]>
git.proxmox.com Git - mirror_ovs.git/blob - lib/daemon.c
2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013 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"
32 #include "ovs-thread.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
)
92 assert_single_threaded();
94 pidfile
= make_pidfile_name(name
);
97 /* Returns an absolute path to the configured pidfile, or a null pointer if no
98 * pidfile is configured. The caller must not modify or free the returned
106 /* Sets that we do not chdir to "/". */
113 /* Will we chdir to "/" as part of daemonizing? */
115 is_chdir_enabled(void)
120 /* Normally, daemonize() or damonize_start() will terminate the program with a
121 * message if a locked pidfile already exists. If this function is called, an
122 * existing pidfile will be replaced, with a warning. */
124 ignore_existing_pidfile(void)
126 overwrite_pidfile
= true;
129 /* Sets up a following call to daemonize() to detach from the foreground
130 * session, running this process in the background. */
137 /* Will daemonize() really detach? */
144 /* Sets up a following call to daemonize() to fork a supervisory process to
145 * monitor the daemon and restart it if it dies due to an error signal. */
147 daemon_set_monitor(void)
152 /* A daemon doesn't normally have any use for the file descriptors for stdin,
153 * stdout, and stderr after it detaches. To keep these file descriptors from
154 * e.g. holding an SSH session open, by default detaching replaces each of
155 * these file descriptors by /dev/null. But a few daemons expect the user to
156 * redirect stdout or stderr to a file, in which case it is desirable to keep
157 * these file descriptors. This function, therefore, disables replacing 'fd'
158 * by /dev/null when the daemon detaches. */
160 daemon_save_fd(int fd
)
162 ovs_assert(fd
== STDIN_FILENO
||
163 fd
== STDOUT_FILENO
||
164 fd
== STDERR_FILENO
);
168 /* Unregisters pidfile from being unlinked when the program terminates via
169 * exit() or a fatal signal. */
171 remove_pidfile_from_unlink(void)
174 fatal_signal_remove_file_to_unlink(pidfile
);
178 /* Registers pidfile to be unlinked when the program terminates via exit() or a
181 add_pidfile_to_unlink(void)
184 fatal_signal_add_file_to_unlink(pidfile
);
188 /* If a pidfile has been configured, creates it and stores the running
189 * process's pid in it. Ensures that the pidfile will be deleted when the
194 long int pid
= getpid();
200 /* Create a temporary pidfile. */
201 if (overwrite_pidfile
) {
202 tmpfile
= xasprintf("%s.tmp%ld", pidfile
, pid
);
203 fatal_signal_add_file_to_unlink(tmpfile
);
205 /* Everyone shares the same file which will be treated as a lock. To
206 * avoid some uncomfortable race conditions, we can't set up the fatal
207 * signal unlink until we've acquired it. */
208 tmpfile
= xasprintf("%s.tmp", pidfile
);
211 file
= fopen(tmpfile
, "a+");
213 VLOG_FATAL("%s: create failed (%s)", tmpfile
, ovs_strerror(errno
));
216 error
= lock_pidfile(file
, F_SETLK
);
218 /* Looks like we failed to acquire the lock. Note that, if we failed
219 * for some other reason (and '!overwrite_pidfile'), we will have
220 * left 'tmpfile' as garbage in the file system. */
221 VLOG_FATAL("%s: fcntl(F_SETLK) failed (%s)", tmpfile
,
222 ovs_strerror(error
));
225 if (!overwrite_pidfile
) {
226 /* We acquired the lock. Make sure to clean up on exit, and verify
227 * that we're allowed to create the actual pidfile. */
228 fatal_signal_add_file_to_unlink(tmpfile
);
229 check_already_running();
232 if (fstat(fileno(file
), &s
) == -1) {
233 VLOG_FATAL("%s: fstat failed (%s)", tmpfile
, ovs_strerror(errno
));
236 if (ftruncate(fileno(file
), 0) == -1) {
237 VLOG_FATAL("%s: truncate failed (%s)", tmpfile
, ovs_strerror(errno
));
240 fprintf(file
, "%ld\n", pid
);
241 if (fflush(file
) == EOF
) {
242 VLOG_FATAL("%s: write failed (%s)", tmpfile
, ovs_strerror(errno
));
245 error
= rename(tmpfile
, pidfile
);
247 /* Due to a race, 'tmpfile' may be owned by a different process, so we
248 * shouldn't delete it on exit. */
249 fatal_signal_remove_file_to_unlink(tmpfile
);
252 VLOG_FATAL("failed to rename \"%s\" to \"%s\" (%s)",
253 tmpfile
, pidfile
, ovs_strerror(errno
));
256 /* Ensure that the pidfile will get deleted on exit. */
257 fatal_signal_add_file_to_unlink(pidfile
);
261 * We don't close 'file' because its file descriptor must remain open to
263 pidfile_dev
= s
.st_dev
;
264 pidfile_ino
= s
.st_ino
;
268 /* If configured with set_pidfile() or set_detach(), creates the pid file and
269 * detaches from the foreground session. */
274 daemonize_complete();
277 /* Calls fork() and on success returns its return value. On failure, logs an
278 * error and exits unsuccessfully.
280 * Post-fork, but before returning, this function calls a few other functions
281 * that are generally useful if the child isn't planning to exec a new
284 fork_and_clean_up(void)
288 /* Running in parent process. */
291 /* Running in child process. */
300 * - In the parent, waits for the child to signal that it has completed its
301 * startup sequence. Then stores -1 in '*fdp' and returns the child's pid.
303 * - In the child, stores a fd in '*fdp' and returns 0. The caller should
304 * pass the fd to fork_notify_startup() after it finishes its startup
307 * If something goes wrong with the fork, logs a critical error and aborts the
310 fork_and_wait_for_startup(int *fdp
)
317 pid
= fork_and_clean_up();
319 /* Running in parent process. */
324 if (read_fully(fds
[0], &c
, 1, &bytes_read
) != 0) {
329 retval
= waitpid(pid
, &status
, 0);
330 } while (retval
== -1 && errno
== EINTR
);
333 if (WIFEXITED(status
) && WEXITSTATUS(status
)) {
334 /* Child exited with an error. Convey the same error
335 * to our parent process as a courtesy. */
336 exit(WEXITSTATUS(status
));
338 char *status_msg
= process_status_msg(status
);
339 VLOG_FATAL("fork child died before signaling startup (%s)",
342 } else if (retval
< 0) {
343 VLOG_FATAL("waitpid failed (%s)", ovs_strerror(errno
));
351 /* Running in child process. */
360 fork_notify_startup(int fd
)
363 size_t bytes_written
;
366 error
= write_fully(fd
, "", 1, &bytes_written
);
368 VLOG_FATAL("pipe write failed (%s)", ovs_strerror(error
));
376 should_restart(int status
)
378 if (WIFSIGNALED(status
)) {
379 static const int error_signals
[] = {
380 SIGABRT
, SIGALRM
, SIGBUS
, SIGFPE
, SIGILL
, SIGPIPE
, SIGSEGV
,
386 for (i
= 0; i
< ARRAY_SIZE(error_signals
); i
++) {
387 if (error_signals
[i
] == WTERMSIG(status
)) {
396 monitor_daemon(pid_t daemon_pid
)
398 /* XXX Should log daemon's stderr output at startup time. */
403 set_subprogram_name("monitor");
404 status_msg
= xstrdup("healthy");
405 last_restart
= TIME_MIN
;
411 proctitle_set("monitoring pid %lu (%s)",
412 (unsigned long int) daemon_pid
, status_msg
);
415 retval
= waitpid(daemon_pid
, &status
, 0);
416 } while (retval
== -1 && errno
== EINTR
);
419 VLOG_FATAL("waitpid failed (%s)", ovs_strerror(errno
));
420 } else if (retval
== daemon_pid
) {
421 char *s
= process_status_msg(status
);
422 if (should_restart(status
)) {
424 status_msg
= xasprintf("%d crashes: pid %lu died, %s",
426 (unsigned long int) daemon_pid
, s
);
429 if (WCOREDUMP(status
)) {
430 /* Disable further core dumps to save disk space. */
435 if (setrlimit(RLIMIT_CORE
, &r
) == -1) {
436 VLOG_WARN("failed to disable core dumps: %s",
437 ovs_strerror(errno
));
441 /* Throttle restarts to no more than once every 10 seconds. */
442 if (time(NULL
) < last_restart
+ 10) {
443 VLOG_WARN("%s, waiting until 10 seconds since last "
444 "restart", status_msg
);
446 time_t now
= time(NULL
);
447 time_t wakeup
= last_restart
+ 10;
454 last_restart
= time(NULL
);
456 VLOG_ERR("%s, restarting", status_msg
);
457 daemon_pid
= fork_and_wait_for_startup(&daemonize_fd
);
462 VLOG_INFO("pid %lu died, %s, exiting",
463 (unsigned long int) daemon_pid
, s
);
471 /* Running in new daemon process. */
473 set_subprogram_name("");
476 /* Close standard file descriptors (except any that the client has requested we
477 * leave open by calling daemon_save_fd()). If we're started from e.g. an SSH
478 * session, then this keeps us from holding that session open artificially. */
480 close_standard_fds(void)
482 int null_fd
= get_null_fd();
486 for (fd
= 0; fd
< 3; fd
++) {
493 /* Disable logging to stderr to avoid wasting CPU time. */
494 vlog_set_levels(NULL
, VLF_CONSOLE
, VLL_OFF
);
497 /* If daemonization is configured, then starts daemonization, by forking and
498 * returning in the child process. The parent process hangs around until the
499 * child lets it know either that it completed startup successfully (by calling
500 * daemon_complete()) or that it failed to start up (by exiting with a nonzero
503 daemonize_start(void)
505 assert_single_threaded();
509 if (fork_and_wait_for_startup(&daemonize_fd
) > 0) {
510 /* Running in parent process. */
514 /* Running in daemon or monitor process. */
519 int saved_daemonize_fd
= daemonize_fd
;
522 daemon_pid
= fork_and_wait_for_startup(&daemonize_fd
);
523 if (daemon_pid
> 0) {
524 /* Running in monitor process. */
525 fork_notify_startup(saved_daemonize_fd
);
526 close_standard_fds();
527 monitor_daemon(daemon_pid
);
529 /* Running in daemon process. */
532 forbid_forking("running in daemon process");
538 /* Make sure that the unixctl commands for vlog get registered in a
539 * daemon, even before the first log message. */
543 /* If daemonization is configured, then this function notifies the parent
544 * process that the child process has completed startup successfully. It also
545 * call daemonize_post_detach().
547 * Calling this function more than once has no additional effect. */
549 daemonize_complete(void)
559 fork_notify_startup(daemonize_fd
);
561 daemonize_post_detach();
565 /* If daemonization is configured, then this function does traditional Unix
566 * daemonization behavior: join a new session, chdir to the root (if not
567 * disabled), and close the standard file descriptors.
569 * It only makes sense to call this function as part of an implementation of a
570 * special daemon subprocess. A normal daemon should just call
571 * daemonize_complete(). */
573 daemonize_post_detach(void)
579 close_standard_fds();
587 "\nDaemon options:\n"
588 " --detach run in background as daemon\n"
589 " --no-chdir do not chdir to '/'\n"
590 " --pidfile[=FILE] create pidfile (default: %s/%s.pid)\n"
591 " --overwrite-pidfile with --pidfile, start even if already "
593 ovs_rundir(), program_name
);
597 lock_pidfile__(FILE *file
, int command
, struct flock
*lck
)
601 lck
->l_type
= F_WRLCK
;
602 lck
->l_whence
= SEEK_SET
;
608 error
= fcntl(fileno(file
), command
, lck
) == -1 ? errno
: 0;
609 } while (error
== EINTR
);
614 lock_pidfile(FILE *file
, int command
)
618 return lock_pidfile__(file
, command
, &lck
);
622 read_pidfile__(const char *pidfile
, bool delete_if_stale
)
630 if ((pidfile_ino
|| pidfile_dev
)
631 && !stat(pidfile
, &s
)
632 && s
.st_ino
== pidfile_ino
&& s
.st_dev
== pidfile_dev
) {
633 /* It's our own pidfile. We can't afford to open it, because closing
634 * *any* fd for a file that a process has locked also releases all the
635 * locks on that file.
637 * Fortunately, we know the associated pid anyhow: */
641 file
= fopen(pidfile
, "r+");
643 if (errno
== ENOENT
&& delete_if_stale
) {
647 VLOG_WARN("%s: open: %s", pidfile
, ovs_strerror(error
));
651 error
= lock_pidfile__(file
, F_GETLK
, &lck
);
653 VLOG_WARN("%s: fcntl: %s", pidfile
, ovs_strerror(error
));
656 if (lck
.l_type
== F_UNLCK
) {
657 /* pidfile exists but it isn't locked by anyone. We need to delete it
658 * so that a new pidfile can go in its place. But just calling
659 * unlink(pidfile) makes a nasty race: what if someone else unlinks it
660 * before we do and then replaces it by a valid pidfile? We'd unlink
661 * their valid pidfile. We do a little dance to avoid the race, by
662 * locking the invalid pidfile. Only one process can have the invalid
663 * pidfile locked, and only that process has the right to unlink it. */
664 if (!delete_if_stale
) {
666 VLOG_DBG("%s: pid file is stale", pidfile
);
671 error
= lock_pidfile(file
, F_SETLK
);
673 /* We lost a race with someone else doing the same thing. */
674 VLOG_WARN("%s: lost race to lock pidfile", pidfile
);
678 /* Is the file we have locked still named 'pidfile'? */
679 if (stat(pidfile
, &s
) || fstat(fileno(file
), &s2
)
680 || s
.st_ino
!= s2
.st_ino
|| s
.st_dev
!= s2
.st_dev
) {
681 /* No. We lost a race with someone else who got the lock before
682 * us, deleted the pidfile, and closed it (releasing the lock). */
684 VLOG_WARN("%s: lost race to delete pidfile", pidfile
);
688 /* We won the right to delete the stale pidfile. */
689 if (unlink(pidfile
)) {
691 VLOG_WARN("%s: failed to delete stale pidfile (%s)",
692 pidfile
, ovs_strerror(error
));
695 VLOG_DBG("%s: deleted stale pidfile", pidfile
);
700 if (!fgets(line
, sizeof line
, file
)) {
703 VLOG_WARN("%s: read: %s", pidfile
, ovs_strerror(error
));
706 VLOG_WARN("%s: read: unexpected end of file", pidfile
);
711 if (lck
.l_pid
!= strtoul(line
, NULL
, 10)) {
712 /* The process that has the pidfile locked is not the process that
713 * created it. It must be stale, with the process that has it locked
714 * preparing to delete it. */
716 VLOG_WARN("%s: stale pidfile for pid %s being deleted by pid %ld",
717 pidfile
, line
, (long int) lck
.l_pid
);
731 /* Opens and reads a PID from 'pidfile'. Returns the positive PID if
732 * successful, otherwise a negative errno value. */
734 read_pidfile(const char *pidfile
)
736 return read_pidfile__(pidfile
, false);
739 /* Checks whether a process with the given 'pidfile' is already running and,
740 * if so, aborts. If 'pidfile' is stale, deletes it. */
742 check_already_running(void)
744 long int pid
= read_pidfile__(pidfile
, true);
746 VLOG_FATAL("%s: already running as pid %ld, aborting", pidfile
, pid
);
747 } else if (pid
< 0) {
748 VLOG_FATAL("%s: pidfile check failed (%s), aborting",
749 pidfile
, ovs_strerror(-pid
));