src/udev/udevd.c

   1 /*
   2  * Copyright (C) 2004-2012 Kay Sievers <kay.sievers@vrfy.org>
   3  * Copyright (C) 2004 Chris Friesen <chris_friesen@sympatico.ca>
   4  * Copyright (C) 2009 Canonical Ltd.
   5  * Copyright (C) 2009 Scott James Remnant <scott@netsplit.com>
   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 2 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 <stddef.h>
  22 #include <signal.h>
  23 #include <unistd.h>
  24 #include <errno.h>
  25 #include <stdio.h>
  26 #include <stdlib.h>
  27 #include <stdbool.h>
  28 #include <string.h>
  29 #include <ctype.h>
  30 #include <fcntl.h>
  31 #include <time.h>
  32 #include <getopt.h>
  33 #include <dirent.h>
  34 #include <sys/time.h>
  35 #include <sys/prctl.h>
  36 #include <sys/socket.h>
  37 #include <sys/un.h>
  38 #include <sys/signalfd.h>
  39 #include <sys/epoll.h>
  40 #include <sys/poll.h>
  41 #include <sys/wait.h>
  42 #include <sys/stat.h>
  43 #include <sys/ioctl.h>
  44 #include <sys/inotify.h>
  45 #include <sys/utsname.h>
  46
  47 #include "udev.h"
  48 #include "sd-daemon.h"
  49 #include "cgroup-util.h"
  50 #include "dev-setup.h"
  51
  52 static bool debug;
  53
  54 void udev_main_log(struct udev *udev, int priority,
  55                    const char *file, int line, const char *fn,
  56                    const char *format, va_list args)
  57 {
  58         log_metav(priority, file, line, fn, format, args);
  59 }
  60
  61 static struct udev_rules *rules;
  62 static struct udev_queue_export *udev_queue_export;
  63 static struct udev_ctrl *udev_ctrl;
  64 static struct udev_monitor *monitor;
  65 static int worker_watch[2] = { -1, -1 };
  66 static int fd_signal = -1;
  67 static int fd_ep = -1;
  68 static int fd_inotify = -1;
  69 static bool stop_exec_queue;
  70 static bool reload;
  71 static int children;
  72 static int children_max;
  73 static int exec_delay;
  74 static sigset_t sigmask_orig;
  75 static UDEV_LIST(event_list);
  76 static UDEV_LIST(worker_list);
  77 char *udev_cgroup;
  78 static bool udev_exit;
  79
  80 enum event_state {
  81         EVENT_UNDEF,
  82         EVENT_QUEUED,
  83         EVENT_RUNNING,
  84 };
  85
  86 struct event {
  87         struct udev_list_node node;
  88         struct udev *udev;
  89         struct udev_device *dev;
  90         enum event_state state;
  91         int exitcode;
  92         unsigned long long int delaying_seqnum;
  93         unsigned long long int seqnum;
  94         const char *devpath;
  95         size_t devpath_len;
  96         const char *devpath_old;
  97         dev_t devnum;
  98         bool is_block;
  99         int ifindex;
 100 };
 101
 102 static inline struct event *node_to_event(struct udev_list_node *node)
 103 {
 104         return container_of(node, struct event, node);
 105 }
 106
 107 static void event_queue_cleanup(struct udev *udev, enum event_state type);
 108
 109 enum worker_state {
 110         WORKER_UNDEF,
 111         WORKER_RUNNING,
 112         WORKER_IDLE,
 113         WORKER_KILLED,
 114 };
 115
 116 struct worker {
 117         struct udev_list_node node;
 118         struct udev *udev;
 119         int refcount;
 120         pid_t pid;
 121         struct udev_monitor *monitor;
 122         enum worker_state state;
 123         struct event *event;
 124         unsigned long long event_start_usec;
 125 };
 126
 127 /* passed from worker to main process */
 128 struct worker_message {
 129         pid_t pid;
 130         int exitcode;
 131 };
 132
 133 static inline struct worker *node_to_worker(struct udev_list_node *node)
 134 {
 135         return container_of(node, struct worker, node);
 136 }
 137
 138 static void event_queue_delete(struct event *event, bool export)
 139 {
 140         udev_list_node_remove(&event->node);
 141
 142         if (export) {
 143                 udev_queue_export_device_finished(udev_queue_export, event->dev);
 144                 log_debug("seq %llu done with %i\n", udev_device_get_seqnum(event->dev), event->exitcode);
 145         }
 146         udev_device_unref(event->dev);
 147         free(event);
 148 }
 149
 150 static struct worker *worker_ref(struct worker *worker)
 151 {
 152         worker->refcount++;
 153         return worker;
 154 }
 155
 156 static void worker_cleanup(struct worker *worker)
 157 {
 158         udev_list_node_remove(&worker->node);
 159         udev_monitor_unref(worker->monitor);
 160         children--;
 161         free(worker);
 162 }
 163
 164 static void worker_unref(struct worker *worker)
 165 {
 166         worker->refcount--;
 167         if (worker->refcount > 0)
 168                 return;
 169         log_debug("worker [%u] cleaned up\n", worker->pid);
 170         worker_cleanup(worker);
 171 }
 172
 173 static void worker_list_cleanup(struct udev *udev)
 174 {
 175         struct udev_list_node *loop, *tmp;
 176
 177         udev_list_node_foreach_safe(loop, tmp, &worker_list) {
 178                 struct worker *worker = node_to_worker(loop);
 179
 180                 worker_cleanup(worker);
 181         }
 182 }
 183
 184 static void worker_new(struct event *event)
 185 {
 186         struct udev *udev = event->udev;
 187         struct worker *worker;
 188         struct udev_monitor *worker_monitor;
 189         pid_t pid;
 190
 191         /* listen for new events */
 192         worker_monitor = udev_monitor_new_from_netlink(udev, NULL);
 193         if (worker_monitor == NULL)
 194                 return;
 195         /* allow the main daemon netlink address to send devices to the worker */
 196         udev_monitor_allow_unicast_sender(worker_monitor, monitor);
 197         udev_monitor_enable_receiving(worker_monitor);
 198
 199         worker = calloc(1, sizeof(struct worker));
 200         if (worker == NULL) {
 201                 udev_monitor_unref(worker_monitor);
 202                 return;
 203         }
 204         /* worker + event reference */
 205         worker->refcount = 2;
 206         worker->udev = udev;
 207
 208         pid = fork();
 209         switch (pid) {
 210         case 0: {
 211                 struct udev_device *dev = NULL;
 212                 int fd_monitor;
 213                 struct epoll_event ep_signal, ep_monitor;
 214                 sigset_t mask;
 215                 int rc = EXIT_SUCCESS;
 216
 217                 /* take initial device from queue */
 218                 dev = event->dev;
 219                 event->dev = NULL;
 220
 221                 free(worker);
 222                 worker_list_cleanup(udev);
 223                 event_queue_cleanup(udev, EVENT_UNDEF);
 224                 udev_queue_export_unref(udev_queue_export);
 225                 udev_monitor_unref(monitor);
 226                 udev_ctrl_unref(udev_ctrl);
 227                 close(fd_signal);
 228                 close(fd_ep);
 229                 close(worker_watch[READ_END]);
 230
 231                 sigfillset(&mask);
 232                 fd_signal = signalfd(-1, &mask, SFD_NONBLOCK|SFD_CLOEXEC);
 233                 if (fd_signal < 0) {
 234                         log_error("error creating signalfd %m\n");
 235                         rc = 2;
 236                         goto out;
 237                 }
 238
 239                 fd_ep = epoll_create1(EPOLL_CLOEXEC);
 240                 if (fd_ep < 0) {
 241                         log_error("error creating epoll fd: %m\n");
 242                         rc = 3;
 243                         goto out;
 244                 }
 245
 246                 memset(&ep_signal, 0, sizeof(struct epoll_event));
 247                 ep_signal.events = EPOLLIN;
 248                 ep_signal.data.fd = fd_signal;
 249
 250                 fd_monitor = udev_monitor_get_fd(worker_monitor);
 251                 memset(&ep_monitor, 0, sizeof(struct epoll_event));
 252                 ep_monitor.events = EPOLLIN;
 253                 ep_monitor.data.fd = fd_monitor;
 254
 255                 if (epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_signal, &ep_signal) < 0 ||
 256                     epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_monitor, &ep_monitor) < 0) {
 257                         log_error("fail to add fds to epoll: %m\n");
 258                         rc = 4;
 259                         goto out;
 260                 }
 261
 262                 /* request TERM signal if parent exits */
 263                 prctl(PR_SET_PDEATHSIG, SIGTERM);
 264
 265                 /* reset OOM score, we only protect the main daemon */
 266                 write_one_line_file("/proc/self/oom_score_adj", "0");
 267
 268                 for (;;) {
 269                         struct udev_event *udev_event;
 270                         struct worker_message msg;
 271                         int err;
 272
 273                         log_debug("seq %llu running\n", udev_device_get_seqnum(dev));
 274                         udev_event = udev_event_new(dev);
 275                         if (udev_event == NULL) {
 276                                 rc = 5;
 277                                 goto out;
 278                         }
 279
 280                         /* needed for SIGCHLD/SIGTERM in spawn() */
 281                         udev_event->fd_signal = fd_signal;
 282
 283                         if (exec_delay > 0)
 284                                 udev_event->exec_delay = exec_delay;
 285
 286                         /* apply rules, create node, symlinks */
 287                         err = udev_event_execute_rules(udev_event, rules, &sigmask_orig);
 288
 289                         if (err == 0)
 290                                 udev_event_execute_run(udev_event, &sigmask_orig);
 291
 292                         /* apply/restore inotify watch */
 293                         if (err == 0 && udev_event->inotify_watch) {
 294                                 udev_watch_begin(udev, dev);
 295                                 udev_device_update_db(dev);
 296                         }
 297
 298                         /* send processed event back to libudev listeners */
 299                         udev_monitor_send_device(worker_monitor, NULL, dev);
 300
 301                         /* send udevd the result of the event execution */
 302                         memset(&msg, 0, sizeof(struct worker_message));
 303                         if (err != 0)
 304                                 msg.exitcode = err;
 305                         msg.pid = getpid();
 306                         send(worker_watch[WRITE_END], &msg, sizeof(struct worker_message), 0);
 307
 308                         log_debug("seq %llu processed with %i\n", udev_device_get_seqnum(dev), err);
 309
 310                         udev_device_unref(dev);
 311                         dev = NULL;
 312
 313                         if (udev_event->sigterm) {
 314                                 udev_event_unref(udev_event);
 315                                 goto out;
 316                         }
 317
 318                         udev_event_unref(udev_event);
 319
 320                         /* wait for more device messages from main udevd, or term signal */
 321                         while (dev == NULL) {
 322                                 struct epoll_event ev[4];
 323                                 int fdcount;
 324                                 int i;
 325
 326                                 fdcount = epoll_wait(fd_ep, ev, ELEMENTSOF(ev), -1);
 327                                 if (fdcount < 0) {
 328                                         if (errno == EINTR)
 329                                                 continue;
 330                                         log_error("failed to poll: %m\n");
 331                                         goto out;
 332                                 }
 333
 334                                 for (i = 0; i < fdcount; i++) {
 335                                         if (ev[i].data.fd == fd_monitor && ev[i].events & EPOLLIN) {
 336                                                 dev = udev_monitor_receive_device(worker_monitor);
 337                                                 break;
 338                                         } else if (ev[i].data.fd == fd_signal && ev[i].events & EPOLLIN) {
 339                                                 struct signalfd_siginfo fdsi;
 340                                                 ssize_t size;
 341
 342                                                 size = read(fd_signal, &fdsi, sizeof(struct signalfd_siginfo));
 343                                                 if (size != sizeof(struct signalfd_siginfo))
 344                                                         continue;
 345                                                 switch (fdsi.ssi_signo) {
 346                                                 case SIGTERM:
 347                                                         goto out;
 348                                                 }
 349                                         }
 350                                 }
 351                         }
 352                 }
 353 out:
 354                 udev_device_unref(dev);
 355                 if (fd_signal >= 0)
 356                         close(fd_signal);
 357                 if (fd_ep >= 0)
 358                         close(fd_ep);
 359                 close(fd_inotify);
 360                 close(worker_watch[WRITE_END]);
 361                 udev_rules_unref(rules);
 362                 udev_builtin_exit(udev);
 363                 udev_monitor_unref(worker_monitor);
 364                 udev_unref(udev);
 365                 log_close();
 366                 exit(rc);
 367         }
 368         case -1:
 369                 udev_monitor_unref(worker_monitor);
 370                 event->state = EVENT_QUEUED;
 371                 free(worker);
 372                 log_error("fork of child failed: %m\n");
 373                 break;
 374         default:
 375                 /* close monitor, but keep address around */
 376                 udev_monitor_disconnect(worker_monitor);
 377                 worker->monitor = worker_monitor;
 378                 worker->pid = pid;
 379                 worker->state = WORKER_RUNNING;
 380                 worker->event_start_usec = now_usec();
 381                 worker->event = event;
 382                 event->state = EVENT_RUNNING;
 383                 udev_list_node_append(&worker->node, &worker_list);
 384                 children++;
 385                 log_debug("seq %llu forked new worker [%u]\n", udev_device_get_seqnum(event->dev), pid);
 386                 break;
 387         }
 388 }
 389
 390 static void event_run(struct event *event)
 391 {
 392         struct udev_list_node *loop;
 393
 394         udev_list_node_foreach(loop, &worker_list) {
 395                 struct worker *worker = node_to_worker(loop);
 396                 ssize_t count;
 397
 398                 if (worker->state != WORKER_IDLE)
 399                         continue;
 400
 401                 count = udev_monitor_send_device(monitor, worker->monitor, event->dev);
 402                 if (count < 0) {
 403                         log_error("worker [%u] did not accept message %zi (%m), kill it\n", worker->pid, count);
 404                         kill(worker->pid, SIGKILL);
 405                         worker->state = WORKER_KILLED;
 406                         continue;
 407                 }
 408                 worker_ref(worker);
 409                 worker->event = event;
 410                 worker->state = WORKER_RUNNING;
 411                 worker->event_start_usec = now_usec();
 412                 event->state = EVENT_RUNNING;
 413                 return;
 414         }
 415
 416         if (children >= children_max) {
 417                 if (children_max > 1)
 418                         log_debug("maximum number (%i) of children reached\n", children);
 419                 return;
 420         }
 421
 422         /* start new worker and pass initial device */
 423         worker_new(event);
 424 }
 425
 426 static int event_queue_insert(struct udev_device *dev)
 427 {
 428         struct event *event;
 429
 430         event = calloc(1, sizeof(struct event));
 431         if (event == NULL)
 432                 return -1;
 433
 434         event->udev = udev_device_get_udev(dev);
 435         event->dev = dev;
 436         event->seqnum = udev_device_get_seqnum(dev);
 437         event->devpath = udev_device_get_devpath(dev);
 438         event->devpath_len = strlen(event->devpath);
 439         event->devpath_old = udev_device_get_devpath_old(dev);
 440         event->devnum = udev_device_get_devnum(dev);
 441         event->is_block = (strcmp("block", udev_device_get_subsystem(dev)) == 0);
 442         event->ifindex = udev_device_get_ifindex(dev);
 443
 444         udev_queue_export_device_queued(udev_queue_export, dev);
 445         log_debug("seq %llu queued, '%s' '%s'\n", udev_device_get_seqnum(dev),
 446              udev_device_get_action(dev), udev_device_get_subsystem(dev));
 447
 448         event->state = EVENT_QUEUED;
 449         udev_list_node_append(&event->node, &event_list);
 450         return 0;
 451 }
 452
 453 static void worker_kill(struct udev *udev)
 454 {
 455         struct udev_list_node *loop;
 456
 457         udev_list_node_foreach(loop, &worker_list) {
 458                 struct worker *worker = node_to_worker(loop);
 459
 460                 if (worker->state == WORKER_KILLED)
 461                         continue;
 462
 463                 worker->state = WORKER_KILLED;
 464                 kill(worker->pid, SIGTERM);
 465         }
 466 }
 467
 468 /* lookup event for identical, parent, child device */
 469 static bool is_devpath_busy(struct event *event)
 470 {
 471         struct udev_list_node *loop;
 472         size_t common;
 473
 474         /* check if queue contains events we depend on */
 475         udev_list_node_foreach(loop, &event_list) {
 476                 struct event *loop_event = node_to_event(loop);
 477
 478                 /* we already found a later event, earlier can not block us, no need to check again */
 479                 if (loop_event->seqnum < event->delaying_seqnum)
 480                         continue;
 481
 482                 /* event we checked earlier still exists, no need to check again */
 483                 if (loop_event->seqnum == event->delaying_seqnum)
 484                         return true;
 485
 486                 /* found ourself, no later event can block us */
 487                 if (loop_event->seqnum >= event->seqnum)
 488                         break;
 489
 490                 /* check major/minor */
 491                 if (major(event->devnum) != 0 && event->devnum == loop_event->devnum && event->is_block == loop_event->is_block)
 492                         return true;
 493
 494                 /* check network device ifindex */
 495                 if (event->ifindex != 0 && event->ifindex == loop_event->ifindex)
 496                         return true;
 497
 498                 /* check our old name */
 499                 if (event->devpath_old != NULL && strcmp(loop_event->devpath, event->devpath_old) == 0) {
 500                         event->delaying_seqnum = loop_event->seqnum;
 501                         return true;
 502                 }
 503
 504                 /* compare devpath */
 505                 common = MIN(loop_event->devpath_len, event->devpath_len);
 506
 507                 /* one devpath is contained in the other? */
 508                 if (memcmp(loop_event->devpath, event->devpath, common) != 0)
 509                         continue;
 510
 511                 /* identical device event found */
 512                 if (loop_event->devpath_len == event->devpath_len) {
 513                         /* devices names might have changed/swapped in the meantime */
 514                         if (major(event->devnum) != 0 && (event->devnum != loop_event->devnum || event->is_block != loop_event->is_block))
 515                                 continue;
 516                         if (event->ifindex != 0 && event->ifindex != loop_event->ifindex)
 517                                 continue;
 518                         event->delaying_seqnum = loop_event->seqnum;
 519                         return true;
 520                 }
 521
 522                 /* parent device event found */
 523                 if (event->devpath[common] == '/') {
 524                         event->delaying_seqnum = loop_event->seqnum;
 525                         return true;
 526                 }
 527
 528                 /* child device event found */
 529                 if (loop_event->devpath[common] == '/') {
 530                         event->delaying_seqnum = loop_event->seqnum;
 531                         return true;
 532                 }
 533
 534                 /* no matching device */
 535                 continue;
 536         }
 537
 538         return false;
 539 }
 540
 541 static void event_queue_start(struct udev *udev)
 542 {
 543         struct udev_list_node *loop;
 544
 545         udev_list_node_foreach(loop, &event_list) {
 546                 struct event *event = node_to_event(loop);
 547
 548                 if (event->state != EVENT_QUEUED)
 549                         continue;
 550
 551                 /* do not start event if parent or child event is still running */
 552                 if (is_devpath_busy(event))
 553                         continue;
 554
 555                 event_run(event);
 556         }
 557 }
 558
 559 static void event_queue_cleanup(struct udev *udev, enum event_state match_type)
 560 {
 561         struct udev_list_node *loop, *tmp;
 562
 563         udev_list_node_foreach_safe(loop, tmp, &event_list) {
 564                 struct event *event = node_to_event(loop);
 565
 566                 if (match_type != EVENT_UNDEF && match_type != event->state)
 567                         continue;
 568
 569                 event_queue_delete(event, false);
 570         }
 571 }
 572
 573 static void worker_returned(int fd_worker)
 574 {
 575         for (;;) {
 576                 struct worker_message msg;
 577                 ssize_t size;
 578                 struct udev_list_node *loop;
 579
 580                 size = recv(fd_worker, &msg, sizeof(struct worker_message), MSG_DONTWAIT);
 581                 if (size != sizeof(struct worker_message))
 582                         break;
 583
 584                 /* lookup worker who sent the signal */
 585                 udev_list_node_foreach(loop, &worker_list) {
 586                         struct worker *worker = node_to_worker(loop);
 587
 588                         if (worker->pid != msg.pid)
 589                                 continue;
 590
 591                         /* worker returned */
 592                         if (worker->event) {
 593                                 worker->event->exitcode = msg.exitcode;
 594                                 event_queue_delete(worker->event, true);
 595                                 worker->event = NULL;
 596                         }
 597                         if (worker->state != WORKER_KILLED)
 598                                 worker->state = WORKER_IDLE;
 599                         worker_unref(worker);
 600                         break;
 601                 }
 602         }
 603 }
 604
 605 /* receive the udevd message from userspace */
 606 static struct udev_ctrl_connection *handle_ctrl_msg(struct udev_ctrl *uctrl)
 607 {
 608         struct udev *udev = udev_ctrl_get_udev(uctrl);
 609         struct udev_ctrl_connection *ctrl_conn;
 610         struct udev_ctrl_msg *ctrl_msg = NULL;
 611         const char *str;
 612         int i;
 613
 614         ctrl_conn = udev_ctrl_get_connection(uctrl);
 615         if (ctrl_conn == NULL)
 616                 goto out;
 617
 618         ctrl_msg = udev_ctrl_receive_msg(ctrl_conn);
 619         if (ctrl_msg == NULL)
 620                 goto out;
 621
 622         i = udev_ctrl_get_set_log_level(ctrl_msg);
 623         if (i >= 0) {
 624                 log_debug("udevd message (SET_LOG_PRIORITY) received, log_priority=%i\n", i);
 625                 log_set_max_level(i);
 626                 udev_set_log_priority(udev, i);
 627                 worker_kill(udev);
 628         }
 629
 630         if (udev_ctrl_get_stop_exec_queue(ctrl_msg) > 0) {
 631                 log_debug("udevd message (STOP_EXEC_QUEUE) received\n");
 632                 stop_exec_queue = true;
 633         }
 634
 635         if (udev_ctrl_get_start_exec_queue(ctrl_msg) > 0) {
 636                 log_debug("udevd message (START_EXEC_QUEUE) received\n");
 637                 stop_exec_queue = false;
 638         }
 639
 640         if (udev_ctrl_get_reload(ctrl_msg) > 0) {
 641                 log_debug("udevd message (RELOAD) received\n");
 642                 reload = true;
 643         }
 644
 645         str = udev_ctrl_get_set_env(ctrl_msg);
 646         if (str != NULL) {
 647                 char *key;
 648
 649                 key = strdup(str);
 650                 if (key != NULL) {
 651                         char *val;
 652
 653                         val = strchr(key, '=');
 654                         if (val != NULL) {
 655                                 val[0] = '\0';
 656                                 val = &val[1];
 657                                 if (val[0] == '\0') {
 658                                         log_debug("udevd message (ENV) received, unset '%s'\n", key);
 659                                         udev_add_property(udev, key, NULL);
 660                                 } else {
 661                                         log_debug("udevd message (ENV) received, set '%s=%s'\n", key, val);
 662                                         udev_add_property(udev, key, val);
 663                                 }
 664                         } else {
 665                                 log_error("wrong key format '%s'\n", key);
 666                         }
 667                         free(key);
 668                 }
 669                 worker_kill(udev);
 670         }
 671
 672         i = udev_ctrl_get_set_children_max(ctrl_msg);
 673         if (i >= 0) {
 674                 log_debug("udevd message (SET_MAX_CHILDREN) received, children_max=%i\n", i);
 675                 children_max = i;
 676         }
 677
 678         if (udev_ctrl_get_ping(ctrl_msg) > 0)
 679                 log_debug("udevd message (SYNC) received\n");
 680
 681         if (udev_ctrl_get_exit(ctrl_msg) > 0) {
 682                 log_debug("udevd message (EXIT) received\n");
 683                 udev_exit = true;
 684                 /* keep reference to block the client until we exit */
 685                 udev_ctrl_connection_ref(ctrl_conn);
 686         }
 687 out:
 688         udev_ctrl_msg_unref(ctrl_msg);
 689         return udev_ctrl_connection_unref(ctrl_conn);
 690 }
 691
 692 /* read inotify messages */
 693 static int handle_inotify(struct udev *udev)
 694 {
 695         int nbytes, pos;
 696         char *buf;
 697         struct inotify_event *ev;
 698
 699         if ((ioctl(fd_inotify, FIONREAD, &nbytes) < 0) || (nbytes <= 0))
 700                 return 0;
 701
 702         buf = malloc(nbytes);
 703         if (buf == NULL) {
 704                 log_error("error getting buffer for inotify\n");
 705                 return -1;
 706         }
 707
 708         nbytes = read(fd_inotify, buf, nbytes);
 709
 710         for (pos = 0; pos < nbytes; pos += sizeof(struct inotify_event) + ev->len) {
 711                 struct udev_device *dev;
 712
 713                 ev = (struct inotify_event *)(buf + pos);
 714                 dev = udev_watch_lookup(udev, ev->wd);
 715                 if (dev != NULL) {
 716                         log_debug("inotify event: %x for %s\n", ev->mask, udev_device_get_devnode(dev));
 717                         if (ev->mask & IN_CLOSE_WRITE) {
 718                                 char filename[UTIL_PATH_SIZE];
 719                                 int fd;
 720
 721                                 log_debug("device %s closed, synthesising 'change'\n", udev_device_get_devnode(dev));
 722                                 util_strscpyl(filename, sizeof(filename), udev_device_get_syspath(dev), "/uevent", NULL);
 723                                 fd = open(filename, O_WRONLY);
 724                                 if (fd >= 0) {
 725                                         if (write(fd, "change", 6) < 0)
 726                                                 log_debug("error writing uevent: %m\n");
 727                                         close(fd);
 728                                 }
 729                         }
 730                         if (ev->mask & IN_IGNORED)
 731                                 udev_watch_end(udev, dev);
 732
 733                         udev_device_unref(dev);
 734                 }
 735
 736         }
 737
 738         free(buf);
 739         return 0;
 740 }
 741
 742 static void handle_signal(struct udev *udev, int signo)
 743 {
 744         switch (signo) {
 745         case SIGINT:
 746         case SIGTERM:
 747                 udev_exit = true;
 748                 break;
 749         case SIGCHLD:
 750                 for (;;) {
 751                         pid_t pid;
 752                         int status;
 753                         struct udev_list_node *loop, *tmp;
 754
 755                         pid = waitpid(-1, &status, WNOHANG);
 756                         if (pid <= 0)
 757                                 break;
 758
 759                         udev_list_node_foreach_safe(loop, tmp, &worker_list) {
 760                                 struct worker *worker = node_to_worker(loop);
 761
 762                                 if (worker->pid != pid)
 763                                         continue;
 764                                 log_debug("worker [%u] exit\n", pid);
 765
 766                                 if (WIFEXITED(status)) {
 767                                         if (WEXITSTATUS(status) != 0)
 768                                                 log_error("worker [%u] exit with return code %i\n", pid, WEXITSTATUS(status));
 769                                 } else if (WIFSIGNALED(status)) {
 770                                         log_error("worker [%u] terminated by signal %i (%s)\n",
 771                                             pid, WTERMSIG(status), strsignal(WTERMSIG(status)));
 772                                 } else if (WIFSTOPPED(status)) {
 773                                         log_error("worker [%u] stopped\n", pid);
 774                                 } else if (WIFCONTINUED(status)) {
 775                                         log_error("worker [%u] continued\n", pid);
 776                                 } else {
 777                                         log_error("worker [%u] exit with status 0x%04x\n", pid, status);
 778                                 }
 779
 780                                 if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 781                                         if (worker->event) {
 782                                                 log_error("worker [%u] failed while handling '%s'\n",
 783                                                           pid, worker->event->devpath);
 784                                                 worker->event->exitcode = -32;
 785                                                 event_queue_delete(worker->event, true);
 786                                                 /* drop reference taken for state 'running' */
 787                                                 worker_unref(worker);
 788                                         }
 789                                 }
 790                                 worker_unref(worker);
 791                                 break;
 792                         }
 793                 }
 794                 break;
 795         case SIGHUP:
 796                 reload = true;
 797                 break;
 798         }
 799 }
 800
 801 static void static_dev_create_from_modules(struct udev *udev)
 802 {
 803         struct utsname kernel;
 804         char modules[UTIL_PATH_SIZE];
 805         char buf[4096];
 806         FILE *f;
 807
 808         uname(&kernel);
 809         util_strscpyl(modules, sizeof(modules), "/lib/modules/", kernel.release, "/modules.devname", NULL);
 810         f = fopen(modules, "r");
 811         if (f == NULL)
 812                 return;
 813
 814         while (fgets(buf, sizeof(buf), f) != NULL) {
 815                 char *s;
 816                 const char *modname;
 817                 const char *devname;
 818                 const char *devno;
 819                 int maj, min;
 820                 char type;
 821                 mode_t mode;
 822                 char filename[UTIL_PATH_SIZE];
 823
 824                 if (buf[0] == '#')
 825                         continue;
 826
 827                 modname = buf;
 828                 s = strchr(modname, ' ');
 829                 if (s == NULL)
 830                         continue;
 831                 s[0] = '\0';
 832
 833                 devname = &s[1];
 834                 s = strchr(devname, ' ');
 835                 if (s == NULL)
 836                         continue;
 837                 s[0] = '\0';
 838
 839                 devno = &s[1];
 840                 s = strchr(devno, ' ');
 841                 if (s == NULL)
 842                         s = strchr(devno, '\n');
 843                 if (s != NULL)
 844                         s[0] = '\0';
 845                 if (sscanf(devno, "%c%u:%u", &type, &maj, &min) != 3)
 846                         continue;
 847
 848                 if (type == 'c')
 849                         mode = S_IFCHR;
 850                 else if (type == 'b')
 851                         mode = S_IFBLK;
 852                 else
 853                         continue;
 854
 855                 util_strscpyl(filename, sizeof(filename), "/dev/", devname, NULL);
 856                 mkdir_parents_label(filename, 0755);
 857                 label_context_set(filename, mode);
 858                 log_debug("mknod '%s' %c%u:%u\n", filename, type, maj, min);
 859                 if (mknod(filename, mode, makedev(maj, min)) < 0 && errno == EEXIST)
 860                         utimensat(AT_FDCWD, filename, NULL, 0);
 861                 label_context_clear();
 862         }
 863
 864         fclose(f);
 865 }
 866
 867 static int mem_size_mb(void)
 868 {
 869         FILE *f;
 870         char buf[4096];
 871         long int memsize = -1;
 872
 873         f = fopen("/proc/meminfo", "r");
 874         if (f == NULL)
 875                 return -1;
 876
 877         while (fgets(buf, sizeof(buf), f) != NULL) {
 878                 long int value;
 879
 880                 if (sscanf(buf, "MemTotal: %ld kB", &value) == 1) {
 881                         memsize = value / 1024;
 882                         break;
 883                 }
 884         }
 885
 886         fclose(f);
 887         return memsize;
 888 }
 889
 890 static int convert_db(struct udev *udev)
 891 {
 892         char filename[UTIL_PATH_SIZE];
 893         FILE *f;
 894         struct udev_enumerate *udev_enumerate;
 895         struct udev_list_entry *list_entry;
 896
 897         /* current database */
 898         if (access("/run/udev/data", F_OK) >= 0)
 899                 return 0;
 900
 901         /* make sure we do not get here again */
 902         mkdir_parents("/run/udev/data", 0755);
 903         mkdir(filename, 0755);
 904
 905         /* old database */
 906         util_strscpyl(filename, sizeof(filename), "/dev/.udev/db", NULL);
 907         if (access(filename, F_OK) < 0)
 908                 return 0;
 909
 910         f = fopen("/dev/kmsg", "w");
 911         if (f != NULL) {
 912                 fprintf(f, "<30>systemd-udevd[%u]: converting old udev database\n", getpid());
 913                 fclose(f);
 914         }
 915
 916         udev_enumerate = udev_enumerate_new(udev);
 917         if (udev_enumerate == NULL)
 918                 return -1;
 919         udev_enumerate_scan_devices(udev_enumerate);
 920         udev_list_entry_foreach(list_entry, udev_enumerate_get_list_entry(udev_enumerate)) {
 921                 struct udev_device *device;
 922
 923                 device = udev_device_new_from_syspath(udev, udev_list_entry_get_name(list_entry));
 924                 if (device == NULL)
 925                         continue;
 926
 927                 /* try to find the old database for devices without a current one */
 928                 if (udev_device_read_db(device, NULL) < 0) {
 929                         bool have_db;
 930                         const char *id;
 931                         struct stat stats;
 932                         char devpath[UTIL_PATH_SIZE];
 933                         char from[UTIL_PATH_SIZE];
 934
 935                         have_db = false;
 936
 937                         /* find database in old location */
 938                         id = udev_device_get_id_filename(device);
 939                         util_strscpyl(from, sizeof(from), "/dev/.udev/db/", id, NULL);
 940                         if (lstat(from, &stats) == 0) {
 941                                 if (!have_db) {
 942                                         udev_device_read_db(device, from);
 943                                         have_db = true;
 944                                 }
 945                                 unlink(from);
 946                         }
 947
 948                         /* find old database with $subsys:$sysname name */
 949                         util_strscpyl(from, sizeof(from), "/dev/.udev/db/",
 950                                       udev_device_get_subsystem(device), ":", udev_device_get_sysname(device), NULL);
 951                         if (lstat(from, &stats) == 0) {
 952                                 if (!have_db) {
 953                                         udev_device_read_db(device, from);
 954                                         have_db = true;
 955                                 }
 956                                 unlink(from);
 957                         }
 958
 959                         /* find old database with the encoded devpath name */
 960                         util_path_encode(udev_device_get_devpath(device), devpath, sizeof(devpath));
 961                         util_strscpyl(from, sizeof(from), "/dev/.udev/db/", devpath, NULL);
 962                         if (lstat(from, &stats) == 0) {
 963                                 if (!have_db) {
 964                                         udev_device_read_db(device, from);
 965                                         have_db = true;
 966                                 }
 967                                 unlink(from);
 968                         }
 969
 970                         /* write out new database */
 971                         if (have_db)
 972                                 udev_device_update_db(device);
 973                 }
 974                 udev_device_unref(device);
 975         }
 976         udev_enumerate_unref(udev_enumerate);
 977         return 0;
 978 }
 979
 980 static int systemd_fds(struct udev *udev, int *rctrl, int *rnetlink)
 981 {
 982         int ctrl = -1, netlink = -1;
 983         int fd, n;
 984
 985         n = sd_listen_fds(true);
 986         if (n <= 0)
 987                 return -1;
 988
 989         for (fd = SD_LISTEN_FDS_START; fd < n + SD_LISTEN_FDS_START; fd++) {
 990                 if (sd_is_socket(fd, AF_LOCAL, SOCK_SEQPACKET, -1)) {
 991                         if (ctrl >= 0)
 992                                 return -1;
 993                         ctrl = fd;
 994                         continue;
 995                 }
 996
 997                 if (sd_is_socket(fd, AF_NETLINK, SOCK_RAW, -1)) {
 998                         if (netlink >= 0)
 999                                 return -1;
1000                         netlink = fd;
1001                         continue;
1002                 }
1003
1004                 return -1;
1005         }
1006
1007         if (ctrl < 0 || netlink < 0)
1008                 return -1;
1009
1010         log_debug("ctrl=%i netlink=%i\n", ctrl, netlink);
1011         *rctrl = ctrl;
1012         *rnetlink = netlink;
1013         return 0;
1014 }
1015
1016 /*
1017  * read the kernel commandline, in case we need to get into debug mode
1018  *   udev.log-priority=<level>              syslog priority
1019  *   udev.children-max=<number of workers>  events are fully serialized if set to 1
1020  *   udev.exec-delay=<number of seconds>    delay execution of every executed program
1021  */
1022 static void kernel_cmdline_options(struct udev *udev)
1023 {
1024         char *line, *w, *state;
1025         size_t l;
1026
1027         if (read_one_line_file("/proc/cmdline", &line) < 0)
1028                 return;
1029
1030         FOREACH_WORD_QUOTED(w, l, line, state) {
1031                 char *s, *opt;
1032
1033                 s = strndup(w, l);
1034                 if (!s)
1035                         break;
1036
1037                 /* accept the same options for the initrd, prefixed with "rd." */
1038                 if (in_initrd() && startswith(s, "rd."))
1039                         opt = s + 3;
1040                 else
1041                         opt = s;
1042
1043                 if (startswith(opt, "udev.log-priority="))
1044                         udev_set_log_priority(udev, util_log_priority(opt + 18));
1045
1046                 if (startswith(opt, "udev.children-max="))
1047                         children_max = strtoul(opt + 18, NULL, 0);
1048
1049                 if (startswith(opt, "udev.exec-delay="))
1050                         exec_delay = strtoul(opt + 16, NULL, 0);
1051
1052                 free(s);
1053         }
1054
1055         free(line);
1056 }
1057
1058 int main(int argc, char *argv[])
1059 {
1060         struct udev *udev;
1061         FILE *f;
1062         sigset_t mask;
1063         int daemonize = false;
1064         int resolve_names = 1;
1065         static const struct option options[] = {
1066                 { "daemon", no_argument, NULL, 'd' },
1067                 { "debug", no_argument, NULL, 'D' },
1068                 { "children-max", required_argument, NULL, 'c' },
1069                 { "exec-delay", required_argument, NULL, 'e' },
1070                 { "resolve-names", required_argument, NULL, 'N' },
1071                 { "help", no_argument, NULL, 'h' },
1072                 { "version", no_argument, NULL, 'V' },
1073                 {}
1074         };
1075         int fd_ctrl = -1;
1076         int fd_netlink = -1;
1077         int fd_worker = -1;
1078         struct epoll_event ep_ctrl, ep_inotify, ep_signal, ep_netlink, ep_worker;
1079         struct udev_ctrl_connection *ctrl_conn = NULL;
1080         int rc = 1;
1081
1082         udev = udev_new();
1083         if (udev == NULL)
1084                 goto exit;
1085
1086         log_open();
1087         log_parse_environment();
1088         udev_set_log_fn(udev, udev_main_log);
1089         log_debug("version %s\n", VERSION);
1090         label_init("/dev");
1091
1092         for (;;) {
1093                 int option;
1094
1095                 option = getopt_long(argc, argv, "c:deDtN:hV", options, NULL);
1096                 if (option == -1)
1097                         break;
1098
1099                 switch (option) {
1100                 case 'd':
1101                         daemonize = true;
1102                         break;
1103                 case 'c':
1104                         children_max = strtoul(optarg, NULL, 0);
1105                         break;
1106                 case 'e':
1107                         exec_delay = strtoul(optarg, NULL, 0);
1108                         break;
1109                 case 'D':
1110                         debug = true;
1111                         log_set_max_level(LOG_DEBUG);
1112                         udev_set_log_priority(udev, LOG_INFO);
1113                         break;
1114                 case 'N':
1115                         if (strcmp (optarg, "early") == 0) {
1116                                 resolve_names = 1;
1117                         } else if (strcmp (optarg, "late") == 0) {
1118                                 resolve_names = 0;
1119                         } else if (strcmp (optarg, "never") == 0) {
1120                                 resolve_names = -1;
1121                         } else {
1122                                 fprintf(stderr, "resolve-names must be early, late or never\n");
1123                                 log_error("resolve-names must be early, late or never\n");
1124                                 goto exit;
1125                         }
1126                         break;
1127                 case 'h':
1128                         printf("Usage: udevd OPTIONS\n"
1129                                "  --daemon\n"
1130                                "  --debug\n"
1131                                "  --children-max=<maximum number of workers>\n"
1132                                "  --exec-delay=<seconds to wait before executing RUN=>\n"
1133                                "  --resolve-names=early|late|never\n"
1134                                "  --version\n"
1135                                "  --help\n"
1136                                "\n");
1137                         goto exit;
1138                 case 'V':
1139                         printf("%s\n", VERSION);
1140                         goto exit;
1141                 default:
1142                         goto exit;
1143                 }
1144         }
1145
1146         kernel_cmdline_options(udev);
1147
1148         if (getuid() != 0) {
1149                 fprintf(stderr, "root privileges required\n");
1150                 log_error("root privileges required\n");
1151                 goto exit;
1152         }
1153
1154         /* set umask before creating any file/directory */
1155         chdir("/");
1156         umask(022);
1157
1158         mkdir("/run/udev", 0755);
1159
1160         dev_setup();
1161         static_dev_create_from_modules(udev);
1162
1163         /* before opening new files, make sure std{in,out,err} fds are in a sane state */
1164         if (daemonize) {
1165                 int fd;
1166
1167                 fd = open("/dev/null", O_RDWR);
1168                 if (fd >= 0) {
1169                         if (write(STDOUT_FILENO, 0, 0) < 0)
1170                                 dup2(fd, STDOUT_FILENO);
1171                         if (write(STDERR_FILENO, 0, 0) < 0)
1172                                 dup2(fd, STDERR_FILENO);
1173                         if (fd > STDERR_FILENO)
1174                                 close(fd);
1175                 } else {
1176                         fprintf(stderr, "cannot open /dev/null\n");
1177                         log_error("cannot open /dev/null\n");
1178                 }
1179         }
1180
1181         if (systemd_fds(udev, &fd_ctrl, &fd_netlink) >= 0) {
1182                 /* get control and netlink socket from from systemd */
1183                 udev_ctrl = udev_ctrl_new_from_fd(udev, fd_ctrl);
1184                 if (udev_ctrl == NULL) {
1185                         log_error("error taking over udev control socket");
1186                         rc = 1;
1187                         goto exit;
1188                 }
1189
1190                 monitor = udev_monitor_new_from_netlink_fd(udev, "kernel", fd_netlink);
1191                 if (monitor == NULL) {
1192                         log_error("error taking over netlink socket\n");
1193                         rc = 3;
1194                         goto exit;
1195                 }
1196
1197                 /* get our own cgroup, we regularly kill everything udev has left behind */
1198                 if (cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, 0, &udev_cgroup) < 0)
1199                         udev_cgroup = NULL;
1200         } else {
1201                 /* open control and netlink socket */
1202                 udev_ctrl = udev_ctrl_new(udev);
1203                 if (udev_ctrl == NULL) {
1204                         fprintf(stderr, "error initializing udev control socket");
1205                         log_error("error initializing udev control socket");
1206                         rc = 1;
1207                         goto exit;
1208                 }
1209                 fd_ctrl = udev_ctrl_get_fd(udev_ctrl);
1210
1211                 monitor = udev_monitor_new_from_netlink(udev, "kernel");
1212                 if (monitor == NULL) {
1213                         fprintf(stderr, "error initializing netlink socket\n");
1214                         log_error("error initializing netlink socket\n");
1215                         rc = 3;
1216                         goto exit;
1217                 }
1218                 fd_netlink = udev_monitor_get_fd(monitor);
1219         }
1220
1221         if (udev_monitor_enable_receiving(monitor) < 0) {
1222                 fprintf(stderr, "error binding netlink socket\n");
1223                 log_error("error binding netlink socket\n");
1224                 rc = 3;
1225                 goto exit;
1226         }
1227
1228         if (udev_ctrl_enable_receiving(udev_ctrl) < 0) {
1229                 fprintf(stderr, "error binding udev control socket\n");
1230                 log_error("error binding udev control socket\n");
1231                 rc = 1;
1232                 goto exit;
1233         }
1234
1235         udev_monitor_set_receive_buffer_size(monitor, 128*1024*1024);
1236
1237         /* create queue file before signalling 'ready', to make sure we block 'settle' */
1238         udev_queue_export = udev_queue_export_new(udev);
1239         if (udev_queue_export == NULL) {
1240                 log_error("error creating queue file\n");
1241                 goto exit;
1242         }
1243
1244         if (daemonize) {
1245                 pid_t pid;
1246
1247                 pid = fork();
1248                 switch (pid) {
1249                 case 0:
1250                         break;
1251                 case -1:
1252                         log_error("fork of daemon failed: %m\n");
1253                         rc = 4;
1254                         goto exit;
1255                 default:
1256                         rc = EXIT_SUCCESS;
1257                         goto exit_daemonize;
1258                 }
1259
1260                 setsid();
1261
1262                 write_one_line_file("/proc/self/oom_score_adj", "-1000");
1263         } else {
1264                 sd_notify(1, "READY=1");
1265         }
1266
1267         f = fopen("/dev/kmsg", "w");
1268         if (f != NULL) {
1269                 fprintf(f, "<30>systemd-udevd[%u]: starting version " VERSION "\n", getpid());
1270                 fclose(f);
1271         }
1272
1273         if (!debug) {
1274                 int fd;
1275
1276                 fd = open("/dev/null", O_RDWR);
1277                 if (fd >= 0) {
1278                         dup2(fd, STDIN_FILENO);
1279                         dup2(fd, STDOUT_FILENO);
1280                         dup2(fd, STDERR_FILENO);
1281                         close(fd);
1282                 }
1283         }
1284
1285         fd_inotify = udev_watch_init(udev);
1286         if (fd_inotify < 0) {
1287                 fprintf(stderr, "error initializing inotify\n");
1288                 log_error("error initializing inotify\n");
1289                 rc = 4;
1290                 goto exit;
1291         }
1292         udev_watch_restore(udev);
1293
1294         /* block and listen to all signals on signalfd */
1295         sigfillset(&mask);
1296         sigprocmask(SIG_SETMASK, &mask, &sigmask_orig);
1297         fd_signal = signalfd(-1, &mask, SFD_NONBLOCK|SFD_CLOEXEC);
1298         if (fd_signal < 0) {
1299                 fprintf(stderr, "error creating signalfd\n");
1300                 log_error("error creating signalfd\n");
1301                 rc = 5;
1302                 goto exit;
1303         }
1304
1305         /* unnamed socket from workers to the main daemon */
1306         if (socketpair(AF_LOCAL, SOCK_DGRAM|SOCK_CLOEXEC, 0, worker_watch) < 0) {
1307                 fprintf(stderr, "error creating socketpair\n");
1308                 log_error("error creating socketpair\n");
1309                 rc = 6;
1310                 goto exit;
1311         }
1312         fd_worker = worker_watch[READ_END];
1313
1314         udev_builtin_init(udev);
1315
1316         rules = udev_rules_new(udev, resolve_names);
1317         if (rules == NULL) {
1318                 log_error("error reading rules\n");
1319                 goto exit;
1320         }
1321
1322         memset(&ep_ctrl, 0, sizeof(struct epoll_event));
1323         ep_ctrl.events = EPOLLIN;
1324         ep_ctrl.data.fd = fd_ctrl;
1325
1326         memset(&ep_inotify, 0, sizeof(struct epoll_event));
1327         ep_inotify.events = EPOLLIN;
1328         ep_inotify.data.fd = fd_inotify;
1329
1330         memset(&ep_signal, 0, sizeof(struct epoll_event));
1331         ep_signal.events = EPOLLIN;
1332         ep_signal.data.fd = fd_signal;
1333
1334         memset(&ep_netlink, 0, sizeof(struct epoll_event));
1335         ep_netlink.events = EPOLLIN;
1336         ep_netlink.data.fd = fd_netlink;
1337
1338         memset(&ep_worker, 0, sizeof(struct epoll_event));
1339         ep_worker.events = EPOLLIN;
1340         ep_worker.data.fd = fd_worker;
1341
1342         fd_ep = epoll_create1(EPOLL_CLOEXEC);
1343         if (fd_ep < 0) {
1344                 log_error("error creating epoll fd: %m\n");
1345                 goto exit;
1346         }
1347         if (epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_ctrl, &ep_ctrl) < 0 ||
1348             epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_inotify, &ep_inotify) < 0 ||
1349             epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_signal, &ep_signal) < 0 ||
1350             epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_netlink, &ep_netlink) < 0 ||
1351             epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_worker, &ep_worker) < 0) {
1352                 log_error("fail to add fds to epoll: %m\n");
1353                 goto exit;
1354         }
1355
1356         /* if needed, convert old database from earlier udev version */
1357         convert_db(udev);
1358
1359         if (children_max <= 0) {
1360                 int memsize = mem_size_mb();
1361
1362                 /* set value depending on the amount of RAM */
1363                 if (memsize > 0)
1364                         children_max = 16 + (memsize / 8);
1365                 else
1366                         children_max = 16;
1367         }
1368         log_debug("set children_max to %u\n", children_max);
1369
1370         udev_rules_apply_static_dev_perms(rules);
1371
1372         udev_list_node_init(&event_list);
1373         udev_list_node_init(&worker_list);
1374
1375         for (;;) {
1376                 static unsigned long long last_usec;
1377                 struct epoll_event ev[8];
1378                 int fdcount;
1379                 int timeout;
1380                 bool is_worker, is_signal, is_inotify, is_netlink, is_ctrl;
1381                 int i;
1382
1383                 if (udev_exit) {
1384                         /* close sources of new events and discard buffered events */
1385                         if (fd_ctrl >= 0) {
1386                                 epoll_ctl(fd_ep, EPOLL_CTL_DEL, fd_ctrl, NULL);
1387                                 fd_ctrl = -1;
1388                         }
1389                         if (monitor != NULL) {
1390                                 epoll_ctl(fd_ep, EPOLL_CTL_DEL, fd_netlink, NULL);
1391                                 udev_monitor_unref(monitor);
1392                                 monitor = NULL;
1393                         }
1394                         if (fd_inotify >= 0) {
1395                                 epoll_ctl(fd_ep, EPOLL_CTL_DEL, fd_inotify, NULL);
1396                                 close(fd_inotify);
1397                                 fd_inotify = -1;
1398                         }
1399
1400                         /* discard queued events and kill workers */
1401                         event_queue_cleanup(udev, EVENT_QUEUED);
1402                         worker_kill(udev);
1403
1404                         /* exit after all has cleaned up */
1405                         if (udev_list_node_is_empty(&event_list) && udev_list_node_is_empty(&worker_list))
1406                                 break;
1407
1408                         /* timeout at exit for workers to finish */
1409                         timeout = 30 * 1000;
1410                 } else if (udev_list_node_is_empty(&event_list) && !children) {
1411                         /* we are idle */
1412                         timeout = -1;
1413
1414                         /* cleanup possible left-over processes in our cgroup */
1415                         if (udev_cgroup)
1416                                 cg_kill(SYSTEMD_CGROUP_CONTROLLER, udev_cgroup, SIGKILL, false, true, NULL);
1417                 } else {
1418                         /* kill idle or hanging workers */
1419                         timeout = 3 * 1000;
1420                 }
1421                 fdcount = epoll_wait(fd_ep, ev, ELEMENTSOF(ev), timeout);
1422                 if (fdcount < 0)
1423                         continue;
1424
1425                 if (fdcount == 0) {
1426                         struct udev_list_node *loop;
1427
1428                         /* timeout */
1429                         if (udev_exit) {
1430                                 log_error("timeout, giving up waiting for workers to finish\n");
1431                                 break;
1432                         }
1433
1434                         /* kill idle workers */
1435                         if (udev_list_node_is_empty(&event_list)) {
1436                                 log_debug("cleanup idle workers\n");
1437                                 worker_kill(udev);
1438                         }
1439
1440                         /* check for hanging events */
1441                         udev_list_node_foreach(loop, &worker_list) {
1442                                 struct worker *worker = node_to_worker(loop);
1443
1444                                 if (worker->state != WORKER_RUNNING)
1445                                         continue;
1446
1447                                 if ((now_usec() - worker->event_start_usec) > 30 * 1000 * 1000) {
1448                                         log_error("worker [%u] %s timeout; kill it\n", worker->pid,
1449                                             worker->event ? worker->event->devpath : "<idle>");
1450                                         kill(worker->pid, SIGKILL);
1451                                         worker->state = WORKER_KILLED;
1452                                         /* drop reference taken for state 'running' */
1453                                         worker_unref(worker);
1454                                         if (worker->event) {
1455                                                 log_error("seq %llu '%s' killed\n",
1456                                                           udev_device_get_seqnum(worker->event->dev), worker->event->devpath);
1457                                                 worker->event->exitcode = -64;
1458                                                 event_queue_delete(worker->event, true);
1459                                                 worker->event = NULL;
1460                                         }
1461                                 }
1462                         }
1463
1464                 }
1465
1466                 is_worker = is_signal = is_inotify = is_netlink = is_ctrl = false;
1467                 for (i = 0; i < fdcount; i++) {
1468                         if (ev[i].data.fd == fd_worker && ev[i].events & EPOLLIN)
1469                                 is_worker = true;
1470                         else if (ev[i].data.fd == fd_netlink && ev[i].events & EPOLLIN)
1471                                 is_netlink = true;
1472                         else if (ev[i].data.fd == fd_signal && ev[i].events & EPOLLIN)
1473                                 is_signal = true;
1474                         else if (ev[i].data.fd == fd_inotify && ev[i].events & EPOLLIN)
1475                                 is_inotify = true;
1476                         else if (ev[i].data.fd == fd_ctrl && ev[i].events & EPOLLIN)
1477                                 is_ctrl = true;
1478                 }
1479
1480                 /* check for changed config, every 3 seconds at most */
1481                 if ((now_usec() - last_usec) > 3 * 1000 * 1000) {
1482                         if (udev_rules_check_timestamp(rules))
1483                                 reload = true;
1484                         if (udev_builtin_validate(udev))
1485                                 reload = true;
1486
1487                         last_usec = now_usec();
1488                 }
1489
1490                 /* reload requested, HUP signal received, rules changed, builtin changed */
1491                 if (reload) {
1492                         worker_kill(udev);
1493                         rules = udev_rules_unref(rules);
1494                         udev_builtin_exit(udev);
1495                         reload = 0;
1496                 }
1497
1498                 /* event has finished */
1499                 if (is_worker)
1500                         worker_returned(fd_worker);
1501
1502                 if (is_netlink) {
1503                         struct udev_device *dev;
1504
1505                         dev = udev_monitor_receive_device(monitor);
1506                         if (dev != NULL) {
1507                                 udev_device_set_usec_initialized(dev, now_usec());
1508                                 if (event_queue_insert(dev) < 0)
1509                                         udev_device_unref(dev);
1510                         }
1511                 }
1512
1513                 /* start new events */
1514                 if (!udev_list_node_is_empty(&event_list) && !udev_exit && !stop_exec_queue) {
1515                         if (rules == NULL)
1516                                 rules = udev_rules_new(udev, resolve_names);
1517                         if (rules != NULL)
1518                                 event_queue_start(udev);
1519                 }
1520
1521                 if (is_signal) {
1522                         struct signalfd_siginfo fdsi;
1523                         ssize_t size;
1524
1525                         size = read(fd_signal, &fdsi, sizeof(struct signalfd_siginfo));
1526                         if (size == sizeof(struct signalfd_siginfo))
1527                                 handle_signal(udev, fdsi.ssi_signo);
1528                 }
1529
1530                 /* we are shutting down, the events below are not handled anymore */
1531                 if (udev_exit)
1532                         continue;
1533
1534                 /* device node watch */
1535                 if (is_inotify)
1536                         handle_inotify(udev);
1537
1538                 /*
1539                  * This needs to be after the inotify handling, to make sure,
1540                  * that the ping is send back after the possibly generated
1541                  * "change" events by the inotify device node watch.
1542                  *
1543                  * A single time we may receive a client connection which we need to
1544                  * keep open to block the client. It will be closed right before we
1545                  * exit.
1546                  */
1547                 if (is_ctrl)
1548                         ctrl_conn = handle_ctrl_msg(udev_ctrl);
1549         }
1550
1551         rc = EXIT_SUCCESS;
1552 exit:
1553         udev_queue_export_cleanup(udev_queue_export);
1554         udev_ctrl_cleanup(udev_ctrl);
1555 exit_daemonize:
1556         if (fd_ep >= 0)
1557                 close(fd_ep);
1558         worker_list_cleanup(udev);
1559         event_queue_cleanup(udev, EVENT_UNDEF);
1560         udev_rules_unref(rules);
1561         udev_builtin_exit(udev);
1562         if (fd_signal >= 0)
1563                 close(fd_signal);
1564         if (worker_watch[READ_END] >= 0)
1565                 close(worker_watch[READ_END]);
1566         if (worker_watch[WRITE_END] >= 0)
1567                 close(worker_watch[WRITE_END]);
1568         udev_monitor_unref(monitor);
1569         udev_queue_export_unref(udev_queue_export);
1570         udev_ctrl_connection_unref(ctrl_conn);
1571         udev_ctrl_unref(udev_ctrl);
1572         label_finish();
1573         udev_unref(udev);
1574         log_close();
1575         return rc;
1576 }