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;             /* Was --detach specified? */
  43 static bool detached;           /* Have we already detached? */
  44
  45 /* --pidfile: Name of pidfile (null if none). */
  46 static char *pidfile;
  47
  48 /* Device and inode of pidfile, so we can avoid reopening it. */
  49 static dev_t pidfile_dev;
  50 static ino_t pidfile_ino;
  51
  52 /* --overwrite-pidfile: Create pidfile even if one already exists and is
  53    locked? */
  54 static bool overwrite_pidfile;
  55
  56 /* --no-chdir: Should we chdir to "/"? */
  57 static bool chdir_ = true;
  58
  59 /* File descriptor used by daemonize_start() and daemonize_complete(). */
  60 static int daemonize_fd = -1;
  61
  62 /* --monitor: Should a supervisory process monitor the daemon and restart it if
  63  * it dies due to an error signal? */
  64 static bool monitor;
  65
  66 /* For each of the standard file descriptors, whether to replace it by
  67  * /dev/null (if false) or keep it for the daemon to use (if true). */
  68 static bool save_fds[3];
  69
  70 static void check_already_running(void);
  71 static int lock_pidfile(FILE *, int command);
  72
  73 /* Returns the file name that would be used for a pidfile if 'name' were
  74  * provided to set_pidfile().  The caller must free the returned string. */
  75 char *
  76 make_pidfile_name(const char *name)
  77 {
  78     return (!name
  79             ? xasprintf("%s/%s.pid", ovs_rundir(), program_name)
  80             : abs_file_name(ovs_rundir(), name));
  81 }
  82
  83 /* Sets up a following call to daemonize() to create a pidfile named 'name'.
  84  * If 'name' begins with '/', then it is treated as an absolute path.
  85  * Otherwise, it is taken relative to RUNDIR, which is $(prefix)/var/run by
  86  * default.
  87  *
  88  * If 'name' is null, then program_name followed by ".pid" is used. */
  89 void
  90 set_pidfile(const char *name)
  91 {
  92     free(pidfile);
  93     pidfile = make_pidfile_name(name);
  94 }
  95
  96 /* Returns an absolute path to the configured pidfile, or a null pointer if no
  97  * pidfile is configured.  The caller must not modify or free the returned
  98  * string. */
  99 const char *
 100 get_pidfile(void)
 101 {
 102     return pidfile;
 103 }
 104
 105 /* Sets that we do not chdir to "/". */
 106 void
 107 set_no_chdir(void)
 108 {
 109     chdir_ = false;
 110 }
 111
 112 /* Will we chdir to "/" as part of daemonizing? */
 113 bool
 114 is_chdir_enabled(void)
 115 {
 116     return chdir_;
 117 }
 118
 119 /* Normally, daemonize() or damonize_start() will terminate the program with a
 120  * message if a locked pidfile already exists.  If this function is called, an
 121  * existing pidfile will be replaced, with a warning. */
 122 void
 123 ignore_existing_pidfile(void)
 124 {
 125     overwrite_pidfile = true;
 126 }
 127
 128 /* Sets up a following call to daemonize() to detach from the foreground
 129  * session, running this process in the background.  */
 130 void
 131 set_detach(void)
 132 {
 133     detach = true;
 134 }
 135
 136 /* Will daemonize() really detach? */
 137 bool
 138 get_detach(void)
 139 {
 140     return detach;
 141 }
 142
 143 /* Sets up a following call to daemonize() to fork a supervisory process to
 144  * monitor the daemon and restart it if it dies due to an error signal.  */
 145 void
 146 daemon_set_monitor(void)
 147 {
 148     monitor = true;
 149 }
 150
 151 /* A daemon doesn't normally have any use for the file descriptors for stdin,
 152  * stdout, and stderr after it detaches.  To keep these file descriptors from
 153  * e.g. holding an SSH session open, by default detaching replaces each of
 154  * these file descriptors by /dev/null.  But a few daemons expect the user to
 155  * redirect stdout or stderr to a file, in which case it is desirable to keep
 156  * these file descriptors.  This function, therefore, disables replacing 'fd'
 157  * by /dev/null when the daemon detaches. */
 158 void
 159 daemon_save_fd(int fd)
 160 {
 161     assert(fd == STDIN_FILENO || fd == STDOUT_FILENO || fd == STDERR_FILENO);
 162     save_fds[fd] = true;
 163 }
 164
 165 /* If a pidfile has been configured, creates it and stores the running
 166  * process's pid in it.  Ensures that the pidfile will be deleted when the
 167  * process exits. */
 168 static void
 169 make_pidfile(void)
 170 {
 171     long int pid = getpid();
 172     struct stat s;
 173     char *tmpfile;
 174     FILE *file;
 175     int error;
 176
 177     /* Create a temporary pidfile. */
 178     tmpfile = xasprintf("%s.tmp%ld", pidfile, pid);
 179     fatal_signal_add_file_to_unlink(tmpfile);
 180     file = fopen(tmpfile, "w+");
 181     if (!file) {
 182         VLOG_FATAL("%s: create failed (%s)", tmpfile, strerror(errno));
 183     }
 184
 185     if (fstat(fileno(file), &s) == -1) {
 186         VLOG_FATAL("%s: fstat failed (%s)", tmpfile, strerror(errno));
 187     }
 188
 189     fprintf(file, "%ld\n", pid);
 190     if (fflush(file) == EOF) {
 191         VLOG_FATAL("%s: write failed (%s)", tmpfile, strerror(errno));
 192     }
 193
 194     error = lock_pidfile(file, F_SETLK);
 195     if (error) {
 196         VLOG_FATAL("%s: fcntl(F_SETLK) failed (%s)", tmpfile, strerror(error));
 197     }
 198
 199     /* Rename or link it to the correct name. */
 200     if (overwrite_pidfile) {
 201         if (rename(tmpfile, pidfile) < 0) {
 202             VLOG_FATAL("failed to rename \"%s\" to \"%s\" (%s)",
 203                        tmpfile, pidfile, strerror(errno));
 204         }
 205     } else {
 206         do {
 207             error = link(tmpfile, pidfile) == -1 ? errno : 0;
 208             if (error == EEXIST) {
 209                 check_already_running();
 210             }
 211         } while (error == EINTR || error == EEXIST);
 212         if (error) {
 213             VLOG_FATAL("failed to link \"%s\" as \"%s\" (%s)",
 214                        tmpfile, pidfile, strerror(error));
 215         }
 216     }
 217
 218     /* Ensure that the pidfile will get deleted on exit. */
 219     fatal_signal_add_file_to_unlink(pidfile);
 220
 221     /* Delete the temporary pidfile if it still exists. */
 222     if (!overwrite_pidfile) {
 223         error = fatal_signal_unlink_file_now(tmpfile);
 224         if (error) {
 225             VLOG_FATAL("%s: unlink failed (%s)", tmpfile, strerror(error));
 226         }
 227     }
 228
 229     /* Clean up.
 230      *
 231      * We don't close 'file' because its file descriptor must remain open to
 232      * hold the lock. */
 233     pidfile_dev = s.st_dev;
 234     pidfile_ino = s.st_ino;
 235     free(tmpfile);
 236     free(pidfile);
 237     pidfile = NULL;
 238 }
 239
 240 /* If configured with set_pidfile() or set_detach(), creates the pid file and
 241  * detaches from the foreground session.  */
 242 void
 243 daemonize(void)
 244 {
 245     daemonize_start();
 246     daemonize_complete();
 247 }
 248
 249 /* Calls fork() and on success returns its return value.  On failure, logs an
 250  * error and exits unsuccessfully.
 251  *
 252  * Post-fork, but before returning, this function calls a few other functions
 253  * that are generally useful if the child isn't planning to exec a new
 254  * process. */
 255 pid_t
 256 fork_and_clean_up(void)
 257 {
 258     pid_t pid;
 259
 260     pid = fork();
 261     if (pid > 0) {
 262         /* Running in parent process. */
 263         fatal_signal_fork();
 264     } else if (!pid) {
 265         /* Running in child process. */
 266         time_postfork();
 267         lockfile_postfork();
 268     } else {
 269         VLOG_FATAL("fork failed (%s)", strerror(errno));
 270     }
 271
 272     return pid;
 273 }
 274
 275 /* Forks, then:
 276  *
 277  *   - In the parent, waits for the child to signal that it has completed its
 278  *     startup sequence.  Then stores -1 in '*fdp' and returns the child's pid.
 279  *
 280  *   - In the child, stores a fd in '*fdp' and returns 0.  The caller should
 281  *     pass the fd to fork_notify_startup() after it finishes its startup
 282  *     sequence.
 283  *
 284  * If something goes wrong with the fork, logs a critical error and aborts the
 285  * process. */
 286 static pid_t
 287 fork_and_wait_for_startup(int *fdp)
 288 {
 289     int fds[2];
 290     pid_t pid;
 291
 292     xpipe(fds);
 293
 294     pid = fork_and_clean_up();
 295     if (pid > 0) {
 296         /* Running in parent process. */
 297         size_t bytes_read;
 298         char c;
 299
 300         close(fds[1]);
 301         if (read_fully(fds[0], &c, 1, &bytes_read) != 0) {
 302             int retval;
 303             int status;
 304
 305             do {
 306                 retval = waitpid(pid, &status, 0);
 307             } while (retval == -1 && errno == EINTR);
 308
 309             if (retval == pid) {
 310                 if (WIFEXITED(status) && WEXITSTATUS(status)) {
 311                     /* Child exited with an error.  Convey the same error
 312                      * to our parent process as a courtesy. */
 313                     exit(WEXITSTATUS(status));
 314                 } else {
 315                     char *status_msg = process_status_msg(status);
 316                     VLOG_FATAL("fork child died before signaling startup (%s)",
 317                                status_msg);
 318                 }
 319             } else if (retval < 0) {
 320                 VLOG_FATAL("waitpid failed (%s)", strerror(errno));
 321             } else {
 322                 NOT_REACHED();
 323             }
 324         }
 325         close(fds[0]);
 326         *fdp = -1;
 327     } else if (!pid) {
 328         /* Running in child process. */
 329         close(fds[0]);
 330         *fdp = fds[1];
 331     }
 332
 333     return pid;
 334 }
 335
 336 static void
 337 fork_notify_startup(int fd)
 338 {
 339     if (fd != -1) {
 340         size_t bytes_written;
 341         int error;
 342
 343         error = write_fully(fd, "", 1, &bytes_written);
 344         if (error) {
 345             VLOG_FATAL("pipe write failed (%s)", strerror(error));
 346         }
 347
 348         close(fd);
 349     }
 350 }
 351
 352 static bool
 353 should_restart(int status)
 354 {
 355     if (WIFSIGNALED(status)) {
 356         static const int error_signals[] = {
 357             SIGABRT, SIGALRM, SIGBUS, SIGFPE, SIGILL, SIGPIPE, SIGSEGV,
 358             SIGXCPU, SIGXFSZ
 359         };
 360
 361         size_t i;
 362
 363         for (i = 0; i < ARRAY_SIZE(error_signals); i++) {
 364             if (error_signals[i] == WTERMSIG(status)) {
 365                 return true;
 366             }
 367         }
 368     }
 369     return false;
 370 }
 371
 372 static void
 373 monitor_daemon(pid_t daemon_pid)
 374 {
 375     /* XXX Should log daemon's stderr output at startup time. */
 376     time_t last_restart;
 377     char *status_msg;
 378     int crashes;
 379
 380     subprogram_name = "monitor";
 381     status_msg = xstrdup("healthy");
 382     last_restart = TIME_MIN;
 383     crashes = 0;
 384     for (;;) {
 385         int retval;
 386         int status;
 387
 388         proctitle_set("%s: monitoring pid %lu (%s)",
 389                       program_name, (unsigned long int) daemon_pid,
 390                       status_msg);
 391
 392         do {
 393             retval = waitpid(daemon_pid, &status, 0);
 394         } while (retval == -1 && errno == EINTR);
 395
 396         if (retval == -1) {
 397             VLOG_FATAL("waitpid failed (%s)", strerror(errno));
 398         } else if (retval == daemon_pid) {
 399             char *s = process_status_msg(status);
 400             if (should_restart(status)) {
 401                 free(status_msg);
 402                 status_msg = xasprintf("%d crashes: pid %lu died, %s",
 403                                        ++crashes,
 404                                        (unsigned long int) daemon_pid, s);
 405                 free(s);
 406
 407                 if (WCOREDUMP(status)) {
 408                     /* Disable further core dumps to save disk space. */
 409                     struct rlimit r;
 410
 411                     r.rlim_cur = 0;
 412                     r.rlim_max = 0;
 413                     if (setrlimit(RLIMIT_CORE, &r) == -1) {
 414                         VLOG_WARN("failed to disable core dumps: %s",
 415                                   strerror(errno));
 416                     }
 417                 }
 418
 419                 /* Throttle restarts to no more than once every 10 seconds. */
 420                 if (time(NULL) < last_restart + 10) {
 421                     VLOG_WARN("%s, waiting until 10 seconds since last "
 422                               "restart", status_msg);
 423                     for (;;) {
 424                         time_t now = time(NULL);
 425                         time_t wakeup = last_restart + 10;
 426                         if (now >= wakeup) {
 427                             break;
 428                         }
 429                         sleep(wakeup - now);
 430                     }
 431                 }
 432                 last_restart = time(NULL);
 433
 434                 VLOG_ERR("%s, restarting", status_msg);
 435                 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
 436                 if (!daemon_pid) {
 437                     break;
 438                 }
 439             } else {
 440                 VLOG_INFO("pid %lu died, %s, exiting",
 441                           (unsigned long int) daemon_pid, s);
 442                 free(s);
 443                 exit(0);
 444             }
 445         }
 446     }
 447     free(status_msg);
 448
 449     /* Running in new daemon process. */
 450     proctitle_restore();
 451     subprogram_name = "";
 452 }
 453
 454 /* Close standard file descriptors (except any that the client has requested we
 455  * leave open by calling daemon_save_fd()).  If we're started from e.g. an SSH
 456  * session, then this keeps us from holding that session open artificially. */
 457 static void
 458 close_standard_fds(void)
 459 {
 460     int null_fd = get_null_fd();
 461     if (null_fd >= 0) {
 462         int fd;
 463
 464         for (fd = 0; fd < 3; fd++) {
 465             if (!save_fds[fd]) {
 466                 dup2(null_fd, fd);
 467             }
 468         }
 469     }
 470
 471     /* Disable logging to stderr to avoid wasting CPU time. */
 472     vlog_set_levels(NULL, VLF_CONSOLE, VLL_OFF);
 473 }
 474
 475 /* If daemonization is configured, then starts daemonization, by forking and
 476  * returning in the child process.  The parent process hangs around until the
 477  * child lets it know either that it completed startup successfully (by calling
 478  * daemon_complete()) or that it failed to start up (by exiting with a nonzero
 479  * exit code). */
 480 void
 481 daemonize_start(void)
 482 {
 483     daemonize_fd = -1;
 484
 485     if (detach) {
 486         if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
 487             /* Running in parent process. */
 488             exit(0);
 489         }
 490         /* Running in daemon or monitor process. */
 491     }
 492
 493     if (monitor) {
 494         int saved_daemonize_fd = daemonize_fd;
 495         pid_t daemon_pid;
 496
 497         daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
 498         if (daemon_pid > 0) {
 499             /* Running in monitor process. */
 500             fork_notify_startup(saved_daemonize_fd);
 501             close_standard_fds();
 502             monitor_daemon(daemon_pid);
 503         }
 504         /* Running in daemon process. */
 505     }
 506
 507     if (pidfile) {
 508         make_pidfile();
 509     }
 510
 511     /* Make sure that the unixctl commands for vlog get registered in a
 512      * daemon, even before the first log message. */
 513     vlog_init();
 514 }
 515
 516 /* If daemonization is configured, then this function notifies the parent
 517  * process that the child process has completed startup successfully.  It also
 518  * call daemonize_post_detach().
 519  *
 520  * Calling this function more than once has no additional effect. */
 521 void
 522 daemonize_complete(void)
 523 {
 524     if (!detached) {
 525         detached = true;
 526
 527         fork_notify_startup(daemonize_fd);
 528         daemonize_fd = -1;
 529         daemonize_post_detach();
 530     }
 531 }
 532
 533 /* If daemonization is configured, then this function does traditional Unix
 534  * daemonization behavior: join a new session, chdir to the root (if not
 535  * disabled), and close the standard file descriptors.
 536  *
 537  * It only makes sense to call this function as part of an implementation of a
 538  * special daemon subprocess.  A normal daemon should just call
 539  * daemonize_complete(). */
 540 void
 541 daemonize_post_detach(void)
 542 {
 543     if (detach) {
 544         setsid();
 545         if (chdir_) {
 546             ignore(chdir("/"));
 547         }
 548         close_standard_fds();
 549     }
 550 }
 551
 552 void
 553 daemon_usage(void)
 554 {
 555     printf(
 556         "\nDaemon options:\n"
 557         "  --detach                run in background as daemon\n"
 558         "  --no-chdir              do not chdir to '/'\n"
 559         "  --pidfile[=FILE]        create pidfile (default: %s/%s.pid)\n"
 560         "  --overwrite-pidfile     with --pidfile, start even if already "
 561                                    "running\n",
 562         ovs_rundir(), program_name);
 563 }
 564
 565 static int
 566 lock_pidfile__(FILE *file, int command, struct flock *lck)
 567 {
 568     int error;
 569
 570     lck->l_type = F_WRLCK;
 571     lck->l_whence = SEEK_SET;
 572     lck->l_start = 0;
 573     lck->l_len = 0;
 574     lck->l_pid = 0;
 575
 576     do {
 577         error = fcntl(fileno(file), command, lck) == -1 ? errno : 0;
 578     } while (error == EINTR);
 579     return error;
 580 }
 581
 582 static int
 583 lock_pidfile(FILE *file, int command)
 584 {
 585     struct flock lck;
 586
 587     return lock_pidfile__(file, command, &lck);
 588 }
 589
 590 static pid_t
 591 read_pidfile__(const char *pidfile, bool delete_if_stale)
 592 {
 593     struct stat s, s2;
 594     struct flock lck;
 595     char line[128];
 596     FILE *file;
 597     int error;
 598
 599     if ((pidfile_ino || pidfile_dev)
 600         && !stat(pidfile, &s)
 601         && s.st_ino == pidfile_ino && s.st_dev == pidfile_dev) {
 602         /* It's our own pidfile.  We can't afford to open it, because closing
 603          * *any* fd for a file that a process has locked also releases all the
 604          * locks on that file.
 605          *
 606          * Fortunately, we know the associated pid anyhow: */
 607         return getpid();
 608     }
 609
 610     file = fopen(pidfile, "r+");
 611     if (!file) {
 612         if (errno == ENOENT && delete_if_stale) {
 613             return 0;
 614         }
 615         error = errno;
 616         VLOG_WARN("%s: open: %s", pidfile, strerror(error));
 617         goto error;
 618     }
 619
 620     error = lock_pidfile__(file, F_GETLK, &lck);
 621     if (error) {
 622         VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
 623         goto error;
 624     }
 625     if (lck.l_type == F_UNLCK) {
 626         /* pidfile exists but it isn't locked by anyone.  We need to delete it
 627          * so that a new pidfile can go in its place.  But just calling
 628          * unlink(pidfile) makes a nasty race: what if someone else unlinks it
 629          * before we do and then replaces it by a valid pidfile?  We'd unlink
 630          * their valid pidfile.  We do a little dance to avoid the race, by
 631          * locking the invalid pidfile.  Only one process can have the invalid
 632          * pidfile locked, and only that process has the right to unlink it. */
 633         if (!delete_if_stale) {
 634             error = ESRCH;
 635             VLOG_DBG("%s: pid file is stale", pidfile);
 636             goto error;
 637         }
 638
 639         /* Get the lock. */
 640         error = lock_pidfile(file, F_SETLK);
 641         if (error) {
 642             /* We lost a race with someone else doing the same thing. */
 643             VLOG_WARN("%s: lost race to lock pidfile", pidfile);
 644             goto error;
 645         }
 646
 647         /* Is the file we have locked still named 'pidfile'? */
 648         if (stat(pidfile, &s) || fstat(fileno(file), &s2)
 649             || s.st_ino != s2.st_ino || s.st_dev != s2.st_dev) {
 650             /* No.  We lost a race with someone else who got the lock before
 651              * us, deleted the pidfile, and closed it (releasing the lock). */
 652             error = EALREADY;
 653             VLOG_WARN("%s: lost race to delete pidfile", pidfile);
 654             goto error;
 655         }
 656
 657         /* We won the right to delete the stale pidfile. */
 658         if (unlink(pidfile)) {
 659             error = errno;
 660             VLOG_WARN("%s: failed to delete stale pidfile (%s)",
 661                       pidfile, strerror(error));
 662             goto error;
 663         }
 664         VLOG_DBG("%s: deleted stale pidfile", pidfile);
 665         fclose(file);
 666         return 0;
 667     }
 668
 669     if (!fgets(line, sizeof line, file)) {
 670         if (ferror(file)) {
 671             error = errno;
 672             VLOG_WARN("%s: read: %s", pidfile, strerror(error));
 673         } else {
 674             error = ESRCH;
 675             VLOG_WARN("%s: read: unexpected end of file", pidfile);
 676         }
 677         goto error;
 678     }
 679
 680     if (lck.l_pid != strtoul(line, NULL, 10)) {
 681         /* The process that has the pidfile locked is not the process that
 682          * created it.  It must be stale, with the process that has it locked
 683          * preparing to delete it. */
 684         error = ESRCH;
 685         VLOG_WARN("%s: stale pidfile for pid %s being deleted by pid %ld",
 686                   pidfile, line, (long int) lck.l_pid);
 687         goto error;
 688     }
 689
 690     fclose(file);
 691     return lck.l_pid;
 692
 693 error:
 694     if (file) {
 695         fclose(file);
 696     }
 697     return -error;
 698 }
 699
 700 /* Opens and reads a PID from 'pidfile'.  Returns the positive PID if
 701  * successful, otherwise a negative errno value. */
 702 pid_t
 703 read_pidfile(const char *pidfile)
 704 {
 705     return read_pidfile__(pidfile, false);
 706 }
 707
 708 /* Checks whether a process with the given 'pidfile' is already running and,
 709  * if so, aborts.  If 'pidfile' is stale, deletes it. */
 710 static void
 711 check_already_running(void)
 712 {
 713     long int pid = read_pidfile__(pidfile, true);
 714     if (pid > 0) {
 715         VLOG_FATAL("%s: already running as pid %ld, aborting", pidfile, pid);
 716     } else if (pid < 0) {
 717         VLOG_FATAL("%s: pidfile check failed (%s), aborting",
 718                    pidfile, strerror(-pid));
 719     }
 720 }