lib/daemon.c

   1 /*
   2  * Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira, Inc.
   3  *
   4  * Licensed under the Apache License, Version 2.0 (the "License");
   5  * you may not use this file except in compliance with the License.
   6  * You may obtain a copy of the License at:
   7  *
   8  *     http://www.apache.org/licenses/LICENSE-2.0
   9  *
  10  * Unless required by applicable law or agreed to in writing, software
  11  * distributed under the License is distributed on an "AS IS" BASIS,
  12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  * See the License for the specific language governing permissions and
  14  * limitations under the License.
  15  */
  16
  17 #include <config.h>
  18 #include "daemon.h"
  19 #include <assert.h>
  20 #include <errno.h>
  21 #include <fcntl.h>
  22 #include <signal.h>
  23 #include <stdlib.h>
  24 #include <string.h>
  25 #include <sys/resource.h>
  26 #include <sys/wait.h>
  27 #include <sys/stat.h>
  28 #include <unistd.h>
  29 #include "command-line.h"
  30 #include "fatal-signal.h"
  31 #include "dirs.h"
  32 #include "lockfile.h"
  33 #include "process.h"
  34 #include "socket-util.h"
  35 #include "timeval.h"
  36 #include "util.h"
  37 #include "vlog.h"
  38
  39 VLOG_DEFINE_THIS_MODULE(daemon);
  40
  41 /* --detach: Should we run in the background? */
  42 static bool detach;
  43
  44 /* --pidfile: Name of pidfile (null if none). */
  45 static char *pidfile;
  46
  47 /* Device and inode of pidfile, so we can avoid reopening it. */
  48 static dev_t pidfile_dev;
  49 static ino_t pidfile_ino;
  50
  51 /* --overwrite-pidfile: Create pidfile even if one already exists and is
  52    locked? */
  53 static bool overwrite_pidfile;
  54
  55 /* --no-chdir: Should we chdir to "/"? */
  56 static bool chdir_ = true;
  57
  58 /* File descriptor used by daemonize_start() and daemonize_complete(). */
  59 static int daemonize_fd = -1;
  60
  61 /* --monitor: Should a supervisory process monitor the daemon and restart it if
  62  * it dies due to an error signal? */
  63 static bool monitor;
  64
  65 /* For each of the standard file descriptors, whether to replace it by
  66  * /dev/null (if false) or keep it for the daemon to use (if true). */
  67 static bool save_fds[3];
  68
  69 static void check_already_running(void);
  70 static int lock_pidfile(FILE *, int command);
  71
  72 /* Returns the file name that would be used for a pidfile if 'name' were
  73  * provided to set_pidfile().  The caller must free the returned string. */
  74 char *
  75 make_pidfile_name(const char *name)
  76 {
  77     return (!name
  78             ? xasprintf("%s/%s.pid", ovs_rundir(), program_name)
  79             : abs_file_name(ovs_rundir(), name));
  80 }
  81
  82 /* Sets up a following call to daemonize() to create a pidfile named 'name'.
  83  * If 'name' begins with '/', then it is treated as an absolute path.
  84  * Otherwise, it is taken relative to RUNDIR, which is $(prefix)/var/run by
  85  * default.
  86  *
  87  * If 'name' is null, then program_name followed by ".pid" is used. */
  88 void
  89 set_pidfile(const char *name)
  90 {
  91     free(pidfile);
  92     pidfile = make_pidfile_name(name);
  93 }
  94
  95 /* Returns an absolute path to the configured pidfile, or a null pointer if no
  96  * pidfile is configured.  The caller must not modify or free the returned
  97  * string. */
  98 const char *
  99 get_pidfile(void)
 100 {
 101     return pidfile;
 102 }
 103
 104 /* Sets that we do not chdir to "/". */
 105 void
 106 set_no_chdir(void)
 107 {
 108     chdir_ = false;
 109 }
 110
 111 /* Will we chdir to "/" as part of daemonizing? */
 112 bool
 113 is_chdir_enabled(void)
 114 {
 115     return chdir_;
 116 }
 117
 118 /* Normally, daemonize() or damonize_start() will terminate the program with a
 119  * message if a locked pidfile already exists.  If this function is called, an
 120  * existing pidfile will be replaced, with a warning. */
 121 void
 122 ignore_existing_pidfile(void)
 123 {
 124     overwrite_pidfile = true;
 125 }
 126
 127 /* Sets up a following call to daemonize() to detach from the foreground
 128  * session, running this process in the background.  */
 129 void
 130 set_detach(void)
 131 {
 132     detach = true;
 133 }
 134
 135 /* Will daemonize() really detach? */
 136 bool
 137 get_detach(void)
 138 {
 139     return detach;
 140 }
 141
 142 /* Sets up a following call to daemonize() to fork a supervisory process to
 143  * monitor the daemon and restart it if it dies due to an error signal.  */
 144 void
 145 daemon_set_monitor(void)
 146 {
 147     monitor = true;
 148 }
 149
 150 /* A daemon doesn't normally have any use for the file descriptors for stdin,
 151  * stdout, and stderr after it detaches.  To keep these file descriptors from
 152  * e.g. holding an SSH session open, by default detaching replaces each of
 153  * these file descriptors by /dev/null.  But a few daemons expect the user to
 154  * redirect stdout or stderr to a file, in which case it is desirable to keep
 155  * these file descriptors.  This function, therefore, disables replacing 'fd'
 156  * by /dev/null when the daemon detaches. */
 157 void
 158 daemon_save_fd(int fd)
 159 {
 160     assert(fd == STDIN_FILENO || fd == STDOUT_FILENO || fd == STDERR_FILENO);
 161     save_fds[fd] = true;
 162 }
 163
 164 /* If a pidfile has been configured, creates it and stores the running
 165  * process's pid in it.  Ensures that the pidfile will be deleted when the
 166  * process exits. */
 167 static void
 168 make_pidfile(void)
 169 {
 170     long int pid = getpid();
 171     struct stat s;
 172     char *tmpfile;
 173     FILE *file;
 174     int error;
 175
 176     /* Create a temporary pidfile. */
 177     tmpfile = xasprintf("%s.tmp%ld", pidfile, pid);
 178     fatal_signal_add_file_to_unlink(tmpfile);
 179     file = fopen(tmpfile, "w+");
 180     if (!file) {
 181         VLOG_FATAL("%s: create failed (%s)", tmpfile, strerror(errno));
 182     }
 183
 184     if (fstat(fileno(file), &s) == -1) {
 185         VLOG_FATAL("%s: fstat failed (%s)", tmpfile, strerror(errno));
 186     }
 187
 188     fprintf(file, "%ld\n", pid);
 189     if (fflush(file) == EOF) {
 190         VLOG_FATAL("%s: write failed (%s)", tmpfile, strerror(errno));
 191     }
 192
 193     error = lock_pidfile(file, F_SETLK);
 194     if (error) {
 195         VLOG_FATAL("%s: fcntl(F_SETLK) failed (%s)", tmpfile, strerror(error));
 196     }
 197
 198     /* Rename or link it to the correct name. */
 199     if (overwrite_pidfile) {
 200         if (rename(tmpfile, pidfile) < 0) {
 201             VLOG_FATAL("failed to rename \"%s\" to \"%s\" (%s)",
 202                        tmpfile, pidfile, strerror(errno));
 203         }
 204     } else {
 205         do {
 206             error = link(tmpfile, pidfile) == -1 ? errno : 0;
 207             if (error == EEXIST) {
 208                 check_already_running();
 209             }
 210         } while (error == EINTR || error == EEXIST);
 211         if (error) {
 212             VLOG_FATAL("failed to link \"%s\" as \"%s\" (%s)",
 213                        tmpfile, pidfile, strerror(error));
 214         }
 215     }
 216
 217     /* Ensure that the pidfile will get deleted on exit. */
 218     fatal_signal_add_file_to_unlink(pidfile);
 219
 220     /* Delete the temporary pidfile if it still exists. */
 221     if (!overwrite_pidfile) {
 222         error = fatal_signal_unlink_file_now(tmpfile);
 223         if (error) {
 224             VLOG_FATAL("%s: unlink failed (%s)", tmpfile, strerror(error));
 225         }
 226     }
 227
 228     /* Clean up.
 229      *
 230      * We don't close 'file' because its file descriptor must remain open to
 231      * hold the lock. */
 232     pidfile_dev = s.st_dev;
 233     pidfile_ino = s.st_ino;
 234     free(tmpfile);
 235     free(pidfile);
 236     pidfile = NULL;
 237 }
 238
 239 /* If configured with set_pidfile() or set_detach(), creates the pid file and
 240  * detaches from the foreground session.  */
 241 void
 242 daemonize(void)
 243 {
 244     daemonize_start();
 245     daemonize_complete();
 246 }
 247
 248 /* Calls fork() and on success returns its return value.  On failure, logs an
 249  * error and exits unsuccessfully.
 250  *
 251  * Post-fork, but before returning, this function calls a few other functions
 252  * that are generally useful if the child isn't planning to exec a new
 253  * process. */
 254 pid_t
 255 fork_and_clean_up(void)
 256 {
 257     pid_t pid;
 258
 259     pid = fork();
 260     if (pid > 0) {
 261         /* Running in parent process. */
 262         fatal_signal_fork();
 263     } else if (!pid) {
 264         /* Running in child process. */
 265         time_postfork();
 266         lockfile_postfork();
 267     } else {
 268         VLOG_FATAL("fork failed (%s)", strerror(errno));
 269     }
 270
 271     return pid;
 272 }
 273
 274 /* Forks, then:
 275  *
 276  *   - In the parent, waits for the child to signal that it has completed its
 277  *     startup sequence.  Then stores -1 in '*fdp' and returns the child's pid.
 278  *
 279  *   - In the child, stores a fd in '*fdp' and returns 0.  The caller should
 280  *     pass the fd to fork_notify_startup() after it finishes its startup
 281  *     sequence.
 282  *
 283  * If something goes wrong with the fork, logs a critical error and aborts the
 284  * process. */
 285 static pid_t
 286 fork_and_wait_for_startup(int *fdp)
 287 {
 288     int fds[2];
 289     pid_t pid;
 290
 291     xpipe(fds);
 292
 293     pid = fork_and_clean_up();
 294     if (pid > 0) {
 295         /* Running in parent process. */
 296         size_t bytes_read;
 297         char c;
 298
 299         close(fds[1]);
 300         if (read_fully(fds[0], &c, 1, &bytes_read) != 0) {
 301             int retval;
 302             int status;
 303
 304             do {
 305                 retval = waitpid(pid, &status, 0);
 306             } while (retval == -1 && errno == EINTR);
 307
 308             if (retval == pid) {
 309                 if (WIFEXITED(status) && WEXITSTATUS(status)) {
 310                     /* Child exited with an error.  Convey the same error
 311                      * to our parent process as a courtesy. */
 312                     exit(WEXITSTATUS(status));
 313                 } else {
 314                     char *status_msg = process_status_msg(status);
 315                     VLOG_FATAL("fork child died before signaling startup (%s)",
 316                                status_msg);
 317                 }
 318             } else if (retval < 0) {
 319                 VLOG_FATAL("waitpid failed (%s)", strerror(errno));
 320             } else {
 321                 NOT_REACHED();
 322             }
 323         }
 324         close(fds[0]);
 325         *fdp = -1;
 326     } else if (!pid) {
 327         /* Running in child process. */
 328         close(fds[0]);
 329         *fdp = fds[1];
 330     }
 331
 332     return pid;
 333 }
 334
 335 static void
 336 fork_notify_startup(int fd)
 337 {
 338     if (fd != -1) {
 339         size_t bytes_written;
 340         int error;
 341
 342         error = write_fully(fd, "", 1, &bytes_written);
 343         if (error) {
 344             VLOG_FATAL("pipe write failed (%s)", strerror(error));
 345         }
 346
 347         close(fd);
 348     }
 349 }
 350
 351 static bool
 352 should_restart(int status)
 353 {
 354     if (WIFSIGNALED(status)) {
 355         static const int error_signals[] = {
 356             SIGABRT, SIGALRM, SIGBUS, SIGFPE, SIGILL, SIGPIPE, SIGSEGV,
 357             SIGXCPU, SIGXFSZ
 358         };
 359
 360         size_t i;
 361
 362         for (i = 0; i < ARRAY_SIZE(error_signals); i++) {
 363             if (error_signals[i] == WTERMSIG(status)) {
 364                 return true;
 365             }
 366         }
 367     }
 368     return false;
 369 }
 370
 371 static void
 372 monitor_daemon(pid_t daemon_pid)
 373 {
 374     /* XXX Should log daemon's stderr output at startup time. */
 375     time_t last_restart;
 376     char *status_msg;
 377     int crashes;
 378
 379     subprogram_name = "monitor";
 380     status_msg = xstrdup("healthy");
 381     last_restart = TIME_MIN;
 382     crashes = 0;
 383     for (;;) {
 384         int retval;
 385         int status;
 386
 387         proctitle_set("%s: monitoring pid %lu (%s)",
 388                       program_name, (unsigned long int) daemon_pid,
 389                       status_msg);
 390
 391         do {
 392             retval = waitpid(daemon_pid, &status, 0);
 393         } while (retval == -1 && errno == EINTR);
 394
 395         if (retval == -1) {
 396             VLOG_FATAL("waitpid failed (%s)", strerror(errno));
 397         } else if (retval == daemon_pid) {
 398             char *s = process_status_msg(status);
 399             if (should_restart(status)) {
 400                 free(status_msg);
 401                 status_msg = xasprintf("%d crashes: pid %lu died, %s",
 402                                        ++crashes,
 403                                        (unsigned long int) daemon_pid, s);
 404                 free(s);
 405
 406                 if (WCOREDUMP(status)) {
 407                     /* Disable further core dumps to save disk space. */
 408                     struct rlimit r;
 409
 410                     r.rlim_cur = 0;
 411                     r.rlim_max = 0;
 412                     if (setrlimit(RLIMIT_CORE, &r) == -1) {
 413                         VLOG_WARN("failed to disable core dumps: %s",
 414                                   strerror(errno));
 415                     }
 416                 }
 417
 418                 /* Throttle restarts to no more than once every 10 seconds. */
 419                 if (time(NULL) < last_restart + 10) {
 420                     VLOG_WARN("%s, waiting until 10 seconds since last "
 421                               "restart", status_msg);
 422                     for (;;) {
 423                         time_t now = time(NULL);
 424                         time_t wakeup = last_restart + 10;
 425                         if (now >= wakeup) {
 426                             break;
 427                         }
 428                         sleep(wakeup - now);
 429                     }
 430                 }
 431                 last_restart = time(NULL);
 432
 433                 VLOG_ERR("%s, restarting", status_msg);
 434                 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
 435                 if (!daemon_pid) {
 436                     break;
 437                 }
 438             } else {
 439                 VLOG_INFO("pid %lu died, %s, exiting",
 440                           (unsigned long int) daemon_pid, s);
 441                 free(s);
 442                 exit(0);
 443             }
 444         }
 445     }
 446     free(status_msg);
 447
 448     /* Running in new daemon process. */
 449     proctitle_restore();
 450     subprogram_name = "";
 451 }
 452
 453 /* Close standard file descriptors (except any that the client has requested we
 454  * leave open by calling daemon_save_fd()).  If we're started from e.g. an SSH
 455  * session, then this keeps us from holding that session open artificially. */
 456 static void
 457 close_standard_fds(void)
 458 {
 459     int null_fd = get_null_fd();
 460     if (null_fd >= 0) {
 461         int fd;
 462
 463         for (fd = 0; fd < 3; fd++) {
 464             if (!save_fds[fd]) {
 465                 dup2(null_fd, fd);
 466             }
 467         }
 468     }
 469
 470     /* Disable logging to stderr to avoid wasting CPU time. */
 471     vlog_set_levels(NULL, VLF_CONSOLE, VLL_OFF);
 472 }
 473
 474 /* If daemonization is configured, then starts daemonization, by forking and
 475  * returning in the child process.  The parent process hangs around until the
 476  * child lets it know either that it completed startup successfully (by calling
 477  * daemon_complete()) or that it failed to start up (by exiting with a nonzero
 478  * exit code). */
 479 void
 480 daemonize_start(void)
 481 {
 482     daemonize_fd = -1;
 483
 484     if (detach) {
 485         if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
 486             /* Running in parent process. */
 487             exit(0);
 488         }
 489         /* Running in daemon or monitor process. */
 490     }
 491
 492     if (monitor) {
 493         int saved_daemonize_fd = daemonize_fd;
 494         pid_t daemon_pid;
 495
 496         daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
 497         if (daemon_pid > 0) {
 498             /* Running in monitor process. */
 499             fork_notify_startup(saved_daemonize_fd);
 500             close_standard_fds();
 501             monitor_daemon(daemon_pid);
 502         }
 503         /* Running in daemon process. */
 504     }
 505
 506     if (pidfile) {
 507         make_pidfile();
 508     }
 509
 510     /* Make sure that the unixctl commands for vlog get registered in a
 511      * daemon, even before the first log message. */
 512     vlog_init();
 513 }
 514
 515 /* If daemonization is configured, then this function notifies the parent
 516  * process that the child process has completed startup successfully.
 517  *
 518  * Calling this function more than once has no additional effect. */
 519 void
 520 daemonize_complete(void)
 521 {
 522     fork_notify_startup(daemonize_fd);
 523     daemonize_fd = -1;
 524
 525     if (detach) {
 526         setsid();
 527         if (chdir_) {
 528             ignore(chdir("/"));
 529         }
 530         close_standard_fds();
 531         detach = false;
 532     }
 533 }
 534
 535 void
 536 daemon_usage(void)
 537 {
 538     printf(
 539         "\nDaemon options:\n"
 540         "  --detach                run in background as daemon\n"
 541         "  --no-chdir              do not chdir to '/'\n"
 542         "  --pidfile[=FILE]        create pidfile (default: %s/%s.pid)\n"
 543         "  --overwrite-pidfile     with --pidfile, start even if already "
 544                                    "running\n",
 545         ovs_rundir(), program_name);
 546 }
 547
 548 static int
 549 lock_pidfile__(FILE *file, int command, struct flock *lck)
 550 {
 551     int error;
 552
 553     lck->l_type = F_WRLCK;
 554     lck->l_whence = SEEK_SET;
 555     lck->l_start = 0;
 556     lck->l_len = 0;
 557     lck->l_pid = 0;
 558
 559     do {
 560         error = fcntl(fileno(file), command, lck) == -1 ? errno : 0;
 561     } while (error == EINTR);
 562     return error;
 563 }
 564
 565 static int
 566 lock_pidfile(FILE *file, int command)
 567 {
 568     struct flock lck;
 569
 570     return lock_pidfile__(file, command, &lck);
 571 }
 572
 573 static pid_t
 574 read_pidfile__(const char *pidfile, bool delete_if_stale)
 575 {
 576     struct stat s, s2;
 577     struct flock lck;
 578     char line[128];
 579     FILE *file;
 580     int error;
 581
 582     if ((pidfile_ino || pidfile_dev)
 583         && !stat(pidfile, &s)
 584         && s.st_ino == pidfile_ino && s.st_dev == pidfile_dev) {
 585         /* It's our own pidfile.  We can't afford to open it, because closing
 586          * *any* fd for a file that a process has locked also releases all the
 587          * locks on that file.
 588          *
 589          * Fortunately, we know the associated pid anyhow: */
 590         return getpid();
 591     }
 592
 593     file = fopen(pidfile, "r+");
 594     if (!file) {
 595         if (errno == ENOENT && delete_if_stale) {
 596             return 0;
 597         }
 598         error = errno;
 599         VLOG_WARN("%s: open: %s", pidfile, strerror(error));
 600         goto error;
 601     }
 602
 603     error = lock_pidfile__(file, F_GETLK, &lck);
 604     if (error) {
 605         VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
 606         goto error;
 607     }
 608     if (lck.l_type == F_UNLCK) {
 609         /* pidfile exists but it isn't locked by anyone.  We need to delete it
 610          * so that a new pidfile can go in its place.  But just calling
 611          * unlink(pidfile) makes a nasty race: what if someone else unlinks it
 612          * before we do and then replaces it by a valid pidfile?  We'd unlink
 613          * their valid pidfile.  We do a little dance to avoid the race, by
 614          * locking the invalid pidfile.  Only one process can have the invalid
 615          * pidfile locked, and only that process has the right to unlink it. */
 616         if (!delete_if_stale) {
 617             error = ESRCH;
 618             VLOG_DBG("%s: pid file is stale", pidfile);
 619             goto error;
 620         }
 621
 622         /* Get the lock. */
 623         error = lock_pidfile(file, F_SETLK);
 624         if (error) {
 625             /* We lost a race with someone else doing the same thing. */
 626             VLOG_WARN("%s: lost race to lock pidfile", pidfile);
 627             goto error;
 628         }
 629
 630         /* Is the file we have locked still named 'pidfile'? */
 631         if (stat(pidfile, &s) || fstat(fileno(file), &s2)
 632             || s.st_ino != s2.st_ino || s.st_dev != s2.st_dev) {
 633             /* No.  We lost a race with someone else who got the lock before
 634              * us, deleted the pidfile, and closed it (releasing the lock). */
 635             error = EALREADY;
 636             VLOG_WARN("%s: lost race to delete pidfile", pidfile);
 637             goto error;
 638         }
 639
 640         /* We won the right to delete the stale pidfile. */
 641         if (unlink(pidfile)) {
 642             error = errno;
 643             VLOG_WARN("%s: failed to delete stale pidfile (%s)",
 644                       pidfile, strerror(error));
 645             goto error;
 646         }
 647         VLOG_DBG("%s: deleted stale pidfile", pidfile);
 648         fclose(file);
 649         return 0;
 650     }
 651
 652     if (!fgets(line, sizeof line, file)) {
 653         if (ferror(file)) {
 654             error = errno;
 655             VLOG_WARN("%s: read: %s", pidfile, strerror(error));
 656         } else {
 657             error = ESRCH;
 658             VLOG_WARN("%s: read: unexpected end of file", pidfile);
 659         }
 660         goto error;
 661     }
 662
 663     if (lck.l_pid != strtoul(line, NULL, 10)) {
 664         /* The process that has the pidfile locked is not the process that
 665          * created it.  It must be stale, with the process that has it locked
 666          * preparing to delete it. */
 667         error = ESRCH;
 668         VLOG_WARN("%s: stale pidfile for pid %s being deleted by pid %ld",
 669                   pidfile, line, (long int) lck.l_pid);
 670         goto error;
 671     }
 672
 673     fclose(file);
 674     return lck.l_pid;
 675
 676 error:
 677     if (file) {
 678         fclose(file);
 679     }
 680     return -error;
 681 }
 682
 683 /* Opens and reads a PID from 'pidfile'.  Returns the positive PID if
 684  * successful, otherwise a negative errno value. */
 685 pid_t
 686 read_pidfile(const char *pidfile)
 687 {
 688     return read_pidfile__(pidfile, false);
 689 }
 690
 691 /* Checks whether a process with the given 'pidfile' is already running and,
 692  * if so, aborts.  If 'pidfile' is stale, deletes it. */
 693 static void
 694 check_already_running(void)
 695 {
 696     long int pid = read_pidfile__(pidfile, true);
 697     if (pid > 0) {
 698         VLOG_FATAL("%s: already running as pid %ld, aborting", pidfile, pid);
 699     } else if (pid < 0) {
 700         VLOG_FATAL("%s: pidfile check failed (%s), aborting",
 701                    pidfile, strerror(-pid));
 702     }
 703 }