
 173         if (r < 0)
 174                 return r;
 175
 176 #ifdef HAVE_SYSV_COMPAT
 177         if (SERVICE(u)->sysv_path) {
 178                 log_error("Using SysV services for socket activation is not supported. Refusing.");
 179                 return -ENOENT;
 180         }
 181 #endif
 182
 183         u->meta.no_gc = true;
 184         s->service = SERVICE(u);
 185         return 0;
 186 }
 187
 188 static bool have_non_accept_socket(Socket *s) {
 189         SocketPort *p;
 190
 191         assert(s);
 192
 193         if (!s->accept)
 194                 return true;
 195
 196         LIST_FOREACH(port, p, s->ports) {
 197
 198                 if (p->type != SOCKET_SOCKET)
 199                         return true;
 200
 201                 if (!socket_address_can_accept(&p->address))
 202                         return true;
 203         }
 204
 205         return false;
 206 }
 207
 208 static int socket_verify(Socket *s) {
 209         assert(s);
 210
 211         if (s->meta.load_state != UNIT_LOADED)
 212                 return 0;
 213
 214         if (!s->ports) {
 215                 log_error("%s lacks Listen setting. Refusing.", s->meta.id);
 216                 return -EINVAL;
 217         }
 218
 219         if (s->accept && have_non_accept_socket(s)) {
 220                 log_error("%s configured for accepting sockets, but sockets are non-accepting. Refusing.", s->meta.id);
 221                 return -EINVAL;
 222         }
 223
 224         if (s->accept && s->max_connections <= 0) {
 225                 log_error("%s's MaxConnection setting too small. Refusing.", s->meta.id);
 226                 return -EINVAL;
 227         }
 228
 229         if (s->accept && s->service) {
 230                 log_error("Explicit service configuration for accepting sockets not supported on %s. Refusing.", s->meta.id);
 231                 return -EINVAL;
 232         }
 233
 234         if (s->exec_context.pam_name && s->exec_context.kill_mode != KILL_CONTROL_GROUP) {
 235                 log_error("%s has PAM enabled. Kill mode must be set to 'control-group'. Refusing.", s->meta.id);
 236                 return -EINVAL;
 237         }
 238
 239         return 0;
 240 }
 241
 242 static bool socket_needs_mount(Socket *s, const char *prefix) {
 243         SocketPort *p;
 244
 245         assert(s);
 246
 247         LIST_FOREACH(port, p, s->ports) {
 248
 249                 if (p->type == SOCKET_SOCKET) {
 250                         if (socket_address_needs_mount(&p->address, prefix))
 251                                 return true;
 252                 } else if (p->type == SOCKET_FIFO || p->type == SOCKET_SPECIAL) {
 253                         if (path_startswith(p->path, prefix))
 254                                 return true;
 255                 }
 256         }
 257
 258         return false;
 259 }
 260
 261 int socket_add_one_mount_link(Socket *s, Mount *m) {
 262         int r;
 263
 264         assert(s);
 265         assert(m);
 266
 267         if (s->meta.load_state != UNIT_LOADED ||
 268             m->meta.load_state != UNIT_LOADED)
 269                 return 0;
 270
 271         if (!socket_needs_mount(s, m->where))
 272                 return 0;
 273
 274         if ((r = unit_add_two_dependencies(UNIT(s), UNIT_AFTER, UNIT_REQUIRES, UNIT(m), true)) < 0)
 275                 return r;
 276
 277         return 0;
 278 }
 279
 280 static int socket_add_mount_links(Socket *s) {
 281         Meta *other;
 282         int r;
 283
 284         assert(s);
 285
 286         LIST_FOREACH(units_per_type, other, s->meta.manager->units_per_type[UNIT_MOUNT])
 287                 if ((r = socket_add_one_mount_link(s, (Mount*) other)) < 0)
 288                         return r;
 289
 290         return 0;
 291 }
 292
 293 static int socket_add_device_link(Socket *s) {
 294         char *t;
 295         int r;
 296
 297         assert(s);
 298
 299         if (!s->bind_to_device)
 300                 return 0;
 301
 302         if (asprintf(&t, "/sys/subsystem/net/devices/%s", s->bind_to_device) < 0)
 303                 return -ENOMEM;
 304
 305         r = unit_add_node_link(UNIT(s), t, false);
 306         free(t);
 307
 308         return r;
 309 }
 310
 311 static int socket_add_default_dependencies(Socket *s) {
 312         int r;
 313         assert(s);
 314
 315         if (s->meta.manager->running_as == MANAGER_SYSTEM) {
 316                 if ((r = unit_add_dependency_by_name(UNIT(s), UNIT_BEFORE, SPECIAL_SOCKETS_TARGET, NULL, true)) < 0)
 317                         return r;
 318
 319                 if ((r = unit_add_two_dependencies_by_name(UNIT(s), UNIT_AFTER, UNIT_REQUIRES, SPECIAL_SYSINIT_TARGET, NULL, true)) < 0)
 320                         return r;
 321         }
 322
 323         return unit_add_two_dependencies_by_name(UNIT(s), UNIT_BEFORE, UNIT_CONFLICTS, SPECIAL_SHUTDOWN_TARGET, NULL, true);
 324 }
 325
 326 static int socket_load(Unit *u) {
 327         Socket *s = SOCKET(u);
 328         int r;
 329
 330         assert(u);
 331         assert(u->meta.load_state == UNIT_STUB);
 332
 333         if ((r = unit_load_fragment_and_dropin(u)) < 0)
 334                 return r;
 335
 336         /* This is a new unit? Then let's add in some extras */
 337         if (u->meta.load_state == UNIT_LOADED) {
 338
 339                 if (have_non_accept_socket(s)) {
 340
 341                         if (!s->service)
 342                                 if ((r = unit_load_related_unit(u, ".service", (Unit**) &s->service)) < 0)
 343                                         return r;
 344
 345                         if ((r = unit_add_dependency(u, UNIT_BEFORE, UNIT(s->service), true)) < 0)
 346                                 return r;
 347                 }
 348
 349                 if ((r = socket_add_mount_links(s)) < 0)
 350                         return r;
 351
 352                 if ((r = socket_add_device_link(s)) < 0)
 353                         return r;
 354
 355                 if ((r = unit_add_exec_dependencies(u, &s->exec_context)) < 0)
 356                         return r;
 357
 358                 if ((r = unit_add_default_cgroups(u)) < 0)
 359                         return r;
 360
 361                 if (s->meta.default_dependencies)
 362                         if ((r = socket_add_default_dependencies(s)) < 0)
 363                                 return r;
 364         }
 365
 366         return socket_verify(s);
 367 }
 368
 369 static const char* listen_lookup(int family, int type) {
 370
 371         if (family == AF_NETLINK)
 372                 return "ListenNetlink";
 373
 374         if (type == SOCK_STREAM)
 375                 return "ListenStream";
 376         else if (type == SOCK_DGRAM)
 377                 return "ListenDatagram";
 378         else if (type == SOCK_SEQPACKET)
 379                 return "ListenSequentialPacket";
 380
 381         assert_not_reached("Unknown socket type");
 382         return NULL;
 383 }
 384
 385 static void socket_dump(Unit *u, FILE *f, const char *prefix) {
 386
 387         SocketExecCommand c;
 388         Socket *s = SOCKET(u);
 389         SocketPort *p;
 390         const char *prefix2;
 391         char *p2;
 392
 393         assert(s);
 394         assert(f);
 395
 396         p2 = strappend(prefix, "\t");
 397         prefix2 = p2 ? p2 : prefix;
 398
 399         fprintf(f,
 400                 "%sSocket State: %s\n"
 401                 "%sBindIPv6Only: %s\n"
 402                 "%sBacklog: %u\n"
 403                 "%sSocketMode: %04o\n"
 404                 "%sDirectoryMode: %04o\n"
 405                 "%sKeepAlive: %s\n"
 406                 "%sFreeBind: %s\n"
 407                 "%sTCPCongestion: %s\n",
 408                 prefix, socket_state_to_string(s->state),
 409                 prefix, socket_address_bind_ipv6_only_to_string(s->bind_ipv6_only),
 410                 prefix, s->backlog,
 411                 prefix, s->socket_mode,
 412                 prefix, s->directory_mode,
 413                 prefix, yes_no(s->keep_alive),
 414                 prefix, yes_no(s->free_bind),
 415                 prefix, strna(s->tcp_congestion));
 416
 417         if (s->control_pid > 0)
 418                 fprintf(f,
 419                         "%sControl PID: %lu\n",
 420                         prefix, (unsigned long) s->control_pid);
 421
 422         if (s->bind_to_device)
 423                 fprintf(f,
 424                         "%sBindToDevice: %s\n",
 425                         prefix, s->bind_to_device);
 426
 427         if (s->accept)
 428                 fprintf(f,
 429                         "%sAccepted: %u\n"
 430                         "%sNConnections: %u\n"
 431                         "%sMaxConnections: %u\n",
 432                         prefix, s->n_accepted,
 433                         prefix, s->n_connections,
 434                         prefix, s->max_connections);
 435
 436         if (s->priority >= 0)
 437                 fprintf(f,
 438                         "%sPriority: %i\n",
 439                         prefix, s->priority);
 440
 441         if (s->receive_buffer > 0)
 442                 fprintf(f,
 443                         "%sReceiveBuffer: %zu\n",
 444                         prefix, s->receive_buffer);
 445
 446         if (s->send_buffer > 0)
 447                 fprintf(f,
 448                         "%sSendBuffer: %zu\n",
 449                         prefix, s->send_buffer);
 450
 451         if (s->ip_tos >= 0)
 452                 fprintf(f,
 453                         "%sIPTOS: %i\n",
 454                         prefix, s->ip_tos);
 455
 456         if (s->ip_ttl >= 0)
 457                 fprintf(f,
 458                         "%sIPTTL: %i\n",
 459                         prefix, s->ip_ttl);
 460
 461         if (s->pipe_size > 0)
 462                 fprintf(f,
 463                         "%sPipeSize: %zu\n",
 464                         prefix, s->pipe_size);
 465
 466         if (s->mark >= 0)
 467                 fprintf(f,
 468                         "%sMark: %i\n",
 469                         prefix, s->mark);
 470
 471         if (s->mq_maxmsg > 0)
 472                 fprintf(f,
 473                         "%sMessageQueueMaxMessages: %li\n",
 474                         prefix, s->mq_maxmsg);
 475
 476         if (s->mq_msgsize > 0)
 477                 fprintf(f,
 478                         "%sMessageQueueMessageSize: %li\n",
 479                         prefix, s->mq_msgsize);
 480
 481         LIST_FOREACH(port, p, s->ports) {
 482
 483                 if (p->type == SOCKET_SOCKET) {
 484                         const char *t;
 485                         int r;
 486                         char *k = NULL;
 487
 488                         if ((r = socket_address_print(&p->address, &k)) < 0)
 489                                 t = strerror(-r);
 490                         else
 491                                 t = k;
 492
 493                         fprintf(f, "%s%s: %s\n", prefix, listen_lookup(socket_address_family(&p->address), p->address.type), t);
 494                         free(k);
 495                 } else if (p->type == SOCKET_SPECIAL)
 496                         fprintf(f, "%sListenSpecial: %s\n", prefix, p->path);
 497                 else if (p->type == SOCKET_MQUEUE)
 498                         fprintf(f, "%sListenMessageQueue: %s\n", prefix, p->path);
 499                 else
 500                         fprintf(f, "%sListenFIFO: %s\n", prefix, p->path);
 501         }
 502
 503         exec_context_dump(&s->exec_context, f, prefix);
 504
 505         for (c = 0; c < _SOCKET_EXEC_COMMAND_MAX; c++) {
 506                 if (!s->exec_command[c])
 507                         continue;
 508
 509                 fprintf(f, "%s-> %s:\n",
 510                         prefix, socket_exec_command_to_string(c));
 511
 512                 exec_command_dump_list(s->exec_command[c], f, prefix2);
 513         }
 514
 515         free(p2);
 516 }
 517
 518 static int instance_from_socket(int fd, unsigned nr, char **instance) {
 519         socklen_t l;
 520         char *r;
 521         union {
 522                 struct sockaddr sa;
 523                 struct sockaddr_un un;
 524                 struct sockaddr_in in;
 525                 struct sockaddr_in6 in6;
 526                 struct sockaddr_storage storage;
 527         } local, remote;
 528
 529         assert(fd >= 0);
 530         assert(instance);
 531
 532         l = sizeof(local);
 533         if (getsockname(fd, &local.sa, &l) < 0)
 534                 return -errno;
 535
 536         l = sizeof(remote);
 537         if (getpeername(fd, &remote.sa, &l) < 0)
 538                 return -errno;
 539
 540         switch (local.sa.sa_family) {
 541
 542         case AF_INET: {
 543                 uint32_t
 544                         a = ntohl(local.in.sin_addr.s_addr),
 545                         b = ntohl(remote.in.sin_addr.s_addr);
 546
 547                 if (asprintf(&r,
 548                              "%u.%u.%u.%u:%u-%u.%u.%u.%u:%u",
 549                              a >> 24, (a >> 16) & 0xFF, (a >> 8) & 0xFF, a & 0xFF,
 550                              ntohs(local.in.sin_port),
 551                              b >> 24, (b >> 16) & 0xFF, (b >> 8) & 0xFF, b & 0xFF,
 552                              ntohs(remote.in.sin_port)) < 0)
 553                         return -ENOMEM;
 554
 555                 break;
 556         }
 557
 558         case AF_INET6: {
 559                 static const char ipv4_prefix[] = {
 560                         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0xFF, 0xFF
 561                 };
 562
 563                 if (memcmp(&local.in6.sin6_addr, ipv4_prefix, sizeof(ipv4_prefix)) == 0 &&
 564                     memcmp(&remote.in6.sin6_addr, ipv4_prefix, sizeof(ipv4_prefix)) == 0) {
 565                         const uint8_t
 566                                 *a = local.in6.sin6_addr.s6_addr+12,
 567                                 *b = remote.in6.sin6_addr.s6_addr+12;
 568
 569                         if (asprintf(&r,
 570                                      "%u.%u.%u.%u:%u-%u.%u.%u.%u:%u",
 571                                      a[0], a[1], a[2], a[3],
 572                                      ntohs(local.in6.sin6_port),
 573                                      b[0], b[1], b[2], b[3],
 574                                      ntohs(remote.in6.sin6_port)) < 0)
 575                                 return -ENOMEM;
 576                 } else {
 577                         char a[INET6_ADDRSTRLEN], b[INET6_ADDRSTRLEN];
 578
 579                         if (asprintf(&r,
 580                                      "%s:%u-%s:%u",
 581                                      inet_ntop(AF_INET6, &local.in6.sin6_addr, a, sizeof(a)),
 582                                      ntohs(local.in6.sin6_port),
 583                                      inet_ntop(AF_INET6, &remote.in6.sin6_addr, b, sizeof(b)),
 584                                      ntohs(remote.in6.sin6_port)) < 0)
 585                                 return -ENOMEM;
 586                 }
 587
 588                 break;
 589         }
 590
 591         case AF_UNIX: {
 592                 struct ucred ucred;
 593
 594                 l = sizeof(ucred);
 595                 if (getsockopt(fd, SOL_SOCKET, SO_PEERCRED, &ucred, &l) < 0)
 596                         return -errno;
 597
 598                 if (asprintf(&r,
 599                              "%u-%lu-%lu",
 600                              nr,
 601                              (unsigned long) ucred.pid,
 602                              (unsigned long) ucred.uid) < 0)
 603                         return -ENOMEM;
 604
 605                 break;
 606         }
 607
 608         default:
 609                 assert_not_reached("Unhandled socket type.");
 610         }
 611
 612         *instance = r;
 613         return 0;
 614 }
 615
 616 static void socket_close_fds(Socket *s) {
 617         SocketPort *p;
 618
 619         assert(s);
 620
 621         LIST_FOREACH(port, p, s->ports) {
 622                 if (p->fd < 0)
 623                         continue;
 624
 625                 unit_unwatch_fd(UNIT(s), &p->fd_watch);
 626                 close_nointr_nofail(p->fd);
 627
 628                 /* One little note: we should never delete any sockets
 629                  * in the file system here! After all some other
 630                  * process we spawned might still have a reference of
 631                  * this fd and wants to continue to use it. Therefore
 632                  * we delete sockets in the file system before we
 633                  * create a new one, not after we stopped using
 634                  * one! */
 635
 636                 p->fd = -1;
 637         }
 638 }
 639
 640 static void socket_apply_socket_options(Socket *s, int fd) {
 641         assert(s);
 642         assert(fd >= 0);
 643
 644         if (s->keep_alive) {
 645                 int b = s->keep_alive;
 646                 if (setsockopt(fd, SOL_SOCKET, SO_KEEPALIVE, &b, sizeof(b)) < 0)
 647                         log_warning("SO_KEEPALIVE failed: %m");
 648         }
 649
 650         if (s->priority >= 0)
 651                 if (setsockopt(fd, SOL_SOCKET, SO_PRIORITY, &s->priority, sizeof(s->priority)) < 0)
 652                         log_warning("SO_PRIORITY failed: %m");
 653
 654         if (s->receive_buffer > 0) {
 655                 int value = (int) s->receive_buffer;
 656                 if (setsockopt(fd, SOL_SOCKET, SO_RCVBUF, &value, sizeof(value)) < 0)
 657                         log_warning("SO_RCVBUF failed: %m");
 658         }
 659
 660         if (s->send_buffer > 0) {
 661                 int value = (int) s->send_buffer;
 662                 if (setsockopt(fd, SOL_SOCKET, SO_SNDBUF, &value, sizeof(value)) < 0)
 663                         log_warning("SO_SNDBUF failed: %m");
 664         }
 665
 666         if (s->mark >= 0)
 667                 if (setsockopt(fd, SOL_SOCKET, SO_MARK, &s->mark, sizeof(s->mark)) < 0)
 668                         log_warning("SO_MARK failed: %m");
 669
 670         if (s->ip_tos >= 0)
 671                 if (setsockopt(fd, IPPROTO_IP, IP_TOS, &s->ip_tos, sizeof(s->ip_tos)) < 0)
 672                         log_warning("IP_TOS failed: %m");
 673
 674         if (s->ip_ttl >= 0) {
 675                 int r, x;
 676
 677                 r = setsockopt(fd, IPPROTO_IP, IP_TTL, &s->ip_ttl, sizeof(s->ip_ttl));
 678
 679                 if (socket_ipv6_is_supported())
 680                         x = setsockopt(fd, IPPROTO_IPV6, IPV6_UNICAST_HOPS, &s->ip_ttl, sizeof(s->ip_ttl));
 681                 else {
 682                         x = -1;
 683                         errno = EAFNOSUPPORT;
 684                 }
 685
 686                 if (r < 0 && x < 0)
 687                         log_warning("IP_TTL/IPV6_UNICAST_HOPS failed: %m");
 688         }
 689
 690         if (s->tcp_congestion)
 691                 if (setsockopt(fd, SOL_TCP, TCP_CONGESTION, s->tcp_congestion, strlen(s->tcp_congestion)+1) < 0)
 692                         log_warning("TCP_CONGESTION failed: %m");
 693 }
 694
 695 static void socket_apply_fifo_options(Socket *s, int fd) {
 696         assert(s);
 697         assert(fd >= 0);
 698
 699         if (s->pipe_size > 0)
 700                 if (fcntl(fd, F_SETPIPE_SZ, s->pipe_size) < 0)
 701                         log_warning("F_SETPIPE_SZ: %m");
 702 }
 703
 704 static int fifo_address_create(
 705                 const char *path,
 706                 mode_t directory_mode,
 707                 mode_t socket_mode,
 708                 int *_fd) {
 709
 710         int fd = -1, r = 0;
 711         struct stat st;
 712         mode_t old_mask;
 713
 714         assert(path);
 715         assert(_fd);
 716
 717         mkdir_parents(path, directory_mode);
 718
 719         if ((r = label_fifofile_set(path)) < 0)
 720                 goto fail;
 721
 722         /* Enforce the right access mode for the fifo */
 723         old_mask = umask(~ socket_mode);
 724
 725         /* Include the original umask in our mask */
 726         umask(~socket_mode | old_mask);
 727
 728         r = mkfifo(path, socket_mode);
 729         umask(old_mask);
 730
 731         if (r < 0 && errno != EEXIST) {
 732                 r = -errno;
 733                 goto fail;
 734         }
 735
 736         if ((fd = open(path, O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK|O_NOFOLLOW)) < 0) {
 737                 r = -errno;
 738                 goto fail;
 739         }
 740
 741         label_file_clear();
 742
 743         if (fstat(fd, &st) < 0) {
 744                 r = -errno;
 745                 goto fail;
 746         }
 747
 748         if (!S_ISFIFO(st.st_mode) ||
 749             (st.st_mode & 0777) != (socket_mode & ~old_mask) ||
 750             st.st_uid != getuid() ||
 751             st.st_gid != getgid()) {
 752
 753                 r = -EEXIST;
 754                 goto fail;
 755         }
 756
 757         *_fd = fd;
 758         return 0;
 759
 760 fail:
 761         label_file_clear();
 762
 763         if (fd >= 0)
 764                 close_nointr_nofail(fd);
 765
 766         return r;
 767 }
 768
 769 static int special_address_create(
 770                 const char *path,
 771                 int *_fd) {
 772
 773         int fd = -1, r = 0;
 774         struct stat st;
 775
 776         assert(path);
 777         assert(_fd);
 778
 779         if ((fd = open(path, O_RDONLY|O_CLOEXEC|O_NOCTTY|O_NONBLOCK|O_NOFOLLOW)) < 0) {
 780                 r = -errno;
 781                 goto fail;
 782         }
 783
 784         if (fstat(fd, &st) < 0) {
 785                 r = -errno;
 786                 goto fail;
 787         }
 788
 789         /* Check whether this is a /proc, /sys or /dev file or char device */
 790         if (!S_ISREG(st.st_mode) && !S_ISCHR(st.st_mode)) {
 791                 r = -EEXIST;
 792                 goto fail;
 793         }
 794
 795         *_fd = fd;
 796         return 0;
 797
 798 fail:
 799         if (fd >= 0)
 800                 close_nointr_nofail(fd);
 801
 802         return r;
 803 }
 804
 805 static int mq_address_create(
 806                 const char *path,
 807                 mode_t mq_mode,
 808                 long maxmsg,
 809                 long msgsize,
 810                 int *_fd) {
 811
 812         int fd = -1, r = 0;
 813         struct stat st;
 814         mode_t old_mask;
 815         struct mq_attr _attr, *attr = NULL;
 816
 817         assert(path);
 818         assert(_fd);
 819
 820         if (maxmsg > 0 && msgsize > 0) {
 821                 zero(_attr);
 822                 _attr.mq_flags = O_NONBLOCK;
 823                 _attr.mq_maxmsg = maxmsg;
 824                 _attr.mq_msgsize = msgsize;
 825                 attr = &_attr;
 826         }
 827
 828         /* Enforce the right access mode for the mq */
 829         old_mask = umask(~ mq_mode);
 830
 831         /* Include the original umask in our mask */
 832         umask(~mq_mode | old_mask);
 833
 834         fd = mq_open(path, O_RDONLY|O_CLOEXEC|O_NONBLOCK|O_CREAT, mq_mode, attr);
 835         umask(old_mask);
 836
 837         if (fd < 0 && errno != EEXIST) {
 838                 r = -errno;
 839                 goto fail;
 840         }
 841
 842         if (fstat(fd, &st) < 0) {
 843                 r = -errno;
 844                 goto fail;
 845         }
 846
 847         if ((st.st_mode & 0777) != (mq_mode & ~old_mask) ||
 848             st.st_uid != getuid() ||
 849             st.st_gid != getgid()) {
 850
 851                 r = -EEXIST;
 852                 goto fail;
 853         }
 854
 855         *_fd = fd;
 856         return 0;
 857
 858 fail:
 859         if (fd >= 0)
 860                 close_nointr_nofail(fd);
 861
 862         return r;
 863 }
 864
 865 static int socket_open_fds(Socket *s) {
 866         SocketPort *p;
 867         int r;
 868         char *label = NULL;
 869         bool know_label = false;
 870
 871         assert(s);
 872
 873         LIST_FOREACH(port, p, s->ports) {
 874
 875                 if (p->fd >= 0)
 876                         continue;
 877
 878                 if (p->type == SOCKET_SOCKET) {
 879
 880                         if (!know_label) {
 881
 882                                 if ((r = socket_instantiate_service(s)) < 0)
 883                                         return r;
 884
 885                                 if (s->service && s->service->exec_command[SERVICE_EXEC_START])
 886                                         if ((r = label_get_socket_label_from_exe(s->service->exec_command[SERVICE_EXEC_START]->path, &label)) < 0) {
 887                                                 if (r != -EPERM)
 888                                                         return r;
 889                                         }
 890
 891                                 know_label = true;
 892                         }
 893
 894                         if ((r = socket_address_listen(
 895                                              &p->address,
 896                                              s->backlog,
 897                                              s->bind_ipv6_only,
 898                                              s->bind_to_device,
 899                                              s->free_bind,
 900                                              s->directory_mode,
 901                                              s->socket_mode,
 902                                              label,
 903                                              &p->fd)) < 0)
 904                                 goto rollback;
 905
 906                         socket_apply_socket_options(s, p->fd);
 907
 908                 } else  if (p->type == SOCKET_SPECIAL) {
 909
 910                         if ((r = special_address_create(
 911                                              p->path,
 912                                              &p->fd)) < 0)
 913                                 goto rollback;
 914
 915                 } else  if (p->type == SOCKET_FIFO) {
 916
 917                         if ((r = fifo_address_create(
 918                                              p->path,
 919                                              s->directory_mode,
 920                                              s->socket_mode,
 921                                              &p->fd)) < 0)
 922                                 goto rollback;
 923
 924                         socket_apply_fifo_options(s, p->fd);
 925                 } else if (p->type == SOCKET_MQUEUE) {
 926
 927                         if ((r = mq_address_create(
 928                                              p->path,
 929                                              s->socket_mode,
 930                                              s->mq_maxmsg,
 931                                              s->mq_msgsize,
 932                                              &p->fd)) < 0)
 933                                 goto rollback;
 934                 } else
 935                         assert_not_reached("Unknown port type");
 936         }
 937
 938         label_free(label);
 939         return 0;
 940
 941 rollback:
 942         socket_close_fds(s);
 943         label_free(label);
 944         return r;
 945 }
 946
 947 static void socket_unwatch_fds(Socket *s) {
 948         SocketPort *p;
 949
 950         assert(s);
 951
 952         LIST_FOREACH(port, p, s->ports) {
 953                 if (p->fd < 0)
 954                         continue;
 955
 956                 unit_unwatch_fd(UNIT(s), &p->fd_watch);
 957         }
 958 }
 959
 960 static int socket_watch_fds(Socket *s) {
 961         SocketPort *p;
 962         int r;
 963
 964         assert(s);
 965
 966         LIST_FOREACH(port, p, s->ports) {
 967                 if (p->fd < 0)
 968                         continue;
 969
 970                 p->fd_watch.socket_accept =
 971                         s->accept &&
 972                         p->type == SOCKET_SOCKET &&
 973                         socket_address_can_accept(&p->address);
 974
 975                 if ((r = unit_watch_fd(UNIT(s), p->fd, EPOLLIN, &p->fd_watch)) < 0)
 976                         goto fail;
 977         }
 978
 979         return 0;
 980
 981 fail:
 982         socket_unwatch_fds(s);
 983         return r;
 984 }
 985
 986 static void socket_set_state(Socket *s, SocketState state) {
 987         SocketState old_state;
 988         assert(s);
 989
 990         old_state = s->state;
 991         s->state = state;
 992
 993         if (state != SOCKET_START_PRE &&
 994             state != SOCKET_START_POST &&
 995             state != SOCKET_STOP_PRE &&
 996             state != SOCKET_STOP_PRE_SIGTERM &&
 997             state != SOCKET_STOP_PRE_SIGKILL &&
 998             state != SOCKET_STOP_POST &&
 999             state != SOCKET_FINAL_SIGTERM &&
1000             state != SOCKET_FINAL_SIGKILL) {
1001                 unit_unwatch_timer(UNIT(s), &s->timer_watch);
1002                 socket_unwatch_control_pid(s);
1003                 s->control_command = NULL;
1004                 s->control_command_id = _SOCKET_EXEC_COMMAND_INVALID;
1005         }
1006
1007         if (state != SOCKET_LISTENING)
1008                 socket_unwatch_fds(s);
1009
1010         if (state != SOCKET_START_POST &&
1011             state != SOCKET_LISTENING &&
1012             state != SOCKET_RUNNING &&
1013             state != SOCKET_STOP_PRE &&
1014             state != SOCKET_STOP_PRE_SIGTERM &&
1015             state != SOCKET_STOP_PRE_SIGKILL)
1016                 socket_close_fds(s);
1017
1018         if (state != old_state)
1019                 log_debug("%s changed %s -> %s",
1020                           s->meta.id,
1021                           socket_state_to_string(old_state),
1022                           socket_state_to_string(state));
1023
1024         unit_notify(UNIT(s), state_translation_table[old_state], state_translation_table[state], true);
1025 }
1026
1027 static int socket_coldplug(Unit *u) {
1028         Socket *s = SOCKET(u);
1029         int r;
1030
1031         assert(s);
1032         assert(s->state == SOCKET_DEAD);
1033
1034         if (s->deserialized_state != s->state) {
1035
1036                 if (s->deserialized_state == SOCKET_START_PRE ||
1037                     s->deserialized_state == SOCKET_START_POST ||
1038                     s->deserialized_state == SOCKET_STOP_PRE ||
1039                     s->deserialized_state == SOCKET_STOP_PRE_SIGTERM ||
1040                     s->deserialized_state == SOCKET_STOP_PRE_SIGKILL ||
1041                     s->deserialized_state == SOCKET_STOP_POST ||
1042                     s->deserialized_state == SOCKET_FINAL_SIGTERM ||
1043                     s->deserialized_state == SOCKET_FINAL_SIGKILL) {
1044
1045                         if (s->control_pid <= 0)
1046                                 return -EBADMSG;
1047
1048                         if ((r = unit_watch_pid(UNIT(s), s->control_pid)) < 0)
1049                                 return r;
1050
1051                         if ((r = unit_watch_timer(UNIT(s), s->timeout_usec, &s->timer_watch)) < 0)
1052                                 return r;
1053                 }
1054
1055                 if (s->deserialized_state == SOCKET_START_POST ||
1056                     s->deserialized_state == SOCKET_LISTENING ||
1057                     s->deserialized_state == SOCKET_RUNNING ||
1058                     s->deserialized_state == SOCKET_STOP_PRE ||
1059                     s->deserialized_state == SOCKET_STOP_PRE_SIGTERM ||
1060                     s->deserialized_state == SOCKET_STOP_PRE_SIGKILL)
1061                         if ((r = socket_open_fds(s)) < 0)
1062                                 return r;
1063
1064                 if (s->deserialized_state == SOCKET_LISTENING)
1065                         if ((r = socket_watch_fds(s)) < 0)
1066                                 return r;
1067
1068                 socket_set_state(s, s->deserialized_state);
1069         }
1070
1071         return 0;
1072 }
1073
1074 static int socket_spawn(Socket *s, ExecCommand *c, pid_t *_pid) {
1075         pid_t pid;
1076         int r;
1077         char **argv;
1078
1079         assert(s);
1080         assert(c);
1081         assert(_pid);
1082
1083         if ((r = unit_watch_timer(UNIT(s), s->timeout_usec, &s->timer_watch)) < 0)
1084                 goto fail;
1085
1086         if (!(argv = unit_full_printf_strv(UNIT(s), c->argv))) {
1087                 r = -ENOMEM;
1088                 goto fail;
1089         }
1090
1091         r = exec_spawn(c,
1092                        argv,
1093                        &s->exec_context,
1094                        NULL, 0,
1095                        s->meta.manager->environment,
1096                        true,
1097                        true,
1098                        true,
1099                        s->meta.manager->confirm_spawn,
1100                        s->meta.cgroup_bondings,
1101                        &pid);
1102
1103         strv_free(argv);
1104         if (r < 0)
1105                 goto fail;
1106
1107         if ((r = unit_watch_pid(UNIT(s), pid)) < 0)
1108                 /* FIXME: we need to do something here */
1109                 goto fail;
1110
1111         *_pid = pid;
1112
1113         return 0;
1114
1115 fail:
1116         unit_unwatch_timer(UNIT(s), &s->timer_watch);
1117
1118         return r;
1119 }
1120
1121 static void socket_enter_dead(Socket *s, bool success) {
1122         assert(s);
1123
1124         if (!success)
1125                 s->failure = true;
1126
1127         socket_set_state(s, s->failure ? SOCKET_FAILED : SOCKET_DEAD);
1128 }
1129
1130 static void socket_enter_signal(Socket *s, SocketState state, bool success);
1131
1132 static void socket_enter_stop_post(Socket *s, bool success) {
1133         int r;
1134         assert(s);
1135
1136         if (!success)
1137                 s->failure = true;
1138
1139         socket_unwatch_control_pid(s);
1140
1141         s->control_command_id = SOCKET_EXEC_STOP_POST;
1142
1143         if ((s->control_command = s->exec_command[SOCKET_EXEC_STOP_POST])) {
1144                 if ((r = socket_spawn(s, s->control_command, &s->control_pid)) < 0)
1145                         goto fail;
1146
1147                 socket_set_state(s, SOCKET_STOP_POST);
1148         } else
1149                 socket_enter_signal(s, SOCKET_FINAL_SIGTERM, true);
1150
1151         return;
1152
1153 fail:
1154         log_warning("%s failed to run 'stop-post' task: %s", s->meta.id, strerror(-r));
1155         socket_enter_signal(s, SOCKET_FINAL_SIGTERM, false);
1156 }
1157
1158 static void socket_enter_signal(Socket *s, SocketState state, bool success) {
1159         int r;
1160         Set *pid_set = NULL;
1161         bool wait_for_exit = false;
1162
1163         assert(s);
1164
1165         if (!success)
1166                 s->failure = true;
1167
1168         if (s->exec_context.kill_mode != KILL_NONE) {
1169                 int sig = (state == SOCKET_STOP_PRE_SIGTERM || state == SOCKET_FINAL_SIGTERM) ? s->exec_context.kill_signal : SIGKILL;
1170
1171                 if (s->control_pid > 0) {
1172                         if (kill_and_sigcont(s->control_pid, sig) < 0 && errno != ESRCH)
1173
1174                                 log_warning("Failed to kill control process %li: %m", (long) s->control_pid);
1175                         else
1176                                 wait_for_exit = true;
1177                 }
1178
1179                 if (s->exec_context.kill_mode == KILL_CONTROL_GROUP) {
1180
1181                         if (!(pid_set = set_new(trivial_hash_func, trivial_compare_func))) {
1182                                 r = -ENOMEM;
1183                                 goto fail;
1184                         }
1185
1186                         /* Exclude the control pid from being killed via the cgroup */
1187                         if (s->control_pid > 0)
1188                                 if ((r = set_put(pid_set, LONG_TO_PTR(s->control_pid))) < 0)
1189                                         goto fail;
1190
1191                         if ((r = cgroup_bonding_kill_list(s->meta.cgroup_bondings, sig, true, pid_set)) < 0) {
1192                                 if (r != -EAGAIN && r != -ESRCH && r != -ENOENT)
1193                                         log_warning("Failed to kill control group: %s", strerror(-r));
1194                         } else if (r > 0)
1195                                 wait_for_exit = true;
1196
1197                         set_free(pid_set);
1198                         pid_set = NULL;
1199                 }
1200         }
1201
1202         if (wait_for_exit) {
1203                 if ((r = unit_watch_timer(UNIT(s), s->timeout_usec, &s->timer_watch)) < 0)
1204                         goto fail;
1205
1206                 socket_set_state(s, state);
1207         } else if (state == SOCKET_STOP_PRE_SIGTERM || state == SOCKET_STOP_PRE_SIGKILL)
1208                 socket_enter_stop_post(s, true);
1209         else
1210                 socket_enter_dead(s, true);
1211
1212         return;
1213
1214 fail:
1215         log_warning("%s failed to kill processes: %s", s->meta.id, strerror(-r));
1216
1217         if (state == SOCKET_STOP_PRE_SIGTERM || state == SOCKET_STOP_PRE_SIGKILL)
1218                 socket_enter_stop_post(s, false);
1219         else
1220                 socket_enter_dead(s, false);
1221
1222         if (pid_set)
1223                 set_free(pid_set);
1224 }
1225
1226 static void socket_enter_stop_pre(Socket *s, bool success) {
1227         int r;
1228         assert(s);
1229
1230         if (!success)
1231                 s->failure = true;
1232
1233         socket_unwatch_control_pid(s);
1234
1235         s->control_command_id = SOCKET_EXEC_STOP_PRE;
1236
1237         if ((s->control_command = s->exec_command[SOCKET_EXEC_STOP_PRE])) {
1238                 if ((r = socket_spawn(s, s->control_command, &s->control_pid)) < 0)
1239                         goto fail;
1240
1241                 socket_set_state(s, SOCKET_STOP_PRE);
1242         } else
1243                 socket_enter_stop_post(s, true);
1244
1245         return;
1246
1247 fail:
1248         log_warning("%s failed to run 'stop-pre' task: %s", s->meta.id, strerror(-r));
1249         socket_enter_stop_post(s, false);
1250 }
1251
1252 static void socket_enter_listening(Socket *s) {
1253         int r;
1254         assert(s);
1255
1256         if ((r = socket_watch_fds(s)) < 0) {
1257                 log_warning("%s failed to watch sockets: %s", s->meta.id, strerror(-r));
1258                 goto fail;
1259         }
1260
1261         socket_set_state(s, SOCKET_LISTENING);
1262         return;
1263
1264 fail:
1265         socket_enter_stop_pre(s, false);
1266 }
1267
1268 static void socket_enter_start_post(Socket *s) {
1269         int r;
1270         assert(s);
1271
1272         if ((r = socket_open_fds(s)) < 0) {
1273                 log_warning("%s failed to listen on sockets: %s", s->meta.id, strerror(-r));
1274                 goto fail;
1275         }
1276
1277         socket_unwatch_control_pid(s);
1278
1279         s->control_command_id = SOCKET_EXEC_START_POST;
1280
1281         if ((s->control_command = s->exec_command[SOCKET_EXEC_START_POST])) {
1282                 if ((r = socket_spawn(s, s->control_command, &s->control_pid)) < 0) {
1283                         log_warning("%s failed to run 'start-post' task: %s", s->meta.id, strerror(-r));
1284                         goto fail;
1285                 }
1286
1287                 socket_set_state(s, SOCKET_START_POST);
1288         } else
1289                 socket_enter_listening(s);
1290
1291         return;
1292
1293 fail:
1294         socket_enter_stop_pre(s, false);
1295 }
1296
1297 static void socket_enter_start_pre(Socket *s) {
1298         int r;
1299         assert(s);
1300
1301         socket_unwatch_control_pid(s);
1302
1303         s->control_command_id = SOCKET_EXEC_START_PRE;
1304
1305         if ((s->control_command = s->exec_command[SOCKET_EXEC_START_PRE])) {
1306                 if ((r = socket_spawn(s, s->control_command, &s->control_pid)) < 0)
1307                         goto fail;
1308
1309                 socket_set_state(s, SOCKET_START_PRE);
1310         } else
1311                 socket_enter_start_post(s);
1312
1313         return;
1314
1315 fail:
1316         log_warning("%s failed to run 'start-pre' task: %s", s->meta.id, strerror(-r));
1317         socket_enter_dead(s, false);
1318 }
1319
1320 static void socket_enter_running(Socket *s, int cfd) {
1321         int r;
1322         DBusError error;
1323
1324         assert(s);
1325         dbus_error_init(&error);
1326
1327         /* We don't take connections anymore if we are supposed to
1328          * shut down anyway */
1329         if (unit_pending_inactive(UNIT(s))) {
1330                 log_debug("Suppressing connection request on %s since unit stop is scheduled.", s->meta.id);
1331
1332                 if (cfd >= 0)
1333                         close_nointr_nofail(cfd);
1334                 else  {
1335                         /* Flush all sockets by closing and reopening them */
1336                         socket_close_fds(s);
1337
1338                         if ((r = socket_watch_fds(s)) < 0) {
1339                                 log_warning("%s failed to watch sockets: %s", s->meta.id, strerror(-r));
1340                                 socket_enter_stop_pre(s, false);
1341                         }
1342                 }
1343
1344                 return;
1345         }
1346
1347         if (cfd < 0) {
1348                 bool pending = false;
1349                 Meta *i;
1350
1351                 /* If there's already a start pending don't bother to
1352                  * do anything */
1353                 LIST_FOREACH(units_per_type, i, s->meta.manager->units_per_type[UNIT_SERVICE]) {
1354                         Service *service = (Service *) i;
1355
1356                         if (!set_get(service->configured_sockets, s))
1357                                 continue;
1358
1359                         if (!unit_pending_active(UNIT(service)))
1360                                 continue;
1361
1362                         pending = true;
1363                         break;
1364                 }
1365
1366                 if (!pending)
1367                         if ((r = manager_add_job(s->meta.manager, JOB_START, UNIT(s->service), JOB_REPLACE, true, &error, NULL)) < 0)
1368                                 goto fail;
1369
1370                 socket_set_state(s, SOCKET_RUNNING);
1371         } else {
1372                 char *prefix, *instance = NULL, *name;
1373                 Service *service;
1374
1375                 if (s->n_connections >= s->max_connections) {
1376                         log_warning("Too many incoming connections (%u)", s->n_connections);
1377                         close_nointr_nofail(cfd);
1378                         return;
1379                 }
1380
1381                 if ((r = socket_instantiate_service(s)) < 0)
1382                         goto fail;
1383
1384                 if ((r = instance_from_socket(cfd, s->n_accepted, &instance)) < 0)
1385                         goto fail;
1386
1387                 if (!(prefix = unit_name_to_prefix(s->meta.id))) {
1388                         free(instance);
1389                         r = -ENOMEM;
1390                         goto fail;
1391                 }
1392
1393                 name = unit_name_build(prefix, instance, ".service");
1394                 free(prefix);
1395                 free(instance);
1396
1397                 if (!name) {
1398                         r = -ENOMEM;
1399                         goto fail;
1400                 }
1401
1402                 if ((r = unit_add_name(UNIT(s->service), name)) < 0) {
1403                         free(name);
1404                         goto fail;
1405                 }
1406
1407                 service = s->service;
1408                 s->service = NULL;
1409                 s->n_accepted ++;
1410
1411                 service->meta.no_gc = false;
1412
1413                 unit_choose_id(UNIT(service), name);
1414                 free(name);
1415
1416                 if ((r = service_set_socket_fd(service, cfd, s)) < 0)
1417                         goto fail;
1418
1419                 cfd = -1;
1420                 s->n_connections ++;
1421
1422                 if ((r = manager_add_job(s->meta.manager, JOB_START, UNIT(service), JOB_REPLACE, true, &error, NULL)) < 0)
1423                         goto fail;
1424
1425                 /* Notify clients about changed counters */
1426                 unit_add_to_dbus_queue(UNIT(s));
1427         }
1428
1429         return;
1430
1431 fail:
1432         log_warning("%s failed to queue socket startup job: %s", s->meta.id, bus_error(&error, r));
1433         socket_enter_stop_pre(s, false);
1434
1435         if (cfd >= 0)
1436                 close_nointr_nofail(cfd);
1437
1438         dbus_error_free(&error);
1439 }
1440
1441 static void socket_run_next(Socket *s, bool success) {
1442         int r;
1443
1444         assert(s);
1445         assert(s->control_command);
1446         assert(s->control_command->command_next);
1447
1448         if (!success)
1449                 s->failure = true;
1450
1451         socket_unwatch_control_pid(s);
1452
1453         s->control_command = s->control_command->command_next;
1454
1455         if ((r = socket_spawn(s, s->control_command, &s->control_pid)) < 0)
1456                 goto fail;
1457
1458         return;
1459
1460 fail:
1461         log_warning("%s failed to run next task: %s", s->meta.id, strerror(-r));
1462
1463         if (s->state == SOCKET_START_POST)
1464                 socket_enter_stop_pre(s, false);
1465         else if (s->state == SOCKET_STOP_POST)
1466                 socket_enter_dead(s, false);
1467         else
1468                 socket_enter_signal(s, SOCKET_FINAL_SIGTERM, false);
1469 }
1470
1471 static int socket_start(Unit *u) {
1472         Socket *s = SOCKET(u);
1473
1474         assert(s);
1475
1476         /* We cannot fulfill this request right now, try again later
1477          * please! */
1478         if (s->state == SOCKET_STOP_PRE ||
1479             s->state == SOCKET_STOP_PRE_SIGKILL ||
1480             s->state == SOCKET_STOP_PRE_SIGTERM ||
1481             s->state == SOCKET_STOP_POST ||
1482             s->state == SOCKET_FINAL_SIGTERM ||
1483             s->state == SOCKET_FINAL_SIGKILL)
1484                 return -EAGAIN;
1485
1486         if (s->state == SOCKET_START_PRE ||
1487             s->state == SOCKET_START_POST)
1488                 return 0;
1489
1490         /* Cannot run this without the service being around */
1491         if (s->service) {
1492                 if (s->service->meta.load_state != UNIT_LOADED) {
1493                         log_error("Socket service %s not loaded, refusing.", s->service->meta.id);
1494                         return -ENOENT;
1495                 }
1496
1497                 /* If the service is already active we cannot start the
1498                  * socket */
1499                 if (s->service->state != SERVICE_DEAD &&
1500                     s->service->state != SERVICE_FAILED &&
1501                     s->service->state != SERVICE_AUTO_RESTART) {
1502                         log_error("Socket service %s already active, refusing.", s->service->meta.id);
1503                         return -EBUSY;
1504                 }
1505
1506 #ifdef HAVE_SYSV_COMPAT
1507                 if (s->service->sysv_path) {
1508                         log_error("Using SysV services for socket activation is not supported. Refusing.");
1509                         return -ENOENT;
1510                 }
1511 #endif
1512         }
1513
1514         assert(s->state == SOCKET_DEAD || s->state == SOCKET_FAILED);
1515
1516         s->failure = false;
1517         socket_enter_start_pre(s);
1518         return 0;
1519 }
1520
1521 static int socket_stop(Unit *u) {
1522         Socket *s = SOCKET(u);
1523
1524         assert(s);
1525
1526         /* Already on it */
1527         if (s->state == SOCKET_STOP_PRE ||
1528             s->state == SOCKET_STOP_PRE_SIGTERM ||
1529             s->state == SOCKET_STOP_PRE_SIGKILL ||
1530             s->state == SOCKET_STOP_POST ||
1531             s->state == SOCKET_FINAL_SIGTERM ||
1532             s->state == SOCKET_FINAL_SIGKILL)
1533                 return 0;
1534
1535         /* If there's already something running we go directly into
1536          * kill mode. */
1537         if (s->state == SOCKET_START_PRE ||
1538             s->state == SOCKET_START_POST) {
1539                 socket_enter_signal(s, SOCKET_STOP_PRE_SIGTERM, true);
1540                 return -EAGAIN;
1541         }
1542
1543         assert(s->state == SOCKET_LISTENING || s->state == SOCKET_RUNNING);
1544
1545         socket_enter_stop_pre(s, true);
1546         return 0;
1547 }
1548
1549 static int socket_serialize(Unit *u, FILE *f, FDSet *fds) {
1550         Socket *s = SOCKET(u);
1551         SocketPort *p;
1552         int r;
1553
1554         assert(u);
1555         assert(f);
1556         assert(fds);
1557
1558         unit_serialize_item(u, f, "state", socket_state_to_string(s->state));
1559         unit_serialize_item(u, f, "failure", yes_no(s->failure));
1560         unit_serialize_item_format(u, f, "n-accepted", "%u", s->n_accepted);
1561
1562         if (s->control_pid > 0)
1563                 unit_serialize_item_format(u, f, "control-pid", "%lu", (unsigned long) s->control_pid);
1564
1565         if (s->control_command_id >= 0)
1566                 unit_serialize_item(u, f, "control-command", socket_exec_command_to_string(s->control_command_id));
1567
1568         LIST_FOREACH(port, p, s->ports) {
1569                 int copy;
1570
1571                 if (p->fd < 0)
1572                         continue;
1573
1574                 if ((copy = fdset_put_dup(fds, p->fd)) < 0)
1575                         return copy;
1576
1577                 if (p->type == SOCKET_SOCKET) {
1578                         char *t;
1579
1580                         if ((r = socket_address_print(&p->address, &t)) < 0)
1581                                 return r;
1582
1583                         if (socket_address_family(&p->address) == AF_NETLINK)
1584                                 unit_serialize_item_format(u, f, "netlink", "%i %s", copy, t);
1585                         else
1586                                 unit_serialize_item_format(u, f, "socket", "%i %i %s", copy, p->address.type, t);
1587                         free(t);
1588                 } else if (p->type == SOCKET_SPECIAL)
1589                         unit_serialize_item_format(u, f, "special", "%i %s", copy, p->path);
1590                 else {
1591                         assert(p->type == SOCKET_FIFO);
1592                         unit_serialize_item_format(u, f, "fifo", "%i %s", copy, p->path);
1593                 }
1594         }
1595
1596         return 0;
1597 }
1598
1599 static int socket_deserialize_item(Unit *u, const char *key, const char *value, FDSet *fds) {
1600         Socket *s = SOCKET(u);
1601
1602         assert(u);
1603         assert(key);
1604         assert(value);
1605         assert(fds);
1606
1607         if (streq(key, "state")) {
1608                 SocketState state;
1609
1610                 if ((state = socket_state_from_string(value)) < 0)
1611                         log_debug("Failed to parse state value %s", value);
1612                 else
1613                         s->deserialized_state = state;
1614         } else if (streq(key, "failure")) {
1615                 int b;
1616
1617                 if ((b = parse_boolean(value)) < 0)
1618                         log_debug("Failed to parse failure value %s", value);
1619                 else
1620                         s->failure = b || s->failure;
1621
1622         } else if (streq(key, "n-accepted")) {
1623                 unsigned k;
1624
1625                 if (safe_atou(value, &k) < 0)
1626                         log_debug("Failed to parse n-accepted value %s", value);
1627                 else
1628                         s->n_accepted += k;
1629         } else if (streq(key, "control-pid")) {
1630                 pid_t pid;
1631
1632                 if (parse_pid(value, &pid) < 0)
1633                         log_debug("Failed to parse control-pid value %s", value);
1634                 else
1635                         s->control_pid = pid;
1636         } else if (streq(key, "control-command")) {
1637                 SocketExecCommand id;
1638
1639                 if ((id = socket_exec_command_from_string(value)) < 0)
1640                         log_debug("Failed to parse exec-command value %s", value);
1641                 else {
1642                         s->control_command_id = id;
1643                         s->control_command = s->exec_command[id];
1644                 }
1645         } else if (streq(key, "fifo")) {
1646                 int fd, skip = 0;
1647                 SocketPort *p;
1648
1649                 if (sscanf(value, "%i %n", &fd, &skip) < 1 || fd < 0 || !fdset_contains(fds, fd))
1650                         log_debug("Failed to parse fifo value %s", value);
1651                 else {
1652
1653                         LIST_FOREACH(port, p, s->ports)
1654                                 if (p->type == SOCKET_FIFO &&
1655                                     streq_ptr(p->path, value+skip))
1656                                         break;
1657
1658                         if (p) {
1659                                 if (p->fd >= 0)
1660                                         close_nointr_nofail(p->fd);
1661                                 p->fd = fdset_remove(fds, fd);
1662                         }
1663                 }
1664
1665         } else if (streq(key, "special")) {
1666                 int fd, skip = 0;
1667                 SocketPort *p;
1668
1669                 if (sscanf(value, "%i %n", &fd, &skip) < 1 || fd < 0 || !fdset_contains(fds, fd))
1670                         log_debug("Failed to parse special value %s", value);
1671                 else {
1672
1673                         LIST_FOREACH(port, p, s->ports)
1674                                 if (p->type == SOCKET_SPECIAL &&
1675                                     streq_ptr(p->path, value+skip))
1676                                         break;
1677
1678                         if (p) {
1679                                 if (p->fd >= 0)
1680                                         close_nointr_nofail(p->fd);
1681                                 p->fd = fdset_remove(fds, fd);
1682                         }
1683                 }
1684
1685         } else if (streq(key, "socket")) {
1686                 int fd, type, skip = 0;
1687                 SocketPort *p;
1688
1689                 if (sscanf(value, "%i %i %n", &fd, &type, &skip) < 2 || fd < 0 || type < 0 || !fdset_contains(fds, fd))
1690                         log_debug("Failed to parse socket value %s", value);
1691                 else {
1692
1693                         LIST_FOREACH(port, p, s->ports)
1694                                 if (socket_address_is(&p->address, value+skip, type))
1695                                         break;
1696
1697                         if (p) {
1698                                 if (p->fd >= 0)
1699                                         close_nointr_nofail(p->fd);
1700                                 p->fd = fdset_remove(fds, fd);
1701                         }
1702                 }
1703
1704         } else if (streq(key, "netlink")) {
1705                 int fd, skip = 0;
1706                 SocketPort *p;
1707
1708                 if (sscanf(value, "%i %n", &fd, &skip) < 1 || fd < 0 || !fdset_contains(fds, fd))
1709                         log_debug("Failed to parse socket value %s", value);
1710                 else {
1711
1712                         LIST_FOREACH(port, p, s->ports)
1713                                 if (socket_address_is_netlink(&p->address, value+skip))
1714                                         break;
1715
1716                         if (p) {
1717                                 if (p->fd >= 0)
1718                                         close_nointr_nofail(p->fd);
1719                                 p->fd = fdset_remove(fds, fd);
1720                         }
1721                 }
1722
1723         } else
1724                 log_debug("Unknown serialization key '%s'", key);
1725
1726         return 0;
1727 }
1728
1729 static UnitActiveState socket_active_state(Unit *u) {
1730         assert(u);
1731
1732         return state_translation_table[SOCKET(u)->state];
1733 }
1734
1735 static const char *socket_sub_state_to_string(Unit *u) {
1736         assert(u);
1737
1738         return socket_state_to_string(SOCKET(u)->state);
1739 }
1740
1741 static bool socket_check_gc(Unit *u) {
1742         Socket *s = SOCKET(u);
1743
1744         assert(u);
1745
1746         return s->n_connections > 0;
1747 }
1748
1749 static void socket_fd_event(Unit *u, int fd, uint32_t events, Watch *w) {
1750         Socket *s = SOCKET(u);
1751         int cfd = -1;
1752
1753         assert(s);
1754         assert(fd >= 0);
1755
1756         if (s->state != SOCKET_LISTENING)
1757                 return;
1758
1759         log_debug("Incoming traffic on %s", u->meta.id);
1760
1761         if (events != EPOLLIN) {
1762
1763                 if (events & EPOLLHUP)
1764                         log_error("%s: Got POLLHUP on a listening socket. The service probably invoked shutdown() on it, and should better not do that.", u->meta.id);
1765                 else
1766                         log_error("%s: Got unexpected poll event (0x%x) on socket.", u->meta.id, events);
1767
1768                 goto fail;
1769         }
1770
1771         if (w->socket_accept) {
1772                 for (;;) {
1773
1774                         if ((cfd = accept4(fd, NULL, NULL, SOCK_NONBLOCK)) < 0) {
1775
1776                                 if (errno == EINTR)
1777                                         continue;
1778
1779                                 log_error("Failed to accept socket: %m");
1780                                 goto fail;
1781                         }
1782
1783                         break;
1784                 }
1785
1786                 socket_apply_socket_options(s, cfd);
1787         }
1788
1789         socket_enter_running(s, cfd);
1790         return;
1791
1792 fail:
1793         socket_enter_stop_pre(s, false);
1794 }
1795
1796 static void socket_sigchld_event(Unit *u, pid_t pid, int code, int status) {
1797         Socket *s = SOCKET(u);
1798         bool success;
1799
1800         assert(s);
1801         assert(pid >= 0);
1802
1803         if (pid != s->control_pid)
1804                 return;
1805
1806         s->control_pid = 0;
1807
1808         success = is_clean_exit(code, status);
1809
1810         if (s->control_command) {
1811                 exec_status_exit(&s->control_command->exec_status, pid, code, status, s->exec_context.utmp_id);
1812
1813                 if (s->control_command->ignore)
1814                         success = true;
1815         }
1816
1817         log_full(success ? LOG_DEBUG : LOG_NOTICE,
1818                  "%s control process exited, code=%s status=%i", u->meta.id, sigchld_code_to_string(code), status);
1819         s->failure = s->failure || !success;
1820
1821         if (s->control_command && s->control_command->command_next && success) {
1822                 log_debug("%s running next command for state %s", u->meta.id, socket_state_to_string(s->state));
1823                 socket_run_next(s, success);
1824         } else {
1825                 s->control_command = NULL;
1826                 s->control_command_id = _SOCKET_EXEC_COMMAND_INVALID;
1827
1828                 /* No further commands for this step, so let's figure
1829                  * out what to do next */
1830
1831                 log_debug("%s got final SIGCHLD for state %s", u->meta.id, socket_state_to_string(s->state));
1832
1833                 switch (s->state) {
1834
1835                 case SOCKET_START_PRE:
1836                         if (success)
1837                                 socket_enter_start_post(s);
1838                         else
1839                                 socket_enter_signal(s, SOCKET_FINAL_SIGTERM, false);
1840                         break;
1841
1842                 case SOCKET_START_POST:
1843                         if (success)
1844                                 socket_enter_listening(s);
1845                         else
1846                                 socket_enter_stop_pre(s, false);
1847                         break;
1848
1849                 case SOCKET_STOP_PRE:
1850                 case SOCKET_STOP_PRE_SIGTERM:
1851                 case SOCKET_STOP_PRE_SIGKILL:
1852                         socket_enter_stop_post(s, success);
1853                         break;
1854
1855                 case SOCKET_STOP_POST:
1856                 case SOCKET_FINAL_SIGTERM:
1857                 case SOCKET_FINAL_SIGKILL:
1858                         socket_enter_dead(s, success);
1859                         break;
1860
1861                 default:
1862                         assert_not_reached("Uh, control process died at wrong time.");
1863                 }
1864         }
1865
1866         /* Notify clients about changed exit status */
1867         unit_add_to_dbus_queue(u);
1868 }
1869
1870 static void socket_timer_event(Unit *u, uint64_t elapsed, Watch *w) {
1871         Socket *s = SOCKET(u);
1872
1873         assert(s);
1874         assert(elapsed == 1);
1875         assert(w == &s->timer_watch);
1876
1877         switch (s->state) {
1878
1879         case SOCKET_START_PRE:
1880                 log_warning("%s starting timed out. Terminating.", u->meta.id);
1881                 socket_enter_signal(s, SOCKET_FINAL_SIGTERM, false);
1882                 break;
1883
1884         case SOCKET_START_POST:
1885                 log_warning("%s starting timed out. Stopping.", u->meta.id);
1886                 socket_enter_stop_pre(s, false);
1887                 break;
1888
1889         case SOCKET_STOP_PRE:
1890                 log_warning("%s stopping timed out. Terminating.", u->meta.id);
1891                 socket_enter_signal(s, SOCKET_STOP_PRE_SIGTERM, false);
1892                 break;
1893
1894         case SOCKET_STOP_PRE_SIGTERM:
1895                 if (s->exec_context.send_sigkill) {
1896                         log_warning("%s stopping timed out. Killing.", u->meta.id);
1897                         socket_enter_signal(s, SOCKET_STOP_PRE_SIGKILL, false);
1898                 } else {
1899                         log_warning("%s stopping timed out. Skipping SIGKILL. Ignoring.", u->meta.id);
1900                         socket_enter_stop_post(s, false);
1901                 }
1902                 break;
1903
1904         case SOCKET_STOP_PRE_SIGKILL:
1905                 log_warning("%s still around after SIGKILL. Ignoring.", u->meta.id);
1906                 socket_enter_stop_post(s, false);
1907                 break;
1908
1909         case SOCKET_STOP_POST:
1910                 log_warning("%s stopping timed out (2). Terminating.", u->meta.id);
1911                 socket_enter_signal(s, SOCKET_FINAL_SIGTERM, false);
1912                 break;
1913
1914         case SOCKET_FINAL_SIGTERM:
1915                 if (s->exec_context.send_sigkill) {
1916                         log_warning("%s stopping timed out (2). Killing.", u->meta.id);
1917                         socket_enter_signal(s, SOCKET_FINAL_SIGKILL, false);
1918                 } else {
1919                         log_warning("%s stopping timed out (2). Skipping SIGKILL. Ignoring.", u->meta.id);
1920                         socket_enter_dead(s, false);
1921                 }
1922                 break;
1923
1924         case SOCKET_FINAL_SIGKILL:
1925                 log_warning("%s still around after SIGKILL (2). Entering failed mode.", u->meta.id);
1926                 socket_enter_dead(s, false);
1927                 break;
1928
1929         default:
1930                 assert_not_reached("Timeout at wrong time.");
1931         }
1932 }
1933
1934 int socket_collect_fds(Socket *s, int **fds, unsigned *n_fds) {
1935         int *rfds;
1936         unsigned rn_fds, k;
1937         SocketPort *p;
1938
1939         assert(s);
1940         assert(fds);
1941         assert(n_fds);
1942
1943         /* Called from the service code for requesting our fds */
1944
1945         rn_fds = 0;
1946         LIST_FOREACH(port, p, s->ports)
1947                 if (p->fd >= 0)
1948                         rn_fds++;
1949
1950         if (!(rfds = new(int, rn_fds)))
1951                 return -ENOMEM;
1952
1953         k = 0;
1954         LIST_FOREACH(port, p, s->ports)
1955                 if (p->fd >= 0)
1956                         rfds[k++] = p->fd;
1957
1958         assert(k == rn_fds);
1959
1960         *fds = rfds;
1961         *n_fds = rn_fds;
1962
1963         return 0;
1964 }
1965
1966 void socket_notify_service_dead(Socket *s) {
1967         assert(s);
1968
1969         /* The service is dead. Dang!
1970          *
1971          * This is strictly for one-instance-for-all-connections
1972          * services. */
1973
1974         if (s->state == SOCKET_RUNNING) {
1975                 log_debug("%s got notified about service death.", s->meta.id);
1976                 socket_enter_listening(s);
1977         }
1978 }
1979
1980 void socket_connection_unref(Socket *s) {
1981         assert(s);
1982
1983         /* The service is dead. Yay!
1984          *
1985          * This is strictly for one-instance-per-connection
1986          * services. */
1987
1988         assert(s->n_connections > 0);
1989         s->n_connections--;
1990
1991         log_debug("%s: One connection closed, %u left.", s->meta.id, s->n_connections);
1992 }
1993
1994 static void socket_reset_failed(Unit *u) {
1995         Socket *s = SOCKET(u);
1996
1997         assert(s);
1998
1999         if (s->state == SOCKET_FAILED)
2000                 socket_set_state(s, SOCKET_DEAD);
2001
2002         s->failure = false;
2003 }
2004
2005 static int socket_kill(Unit *u, KillWho who, KillMode mode, int signo, DBusError *error) {
2006         Socket *s = SOCKET(u);
2007         int r = 0;
2008         Set *pid_set = NULL;
2009
2010         assert(s);
2011
2012         if (who == KILL_MAIN) {
2013                 dbus_set_error(error, BUS_ERROR_NO_SUCH_PROCESS, "Socket units have no main processes");
2014                 return -EINVAL;
2015         }
2016
2017         if (s->control_pid <= 0 && who == KILL_CONTROL) {
2018                 dbus_set_error(error, BUS_ERROR_NO_SUCH_PROCESS, "No control process to kill");
2019                 return -ENOENT;
2020         }
2021
2022         if (s->control_pid > 0)
2023                 if (kill(s->control_pid, signo) < 0)
2024                         r = -errno;
2025
2026         if (mode == KILL_CONTROL_GROUP) {
2027                 int q;
2028
2029                 if (!(pid_set = set_new(trivial_hash_func, trivial_compare_func)))
2030                         return -ENOMEM;
2031
2032                 /* Exclude the control pid from being killed via the cgroup */
2033                 if (s->control_pid > 0)
2034                         if ((q = set_put(pid_set, LONG_TO_PTR(s->control_pid))) < 0) {
2035                                 r = q;
2036                                 goto finish;
2037                         }
2038
2039                 if ((q = cgroup_bonding_kill_list(s->meta.cgroup_bondings, signo, false, pid_set)) < 0)
2040                         if (r != -EAGAIN && r != -ESRCH && r != -ENOENT)
2041                                 r = q;
2042         }
2043
2044 finish:
2045         if (pid_set)
2046                 set_free(pid_set);
2047
2048         return r;
2049 }
2050
2051 static const char* const socket_state_table[_SOCKET_STATE_MAX] = {
2052         [SOCKET_DEAD] = "dead",
2053         [SOCKET_START_PRE] = "start-pre",
2054         [SOCKET_START_POST] = "start-post",
2055         [SOCKET_LISTENING] = "listening",
2056         [SOCKET_RUNNING] = "running",
2057         [SOCKET_STOP_PRE] = "stop-pre",
2058         [SOCKET_STOP_PRE_SIGTERM] = "stop-pre-sigterm",
2059         [SOCKET_STOP_PRE_SIGKILL] = "stop-pre-sigkill",
2060         [SOCKET_STOP_POST] = "stop-post",
2061         [SOCKET_FINAL_SIGTERM] = "final-sigterm",
2062         [SOCKET_FINAL_SIGKILL] = "final-sigkill",
2063         [SOCKET_FAILED] = "failed"
2064 };
2065
2066 DEFINE_STRING_TABLE_LOOKUP(socket_state, SocketState);
2067
2068 static const char* const socket_exec_command_table[_SOCKET_EXEC_COMMAND_MAX] = {
2069         [SOCKET_EXEC_START_PRE] = "StartPre",
2070         [SOCKET_EXEC_START_POST] = "StartPost",
2071         [SOCKET_EXEC_STOP_PRE] = "StopPre",
2072         [SOCKET_EXEC_STOP_POST] = "StopPost"
2073 };
2074
2075 DEFINE_STRING_TABLE_LOOKUP(socket_exec_command, SocketExecCommand);
2076
2077 const UnitVTable socket_vtable = {
2078         .suffix = ".socket",
2079
2080         .init = socket_init,
2081         .done = socket_done,
2082         .load = socket_load,
2083
2084         .kill = socket_kill,
2085
2086         .coldplug = socket_coldplug,
2087
2088         .dump = socket_dump,
2089
2090         .start = socket_start,
2091         .stop = socket_stop,
2092
2093         .serialize = socket_serialize,
2094         .deserialize_item = socket_deserialize_item,
2095
2096         .active_state = socket_active_state,
2097         .sub_state_to_string = socket_sub_state_to_string,
2098
2099         .check_gc = socket_check_gc,
2100
2101         .fd_event = socket_fd_event,
2102         .sigchld_event = socket_sigchld_event,
2103         .timer_event = socket_timer_event,
2104
2105         .reset_failed = socket_reset_failed,
2106
2107         .bus_interface = "org.freedesktop.systemd1.Socket",
2108         .bus_message_handler = bus_socket_message_handler,
2109         .bus_invalidating_properties =  bus_socket_invalidating_properties
2110 };