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.
25 #include <sys/resource.h>
29 #include "command-line.h"
30 #include "fatal-signal.h"
34 #include "socket-util.h"
39 VLOG_DEFINE_THIS_MODULE(daemon);
41 /* --detach: Should we run in the background? */
42 static bool detach; /* Was --detach specified? */
43 static bool detached; /* Have we already detached? */
45 /* --pidfile: Name of pidfile (null if none). */
48 /* Device and inode of pidfile, so we can avoid reopening it. */
49 static dev_t pidfile_dev;
50 static ino_t pidfile_ino;
52 /* --overwrite-pidfile: Create pidfile even if one already exists and is
54 static bool overwrite_pidfile;
56 /* --no-chdir: Should we chdir to "/"? */
57 static bool chdir_ = true;
59 /* File descriptor used by daemonize_start() and daemonize_complete(). */
60 static int daemonize_fd = -1;
62 /* --monitor: Should a supervisory process monitor the daemon and restart it if
63 * it dies due to an error signal? */
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];
70 static void check_already_running(void);
71 static int lock_pidfile(FILE *, int command);
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. */
76 make_pidfile_name(const char *name)
79 ? xasprintf("%s/%s.pid", ovs_rundir(), program_name)
80 : abs_file_name(ovs_rundir(), name));
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
88 * If 'name' is null, then program_name followed by ".pid" is used. */
90 set_pidfile(const char *name)
93 pidfile = make_pidfile_name(name);
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
105 /* Sets that we do not chdir to "/". */
112 /* Will we chdir to "/" as part of daemonizing? */
114 is_chdir_enabled(void)
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. */
123 ignore_existing_pidfile(void)
125 overwrite_pidfile = true;
128 /* Sets up a following call to daemonize() to detach from the foreground
129 * session, running this process in the background. */
136 /* Will daemonize() really detach? */
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. */
146 daemon_set_monitor(void)
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. */
159 daemon_save_fd(int fd)
161 assert(fd == STDIN_FILENO || fd == STDOUT_FILENO || fd == STDERR_FILENO);
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
171 long int pid = getpid();
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+");
182 VLOG_FATAL("%s: create failed (%s)", tmpfile, strerror(errno));
185 if (fstat(fileno(file), &s) == -1) {
186 VLOG_FATAL("%s: fstat failed (%s)", tmpfile, strerror(errno));
189 fprintf(file, "%ld\n", pid);
190 if (fflush(file) == EOF) {
191 VLOG_FATAL("%s: write failed (%s)", tmpfile, strerror(errno));
194 error = lock_pidfile(file, F_SETLK);
196 VLOG_FATAL("%s: fcntl(F_SETLK) failed (%s)", tmpfile, strerror(error));
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));
207 error = link(tmpfile, pidfile) == -1 ? errno : 0;
208 if (error == EEXIST) {
209 check_already_running();
211 } while (error == EINTR || error == EEXIST);
213 VLOG_FATAL("failed to link \"%s\" as \"%s\" (%s)",
214 tmpfile, pidfile, strerror(error));
218 /* Ensure that the pidfile will get deleted on exit. */
219 fatal_signal_add_file_to_unlink(pidfile);
221 /* Delete the temporary pidfile if it still exists. */
222 if (!overwrite_pidfile) {
223 error = fatal_signal_unlink_file_now(tmpfile);
225 VLOG_FATAL("%s: unlink failed (%s)", tmpfile, strerror(error));
231 * We don't close 'file' because its file descriptor must remain open to
233 pidfile_dev = s.st_dev;
234 pidfile_ino = s.st_ino;
240 /* If configured with set_pidfile() or set_detach(), creates the pid file and
241 * detaches from the foreground session. */
246 daemonize_complete();
249 /* Calls fork() and on success returns its return value. On failure, logs an
250 * error and exits unsuccessfully.
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
256 fork_and_clean_up(void)
262 /* Running in parent process. */
265 /* Running in child process. */
269 VLOG_FATAL("fork failed (%s)", strerror(errno));
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.
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
284 * If something goes wrong with the fork, logs a critical error and aborts the
287 fork_and_wait_for_startup(int *fdp)
294 pid = fork_and_clean_up();
296 /* Running in parent process. */
301 if (read_fully(fds[0], &c, 1, &bytes_read) != 0) {
306 retval = waitpid(pid, &status, 0);
307 } while (retval == -1 && errno == EINTR);
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));
315 char *status_msg = process_status_msg(status);
316 VLOG_FATAL("fork child died before signaling startup (%s)",
319 } else if (retval < 0) {
320 VLOG_FATAL("waitpid failed (%s)", strerror(errno));
328 /* Running in child process. */
337 fork_notify_startup(int fd)
340 size_t bytes_written;
343 error = write_fully(fd, "", 1, &bytes_written);
345 VLOG_FATAL("pipe write failed (%s)", strerror(error));
353 should_restart(int status)
355 if (WIFSIGNALED(status)) {
356 static const int error_signals[] = {
357 SIGABRT, SIGALRM, SIGBUS, SIGFPE, SIGILL, SIGPIPE, SIGSEGV,
363 for (i = 0; i < ARRAY_SIZE(error_signals); i++) {
364 if (error_signals[i] == WTERMSIG(status)) {
373 monitor_daemon(pid_t daemon_pid)
375 /* XXX Should log daemon's stderr output at startup time. */
380 subprogram_name = "monitor";
381 status_msg = xstrdup("healthy");
382 last_restart = TIME_MIN;
388 proctitle_set("monitoring pid %lu (%s)",
389 (unsigned long int) daemon_pid, status_msg);
392 retval = waitpid(daemon_pid, &status, 0);
393 } while (retval == -1 && errno == EINTR);
396 VLOG_FATAL("waitpid failed (%s)", strerror(errno));
397 } else if (retval == daemon_pid) {
398 char *s = process_status_msg(status);
399 if (should_restart(status)) {
401 status_msg = xasprintf("%d crashes: pid %lu died, %s",
403 (unsigned long int) daemon_pid, s);
406 if (WCOREDUMP(status)) {
407 /* Disable further core dumps to save disk space. */
412 if (setrlimit(RLIMIT_CORE, &r) == -1) {
413 VLOG_WARN("failed to disable core dumps: %s",
418 /* Throttle restarts to no more than once every 10 seconds. */
419 if (time(NULL) < last_restart + 10) {
420 VLOG_WARN("%s, waiting until 10 seconds since last "
421 "restart", status_msg);
423 time_t now = time(NULL);
424 time_t wakeup = last_restart + 10;
431 last_restart = time(NULL);
433 VLOG_ERR("%s, restarting", status_msg);
434 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
439 VLOG_INFO("pid %lu died, %s, exiting",
440 (unsigned long int) daemon_pid, s);
448 /* Running in new daemon process. */
450 subprogram_name = "";
453 /* Close standard file descriptors (except any that the client has requested we
454 * leave open by calling daemon_save_fd()). If we're started from e.g. an SSH
455 * session, then this keeps us from holding that session open artificially. */
457 close_standard_fds(void)
459 int null_fd = get_null_fd();
463 for (fd = 0; fd < 3; fd++) {
470 /* Disable logging to stderr to avoid wasting CPU time. */
471 vlog_set_levels(NULL, VLF_CONSOLE, VLL_OFF);
474 /* If daemonization is configured, then starts daemonization, by forking and
475 * returning in the child process. The parent process hangs around until the
476 * child lets it know either that it completed startup successfully (by calling
477 * daemon_complete()) or that it failed to start up (by exiting with a nonzero
480 daemonize_start(void)
485 if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
486 /* Running in parent process. */
489 /* Running in daemon or monitor process. */
493 int saved_daemonize_fd = daemonize_fd;
496 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
497 if (daemon_pid > 0) {
498 /* Running in monitor process. */
499 fork_notify_startup(saved_daemonize_fd);
500 close_standard_fds();
501 monitor_daemon(daemon_pid);
503 /* Running in daemon process. */
510 /* Make sure that the unixctl commands for vlog get registered in a
511 * daemon, even before the first log message. */
515 /* If daemonization is configured, then this function notifies the parent
516 * process that the child process has completed startup successfully. It also
517 * call daemonize_post_detach().
519 * Calling this function more than once has no additional effect. */
521 daemonize_complete(void)
526 fork_notify_startup(daemonize_fd);
528 daemonize_post_detach();
532 /* If daemonization is configured, then this function does traditional Unix
533 * daemonization behavior: join a new session, chdir to the root (if not
534 * disabled), and close the standard file descriptors.
536 * It only makes sense to call this function as part of an implementation of a
537 * special daemon subprocess. A normal daemon should just call
538 * daemonize_complete(). */
540 daemonize_post_detach(void)
547 close_standard_fds();
555 "\nDaemon options:\n"
556 " --detach run in background as daemon\n"
557 " --no-chdir do not chdir to '/'\n"
558 " --pidfile[=FILE] create pidfile (default: %s/%s.pid)\n"
559 " --overwrite-pidfile with --pidfile, start even if already "
561 ovs_rundir(), program_name);
565 lock_pidfile__(FILE *file, int command, struct flock *lck)
569 lck->l_type = F_WRLCK;
570 lck->l_whence = SEEK_SET;
576 error = fcntl(fileno(file), command, lck) == -1 ? errno : 0;
577 } while (error == EINTR);
582 lock_pidfile(FILE *file, int command)
586 return lock_pidfile__(file, command, &lck);
590 read_pidfile__(const char *pidfile, bool delete_if_stale)
598 if ((pidfile_ino || pidfile_dev)
599 && !stat(pidfile, &s)
600 && s.st_ino == pidfile_ino && s.st_dev == pidfile_dev) {
601 /* It's our own pidfile. We can't afford to open it, because closing
602 * *any* fd for a file that a process has locked also releases all the
603 * locks on that file.
605 * Fortunately, we know the associated pid anyhow: */
609 file = fopen(pidfile, "r+");
611 if (errno == ENOENT && delete_if_stale) {
615 VLOG_WARN("%s: open: %s", pidfile, strerror(error));
619 error = lock_pidfile__(file, F_GETLK, &lck);
621 VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
624 if (lck.l_type == F_UNLCK) {
625 /* pidfile exists but it isn't locked by anyone. We need to delete it
626 * so that a new pidfile can go in its place. But just calling
627 * unlink(pidfile) makes a nasty race: what if someone else unlinks it
628 * before we do and then replaces it by a valid pidfile? We'd unlink
629 * their valid pidfile. We do a little dance to avoid the race, by
630 * locking the invalid pidfile. Only one process can have the invalid
631 * pidfile locked, and only that process has the right to unlink it. */
632 if (!delete_if_stale) {
634 VLOG_DBG("%s: pid file is stale", pidfile);
639 error = lock_pidfile(file, F_SETLK);
641 /* We lost a race with someone else doing the same thing. */
642 VLOG_WARN("%s: lost race to lock pidfile", pidfile);
646 /* Is the file we have locked still named 'pidfile'? */
647 if (stat(pidfile, &s) || fstat(fileno(file), &s2)
648 || s.st_ino != s2.st_ino || s.st_dev != s2.st_dev) {
649 /* No. We lost a race with someone else who got the lock before
650 * us, deleted the pidfile, and closed it (releasing the lock). */
652 VLOG_WARN("%s: lost race to delete pidfile", pidfile);
656 /* We won the right to delete the stale pidfile. */
657 if (unlink(pidfile)) {
659 VLOG_WARN("%s: failed to delete stale pidfile (%s)",
660 pidfile, strerror(error));
663 VLOG_DBG("%s: deleted stale pidfile", pidfile);
668 if (!fgets(line, sizeof line, file)) {
671 VLOG_WARN("%s: read: %s", pidfile, strerror(error));
674 VLOG_WARN("%s: read: unexpected end of file", pidfile);
679 if (lck.l_pid != strtoul(line, NULL, 10)) {
680 /* The process that has the pidfile locked is not the process that
681 * created it. It must be stale, with the process that has it locked
682 * preparing to delete it. */
684 VLOG_WARN("%s: stale pidfile for pid %s being deleted by pid %ld",
685 pidfile, line, (long int) lck.l_pid);
699 /* Opens and reads a PID from 'pidfile'. Returns the positive PID if
700 * successful, otherwise a negative errno value. */
702 read_pidfile(const char *pidfile)
704 return read_pidfile__(pidfile, false);
707 /* Checks whether a process with the given 'pidfile' is already running and,
708 * if so, aborts. If 'pidfile' is stale, deletes it. */
710 check_already_running(void)
712 long int pid = read_pidfile__(pidfile, true);
714 VLOG_FATAL("%s: already running as pid %ld, aborting", pidfile, pid);
715 } else if (pid < 0) {
716 VLOG_FATAL("%s: pidfile check failed (%s), aborting",
717 pidfile, strerror(-pid));