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_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 base_process_service::do_smooth_recovery() noexcept
168 if (! restart_ps_process()) {
170 services->process_queues();
174 void process_service::handle_exit_status(int exit_status) noexcept
176 bool did_exit = WIFEXITED(exit_status);
177 bool was_signalled = WIFSIGNALED(exit_status);
179 auto service_state = get_state();
181 if (exit_status != 0 && service_state != service_state_t::STOPPING) {
183 log(loglevel_t::ERROR, "Service ", get_name(), " process terminated with exit code ",
184 WEXITSTATUS(exit_status));
186 else if (was_signalled) {
187 log(loglevel_t::ERROR, "Service ", get_name(), " terminated due to signal ",
188 WTERMSIG(exit_status));
192 if (service_state == service_state_t::STARTING) {
193 if (did_exit && WEXITSTATUS(exit_status) == 0) {
200 else if (service_state == service_state_t::STOPPING) {
201 // We won't log a non-zero exit status or termination due to signal here -
202 // we assume that the process died because we signalled it.
205 else if (smooth_recovery && service_state == service_state_t::STARTED
206 && get_target_state() == service_state_t::STARTED) {
207 do_smooth_recovery();
213 services->process_queues();
216 void process_service::exec_failed(int errcode) noexcept
218 log(loglevel_t::ERROR, get_name(), ": execution failed: ", strerror(errcode));
219 if (get_state() == service_state_t::STARTING) {
223 // Process service in smooth recovery:
228 void bgproc_service::handle_exit_status(int exit_status) noexcept
231 bool did_exit = WIFEXITED(exit_status);
232 bool was_signalled = WIFSIGNALED(exit_status);
233 auto service_state = get_state();
235 if (exit_status != 0 && service_state != service_state_t::STOPPING) {
237 log(loglevel_t::ERROR, "Service ", get_name(), " process terminated with exit code ",
238 WEXITSTATUS(exit_status));
240 else if (was_signalled) {
241 log(loglevel_t::ERROR, "Service ", get_name(), " terminated due to signal ",
242 WTERMSIG(exit_status));
246 // This may be a "smooth recovery" where we are restarting the process while leaving the
247 // service in the STARTED state.
248 if (restarting && service_state == service_state_t::STARTED) {
250 bool need_stop = false;
251 if ((did_exit && WEXITSTATUS(exit_status) != 0) || was_signalled) {
255 // We need to re-read the PID, since it has now changed.
256 if (pid_file.length() != 0) {
257 auto pid_result = read_pid_file(&exit_status);
258 switch (pid_result) {
259 case pid_result_t::FAILED:
260 // Failed startup: no auto-restart.
263 case pid_result_t::TERMINATED:
265 case pid_result_t::OK:
272 // Failed startup: no auto-restart.
274 services->process_queues();
281 if (service_state == service_state_t::STARTING) {
282 // POSIX requires that if the process exited clearly with a status code of 0,
283 // the exit status value will be 0:
284 if (exit_status == 0) {
285 auto pid_result = read_pid_file(&exit_status);
286 switch (pid_result) {
287 case pid_result_t::FAILED:
288 // Failed startup: no auto-restart.
291 case pid_result_t::TERMINATED:
292 // started, but immediately terminated
295 case pid_result_t::OK:
304 else if (service_state == service_state_t::STOPPING) {
305 // We won't log a non-zero exit status or termination due to signal here -
306 // we assume that the process died because we signalled it.
310 // we must be STARTED
311 if (smooth_recovery && get_target_state() == service_state_t::STARTED) {
312 do_smooth_recovery();
315 if (! do_auto_restart() && start_explicit) {
316 start_explicit = false;
323 services->process_queues();
326 void bgproc_service::exec_failed(int errcode) noexcept
328 log(loglevel_t::ERROR, get_name(), ": execution failed: ", strerror(errcode));
329 // Only time we execute is for startup:
333 void scripted_service::handle_exit_status(int exit_status) noexcept
335 bool did_exit = WIFEXITED(exit_status);
336 bool was_signalled = WIFSIGNALED(exit_status);
337 auto service_state = get_state();
339 // For a scripted service, a termination occurs in one of three main cases:
340 // - the start script completed (or failed), when service was STARTING
341 // - the start script was interrupted to cancel startup; state is STOPPING
342 // - the stop script complete (or failed), state is STOPPING
344 if (service_state == service_state_t::STOPPING) {
345 // We might be running the stop script, or we might be running the start script and have issued
346 // a cancel order via SIGINT:
347 if (did_exit && WEXITSTATUS(exit_status) == 0) {
348 if (interrupting_start) {
349 interrupting_start = false;
350 // launch stop script:
354 // We were running the stop script and finished successfully
359 if (interrupting_start) {
360 // We issued a start interrupt, so we expected this failure:
362 log(loglevel_t::INFO, "Service ", get_name(), " start cancelled; exit code ",
363 WEXITSTATUS(exit_status));
365 else if (was_signalled) {
366 log(loglevel_t::INFO, "Service ", get_name(), " start cancelled from signal ",
367 WTERMSIG(exit_status));
371 // ??? failed to stop! Let's log it as warning:
373 log(loglevel_t::WARN, "Service ", get_name(), " stop command failed with exit code ",
374 WEXITSTATUS(exit_status));
376 else if (was_signalled) {
377 log(loglevel_t::WARN, "Service ", get_name(), " stop command terminated due to signal ",
378 WTERMSIG(exit_status));
381 // Even if the stop script failed, assume that service is now stopped, so that any dependencies
382 // can be stopped. There's not really any other useful course of action here.
383 interrupting_start = false;
386 services->process_queues();
389 if (exit_status == 0) {
395 log(loglevel_t::ERROR, "Service ", get_name(), " command failed with exit code ",
396 WEXITSTATUS(exit_status));
398 else if (was_signalled) {
399 log(loglevel_t::ERROR, "Service ", get_name(), " command terminated due to signal ",
400 WTERMSIG(exit_status));
404 services->process_queues();
408 void scripted_service::exec_failed(int errcode) noexcept
410 log(loglevel_t::ERROR, get_name(), ": execution failed: ", strerror(errcode));
411 auto service_state = get_state();
412 if (service_state == service_state_t::STARTING) {
415 else if (service_state == service_state_t::STOPPING) {
416 // We've logged the failure, but it's probably better not to leave the service in
422 rearm exec_status_pipe_watcher::fd_event(eventloop_t &loop, int fd, int flags) noexcept
424 base_process_service *sr = service;
425 sr->waiting_for_execstat = false;
428 int r = read(get_watched_fd(), &exec_status, sizeof(int));
430 close(get_watched_fd());
433 // We read an errno code; exec() failed, and the service startup failed.
435 sr->child_listener.deregister(eventLoop, sr->pid);
436 sr->reserved_child_watch = false;
437 if (sr->stop_timer_armed) {
438 sr->restart_timer.stop_timer(loop);
439 sr->stop_timer_armed = false;
443 sr->exec_failed(exec_status);
447 if (sr->get_type() == service_type_t::PROCESS) {
448 // This could be a smooth recovery (state already STARTED). Even more, the process
449 // might be stopped (and killed via a signal) during smooth recovery. We don't to
450 // process startup again in either case, so we check for state STARTING:
451 if (sr->get_state() == service_state_t::STARTING) {
454 else if (sr->get_state() == service_state_t::STOPPING) {
455 // stopping, but smooth recovery was in process. That's now over so we can
456 // commence normal stop. Note that if pid == -1 the process already stopped(!),
457 // that's handled below.
458 if (sr->pid != -1 && sr->stop_check_dependents()) {
465 // Somehow the process managed to complete before we even saw the status.
466 sr->handle_exit_status(sr->exit_status);
470 sr->services->process_queues();
472 return rearm::REMOVED;
475 void service_record::require() noexcept
477 if (required_by++ == 0) {
478 prop_require = !prop_release;
479 prop_release = false;
480 services->add_prop_queue(this);
484 void service_record::release() noexcept
486 if (--required_by == 0) {
487 desired_state = service_state_t::STOPPED;
489 // Can stop, and can release dependencies now. We don't need to issue a release if
490 // the require was pending though:
491 prop_release = !prop_require;
492 prop_require = false;
493 services->add_prop_queue(this);
495 if (service_state == service_state_t::STOPPED) {
496 services->service_inactive(this);
504 void service_record::release_dependencies() noexcept
506 for (auto & dependency : depends_on) {
507 service_record * dep_to = dependency.get_to();
508 if (dependency.holding_acq) {
510 dependency.holding_acq = false;
515 void service_record::start(bool activate) noexcept
517 if (activate && ! start_explicit) {
519 start_explicit = true;
522 if (desired_state == service_state_t::STARTED && service_state != service_state_t::STOPPED) return;
524 bool was_active = service_state != service_state_t::STOPPED || desired_state != service_state_t::STOPPED;
525 desired_state = service_state_t::STARTED;
527 if (service_state != service_state_t::STOPPED) {
528 // We're already starting/started, or we are stopping and need to wait for
529 // that the complete.
530 if (service_state != service_state_t::STOPPING || ! can_interrupt_stop()) {
533 // We're STOPPING, and that can be interrupted. Our dependencies might be STOPPING,
534 // but if so they are waiting (for us), so they too can be instantly returned to
536 notify_listeners(service_event_t::STOPCANCELLED);
538 else if (! was_active) {
539 services->service_active(this);
542 service_state = service_state_t::STARTING;
543 waiting_for_deps = true;
545 if (start_check_dependencies()) {
546 services->add_transition_queue(this);
550 void service_record::do_propagation() noexcept
553 // Need to require all our dependencies
554 for (auto & dep : depends_on) {
555 dep.get_to()->require();
556 dep.holding_acq = true;
558 prop_require = false;
562 release_dependencies();
563 prop_release = false;
567 prop_failure = false;
568 failed_to_start(true);
582 void service_record::execute_transition() noexcept
584 // state is STARTED with restarting set true if we are running a smooth recovery.
585 if (service_state == service_state_t::STARTING || (service_state == service_state_t::STARTED
587 if (check_deps_started()) {
588 bool have_console = service_state == service_state_t::STARTED && onstart_flags.runs_on_console;
589 all_deps_started(have_console);
592 else if (service_state == service_state_t::STOPPING) {
593 if (stop_check_dependents()) {
599 void service_record::do_start() noexcept
601 if (pinned_stopped) return;
603 if (service_state != service_state_t::STARTING) {
607 service_state = service_state_t::STARTING;
609 waiting_for_deps = true;
611 // Ask dependencies to start, mark them as being waited on.
612 if (check_deps_started()) {
613 // Once all dependencies are started, we start properly:
618 void service_record::dependency_started() noexcept
620 if ((service_state == service_state_t::STARTING || service_state == service_state_t::STARTED)
621 && waiting_for_deps) {
622 services->add_transition_queue(this);
626 bool service_record::start_check_dependencies() noexcept
628 bool all_deps_started = true;
630 for (auto & dep : depends_on) {
631 service_record * to = dep.get_to();
632 if (to->service_state != service_state_t::STARTED) {
633 if (to->service_state != service_state_t::STARTING) {
634 to->prop_start = true;
635 services->add_prop_queue(to);
637 dep.waiting_on = true;
638 all_deps_started = false;
642 return all_deps_started;
645 bool service_record::check_deps_started() noexcept
647 for (auto & dep : depends_on) {
648 if (dep.waiting_on) {
656 bool service_record::open_socket() noexcept
658 if (socket_path.empty() || socket_fd != -1) {
659 // No socket, or already open
663 const char * saddrname = socket_path.c_str();
665 // Check the specified socket path
666 struct stat stat_buf;
667 if (stat(saddrname, &stat_buf) == 0) {
668 if ((stat_buf.st_mode & S_IFSOCK) == 0) {
670 log(loglevel_t::ERROR, service_name, ": Activation socket file exists (and is not a socket)");
674 else if (errno != ENOENT) {
676 log(loglevel_t::ERROR, service_name, ": Error checking activation socket: ", strerror(errno));
680 // Remove stale socket file (if it exists).
681 // We won't test the return from unlink - if it fails other than due to ENOENT, we should get an
682 // error when we try to create the socket anyway.
685 uint sockaddr_size = offsetof(struct sockaddr_un, sun_path) + socket_path.length() + 1;
686 struct sockaddr_un * name = static_cast<sockaddr_un *>(malloc(sockaddr_size));
687 if (name == nullptr) {
688 log(loglevel_t::ERROR, service_name, ": Opening activation socket: out of memory");
692 name->sun_family = AF_UNIX;
693 strcpy(name->sun_path, saddrname);
695 int sockfd = dinit_socket(AF_UNIX, SOCK_STREAM, 0, SOCK_NONBLOCK | SOCK_CLOEXEC);
697 log(loglevel_t::ERROR, service_name, ": Error creating activation socket: ", strerror(errno));
702 if (bind(sockfd, (struct sockaddr *) name, sockaddr_size) == -1) {
703 log(loglevel_t::ERROR, service_name, ": Error binding activation socket: ", strerror(errno));
711 // POSIX (1003.1, 2013) says that fchown and fchmod don't necessarily work on sockets. We have to
712 // use chown and chmod instead.
713 if (chown(saddrname, socket_uid, socket_gid)) {
714 log(loglevel_t::ERROR, service_name, ": Error setting activation socket owner/group: ", strerror(errno));
719 if (chmod(saddrname, socket_perms) == -1) {
720 log(loglevel_t::ERROR, service_name, ": Error setting activation socket permissions: ", strerror(errno));
725 if (listen(sockfd, 128) == -1) { // 128 "seems reasonable".
726 log(loglevel_t::ERROR, ": Error listening on activation socket: ", strerror(errno));
735 void service_record::all_deps_started(bool has_console) noexcept
737 if (onstart_flags.starts_on_console && ! has_console) {
738 waiting_for_deps = true;
743 waiting_for_deps = false;
745 if (! can_proceed_to_start()) {
746 waiting_for_deps = true;
750 if (! open_socket()) {
754 bool start_success = bring_up();
755 if (! start_success) {
760 void service_record::acquired_console() noexcept
762 if (service_state != service_state_t::STARTING) {
763 // We got the console but no longer want it.
766 else if (check_deps_started()) {
767 all_deps_started(true);
770 // We got the console but can't use it yet.
775 bgproc_service::pid_result_t
776 bgproc_service::read_pid_file(int *exit_status) noexcept
778 const char *pid_file_c = pid_file.c_str();
779 int fd = open(pid_file_c, O_CLOEXEC);
781 log(loglevel_t::ERROR, get_name(), ": read pid file: ", strerror(errno));
782 return pid_result_t::FAILED;
785 char pidbuf[21]; // just enough to hold any 64-bit integer
786 int r = ss_read(fd, pidbuf, 20);
788 // Could not read from PID file
789 log(loglevel_t::ERROR, get_name(), ": could not read from pidfile; ", strerror(errno));
791 return pid_result_t::FAILED;
795 pidbuf[r] = 0; // store nul terminator
797 bool valid_pid = false;
799 unsigned long long v = std::stoull(pidbuf, nullptr, 0);
800 if (v <= std::numeric_limits<pid_t>::max()) {
805 catch (std::out_of_range &exc) {
808 catch (std::invalid_argument &exc) {
809 // Ok, so it doesn't look like a number: proceed...
813 pid_t wait_r = waitpid(pid, exit_status, WNOHANG);
814 if (wait_r == -1 && errno == ECHILD) {
815 // We can't track this child - check process exists:
816 if (kill(pid, 0) == 0 || errno != ESRCH) {
817 tracking_child = false;
818 return pid_result_t::OK;
821 log(loglevel_t::ERROR, get_name(), ": pid read from pidfile (", pid, ") is not valid");
823 return pid_result_t::FAILED;
826 else if (wait_r == pid) {
828 return pid_result_t::TERMINATED;
830 else if (wait_r == 0) {
831 // We can track the child
832 child_listener.add_reserved(eventLoop, pid, DEFAULT_PRIORITY - 10);
833 tracking_child = true;
834 reserved_child_watch = true;
835 return pid_result_t::OK;
839 log(loglevel_t::ERROR, get_name(), ": pid read from pidfile (", pid, ") is not valid");
841 return pid_result_t::FAILED;
844 void service_record::started() noexcept
846 if (onstart_flags.starts_on_console && ! onstart_flags.runs_on_console) {
847 tcsetpgrp(0, getpgrp());
851 log_service_started(get_name());
852 service_state = service_state_t::STARTED;
853 notify_listeners(service_event_t::STARTED);
855 if (onstart_flags.rw_ready) {
856 open_control_socket();
858 if (onstart_flags.log_ready) {
859 setup_external_log();
862 if (force_stop || desired_state == service_state_t::STOPPED) {
868 // Notify any dependents whose desired state is STARTED:
869 for (auto dept : dependents) {
870 dept->get_from()->dependency_started();
871 dept->waiting_on = false;
875 void service_record::failed_to_start(bool depfailed) noexcept
877 if (!depfailed && onstart_flags.starts_on_console) {
878 tcsetpgrp(0, getpgrp());
882 log_service_failed(get_name());
883 service_state = service_state_t::STOPPED;
884 if (start_explicit) {
885 start_explicit = false;
888 notify_listeners(service_event_t::FAILEDSTART);
890 // Cancel start of dependents:
891 for (auto & dept : dependents) {
892 switch (dept->dep_type) {
893 case dependency_type::REGULAR:
894 case dependency_type::MILESTONE:
895 if (dept->get_from()->service_state == service_state_t::STARTING) {
896 dept->get_from()->prop_failure = true;
897 services->add_prop_queue(dept->get_from());
900 case dependency_type::WAITS_FOR:
901 case dependency_type::SOFT:
902 if (dept->waiting_on) {
903 dept->waiting_on = false;
904 dept->get_from()->dependency_started();
906 if (dept->holding_acq) {
907 dept->holding_acq = false;
914 bool service_record::bring_up() noexcept
916 // default implementation: there is no process, so we are started.
921 bool base_process_service::bring_up() noexcept
925 return restart_ps_process();
930 eventLoop.get_time(restart_interval_time, clock_type::MONOTONIC);
931 restart_interval_count = 0;
932 if (start_ps_process(exec_arg_parts, onstart_flags.starts_on_console)) {
933 if (start_timeout != time_val(0,0)) {
934 restart_timer.arm_timer_rel(eventLoop, start_timeout);
935 stop_timer_armed = true;
937 else if (stop_timer_armed) {
938 restart_timer.stop_timer(eventLoop);
939 stop_timer_armed = false;
947 bool base_process_service::start_ps_process(const std::vector<const char *> &cmd, bool on_console) noexcept
949 // In general, you can't tell whether fork/exec is successful. We use a pipe to communicate
950 // success/failure from the child to the parent. The pipe is set CLOEXEC so a successful
951 // exec closes the pipe, and the parent sees EOF. If the exec is unsuccessful, the errno
952 // is written to the pipe, and the parent can read it.
954 eventLoop.get_time(last_start_time, clock_type::MONOTONIC);
957 if (pipe2(pipefd, O_CLOEXEC)) {
958 log(loglevel_t::ERROR, get_name(), ": can't create status check pipe: ", strerror(errno));
962 const char * logfile = this->logfile.c_str();
964 logfile = "/dev/null";
967 bool child_status_registered = false;
968 control_conn_t *control_conn = nullptr;
970 int control_socket[2] = {-1, -1};
971 if (onstart_flags.pass_cs_fd) {
972 if (dinit_socketpair(AF_UNIX, SOCK_STREAM, /* protocol */ 0, control_socket, SOCK_NONBLOCK)) {
973 log(loglevel_t::ERROR, get_name(), ": can't create control socket: ", strerror(errno));
977 // Make the server side socket close-on-exec:
978 int fdflags = fcntl(control_socket[0], F_GETFD);
979 fcntl(control_socket[0], F_SETFD, fdflags | FD_CLOEXEC);
982 control_conn = new control_conn_t(eventLoop, services, control_socket[0]);
984 catch (std::exception &exc) {
985 log(loglevel_t::ERROR, get_name(), ": can't launch process; out of memory");
990 // Set up complete, now fork and exec:
995 child_status_listener.add_watch(eventLoop, pipefd[0], IN_EVENTS);
996 child_status_registered = true;
998 // We specify a high priority (i.e. low priority value) so that process termination is
999 // handled early. This means we have always recorded that the process is terminated by the
1000 // time that we handle events that might otherwise cause us to signal the process, so we
1001 // avoid sending a signal to an invalid (and possibly recycled) process ID.
1002 forkpid = child_listener.fork(eventLoop, reserved_child_watch, DEFAULT_PRIORITY - 10);
1003 reserved_child_watch = true;
1005 catch (std::exception &e) {
1006 log(loglevel_t::ERROR, get_name(), ": Could not fork: ", e.what());
1011 run_child_proc(cmd.data(), logfile, on_console, pipefd[1], control_socket[1]);
1015 close(pipefd[1]); // close the 'other end' fd
1016 if (control_socket[1] != -1) {
1017 close(control_socket[1]);
1021 waiting_for_execstat = true;
1028 if (child_status_registered) {
1029 child_status_listener.deregister(eventLoop);
1032 if (onstart_flags.pass_cs_fd) {
1033 delete control_conn;
1036 close(control_socket[0]);
1037 close(control_socket[1]);
1047 void service_record::run_child_proc(const char * const *args, const char *logfile, bool on_console,
1048 int wpipefd, int csfd) noexcept
1050 // Child process. Must not allocate memory (or otherwise risk throwing any exception)
1051 // from here until exit().
1053 // If the console already has a session leader, presumably it is us. On the other hand
1054 // if it has no session leader, and we don't create one, then control inputs such as
1055 // ^C will have no effect.
1056 bool do_set_ctty = (tcgetsid(0) == -1);
1058 // Copy signal mask, but unmask signals that we masked on startup. For the moment, we'll
1059 // also block all signals, since apparently dup() can be interrupted (!!! really, POSIX??).
1060 sigset_t sigwait_set;
1061 sigset_t sigall_set;
1062 sigfillset(&sigall_set);
1063 sigprocmask(SIG_SETMASK, &sigall_set, &sigwait_set);
1064 sigdelset(&sigwait_set, SIGCHLD);
1065 sigdelset(&sigwait_set, SIGINT);
1066 sigdelset(&sigwait_set, SIGTERM);
1067 sigdelset(&sigwait_set, SIGQUIT);
1069 constexpr int bufsz = ((CHAR_BIT * sizeof(pid_t)) / 3 + 2) + 11;
1070 // "LISTEN_PID=" - 11 characters; the expression above gives a conservative estimate
1071 // on the maxiumum number of bytes required for LISTEN=nnn, including nul terminator,
1072 // where nnn is a pid_t in decimal (i.e. one decimal digit is worth just over 3 bits).
1075 // "DINIT_CS_FD=" - 12 bytes. (we -1 from sizeof(int) in account of sign bit).
1076 constexpr int csenvbufsz = ((CHAR_BIT * sizeof(int) - 1) / 3 + 2) + 12;
1077 char csenvbuf[csenvbufsz];
1079 int minfd = (socket_fd == -1) ? 3 : 4;
1081 // Move wpipefd/csfd to another fd if necessary
1082 if (wpipefd < minfd) {
1083 wpipefd = fcntl(wpipefd, F_DUPFD_CLOEXEC, minfd);
1084 if (wpipefd == -1) goto failure_out;
1087 if (csfd != -1 && csfd < minfd) {
1088 csfd = fcntl(csfd, F_DUPFD, minfd);
1089 if (csfd == -1) goto failure_out;
1092 if (socket_fd != -1) {
1094 if (dup2(socket_fd, 3) == -1) goto failure_out;
1095 if (socket_fd != 3) {
1099 if (putenv(const_cast<char *>("LISTEN_FDS=1"))) goto failure_out;
1100 snprintf(nbuf, bufsz, "LISTEN_PID=%jd", static_cast<intmax_t>(getpid()));
1101 if (putenv(nbuf)) goto failure_out;
1105 snprintf(csenvbuf, csenvbufsz, "DINIT_CS_FD=%d", csfd);
1106 if (putenv(csenvbuf)) goto failure_out;
1110 // Re-set stdin, stdout, stderr
1111 close(0); close(1); close(2);
1113 if (open("/dev/null", O_RDONLY) == 0) {
1114 // stdin = 0. That's what we should have; proceed with opening
1115 // stdout and stderr.
1116 if (open(logfile, O_WRONLY | O_CREAT | O_APPEND, S_IRUSR | S_IWUSR) != 1) {
1119 if (dup2(1, 2) != 2) {
1123 else goto failure_out;
1125 // We have the option of creating a session and process group, or just a new process
1126 // group. If we just create a new process group, the child process cannot make itself
1127 // a session leader if it wants to do that (eg getty/login will generally want this).
1128 // If we do neither, and we are running with a controlling terminal, a ^C or similar
1129 // will also affect the child process (which probably isn't so bad, though since we
1130 // will handle the shutdown ourselves it's not necessary). Creating a new session
1131 // (and a new process group as part of that) seems like a safe bet, and has the
1132 // advantage of letting us signal the process as part of a process group.
1136 // "run on console" - run as a foreground job on the terminal/console device
1138 // if do_set_ctty is false, we are the session leader; we are probably running
1139 // as a user process. Don't create a new session leader in that case, and run
1140 // as part of the parent session. Otherwise, the new session cannot claim the
1141 // terminal as a controlling terminal (it is already claimed), meaning that it
1142 // will not see control signals from ^C etc.
1145 // Disable suspend (^Z) (and on some systems, delayed suspend / ^Y)
1146 signal(SIGTSTP, SIG_IGN);
1148 // Become session leader
1150 ioctl(0, TIOCSCTTY, 0);
1153 tcsetpgrp(0, getpgrp());
1156 sigprocmask(SIG_SETMASK, &sigwait_set, nullptr);
1158 execvp(args[0], const_cast<char **>(args));
1160 // If we got here, the exec failed:
1162 int exec_status = errno;
1163 write(wpipefd, &exec_status, sizeof(int));
1167 // Mark this and all dependent services as force-stopped.
1168 void service_record::forced_stop() noexcept
1170 if (service_state != service_state_t::STOPPED) {
1172 services->add_transition_queue(this);
1176 void service_record::dependent_stopped() noexcept
1178 if (service_state == service_state_t::STOPPING && waiting_for_deps) {
1179 services->add_transition_queue(this);
1183 void service_record::stop(bool bring_down) noexcept
1185 if (start_explicit) {
1186 start_explicit = false;
1195 void service_record::do_stop() noexcept
1197 if (pinned_started) return;
1199 if (start_explicit && ! do_auto_restart()) {
1200 start_explicit = false;
1202 if (required_by == 0) return; // release will re-call us anyway
1205 bool all_deps_stopped = stop_dependents();
1207 if (service_state != service_state_t::STARTED) {
1208 if (service_state == service_state_t::STARTING) {
1209 if (! can_interrupt_start()) {
1210 // Well this is awkward: we're going to have to continue starting. We can stop once we've
1211 // reached the started state.
1215 if (! interrupt_start()) {
1216 // Now wait for service startup to actually end; we don't need to handle it here.
1220 // We must have had desired_state == STARTED.
1221 notify_listeners(service_event_t::STARTCANCELLED);
1223 // Reaching this point, we are starting interruptibly - so we
1224 // stop now (by falling through to below).
1227 // If we're starting we need to wait for that to complete.
1228 // If we're already stopping/stopped there's nothing to do.
1233 service_state = service_state_t::STOPPING;
1234 waiting_for_deps = true;
1235 if (all_deps_stopped) {
1236 services->add_transition_queue(this);
1240 bool service_record::stop_check_dependents() noexcept
1242 bool all_deps_stopped = true;
1243 for (auto dept : dependents) {
1244 if (dept->dep_type == dependency_type::REGULAR && ! dept->get_from()->is_stopped()) {
1245 all_deps_stopped = false;
1250 return all_deps_stopped;
1253 bool service_record::stop_dependents() noexcept
1255 bool all_deps_stopped = true;
1256 for (auto dept : dependents) {
1257 if (dept->dep_type == dependency_type::REGULAR) {
1258 if (! dept->get_from()->is_stopped()) {
1259 // Note we check *first* since if the dependent service is not stopped,
1260 // 1. We will issue a stop to it shortly and
1261 // 2. It will notify us when stopped, at which point the stop_check_dependents()
1262 // check is run anyway.
1263 all_deps_stopped = false;
1267 // If this service is to be forcefully stopped, dependents must also be.
1268 dept->get_from()->forced_stop();
1271 dept->get_from()->prop_stop = true;
1272 services->add_prop_queue(dept->get_from());
1276 return all_deps_stopped;
1279 // All dependents have stopped; we can stop now, too. Only called when STOPPING.
1280 void service_record::bring_down() noexcept
1282 waiting_for_deps = false;
1286 void base_process_service::kill_pg(int signo) noexcept
1288 pid_t pgid = getpgid(pid);
1290 // only should happen if pid is invalid, which should never happen...
1291 log(loglevel_t::ERROR, get_name(), ": can't signal process: ", strerror(errno));
1297 void base_process_service::bring_down() noexcept
1299 waiting_for_deps = false;
1301 // The process is still kicking on - must actually kill it. We signal the process
1302 // group (-pid) rather than just the process as there's less risk then of creating
1303 // an orphaned process group:
1304 if (! onstart_flags.no_sigterm) {
1307 if (term_signal != -1) {
1308 kill_pg(term_signal);
1311 // In most cases, the rest is done in handle_exit_status.
1312 // If we are a BGPROCESS and the process is not our immediate child, however, that
1313 // won't work - check for this now:
1314 if (get_type() == service_type_t::BGPROCESS && ! tracking_child) {
1317 else if (stop_timeout != time_val(0,0)) {
1318 restart_timer.arm_timer_rel(eventLoop, stop_timeout);
1319 stop_timer_armed = true;
1323 // The process is already dead.
1328 void process_service::bring_down() noexcept
1330 waiting_for_deps = false;
1331 if (waiting_for_execstat) {
1332 // The process is still starting. This should be uncommon, but can occur during
1333 // smooth recovery. We can't do much now; we have to wait until we get the
1334 // status, and then act appropriately.
1337 else if (pid != -1) {
1338 // The process is still kicking on - must actually kill it. We signal the process
1339 // group (-pid) rather than just the process as there's less risk then of creating
1340 // an orphaned process group:
1341 if (! onstart_flags.no_sigterm) {
1344 if (term_signal != -1) {
1345 kill_pg(term_signal);
1348 // In most cases, the rest is done in handle_exit_status.
1349 // If we are a BGPROCESS and the process is not our immediate child, however, that
1350 // won't work - check for this now:
1351 if (get_type() == service_type_t::BGPROCESS && ! tracking_child) {
1354 else if (stop_timeout != time_val(0,0)) {
1355 restart_timer.arm_timer_rel(eventLoop, stop_timeout);
1356 stop_timer_armed = true;
1360 // The process is already dead.
1365 void scripted_service::bring_down() noexcept
1367 waiting_for_deps = false;
1368 if (stop_command.length() == 0) {
1371 else if (! start_ps_process(stop_arg_parts, false)) {
1372 // Couldn't execute stop script, but there's not much we can do:
1376 // successfully started stop script: start kill timer:
1377 if (stop_timeout != time_val(0,0)) {
1378 restart_timer.arm_timer_rel(eventLoop, stop_timeout);
1379 stop_timer_armed = true;
1384 void service_record::unpin() noexcept
1386 if (pinned_started) {
1387 pinned_started = false;
1388 if (desired_state == service_state_t::STOPPED || force_stop) {
1390 services->process_queues();
1393 if (pinned_stopped) {
1394 pinned_stopped = false;
1395 if (desired_state == service_state_t::STARTED) {
1397 services->process_queues();
1402 void service_record::queue_for_console() noexcept
1404 services->append_console_queue(this);
1407 void service_record::release_console() noexcept
1409 services->pull_console_queue();
1412 bool service_record::interrupt_start() noexcept
1414 services->unqueue_console(this);
1418 void service_set::service_active(service_record *sr) noexcept
1423 void service_set::service_inactive(service_record *sr) noexcept
1428 base_process_service::base_process_service(service_set *sset, string name,
1429 service_type_t service_type_p, string &&command,
1430 std::list<std::pair<unsigned,unsigned>> &command_offsets,
1431 const std::list<prelim_dep> &deplist_p)
1432 : service_record(sset, name, service_type_p, deplist_p), child_listener(this),
1433 child_status_listener(this), restart_timer(this)
1435 program_name = std::move(command);
1436 exec_arg_parts = separate_args(program_name, command_offsets);
1438 restart_interval_count = 0;
1439 restart_interval_time = {0, 0};
1440 restart_timer.service = this;
1441 restart_timer.add_timer(eventLoop);
1443 // By default, allow a maximum of 3 restarts within 10.0 seconds:
1444 restart_interval.seconds() = 10;
1445 restart_interval.nseconds() = 0;
1446 max_restart_interval_count = 3;
1448 waiting_restart_timer = false;
1449 reserved_child_watch = false;
1450 tracking_child = false;
1451 stop_timer_armed = false;
1452 start_is_interruptible = false;
1455 void base_process_service::do_restart() noexcept
1457 waiting_restart_timer = false;
1458 restart_interval_count++;
1459 auto service_state = get_state();
1461 // We may be STARTING (regular restart) or STARTED ("smooth recovery"). This affects whether
1462 // the process should be granted access to the console:
1463 bool on_console = service_state == service_state_t::STARTING
1464 ? onstart_flags.starts_on_console : onstart_flags.runs_on_console;
1466 if (service_state == service_state_t::STARTING) {
1467 // for a smooth recovery, we want to check dependencies are available before actually
1469 if (! check_deps_started()) {
1470 waiting_for_deps = true;
1475 if (! start_ps_process(exec_arg_parts, on_console)) {
1477 if (service_state == service_state_t::STARTING) {
1481 // desired_state = service_state_t::STOPPED;
1484 services->process_queues();
1488 bool base_process_service::restart_ps_process() noexcept
1490 using time_val = dasynq::time_val;
1492 time_val current_time;
1493 eventLoop.get_time(current_time, clock_type::MONOTONIC);
1495 if (max_restart_interval_count != 0) {
1496 // Check whether we're still in the most recent restart check interval:
1497 time_val int_diff = current_time - restart_interval_time;
1498 if (int_diff < restart_interval) {
1499 if (restart_interval_count >= max_restart_interval_count) {
1500 log(loglevel_t::ERROR, "Service ", get_name(), " restarting too quickly; stopping.");
1505 restart_interval_time = current_time;
1506 restart_interval_count = 0;
1510 // Check if enough time has lapsed since the prevous restart. If not, start a timer:
1511 time_val tdiff = current_time - last_start_time;
1512 if (restart_delay <= tdiff) {
1513 // > restart delay (normally 200ms)
1517 time_val timeout = restart_delay - tdiff;
1518 restart_timer.arm_timer_rel(eventLoop, timeout);
1519 waiting_restart_timer = true;
1524 bool base_process_service::interrupt_start() noexcept
1526 if (waiting_restart_timer) {
1527 restart_timer.stop_timer(eventLoop);
1528 waiting_restart_timer = false;
1529 return service_record::interrupt_start();
1532 log(loglevel_t::WARN, "Interrupting start of service ", get_name(), " with pid ", pid, " (with SIGINT).");
1534 if (stop_timeout != time_val(0,0)) {
1535 restart_timer.arm_timer(eventLoop, stop_timeout);
1536 stop_timer_armed = true;
1538 else if (stop_timer_armed) {
1539 restart_timer.stop_timer(eventLoop);
1540 stop_timer_armed = false;
1542 set_state(service_state_t::STOPPING);
1547 void base_process_service::kill_with_fire() noexcept
1550 log(loglevel_t::WARN, "Service ", get_name(), " with pid ", pid, " exceeded allowed stop time; killing.");
1555 dasynq::rearm process_restart_timer::timer_expiry(eventloop_t &, int expiry_count)
1557 service->stop_timer_armed = false;
1559 // Timer expires if:
1560 // We are stopping, including after having startup cancelled (stop timeout, state is STOPPING); We are
1561 // starting (start timeout, state is STARTING); We are waiting for restart timer before restarting,
1562 // including smooth recovery (restart timeout, state is STARTING or STARTED).
1563 if (service->get_state() == service_state_t::STOPPING) {
1564 service->kill_with_fire();
1566 else if (service->pid != -1) {
1567 // Starting, start timed out.
1568 service->interrupt_start();
1571 // STARTING / STARTED, and we have a pid: must be restarting (smooth recovery if STARTED)
1572 service->do_restart();
1575 // Leave the timer disabled, or, if it has been reset by any processing above, leave it armed:
1576 return dasynq::rearm::NOOP;