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? */
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 assert(fd == STDIN_FILENO || fd == STDOUT_FILENO || fd == STDERR_FILENO);
164 /* If a pidfile has been configured, creates it and stores the running
165 * process's pid in it. Ensures that the pidfile will be deleted when the
170 long int pid = getpid();
176 /* Create a temporary pidfile. */
177 tmpfile = xasprintf("%s.tmp%ld", pidfile, pid);
178 fatal_signal_add_file_to_unlink(tmpfile);
179 file = fopen(tmpfile, "w+");
181 VLOG_FATAL("%s: create failed (%s)", tmpfile, strerror(errno));
184 if (fstat(fileno(file), &s) == -1) {
185 VLOG_FATAL("%s: fstat failed (%s)", tmpfile, strerror(errno));
188 fprintf(file, "%ld\n", pid);
189 if (fflush(file) == EOF) {
190 VLOG_FATAL("%s: write failed (%s)", tmpfile, strerror(errno));
193 error = lock_pidfile(file, F_SETLK);
195 VLOG_FATAL("%s: fcntl(F_SETLK) failed (%s)", tmpfile, strerror(error));
198 /* Rename or link it to the correct name. */
199 if (overwrite_pidfile) {
200 if (rename(tmpfile, pidfile) < 0) {
201 VLOG_FATAL("failed to rename \"%s\" to \"%s\" (%s)",
202 tmpfile, pidfile, strerror(errno));
206 error = link(tmpfile, pidfile) == -1 ? errno : 0;
207 if (error == EEXIST) {
208 check_already_running();
210 } while (error == EINTR || error == EEXIST);
212 VLOG_FATAL("failed to link \"%s\" as \"%s\" (%s)",
213 tmpfile, pidfile, strerror(error));
217 /* Ensure that the pidfile will get deleted on exit. */
218 fatal_signal_add_file_to_unlink(pidfile);
220 /* Delete the temporary pidfile if it still exists. */
221 if (!overwrite_pidfile) {
222 error = fatal_signal_unlink_file_now(tmpfile);
224 VLOG_FATAL("%s: unlink failed (%s)", tmpfile, strerror(error));
230 * We don't close 'file' because its file descriptor must remain open to
232 pidfile_dev = s.st_dev;
233 pidfile_ino = s.st_ino;
239 /* If configured with set_pidfile() or set_detach(), creates the pid file and
240 * detaches from the foreground session. */
245 daemonize_complete();
250 * - In the parent, waits for the child to signal that it has completed its
251 * startup sequence. Then stores -1 in '*fdp' and returns the child's pid.
253 * - In the child, stores a fd in '*fdp' and returns 0. The caller should
254 * pass the fd to fork_notify_startup() after it finishes its startup
257 * If something goes wrong with the fork, logs a critical error and aborts the
260 fork_and_wait_for_startup(int *fdp)
269 /* Running in parent process. */
275 if (read_fully(fds[0], &c, 1, &bytes_read) != 0) {
280 retval = waitpid(pid, &status, 0);
281 } while (retval == -1 && errno == EINTR);
284 if (WIFEXITED(status) && WEXITSTATUS(status)) {
285 /* Child exited with an error. Convey the same error
286 * to our parent process as a courtesy. */
287 exit(WEXITSTATUS(status));
289 char *status_msg = process_status_msg(status);
290 VLOG_FATAL("fork child died before signaling startup (%s)",
293 } else if (retval < 0) {
294 VLOG_FATAL("waitpid failed (%s)", strerror(errno));
302 /* Running in child process. */
308 VLOG_FATAL("fork failed (%s)", strerror(errno));
315 fork_notify_startup(int fd)
318 size_t bytes_written;
321 error = write_fully(fd, "", 1, &bytes_written);
323 VLOG_FATAL("pipe write failed (%s)", strerror(error));
331 should_restart(int status)
333 if (WIFSIGNALED(status)) {
334 static const int error_signals[] = {
335 SIGABRT, SIGALRM, SIGBUS, SIGFPE, SIGILL, SIGPIPE, SIGSEGV,
341 for (i = 0; i < ARRAY_SIZE(error_signals); i++) {
342 if (error_signals[i] == WTERMSIG(status)) {
351 monitor_daemon(pid_t daemon_pid)
353 /* XXX Should log daemon's stderr output at startup time. */
354 const char *saved_program_name;
359 saved_program_name = program_name;
360 program_name = xasprintf("monitor(%s)", program_name);
361 status_msg = xstrdup("healthy");
362 last_restart = TIME_MIN;
368 proctitle_set("%s: monitoring pid %lu (%s)",
369 saved_program_name, (unsigned long int) daemon_pid,
373 retval = waitpid(daemon_pid, &status, 0);
374 } while (retval == -1 && errno == EINTR);
377 VLOG_FATAL("waitpid failed (%s)", strerror(errno));
378 } else if (retval == daemon_pid) {
379 char *s = process_status_msg(status);
380 if (should_restart(status)) {
382 status_msg = xasprintf("%d crashes: pid %lu died, %s",
384 (unsigned long int) daemon_pid, s);
387 if (WCOREDUMP(status)) {
388 /* Disable further core dumps to save disk space. */
393 if (setrlimit(RLIMIT_CORE, &r) == -1) {
394 VLOG_WARN("failed to disable core dumps: %s",
399 /* Throttle restarts to no more than once every 10 seconds. */
400 if (time(NULL) < last_restart + 10) {
401 VLOG_WARN("%s, waiting until 10 seconds since last "
402 "restart", status_msg);
404 time_t now = time(NULL);
405 time_t wakeup = last_restart + 10;
412 last_restart = time(NULL);
414 VLOG_ERR("%s, restarting", status_msg);
415 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
420 VLOG_INFO("pid %lu died, %s, exiting",
421 (unsigned long int) daemon_pid, s);
429 /* Running in new daemon process. */
431 free((char *) program_name);
432 program_name = saved_program_name;
435 /* Close standard file descriptors (except any that the client has requested we
436 * leave open by calling daemon_save_fd()). If we're started from e.g. an SSH
437 * session, then this keeps us from holding that session open artificially. */
439 close_standard_fds(void)
441 int null_fd = get_null_fd();
445 for (fd = 0; fd < 3; fd++) {
452 /* Disable logging to stderr to avoid wasting CPU time. */
453 vlog_set_levels(NULL, VLF_CONSOLE, VLL_OFF);
456 /* If daemonization is configured, then starts daemonization, by forking and
457 * returning in the child process. The parent process hangs around until the
458 * child lets it know either that it completed startup successfully (by calling
459 * daemon_complete()) or that it failed to start up (by exiting with a nonzero
462 daemonize_start(void)
467 if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
468 /* Running in parent process. */
471 /* Running in daemon or monitor process. */
475 int saved_daemonize_fd = daemonize_fd;
478 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
479 if (daemon_pid > 0) {
480 /* Running in monitor process. */
481 fork_notify_startup(saved_daemonize_fd);
482 close_standard_fds();
483 monitor_daemon(daemon_pid);
485 /* Running in daemon process. */
492 /* Make sure that the unixctl commands for vlog get registered in a
493 * daemon, even before the first log message. */
497 /* If daemonization is configured, then this function notifies the parent
498 * process that the child process has completed startup successfully.
500 * Calling this function more than once has no additional effect. */
502 daemonize_complete(void)
504 fork_notify_startup(daemonize_fd);
512 close_standard_fds();
521 "\nDaemon options:\n"
522 " --detach run in background as daemon\n"
523 " --no-chdir do not chdir to '/'\n"
524 " --pidfile[=FILE] create pidfile (default: %s/%s.pid)\n"
525 " --overwrite-pidfile with --pidfile, start even if already "
527 ovs_rundir(), program_name);
531 lock_pidfile__(FILE *file, int command, struct flock *lck)
535 lck->l_type = F_WRLCK;
536 lck->l_whence = SEEK_SET;
542 error = fcntl(fileno(file), command, lck) == -1 ? errno : 0;
543 } while (error == EINTR);
548 lock_pidfile(FILE *file, int command)
552 return lock_pidfile__(file, command, &lck);
556 read_pidfile__(const char *pidfile, bool delete_if_stale)
564 if ((pidfile_ino || pidfile_dev)
565 && !stat(pidfile, &s)
566 && s.st_ino == pidfile_ino && s.st_dev == pidfile_dev) {
567 /* It's our own pidfile. We can't afford to open it, because closing
568 * *any* fd for a file that a process has locked also releases all the
569 * locks on that file.
571 * Fortunately, we know the associated pid anyhow: */
575 file = fopen(pidfile, "r+");
577 if (errno == ENOENT && delete_if_stale) {
581 VLOG_WARN("%s: open: %s", pidfile, strerror(error));
585 error = lock_pidfile__(file, F_GETLK, &lck);
587 VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
590 if (lck.l_type == F_UNLCK) {
591 /* pidfile exists but it isn't locked by anyone. We need to delete it
592 * so that a new pidfile can go in its place. But just calling
593 * unlink(pidfile) makes a nasty race: what if someone else unlinks it
594 * before we do and then replaces it by a valid pidfile? We'd unlink
595 * their valid pidfile. We do a little dance to avoid the race, by
596 * locking the invalid pidfile. Only one process can have the invalid
597 * pidfile locked, and only that process has the right to unlink it. */
598 if (!delete_if_stale) {
600 VLOG_DBG("%s: pid file is stale", pidfile);
605 error = lock_pidfile(file, F_SETLK);
607 /* We lost a race with someone else doing the same thing. */
608 VLOG_WARN("%s: lost race to lock pidfile", pidfile);
612 /* Is the file we have locked still named 'pidfile'? */
613 if (stat(pidfile, &s) || fstat(fileno(file), &s2)
614 || s.st_ino != s2.st_ino || s.st_dev != s2.st_dev) {
615 /* No. We lost a race with someone else who got the lock before
616 * us, deleted the pidfile, and closed it (releasing the lock). */
618 VLOG_WARN("%s: lost race to delete pidfile", pidfile);
622 /* We won the right to delete the stale pidfile. */
623 if (unlink(pidfile)) {
625 VLOG_WARN("%s: failed to delete stale pidfile (%s)",
626 pidfile, strerror(error));
629 VLOG_DBG("%s: deleted stale pidfile", pidfile);
634 if (!fgets(line, sizeof line, file)) {
637 VLOG_WARN("%s: read: %s", pidfile, strerror(error));
640 VLOG_WARN("%s: read: unexpected end of file", pidfile);
645 if (lck.l_pid != strtoul(line, NULL, 10)) {
646 /* The process that has the pidfile locked is not the process that
647 * created it. It must be stale, with the process that has it locked
648 * preparing to delete it. */
650 VLOG_WARN("%s: stale pidfile for pid %s being deleted by pid %ld",
651 pidfile, line, (long int) lck.l_pid);
665 /* Opens and reads a PID from 'pidfile'. Returns the positive PID if
666 * successful, otherwise a negative errno value. */
668 read_pidfile(const char *pidfile)
670 return read_pidfile__(pidfile, false);
673 /* Checks whether a process with the given 'pidfile' is already running and,
674 * if so, aborts. If 'pidfile' is stale, deletes it. */
676 check_already_running(void)
678 long int pid = read_pidfile__(pidfile, true);
680 VLOG_FATAL("%s: already running as pid %ld, aborting", pidfile, pid);
681 } else if (pid < 0) {
682 VLOG_FATAL("%s: pidfile check failed (%s), aborting",
683 pidfile, strerror(-pid));