10 #include <sys/ioctl.h>
12 #include <sys/socket.h>
18 #include "dinit-log.h"
19 #include "dinit-socket.h"
20 #include "dinit-util.h"
23 * service.cc - Service management.
24 * See service.h for details.
28 void open_control_socket(bool report_ro_failure = true) noexcept;
29 void setup_external_log() noexcept;
30 extern eventloop_t eventLoop;
32 // Find the requested service by name
33 static service_record * find_service(const std::list<service_record *> & records,
34 const char *name) noexcept
37 list<service_record *>::const_iterator i = records.begin();
38 for ( ; i != records.end(); i++ ) {
39 if (strcmp((*i)->get_service_name().c_str(), name) == 0) {
46 service_record * service_set::find_service(const std::string &name) noexcept
48 return ::find_service(records, name.c_str());
51 void service_set::stop_service(const std::string & name) noexcept
53 service_record *record = find_service(name);
54 if (record != nullptr) {
60 // Called when a service has actually stopped; dependents have stopped already, unless this stop
61 // is due to an unexpected process termination.
62 void service_record::stopped() noexcept
64 if (onstart_flags.runs_on_console) {
65 tcsetpgrp(0, getpgrp());
66 discard_console_log_buffer();
72 // If we are a soft dependency of another target, break the acquisition from that target now:
73 for (auto & dependent : dependents) {
74 if (dependent->dep_type != dependency_type::REGULAR) {
75 if (dependent->holding_acq) {
76 dependent->holding_acq = false;
82 bool will_restart = (desired_state == service_state_t::STARTED)
83 && services->get_auto_restart();
85 for (auto dependency : depends_on) {
86 // we signal dependencies in case they are waiting for us to stop:
87 dependency.get_to()->dependent_stopped();
90 service_state = service_state_t::STOPPED;
93 // Desired state is "started".
98 if (socket_fd != -1) {
103 if (start_explicit) {
104 start_explicit = false;
107 else if (required_by == 0) {
108 services->service_inactive(this);
112 log_service_stopped(service_name);
113 notify_listeners(service_event_t::STOPPED);
116 dasynq::rearm service_child_watcher::status_change(eventloop_t &loop, pid_t child, int status) noexcept
118 base_process_service *sr = service;
121 sr->exit_status = status;
123 // Ok, for a process service, any process death which we didn't rig
124 // ourselves is a bit... unexpected. Probably, the child died because
125 // we asked it to (sr->service_state == STOPPING). But even if
126 // we didn't, there's not much we can do.
128 if (sr->waiting_for_execstat) {
129 // We still don't have an exec() status from the forked child, wait for that
130 // before doing any further processing.
131 return rearm::NOOP; // hold watch reservation
134 // Must stop watch now since handle_exit_status might result in re-launch:
135 // (stop_watch instead of deregister, so that we hold watch reservation).
138 if (sr->stop_timer_armed) {
139 sr->restart_timer.stop_timer(loop);
140 sr->stop_timer_armed = false;
143 sr->handle_exit_status(status);
147 bool service_record::do_auto_restart() noexcept
150 return services->get_auto_restart();
155 void service_record::emergency_stop() noexcept
157 if (! do_auto_restart() && start_explicit) {
158 start_explicit = false;
166 void process_service::handle_exit_status(int exit_status) noexcept
168 bool did_exit = WIFEXITED(exit_status);
169 bool was_signalled = WIFSIGNALED(exit_status);
172 if (exit_status != 0 && service_state != service_state_t::STOPPING) {
174 log(loglevel_t::ERROR, "Service ", service_name, " process terminated with exit code ",
175 WEXITSTATUS(exit_status));
177 else if (was_signalled) {
178 log(loglevel_t::ERROR, "Service ", service_name, " terminated due to signal ",
179 WTERMSIG(exit_status));
183 if (service_state == service_state_t::STARTING) {
184 if (did_exit && WEXITSTATUS(exit_status) == 0) {
191 else if (service_state == service_state_t::STOPPING) {
192 // We won't log a non-zero exit status or termination due to signal here -
193 // we assume that the process died because we signalled it.
196 else if (smooth_recovery && service_state == service_state_t::STARTED
197 && desired_state == service_state_t::STARTED) {
198 do_smooth_recovery();
204 services->process_queues();
207 void base_process_service::do_smooth_recovery() noexcept
209 if (! restart_ps_process()) {
211 services->process_queues();
215 void bgproc_service::handle_exit_status(int exit_status) noexcept
218 bool did_exit = WIFEXITED(exit_status);
219 bool was_signalled = WIFSIGNALED(exit_status);
221 if (exit_status != 0 && service_state != service_state_t::STOPPING) {
223 log(loglevel_t::ERROR, "Service ", service_name, " process terminated with exit code ",
224 WEXITSTATUS(exit_status));
226 else if (was_signalled) {
227 log(loglevel_t::ERROR, "Service ", service_name, " terminated due to signal ",
228 WTERMSIG(exit_status));
232 // This may be a "smooth recovery" where we are restarting the process while leaving the
233 // service in the STARTED state.
234 if (restarting && service_state == service_state_t::STARTED) {
236 bool need_stop = false;
237 if ((did_exit && WEXITSTATUS(exit_status) != 0) || was_signalled) {
241 // We need to re-read the PID, since it has now changed.
242 if (pid_file.length() != 0) {
243 auto pid_result = read_pid_file(&exit_status);
244 switch (pid_result) {
245 case pid_result_t::FAILED:
246 // Failed startup: no auto-restart.
249 case pid_result_t::TERMINATED:
251 case pid_result_t::OK:
258 // Failed startup: no auto-restart.
260 services->process_queues();
267 if (service_state == service_state_t::STARTING) {
268 // POSIX requires that if the process exited clearly with a status code of 0,
269 // the exit status value will be 0:
270 if (exit_status == 0) {
271 auto pid_result = read_pid_file(&exit_status);
272 switch (pid_result) {
273 case pid_result_t::FAILED:
274 // Failed startup: no auto-restart.
277 case pid_result_t::TERMINATED:
278 // started, but immediately terminated
281 case pid_result_t::OK:
290 else if (service_state == service_state_t::STOPPING) {
291 // We won't log a non-zero exit status or termination due to signal here -
292 // we assume that the process died because we signalled it.
295 else if (smooth_recovery && service_state == service_state_t::STARTED && desired_state == service_state_t::STARTED) {
296 do_smooth_recovery();
300 // we must be STARTED
301 if (! do_auto_restart() && start_explicit) {
302 start_explicit = false;
309 services->process_queues();
312 void scripted_service::handle_exit_status(int exit_status) noexcept
314 bool did_exit = WIFEXITED(exit_status);
315 bool was_signalled = WIFSIGNALED(exit_status);
317 if (service_state == service_state_t::STOPPING) {
318 if (did_exit && WEXITSTATUS(exit_status) == 0) {
322 // ??? failed to stop! Let's log it as info:
324 log(loglevel_t::INFO, "Service ", service_name, " stop command failed with exit code ",
325 WEXITSTATUS(exit_status));
327 else if (was_signalled) {
328 log(loglevel_t::INFO, "Serivice ", service_name, " stop command terminated due to signal ",
329 WTERMSIG(exit_status));
331 // Just assume that we stopped, so that any dependencies
335 services->process_queues();
338 if (exit_status == 0) {
344 log(loglevel_t::ERROR, "Service ", service_name, " command failed with exit code ",
345 WEXITSTATUS(exit_status));
347 else if (was_signalled) {
348 log(loglevel_t::ERROR, "Service ", service_name, " command terminated due to signal ",
349 WTERMSIG(exit_status));
353 services->process_queues();
357 rearm exec_status_pipe_watcher::fd_event(eventloop_t &loop, int fd, int flags) noexcept
359 base_process_service *sr = service;
360 sr->waiting_for_execstat = false;
363 int r = read(get_watched_fd(), &exec_status, sizeof(int));
365 close(get_watched_fd());
368 // We read an errno code; exec() failed, and the service startup failed.
370 sr->child_listener.deregister(eventLoop, sr->pid);
371 sr->reserved_child_watch = false;
372 if (sr->stop_timer_armed) {
373 sr->restart_timer.stop_timer(loop);
374 sr->stop_timer_armed = false;
378 log(loglevel_t::ERROR, sr->service_name, ": execution failed: ", strerror(exec_status));
379 if (sr->service_state == service_state_t::STARTING) {
380 sr->failed_to_start();
382 else if (sr->service_state == service_state_t::STOPPING) {
383 // Must be a scripted service, or a regular process service that happened to be in smooth
384 // recovery when the stop was issued.
385 if (sr->record_type == service_type::PROCESS) {
386 if (sr->stop_check_dependents()) {
391 // We've logged the failure, but it's probably better not to leave the service in
396 else if (sr->service_state == service_state_t::STARTED) {
397 // Process service in smooth recovery:
398 sr->emergency_stop();
403 if (sr->record_type == service_type::PROCESS) {
404 // This could be a smooth recovery (state already STARTED). Even more, the process
405 // might be stopped (and killed via a signal) during smooth recovery. We don't to
406 // process startup again in either case, so we check for state STARTING:
407 if (sr->service_state == service_state_t::STARTING) {
410 else if (sr->service_state == service_state_t::STOPPING) {
411 // stopping, but smooth recovery was in process. That's now over so we can
412 // commence normal stop. Note that if pid == -1 the process already stopped(!),
413 // that's handled below.
414 if (sr->pid != -1 && sr->stop_check_dependents()) {
415 sr->all_deps_stopped();
421 // Somehow the process managed to complete before we even saw the status.
422 sr->handle_exit_status(sr->exit_status);
426 sr->services->process_queues();
428 return rearm::REMOVED;
431 void service_record::require() noexcept
433 if (required_by++ == 0) {
434 prop_require = !prop_release;
435 prop_release = false;
436 services->add_prop_queue(this);
440 void service_record::release() noexcept
442 if (--required_by == 0) {
443 desired_state = service_state_t::STOPPED;
445 // Can stop, and can release dependencies now. We don't need to issue a release if
446 // the require was pending though:
447 prop_release = !prop_require;
448 prop_require = false;
449 services->add_prop_queue(this);
451 if (service_state == service_state_t::STOPPED) {
452 services->service_inactive(this);
460 void service_record::release_dependencies() noexcept
462 for (auto & dependency : depends_on) {
463 service_record * dep_to = dependency.get_to();
464 if (dependency.holding_acq) {
466 dependency.holding_acq = false;
471 void service_record::start(bool activate) noexcept
473 if (activate && ! start_explicit) {
475 start_explicit = true;
478 if (desired_state == service_state_t::STARTED && service_state != service_state_t::STOPPED) return;
480 bool was_active = service_state != service_state_t::STOPPED || desired_state != service_state_t::STOPPED;
481 desired_state = service_state_t::STARTED;
483 if (service_state != service_state_t::STOPPED) {
484 // We're already starting/started, or we are stopping and need to wait for
485 // that the complete.
486 if (service_state != service_state_t::STOPPING || ! can_interrupt_stop()) {
489 // We're STOPPING, and that can be interrupted. Our dependencies might be STOPPING,
490 // but if so they are waiting (for us), so they too can be instantly returned to
492 notify_listeners(service_event_t::STOPCANCELLED);
494 else if (! was_active) {
495 services->service_active(this);
498 service_state = service_state_t::STARTING;
499 waiting_for_deps = true;
501 if (start_check_dependencies()) {
502 services->add_transition_queue(this);
506 void service_record::do_propagation() noexcept
509 // Need to require all our dependencies
510 for (auto & dep : depends_on) {
511 dep.get_to()->require();
512 dep.holding_acq = true;
514 prop_require = false;
518 release_dependencies();
519 prop_release = false;
523 prop_failure = false;
524 failed_to_start(true);
538 void service_record::execute_transition() noexcept
540 // state is STARTED with restarting set true if we are running a smooth recovery.
541 if (service_state == service_state_t::STARTING || (service_state == service_state_t::STARTED
543 if (check_deps_started()) {
544 bool have_console = service_state == service_state_t::STARTED && onstart_flags.runs_on_console;
545 all_deps_started(have_console);
548 else if (service_state == service_state_t::STOPPING) {
549 if (stop_check_dependents()) {
555 void service_record::do_start() noexcept
557 if (pinned_stopped) return;
559 if (service_state != service_state_t::STARTING) {
563 service_state = service_state_t::STARTING;
565 waiting_for_deps = true;
567 // Ask dependencies to start, mark them as being waited on.
568 if (check_deps_started()) {
569 // Once all dependencies are started, we start properly:
574 void service_record::dependency_started() noexcept
576 if ((service_state == service_state_t::STARTING || service_state == service_state_t::STARTED)
577 && waiting_for_deps) {
578 services->add_transition_queue(this);
582 bool service_record::start_check_dependencies() noexcept
584 bool all_deps_started = true;
586 for (auto & dep : depends_on) {
587 service_record * to = dep.get_to();
588 if (to->service_state != service_state_t::STARTED) {
589 if (to->service_state != service_state_t::STARTING) {
590 to->prop_start = true;
591 services->add_prop_queue(to);
593 dep.waiting_on = true;
594 all_deps_started = false;
598 return all_deps_started;
601 bool service_record::check_deps_started() noexcept
603 for (auto & dep : depends_on) {
604 if (dep.waiting_on) {
612 bool service_record::open_socket() noexcept
614 if (socket_path.empty() || socket_fd != -1) {
615 // No socket, or already open
619 const char * saddrname = socket_path.c_str();
621 // Check the specified socket path
622 struct stat stat_buf;
623 if (stat(saddrname, &stat_buf) == 0) {
624 if ((stat_buf.st_mode & S_IFSOCK) == 0) {
626 log(loglevel_t::ERROR, service_name, ": Activation socket file exists (and is not a socket)");
630 else if (errno != ENOENT) {
632 log(loglevel_t::ERROR, service_name, ": Error checking activation socket: ", strerror(errno));
636 // Remove stale socket file (if it exists).
637 // We won't test the return from unlink - if it fails other than due to ENOENT, we should get an
638 // error when we try to create the socket anyway.
641 uint sockaddr_size = offsetof(struct sockaddr_un, sun_path) + socket_path.length() + 1;
642 struct sockaddr_un * name = static_cast<sockaddr_un *>(malloc(sockaddr_size));
643 if (name == nullptr) {
644 log(loglevel_t::ERROR, service_name, ": Opening activation socket: out of memory");
648 name->sun_family = AF_UNIX;
649 strcpy(name->sun_path, saddrname);
651 int sockfd = dinit_socket(AF_UNIX, SOCK_STREAM, 0, SOCK_NONBLOCK | SOCK_CLOEXEC);
653 log(loglevel_t::ERROR, service_name, ": Error creating activation socket: ", strerror(errno));
658 if (bind(sockfd, (struct sockaddr *) name, sockaddr_size) == -1) {
659 log(loglevel_t::ERROR, service_name, ": Error binding activation socket: ", strerror(errno));
667 // POSIX (1003.1, 2013) says that fchown and fchmod don't necessarily work on sockets. We have to
668 // use chown and chmod instead.
669 if (chown(saddrname, socket_uid, socket_gid)) {
670 log(loglevel_t::ERROR, service_name, ": Error setting activation socket owner/group: ", strerror(errno));
675 if (chmod(saddrname, socket_perms) == -1) {
676 log(loglevel_t::ERROR, service_name, ": Error setting activation socket permissions: ", strerror(errno));
681 if (listen(sockfd, 128) == -1) { // 128 "seems reasonable".
682 log(loglevel_t::ERROR, ": Error listening on activation socket: ", strerror(errno));
691 void service_record::all_deps_started(bool has_console) noexcept
693 if (onstart_flags.starts_on_console && ! has_console) {
694 waiting_for_deps = true;
699 waiting_for_deps = false;
701 if (! can_proceed_to_start()) {
702 waiting_for_deps = true;
706 if (! open_socket()) {
710 bool start_success = start_ps_process();
711 if (! start_success) {
716 void service_record::acquired_console() noexcept
718 if (service_state != service_state_t::STARTING) {
719 // We got the console but no longer want it.
722 else if (check_deps_started()) {
723 all_deps_started(true);
726 // We got the console but can't use it yet.
731 bgproc_service::pid_result_t
732 bgproc_service::read_pid_file(int *exit_status) noexcept
734 const char *pid_file_c = pid_file.c_str();
735 int fd = open(pid_file_c, O_CLOEXEC);
737 log(loglevel_t::ERROR, service_name, ": read pid file: ", strerror(errno));
738 return pid_result_t::FAILED;
741 char pidbuf[21]; // just enough to hold any 64-bit integer
742 int r = ss_read(fd, pidbuf, 20);
744 // Could not read from PID file
745 log(loglevel_t::ERROR, service_name, ": could not read from pidfile; ", strerror(errno));
747 return pid_result_t::FAILED;
751 pidbuf[r] = 0; // store nul terminator
753 bool valid_pid = false;
755 unsigned long long v = std::stoull(pidbuf, nullptr, 0);
756 if (v <= std::numeric_limits<pid_t>::max()) {
761 catch (std::out_of_range &exc) {
764 catch (std::invalid_argument &exc) {
765 // Ok, so it doesn't look like a number: proceed...
769 pid_t wait_r = waitpid(pid, exit_status, WNOHANG);
770 if (wait_r == -1 && errno == ECHILD) {
771 // We can't track this child - check process exists:
772 if (kill(pid, 0) == 0 || errno != ESRCH) {
773 tracking_child = false;
774 return pid_result_t::OK;
777 log(loglevel_t::ERROR, service_name, ": pid read from pidfile (", pid, ") is not valid");
779 return pid_result_t::FAILED;
782 else if (wait_r == pid) {
784 return pid_result_t::TERMINATED;
786 else if (wait_r == 0) {
787 // We can track the child
788 child_listener.add_reserved(eventLoop, pid, DEFAULT_PRIORITY - 10);
789 tracking_child = true;
790 reserved_child_watch = true;
791 return pid_result_t::OK;
795 log(loglevel_t::ERROR, service_name, ": pid read from pidfile (", pid, ") is not valid");
797 return pid_result_t::FAILED;
800 void service_record::started() noexcept
802 if (onstart_flags.starts_on_console && ! onstart_flags.runs_on_console) {
803 tcsetpgrp(0, getpgrp());
807 log_service_started(service_name);
808 service_state = service_state_t::STARTED;
809 notify_listeners(service_event_t::STARTED);
811 if (onstart_flags.rw_ready) {
812 open_control_socket();
814 if (onstart_flags.log_ready) {
815 setup_external_log();
818 if (force_stop || desired_state == service_state_t::STOPPED) {
824 // Notify any dependents whose desired state is STARTED:
825 for (auto dept : dependents) {
826 dept->get_from()->dependency_started();
827 dept->waiting_on = false;
831 void service_record::failed_to_start(bool depfailed) noexcept
833 if (!depfailed && onstart_flags.starts_on_console) {
834 tcsetpgrp(0, getpgrp());
838 log_service_failed(service_name);
839 service_state = service_state_t::STOPPED;
840 if (start_explicit) {
841 start_explicit = false;
844 notify_listeners(service_event_t::FAILEDSTART);
846 // Cancel start of dependents:
847 for (auto & dept : dependents) {
848 switch (dept->dep_type) {
849 case dependency_type::REGULAR:
850 case dependency_type::MILESTONE:
851 if (dept->get_from()->service_state == service_state_t::STARTING) {
852 dept->get_from()->prop_failure = true;
853 services->add_prop_queue(dept->get_from());
856 case dependency_type::WAITS_FOR:
857 case dependency_type::SOFT:
858 if (dept->waiting_on) {
859 dept->waiting_on = false;
860 dept->get_from()->dependency_started();
862 if (dept->holding_acq) {
863 dept->holding_acq = false;
870 bool service_record::start_ps_process() noexcept
872 // default implementation: there is no process, so we are started.
877 bool base_process_service::start_ps_process() noexcept
881 return restart_ps_process();
886 eventLoop.get_time(restart_interval_time, clock_type::MONOTONIC);
887 restart_interval_count = 0;
888 return start_ps_process(exec_arg_parts, onstart_flags.starts_on_console);
892 bool base_process_service::start_ps_process(const std::vector<const char *> &cmd, bool on_console) noexcept
894 // In general, you can't tell whether fork/exec is successful. We use a pipe to communicate
895 // success/failure from the child to the parent. The pipe is set CLOEXEC so a successful
896 // exec closes the pipe, and the parent sees EOF. If the exec is unsuccessful, the errno
897 // is written to the pipe, and the parent can read it.
899 eventLoop.get_time(last_start_time, clock_type::MONOTONIC);
902 if (pipe2(pipefd, O_CLOEXEC)) {
903 log(loglevel_t::ERROR, service_name, ": can't create status check pipe: ", strerror(errno));
907 const char * logfile = this->logfile.c_str();
909 logfile = "/dev/null";
912 bool child_status_registered = false;
913 control_conn_t *control_conn = nullptr;
915 int control_socket[2] = {-1, -1};
916 if (onstart_flags.pass_cs_fd) {
917 if (dinit_socketpair(AF_UNIX, SOCK_STREAM, /* protocol */ 0, control_socket, SOCK_NONBLOCK)) {
918 log(loglevel_t::ERROR, service_name, ": can't create control socket: ", strerror(errno));
922 // Make the server side socket close-on-exec:
923 int fdflags = fcntl(control_socket[0], F_GETFD);
924 fcntl(control_socket[0], F_SETFD, fdflags | FD_CLOEXEC);
927 control_conn = new control_conn_t(eventLoop, services, control_socket[0]);
929 catch (std::exception &exc) {
930 log(loglevel_t::ERROR, service_name, ": can't launch process; out of memory");
935 // Set up complete, now fork and exec:
940 child_status_listener.add_watch(eventLoop, pipefd[0], IN_EVENTS);
941 child_status_registered = true;
943 // We specify a high priority (i.e. low priority value) so that process termination is
944 // handled early. This means we have always recorded that the process is terminated by the
945 // time that we handle events that might otherwise cause us to signal the process, so we
946 // avoid sending a signal to an invalid (and possibly recycled) process ID.
947 forkpid = child_listener.fork(eventLoop, reserved_child_watch, DEFAULT_PRIORITY - 10);
948 reserved_child_watch = true;
950 catch (std::exception &e) {
951 log(loglevel_t::ERROR, service_name, ": Could not fork: ", e.what());
956 run_child_proc(cmd.data(), logfile, on_console, pipefd[1], control_socket[1]);
960 close(pipefd[1]); // close the 'other end' fd
961 if (control_socket[1] != -1) {
962 close(control_socket[1]);
966 waiting_for_execstat = true;
973 if (child_status_registered) {
974 child_status_listener.deregister(eventLoop);
977 if (onstart_flags.pass_cs_fd) {
981 close(control_socket[0]);
982 close(control_socket[1]);
992 void service_record::run_child_proc(const char * const *args, const char *logfile, bool on_console,
993 int wpipefd, int csfd) noexcept
995 // Child process. Must not allocate memory (or otherwise risk throwing any exception)
996 // from here until exit().
998 // If the console already has a session leader, presumably it is us. On the other hand
999 // if it has no session leader, and we don't create one, then control inputs such as
1000 // ^C will have no effect.
1001 bool do_set_ctty = (tcgetsid(0) == -1);
1003 // Copy signal mask, but unmask signals that we masked on startup. For the moment, we'll
1004 // also block all signals, since apparently dup() can be interrupted (!!! really, POSIX??).
1005 sigset_t sigwait_set;
1006 sigset_t sigall_set;
1007 sigfillset(&sigall_set);
1008 sigprocmask(SIG_SETMASK, &sigall_set, &sigwait_set);
1009 sigdelset(&sigwait_set, SIGCHLD);
1010 sigdelset(&sigwait_set, SIGINT);
1011 sigdelset(&sigwait_set, SIGTERM);
1012 sigdelset(&sigwait_set, SIGQUIT);
1014 constexpr int bufsz = ((CHAR_BIT * sizeof(pid_t)) / 3 + 2) + 11;
1015 // "LISTEN_PID=" - 11 characters; the expression above gives a conservative estimate
1016 // on the maxiumum number of bytes required for LISTEN=nnn, including nul terminator,
1017 // where nnn is a pid_t in decimal (i.e. one decimal digit is worth just over 3 bits).
1020 // "DINIT_CS_FD=" - 12 bytes. (we -1 from sizeof(int) in account of sign bit).
1021 constexpr int csenvbufsz = ((CHAR_BIT * sizeof(int) - 1) / 3 + 2) + 12;
1022 char csenvbuf[csenvbufsz];
1024 int minfd = (socket_fd == -1) ? 3 : 4;
1026 // Move wpipefd/csfd to another fd if necessary
1027 if (wpipefd < minfd) {
1028 wpipefd = fcntl(wpipefd, F_DUPFD_CLOEXEC, minfd);
1029 if (wpipefd == -1) goto failure_out;
1032 if (csfd != -1 && csfd < minfd) {
1033 csfd = fcntl(csfd, F_DUPFD, minfd);
1034 if (csfd == -1) goto failure_out;
1037 if (socket_fd != -1) {
1039 if (dup2(socket_fd, 3) == -1) goto failure_out;
1040 if (socket_fd != 3) {
1044 if (putenv(const_cast<char *>("LISTEN_FDS=1"))) goto failure_out;
1045 snprintf(nbuf, bufsz, "LISTEN_PID=%jd", static_cast<intmax_t>(getpid()));
1046 if (putenv(nbuf)) goto failure_out;
1050 snprintf(csenvbuf, csenvbufsz, "DINIT_CS_FD=%d", csfd);
1051 if (putenv(csenvbuf)) goto failure_out;
1055 // Re-set stdin, stdout, stderr
1056 close(0); close(1); close(2);
1058 if (open("/dev/null", O_RDONLY) == 0) {
1059 // stdin = 0. That's what we should have; proceed with opening
1060 // stdout and stderr.
1061 if (open(logfile, O_WRONLY | O_CREAT | O_APPEND, S_IRUSR | S_IWUSR) != 1) {
1064 if (dup2(1, 2) != 2) {
1068 else goto failure_out;
1070 // We have the option of creating a session and process group, or just a new process
1071 // group. If we just create a new process group, the child process cannot make itself
1072 // a session leader if it wants to do that (eg getty/login will generally want this).
1073 // If we do neither, and we are running with a controlling terminal, a ^C or similar
1074 // will also affect the child process (which probably isn't so bad, though since we
1075 // will handle the shutdown ourselves it's not necessary). Creating a new session
1076 // (and a new process group as part of that) seems like a safe bet, and has the
1077 // advantage of letting us signal the process as part of a process group.
1081 // "run on console" - run as a foreground job on the terminal/console device
1083 // if do_set_ctty is false, we are the session leader; we are probably running
1084 // as a user process. Don't create a new session leader in that case, and run
1085 // as part of the parent session. Otherwise, the new session cannot claim the
1086 // terminal as a controlling terminal (it is already claimed), meaning that it
1087 // will not see control signals from ^C etc.
1090 // Disable suspend (^Z) (and on some systems, delayed suspend / ^Y)
1091 signal(SIGTSTP, SIG_IGN);
1093 // Become session leader
1095 ioctl(0, TIOCSCTTY, 0);
1098 tcsetpgrp(0, getpgrp());
1101 sigprocmask(SIG_SETMASK, &sigwait_set, nullptr);
1103 execvp(args[0], const_cast<char **>(args));
1105 // If we got here, the exec failed:
1107 int exec_status = errno;
1108 write(wpipefd, &exec_status, sizeof(int));
1112 // Mark this and all dependent services as force-stopped.
1113 void service_record::forced_stop() noexcept
1115 if (service_state != service_state_t::STOPPED) {
1117 services->add_transition_queue(this);
1121 void service_record::dependent_stopped() noexcept
1123 if (service_state == service_state_t::STOPPING && waiting_for_deps) {
1124 services->add_transition_queue(this);
1128 void service_record::stop(bool bring_down) noexcept
1130 if (start_explicit) {
1131 start_explicit = false;
1140 void service_record::do_stop() noexcept
1142 if (pinned_started) return;
1144 if (start_explicit && ! do_auto_restart()) {
1145 start_explicit = false;
1147 if (required_by == 0) return; // release will re-call us anyway
1150 if (service_state != service_state_t::STARTED) {
1151 if (service_state == service_state_t::STARTING) {
1152 if (! can_interrupt_start()) {
1153 // Well this is awkward: we're going to have to continue
1154 // starting, but we don't want any dependents to think that
1155 // they are still waiting to start.
1156 // Make sure they remain stopped:
1161 // We must have had desired_state == STARTED.
1162 notify_listeners(service_event_t::STARTCANCELLED);
1166 // Reaching this point, we are starting interruptibly - so we
1167 // stop now (by falling through to below).
1170 // If we're starting we need to wait for that to complete.
1171 // If we're already stopping/stopped there's nothing to do.
1176 service_state = service_state_t::STOPPING;
1177 waiting_for_deps = true;
1178 if (stop_dependents()) {
1179 services->add_transition_queue(this);
1183 bool service_record::stop_check_dependents() noexcept
1185 bool all_deps_stopped = true;
1186 for (auto dept : dependents) {
1187 if (dept->dep_type == dependency_type::REGULAR && ! dept->get_from()->is_stopped()) {
1188 all_deps_stopped = false;
1193 return all_deps_stopped;
1196 bool service_record::stop_dependents() noexcept
1198 bool all_deps_stopped = true;
1199 for (auto dept : dependents) {
1200 if (dept->dep_type == dependency_type::REGULAR) {
1201 if (! dept->get_from()->is_stopped()) {
1202 // Note we check *first* since if the dependent service is not stopped,
1203 // 1. We will issue a stop to it shortly and
1204 // 2. It will notify us when stopped, at which point the stop_check_dependents()
1205 // check is run anyway.
1206 all_deps_stopped = false;
1210 // If this service is to be forcefully stopped, dependents must also be.
1211 dept->get_from()->forced_stop();
1214 dept->get_from()->prop_stop = true;
1215 services->add_prop_queue(dept->get_from());
1219 return all_deps_stopped;
1222 // All dependents have stopped; we can stop now, too. Only called when STOPPING.
1223 void service_record::all_deps_stopped() noexcept
1225 waiting_for_deps = false;
1229 void base_process_service::kill_pg(int signo) noexcept
1231 pid_t pgid = getpgid(pid);
1233 // only should happen if pid is invalid, which should never happen...
1234 log(loglevel_t::ERROR, service_name, ": can't signal process: ", strerror(errno));
1240 void base_process_service::all_deps_stopped() noexcept
1242 waiting_for_deps = false;
1244 // The process is still kicking on - must actually kill it. We signal the process
1245 // group (-pid) rather than just the process as there's less risk then of creating
1246 // an orphaned process group:
1247 if (! onstart_flags.no_sigterm) {
1250 if (term_signal != -1) {
1251 kill_pg(term_signal);
1254 // In most cases, the rest is done in handle_exit_status.
1255 // If we are a BGPROCESS and the process is not our immediate child, however, that
1256 // won't work - check for this now:
1257 if (record_type == service_type::BGPROCESS && ! tracking_child) {
1260 else if (stop_timeout != time_val(0,0)) {
1261 restart_timer.arm_timer_rel(eventLoop, stop_timeout);
1262 stop_timer_armed = true;
1266 // The process is already dead.
1271 void process_service::all_deps_stopped() noexcept
1273 waiting_for_deps = false;
1274 if (waiting_for_execstat) {
1275 // The process is still starting. This should be uncommon, but can occur during
1276 // smooth recovery. We can't do much now; we have to wait until we get the
1277 // status, and then act appropriately.
1280 else if (pid != -1) {
1281 // The process is still kicking on - must actually kill it. We signal the process
1282 // group (-pid) rather than just the process as there's less risk then of creating
1283 // an orphaned process group:
1284 if (! onstart_flags.no_sigterm) {
1287 if (term_signal != -1) {
1288 kill_pg(term_signal);
1291 // In most cases, the rest is done in handle_exit_status.
1292 // If we are a BGPROCESS and the process is not our immediate child, however, that
1293 // won't work - check for this now:
1294 if (record_type == service_type::BGPROCESS && ! tracking_child) {
1297 else if (stop_timeout != time_val(0,0)) {
1298 restart_timer.arm_timer_rel(eventLoop, stop_timeout);
1299 stop_timer_armed = true;
1303 // The process is already dead.
1308 void scripted_service::all_deps_stopped() noexcept
1310 waiting_for_deps = false;
1311 if (stop_command.length() == 0) {
1314 else if (! start_ps_process(stop_arg_parts, false)) {
1315 // Couldn't execute stop script, but there's not much we can do:
1319 // successfully started stop script: start kill timer:
1320 if (stop_timeout != time_val(0,0)) {
1321 restart_timer.arm_timer_rel(eventLoop, stop_timeout);
1322 stop_timer_armed = true;
1327 void service_record::unpin() noexcept
1329 if (pinned_started) {
1330 pinned_started = false;
1331 if (desired_state == service_state_t::STOPPED || force_stop) {
1333 services->process_queues();
1336 if (pinned_stopped) {
1337 pinned_stopped = false;
1338 if (desired_state == service_state_t::STARTED) {
1340 services->process_queues();
1345 void service_record::queue_for_console() noexcept
1347 services->append_console_queue(this);
1350 void service_record::release_console() noexcept
1352 services->pull_console_queue();
1355 void service_record::interrupt_start() noexcept
1357 services->unqueue_console(this);
1360 void service_set::service_active(service_record *sr) noexcept
1365 void service_set::service_inactive(service_record *sr) noexcept
1370 base_process_service::base_process_service(service_set *sset, string name,
1371 service_type service_type_p, string &&command,
1372 std::list<std::pair<unsigned,unsigned>> &command_offsets,
1373 const std::list<prelim_dep> &deplist_p)
1374 : service_record(sset, name, service_type_p, std::move(command), command_offsets,
1375 deplist_p), child_listener(this), child_status_listener(this)
1377 restart_interval_count = 0;
1378 restart_interval_time = {0, 0};
1379 restart_timer.service = this;
1380 restart_timer.add_timer(eventLoop);
1382 // By default, allow a maximum of 3 restarts within 10.0 seconds:
1383 restart_interval.seconds() = 10;
1384 restart_interval.nseconds() = 0;
1385 max_restart_interval_count = 3;
1387 waiting_restart_timer = false;
1388 reserved_child_watch = false;
1389 tracking_child = false;
1390 stop_timer_armed = false;
1393 void base_process_service::do_restart() noexcept
1395 waiting_restart_timer = false;
1396 restart_interval_count++;
1398 // We may be STARTING (regular restart) or STARTED ("smooth recovery"). This affects whether
1399 // the process should be granted access to the console:
1400 bool on_console = service_state == service_state_t::STARTING
1401 ? onstart_flags.starts_on_console : onstart_flags.runs_on_console;
1403 if (service_state == service_state_t::STARTING) {
1404 // for a smooth recovery, we want to check dependencies are available before actually
1406 if (! check_deps_started()) {
1407 waiting_for_deps = true;
1412 if (! start_ps_process(exec_arg_parts, on_console)) {
1414 if (service_state == service_state_t::STARTING) {
1418 desired_state = service_state_t::STOPPED;
1421 services->process_queues();
1425 bool base_process_service::restart_ps_process() noexcept
1427 using time_val = dasynq::time_val;
1429 time_val current_time;
1430 eventLoop.get_time(current_time, clock_type::MONOTONIC);
1432 if (max_restart_interval_count != 0) {
1433 // Check whether we're still in the most recent restart check interval:
1434 time_val int_diff = current_time - restart_interval_time;
1435 if (int_diff < restart_interval) {
1436 if (restart_interval_count >= max_restart_interval_count) {
1437 log(loglevel_t::ERROR, "Service ", service_name, " restarting too quickly; stopping.");
1442 restart_interval_time = current_time;
1443 restart_interval_count = 0;
1447 // Check if enough time has lapsed since the prevous restart. If not, start a timer:
1448 time_val tdiff = current_time - last_start_time;
1449 if (restart_delay <= tdiff) {
1450 // > restart delay (normally 200ms)
1454 time_val timeout = restart_delay - tdiff;
1455 restart_timer.arm_timer_rel(eventLoop, timeout);
1456 waiting_restart_timer = true;
1461 void base_process_service::interrupt_start() noexcept
1463 // overridden in subclasses
1464 if (waiting_restart_timer) {
1465 restart_timer.stop_timer(eventLoop);
1466 waiting_restart_timer = false;
1468 service_record::interrupt_start();
1471 void base_process_service::kill_with_fire() noexcept
1474 log(loglevel_t::WARN, "Service ", service_name, "with pid ", pid, " exceeded allowed stop time; killing.");
1479 dasynq::rearm process_restart_timer::timer_expiry(eventloop_t &, int expiry_count)
1481 if (service->service_state == service_state_t::STOPPING) {
1482 service->kill_with_fire();
1483 service->stop_timer_armed = false;
1486 // STARTING / STARTED:
1487 service->do_restart();
1489 return dasynq::rearm::DISARM;