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         size_t bytes_read;
 248         char c;
 249
 250         close(fds[1]);
 251         fatal_signal_fork();
 252         if (read_fully(fds[0], &c, 1, &bytes_read) != 0) {
 253             int retval;
 254             int status;
 255
 256             do {
 257                 retval = waitpid(pid, &status, 0);
 258             } while (retval == -1 && errno == EINTR);
 259
 260             if (retval == pid
 261                 && WIFEXITED(status)
 262                 && WEXITSTATUS(status)) {
 263                 /* Child exited with an error.  Convey the same error to
 264                  * our parent process as a courtesy. */
 265                 exit(WEXITSTATUS(status));
 266             }
 267
 268             VLOG_FATAL("fork child failed to signal startup (%s)",
 269                        strerror(errno));
 270         }
 271         close(fds[0]);
 272         *fdp = -1;
 273     } else if (!pid) {
 274         /* Running in child process. */
 275         close(fds[0]);
 276         time_postfork();
 277         lockfile_postfork();
 278         *fdp = fds[1];
 279     } else {
 280         VLOG_FATAL("fork failed (%s)", strerror(errno));
 281     }
 282
 283     return pid;
 284 }
 285
 286 static void
 287 fork_notify_startup(int fd)
 288 {
 289     if (fd != -1) {
 290         size_t bytes_written;
 291         int error;
 292
 293         error = write_fully(fd, "", 1, &bytes_written);
 294         if (error) {
 295             VLOG_FATAL("pipe write failed (%s)", strerror(error));
 296         }
 297
 298         close(fd);
 299     }
 300 }
 301
 302 static bool
 303 should_restart(int status)
 304 {
 305     if (WIFSIGNALED(status)) {
 306         static const int error_signals[] = {
 307             SIGABRT, SIGALRM, SIGBUS, SIGFPE, SIGILL, SIGPIPE, SIGSEGV,
 308             SIGXCPU, SIGXFSZ
 309         };
 310
 311         size_t i;
 312
 313         for (i = 0; i < ARRAY_SIZE(error_signals); i++) {
 314             if (error_signals[i] == WTERMSIG(status)) {
 315                 return true;
 316             }
 317         }
 318     }
 319     return false;
 320 }
 321
 322 static void
 323 monitor_daemon(pid_t daemon_pid)
 324 {
 325     /* XXX Should log daemon's stderr output at startup time. */
 326     const char *saved_program_name;
 327     time_t last_restart;
 328     char *status_msg;
 329     int crashes;
 330
 331     saved_program_name = program_name;
 332     program_name = xasprintf("monitor(%s)", program_name);
 333     status_msg = xstrdup("healthy");
 334     last_restart = TIME_MIN;
 335     crashes = 0;
 336     for (;;) {
 337         int retval;
 338         int status;
 339
 340         proctitle_set("%s: monitoring pid %lu (%s)",
 341                       saved_program_name, (unsigned long int) daemon_pid,
 342                       status_msg);
 343
 344         do {
 345             retval = waitpid(daemon_pid, &status, 0);
 346         } while (retval == -1 && errno == EINTR);
 347
 348         if (retval == -1) {
 349             VLOG_FATAL("waitpid failed (%s)", strerror(errno));
 350         } else if (retval == daemon_pid) {
 351             char *s = process_status_msg(status);
 352             if (should_restart(status)) {
 353                 free(status_msg);
 354                 status_msg = xasprintf("%d crashes: pid %lu died, %s",
 355                                        ++crashes,
 356                                        (unsigned long int) daemon_pid, s);
 357                 free(s);
 358
 359                 if (WCOREDUMP(status)) {
 360                     /* Disable further core dumps to save disk space. */
 361                     struct rlimit r;
 362
 363                     r.rlim_cur = 0;
 364                     r.rlim_max = 0;
 365                     if (setrlimit(RLIMIT_CORE, &r) == -1) {
 366                         VLOG_WARN("failed to disable core dumps: %s",
 367                                   strerror(errno));
 368                     }
 369                 }
 370
 371                 /* Throttle restarts to no more than once every 10 seconds. */
 372                 if (time(NULL) < last_restart + 10) {
 373                     VLOG_WARN("%s, waiting until 10 seconds since last "
 374                               "restart", status_msg);
 375                     for (;;) {
 376                         time_t now = time(NULL);
 377                         time_t wakeup = last_restart + 10;
 378                         if (now >= wakeup) {
 379                             break;
 380                         }
 381                         sleep(wakeup - now);
 382                     }
 383                 }
 384                 last_restart = time(NULL);
 385
 386                 VLOG_ERR("%s, restarting", status_msg);
 387                 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
 388                 if (!daemon_pid) {
 389                     break;
 390                 }
 391             } else {
 392                 VLOG_INFO("pid %lu died, %s, exiting",
 393                           (unsigned long int) daemon_pid, s);
 394                 free(s);
 395                 exit(0);
 396             }
 397         }
 398     }
 399     free(status_msg);
 400
 401     /* Running in new daemon process. */
 402     proctitle_restore();
 403     free((char *) program_name);
 404     program_name = saved_program_name;
 405 }
 406
 407 /* Close stdin, stdout, stderr.  If we're started from e.g. an SSH session,
 408  * then this keeps us from holding that session open artificially. */
 409 static void
 410 close_standard_fds(void)
 411 {
 412     int null_fd = get_null_fd();
 413     if (null_fd >= 0) {
 414         dup2(null_fd, STDIN_FILENO);
 415         dup2(null_fd, STDOUT_FILENO);
 416         dup2(null_fd, STDERR_FILENO);
 417     }
 418 }
 419
 420 /* If daemonization is configured, then starts daemonization, by forking and
 421  * returning in the child process.  The parent process hangs around until the
 422  * child lets it know either that it completed startup successfully (by calling
 423  * daemon_complete()) or that it failed to start up (by exiting with a nonzero
 424  * exit code). */
 425 void
 426 daemonize_start(void)
 427 {
 428     daemonize_fd = -1;
 429
 430     if (detach) {
 431         if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
 432             /* Running in parent process. */
 433             exit(0);
 434         }
 435         /* Running in daemon or monitor process. */
 436     }
 437
 438     if (monitor) {
 439         int saved_daemonize_fd = daemonize_fd;
 440         pid_t daemon_pid;
 441
 442         daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
 443         if (daemon_pid > 0) {
 444             /* Running in monitor process. */
 445             fork_notify_startup(saved_daemonize_fd);
 446             close_standard_fds();
 447             monitor_daemon(daemon_pid);
 448         }
 449         /* Running in daemon process. */
 450     }
 451
 452     make_pidfile();
 453
 454     /* Make sure that the unixctl commands for vlog get registered in a
 455      * daemon, even before the first log message. */
 456     vlog_init();
 457 }
 458
 459 /* If daemonization is configured, then this function notifies the parent
 460  * process that the child process has completed startup successfully.
 461  *
 462  * Calling this function more than once has no additional effect. */
 463 void
 464 daemonize_complete(void)
 465 {
 466     fork_notify_startup(daemonize_fd);
 467     daemonize_fd = -1;
 468
 469     if (detach) {
 470         setsid();
 471         if (chdir_) {
 472             ignore(chdir("/"));
 473         }
 474         close_standard_fds();
 475         detach = false;
 476     }
 477 }
 478
 479 void
 480 daemon_usage(void)
 481 {
 482     printf(
 483         "\nDaemon options:\n"
 484         "  --detach                run in background as daemon\n"
 485         "  --no-chdir              do not chdir to '/'\n"
 486         "  --pidfile[=FILE]        create pidfile (default: %s/%s.pid)\n"
 487         "  --overwrite-pidfile     with --pidfile, start even if already "
 488                                    "running\n",
 489         ovs_rundir(), program_name);
 490 }
 491
 492 static pid_t
 493 read_pidfile__(const char *pidfile, bool must_exist)
 494 {
 495     char line[128];
 496     struct flock lck;
 497     struct stat s;
 498     FILE *file;
 499     int error;
 500
 501     if ((pidfile_ino || pidfile_dev)
 502         && !stat(pidfile, &s)
 503         && s.st_ino == pidfile_ino && s.st_dev == pidfile_dev) {
 504         /* It's our own pidfile.  We can't afford to open it, because closing
 505          * *any* fd for a file that a process has locked also releases all the
 506          * locks on that file.
 507          *
 508          * Fortunately, we know the associated pid anyhow: */
 509         return getpid();
 510     }
 511
 512     file = fopen(pidfile, "r");
 513     if (!file) {
 514         if (errno == ENOENT && !must_exist) {
 515             return 0;
 516         }
 517         error = errno;
 518         VLOG_WARN("%s: open: %s", pidfile, strerror(error));
 519         goto error;
 520     }
 521
 522     lck.l_type = F_WRLCK;
 523     lck.l_whence = SEEK_SET;
 524     lck.l_start = 0;
 525     lck.l_len = 0;
 526     lck.l_pid = 0;
 527     if (fcntl(fileno(file), F_GETLK, &lck)) {
 528         error = errno;
 529         VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
 530         goto error;
 531     }
 532     if (lck.l_type == F_UNLCK) {
 533         error = ESRCH;
 534         VLOG_WARN("%s: pid file is not locked", pidfile);
 535         goto error;
 536     }
 537
 538     if (!fgets(line, sizeof line, file)) {
 539         if (ferror(file)) {
 540             error = errno;
 541             VLOG_WARN("%s: read: %s", pidfile, strerror(error));
 542         } else {
 543             error = ESRCH;
 544             VLOG_WARN("%s: read: unexpected end of file", pidfile);
 545         }
 546         goto error;
 547     }
 548
 549     if (lck.l_pid != strtoul(line, NULL, 10)) {
 550         error = ESRCH;
 551         VLOG_WARN("l_pid (%ld) != %s pid (%s)",
 552                    (long int) lck.l_pid, pidfile, line);
 553         goto error;
 554     }
 555
 556     fclose(file);
 557     return lck.l_pid;
 558
 559 error:
 560     if (file) {
 561         fclose(file);
 562     }
 563     return -error;
 564 }
 565
 566 /* Opens and reads a PID from 'pidfile'.  Returns the positive PID if
 567  * successful, otherwise a negative errno value. */
 568 pid_t
 569 read_pidfile(const char *pidfile)
 570 {
 571     return read_pidfile__(pidfile, true);
 572 }
 573
 574
 575 /* Opens and reads a PID from 'pidfile', if it exists.  Returns 0 if 'pidfile'
 576  * doesn't exist, the positive PID if successful, otherwise a negative errno
 577  * value. */
 578 pid_t
 579 read_pidfile_if_exists(const char *pidfile)
 580 {
 581     return read_pidfile__(pidfile, false);
 582 }