lib/daemon.c

   1 /*
   2  * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks.
   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 <errno.h>
  20 #include <fcntl.h>
  21 #include <signal.h>
  22 #include <stdlib.h>
  23 #include <string.h>
  24 #include <sys/resource.h>
  25 #include <sys/wait.h>
  26 #include <sys/stat.h>
  27 #include <unistd.h>
  28 #include "command-line.h"
  29 #include "fatal-signal.h"
  30 #include "dirs.h"
  31 #include "lockfile.h"
  32 #include "process.h"
  33 #include "socket-util.h"
  34 #include "timeval.h"
  35 #include "util.h"
  36 #include "vlog.h"
  37
  38 VLOG_DEFINE_THIS_MODULE(daemon);
  39
  40 /* --detach: Should we run in the background? */
  41 static bool detach;
  42
  43 /* --pidfile: Name of pidfile (null if none). */
  44 static char *pidfile;
  45
  46 /* Device and inode of pidfile, so we can avoid reopening it. */
  47 static dev_t pidfile_dev;
  48 static ino_t pidfile_ino;
  49
  50 /* --overwrite-pidfile: Create pidfile even if one already exists and is
  51    locked? */
  52 static bool overwrite_pidfile;
  53
  54 /* --no-chdir: Should we chdir to "/"? */
  55 static bool chdir_ = true;
  56
  57 /* File descriptor used by daemonize_start() and daemonize_complete(). */
  58 static int daemonize_fd = -1;
  59
  60 /* --monitor: Should a supervisory process monitor the daemon and restart it if
  61  * it dies due to an error signal? */
  62 static bool monitor;
  63
  64 /* Returns the file name that would be used for a pidfile if 'name' were
  65  * provided to set_pidfile().  The caller must free the returned string. */
  66 char *
  67 make_pidfile_name(const char *name)
  68 {
  69     return (!name
  70             ? xasprintf("%s/%s.pid", ovs_rundir(), program_name)
  71             : abs_file_name(ovs_rundir(), name));
  72 }
  73
  74 /* Sets up a following call to daemonize() to create a pidfile named 'name'.
  75  * If 'name' begins with '/', then it is treated as an absolute path.
  76  * Otherwise, it is taken relative to RUNDIR, which is $(prefix)/var/run by
  77  * default.
  78  *
  79  * If 'name' is null, then program_name followed by ".pid" is used. */
  80 void
  81 set_pidfile(const char *name)
  82 {
  83     free(pidfile);
  84     pidfile = make_pidfile_name(name);
  85 }
  86
  87 /* Returns an absolute path to the configured pidfile, or a null pointer if no
  88  * pidfile is configured.  The caller must not modify or free the returned
  89  * string. */
  90 const char *
  91 get_pidfile(void)
  92 {
  93     return pidfile;
  94 }
  95
  96 /* Sets that we do not chdir to "/". */
  97 void
  98 set_no_chdir(void)
  99 {
 100     chdir_ = false;
 101 }
 102
 103 /* Will we chdir to "/" as part of daemonizing? */
 104 bool
 105 is_chdir_enabled(void)
 106 {
 107     return chdir_;
 108 }
 109
 110 /* Normally, die_if_already_running() will terminate the program with a message
 111  * if a locked pidfile already exists.  If this function is called,
 112  * die_if_already_running() will merely log a warning. */
 113 void
 114 ignore_existing_pidfile(void)
 115 {
 116     overwrite_pidfile = true;
 117 }
 118
 119 /* Sets up a following call to daemonize() to detach from the foreground
 120  * session, running this process in the background.  */
 121 void
 122 set_detach(void)
 123 {
 124     detach = true;
 125 }
 126
 127 /* Will daemonize() really detach? */
 128 bool
 129 get_detach(void)
 130 {
 131     return detach;
 132 }
 133
 134 /* Sets up a following call to daemonize() to fork a supervisory process to
 135  * monitor the daemon and restart it if it dies due to an error signal.  */
 136 void
 137 daemon_set_monitor(void)
 138 {
 139     monitor = true;
 140 }
 141
 142 /* If a locked pidfile exists, issue a warning message and, unless
 143  * ignore_existing_pidfile() has been called, terminate the program. */
 144 void
 145 die_if_already_running(void)
 146 {
 147     pid_t pid;
 148     if (!pidfile) {
 149         return;
 150     }
 151     pid = read_pidfile_if_exists(pidfile);
 152     if (pid > 0) {
 153         if (!overwrite_pidfile) {
 154             VLOG_ERR("%s: %s already running as pid %ld, aborting",
 155                       get_pidfile(), program_name, (long int) pid);
 156             ovs_fatal(0, "%s: already running as pid %ld",
 157                       get_pidfile(), (long int) pid);
 158         } else {
 159             VLOG_WARN("%s: %s already running as pid %ld",
 160                       get_pidfile(), program_name, (long int) pid);
 161         }
 162     }
 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     if (pidfile) {
 172         /* Create pidfile via temporary file, so that observers never see an
 173          * empty pidfile or an unlocked pidfile. */
 174         long int pid = getpid();
 175         char *tmpfile;
 176         int fd;
 177
 178         tmpfile = xasprintf("%s.tmp%ld", pidfile, pid);
 179         fatal_signal_add_file_to_unlink(tmpfile);
 180         fd = open(tmpfile, O_CREAT | O_WRONLY | O_TRUNC, 0666);
 181         if (fd >= 0) {
 182             struct flock lck;
 183             lck.l_type = F_WRLCK;
 184             lck.l_whence = SEEK_SET;
 185             lck.l_start = 0;
 186             lck.l_len = 0;
 187             if (fcntl(fd, F_SETLK, &lck) != -1) {
 188                 char *text = xasprintf("%ld\n", pid);
 189                 if (write(fd, text, strlen(text)) == strlen(text)) {
 190                     fatal_signal_add_file_to_unlink(pidfile);
 191                     if (rename(tmpfile, pidfile) < 0) {
 192                         VLOG_ERR("failed to rename \"%s\" to \"%s\": %s",
 193                                  tmpfile, pidfile, strerror(errno));
 194                         fatal_signal_remove_file_to_unlink(pidfile);
 195                         close(fd);
 196                     } else {
 197                         /* Keep 'fd' open to retain the lock. */
 198                         struct stat s;
 199
 200                         if (!fstat(fd, &s)) {
 201                             pidfile_dev = s.st_dev;
 202                             pidfile_ino = s.st_ino;
 203                         } else {
 204                             VLOG_ERR("%s: fstat failed: %s",
 205                                      pidfile, strerror(errno));
 206                         }
 207                     }
 208                 } else {
 209                     VLOG_ERR("%s: write failed: %s", tmpfile, strerror(errno));
 210                     close(fd);
 211                 }
 212                 free(text);
 213             } else {
 214                 VLOG_ERR("%s: fcntl failed: %s", tmpfile, strerror(errno));
 215                 close(fd);
 216             }
 217         } else {
 218             VLOG_ERR("%s: create failed: %s", tmpfile, strerror(errno));
 219         }
 220         fatal_signal_remove_file_to_unlink(tmpfile);
 221         free(tmpfile);
 222     }
 223     free(pidfile);
 224     pidfile = NULL;
 225 }
 226
 227 /* If configured with set_pidfile() or set_detach(), creates the pid file and
 228  * detaches from the foreground session.  */
 229 void
 230 daemonize(void)
 231 {
 232     daemonize_start();
 233     daemonize_complete();
 234 }
 235
 236 static pid_t
 237 fork_and_wait_for_startup(int *fdp)
 238 {
 239     int fds[2];
 240     pid_t pid;
 241
 242     xpipe(fds);
 243
 244     pid = fork();
 245     if (pid > 0) {
 246         /* Running in parent process. */
 247         char c;
 248
 249         close(fds[1]);
 250         fatal_signal_fork();
 251         if (read(fds[0], &c, 1) != 1) {
 252             int retval;
 253             int status;
 254
 255             do {
 256                 retval = waitpid(pid, &status, 0);
 257             } while (retval == -1 && errno == EINTR);
 258
 259             if (retval == pid
 260                 && WIFEXITED(status)
 261                 && WEXITSTATUS(status)) {
 262                 /* Child exited with an error.  Convey the same error to
 263                  * our parent process as a courtesy. */
 264                 exit(WEXITSTATUS(status));
 265             }
 266
 267             VLOG_FATAL("fork child failed to signal startup (%s)",
 268                        strerror(errno));
 269         }
 270         close(fds[0]);
 271         *fdp = -1;
 272     } else if (!pid) {
 273         /* Running in child process. */
 274         close(fds[0]);
 275         time_postfork();
 276         lockfile_postfork();
 277         *fdp = fds[1];
 278     } else {
 279         VLOG_FATAL("fork failed (%s)", strerror(errno));
 280     }
 281
 282     return pid;
 283 }
 284
 285 static void
 286 fork_notify_startup(int fd)
 287 {
 288     if (fd != -1) {
 289         size_t bytes_written;
 290         int error;
 291
 292         error = write_fully(fd, "", 1, &bytes_written);
 293         if (error) {
 294             VLOG_FATAL("pipe write failed (%s)", strerror(error));
 295         }
 296
 297         close(fd);
 298     }
 299 }
 300
 301 static bool
 302 should_restart(int status)
 303 {
 304     if (WIFSIGNALED(status)) {
 305         static const int error_signals[] = {
 306             SIGABRT, SIGALRM, SIGBUS, SIGFPE, SIGILL, SIGPIPE, SIGSEGV,
 307             SIGXCPU, SIGXFSZ
 308         };
 309
 310         size_t i;
 311
 312         for (i = 0; i < ARRAY_SIZE(error_signals); i++) {
 313             if (error_signals[i] == WTERMSIG(status)) {
 314                 return true;
 315             }
 316         }
 317     }
 318     return false;
 319 }
 320
 321 static void
 322 monitor_daemon(pid_t daemon_pid)
 323 {
 324     /* XXX Should log daemon's stderr output at startup time. */
 325     const char *saved_program_name;
 326     time_t last_restart;
 327     char *status_msg;
 328     int crashes;
 329
 330     saved_program_name = program_name;
 331     program_name = xasprintf("monitor(%s)", program_name);
 332     status_msg = xstrdup("healthy");
 333     last_restart = TIME_MIN;
 334     crashes = 0;
 335     for (;;) {
 336         int retval;
 337         int status;
 338
 339         proctitle_set("%s: monitoring pid %lu (%s)",
 340                       saved_program_name, (unsigned long int) daemon_pid,
 341                       status_msg);
 342
 343         do {
 344             retval = waitpid(daemon_pid, &status, 0);
 345         } while (retval == -1 && errno == EINTR);
 346
 347         if (retval == -1) {
 348             VLOG_FATAL("waitpid failed (%s)", strerror(errno));
 349         } else if (retval == daemon_pid) {
 350             char *s = process_status_msg(status);
 351             if (should_restart(status)) {
 352                 free(status_msg);
 353                 status_msg = xasprintf("%d crashes: pid %lu died, %s",
 354                                        ++crashes,
 355                                        (unsigned long int) daemon_pid, s);
 356                 free(s);
 357
 358                 if (WCOREDUMP(status)) {
 359                     /* Disable further core dumps to save disk space. */
 360                     struct rlimit r;
 361
 362                     r.rlim_cur = 0;
 363                     r.rlim_max = 0;
 364                     if (setrlimit(RLIMIT_CORE, &r) == -1) {
 365                         VLOG_WARN("failed to disable core dumps: %s",
 366                                   strerror(errno));
 367                     }
 368                 }
 369
 370                 /* Throttle restarts to no more than once every 10 seconds. */
 371                 if (time(NULL) < last_restart + 10) {
 372                     VLOG_WARN("%s, waiting until 10 seconds since last "
 373                               "restart", status_msg);
 374                     for (;;) {
 375                         time_t now = time(NULL);
 376                         time_t wakeup = last_restart + 10;
 377                         if (now >= wakeup) {
 378                             break;
 379                         }
 380                         sleep(wakeup - now);
 381                     }
 382                 }
 383                 last_restart = time(NULL);
 384
 385                 VLOG_ERR("%s, restarting", status_msg);
 386                 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
 387                 if (!daemon_pid) {
 388                     break;
 389                 }
 390             } else {
 391                 VLOG_INFO("pid %lu died, %s, exiting",
 392                           (unsigned long int) daemon_pid, s);
 393                 free(s);
 394                 exit(0);
 395             }
 396         }
 397     }
 398     free(status_msg);
 399
 400     /* Running in new daemon process. */
 401     proctitle_restore();
 402     free((char *) program_name);
 403     program_name = saved_program_name;
 404 }
 405
 406 /* Close stdin, stdout, stderr.  If we're started from e.g. an SSH session,
 407  * then this keeps us from holding that session open artificially. */
 408 static void
 409 close_standard_fds(void)
 410 {
 411     int null_fd = get_null_fd();
 412     if (null_fd >= 0) {
 413         dup2(null_fd, STDIN_FILENO);
 414         dup2(null_fd, STDOUT_FILENO);
 415         dup2(null_fd, STDERR_FILENO);
 416     }
 417 }
 418
 419 /* If daemonization is configured, then starts daemonization, by forking and
 420  * returning in the child process.  The parent process hangs around until the
 421  * child lets it know either that it completed startup successfully (by calling
 422  * daemon_complete()) or that it failed to start up (by exiting with a nonzero
 423  * exit code). */
 424 void
 425 daemonize_start(void)
 426 {
 427     daemonize_fd = -1;
 428
 429     if (detach) {
 430         if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
 431             /* Running in parent process. */
 432             exit(0);
 433         }
 434         /* Running in daemon or monitor process. */
 435     }
 436
 437     if (monitor) {
 438         int saved_daemonize_fd = daemonize_fd;
 439         pid_t daemon_pid;
 440
 441         daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
 442         if (daemon_pid > 0) {
 443             /* Running in monitor process. */
 444             fork_notify_startup(saved_daemonize_fd);
 445             close_standard_fds();
 446             monitor_daemon(daemon_pid);
 447         }
 448         /* Running in daemon process. */
 449     }
 450
 451     make_pidfile();
 452
 453     /* Make sure that the unixctl commands for vlog get registered in a
 454      * daemon, even before the first log message. */
 455     vlog_init();
 456 }
 457
 458 /* If daemonization is configured, then this function notifies the parent
 459  * process that the child process has completed startup successfully.
 460  *
 461  * Calling this function more than once has no additional effect. */
 462 void
 463 daemonize_complete(void)
 464 {
 465     fork_notify_startup(daemonize_fd);
 466     daemonize_fd = -1;
 467
 468     if (detach) {
 469         setsid();
 470         if (chdir_) {
 471             ignore(chdir("/"));
 472         }
 473         close_standard_fds();
 474         detach = false;
 475     }
 476 }
 477
 478 void
 479 daemon_usage(void)
 480 {
 481     printf(
 482         "\nDaemon options:\n"
 483         "  --detach                run in background as daemon\n"
 484         "  --no-chdir              do not chdir to '/'\n"
 485         "  --pidfile[=FILE]        create pidfile (default: %s/%s.pid)\n"
 486         "  --overwrite-pidfile     with --pidfile, start even if already "
 487                                    "running\n",
 488         ovs_rundir(), program_name);
 489 }
 490
 491 static pid_t
 492 read_pidfile__(const char *pidfile, bool must_exist)
 493 {
 494     char line[128];
 495     struct flock lck;
 496     struct stat s;
 497     FILE *file;
 498     int error;
 499
 500     if ((pidfile_ino || pidfile_dev)
 501         && !stat(pidfile, &s)
 502         && s.st_ino == pidfile_ino && s.st_dev == pidfile_dev) {
 503         /* It's our own pidfile.  We can't afford to open it, because closing
 504          * *any* fd for a file that a process has locked also releases all the
 505          * locks on that file.
 506          *
 507          * Fortunately, we know the associated pid anyhow: */
 508         return getpid();
 509     }
 510
 511     file = fopen(pidfile, "r");
 512     if (!file) {
 513         if (errno == ENOENT && !must_exist) {
 514             return 0;
 515         }
 516         error = errno;
 517         VLOG_WARN("%s: open: %s", pidfile, strerror(error));
 518         goto error;
 519     }
 520
 521     lck.l_type = F_WRLCK;
 522     lck.l_whence = SEEK_SET;
 523     lck.l_start = 0;
 524     lck.l_len = 0;
 525     lck.l_pid = 0;
 526     if (fcntl(fileno(file), F_GETLK, &lck)) {
 527         error = errno;
 528         VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
 529         goto error;
 530     }
 531     if (lck.l_type == F_UNLCK) {
 532         error = ESRCH;
 533         VLOG_WARN("%s: pid file is not locked", pidfile);
 534         goto error;
 535     }
 536
 537     if (!fgets(line, sizeof line, file)) {
 538         if (ferror(file)) {
 539             error = errno;
 540             VLOG_WARN("%s: read: %s", pidfile, strerror(error));
 541         } else {
 542             error = ESRCH;
 543             VLOG_WARN("%s: read: unexpected end of file", pidfile);
 544         }
 545         goto error;
 546     }
 547
 548     if (lck.l_pid != strtoul(line, NULL, 10)) {
 549         error = ESRCH;
 550         VLOG_WARN("l_pid (%ld) != %s pid (%s)",
 551                    (long int) lck.l_pid, pidfile, line);
 552         goto error;
 553     }
 554
 555     fclose(file);
 556     return lck.l_pid;
 557
 558 error:
 559     if (file) {
 560         fclose(file);
 561     }
 562     return -error;
 563 }
 564
 565 /* Opens and reads a PID from 'pidfile'.  Returns the positive PID if
 566  * successful, otherwise a negative errno value. */
 567 pid_t
 568 read_pidfile(const char *pidfile)
 569 {
 570     return read_pidfile__(pidfile, true);
 571 }
 572
 573
 574 /* Opens and reads a PID from 'pidfile', if it exists.  Returns 0 if 'pidfile'
 575  * doesn't exist, the positive PID if successful, otherwise a negative errno
 576  * value. */
 577 pid_t
 578 read_pidfile_if_exists(const char *pidfile)
 579 {
 580     return read_pidfile__(pidfile, false);
 581 }