udev/udevd.c

   1 /*
   2  * Copyright (C) 2004-2011 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
  50 #define UDEVD_PRIORITY                  -4
  51 #define UDEV_PRIORITY                   -2
  52
  53 static bool debug;
  54
  55 static void log_fn(struct udev *udev, int priority,
  56                    const char *file, int line, const char *fn,
  57                    const char *format, va_list args)
  58 {
  59         if (debug) {
  60                 char buf[1024];
  61                 struct timeval tv;
  62                 struct timezone tz;
  63
  64                 vsnprintf(buf, sizeof(buf), format, args);
  65                 gettimeofday(&tv, &tz);
  66                 fprintf(stderr, "%llu.%06u [%u] %s: %s",
  67                         (unsigned long long) tv.tv_sec, (unsigned int) tv.tv_usec,
  68                         (int) getpid(), fn, buf);
  69         } else {
  70                 vsyslog(priority, format, args);
  71         }
  72 }
  73
  74 static struct udev_rules *rules;
  75 static struct udev_queue_export *udev_queue_export;
  76 static struct udev_ctrl *udev_ctrl;
  77 static struct udev_monitor *monitor;
  78 static int worker_watch[2] = { -1, -1 };
  79 static int fd_signal = -1;
  80 static int fd_ep = -1;
  81 static int fd_inotify = -1;
  82 static bool stop_exec_queue;
  83 static bool reload_config;
  84 static int children;
  85 static int children_max;
  86 static int exec_delay;
  87 static sigset_t orig_sigmask;
  88 static UDEV_LIST(event_list);
  89 static UDEV_LIST(worker_list);
  90 static bool udev_exit;
  91 static volatile sig_atomic_t worker_exit;
  92
  93 enum event_state {
  94         EVENT_UNDEF,
  95         EVENT_QUEUED,
  96         EVENT_RUNNING,
  97 };
  98
  99 struct event {
 100         struct udev_list_node node;
 101         struct udev *udev;
 102         struct udev_device *dev;
 103         enum event_state state;
 104         int exitcode;
 105         unsigned long long int delaying_seqnum;
 106         unsigned long long int seqnum;
 107         const char *devpath;
 108         size_t devpath_len;
 109         const char *devpath_old;
 110         dev_t devnum;
 111         bool is_block;
 112         int ifindex;
 113 };
 114
 115 static struct event *node_to_event(struct udev_list_node *node)
 116 {
 117         char *event;
 118
 119         event = (char *)node;
 120         event -= offsetof(struct event, node);
 121         return (struct event *)event;
 122 }
 123
 124 static void event_queue_cleanup(struct udev *udev, enum event_state type);
 125
 126 enum worker_state {
 127         WORKER_UNDEF,
 128         WORKER_RUNNING,
 129         WORKER_IDLE,
 130         WORKER_KILLED,
 131 };
 132
 133 struct worker {
 134         struct udev_list_node node;
 135         struct udev *udev;
 136         int refcount;
 137         pid_t pid;
 138         struct udev_monitor *monitor;
 139         enum worker_state state;
 140         struct event *event;
 141 };
 142
 143 /* passed from worker to main process */
 144 struct worker_message {
 145         pid_t pid;
 146         int exitcode;
 147 };
 148
 149 static struct worker *node_to_worker(struct udev_list_node *node)
 150 {
 151         char *worker;
 152
 153         worker = (char *)node;
 154         worker -= offsetof(struct worker, node);
 155         return (struct worker *)worker;
 156 }
 157
 158 static void event_queue_delete(struct event *event, bool export)
 159 {
 160         udev_list_node_remove(&event->node);
 161
 162         if (export) {
 163                 /* mark as failed, if "add" event returns non-zero */
 164                 if (event->exitcode != 0 && strcmp(udev_device_get_action(event->dev), "remove") != 0)
 165                         udev_queue_export_device_failed(udev_queue_export, event->dev);
 166                 else
 167                         udev_queue_export_device_finished(udev_queue_export, event->dev);
 168                 info(event->udev, "seq %llu done with %i\n", udev_device_get_seqnum(event->dev), event->exitcode);
 169         }
 170         udev_device_unref(event->dev);
 171         free(event);
 172 }
 173
 174 static void event_sig_handler(int signum)
 175 {
 176         switch (signum) {
 177         case SIGALRM:
 178                 _exit(1);
 179                 break;
 180         case SIGTERM:
 181                 worker_exit = true;
 182                 break;
 183         }
 184 }
 185
 186 static struct worker *worker_ref(struct worker *worker)
 187 {
 188         worker->refcount++;
 189         return worker;
 190 }
 191
 192 static void worker_cleanup(struct worker *worker)
 193 {
 194         udev_list_node_remove(&worker->node);
 195         udev_monitor_unref(worker->monitor);
 196         children--;
 197         free(worker);
 198 }
 199
 200 static void worker_unref(struct worker *worker)
 201 {
 202         worker->refcount--;
 203         if (worker->refcount > 0)
 204                 return;
 205         info(worker->udev, "worker [%u] cleaned up\n", worker->pid);
 206         worker_cleanup(worker);
 207 }
 208
 209 static void worker_list_cleanup(struct udev *udev)
 210 {
 211         struct udev_list_node *loop, *tmp;
 212
 213         udev_list_node_foreach_safe(loop, tmp, &worker_list) {
 214                 struct worker *worker = node_to_worker(loop);
 215
 216                 worker_cleanup(worker);
 217         }
 218 }
 219
 220 static void worker_new(struct event *event)
 221 {
 222         struct udev *udev = event->udev;
 223         struct worker *worker;
 224         struct udev_monitor *worker_monitor;
 225         pid_t pid;
 226         struct sigaction act;
 227
 228         /* listen for new events */
 229         worker_monitor = udev_monitor_new_from_netlink(udev, NULL);
 230         if (worker_monitor == NULL)
 231                 return;
 232         /* allow the main daemon netlink address to send devices to the worker */
 233         udev_monitor_allow_unicast_sender(worker_monitor, monitor);
 234         udev_monitor_enable_receiving(worker_monitor);
 235
 236         worker = calloc(1, sizeof(struct worker));
 237         if (worker == NULL) {
 238                 udev_monitor_unref(worker_monitor);
 239                 return;
 240         }
 241         /* worker + event reference */
 242         worker->refcount = 2;
 243         worker->udev = udev;
 244
 245         pid = fork();
 246         switch (pid) {
 247         case 0: {
 248                 sigset_t sigmask;
 249                 struct udev_device *dev;
 250                 struct pollfd pmon = {
 251                         .fd = udev_monitor_get_fd(worker_monitor),
 252                         .events = POLLIN,
 253                 };
 254
 255                 /* move initial device from queue */
 256                 dev = event->dev;
 257                 event->dev = NULL;
 258
 259                 free(worker);
 260                 worker_list_cleanup(udev);
 261                 event_queue_cleanup(udev, EVENT_UNDEF);
 262                 udev_queue_export_unref(udev_queue_export);
 263                 udev_monitor_unref(monitor);
 264                 udev_ctrl_unref(udev_ctrl);
 265                 close(fd_signal);
 266                 close(fd_ep);
 267                 close(worker_watch[READ_END]);
 268                 udev_log_close();
 269                 udev_log_init("udevd-work");
 270                 setpriority(PRIO_PROCESS, 0, UDEV_PRIORITY);
 271
 272                 /* set signal handlers */
 273                 memset(&act, 0x00, sizeof(act));
 274                 act.sa_handler = event_sig_handler;
 275                 sigemptyset (&act.sa_mask);
 276                 act.sa_flags = 0;
 277                 sigaction(SIGTERM, &act, NULL);
 278                 sigaction(SIGALRM, &act, NULL);
 279
 280                 /* unblock SIGALRM */
 281                 sigfillset(&sigmask);
 282                 sigdelset(&sigmask, SIGALRM);
 283                 sigprocmask(SIG_SETMASK, &sigmask, NULL);
 284                 /* SIGTERM is unblocked in ppoll() */
 285                 sigdelset(&sigmask, SIGTERM);
 286
 287                 /* request TERM signal if parent exits */
 288                 prctl(PR_SET_PDEATHSIG, SIGTERM);
 289
 290                 do {
 291                         struct udev_event *udev_event;
 292                         struct worker_message msg = {};
 293                         int err;
 294                         int failed = 0;
 295
 296                         info(udev, "seq %llu running\n", udev_device_get_seqnum(dev));
 297                         udev_event = udev_event_new(dev);
 298                         if (udev_event == NULL)
 299                                 _exit(3);
 300
 301                         /* set timeout to prevent hanging processes */
 302                         alarm(UDEV_EVENT_TIMEOUT);
 303
 304                         if (exec_delay > 0)
 305                                 udev_event->exec_delay = exec_delay;
 306
 307                         /* apply rules, create node, symlinks */
 308                         err = udev_event_execute_rules(udev_event, rules);
 309
 310                         /* rules may change/disable the timeout */
 311                         if (udev_device_get_event_timeout(dev) >= 0)
 312                                 alarm(udev_device_get_event_timeout(dev));
 313
 314                         if (err == 0)
 315                                 failed = udev_event_execute_run(udev_event, &orig_sigmask);
 316
 317                         alarm(0);
 318
 319                         /* apply/restore inotify watch */
 320                         if (err == 0 && udev_event->inotify_watch) {
 321                                 udev_watch_begin(udev, dev);
 322                                 udev_device_update_db(dev);
 323                         }
 324
 325                         /* send processed event back to libudev listeners */
 326                         udev_monitor_send_device(worker_monitor, NULL, dev);
 327
 328                         /* send udevd the result of the event execution */
 329                         if (err != 0)
 330                                 msg.exitcode = err;
 331                         else if (failed != 0)
 332                                 msg.exitcode = failed;
 333                         msg.pid = getpid();
 334                         send(worker_watch[WRITE_END], &msg, sizeof(struct worker_message), 0);
 335
 336                         info(udev, "seq %llu processed with %i\n", udev_device_get_seqnum(dev), err);
 337                         udev_event_unref(udev_event);
 338                         udev_device_unref(dev);
 339                         dev = NULL;
 340
 341                         /* wait for more device messages or signal from udevd */
 342                         while (!worker_exit) {
 343                                 int fdcount;
 344
 345                                 fdcount = ppoll(&pmon, 1, NULL, &sigmask);
 346                                 if (fdcount < 0)
 347                                         continue;
 348
 349                                 if (pmon.revents & POLLIN) {
 350                                         dev = udev_monitor_receive_device(worker_monitor);
 351                                         if (dev != NULL)
 352                                                 break;
 353                                 }
 354                         }
 355                 } while (dev != NULL);
 356
 357                 close(fd_inotify);
 358                 close(worker_watch[WRITE_END]);
 359                 udev_rules_unref(rules);
 360                 udev_monitor_unref(worker_monitor);
 361                 udev_unref(udev);
 362                 udev_log_close();
 363                 exit(EXIT_SUCCESS);
 364         }
 365         case -1:
 366                 udev_monitor_unref(worker_monitor);
 367                 event->state = EVENT_QUEUED;
 368                 free(worker);
 369                 err(udev, "fork of child failed: %m\n");
 370                 break;
 371         default:
 372                 /* close monitor, but keep address around */
 373                 udev_monitor_disconnect(worker_monitor);
 374                 worker->monitor = worker_monitor;
 375                 worker->pid = pid;
 376                 worker->state = WORKER_RUNNING;
 377                 worker->event = event;
 378                 event->state = EVENT_RUNNING;
 379                 udev_list_node_append(&worker->node, &worker_list);
 380                 children++;
 381                 info(udev, "seq %llu forked new worker [%u]\n", udev_device_get_seqnum(event->dev), pid);
 382                 break;
 383         }
 384 }
 385
 386 static void event_run(struct event *event, bool force)
 387 {
 388         struct udev_list_node *loop;
 389
 390         udev_list_node_foreach(loop, &worker_list) {
 391                 struct worker *worker = node_to_worker(loop);
 392                 ssize_t count;
 393
 394                 if (worker->state != WORKER_IDLE)
 395                         continue;
 396
 397                 count = udev_monitor_send_device(monitor, worker->monitor, event->dev);
 398                 if (count < 0) {
 399                         err(event->udev, "worker [%u] did not accept message %zi (%m), kill it\n", worker->pid, count);
 400                         kill(worker->pid, SIGKILL);
 401                         worker->state = WORKER_KILLED;
 402                         continue;
 403                 }
 404                 worker_ref(worker);
 405                 worker->event = event;
 406                 worker->state = WORKER_RUNNING;
 407                 event->state = EVENT_RUNNING;
 408                 return;
 409         }
 410
 411         if (!force && children >= children_max) {
 412                 if (children_max > 1)
 413                         info(event->udev, "maximum number (%i) of children reached\n", children);
 414                 return;
 415         }
 416
 417         /* start new worker and pass initial device */
 418         worker_new(event);
 419 }
 420
 421 static int event_queue_insert(struct udev_device *dev)
 422 {
 423         struct event *event;
 424
 425         event = calloc(1, sizeof(struct event));
 426         if (event == NULL)
 427                 return -1;
 428
 429         event->udev = udev_device_get_udev(dev);
 430         event->dev = dev;
 431         event->seqnum = udev_device_get_seqnum(dev);
 432         event->devpath = udev_device_get_devpath(dev);
 433         event->devpath_len = strlen(event->devpath);
 434         event->devpath_old = udev_device_get_devpath_old(dev);
 435         event->devnum = udev_device_get_devnum(dev);
 436         event->is_block = (strcmp("block", udev_device_get_subsystem(dev)) == 0);
 437         event->ifindex = udev_device_get_ifindex(dev);
 438
 439         udev_queue_export_device_queued(udev_queue_export, dev);
 440         info(event->udev, "seq %llu queued, '%s' '%s'\n", udev_device_get_seqnum(dev),
 441              udev_device_get_action(dev), udev_device_get_subsystem(dev));
 442
 443         event->state = EVENT_QUEUED;
 444         udev_list_node_append(&event->node, &event_list);
 445
 446         /* run all events with a timeout set immediately */
 447         if (udev_device_get_timeout(dev) > 0) {
 448                 event_run(event, true);
 449                 return 0;
 450         }
 451
 452         return 0;
 453 }
 454
 455 static void worker_kill(struct udev *udev, int retain)
 456 {
 457         struct udev_list_node *loop;
 458         int max;
 459
 460         if (children <= retain)
 461                 return;
 462
 463         max = children - retain;
 464
 465         udev_list_node_foreach(loop, &worker_list) {
 466                 struct worker *worker = node_to_worker(loop);
 467
 468                 if (max-- <= 0)
 469                         break;
 470
 471                 if (worker->state == WORKER_KILLED)
 472                         continue;
 473
 474                 worker->state = WORKER_KILLED;
 475                 kill(worker->pid, SIGTERM);
 476         }
 477 }
 478
 479 /* lookup event for identical, parent, child device */
 480 static bool is_devpath_busy(struct event *event)
 481 {
 482         struct udev_list_node *loop;
 483         size_t common;
 484
 485         /* check if queue contains events we depend on */
 486         udev_list_node_foreach(loop, &event_list) {
 487                 struct event *loop_event = node_to_event(loop);
 488
 489                 /* we already found a later event, earlier can not block us, no need to check again */
 490                 if (loop_event->seqnum < event->delaying_seqnum)
 491                         continue;
 492
 493                 /* event we checked earlier still exists, no need to check again */
 494                 if (loop_event->seqnum == event->delaying_seqnum)
 495                         return true;
 496
 497                 /* found ourself, no later event can block us */
 498                 if (loop_event->seqnum >= event->seqnum)
 499                         break;
 500
 501                 /* check major/minor */
 502                 if (major(event->devnum) != 0 && event->devnum == loop_event->devnum && event->is_block == loop_event->is_block)
 503                         return true;
 504
 505                 /* check network device ifindex */
 506                 if (event->ifindex != 0 && event->ifindex == loop_event->ifindex)
 507                         return true;
 508
 509                 /* check our old name */
 510                 if (event->devpath_old != NULL && strcmp(loop_event->devpath, event->devpath_old) == 0) {
 511                         event->delaying_seqnum = loop_event->seqnum;
 512                         return true;
 513                 }
 514
 515                 /* compare devpath */
 516                 common = MIN(loop_event->devpath_len, event->devpath_len);
 517
 518                 /* one devpath is contained in the other? */
 519                 if (memcmp(loop_event->devpath, event->devpath, common) != 0)
 520                         continue;
 521
 522                 /* identical device event found */
 523                 if (loop_event->devpath_len == event->devpath_len) {
 524                         /* devices names might have changed/swapped in the meantime */
 525                         if (major(event->devnum) != 0 && (event->devnum != loop_event->devnum || event->is_block != loop_event->is_block))
 526                                 continue;
 527                         if (event->ifindex != 0 && event->ifindex != loop_event->ifindex)
 528                                 continue;
 529                         event->delaying_seqnum = loop_event->seqnum;
 530                         return true;
 531                 }
 532
 533                 /* parent device event found */
 534                 if (event->devpath[common] == '/') {
 535                         event->delaying_seqnum = loop_event->seqnum;
 536                         return true;
 537                 }
 538
 539                 /* child device event found */
 540                 if (loop_event->devpath[common] == '/') {
 541                         event->delaying_seqnum = loop_event->seqnum;
 542                         return true;
 543                 }
 544
 545                 /* no matching device */
 546                 continue;
 547         }
 548
 549         return false;
 550 }
 551
 552 static void event_queue_start(struct udev *udev)
 553 {
 554         struct udev_list_node *loop;
 555
 556         udev_list_node_foreach(loop, &event_list) {
 557                 struct event *event = node_to_event(loop);
 558
 559                 if (event->state != EVENT_QUEUED)
 560                         continue;
 561
 562                 /* do not start event if parent or child event is still running */
 563                 if (is_devpath_busy(event)) {
 564                         dbg(udev, "delay seq %llu (%s)\n", event->seqnum, event->devpath);
 565                         continue;
 566                 }
 567
 568                 event_run(event, false);
 569         }
 570 }
 571
 572 static void event_queue_cleanup(struct udev *udev, enum event_state match_type)
 573 {
 574         struct udev_list_node *loop, *tmp;
 575
 576         udev_list_node_foreach_safe(loop, tmp, &event_list) {
 577                 struct event *event = node_to_event(loop);
 578
 579                 if (match_type != EVENT_UNDEF && match_type != event->state)
 580                         continue;
 581
 582                 event_queue_delete(event, false);
 583         }
 584 }
 585
 586 static void worker_returned(int fd_worker)
 587 {
 588         for (;;) {
 589                 struct worker_message msg;
 590                 ssize_t size;
 591                 struct udev_list_node *loop;
 592
 593                 size = recv(fd_worker, &msg, sizeof(struct worker_message), MSG_DONTWAIT);
 594                 if (size != sizeof(struct worker_message))
 595                         break;
 596
 597                 /* lookup worker who sent the signal */
 598                 udev_list_node_foreach(loop, &worker_list) {
 599                         struct worker *worker = node_to_worker(loop);
 600
 601                         if (worker->pid != msg.pid)
 602                                 continue;
 603
 604                         /* worker returned */
 605                         worker->event->exitcode = msg.exitcode;
 606                         event_queue_delete(worker->event, true);
 607                         worker->event = NULL;
 608                         if (worker->state != WORKER_KILLED)
 609                                 worker->state = WORKER_IDLE;
 610                         worker_unref(worker);
 611                         break;
 612                 }
 613         }
 614 }
 615
 616 /* receive the udevd message from userspace */
 617 static struct udev_ctrl_connection *handle_ctrl_msg(struct udev_ctrl *uctrl)
 618 {
 619         struct udev *udev = udev_ctrl_get_udev(uctrl);
 620         struct udev_ctrl_connection *ctrl_conn;
 621         struct udev_ctrl_msg *ctrl_msg = NULL;
 622         const char *str;
 623         int i;
 624
 625         ctrl_conn = udev_ctrl_get_connection(uctrl);
 626         if (ctrl_conn == NULL)
 627                 goto out;
 628
 629         ctrl_msg = udev_ctrl_receive_msg(ctrl_conn);
 630         if (ctrl_msg == NULL)
 631                 goto out;
 632
 633         i = udev_ctrl_get_set_log_level(ctrl_msg);
 634         if (i >= 0) {
 635                 info(udev, "udevd message (SET_LOG_PRIORITY) received, log_priority=%i\n", i);
 636                 udev_set_log_priority(udev, i);
 637                 worker_kill(udev, 0);
 638         }
 639
 640         if (udev_ctrl_get_stop_exec_queue(ctrl_msg) > 0) {
 641                 info(udev, "udevd message (STOP_EXEC_QUEUE) received\n");
 642                 stop_exec_queue = true;
 643         }
 644
 645         if (udev_ctrl_get_start_exec_queue(ctrl_msg) > 0) {
 646                 info(udev, "udevd message (START_EXEC_QUEUE) received\n");
 647                 stop_exec_queue = false;
 648         }
 649
 650         if (udev_ctrl_get_reload_rules(ctrl_msg) > 0) {
 651                 info(udev, "udevd message (RELOAD_RULES) received\n");
 652                 reload_config = true;
 653         }
 654
 655         str = udev_ctrl_get_set_env(ctrl_msg);
 656         if (str != NULL) {
 657                 char *key;
 658
 659                 key = strdup(str);
 660                 if (key != NULL) {
 661                         char *val;
 662
 663                         val = strchr(key, '=');
 664                         if (val != NULL) {
 665                                 val[0] = '\0';
 666                                 val = &val[1];
 667                                 if (val[0] == '\0') {
 668                                         info(udev, "udevd message (ENV) received, unset '%s'\n", key);
 669                                         udev_add_property(udev, key, NULL);
 670                                 } else {
 671                                         info(udev, "udevd message (ENV) received, set '%s=%s'\n", key, val);
 672                                         udev_add_property(udev, key, val);
 673                                 }
 674                         } else {
 675                                 err(udev, "wrong key format '%s'\n", key);
 676                         }
 677                         free(key);
 678                 }
 679                 worker_kill(udev, 0);
 680         }
 681
 682         i = udev_ctrl_get_set_children_max(ctrl_msg);
 683         if (i >= 0) {
 684                 info(udev, "udevd message (SET_MAX_CHILDREN) received, children_max=%i\n", i);
 685                 children_max = i;
 686         }
 687
 688         if (udev_ctrl_get_ping(ctrl_msg) > 0)
 689                 info(udev, "udevd message (SYNC) received\n");
 690
 691         if (udev_ctrl_get_exit(ctrl_msg) > 0) {
 692                 info(udev, "udevd message (EXIT) received\n");
 693                 udev_exit = true;
 694                 /* keep reference to block the client until we exit */
 695                 udev_ctrl_connection_ref(ctrl_conn);
 696         }
 697 out:
 698         udev_ctrl_msg_unref(ctrl_msg);
 699         return udev_ctrl_connection_unref(ctrl_conn);
 700 }
 701
 702 /* read inotify messages */
 703 static int handle_inotify(struct udev *udev)
 704 {
 705         int nbytes, pos;
 706         char *buf;
 707         struct inotify_event *ev;
 708
 709         if ((ioctl(fd_inotify, FIONREAD, &nbytes) < 0) || (nbytes <= 0))
 710                 return 0;
 711
 712         buf = malloc(nbytes);
 713         if (buf == NULL) {
 714                 err(udev, "error getting buffer for inotify\n");
 715                 return -1;
 716         }
 717
 718         nbytes = read(fd_inotify, buf, nbytes);
 719
 720         for (pos = 0; pos < nbytes; pos += sizeof(struct inotify_event) + ev->len) {
 721                 struct udev_device *dev;
 722
 723                 ev = (struct inotify_event *)(buf + pos);
 724                 if (ev->len) {
 725                         const char *s;
 726
 727                         info(udev, "inotify event: %x for %s\n", ev->mask, ev->name);
 728                         s = strstr(ev->name, ".rules");
 729                         if (s == NULL)
 730                                 continue;
 731                         if (strlen(s) != strlen(".rules"))
 732                                 continue;
 733                         reload_config = true;
 734                         continue;
 735                 }
 736
 737                 dev = udev_watch_lookup(udev, ev->wd);
 738                 if (dev != NULL) {
 739                         info(udev, "inotify event: %x for %s\n", ev->mask, udev_device_get_devnode(dev));
 740                         if (ev->mask & IN_CLOSE_WRITE) {
 741                                 char filename[UTIL_PATH_SIZE];
 742                                 int fd;
 743
 744                                 info(udev, "device %s closed, synthesising 'change'\n", udev_device_get_devnode(dev));
 745                                 util_strscpyl(filename, sizeof(filename), udev_device_get_syspath(dev), "/uevent", NULL);
 746                                 fd = open(filename, O_WRONLY);
 747                                 if (fd < 0 || write(fd, "change", 6) < 0)
 748                                         info(udev, "error writing uevent: %m\n");
 749                                 close(fd);
 750                         }
 751                         if (ev->mask & IN_IGNORED)
 752                                 udev_watch_end(udev, dev);
 753
 754                         udev_device_unref(dev);
 755                 }
 756
 757         }
 758
 759         free(buf);
 760         return 0;
 761 }
 762
 763 static void handle_signal(struct udev *udev, int signo)
 764 {
 765         switch (signo) {
 766         case SIGINT:
 767         case SIGTERM:
 768                 udev_exit = true;
 769                 break;
 770         case SIGCHLD:
 771                 for (;;) {
 772                         pid_t pid;
 773                         int status;
 774                         struct udev_list_node *loop, *tmp;
 775
 776                         pid = waitpid(-1, &status, WNOHANG);
 777                         if (pid <= 0)
 778                                 break;
 779
 780                         udev_list_node_foreach_safe(loop, tmp, &worker_list) {
 781                                 struct worker *worker = node_to_worker(loop);
 782
 783                                 if (worker->pid != pid)
 784                                         continue;
 785
 786                                 info(udev, "worker [%u] exit\n", pid);
 787                                 if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 788                                         err(udev, "worker [%u] unexpectedly returned with status 0x%04x\n", pid, status);
 789                                         if (worker->event != NULL) {
 790                                                 err(udev, "worker [%u] failed while handling '%s'\n", pid, worker->event->devpath);
 791                                                 worker->event->exitcode = -32;
 792                                                 event_queue_delete(worker->event, true);
 793                                                 /* drop reference from running event */
 794                                                 worker_unref(worker);
 795                                         }
 796                                 }
 797                                 worker_unref(worker);
 798                                 break;
 799                         }
 800                 }
 801                 break;
 802         case SIGHUP:
 803                 reload_config = true;
 804                 break;
 805         }
 806 }
 807
 808 static void static_dev_create_from_modules(struct udev *udev)
 809 {
 810         struct utsname kernel;
 811         char modules[UTIL_PATH_SIZE];
 812         char buf[4096];
 813         FILE *f;
 814
 815         uname(&kernel);
 816         util_strscpyl(modules, sizeof(modules), "/lib/modules/", kernel.release, "/modules.devname", NULL);
 817         f = fopen(modules, "r");
 818         if (f == NULL)
 819                 return;
 820
 821         while (fgets(buf, sizeof(buf), f) != NULL) {
 822                 char *s;
 823                 const char *modname;
 824                 const char *devname;
 825                 const char *devno;
 826                 int maj, min;
 827                 char type;
 828                 mode_t mode;
 829                 char filename[UTIL_PATH_SIZE];
 830
 831                 if (buf[0] == '#')
 832                         continue;
 833
 834                 modname = buf;
 835                 s = strchr(modname, ' ');
 836                 if (s == NULL)
 837                         continue;
 838                 s[0] = '\0';
 839
 840                 devname = &s[1];
 841                 s = strchr(devname, ' ');
 842                 if (s == NULL)
 843                         continue;
 844                 s[0] = '\0';
 845
 846                 devno = &s[1];
 847                 s = strchr(devno, ' ');
 848                 if (s == NULL)
 849                         s = strchr(devno, '\n');
 850                 if (s != NULL)
 851                         s[0] = '\0';
 852                 if (sscanf(devno, "%c%u:%u", &type, &maj, &min) != 3)
 853                         continue;
 854
 855                 if (type == 'c')
 856                         mode = 0600 | S_IFCHR;
 857                 else if (type == 'b')
 858                         mode = 0600 | S_IFBLK;
 859                 else
 860                         continue;
 861
 862                 util_strscpyl(filename, sizeof(filename), udev_get_dev_path(udev), "/", devname, NULL);
 863                 util_create_path_selinux(udev, filename);
 864                 udev_selinux_setfscreatecon(udev, filename, mode);
 865                 info(udev, "mknod '%s' %c%u:%u\n", filename, type, maj, min);
 866                 if (mknod(filename, mode, makedev(maj, min)) < 0 && errno == EEXIST)
 867                         utimensat(AT_FDCWD, filename, NULL, 0);
 868                 udev_selinux_resetfscreatecon(udev);
 869         }
 870
 871         fclose(f);
 872 }
 873
 874 static int copy_dev_dir(struct udev *udev, DIR *dir_from, DIR *dir_to, int maxdepth)
 875 {
 876         struct dirent *dent;
 877
 878         for (dent = readdir(dir_from); dent != NULL; dent = readdir(dir_from)) {
 879                 struct stat stats;
 880
 881                 if (dent->d_name[0] == '.')
 882                         continue;
 883                 if (fstatat(dirfd(dir_from), dent->d_name, &stats, AT_SYMLINK_NOFOLLOW) != 0)
 884                         continue;
 885
 886                 if (S_ISBLK(stats.st_mode) || S_ISCHR(stats.st_mode)) {
 887                         udev_selinux_setfscreateconat(udev, dirfd(dir_to), dent->d_name, stats.st_mode & 0777);
 888                         if (mknodat(dirfd(dir_to), dent->d_name, stats.st_mode, stats.st_rdev) == 0) {
 889                                 fchmodat(dirfd(dir_to), dent->d_name, stats.st_mode & 0777, 0);
 890                                 fchownat(dirfd(dir_to), dent->d_name, stats.st_uid, stats.st_gid, 0);
 891                         } else {
 892                                 utimensat(dirfd(dir_to), dent->d_name, NULL, 0);
 893                         }
 894                         udev_selinux_resetfscreatecon(udev);
 895                 } else if (S_ISLNK(stats.st_mode)) {
 896                         char target[UTIL_PATH_SIZE];
 897                         ssize_t len;
 898
 899                         len = readlinkat(dirfd(dir_from), dent->d_name, target, sizeof(target));
 900                         if (len <= 0 || len == (ssize_t)sizeof(target))
 901                                 continue;
 902                         target[len] = '\0';
 903                         udev_selinux_setfscreateconat(udev, dirfd(dir_to), dent->d_name, S_IFLNK);
 904                         if (symlinkat(target, dirfd(dir_to), dent->d_name) < 0 && errno == EEXIST)
 905                                 utimensat(dirfd(dir_to), dent->d_name, NULL, AT_SYMLINK_NOFOLLOW);
 906                         udev_selinux_resetfscreatecon(udev);
 907                 } else if (S_ISDIR(stats.st_mode)) {
 908                         DIR *dir2_from, *dir2_to;
 909
 910                         if (maxdepth == 0)
 911                                 continue;
 912
 913                         udev_selinux_setfscreateconat(udev, dirfd(dir_to), dent->d_name, S_IFDIR|0755);
 914                         mkdirat(dirfd(dir_to), dent->d_name, 0755);
 915                         udev_selinux_resetfscreatecon(udev);
 916
 917                         dir2_to = fdopendir(openat(dirfd(dir_to), dent->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC));
 918                         if (dir2_to == NULL)
 919                                 continue;
 920
 921                         dir2_from = fdopendir(openat(dirfd(dir_from), dent->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC));
 922                         if (dir2_from == NULL) {
 923                                 closedir(dir2_to);
 924                                 continue;
 925                         }
 926
 927                         copy_dev_dir(udev, dir2_from, dir2_to, maxdepth-1);
 928
 929                         closedir(dir2_to);
 930                         closedir(dir2_from);
 931                 }
 932         }
 933
 934         return 0;
 935 }
 936
 937 static void static_dev_create_links(struct udev *udev, DIR *dir)
 938 {
 939         struct stdlinks {
 940                 const char *link;
 941                 const char *target;
 942         };
 943         static const struct stdlinks stdlinks[] = {
 944                 { "core", "/proc/kcore" },
 945                 { "fd", "/proc/self/fd" },
 946                 { "stdin", "/proc/self/fd/0" },
 947                 { "stdout", "/proc/self/fd/1" },
 948                 { "stderr", "/proc/self/fd/2" },
 949         };
 950         unsigned int i;
 951
 952         for (i = 0; i < ARRAY_SIZE(stdlinks); i++) {
 953                 struct stat sb;
 954
 955                 if (stat(stdlinks[i].target, &sb) == 0) {
 956                         udev_selinux_setfscreateconat(udev, dirfd(dir), stdlinks[i].link, S_IFLNK);
 957                         if (symlinkat(stdlinks[i].target, dirfd(dir), stdlinks[i].link) < 0 && errno == EEXIST)
 958                                 utimensat(dirfd(dir), stdlinks[i].link, NULL, AT_SYMLINK_NOFOLLOW);
 959                         udev_selinux_resetfscreatecon(udev);
 960                 }
 961         }
 962 }
 963
 964 static void static_dev_create_from_devices(struct udev *udev, DIR *dir)
 965 {
 966         DIR *dir_from;
 967
 968         dir_from = opendir(LIBEXECDIR "/devices");
 969         if (dir_from == NULL)
 970                 return;
 971         copy_dev_dir(udev, dir_from, dir, 8);
 972         closedir(dir_from);
 973 }
 974
 975 static void static_dev_create(struct udev *udev)
 976 {
 977         DIR *dir;
 978
 979         dir = opendir(udev_get_dev_path(udev));
 980         if (dir == NULL)
 981                 return;
 982
 983         static_dev_create_links(udev, dir);
 984         static_dev_create_from_devices(udev, dir);
 985
 986         closedir(dir);
 987 }
 988
 989 static int mem_size_mb(void)
 990 {
 991         FILE *f;
 992         char buf[4096];
 993         long int memsize = -1;
 994
 995         f = fopen("/proc/meminfo", "r");
 996         if (f == NULL)
 997                 return -1;
 998
 999         while (fgets(buf, sizeof(buf), f) != NULL) {
1000                 long int value;
1001
1002                 if (sscanf(buf, "MemTotal: %ld kB", &value) == 1) {
1003                         memsize = value / 1024;
1004                         break;
1005                 }
1006         }
1007
1008         fclose(f);
1009         return memsize;
1010 }
1011
1012 static int convert_db(struct udev *udev)
1013 {
1014         char filename[UTIL_PATH_SIZE];
1015         FILE *f;
1016         struct udev_enumerate *udev_enumerate;
1017         struct udev_list_entry *list_entry;
1018
1019         /* current database */
1020         util_strscpyl(filename, sizeof(filename), udev_get_run_path(udev), "/data", NULL);
1021         if (access(filename, F_OK) >= 0)
1022                 return 0;
1023
1024         /* make sure we do not get here again */
1025         util_create_path(udev, filename);
1026         mkdir(filename, 0755);
1027
1028         /* old database */
1029         util_strscpyl(filename, sizeof(filename), udev_get_dev_path(udev), "/.udev/db", NULL);
1030         if (access(filename, F_OK) < 0)
1031                 return 0;
1032
1033         f = fopen("/dev/kmsg", "w");
1034         if (f != NULL) {
1035                 fprintf(f, "<30>udev[%u]: converting old udev database\n", getpid());
1036                 fclose(f);
1037         }
1038
1039         udev_enumerate = udev_enumerate_new(udev);
1040         if (udev_enumerate == NULL)
1041                 return -1;
1042         udev_enumerate_scan_devices(udev_enumerate);
1043         udev_list_entry_foreach(list_entry, udev_enumerate_get_list_entry(udev_enumerate)) {
1044                 struct udev_device *device;
1045
1046                 device = udev_device_new_from_syspath(udev, udev_list_entry_get_name(list_entry));
1047                 if (device == NULL)
1048                         continue;
1049
1050                 /* try to find the old database for devices without a current one */
1051                 if (udev_device_read_db(device, NULL) < 0) {
1052                         bool have_db;
1053                         const char *id;
1054                         struct stat stats;
1055                         char devpath[UTIL_PATH_SIZE];
1056                         char from[UTIL_PATH_SIZE];
1057
1058                         have_db = false;
1059
1060                         /* find database in old location */
1061                         id = udev_device_get_id_filename(device);
1062                         util_strscpyl(from, sizeof(from), udev_get_dev_path(udev), "/.udev/db/", id, NULL);
1063                         if (lstat(from, &stats) == 0) {
1064                                 if (!have_db) {
1065                                         udev_device_read_db(device, from);
1066                                         have_db = true;
1067                                 }
1068                                 unlink(from);
1069                         }
1070
1071                         /* find old database with $subsys:$sysname name */
1072                         util_strscpyl(from, sizeof(from), udev_get_dev_path(udev),
1073                                      "/.udev/db/", udev_device_get_subsystem(device), ":",
1074                                      udev_device_get_sysname(device), NULL);
1075                         if (lstat(from, &stats) == 0) {
1076                                 if (!have_db) {
1077                                         udev_device_read_db(device, from);
1078                                         have_db = true;
1079                                 }
1080                                 unlink(from);
1081                         }
1082
1083                         /* find old database with the encoded devpath name */
1084                         util_path_encode(udev_device_get_devpath(device), devpath, sizeof(devpath));
1085                         util_strscpyl(from, sizeof(from), udev_get_dev_path(udev), "/.udev/db/", devpath, NULL);
1086                         if (lstat(from, &stats) == 0) {
1087                                 if (!have_db) {
1088                                         udev_device_read_db(device, from);
1089                                         have_db = true;
1090                                 }
1091                                 unlink(from);
1092                         }
1093
1094                         /* write out new database */
1095                         if (have_db)
1096                                 udev_device_update_db(device);
1097                 }
1098                 udev_device_unref(device);
1099         }
1100         udev_enumerate_unref(udev_enumerate);
1101         return 0;
1102 }
1103
1104 int main(int argc, char *argv[])
1105 {
1106         struct udev *udev;
1107         int fd;
1108         FILE *f;
1109         sigset_t mask;
1110         int daemonize = false;
1111         int resolve_names = 1;
1112         static const struct option options[] = {
1113                 { "daemon", no_argument, NULL, 'd' },
1114                 { "debug", no_argument, NULL, 'D' },
1115                 { "children-max", required_argument, NULL, 'c' },
1116                 { "exec-delay", required_argument, NULL, 'e' },
1117                 { "resolve-names", required_argument, NULL, 'N' },
1118                 { "help", no_argument, NULL, 'h' },
1119                 { "version", no_argument, NULL, 'V' },
1120                 {}
1121         };
1122         int fd_ctrl = -1;
1123         int fd_netlink = -1;
1124         int fd_worker = -1;
1125         struct epoll_event ep_ctrl, ep_inotify, ep_signal, ep_netlink, ep_worker;
1126         struct udev_ctrl_connection *ctrl_conn = NULL;
1127         int rc = 1;
1128
1129         udev = udev_new();
1130         if (udev == NULL)
1131                 goto exit;
1132
1133         udev_log_init("udevd");
1134         udev_set_log_fn(udev, log_fn);
1135         info(udev, "version %s\n", VERSION);
1136         udev_selinux_init(udev);
1137
1138         /* make sure, that our runtime dir exists and is writable */
1139         if (utimensat(AT_FDCWD, udev_get_run_config_path(udev), NULL, 0) < 0) {
1140                 /* try to create our own subdirectory, do not create parent directories */
1141                 mkdir(udev_get_run_config_path(udev), 0755);
1142
1143                 if (utimensat(AT_FDCWD, udev_get_run_config_path(udev), NULL, 0) >= 0) {
1144                         /* directory seems writable now */
1145                         udev_set_run_path(udev, udev_get_run_config_path(udev));
1146                 } else {
1147                         /* fall back to /dev/.udev */
1148                         char filename[UTIL_PATH_SIZE];
1149
1150                         util_strscpyl(filename, sizeof(filename), udev_get_dev_path(udev), "/.udev", NULL);
1151                         if (udev_set_run_path(udev, filename) == NULL)
1152                                 goto exit;
1153                         mkdir(udev_get_run_path(udev), 0755);
1154                 }
1155         }
1156         /* relabel runtime dir only if it resides below /dev */
1157         if (strncmp(udev_get_run_path(udev), udev_get_dev_path(udev), strlen(udev_get_dev_path(udev))) == 0)
1158                 udev_selinux_lsetfilecon(udev, udev_get_run_path(udev), 0755);
1159         info(udev, "runtime dir '%s'\n", udev_get_run_path(udev));
1160
1161         for (;;) {
1162                 int option;
1163
1164                 option = getopt_long(argc, argv, "c:deDtN:hV", options, NULL);
1165                 if (option == -1)
1166                         break;
1167
1168                 switch (option) {
1169                 case 'd':
1170                         daemonize = true;
1171                         break;
1172                 case 'c':
1173                         children_max = strtoul(optarg, NULL, 0);
1174                         break;
1175                 case 'e':
1176                         exec_delay = strtoul(optarg, NULL, 0);
1177                         break;
1178                 case 'D':
1179                         debug = true;
1180                         if (udev_get_log_priority(udev) < LOG_INFO)
1181                                 udev_set_log_priority(udev, LOG_INFO);
1182                         break;
1183                 case 'N':
1184                         if (strcmp (optarg, "early") == 0) {
1185                                 resolve_names = 1;
1186                         } else if (strcmp (optarg, "late") == 0) {
1187                                 resolve_names = 0;
1188                         } else if (strcmp (optarg, "never") == 0) {
1189                                 resolve_names = -1;
1190                         } else {
1191                                 fprintf(stderr, "resolve-names must be early, late or never\n");
1192                                 err(udev, "resolve-names must be early, late or never\n");
1193                                 goto exit;
1194                         }
1195                         break;
1196                 case 'h':
1197                         printf("Usage: udevd OPTIONS\n"
1198                                "  --daemon\n"
1199                                "  --debug\n"
1200                                "  --children-max=<maximum number of workers>\n"
1201                                "  --exec-delay=<seconds to wait before executing RUN=>\n"
1202                                "  --resolve-names=early|late|never\n"
1203                                "  --version\n"
1204                                "  --help\n"
1205                                "\n");
1206                         goto exit;
1207                 case 'V':
1208                         printf("%s\n", VERSION);
1209                         goto exit;
1210                 default:
1211                         goto exit;
1212                 }
1213         }
1214
1215         /*
1216          * read the kernel commandline, in case we need to get into debug mode
1217          *   udev.log-priority=<level>              syslog priority
1218          *   udev.children-max=<number of workers>  events are fully serialized if set to 1
1219          *
1220          */
1221         f = fopen("/proc/cmdline", "r");
1222         if (f != NULL) {
1223                 char cmdline[4096];
1224
1225                 if (fgets(cmdline, sizeof(cmdline), f) != NULL) {
1226                         char *pos;
1227
1228                         pos = strstr(cmdline, "udev.log-priority=");
1229                         if (pos != NULL) {
1230                                 pos += strlen("udev.log-priority=");
1231                                 udev_set_log_priority(udev, util_log_priority(pos));
1232                         }
1233
1234                         pos = strstr(cmdline, "udev.children-max=");
1235                         if (pos != NULL) {
1236                                 pos += strlen("udev.children-max=");
1237                                 children_max = strtoul(pos, NULL, 0);
1238                         }
1239
1240                         pos = strstr(cmdline, "udev.exec-delay=");
1241                         if (pos != NULL) {
1242                                 pos += strlen("udev.exec-delay=");
1243                                 exec_delay = strtoul(pos, NULL, 0);
1244                         }
1245                 }
1246                 fclose(f);
1247         }
1248
1249         if (getuid() != 0) {
1250                 fprintf(stderr, "root privileges required\n");
1251                 err(udev, "root privileges required\n");
1252                 goto exit;
1253         }
1254
1255         /* set umask before creating any file/directory */
1256         chdir("/");
1257         umask(022);
1258
1259         /* create standard links, copy static nodes, create nodes from modules */
1260         static_dev_create(udev);
1261         static_dev_create_from_modules(udev);
1262
1263         /* before opening new files, make sure std{in,out,err} fds are in a sane state */
1264         fd = open("/dev/null", O_RDWR);
1265         if (fd < 0) {
1266                 fprintf(stderr, "cannot open /dev/null\n");
1267                 err(udev, "cannot open /dev/null\n");
1268         }
1269         if (write(STDOUT_FILENO, 0, 0) < 0)
1270                 dup2(fd, STDOUT_FILENO);
1271         if (write(STDERR_FILENO, 0, 0) < 0)
1272                 dup2(fd, STDERR_FILENO);
1273
1274         /* udevadm control socket */
1275         if (sd_listen_fds(true) == 1 && sd_is_socket(SD_LISTEN_FDS_START, AF_LOCAL, SOCK_SEQPACKET, -1))
1276                 udev_ctrl = udev_ctrl_new_from_fd(udev, SD_LISTEN_FDS_START);
1277         else
1278                 udev_ctrl = udev_ctrl_new_from_socket(udev, UDEV_CTRL_SOCK_PATH);
1279         if (udev_ctrl == NULL) {
1280                 fprintf(stderr, "error initializing udev control socket");
1281                 err(udev, "error initializing udev control socket");
1282                 rc = 1;
1283                 goto exit;
1284         }
1285         if (udev_ctrl_enable_receiving(udev_ctrl) < 0) {
1286                 fprintf(stderr, "error binding udev control socket\n");
1287                 err(udev, "error binding udev control socket\n");
1288                 rc = 1;
1289                 goto exit;
1290         }
1291         fd_ctrl = udev_ctrl_get_fd(udev_ctrl);
1292
1293         monitor = udev_monitor_new_from_netlink(udev, "kernel");
1294         if (monitor == NULL || udev_monitor_enable_receiving(monitor) < 0) {
1295                 fprintf(stderr, "error initializing netlink socket\n");
1296                 err(udev, "error initializing netlink socket\n");
1297                 rc = 3;
1298                 goto exit;
1299         }
1300         udev_monitor_set_receive_buffer_size(monitor, 128*1024*1024);
1301         fd_netlink = udev_monitor_get_fd(monitor);
1302
1303         if (daemonize) {
1304                 pid_t pid;
1305
1306                 pid = fork();
1307                 switch (pid) {
1308                 case 0:
1309                         break;
1310                 case -1:
1311                         err(udev, "fork of daemon failed: %m\n");
1312                         rc = 4;
1313                         goto exit;
1314                 default:
1315                         rc = 0;
1316                         goto exit;
1317                 }
1318         } else {
1319                 sd_notify(1, "READY=1");
1320         }
1321
1322         fd_inotify = udev_watch_init(udev);
1323         if (fd_inotify < 0) {
1324                 fprintf(stderr, "error initializing inotify\n");
1325                 err(udev, "error initializing inotify\n");
1326                 rc = 4;
1327                 goto exit;
1328         }
1329
1330         if (udev_get_rules_path(udev) != NULL) {
1331                 inotify_add_watch(fd_inotify, udev_get_rules_path(udev),
1332                                   IN_DELETE | IN_MOVE | IN_CLOSE_WRITE);
1333         } else {
1334                 char filename[UTIL_PATH_SIZE];
1335                 struct stat statbuf;
1336
1337                 inotify_add_watch(fd_inotify, LIBEXECDIR "/rules.d",
1338                                   IN_DELETE | IN_MOVE | IN_CLOSE_WRITE);
1339                 inotify_add_watch(fd_inotify, SYSCONFDIR "/udev/rules.d",
1340                                   IN_DELETE | IN_MOVE | IN_CLOSE_WRITE);
1341
1342                 /* watch dynamic rules directory */
1343                 util_strscpyl(filename, sizeof(filename), udev_get_run_path(udev), "/rules.d", NULL);
1344                 if (stat(filename, &statbuf) != 0) {
1345                         util_create_path(udev, filename);
1346                         mkdir(filename, 0755);
1347                 }
1348                 inotify_add_watch(fd_inotify, filename,
1349                                   IN_DELETE | IN_MOVE | IN_CLOSE_WRITE);
1350         }
1351         udev_watch_restore(udev);
1352
1353         /* block and listen to all signals on signalfd */
1354         sigfillset(&mask);
1355         sigprocmask(SIG_SETMASK, &mask, &orig_sigmask);
1356         fd_signal = signalfd(-1, &mask, SFD_CLOEXEC);
1357         if (fd_signal < 0) {
1358                 fprintf(stderr, "error getting signalfd\n");
1359                 err(udev, "error getting signalfd\n");
1360                 rc = 5;
1361                 goto exit;
1362         }
1363
1364         /* unnamed socket from workers to the main daemon */
1365         if (socketpair(AF_LOCAL, SOCK_DGRAM|SOCK_CLOEXEC, 0, worker_watch) < 0) {
1366                 fprintf(stderr, "error getting socketpair\n");
1367                 err(udev, "error getting socketpair\n");
1368                 rc = 6;
1369                 goto exit;
1370         }
1371         fd_worker = worker_watch[READ_END];
1372
1373         rules = udev_rules_new(udev, resolve_names);
1374         if (rules == NULL) {
1375                 err(udev, "error reading rules\n");
1376                 goto exit;
1377         }
1378
1379         udev_queue_export = udev_queue_export_new(udev);
1380         if (udev_queue_export == NULL) {
1381                 err(udev, "error creating queue file\n");
1382                 goto exit;
1383         }
1384
1385         memset(&ep_ctrl, 0, sizeof(struct epoll_event));
1386         ep_ctrl.events = EPOLLIN;
1387         ep_ctrl.data.fd = fd_ctrl;
1388         memset(&ep_inotify, 0, sizeof(struct epoll_event));
1389         ep_inotify.events = EPOLLIN;
1390         ep_inotify.data.fd = fd_inotify;
1391         memset(&ep_signal, 0, sizeof(struct epoll_event));
1392         ep_signal.events = EPOLLIN;
1393         ep_signal.data.fd = fd_signal;
1394         memset(&ep_netlink, 0, sizeof(struct epoll_event));
1395         ep_netlink.events = EPOLLIN;
1396         ep_netlink.data.fd = fd_netlink;
1397         memset(&ep_worker, 0, sizeof(struct epoll_event));
1398         ep_worker.events = EPOLLIN;
1399         ep_worker.data.fd = fd_worker;
1400         fd_ep = epoll_create1(EPOLL_CLOEXEC);
1401         if (fd_ep < 0) {
1402                 err(udev, "error creating epoll fd: %m\n");
1403                 goto exit;
1404         }
1405         if (epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_ctrl, &ep_ctrl) < 0 ||
1406             epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_inotify, &ep_inotify) < 0 ||
1407             epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_signal, &ep_signal) < 0 ||
1408             epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_netlink, &ep_netlink) < 0 ||
1409             epoll_ctl(fd_ep, EPOLL_CTL_ADD, fd_worker, &ep_worker) < 0) {
1410                 err(udev, "fail to add fds to epoll: %m\n");
1411                 goto exit;
1412         }
1413
1414         /* if needed, convert old database from earlier udev version */
1415         convert_db(udev);
1416
1417         if (!debug) {
1418                 dup2(fd, STDIN_FILENO);
1419                 dup2(fd, STDOUT_FILENO);
1420                 dup2(fd, STDERR_FILENO);
1421         }
1422         if (fd > STDERR_FILENO)
1423                 close(fd);
1424
1425         /* set scheduling priority for the main daemon process */
1426         setpriority(PRIO_PROCESS, 0, UDEVD_PRIORITY);
1427
1428         setsid();
1429
1430         f = fopen("/dev/kmsg", "w");
1431         if (f != NULL) {
1432                 fprintf(f, "<30>udev[%u]: starting version " VERSION "\n", getpid());
1433                 fclose(f);
1434         }
1435
1436         fd = open("/proc/self/oom_score_adj", O_RDWR);
1437         if (fd < 0) {
1438                 /* Fallback to old interface */
1439                 fd = open("/proc/self/oom_adj", O_RDWR);
1440                 if (fd < 0) {
1441                         err(udev, "error disabling OOM: %m\n");
1442                 } else {
1443                         /* OOM_DISABLE == -17 */
1444                         write(fd, "-17", 3);
1445                         close(fd);
1446                 }
1447         } else {
1448                 write(fd, "-1000", 5);
1449                 close(fd);
1450         }
1451
1452         if (children_max <= 0) {
1453                 int memsize = mem_size_mb();
1454
1455                 /* set value depending on the amount of RAM */
1456                 if (memsize > 0)
1457                         children_max = 128 + (memsize / 8);
1458                 else
1459                         children_max = 128;
1460         }
1461         info(udev, "set children_max to %u\n", children_max);
1462
1463         udev_rules_apply_static_dev_perms(rules);
1464
1465         udev_list_init(&event_list);
1466         udev_list_init(&worker_list);
1467
1468         for (;;) {
1469                 struct epoll_event ev[8];
1470                 int fdcount;
1471                 int timeout;
1472                 bool is_worker, is_signal, is_inotify, is_netlink, is_ctrl;
1473                 int i;
1474
1475                 if (udev_exit) {
1476                         /* close sources of new events and discard buffered events */
1477                         if (fd_ctrl >= 0) {
1478                                 epoll_ctl(fd_ep, EPOLL_CTL_DEL, fd_ctrl, NULL);
1479                                 fd_ctrl = -1;
1480                         }
1481                         if (monitor != NULL) {
1482                                 epoll_ctl(fd_ep, EPOLL_CTL_DEL, fd_netlink, NULL);
1483                                 udev_monitor_unref(monitor);
1484                                 monitor = NULL;
1485                         }
1486                         if (fd_inotify >= 0) {
1487                                 epoll_ctl(fd_ep, EPOLL_CTL_DEL, fd_inotify, NULL);
1488                                 close(fd_inotify);
1489                                 fd_inotify = -1;
1490                         }
1491
1492                         /* discard queued events and kill workers */
1493                         event_queue_cleanup(udev, EVENT_QUEUED);
1494                         worker_kill(udev, 0);
1495
1496                         /* exit after all has cleaned up */
1497                         if (udev_list_is_empty(&event_list) && udev_list_is_empty(&worker_list))
1498                                 break;
1499
1500                         /* timeout at exit for workers to finish */
1501                         timeout = 60 * 1000;
1502                 } else if (udev_list_is_empty(&event_list) && children > 2) {
1503                         /* set timeout to kill idle workers */
1504                         timeout = 3 * 1000;
1505                 } else {
1506                         timeout = -1;
1507                 }
1508                 fdcount = epoll_wait(fd_ep, ev, ARRAY_SIZE(ev), timeout);
1509                 if (fdcount < 0)
1510                         continue;
1511
1512                 if (fdcount == 0) {
1513                         if (udev_exit) {
1514                                 info(udev, "timeout, giving up waiting for workers to finish\n");
1515                                 break;
1516                         }
1517
1518                         /* timeout - kill idle workers */
1519                         worker_kill(udev, 2);
1520                 }
1521
1522                 is_worker = is_signal = is_inotify = is_netlink = is_ctrl = false;
1523                 for (i = 0; i < fdcount; i++) {
1524                         if (ev[i].data.fd == fd_worker && ev[i].events & EPOLLIN)
1525                                 is_worker = true;
1526                         else if (ev[i].data.fd == fd_netlink && ev[i].events & EPOLLIN)
1527                                 is_netlink = true;
1528                         else if (ev[i].data.fd == fd_signal && ev[i].events & EPOLLIN)
1529                                 is_signal = true;
1530                         else if (ev[i].data.fd == fd_inotify && ev[i].events & EPOLLIN)
1531                                 is_inotify = true;
1532                         else if (ev[i].data.fd == fd_ctrl && ev[i].events & EPOLLIN)
1533                                 is_ctrl = true;
1534                 }
1535
1536                 /* event has finished */
1537                 if (is_worker)
1538                         worker_returned(fd_worker);
1539
1540                 if (is_netlink) {
1541                         struct udev_device *dev;
1542
1543                         dev = udev_monitor_receive_device(monitor);
1544                         if (dev != NULL)
1545                                 if (event_queue_insert(dev) < 0)
1546                                         udev_device_unref(dev);
1547                 }
1548
1549                 /* start new events */
1550                 if (!udev_list_is_empty(&event_list) && !udev_exit && !stop_exec_queue)
1551                         event_queue_start(udev);
1552
1553                 if (is_signal) {
1554                         struct signalfd_siginfo fdsi;
1555                         ssize_t size;
1556
1557                         size = read(fd_signal, &fdsi, sizeof(struct signalfd_siginfo));
1558                         if (size == sizeof(struct signalfd_siginfo))
1559                                 handle_signal(udev, fdsi.ssi_signo);
1560                 }
1561
1562                 /* we are shutting down, the events below are not handled anymore */
1563                 if (udev_exit)
1564                         continue;
1565
1566                 /* device node and rules directory inotify watch */
1567                 if (is_inotify)
1568                         handle_inotify(udev);
1569
1570                 /*
1571                  * This needs to be after the inotify handling, to make sure,
1572                  * that the ping is send back after the possibly generated
1573                  * "change" events by the inotify device node watch.
1574                  *
1575                  * A single time we may receive a client connection which we need to
1576                  * keep open to block the client. It will be closed right before we
1577                  * exit.
1578                  */
1579                 if (is_ctrl)
1580                         ctrl_conn = handle_ctrl_msg(udev_ctrl);
1581
1582                 /* rules changed, set by inotify or a HUP signal */
1583                 if (reload_config) {
1584                         struct udev_rules *rules_new;
1585
1586                         worker_kill(udev, 0);
1587                         rules_new = udev_rules_new(udev, resolve_names);
1588                         if (rules_new != NULL) {
1589                                 udev_rules_unref(rules);
1590                                 rules = rules_new;
1591                         }
1592                         reload_config = 0;
1593                 }
1594         }
1595
1596         udev_queue_export_cleanup(udev_queue_export);
1597         rc = 0;
1598 exit:
1599         if (fd_ep >= 0)
1600                 close(fd_ep);
1601         worker_list_cleanup(udev);
1602         event_queue_cleanup(udev, EVENT_UNDEF);
1603         udev_rules_unref(rules);
1604         if (fd_signal >= 0)
1605                 close(fd_signal);
1606         if (worker_watch[READ_END] >= 0)
1607                 close(worker_watch[READ_END]);
1608         if (worker_watch[WRITE_END] >= 0)
1609                 close(worker_watch[WRITE_END]);
1610         udev_monitor_unref(monitor);
1611         udev_queue_export_unref(udev_queue_export);
1612         udev_ctrl_connection_unref(ctrl_conn);
1613         udev_ctrl_unref(udev_ctrl);
1614         udev_selinux_exit(udev);
1615         udev_unref(udev);
1616         udev_log_close();
1617         return rc;
1618 }