2 * Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira, Inc.
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:
8 * http://www.apache.org/licenses/LICENSE-2.0
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.
24 #include <sys/resource.h>
28 #include "command-line.h"
29 #include "fatal-signal.h"
33 #include "socket-util.h"
38 VLOG_DEFINE_THIS_MODULE(daemon);
40 /* --detach: Should we run in the background? */
41 static bool detach; /* Was --detach specified? */
42 static bool detached; /* Have we already detached? */
44 /* --pidfile: Name of pidfile (null if none). */
47 /* Device and inode of pidfile, so we can avoid reopening it. */
48 static dev_t pidfile_dev;
49 static ino_t pidfile_ino;
51 /* --overwrite-pidfile: Create pidfile even if one already exists and is
53 static bool overwrite_pidfile;
55 /* --no-chdir: Should we chdir to "/"? */
56 static bool chdir_ = true;
58 /* File descriptor used by daemonize_start() and daemonize_complete(). */
59 static int daemonize_fd = -1;
61 /* --monitor: Should a supervisory process monitor the daemon and restart it if
62 * it dies due to an error signal? */
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];
69 static void check_already_running(void);
70 static int lock_pidfile(FILE *, int command);
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. */
75 make_pidfile_name(const char *name)
78 ? xasprintf("%s/%s.pid", ovs_rundir(), program_name)
79 : abs_file_name(ovs_rundir(), name));
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
87 * If 'name' is null, then program_name followed by ".pid" is used. */
89 set_pidfile(const char *name)
92 pidfile = make_pidfile_name(name);
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
104 /* Sets that we do not chdir to "/". */
111 /* Will we chdir to "/" as part of daemonizing? */
113 is_chdir_enabled(void)
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. */
122 ignore_existing_pidfile(void)
124 overwrite_pidfile = true;
127 /* Sets up a following call to daemonize() to detach from the foreground
128 * session, running this process in the background. */
135 /* Will daemonize() really detach? */
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. */
145 daemon_set_monitor(void)
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. */
158 daemon_save_fd(int fd)
160 ovs_assert(fd == STDIN_FILENO ||
161 fd == STDOUT_FILENO ||
162 fd == STDERR_FILENO);
166 /* Unregisters pidfile from being unlinked when the program terminates via
167 * exit() or a fatal signal. */
169 remove_pidfile_from_unlink(void)
172 fatal_signal_remove_file_to_unlink(pidfile);
176 /* Registers pidfile to be unlinked when the program terminates via exit() or a
179 add_pidfile_to_unlink(void)
182 fatal_signal_add_file_to_unlink(pidfile);
186 /* If a pidfile has been configured, creates it and stores the running
187 * process's pid in it. Ensures that the pidfile will be deleted when the
192 long int pid = getpid();
198 /* Create a temporary pidfile. */
199 if (overwrite_pidfile) {
200 tmpfile = xasprintf("%s.tmp%ld", pidfile, pid);
201 fatal_signal_add_file_to_unlink(tmpfile);
203 /* Everyone shares the same file which will be treated as a lock. To
204 * avoid some uncomfortable race conditions, we can't set up the fatal
205 * signal unlink until we've acquired it. */
206 tmpfile = xasprintf("%s.tmp", pidfile);
209 file = fopen(tmpfile, "a+");
211 VLOG_FATAL("%s: create failed (%s)", tmpfile, strerror(errno));
214 error = lock_pidfile(file, F_SETLK);
216 /* Looks like we failed to acquire the lock. Note that, if we failed
217 * for some other reason (and '!overwrite_pidfile'), we will have
218 * left 'tmpfile' as garbage in the file system. */
219 VLOG_FATAL("%s: fcntl(F_SETLK) failed (%s)", tmpfile, strerror(error));
222 if (!overwrite_pidfile) {
223 /* We acquired the lock. Make sure to clean up on exit, and verify
224 * that we're allowed to create the actual pidfile. */
225 fatal_signal_add_file_to_unlink(tmpfile);
226 check_already_running();
229 if (fstat(fileno(file), &s) == -1) {
230 VLOG_FATAL("%s: fstat failed (%s)", tmpfile, strerror(errno));
233 if (ftruncate(fileno(file), 0) == -1) {
234 VLOG_FATAL("%s: truncate failed (%s)", tmpfile, strerror(errno));
237 fprintf(file, "%ld\n", pid);
238 if (fflush(file) == EOF) {
239 VLOG_FATAL("%s: write failed (%s)", tmpfile, strerror(errno));
242 error = rename(tmpfile, pidfile);
244 /* Due to a race, 'tmpfile' may be owned by a different process, so we
245 * shouldn't delete it on exit. */
246 fatal_signal_remove_file_to_unlink(tmpfile);
249 VLOG_FATAL("failed to rename \"%s\" to \"%s\" (%s)",
250 tmpfile, pidfile, strerror(errno));
253 /* Ensure that the pidfile will get deleted on exit. */
254 fatal_signal_add_file_to_unlink(pidfile);
258 * We don't close 'file' because its file descriptor must remain open to
260 pidfile_dev = s.st_dev;
261 pidfile_ino = s.st_ino;
265 /* If configured with set_pidfile() or set_detach(), creates the pid file and
266 * detaches from the foreground session. */
271 daemonize_complete();
274 /* Calls fork() and on success returns its return value. On failure, logs an
275 * error and exits unsuccessfully.
277 * Post-fork, but before returning, this function calls a few other functions
278 * that are generally useful if the child isn't planning to exec a new
281 fork_and_clean_up(void)
287 /* Running in parent process. */
290 /* Running in child process. */
294 VLOG_FATAL("fork failed (%s)", strerror(errno));
302 * - In the parent, waits for the child to signal that it has completed its
303 * startup sequence. Then stores -1 in '*fdp' and returns the child's pid.
305 * - In the child, stores a fd in '*fdp' and returns 0. The caller should
306 * pass the fd to fork_notify_startup() after it finishes its startup
309 * If something goes wrong with the fork, logs a critical error and aborts the
312 fork_and_wait_for_startup(int *fdp)
319 pid = fork_and_clean_up();
321 /* Running in parent process. */
326 if (read_fully(fds[0], &c, 1, &bytes_read) != 0) {
331 retval = waitpid(pid, &status, 0);
332 } while (retval == -1 && errno == EINTR);
335 if (WIFEXITED(status) && WEXITSTATUS(status)) {
336 /* Child exited with an error. Convey the same error
337 * to our parent process as a courtesy. */
338 exit(WEXITSTATUS(status));
340 char *status_msg = process_status_msg(status);
341 VLOG_FATAL("fork child died before signaling startup (%s)",
344 } else if (retval < 0) {
345 VLOG_FATAL("waitpid failed (%s)", strerror(errno));
353 /* Running in child process. */
362 fork_notify_startup(int fd)
365 size_t bytes_written;
368 error = write_fully(fd, "", 1, &bytes_written);
370 VLOG_FATAL("pipe write failed (%s)", strerror(error));
378 should_restart(int status)
380 if (WIFSIGNALED(status)) {
381 static const int error_signals[] = {
382 SIGABRT, SIGALRM, SIGBUS, SIGFPE, SIGILL, SIGPIPE, SIGSEGV,
388 for (i = 0; i < ARRAY_SIZE(error_signals); i++) {
389 if (error_signals[i] == WTERMSIG(status)) {
398 monitor_daemon(pid_t daemon_pid)
400 /* XXX Should log daemon's stderr output at startup time. */
405 subprogram_name = "monitor";
406 status_msg = xstrdup("healthy");
407 last_restart = TIME_MIN;
413 proctitle_set("monitoring pid %lu (%s)",
414 (unsigned long int) daemon_pid, status_msg);
417 retval = waitpid(daemon_pid, &status, 0);
418 } while (retval == -1 && errno == EINTR);
421 VLOG_FATAL("waitpid failed (%s)", strerror(errno));
422 } else if (retval == daemon_pid) {
423 char *s = process_status_msg(status);
424 if (should_restart(status)) {
426 status_msg = xasprintf("%d crashes: pid %lu died, %s",
428 (unsigned long int) daemon_pid, s);
431 if (WCOREDUMP(status)) {
432 /* Disable further core dumps to save disk space. */
437 if (setrlimit(RLIMIT_CORE, &r) == -1) {
438 VLOG_WARN("failed to disable core dumps: %s",
443 /* Throttle restarts to no more than once every 10 seconds. */
444 if (time(NULL) < last_restart + 10) {
445 VLOG_WARN("%s, waiting until 10 seconds since last "
446 "restart", status_msg);
448 time_t now = time(NULL);
449 time_t wakeup = last_restart + 10;
456 last_restart = time(NULL);
458 VLOG_ERR("%s, restarting", status_msg);
459 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
464 VLOG_INFO("pid %lu died, %s, exiting",
465 (unsigned long int) daemon_pid, s);
473 /* Running in new daemon process. */
475 subprogram_name = "";
478 /* Close standard file descriptors (except any that the client has requested we
479 * leave open by calling daemon_save_fd()). If we're started from e.g. an SSH
480 * session, then this keeps us from holding that session open artificially. */
482 close_standard_fds(void)
484 int null_fd = get_null_fd();
488 for (fd = 0; fd < 3; fd++) {
495 /* Disable logging to stderr to avoid wasting CPU time. */
496 vlog_set_levels(NULL, VLF_CONSOLE, VLL_OFF);
499 /* If daemonization is configured, then starts daemonization, by forking and
500 * returning in the child process. The parent process hangs around until the
501 * child lets it know either that it completed startup successfully (by calling
502 * daemon_complete()) or that it failed to start up (by exiting with a nonzero
505 daemonize_start(void)
510 if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
511 /* Running in parent process. */
515 /* Running in daemon or monitor process. */
520 int saved_daemonize_fd = daemonize_fd;
523 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
524 if (daemon_pid > 0) {
525 /* Running in monitor process. */
526 fork_notify_startup(saved_daemonize_fd);
527 close_standard_fds();
528 monitor_daemon(daemon_pid);
530 /* Running in daemon process. */
537 /* Make sure that the unixctl commands for vlog get registered in a
538 * daemon, even before the first log message. */
542 /* If daemonization is configured, then this function notifies the parent
543 * process that the child process has completed startup successfully. It also
544 * call daemonize_post_detach().
546 * Calling this function more than once has no additional effect. */
548 daemonize_complete(void)
558 fork_notify_startup(daemonize_fd);
560 daemonize_post_detach();
564 /* If daemonization is configured, then this function does traditional Unix
565 * daemonization behavior: join a new session, chdir to the root (if not
566 * disabled), and close the standard file descriptors.
568 * It only makes sense to call this function as part of an implementation of a
569 * special daemon subprocess. A normal daemon should just call
570 * daemonize_complete(). */
572 daemonize_post_detach(void)
578 close_standard_fds();
586 "\nDaemon options:\n"
587 " --detach run in background as daemon\n"
588 " --no-chdir do not chdir to '/'\n"
589 " --pidfile[=FILE] create pidfile (default: %s/%s.pid)\n"
590 " --overwrite-pidfile with --pidfile, start even if already "
592 ovs_rundir(), program_name);
596 lock_pidfile__(FILE *file, int command, struct flock *lck)
600 lck->l_type = F_WRLCK;
601 lck->l_whence = SEEK_SET;
607 error = fcntl(fileno(file), command, lck) == -1 ? errno : 0;
608 } while (error == EINTR);
613 lock_pidfile(FILE *file, int command)
617 return lock_pidfile__(file, command, &lck);
621 read_pidfile__(const char *pidfile, bool delete_if_stale)
629 if ((pidfile_ino || pidfile_dev)
630 && !stat(pidfile, &s)
631 && s.st_ino == pidfile_ino && s.st_dev == pidfile_dev) {
632 /* It's our own pidfile. We can't afford to open it, because closing
633 * *any* fd for a file that a process has locked also releases all the
634 * locks on that file.
636 * Fortunately, we know the associated pid anyhow: */
640 file = fopen(pidfile, "r+");
642 if (errno == ENOENT && delete_if_stale) {
646 VLOG_WARN("%s: open: %s", pidfile, strerror(error));
650 error = lock_pidfile__(file, F_GETLK, &lck);
652 VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
655 if (lck.l_type == F_UNLCK) {
656 /* pidfile exists but it isn't locked by anyone. We need to delete it
657 * so that a new pidfile can go in its place. But just calling
658 * unlink(pidfile) makes a nasty race: what if someone else unlinks it
659 * before we do and then replaces it by a valid pidfile? We'd unlink
660 * their valid pidfile. We do a little dance to avoid the race, by
661 * locking the invalid pidfile. Only one process can have the invalid
662 * pidfile locked, and only that process has the right to unlink it. */
663 if (!delete_if_stale) {
665 VLOG_DBG("%s: pid file is stale", pidfile);
670 error = lock_pidfile(file, F_SETLK);
672 /* We lost a race with someone else doing the same thing. */
673 VLOG_WARN("%s: lost race to lock pidfile", pidfile);
677 /* Is the file we have locked still named 'pidfile'? */
678 if (stat(pidfile, &s) || fstat(fileno(file), &s2)
679 || s.st_ino != s2.st_ino || s.st_dev != s2.st_dev) {
680 /* No. We lost a race with someone else who got the lock before
681 * us, deleted the pidfile, and closed it (releasing the lock). */
683 VLOG_WARN("%s: lost race to delete pidfile", pidfile);
687 /* We won the right to delete the stale pidfile. */
688 if (unlink(pidfile)) {
690 VLOG_WARN("%s: failed to delete stale pidfile (%s)",
691 pidfile, strerror(error));
694 VLOG_DBG("%s: deleted stale pidfile", pidfile);
699 if (!fgets(line, sizeof line, file)) {
702 VLOG_WARN("%s: read: %s", pidfile, strerror(error));
705 VLOG_WARN("%s: read: unexpected end of file", pidfile);
710 if (lck.l_pid != strtoul(line, NULL, 10)) {
711 /* The process that has the pidfile locked is not the process that
712 * created it. It must be stale, with the process that has it locked
713 * preparing to delete it. */
715 VLOG_WARN("%s: stale pidfile for pid %s being deleted by pid %ld",
716 pidfile, line, (long int) lck.l_pid);
730 /* Opens and reads a PID from 'pidfile'. Returns the positive PID if
731 * successful, otherwise a negative errno value. */
733 read_pidfile(const char *pidfile)
735 return read_pidfile__(pidfile, false);
738 /* Checks whether a process with the given 'pidfile' is already running and,
739 * if so, aborts. If 'pidfile' is stale, deletes it. */
741 check_already_running(void)
743 long int pid = read_pidfile__(pidfile, true);
745 VLOG_FATAL("%s: already running as pid %ld, aborting", pidfile, pid);
746 } else if (pid < 0) {
747 VLOG_FATAL("%s: pidfile check failed (%s), aborting",
748 pidfile, strerror(-pid));