lib/daemon.c

   1 /*
   2  * Copyright (c) 2008, 2009, 2010 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 pidfile has been configured and that pidfile already exists and is
 143  * locked by a running process, returns the pid of the running process.
 144  * Otherwise, returns 0. */
 145 static pid_t
 146 already_running(void)
 147 {
 148     pid_t pid = 0;
 149     if (pidfile) {
 150         int fd = open(pidfile, O_RDWR);
 151         if (fd >= 0) {
 152             struct flock lck;
 153             lck.l_type = F_WRLCK;
 154             lck.l_whence = SEEK_SET;
 155             lck.l_start = 0;
 156             lck.l_len = 0;
 157             if (fcntl(fd, F_GETLK, &lck) != -1 && lck.l_type != F_UNLCK) {
 158                 pid = lck.l_pid;
 159             }
 160             close(fd);
 161         }
 162     }
 163     return pid;
 164 }
 165
 166 /* If a locked pidfile exists, issue a warning message and, unless
 167  * ignore_existing_pidfile() has been called, terminate the program. */
 168 void
 169 die_if_already_running(void)
 170 {
 171     pid_t pid = already_running();
 172     if (pid) {
 173         if (!overwrite_pidfile) {
 174             ovs_fatal(0, "%s: already running as pid %ld",
 175                       get_pidfile(), (long int) pid);
 176         } else {
 177             VLOG_WARN("%s: %s already running as pid %ld",
 178                       get_pidfile(), program_name, (long int) pid);
 179         }
 180     }
 181 }
 182
 183 /* If a pidfile has been configured, creates it and stores the running
 184  * process's pid in it.  Ensures that the pidfile will be deleted when the
 185  * process exits. */
 186 static void
 187 make_pidfile(void)
 188 {
 189     if (pidfile) {
 190         /* Create pidfile via temporary file, so that observers never see an
 191          * empty pidfile or an unlocked pidfile. */
 192         long int pid = getpid();
 193         char *tmpfile;
 194         int fd;
 195
 196         tmpfile = xasprintf("%s.tmp%ld", pidfile, pid);
 197         fatal_signal_add_file_to_unlink(tmpfile);
 198         fd = open(tmpfile, O_CREAT | O_WRONLY | O_TRUNC, 0666);
 199         if (fd >= 0) {
 200             struct flock lck;
 201             lck.l_type = F_WRLCK;
 202             lck.l_whence = SEEK_SET;
 203             lck.l_start = 0;
 204             lck.l_len = 0;
 205             if (fcntl(fd, F_SETLK, &lck) != -1) {
 206                 char *text = xasprintf("%ld\n", pid);
 207                 if (write(fd, text, strlen(text)) == strlen(text)) {
 208                     fatal_signal_add_file_to_unlink(pidfile);
 209                     if (rename(tmpfile, pidfile) < 0) {
 210                         VLOG_ERR("failed to rename \"%s\" to \"%s\": %s",
 211                                  tmpfile, pidfile, strerror(errno));
 212                         fatal_signal_remove_file_to_unlink(pidfile);
 213                         close(fd);
 214                     } else {
 215                         /* Keep 'fd' open to retain the lock. */
 216                         struct stat s;
 217
 218                         if (!fstat(fd, &s)) {
 219                             pidfile_dev = s.st_dev;
 220                             pidfile_ino = s.st_ino;
 221                         } else {
 222                             VLOG_ERR("%s: fstat failed: %s",
 223                                      pidfile, strerror(errno));
 224                         }
 225                     }
 226                     free(text);
 227                 } else {
 228                     VLOG_ERR("%s: write failed: %s", tmpfile, strerror(errno));
 229                     close(fd);
 230                 }
 231             } else {
 232                 VLOG_ERR("%s: fcntl failed: %s", tmpfile, strerror(errno));
 233                 close(fd);
 234             }
 235         } else {
 236             VLOG_ERR("%s: create failed: %s", tmpfile, strerror(errno));
 237         }
 238         fatal_signal_remove_file_to_unlink(tmpfile);
 239         free(tmpfile);
 240     }
 241     free(pidfile);
 242     pidfile = NULL;
 243 }
 244
 245 /* If configured with set_pidfile() or set_detach(), creates the pid file and
 246  * detaches from the foreground session.  */
 247 void
 248 daemonize(void)
 249 {
 250     daemonize_start();
 251     daemonize_complete();
 252 }
 253
 254 static pid_t
 255 fork_and_wait_for_startup(int *fdp)
 256 {
 257     int fds[2];
 258     pid_t pid;
 259
 260     if (pipe(fds) < 0) {
 261         ovs_fatal(errno, "pipe failed");
 262     }
 263
 264     pid = fork();
 265     if (pid > 0) {
 266         /* Running in parent process. */
 267         char c;
 268
 269         close(fds[1]);
 270         fatal_signal_fork();
 271         if (read(fds[0], &c, 1) != 1) {
 272             int retval;
 273             int status;
 274
 275             do {
 276                 retval = waitpid(pid, &status, 0);
 277             } while (retval == -1 && errno == EINTR);
 278
 279             if (retval == pid
 280                 && WIFEXITED(status)
 281                 && WEXITSTATUS(status)) {
 282                 /* Child exited with an error.  Convey the same error to
 283                  * our parent process as a courtesy. */
 284                 exit(WEXITSTATUS(status));
 285             }
 286
 287             ovs_fatal(errno, "fork child failed to signal startup");
 288         }
 289         close(fds[0]);
 290         *fdp = -1;
 291     } else if (!pid) {
 292         /* Running in child process. */
 293         close(fds[0]);
 294         time_postfork();
 295         lockfile_postfork();
 296         *fdp = fds[1];
 297     } else {
 298         ovs_fatal(errno, "could not fork");
 299     }
 300
 301     return pid;
 302 }
 303
 304 static void
 305 fork_notify_startup(int fd)
 306 {
 307     if (fd != -1) {
 308         size_t bytes_written;
 309         int error;
 310
 311         error = write_fully(fd, "", 1, &bytes_written);
 312         if (error) {
 313             ovs_fatal(error, "could not write to pipe");
 314         }
 315
 316         close(fd);
 317     }
 318 }
 319
 320 static bool
 321 should_restart(int status)
 322 {
 323     if (WIFSIGNALED(status)) {
 324         static const int error_signals[] = {
 325             SIGABRT, SIGALRM, SIGBUS, SIGFPE, SIGILL, SIGPIPE, SIGSEGV,
 326             SIGXCPU, SIGXFSZ
 327         };
 328
 329         size_t i;
 330
 331         for (i = 0; i < ARRAY_SIZE(error_signals); i++) {
 332             if (error_signals[i] == WTERMSIG(status)) {
 333                 return true;
 334             }
 335         }
 336     }
 337     return false;
 338 }
 339
 340 static void
 341 monitor_daemon(pid_t daemon_pid)
 342 {
 343     /* XXX Should log daemon's stderr output at startup time. */
 344     const char *saved_program_name;
 345     time_t last_restart;
 346     char *status_msg;
 347     int crashes;
 348
 349     saved_program_name = program_name;
 350     program_name = xasprintf("monitor(%s)", program_name);
 351     status_msg = xstrdup("healthy");
 352     last_restart = TIME_MIN;
 353     crashes = 0;
 354     for (;;) {
 355         int retval;
 356         int status;
 357
 358         proctitle_set("%s: monitoring pid %lu (%s)",
 359                       saved_program_name, (unsigned long int) daemon_pid,
 360                       status_msg);
 361
 362         do {
 363             retval = waitpid(daemon_pid, &status, 0);
 364         } while (retval == -1 && errno == EINTR);
 365
 366         if (retval == -1) {
 367             ovs_fatal(errno, "waitpid failed");
 368         } else if (retval == daemon_pid) {
 369             char *s = process_status_msg(status);
 370             if (should_restart(status)) {
 371                 free(status_msg);
 372                 status_msg = xasprintf("%d crashes: pid %lu died, %s",
 373                                        ++crashes,
 374                                        (unsigned long int) daemon_pid, s);
 375                 free(s);
 376
 377                 if (WCOREDUMP(status)) {
 378                     /* Disable further core dumps to save disk space. */
 379                     struct rlimit r;
 380
 381                     r.rlim_cur = 0;
 382                     r.rlim_max = 0;
 383                     if (setrlimit(RLIMIT_CORE, &r) == -1) {
 384                         VLOG_WARN("failed to disable core dumps: %s",
 385                                   strerror(errno));
 386                     }
 387                 }
 388
 389                 /* Throttle restarts to no more than once every 10 seconds. */
 390                 if (time(NULL) < last_restart + 10) {
 391                     VLOG_WARN("%s, waiting until 10 seconds since last "
 392                               "restart", status_msg);
 393                     for (;;) {
 394                         time_t now = time(NULL);
 395                         time_t wakeup = last_restart + 10;
 396                         if (now >= wakeup) {
 397                             break;
 398                         }
 399                         sleep(wakeup - now);
 400                     }
 401                 }
 402                 last_restart = time(NULL);
 403
 404                 VLOG_ERR("%s, restarting", status_msg);
 405                 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
 406                 if (!daemon_pid) {
 407                     break;
 408                 }
 409             } else {
 410                 VLOG_INFO("pid %lu died, %s, exiting",
 411                           (unsigned long int) daemon_pid, s);
 412                 free(s);
 413                 exit(0);
 414             }
 415         }
 416     }
 417     free(status_msg);
 418
 419     /* Running in new daemon process. */
 420     proctitle_restore();
 421     free((char *) program_name);
 422     program_name = saved_program_name;
 423 }
 424
 425 /* Close stdin, stdout, stderr.  If we're started from e.g. an SSH session,
 426  * then this keeps us from holding that session open artificially. */
 427 static void
 428 close_standard_fds(void)
 429 {
 430     int null_fd = get_null_fd();
 431     if (null_fd >= 0) {
 432         dup2(null_fd, STDIN_FILENO);
 433         dup2(null_fd, STDOUT_FILENO);
 434         dup2(null_fd, STDERR_FILENO);
 435     }
 436 }
 437
 438 /* If daemonization is configured, then starts daemonization, by forking and
 439  * returning in the child process.  The parent process hangs around until the
 440  * child lets it know either that it completed startup successfully (by calling
 441  * daemon_complete()) or that it failed to start up (by exiting with a nonzero
 442  * exit code). */
 443 void
 444 daemonize_start(void)
 445 {
 446     daemonize_fd = -1;
 447
 448     if (detach) {
 449         if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
 450             /* Running in parent process. */
 451             exit(0);
 452         }
 453         /* Running in daemon or monitor process. */
 454     }
 455
 456     if (monitor) {
 457         int saved_daemonize_fd = daemonize_fd;
 458         pid_t daemon_pid;
 459
 460         daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
 461         if (daemon_pid > 0) {
 462             /* Running in monitor process. */
 463             fork_notify_startup(saved_daemonize_fd);
 464             close_standard_fds();
 465             monitor_daemon(daemon_pid);
 466         }
 467         /* Running in daemon process. */
 468     }
 469
 470     make_pidfile();
 471
 472     /* Make sure that the unixctl commands for vlog get registered in a
 473      * daemon, even before the first log message. */
 474     vlog_init();
 475 }
 476
 477 /* If daemonization is configured, then this function notifies the parent
 478  * process that the child process has completed startup successfully. */
 479 void
 480 daemonize_complete(void)
 481 {
 482     fork_notify_startup(daemonize_fd);
 483
 484     if (detach) {
 485         setsid();
 486         if (chdir_) {
 487             ignore(chdir("/"));
 488         }
 489         close_standard_fds();
 490     }
 491 }
 492
 493 void
 494 daemon_usage(void)
 495 {
 496     printf(
 497         "\nDaemon options:\n"
 498         "  --detach                run in background as daemon\n"
 499         "  --no-chdir              do not chdir to '/'\n"
 500         "  --pidfile[=FILE]        create pidfile (default: %s/%s.pid)\n"
 501         "  --overwrite-pidfile     with --pidfile, start even if already "
 502                                    "running\n",
 503         ovs_rundir, program_name);
 504 }
 505
 506 /* Opens and reads a PID from 'pidfile'.  Returns the nonnegative PID if
 507  * successful, otherwise a negative errno value. */
 508 pid_t
 509 read_pidfile(const char *pidfile)
 510 {
 511     char line[128];
 512     struct flock lck;
 513     struct stat s;
 514     FILE *file;
 515     int error;
 516
 517     if ((pidfile_ino || pidfile_dev)
 518         && !stat(pidfile, &s)
 519         && s.st_ino == pidfile_ino && s.st_dev == pidfile_dev) {
 520         /* It's our own pidfile.  We can't afford to open it, because closing
 521          * *any* fd for a file that a process has locked also releases all the
 522          * locks on that file.
 523          *
 524          * Fortunately, we know the associated pid anyhow: */
 525         return getpid();
 526     }
 527
 528     file = fopen(pidfile, "r");
 529     if (!file) {
 530         error = errno;
 531         VLOG_WARN("%s: open: %s", pidfile, strerror(error));
 532         goto error;
 533     }
 534
 535     lck.l_type = F_WRLCK;
 536     lck.l_whence = SEEK_SET;
 537     lck.l_start = 0;
 538     lck.l_len = 0;
 539     if (fcntl(fileno(file), F_GETLK, &lck)) {
 540         error = errno;
 541         VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
 542         goto error;
 543     }
 544     if (lck.l_type == F_UNLCK) {
 545         error = ESRCH;
 546         VLOG_WARN("%s: pid file is not locked", pidfile);
 547         goto error;
 548     }
 549
 550     if (!fgets(line, sizeof line, file)) {
 551         if (ferror(file)) {
 552             error = errno;
 553             VLOG_WARN("%s: read: %s", pidfile, strerror(error));
 554         } else {
 555             error = ESRCH;
 556             VLOG_WARN("%s: read: unexpected end of file", pidfile);
 557         }
 558         goto error;
 559     }
 560
 561     if (lck.l_pid != strtoul(line, NULL, 10)) {
 562         error = ESRCH;
 563         VLOG_WARN("l_pid (%ld) != %s pid (%s)",
 564                    (long int) lck.l_pid, pidfile, line);
 565         goto error;
 566     }
 567
 568     fclose(file);
 569     return lck.l_pid;
 570
 571 error:
 572     if (file) {
 573         fclose(file);
 574     }
 575     return -error;
 576 }