2 * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks.
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? */
43 /* --pidfile: Name of pidfile (null if none). */
46 /* Device and inode of pidfile, so we can avoid reopening it. */
47 static dev_t pidfile_dev;
48 static ino_t pidfile_ino;
50 /* --overwrite-pidfile: Create pidfile even if one already exists and is
52 static bool overwrite_pidfile;
54 /* --no-chdir: Should we chdir to "/"? */
55 static bool chdir_ = true;
57 /* File descriptor used by daemonize_start() and daemonize_complete(). */
58 static int daemonize_fd = -1;
60 /* --monitor: Should a supervisory process monitor the daemon and restart it if
61 * it dies due to an error signal? */
64 static void check_already_running(void);
65 static int lock_pidfile(FILE *, int command);
67 /* Returns the file name that would be used for a pidfile if 'name' were
68 * provided to set_pidfile(). The caller must free the returned string. */
70 make_pidfile_name(const char *name)
73 ? xasprintf("%s/%s.pid", ovs_rundir(), program_name)
74 : abs_file_name(ovs_rundir(), name));
77 /* Sets up a following call to daemonize() to create a pidfile named 'name'.
78 * If 'name' begins with '/', then it is treated as an absolute path.
79 * Otherwise, it is taken relative to RUNDIR, which is $(prefix)/var/run by
82 * If 'name' is null, then program_name followed by ".pid" is used. */
84 set_pidfile(const char *name)
87 pidfile = make_pidfile_name(name);
90 /* Returns an absolute path to the configured pidfile, or a null pointer if no
91 * pidfile is configured. The caller must not modify or free the returned
99 /* Sets that we do not chdir to "/". */
106 /* Will we chdir to "/" as part of daemonizing? */
108 is_chdir_enabled(void)
113 /* Normally, daemonize() or damonize_start() will terminate the program with a
114 * message if a locked pidfile already exists. If this function is called, an
115 * existing pidfile will be replaced, with a warning. */
117 ignore_existing_pidfile(void)
119 overwrite_pidfile = true;
122 /* Sets up a following call to daemonize() to detach from the foreground
123 * session, running this process in the background. */
130 /* Will daemonize() really detach? */
137 /* Sets up a following call to daemonize() to fork a supervisory process to
138 * monitor the daemon and restart it if it dies due to an error signal. */
140 daemon_set_monitor(void)
145 /* If a pidfile has been configured, creates it and stores the running
146 * process's pid in it. Ensures that the pidfile will be deleted when the
151 long int pid = getpid();
157 /* Create a temporary pidfile. */
158 tmpfile = xasprintf("%s.tmp%ld", pidfile, pid);
159 fatal_signal_add_file_to_unlink(tmpfile);
160 file = fopen(tmpfile, "w+");
162 VLOG_FATAL("%s: create failed (%s)", tmpfile, strerror(errno));
165 if (fstat(fileno(file), &s) == -1) {
166 VLOG_FATAL("%s: fstat failed (%s)", tmpfile, strerror(errno));
169 fprintf(file, "%ld\n", pid);
170 if (fflush(file) == EOF) {
171 VLOG_FATAL("%s: write failed (%s)", tmpfile, strerror(errno));
174 error = lock_pidfile(file, F_SETLK);
176 VLOG_FATAL("%s: fcntl(F_SETLK) failed (%s)", tmpfile, strerror(error));
179 /* Rename or link it to the correct name. */
180 if (overwrite_pidfile) {
181 if (rename(tmpfile, pidfile) < 0) {
182 VLOG_FATAL("failed to rename \"%s\" to \"%s\" (%s)",
183 tmpfile, pidfile, strerror(errno));
187 error = link(tmpfile, pidfile) == -1 ? errno : 0;
188 if (error == EEXIST) {
189 check_already_running();
191 } while (error == EINTR || error == EEXIST);
193 VLOG_FATAL("failed to link \"%s\" as \"%s\" (%s)",
194 tmpfile, pidfile, strerror(error));
198 /* Ensure that the pidfile will get deleted on exit. */
199 fatal_signal_add_file_to_unlink(pidfile);
201 /* Delete the temporary pidfile if it still exists. */
202 if (!overwrite_pidfile) {
203 error = fatal_signal_unlink_file_now(tmpfile);
205 VLOG_FATAL("%s: unlink failed (%s)", tmpfile, strerror(error));
211 * We don't close 'file' because its file descriptor must remain open to
213 pidfile_dev = s.st_dev;
214 pidfile_ino = s.st_ino;
220 /* If configured with set_pidfile() or set_detach(), creates the pid file and
221 * detaches from the foreground session. */
226 daemonize_complete();
230 fork_and_wait_for_startup(int *fdp)
239 /* Running in parent process. */
245 if (read_fully(fds[0], &c, 1, &bytes_read) != 0) {
250 retval = waitpid(pid, &status, 0);
251 } while (retval == -1 && errno == EINTR);
254 if (WIFEXITED(status) && WEXITSTATUS(status)) {
255 /* Child exited with an error. Convey the same error
256 * to our parent process as a courtesy. */
257 exit(WEXITSTATUS(status));
259 char *status_msg = process_status_msg(status);
260 VLOG_FATAL("fork child died before signaling startup (%s)",
263 } else if (retval < 0) {
264 VLOG_FATAL("waitpid failed (%s)", strerror(errno));
272 /* Running in child process. */
278 VLOG_FATAL("fork failed (%s)", strerror(errno));
285 fork_notify_startup(int fd)
288 size_t bytes_written;
291 error = write_fully(fd, "", 1, &bytes_written);
293 VLOG_FATAL("pipe write failed (%s)", strerror(error));
301 should_restart(int status)
303 if (WIFSIGNALED(status)) {
304 static const int error_signals[] = {
305 SIGABRT, SIGALRM, SIGBUS, SIGFPE, SIGILL, SIGPIPE, SIGSEGV,
311 for (i = 0; i < ARRAY_SIZE(error_signals); i++) {
312 if (error_signals[i] == WTERMSIG(status)) {
321 monitor_daemon(pid_t daemon_pid)
323 /* XXX Should log daemon's stderr output at startup time. */
324 const char *saved_program_name;
329 saved_program_name = program_name;
330 program_name = xasprintf("monitor(%s)", program_name);
331 status_msg = xstrdup("healthy");
332 last_restart = TIME_MIN;
338 proctitle_set("%s: monitoring pid %lu (%s)",
339 saved_program_name, (unsigned long int) daemon_pid,
343 retval = waitpid(daemon_pid, &status, 0);
344 } while (retval == -1 && errno == EINTR);
347 VLOG_FATAL("waitpid failed (%s)", strerror(errno));
348 } else if (retval == daemon_pid) {
349 char *s = process_status_msg(status);
350 if (should_restart(status)) {
352 status_msg = xasprintf("%d crashes: pid %lu died, %s",
354 (unsigned long int) daemon_pid, s);
357 if (WCOREDUMP(status)) {
358 /* Disable further core dumps to save disk space. */
363 if (setrlimit(RLIMIT_CORE, &r) == -1) {
364 VLOG_WARN("failed to disable core dumps: %s",
369 /* Throttle restarts to no more than once every 10 seconds. */
370 if (time(NULL) < last_restart + 10) {
371 VLOG_WARN("%s, waiting until 10 seconds since last "
372 "restart", status_msg);
374 time_t now = time(NULL);
375 time_t wakeup = last_restart + 10;
382 last_restart = time(NULL);
384 VLOG_ERR("%s, restarting", status_msg);
385 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
390 VLOG_INFO("pid %lu died, %s, exiting",
391 (unsigned long int) daemon_pid, s);
399 /* Running in new daemon process. */
401 free((char *) program_name);
402 program_name = saved_program_name;
405 /* Close stdin, stdout, stderr. If we're started from e.g. an SSH session,
406 * then this keeps us from holding that session open artificially. */
408 close_standard_fds(void)
410 int null_fd = get_null_fd();
412 dup2(null_fd, STDIN_FILENO);
413 dup2(null_fd, STDOUT_FILENO);
414 dup2(null_fd, STDERR_FILENO);
417 /* Disable logging to stderr to avoid wasting CPU time. */
418 vlog_set_levels(NULL, VLF_CONSOLE, VLL_OFF);
421 /* If daemonization is configured, then starts daemonization, by forking and
422 * returning in the child process. The parent process hangs around until the
423 * child lets it know either that it completed startup successfully (by calling
424 * daemon_complete()) or that it failed to start up (by exiting with a nonzero
427 daemonize_start(void)
432 if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
433 /* Running in parent process. */
436 /* Running in daemon or monitor process. */
440 int saved_daemonize_fd = daemonize_fd;
443 daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
444 if (daemon_pid > 0) {
445 /* Running in monitor process. */
446 fork_notify_startup(saved_daemonize_fd);
447 close_standard_fds();
448 monitor_daemon(daemon_pid);
450 /* Running in daemon process. */
457 /* Make sure that the unixctl commands for vlog get registered in a
458 * daemon, even before the first log message. */
462 /* If daemonization is configured, then this function notifies the parent
463 * process that the child process has completed startup successfully.
465 * Calling this function more than once has no additional effect. */
467 daemonize_complete(void)
469 fork_notify_startup(daemonize_fd);
477 close_standard_fds();
486 "\nDaemon options:\n"
487 " --detach run in background as daemon\n"
488 " --no-chdir do not chdir to '/'\n"
489 " --pidfile[=FILE] create pidfile (default: %s/%s.pid)\n"
490 " --overwrite-pidfile with --pidfile, start even if already "
492 ovs_rundir(), program_name);
496 lock_pidfile__(FILE *file, int command, struct flock *lck)
500 lck->l_type = F_WRLCK;
501 lck->l_whence = SEEK_SET;
507 error = fcntl(fileno(file), command, lck) == -1 ? errno : 0;
508 } while (error == EINTR);
513 lock_pidfile(FILE *file, int command)
517 return lock_pidfile__(file, command, &lck);
521 read_pidfile__(const char *pidfile, bool delete_if_stale)
529 if ((pidfile_ino || pidfile_dev)
530 && !stat(pidfile, &s)
531 && s.st_ino == pidfile_ino && s.st_dev == pidfile_dev) {
532 /* It's our own pidfile. We can't afford to open it, because closing
533 * *any* fd for a file that a process has locked also releases all the
534 * locks on that file.
536 * Fortunately, we know the associated pid anyhow: */
540 file = fopen(pidfile, "r+");
542 if (errno == ENOENT && delete_if_stale) {
546 VLOG_WARN("%s: open: %s", pidfile, strerror(error));
550 error = lock_pidfile__(file, F_GETLK, &lck);
552 VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
555 if (lck.l_type == F_UNLCK) {
556 /* pidfile exists but it isn't locked by anyone. We need to delete it
557 * so that a new pidfile can go in its place. But just calling
558 * unlink(pidfile) makes a nasty race: what if someone else unlinks it
559 * before we do and then replaces it by a valid pidfile? We'd unlink
560 * their valid pidfile. We do a little dance to avoid the race, by
561 * locking the invalid pidfile. Only one process can have the invalid
562 * pidfile locked, and only that process has the right to unlink it. */
563 if (!delete_if_stale) {
565 VLOG_DBG("%s: pid file is stale", pidfile);
570 error = lock_pidfile(file, F_SETLK);
572 /* We lost a race with someone else doing the same thing. */
573 VLOG_WARN("%s: lost race to lock pidfile", pidfile);
577 /* Is the file we have locked still named 'pidfile'? */
578 if (stat(pidfile, &s) || fstat(fileno(file), &s2)
579 || s.st_ino != s2.st_ino || s.st_dev != s2.st_dev) {
580 /* No. We lost a race with someone else who got the lock before
581 * us, deleted the pidfile, and closed it (releasing the lock). */
583 VLOG_WARN("%s: lost race to delete pidfile", pidfile);
587 /* We won the right to delete the stale pidfile. */
588 if (unlink(pidfile)) {
590 VLOG_WARN("%s: failed to delete stale pidfile (%s)",
591 pidfile, strerror(error));
594 VLOG_DBG("%s: deleted stale pidfile", pidfile);
599 if (!fgets(line, sizeof line, file)) {
602 VLOG_WARN("%s: read: %s", pidfile, strerror(error));
605 VLOG_WARN("%s: read: unexpected end of file", pidfile);
610 if (lck.l_pid != strtoul(line, NULL, 10)) {
611 /* The process that has the pidfile locked is not the process that
612 * created it. It must be stale, with the process that has it locked
613 * preparing to delete it. */
615 VLOG_WARN("%s: stale pidfile for pid %s being deleted by pid %ld",
616 pidfile, line, (long int) lck.l_pid);
630 /* Opens and reads a PID from 'pidfile'. Returns the positive PID if
631 * successful, otherwise a negative errno value. */
633 read_pidfile(const char *pidfile)
635 return read_pidfile__(pidfile, false);
638 /* Checks whether a process with the given 'pidfile' is already running and,
639 * if so, aborts. If 'pidfile' is stale, deletes it. */
641 check_already_running(void)
643 long int pid = read_pidfile__(pidfile, true);
645 VLOG_FATAL("%s: already running as pid %ld, aborting", pidfile, pid);
646 } else if (pid < 0) {
647 VLOG_FATAL("%s: pidfile check failed (%s), aborting",
648 pidfile, strerror(-pid));