2 This file is part of systemd.
4 Copyright 2013 Lennart Poettering
6 systemd is free software; you can redistribute it and/or modify it
7 under the terms of the GNU Lesser General Public License as published by
8 the Free Software Foundation; either version 2.1 of the License, or
9 (at your option) any later version.
11 systemd is distributed in the hope that it will be useful, but
12 WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
16 You should have received a copy of the GNU Lesser General Public License
17 along with systemd; If not, see <http://www.gnu.org/licenses/>.
20 #include <sys/epoll.h>
21 #include <sys/timerfd.h>
24 #include "sd-daemon.h"
28 #include "alloc-util.h"
35 #include "process-util.h"
37 #include "signal-util.h"
38 #include "string-table.h"
39 #include "string-util.h"
40 #include "time-util.h"
43 #define DEFAULT_ACCURACY_USEC (250 * USEC_PER_MSEC)
45 typedef enum EventSourceType {
49 SOURCE_TIME_MONOTONIC,
50 SOURCE_TIME_REALTIME_ALARM,
51 SOURCE_TIME_BOOTTIME_ALARM,
58 _SOURCE_EVENT_SOURCE_TYPE_MAX,
59 _SOURCE_EVENT_SOURCE_TYPE_INVALID = -1
62 static const char* const event_source_type_table[_SOURCE_EVENT_SOURCE_TYPE_MAX] = {
64 [SOURCE_TIME_REALTIME] = "realtime",
65 [SOURCE_TIME_BOOTTIME] = "bootime",
66 [SOURCE_TIME_MONOTONIC] = "monotonic",
67 [SOURCE_TIME_REALTIME_ALARM] = "realtime-alarm",
68 [SOURCE_TIME_BOOTTIME_ALARM] = "boottime-alarm",
69 [SOURCE_SIGNAL] = "signal",
70 [SOURCE_CHILD] = "child",
71 [SOURCE_DEFER] = "defer",
72 [SOURCE_POST] = "post",
73 [SOURCE_EXIT] = "exit",
74 [SOURCE_WATCHDOG] = "watchdog",
77 DEFINE_PRIVATE_STRING_TABLE_LOOKUP_TO_STRING(event_source_type, int);
79 /* All objects we use in epoll events start with this value, so that
80 * we know how to dispatch it */
81 typedef enum WakeupType {
87 _WAKEUP_TYPE_INVALID = -1,
90 #define EVENT_SOURCE_IS_TIME(t) IN_SET((t), SOURCE_TIME_REALTIME, SOURCE_TIME_BOOTTIME, SOURCE_TIME_MONOTONIC, SOURCE_TIME_REALTIME_ALARM, SOURCE_TIME_BOOTTIME_ALARM)
92 struct sd_event_source {
99 sd_event_handler_t prepare;
103 EventSourceType type:5;
110 unsigned pending_index;
111 unsigned prepare_index;
112 unsigned pending_iteration;
113 unsigned prepare_iteration;
115 LIST_FIELDS(sd_event_source, sources);
119 sd_event_io_handler_t callback;
126 sd_event_time_handler_t callback;
127 usec_t next, accuracy;
128 unsigned earliest_index;
129 unsigned latest_index;
132 sd_event_signal_handler_t callback;
133 struct signalfd_siginfo siginfo;
137 sd_event_child_handler_t callback;
143 sd_event_handler_t callback;
146 sd_event_handler_t callback;
149 sd_event_handler_t callback;
150 unsigned prioq_index;
159 /* For all clocks we maintain two priority queues each, one
160 * ordered for the earliest times the events may be
161 * dispatched, and one ordered by the latest times they must
162 * have been dispatched. The range between the top entries in
163 * the two prioqs is the time window we can freely schedule
176 /* For each priority we maintain one signal fd, so that we
177 * only have to dequeue a single event per priority at a
183 sd_event_source *current;
195 /* timerfd_create() only supports these five clocks so far. We
196 * can add support for more clocks when the kernel learns to
197 * deal with them, too. */
198 struct clock_data realtime;
199 struct clock_data boottime;
200 struct clock_data monotonic;
201 struct clock_data realtime_alarm;
202 struct clock_data boottime_alarm;
206 sd_event_source **signal_sources; /* indexed by signal number */
207 Hashmap *signal_data; /* indexed by priority */
209 Hashmap *child_sources;
210 unsigned n_enabled_child_sources;
219 dual_timestamp timestamp;
220 usec_t timestamp_boottime;
223 bool exit_requested:1;
224 bool need_process_child:1;
226 bool profile_delays:1;
231 sd_event **default_event_ptr;
233 usec_t watchdog_last, watchdog_period;
237 LIST_HEAD(sd_event_source, sources);
239 usec_t last_run, last_log;
240 unsigned delays[sizeof(usec_t) * 8];
243 static void source_disconnect(sd_event_source *s);
245 static int pending_prioq_compare(const void *a, const void *b) {
246 const sd_event_source *x = a, *y = b;
251 /* Enabled ones first */
252 if (x->enabled != SD_EVENT_OFF && y->enabled == SD_EVENT_OFF)
254 if (x->enabled == SD_EVENT_OFF && y->enabled != SD_EVENT_OFF)
257 /* Lower priority values first */
258 if (x->priority < y->priority)
260 if (x->priority > y->priority)
263 /* Older entries first */
264 if (x->pending_iteration < y->pending_iteration)
266 if (x->pending_iteration > y->pending_iteration)
272 static int prepare_prioq_compare(const void *a, const void *b) {
273 const sd_event_source *x = a, *y = b;
278 /* Enabled ones first */
279 if (x->enabled != SD_EVENT_OFF && y->enabled == SD_EVENT_OFF)
281 if (x->enabled == SD_EVENT_OFF && y->enabled != SD_EVENT_OFF)
284 /* Move most recently prepared ones last, so that we can stop
285 * preparing as soon as we hit one that has already been
286 * prepared in the current iteration */
287 if (x->prepare_iteration < y->prepare_iteration)
289 if (x->prepare_iteration > y->prepare_iteration)
292 /* Lower priority values first */
293 if (x->priority < y->priority)
295 if (x->priority > y->priority)
301 static int earliest_time_prioq_compare(const void *a, const void *b) {
302 const sd_event_source *x = a, *y = b;
304 assert(EVENT_SOURCE_IS_TIME(x->type));
305 assert(x->type == y->type);
307 /* Enabled ones first */
308 if (x->enabled != SD_EVENT_OFF && y->enabled == SD_EVENT_OFF)
310 if (x->enabled == SD_EVENT_OFF && y->enabled != SD_EVENT_OFF)
313 /* Move the pending ones to the end */
314 if (!x->pending && y->pending)
316 if (x->pending && !y->pending)
320 if (x->time.next < y->time.next)
322 if (x->time.next > y->time.next)
328 static usec_t time_event_source_latest(const sd_event_source *s) {
329 return usec_add(s->time.next, s->time.accuracy);
332 static int latest_time_prioq_compare(const void *a, const void *b) {
333 const sd_event_source *x = a, *y = b;
335 assert(EVENT_SOURCE_IS_TIME(x->type));
336 assert(x->type == y->type);
338 /* Enabled ones first */
339 if (x->enabled != SD_EVENT_OFF && y->enabled == SD_EVENT_OFF)
341 if (x->enabled == SD_EVENT_OFF && y->enabled != SD_EVENT_OFF)
344 /* Move the pending ones to the end */
345 if (!x->pending && y->pending)
347 if (x->pending && !y->pending)
351 if (time_event_source_latest(x) < time_event_source_latest(y))
353 if (time_event_source_latest(x) > time_event_source_latest(y))
359 static int exit_prioq_compare(const void *a, const void *b) {
360 const sd_event_source *x = a, *y = b;
362 assert(x->type == SOURCE_EXIT);
363 assert(y->type == SOURCE_EXIT);
365 /* Enabled ones first */
366 if (x->enabled != SD_EVENT_OFF && y->enabled == SD_EVENT_OFF)
368 if (x->enabled == SD_EVENT_OFF && y->enabled != SD_EVENT_OFF)
371 /* Lower priority values first */
372 if (x->priority < y->priority)
374 if (x->priority > y->priority)
380 static void free_clock_data(struct clock_data *d) {
382 assert(d->wakeup == WAKEUP_CLOCK_DATA);
385 prioq_free(d->earliest);
386 prioq_free(d->latest);
389 static void event_free(sd_event *e) {
394 while ((s = e->sources)) {
396 source_disconnect(s);
397 sd_event_source_unref(s);
400 assert(e->n_sources == 0);
402 if (e->default_event_ptr)
403 *(e->default_event_ptr) = NULL;
405 safe_close(e->epoll_fd);
406 safe_close(e->watchdog_fd);
408 free_clock_data(&e->realtime);
409 free_clock_data(&e->boottime);
410 free_clock_data(&e->monotonic);
411 free_clock_data(&e->realtime_alarm);
412 free_clock_data(&e->boottime_alarm);
414 prioq_free(e->pending);
415 prioq_free(e->prepare);
418 free(e->signal_sources);
419 hashmap_free(e->signal_data);
421 hashmap_free(e->child_sources);
422 set_free(e->post_sources);
426 _public_ int sd_event_new(sd_event** ret) {
430 assert_return(ret, -EINVAL);
432 e = new0(sd_event, 1);
437 e->watchdog_fd = e->epoll_fd = e->realtime.fd = e->boottime.fd = e->monotonic.fd = e->realtime_alarm.fd = e->boottime_alarm.fd = -1;
438 e->realtime.next = e->boottime.next = e->monotonic.next = e->realtime_alarm.next = e->boottime_alarm.next = USEC_INFINITY;
439 e->realtime.wakeup = e->boottime.wakeup = e->monotonic.wakeup = e->realtime_alarm.wakeup = e->boottime_alarm.wakeup = WAKEUP_CLOCK_DATA;
440 e->original_pid = getpid();
441 e->perturb = USEC_INFINITY;
443 r = prioq_ensure_allocated(&e->pending, pending_prioq_compare);
447 e->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
448 if (e->epoll_fd < 0) {
453 if (secure_getenv("SD_EVENT_PROFILE_DELAYS")) {
454 log_debug("Event loop profiling enabled. Logarithmic histogram of event loop iterations in the range 2^0 ... 2^63 us will be logged every 5s.");
455 e->profile_delays = true;
466 _public_ sd_event* sd_event_ref(sd_event *e) {
471 assert(e->n_ref >= 1);
477 _public_ sd_event* sd_event_unref(sd_event *e) {
482 assert(e->n_ref >= 1);
491 static bool event_pid_changed(sd_event *e) {
494 /* We don't support people creating an event loop and keeping
495 * it around over a fork(). Let's complain. */
497 return e->original_pid != getpid();
500 static void source_io_unregister(sd_event_source *s) {
504 assert(s->type == SOURCE_IO);
506 if (event_pid_changed(s->event))
509 if (!s->io.registered)
512 r = epoll_ctl(s->event->epoll_fd, EPOLL_CTL_DEL, s->io.fd, NULL);
514 log_debug_errno(errno, "Failed to remove source %s (type %s) from epoll: %m",
515 strna(s->description), event_source_type_to_string(s->type));
517 s->io.registered = false;
520 static int source_io_register(
525 struct epoll_event ev = {};
529 assert(s->type == SOURCE_IO);
530 assert(enabled != SD_EVENT_OFF);
535 if (enabled == SD_EVENT_ONESHOT)
536 ev.events |= EPOLLONESHOT;
538 if (s->io.registered)
539 r = epoll_ctl(s->event->epoll_fd, EPOLL_CTL_MOD, s->io.fd, &ev);
541 r = epoll_ctl(s->event->epoll_fd, EPOLL_CTL_ADD, s->io.fd, &ev);
545 s->io.registered = true;
550 #if 0 /// UNNEEDED by elogind
551 static clockid_t event_source_type_to_clock(EventSourceType t) {
555 case SOURCE_TIME_REALTIME:
556 return CLOCK_REALTIME;
558 case SOURCE_TIME_BOOTTIME:
559 return CLOCK_BOOTTIME;
561 case SOURCE_TIME_MONOTONIC:
562 return CLOCK_MONOTONIC;
564 case SOURCE_TIME_REALTIME_ALARM:
565 return CLOCK_REALTIME_ALARM;
567 case SOURCE_TIME_BOOTTIME_ALARM:
568 return CLOCK_BOOTTIME_ALARM;
571 return (clockid_t) -1;
576 static EventSourceType clock_to_event_source_type(clockid_t clock) {
581 return SOURCE_TIME_REALTIME;
584 return SOURCE_TIME_BOOTTIME;
586 case CLOCK_MONOTONIC:
587 return SOURCE_TIME_MONOTONIC;
589 case CLOCK_REALTIME_ALARM:
590 return SOURCE_TIME_REALTIME_ALARM;
592 case CLOCK_BOOTTIME_ALARM:
593 return SOURCE_TIME_BOOTTIME_ALARM;
596 return _SOURCE_EVENT_SOURCE_TYPE_INVALID;
600 static struct clock_data* event_get_clock_data(sd_event *e, EventSourceType t) {
605 case SOURCE_TIME_REALTIME:
608 case SOURCE_TIME_BOOTTIME:
611 case SOURCE_TIME_MONOTONIC:
612 return &e->monotonic;
614 case SOURCE_TIME_REALTIME_ALARM:
615 return &e->realtime_alarm;
617 case SOURCE_TIME_BOOTTIME_ALARM:
618 return &e->boottime_alarm;
625 static int event_make_signal_data(
628 struct signal_data **ret) {
630 struct epoll_event ev = {};
631 struct signal_data *d;
639 if (event_pid_changed(e))
642 if (e->signal_sources && e->signal_sources[sig])
643 priority = e->signal_sources[sig]->priority;
647 d = hashmap_get(e->signal_data, &priority);
649 if (sigismember(&d->sigset, sig) > 0) {
655 r = hashmap_ensure_allocated(&e->signal_data, &uint64_hash_ops);
659 d = new0(struct signal_data, 1);
663 d->wakeup = WAKEUP_SIGNAL_DATA;
665 d->priority = priority;
667 r = hashmap_put(e->signal_data, &d->priority, d);
677 assert_se(sigaddset(&ss_copy, sig) >= 0);
679 r = signalfd(d->fd, &ss_copy, SFD_NONBLOCK|SFD_CLOEXEC);
698 r = epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, d->fd, &ev);
711 d->fd = safe_close(d->fd);
712 hashmap_remove(e->signal_data, &d->priority);
719 static void event_unmask_signal_data(sd_event *e, struct signal_data *d, int sig) {
723 /* Turns off the specified signal in the signal data
724 * object. If the signal mask of the object becomes empty that
727 if (sigismember(&d->sigset, sig) == 0)
730 assert_se(sigdelset(&d->sigset, sig) >= 0);
732 if (sigisemptyset(&d->sigset)) {
734 /* If all the mask is all-zero we can get rid of the structure */
735 hashmap_remove(e->signal_data, &d->priority);
744 if (signalfd(d->fd, &d->sigset, SFD_NONBLOCK|SFD_CLOEXEC) < 0)
745 log_debug_errno(errno, "Failed to unset signal bit, ignoring: %m");
748 static void event_gc_signal_data(sd_event *e, const int64_t *priority, int sig) {
749 struct signal_data *d;
750 static const int64_t zero_priority = 0;
754 /* Rechecks if the specified signal is still something we are
755 * interested in. If not, we'll unmask it, and possibly drop
756 * the signalfd for it. */
758 if (sig == SIGCHLD &&
759 e->n_enabled_child_sources > 0)
762 if (e->signal_sources &&
763 e->signal_sources[sig] &&
764 e->signal_sources[sig]->enabled != SD_EVENT_OFF)
768 * The specified signal might be enabled in three different queues:
770 * 1) the one that belongs to the priority passed (if it is non-NULL)
771 * 2) the one that belongs to the priority of the event source of the signal (if there is one)
772 * 3) the 0 priority (to cover the SIGCHLD case)
774 * Hence, let's remove it from all three here.
778 d = hashmap_get(e->signal_data, priority);
780 event_unmask_signal_data(e, d, sig);
783 if (e->signal_sources && e->signal_sources[sig]) {
784 d = hashmap_get(e->signal_data, &e->signal_sources[sig]->priority);
786 event_unmask_signal_data(e, d, sig);
789 d = hashmap_get(e->signal_data, &zero_priority);
791 event_unmask_signal_data(e, d, sig);
794 static void source_disconnect(sd_event_source *s) {
802 assert(s->event->n_sources > 0);
808 source_io_unregister(s);
812 case SOURCE_TIME_REALTIME:
813 case SOURCE_TIME_BOOTTIME:
814 case SOURCE_TIME_MONOTONIC:
815 case SOURCE_TIME_REALTIME_ALARM:
816 case SOURCE_TIME_BOOTTIME_ALARM: {
817 struct clock_data *d;
819 d = event_get_clock_data(s->event, s->type);
822 prioq_remove(d->earliest, s, &s->time.earliest_index);
823 prioq_remove(d->latest, s, &s->time.latest_index);
824 d->needs_rearm = true;
829 if (s->signal.sig > 0) {
831 if (s->event->signal_sources)
832 s->event->signal_sources[s->signal.sig] = NULL;
834 event_gc_signal_data(s->event, &s->priority, s->signal.sig);
840 if (s->child.pid > 0) {
841 if (s->enabled != SD_EVENT_OFF) {
842 assert(s->event->n_enabled_child_sources > 0);
843 s->event->n_enabled_child_sources--;
846 (void) hashmap_remove(s->event->child_sources, PID_TO_PTR(s->child.pid));
847 event_gc_signal_data(s->event, &s->priority, SIGCHLD);
857 set_remove(s->event->post_sources, s);
861 prioq_remove(s->event->exit, s, &s->exit.prioq_index);
865 assert_not_reached("Wut? I shouldn't exist.");
869 prioq_remove(s->event->pending, s, &s->pending_index);
872 prioq_remove(s->event->prepare, s, &s->prepare_index);
876 s->type = _SOURCE_EVENT_SOURCE_TYPE_INVALID;
878 LIST_REMOVE(sources, event->sources, s);
882 sd_event_unref(event);
885 static void source_free(sd_event_source *s) {
888 source_disconnect(s);
889 free(s->description);
893 static int source_set_pending(sd_event_source *s, bool b) {
897 assert(s->type != SOURCE_EXIT);
905 s->pending_iteration = s->event->iteration;
907 r = prioq_put(s->event->pending, s, &s->pending_index);
913 assert_se(prioq_remove(s->event->pending, s, &s->pending_index));
915 if (EVENT_SOURCE_IS_TIME(s->type)) {
916 struct clock_data *d;
918 d = event_get_clock_data(s->event, s->type);
921 prioq_reshuffle(d->earliest, s, &s->time.earliest_index);
922 prioq_reshuffle(d->latest, s, &s->time.latest_index);
923 d->needs_rearm = true;
926 if (s->type == SOURCE_SIGNAL && !b) {
927 struct signal_data *d;
929 d = hashmap_get(s->event->signal_data, &s->priority);
930 if (d && d->current == s)
937 static sd_event_source *source_new(sd_event *e, bool floating, EventSourceType type) {
942 s = new0(sd_event_source, 1);
948 s->floating = floating;
950 s->pending_index = s->prepare_index = PRIOQ_IDX_NULL;
955 LIST_PREPEND(sources, e->sources, s);
961 _public_ int sd_event_add_io(
963 sd_event_source **ret,
966 sd_event_io_handler_t callback,
972 assert_return(e, -EINVAL);
973 assert_return(fd >= 0, -EBADF);
974 assert_return(!(events & ~(EPOLLIN|EPOLLOUT|EPOLLRDHUP|EPOLLPRI|EPOLLERR|EPOLLHUP|EPOLLET)), -EINVAL);
975 assert_return(callback, -EINVAL);
976 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
977 assert_return(!event_pid_changed(e), -ECHILD);
979 s = source_new(e, !ret, SOURCE_IO);
983 s->wakeup = WAKEUP_EVENT_SOURCE;
985 s->io.events = events;
986 s->io.callback = callback;
987 s->userdata = userdata;
988 s->enabled = SD_EVENT_ON;
990 r = source_io_register(s, s->enabled, events);
1002 static void initialize_perturb(sd_event *e) {
1003 sd_id128_t bootid = {};
1005 /* When we sleep for longer, we try to realign the wakeup to
1006 the same time wihtin each minute/second/250ms, so that
1007 events all across the system can be coalesced into a single
1008 CPU wakeup. However, let's take some system-specific
1009 randomness for this value, so that in a network of systems
1010 with synced clocks timer events are distributed a
1011 bit. Here, we calculate a perturbation usec offset from the
1014 if (_likely_(e->perturb != USEC_INFINITY))
1017 if (sd_id128_get_boot(&bootid) >= 0)
1018 e->perturb = (bootid.qwords[0] ^ bootid.qwords[1]) % USEC_PER_MINUTE;
1021 static int event_setup_timer_fd(
1023 struct clock_data *d,
1026 struct epoll_event ev = {};
1032 if (_likely_(d->fd >= 0))
1035 fd = timerfd_create(clock, TFD_NONBLOCK|TFD_CLOEXEC);
1039 ev.events = EPOLLIN;
1042 r = epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, fd, &ev);
1052 static int time_exit_callback(sd_event_source *s, uint64_t usec, void *userdata) {
1055 return sd_event_exit(sd_event_source_get_event(s), PTR_TO_INT(userdata));
1058 _public_ int sd_event_add_time(
1060 sd_event_source **ret,
1064 sd_event_time_handler_t callback,
1067 EventSourceType type;
1069 struct clock_data *d;
1072 assert_return(e, -EINVAL);
1073 assert_return(accuracy != (uint64_t) -1, -EINVAL);
1074 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1075 assert_return(!event_pid_changed(e), -ECHILD);
1078 callback = time_exit_callback;
1080 type = clock_to_event_source_type(clock);
1081 assert_return(type >= 0, -EOPNOTSUPP);
1083 d = event_get_clock_data(e, type);
1086 r = prioq_ensure_allocated(&d->earliest, earliest_time_prioq_compare);
1090 r = prioq_ensure_allocated(&d->latest, latest_time_prioq_compare);
1095 r = event_setup_timer_fd(e, d, clock);
1100 s = source_new(e, !ret, type);
1104 s->time.next = usec;
1105 s->time.accuracy = accuracy == 0 ? DEFAULT_ACCURACY_USEC : accuracy;
1106 s->time.callback = callback;
1107 s->time.earliest_index = s->time.latest_index = PRIOQ_IDX_NULL;
1108 s->userdata = userdata;
1109 s->enabled = SD_EVENT_ONESHOT;
1111 d->needs_rearm = true;
1113 r = prioq_put(d->earliest, s, &s->time.earliest_index);
1117 r = prioq_put(d->latest, s, &s->time.latest_index);
1131 static int signal_exit_callback(sd_event_source *s, const struct signalfd_siginfo *si, void *userdata) {
1134 return sd_event_exit(sd_event_source_get_event(s), PTR_TO_INT(userdata));
1137 _public_ int sd_event_add_signal(
1139 sd_event_source **ret,
1141 sd_event_signal_handler_t callback,
1145 struct signal_data *d;
1149 assert_return(e, -EINVAL);
1150 assert_return(sig > 0, -EINVAL);
1151 assert_return(sig < _NSIG, -EINVAL);
1152 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1153 assert_return(!event_pid_changed(e), -ECHILD);
1156 callback = signal_exit_callback;
1158 r = pthread_sigmask(SIG_SETMASK, NULL, &ss);
1162 if (!sigismember(&ss, sig))
1165 if (!e->signal_sources) {
1166 e->signal_sources = new0(sd_event_source*, _NSIG);
1167 if (!e->signal_sources)
1169 } else if (e->signal_sources[sig])
1172 s = source_new(e, !ret, SOURCE_SIGNAL);
1176 s->signal.sig = sig;
1177 s->signal.callback = callback;
1178 s->userdata = userdata;
1179 s->enabled = SD_EVENT_ON;
1181 e->signal_sources[sig] = s;
1183 r = event_make_signal_data(e, sig, &d);
1189 /* Use the signal name as description for the event source by default */
1190 (void) sd_event_source_set_description(s, signal_to_string(sig));
1198 #if 0 /// UNNEEDED by elogind
1199 _public_ int sd_event_add_child(
1201 sd_event_source **ret,
1204 sd_event_child_handler_t callback,
1210 assert_return(e, -EINVAL);
1211 assert_return(pid > 1, -EINVAL);
1212 assert_return(!(options & ~(WEXITED|WSTOPPED|WCONTINUED)), -EINVAL);
1213 assert_return(options != 0, -EINVAL);
1214 assert_return(callback, -EINVAL);
1215 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1216 assert_return(!event_pid_changed(e), -ECHILD);
1218 r = hashmap_ensure_allocated(&e->child_sources, NULL);
1222 if (hashmap_contains(e->child_sources, PID_TO_PTR(pid)))
1225 s = source_new(e, !ret, SOURCE_CHILD);
1230 s->child.options = options;
1231 s->child.callback = callback;
1232 s->userdata = userdata;
1233 s->enabled = SD_EVENT_ONESHOT;
1235 r = hashmap_put(e->child_sources, PID_TO_PTR(pid), s);
1241 e->n_enabled_child_sources ++;
1243 r = event_make_signal_data(e, SIGCHLD, NULL);
1245 e->n_enabled_child_sources--;
1250 e->need_process_child = true;
1258 _public_ int sd_event_add_defer(
1260 sd_event_source **ret,
1261 sd_event_handler_t callback,
1267 assert_return(e, -EINVAL);
1268 assert_return(callback, -EINVAL);
1269 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1270 assert_return(!event_pid_changed(e), -ECHILD);
1272 s = source_new(e, !ret, SOURCE_DEFER);
1276 s->defer.callback = callback;
1277 s->userdata = userdata;
1278 s->enabled = SD_EVENT_ONESHOT;
1280 r = source_set_pending(s, true);
1293 _public_ int sd_event_add_post(
1295 sd_event_source **ret,
1296 sd_event_handler_t callback,
1302 assert_return(e, -EINVAL);
1303 assert_return(callback, -EINVAL);
1304 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1305 assert_return(!event_pid_changed(e), -ECHILD);
1307 r = set_ensure_allocated(&e->post_sources, NULL);
1311 s = source_new(e, !ret, SOURCE_POST);
1315 s->post.callback = callback;
1316 s->userdata = userdata;
1317 s->enabled = SD_EVENT_ON;
1319 r = set_put(e->post_sources, s);
1331 _public_ int sd_event_add_exit(
1333 sd_event_source **ret,
1334 sd_event_handler_t callback,
1340 assert_return(e, -EINVAL);
1341 assert_return(callback, -EINVAL);
1342 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1343 assert_return(!event_pid_changed(e), -ECHILD);
1345 r = prioq_ensure_allocated(&e->exit, exit_prioq_compare);
1349 s = source_new(e, !ret, SOURCE_EXIT);
1353 s->exit.callback = callback;
1354 s->userdata = userdata;
1355 s->exit.prioq_index = PRIOQ_IDX_NULL;
1356 s->enabled = SD_EVENT_ONESHOT;
1358 r = prioq_put(s->event->exit, s, &s->exit.prioq_index);
1370 #if 0 /// UNNEEDED by elogind
1371 _public_ sd_event_source* sd_event_source_ref(sd_event_source *s) {
1376 assert(s->n_ref >= 1);
1383 _public_ sd_event_source* sd_event_source_unref(sd_event_source *s) {
1388 assert(s->n_ref >= 1);
1391 if (s->n_ref <= 0) {
1392 /* Here's a special hack: when we are called from a
1393 * dispatch handler we won't free the event source
1394 * immediately, but we will detach the fd from the
1395 * epoll. This way it is safe for the caller to unref
1396 * the event source and immediately close the fd, but
1397 * we still retain a valid event source object after
1400 if (s->dispatching) {
1401 if (s->type == SOURCE_IO)
1402 source_io_unregister(s);
1404 source_disconnect(s);
1412 _public_ int sd_event_source_set_description(sd_event_source *s, const char *description) {
1413 assert_return(s, -EINVAL);
1414 assert_return(!event_pid_changed(s->event), -ECHILD);
1416 return free_and_strdup(&s->description, description);
1419 #if 0 /// UNNEEDED by elogind
1420 _public_ int sd_event_source_get_description(sd_event_source *s, const char **description) {
1421 assert_return(s, -EINVAL);
1422 assert_return(description, -EINVAL);
1423 assert_return(s->description, -ENXIO);
1424 assert_return(!event_pid_changed(s->event), -ECHILD);
1426 *description = s->description;
1431 _public_ sd_event *sd_event_source_get_event(sd_event_source *s) {
1432 assert_return(s, NULL);
1437 #if 0 /// UNNEEDED by elogind
1438 _public_ int sd_event_source_get_pending(sd_event_source *s) {
1439 assert_return(s, -EINVAL);
1440 assert_return(s->type != SOURCE_EXIT, -EDOM);
1441 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1442 assert_return(!event_pid_changed(s->event), -ECHILD);
1447 _public_ int sd_event_source_get_io_fd(sd_event_source *s) {
1448 assert_return(s, -EINVAL);
1449 assert_return(s->type == SOURCE_IO, -EDOM);
1450 assert_return(!event_pid_changed(s->event), -ECHILD);
1456 _public_ int sd_event_source_set_io_fd(sd_event_source *s, int fd) {
1459 assert_return(s, -EINVAL);
1460 assert_return(fd >= 0, -EBADF);
1461 assert_return(s->type == SOURCE_IO, -EDOM);
1462 assert_return(!event_pid_changed(s->event), -ECHILD);
1467 if (s->enabled == SD_EVENT_OFF) {
1469 s->io.registered = false;
1473 saved_fd = s->io.fd;
1474 assert(s->io.registered);
1477 s->io.registered = false;
1479 r = source_io_register(s, s->enabled, s->io.events);
1481 s->io.fd = saved_fd;
1482 s->io.registered = true;
1486 epoll_ctl(s->event->epoll_fd, EPOLL_CTL_DEL, saved_fd, NULL);
1492 #if 0 /// UNNEEDED by elogind
1493 _public_ int sd_event_source_get_io_events(sd_event_source *s, uint32_t* events) {
1494 assert_return(s, -EINVAL);
1495 assert_return(events, -EINVAL);
1496 assert_return(s->type == SOURCE_IO, -EDOM);
1497 assert_return(!event_pid_changed(s->event), -ECHILD);
1499 *events = s->io.events;
1504 _public_ int sd_event_source_set_io_events(sd_event_source *s, uint32_t events) {
1507 assert_return(s, -EINVAL);
1508 assert_return(s->type == SOURCE_IO, -EDOM);
1509 assert_return(!(events & ~(EPOLLIN|EPOLLOUT|EPOLLRDHUP|EPOLLPRI|EPOLLERR|EPOLLHUP|EPOLLET)), -EINVAL);
1510 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1511 assert_return(!event_pid_changed(s->event), -ECHILD);
1513 /* edge-triggered updates are never skipped, so we can reset edges */
1514 if (s->io.events == events && !(events & EPOLLET))
1517 if (s->enabled != SD_EVENT_OFF) {
1518 r = source_io_register(s, s->enabled, events);
1523 s->io.events = events;
1524 source_set_pending(s, false);
1529 #if 0 /// UNNEEDED by elogind
1530 _public_ int sd_event_source_get_io_revents(sd_event_source *s, uint32_t* revents) {
1531 assert_return(s, -EINVAL);
1532 assert_return(revents, -EINVAL);
1533 assert_return(s->type == SOURCE_IO, -EDOM);
1534 assert_return(s->pending, -ENODATA);
1535 assert_return(!event_pid_changed(s->event), -ECHILD);
1537 *revents = s->io.revents;
1541 _public_ int sd_event_source_get_signal(sd_event_source *s) {
1542 assert_return(s, -EINVAL);
1543 assert_return(s->type == SOURCE_SIGNAL, -EDOM);
1544 assert_return(!event_pid_changed(s->event), -ECHILD);
1546 return s->signal.sig;
1549 _public_ int sd_event_source_get_priority(sd_event_source *s, int64_t *priority) {
1550 assert_return(s, -EINVAL);
1551 assert_return(!event_pid_changed(s->event), -ECHILD);
1557 _public_ int sd_event_source_set_priority(sd_event_source *s, int64_t priority) {
1560 assert_return(s, -EINVAL);
1561 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1562 assert_return(!event_pid_changed(s->event), -ECHILD);
1564 if (s->priority == priority)
1567 if (s->type == SOURCE_SIGNAL && s->enabled != SD_EVENT_OFF) {
1568 struct signal_data *old, *d;
1570 /* Move us from the signalfd belonging to the old
1571 * priority to the signalfd of the new priority */
1573 assert_se(old = hashmap_get(s->event->signal_data, &s->priority));
1575 s->priority = priority;
1577 r = event_make_signal_data(s->event, s->signal.sig, &d);
1579 s->priority = old->priority;
1583 event_unmask_signal_data(s->event, old, s->signal.sig);
1585 s->priority = priority;
1588 prioq_reshuffle(s->event->pending, s, &s->pending_index);
1591 prioq_reshuffle(s->event->prepare, s, &s->prepare_index);
1593 if (s->type == SOURCE_EXIT)
1594 prioq_reshuffle(s->event->exit, s, &s->exit.prioq_index);
1599 #if 0 /// UNNEEDED by elogind
1600 _public_ int sd_event_source_get_enabled(sd_event_source *s, int *m) {
1601 assert_return(s, -EINVAL);
1602 assert_return(m, -EINVAL);
1603 assert_return(!event_pid_changed(s->event), -ECHILD);
1610 _public_ int sd_event_source_set_enabled(sd_event_source *s, int m) {
1613 assert_return(s, -EINVAL);
1614 assert_return(m == SD_EVENT_OFF || m == SD_EVENT_ON || m == SD_EVENT_ONESHOT, -EINVAL);
1615 assert_return(!event_pid_changed(s->event), -ECHILD);
1617 /* If we are dead anyway, we are fine with turning off
1618 * sources, but everything else needs to fail. */
1619 if (s->event->state == SD_EVENT_FINISHED)
1620 return m == SD_EVENT_OFF ? 0 : -ESTALE;
1622 if (s->enabled == m)
1625 if (m == SD_EVENT_OFF) {
1630 source_io_unregister(s);
1634 case SOURCE_TIME_REALTIME:
1635 case SOURCE_TIME_BOOTTIME:
1636 case SOURCE_TIME_MONOTONIC:
1637 case SOURCE_TIME_REALTIME_ALARM:
1638 case SOURCE_TIME_BOOTTIME_ALARM: {
1639 struct clock_data *d;
1642 d = event_get_clock_data(s->event, s->type);
1645 prioq_reshuffle(d->earliest, s, &s->time.earliest_index);
1646 prioq_reshuffle(d->latest, s, &s->time.latest_index);
1647 d->needs_rearm = true;
1654 event_gc_signal_data(s->event, &s->priority, s->signal.sig);
1660 assert(s->event->n_enabled_child_sources > 0);
1661 s->event->n_enabled_child_sources--;
1663 event_gc_signal_data(s->event, &s->priority, SIGCHLD);
1668 prioq_reshuffle(s->event->exit, s, &s->exit.prioq_index);
1677 assert_not_reached("Wut? I shouldn't exist.");
1684 r = source_io_register(s, m, s->io.events);
1691 case SOURCE_TIME_REALTIME:
1692 case SOURCE_TIME_BOOTTIME:
1693 case SOURCE_TIME_MONOTONIC:
1694 case SOURCE_TIME_REALTIME_ALARM:
1695 case SOURCE_TIME_BOOTTIME_ALARM: {
1696 struct clock_data *d;
1699 d = event_get_clock_data(s->event, s->type);
1702 prioq_reshuffle(d->earliest, s, &s->time.earliest_index);
1703 prioq_reshuffle(d->latest, s, &s->time.latest_index);
1704 d->needs_rearm = true;
1712 r = event_make_signal_data(s->event, s->signal.sig, NULL);
1714 s->enabled = SD_EVENT_OFF;
1715 event_gc_signal_data(s->event, &s->priority, s->signal.sig);
1723 if (s->enabled == SD_EVENT_OFF)
1724 s->event->n_enabled_child_sources++;
1728 r = event_make_signal_data(s->event, SIGCHLD, NULL);
1730 s->enabled = SD_EVENT_OFF;
1731 s->event->n_enabled_child_sources--;
1732 event_gc_signal_data(s->event, &s->priority, SIGCHLD);
1740 prioq_reshuffle(s->event->exit, s, &s->exit.prioq_index);
1749 assert_not_reached("Wut? I shouldn't exist.");
1754 prioq_reshuffle(s->event->pending, s, &s->pending_index);
1757 prioq_reshuffle(s->event->prepare, s, &s->prepare_index);
1762 _public_ int sd_event_source_get_time(sd_event_source *s, uint64_t *usec) {
1763 assert_return(s, -EINVAL);
1764 assert_return(usec, -EINVAL);
1765 assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
1766 assert_return(!event_pid_changed(s->event), -ECHILD);
1768 *usec = s->time.next;
1772 _public_ int sd_event_source_set_time(sd_event_source *s, uint64_t usec) {
1773 struct clock_data *d;
1775 assert_return(s, -EINVAL);
1776 assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
1777 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1778 assert_return(!event_pid_changed(s->event), -ECHILD);
1780 s->time.next = usec;
1782 source_set_pending(s, false);
1784 d = event_get_clock_data(s->event, s->type);
1787 prioq_reshuffle(d->earliest, s, &s->time.earliest_index);
1788 prioq_reshuffle(d->latest, s, &s->time.latest_index);
1789 d->needs_rearm = true;
1794 #if 0 /// UNNEEDED by elogind
1795 _public_ int sd_event_source_get_time_accuracy(sd_event_source *s, uint64_t *usec) {
1796 assert_return(s, -EINVAL);
1797 assert_return(usec, -EINVAL);
1798 assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
1799 assert_return(!event_pid_changed(s->event), -ECHILD);
1801 *usec = s->time.accuracy;
1805 _public_ int sd_event_source_set_time_accuracy(sd_event_source *s, uint64_t usec) {
1806 struct clock_data *d;
1808 assert_return(s, -EINVAL);
1809 assert_return(usec != (uint64_t) -1, -EINVAL);
1810 assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
1811 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1812 assert_return(!event_pid_changed(s->event), -ECHILD);
1815 usec = DEFAULT_ACCURACY_USEC;
1817 s->time.accuracy = usec;
1819 source_set_pending(s, false);
1821 d = event_get_clock_data(s->event, s->type);
1824 prioq_reshuffle(d->latest, s, &s->time.latest_index);
1825 d->needs_rearm = true;
1830 _public_ int sd_event_source_get_time_clock(sd_event_source *s, clockid_t *clock) {
1831 assert_return(s, -EINVAL);
1832 assert_return(clock, -EINVAL);
1833 assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
1834 assert_return(!event_pid_changed(s->event), -ECHILD);
1836 *clock = event_source_type_to_clock(s->type);
1840 _public_ int sd_event_source_get_child_pid(sd_event_source *s, pid_t *pid) {
1841 assert_return(s, -EINVAL);
1842 assert_return(pid, -EINVAL);
1843 assert_return(s->type == SOURCE_CHILD, -EDOM);
1844 assert_return(!event_pid_changed(s->event), -ECHILD);
1846 *pid = s->child.pid;
1851 _public_ int sd_event_source_set_prepare(sd_event_source *s, sd_event_handler_t callback) {
1854 assert_return(s, -EINVAL);
1855 assert_return(s->type != SOURCE_EXIT, -EDOM);
1856 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1857 assert_return(!event_pid_changed(s->event), -ECHILD);
1859 if (s->prepare == callback)
1862 if (callback && s->prepare) {
1863 s->prepare = callback;
1867 r = prioq_ensure_allocated(&s->event->prepare, prepare_prioq_compare);
1871 s->prepare = callback;
1874 r = prioq_put(s->event->prepare, s, &s->prepare_index);
1878 prioq_remove(s->event->prepare, s, &s->prepare_index);
1883 #if 0 /// UNNEEDED by elogind
1884 _public_ void* sd_event_source_get_userdata(sd_event_source *s) {
1885 assert_return(s, NULL);
1890 _public_ void *sd_event_source_set_userdata(sd_event_source *s, void *userdata) {
1893 assert_return(s, NULL);
1896 s->userdata = userdata;
1902 static usec_t sleep_between(sd_event *e, usec_t a, usec_t b) {
1909 if (a >= USEC_INFINITY)
1910 return USEC_INFINITY;
1915 initialize_perturb(e);
1918 Find a good time to wake up again between times a and b. We
1919 have two goals here:
1921 a) We want to wake up as seldom as possible, hence prefer
1922 later times over earlier times.
1924 b) But if we have to wake up, then let's make sure to
1925 dispatch as much as possible on the entire system.
1927 We implement this by waking up everywhere at the same time
1928 within any given minute if we can, synchronised via the
1929 perturbation value determined from the boot ID. If we can't,
1930 then we try to find the same spot in every 10s, then 1s and
1931 then 250ms step. Otherwise, we pick the last possible time
1935 c = (b / USEC_PER_MINUTE) * USEC_PER_MINUTE + e->perturb;
1937 if (_unlikely_(c < USEC_PER_MINUTE))
1940 c -= USEC_PER_MINUTE;
1946 c = (b / (USEC_PER_SEC*10)) * (USEC_PER_SEC*10) + (e->perturb % (USEC_PER_SEC*10));
1948 if (_unlikely_(c < USEC_PER_SEC*10))
1951 c -= USEC_PER_SEC*10;
1957 c = (b / USEC_PER_SEC) * USEC_PER_SEC + (e->perturb % USEC_PER_SEC);
1959 if (_unlikely_(c < USEC_PER_SEC))
1968 c = (b / (USEC_PER_MSEC*250)) * (USEC_PER_MSEC*250) + (e->perturb % (USEC_PER_MSEC*250));
1970 if (_unlikely_(c < USEC_PER_MSEC*250))
1973 c -= USEC_PER_MSEC*250;
1982 static int event_arm_timer(
1984 struct clock_data *d) {
1986 struct itimerspec its = {};
1987 sd_event_source *a, *b;
1994 if (!d->needs_rearm)
1997 d->needs_rearm = false;
1999 a = prioq_peek(d->earliest);
2000 if (!a || a->enabled == SD_EVENT_OFF || a->time.next == USEC_INFINITY) {
2005 if (d->next == USEC_INFINITY)
2009 r = timerfd_settime(d->fd, TFD_TIMER_ABSTIME, &its, NULL);
2013 d->next = USEC_INFINITY;
2017 b = prioq_peek(d->latest);
2018 assert_se(b && b->enabled != SD_EVENT_OFF);
2020 t = sleep_between(e, a->time.next, time_event_source_latest(b));
2024 assert_se(d->fd >= 0);
2027 /* We don' want to disarm here, just mean some time looooong ago. */
2028 its.it_value.tv_sec = 0;
2029 its.it_value.tv_nsec = 1;
2031 timespec_store(&its.it_value, t);
2033 r = timerfd_settime(d->fd, TFD_TIMER_ABSTIME, &its, NULL);
2041 static int process_io(sd_event *e, sd_event_source *s, uint32_t revents) {
2044 assert(s->type == SOURCE_IO);
2046 /* If the event source was already pending, we just OR in the
2047 * new revents, otherwise we reset the value. The ORing is
2048 * necessary to handle EPOLLONESHOT events properly where
2049 * readability might happen independently of writability, and
2050 * we need to keep track of both */
2053 s->io.revents |= revents;
2055 s->io.revents = revents;
2057 return source_set_pending(s, true);
2060 static int flush_timer(sd_event *e, int fd, uint32_t events, usec_t *next) {
2067 assert_return(events == EPOLLIN, -EIO);
2069 ss = read(fd, &x, sizeof(x));
2071 if (errno == EAGAIN || errno == EINTR)
2077 if (_unlikely_(ss != sizeof(x)))
2081 *next = USEC_INFINITY;
2086 static int process_timer(
2089 struct clock_data *d) {
2098 s = prioq_peek(d->earliest);
2101 s->enabled == SD_EVENT_OFF ||
2105 r = source_set_pending(s, true);
2109 prioq_reshuffle(d->earliest, s, &s->time.earliest_index);
2110 prioq_reshuffle(d->latest, s, &s->time.latest_index);
2111 d->needs_rearm = true;
2117 static int process_child(sd_event *e) {
2124 e->need_process_child = false;
2127 So, this is ugly. We iteratively invoke waitid() with P_PID
2128 + WNOHANG for each PID we wait for, instead of using
2129 P_ALL. This is because we only want to get child
2130 information of very specific child processes, and not all
2131 of them. We might not have processed the SIGCHLD even of a
2132 previous invocation and we don't want to maintain a
2133 unbounded *per-child* event queue, hence we really don't
2134 want anything flushed out of the kernel's queue that we
2135 don't care about. Since this is O(n) this means that if you
2136 have a lot of processes you probably want to handle SIGCHLD
2139 We do not reap the children here (by using WNOWAIT), this
2140 is only done after the event source is dispatched so that
2141 the callback still sees the process as a zombie.
2144 HASHMAP_FOREACH(s, e->child_sources, i) {
2145 assert(s->type == SOURCE_CHILD);
2150 if (s->enabled == SD_EVENT_OFF)
2153 zero(s->child.siginfo);
2154 r = waitid(P_PID, s->child.pid, &s->child.siginfo,
2155 WNOHANG | (s->child.options & WEXITED ? WNOWAIT : 0) | s->child.options);
2159 if (s->child.siginfo.si_pid != 0) {
2161 s->child.siginfo.si_code == CLD_EXITED ||
2162 s->child.siginfo.si_code == CLD_KILLED ||
2163 s->child.siginfo.si_code == CLD_DUMPED;
2165 if (!zombie && (s->child.options & WEXITED)) {
2166 /* If the child isn't dead then let's
2167 * immediately remove the state change
2168 * from the queue, since there's no
2169 * benefit in leaving it queued */
2171 assert(s->child.options & (WSTOPPED|WCONTINUED));
2172 waitid(P_PID, s->child.pid, &s->child.siginfo, WNOHANG|(s->child.options & (WSTOPPED|WCONTINUED)));
2175 r = source_set_pending(s, true);
2184 static int process_signal(sd_event *e, struct signal_data *d, uint32_t events) {
2185 bool read_one = false;
2189 assert_return(events == EPOLLIN, -EIO);
2191 /* If there's a signal queued on this priority and SIGCHLD is
2192 on this priority too, then make sure to recheck the
2193 children we watch. This is because we only ever dequeue
2194 the first signal per priority, and if we dequeue one, and
2195 SIGCHLD might be enqueued later we wouldn't know, but we
2196 might have higher priority children we care about hence we
2197 need to check that explicitly. */
2199 if (sigismember(&d->sigset, SIGCHLD))
2200 e->need_process_child = true;
2202 /* If there's already an event source pending for this
2203 * priority we don't read another */
2208 struct signalfd_siginfo si;
2210 sd_event_source *s = NULL;
2212 n = read(d->fd, &si, sizeof(si));
2214 if (errno == EAGAIN || errno == EINTR)
2220 if (_unlikely_(n != sizeof(si)))
2223 assert(si.ssi_signo < _NSIG);
2227 if (e->signal_sources)
2228 s = e->signal_sources[si.ssi_signo];
2234 s->signal.siginfo = si;
2237 r = source_set_pending(s, true);
2245 static int source_dispatch(sd_event_source *s) {
2249 assert(s->pending || s->type == SOURCE_EXIT);
2251 if (s->type != SOURCE_DEFER && s->type != SOURCE_EXIT) {
2252 r = source_set_pending(s, false);
2257 if (s->type != SOURCE_POST) {
2261 /* If we execute a non-post source, let's mark all
2262 * post sources as pending */
2264 SET_FOREACH(z, s->event->post_sources, i) {
2265 if (z->enabled == SD_EVENT_OFF)
2268 r = source_set_pending(z, true);
2274 if (s->enabled == SD_EVENT_ONESHOT) {
2275 r = sd_event_source_set_enabled(s, SD_EVENT_OFF);
2280 s->dispatching = true;
2285 r = s->io.callback(s, s->io.fd, s->io.revents, s->userdata);
2288 case SOURCE_TIME_REALTIME:
2289 case SOURCE_TIME_BOOTTIME:
2290 case SOURCE_TIME_MONOTONIC:
2291 case SOURCE_TIME_REALTIME_ALARM:
2292 case SOURCE_TIME_BOOTTIME_ALARM:
2293 r = s->time.callback(s, s->time.next, s->userdata);
2297 r = s->signal.callback(s, &s->signal.siginfo, s->userdata);
2300 case SOURCE_CHILD: {
2303 zombie = s->child.siginfo.si_code == CLD_EXITED ||
2304 s->child.siginfo.si_code == CLD_KILLED ||
2305 s->child.siginfo.si_code == CLD_DUMPED;
2307 r = s->child.callback(s, &s->child.siginfo, s->userdata);
2309 /* Now, reap the PID for good. */
2311 waitid(P_PID, s->child.pid, &s->child.siginfo, WNOHANG|WEXITED);
2317 r = s->defer.callback(s, s->userdata);
2321 r = s->post.callback(s, s->userdata);
2325 r = s->exit.callback(s, s->userdata);
2328 case SOURCE_WATCHDOG:
2329 case _SOURCE_EVENT_SOURCE_TYPE_MAX:
2330 case _SOURCE_EVENT_SOURCE_TYPE_INVALID:
2331 assert_not_reached("Wut? I shouldn't exist.");
2334 s->dispatching = false;
2337 log_debug_errno(r, "Event source %s (type %s) returned error, disabling: %m",
2338 strna(s->description), event_source_type_to_string(s->type));
2343 sd_event_source_set_enabled(s, SD_EVENT_OFF);
2348 static int event_prepare(sd_event *e) {
2356 s = prioq_peek(e->prepare);
2357 if (!s || s->prepare_iteration == e->iteration || s->enabled == SD_EVENT_OFF)
2360 s->prepare_iteration = e->iteration;
2361 r = prioq_reshuffle(e->prepare, s, &s->prepare_index);
2367 s->dispatching = true;
2368 r = s->prepare(s, s->userdata);
2369 s->dispatching = false;
2372 log_debug_errno(r, "Prepare callback of event source %s (type %s) returned error, disabling: %m",
2373 strna(s->description), event_source_type_to_string(s->type));
2378 sd_event_source_set_enabled(s, SD_EVENT_OFF);
2384 static int dispatch_exit(sd_event *e) {
2390 p = prioq_peek(e->exit);
2391 if (!p || p->enabled == SD_EVENT_OFF) {
2392 e->state = SD_EVENT_FINISHED;
2398 e->state = SD_EVENT_EXITING;
2400 r = source_dispatch(p);
2402 e->state = SD_EVENT_INITIAL;
2408 static sd_event_source* event_next_pending(sd_event *e) {
2413 p = prioq_peek(e->pending);
2417 if (p->enabled == SD_EVENT_OFF)
2423 static int arm_watchdog(sd_event *e) {
2424 struct itimerspec its = {};
2429 assert(e->watchdog_fd >= 0);
2431 t = sleep_between(e,
2432 e->watchdog_last + (e->watchdog_period / 2),
2433 e->watchdog_last + (e->watchdog_period * 3 / 4));
2435 timespec_store(&its.it_value, t);
2437 /* Make sure we never set the watchdog to 0, which tells the
2438 * kernel to disable it. */
2439 if (its.it_value.tv_sec == 0 && its.it_value.tv_nsec == 0)
2440 its.it_value.tv_nsec = 1;
2442 r = timerfd_settime(e->watchdog_fd, TFD_TIMER_ABSTIME, &its, NULL);
2449 static int process_watchdog(sd_event *e) {
2455 /* Don't notify watchdog too often */
2456 if (e->watchdog_last + e->watchdog_period / 4 > e->timestamp.monotonic)
2459 sd_notify(false, "WATCHDOG=1");
2460 e->watchdog_last = e->timestamp.monotonic;
2462 return arm_watchdog(e);
2465 _public_ int sd_event_prepare(sd_event *e) {
2468 assert_return(e, -EINVAL);
2469 assert_return(!event_pid_changed(e), -ECHILD);
2470 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
2471 assert_return(e->state == SD_EVENT_INITIAL, -EBUSY);
2473 if (e->exit_requested)
2478 e->state = SD_EVENT_PREPARING;
2479 r = event_prepare(e);
2480 e->state = SD_EVENT_INITIAL;
2484 r = event_arm_timer(e, &e->realtime);
2488 r = event_arm_timer(e, &e->boottime);
2492 r = event_arm_timer(e, &e->monotonic);
2496 r = event_arm_timer(e, &e->realtime_alarm);
2500 r = event_arm_timer(e, &e->boottime_alarm);
2504 if (event_next_pending(e) || e->need_process_child)
2507 e->state = SD_EVENT_ARMED;
2512 e->state = SD_EVENT_ARMED;
2513 r = sd_event_wait(e, 0);
2515 e->state = SD_EVENT_ARMED;
2520 _public_ int sd_event_wait(sd_event *e, uint64_t timeout) {
2521 struct epoll_event *ev_queue;
2522 unsigned ev_queue_max;
2525 assert_return(e, -EINVAL);
2526 assert_return(!event_pid_changed(e), -ECHILD);
2527 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
2528 assert_return(e->state == SD_EVENT_ARMED, -EBUSY);
2530 if (e->exit_requested) {
2531 e->state = SD_EVENT_PENDING;
2535 ev_queue_max = MAX(e->n_sources, 1u);
2536 ev_queue = newa(struct epoll_event, ev_queue_max);
2538 m = epoll_wait(e->epoll_fd, ev_queue, ev_queue_max,
2539 timeout == (uint64_t) -1 ? -1 : (int) ((timeout + USEC_PER_MSEC - 1) / USEC_PER_MSEC));
2541 if (errno == EINTR) {
2542 e->state = SD_EVENT_PENDING;
2550 dual_timestamp_get(&e->timestamp);
2551 e->timestamp_boottime = now(CLOCK_BOOTTIME);
2553 for (i = 0; i < m; i++) {
2555 if (ev_queue[i].data.ptr == INT_TO_PTR(SOURCE_WATCHDOG))
2556 r = flush_timer(e, e->watchdog_fd, ev_queue[i].events, NULL);
2558 WakeupType *t = ev_queue[i].data.ptr;
2562 case WAKEUP_EVENT_SOURCE:
2563 r = process_io(e, ev_queue[i].data.ptr, ev_queue[i].events);
2566 case WAKEUP_CLOCK_DATA: {
2567 struct clock_data *d = ev_queue[i].data.ptr;
2568 r = flush_timer(e, d->fd, ev_queue[i].events, &d->next);
2572 case WAKEUP_SIGNAL_DATA:
2573 r = process_signal(e, ev_queue[i].data.ptr, ev_queue[i].events);
2577 assert_not_reached("Invalid wake-up pointer");
2584 r = process_watchdog(e);
2588 r = process_timer(e, e->timestamp.realtime, &e->realtime);
2592 r = process_timer(e, e->timestamp_boottime, &e->boottime);
2596 r = process_timer(e, e->timestamp.monotonic, &e->monotonic);
2600 r = process_timer(e, e->timestamp.realtime, &e->realtime_alarm);
2604 r = process_timer(e, e->timestamp_boottime, &e->boottime_alarm);
2608 if (e->need_process_child) {
2609 r = process_child(e);
2614 if (event_next_pending(e)) {
2615 e->state = SD_EVENT_PENDING;
2623 e->state = SD_EVENT_INITIAL;
2628 _public_ int sd_event_dispatch(sd_event *e) {
2632 assert_return(e, -EINVAL);
2633 assert_return(!event_pid_changed(e), -ECHILD);
2634 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
2635 assert_return(e->state == SD_EVENT_PENDING, -EBUSY);
2637 if (e->exit_requested)
2638 return dispatch_exit(e);
2640 p = event_next_pending(e);
2644 e->state = SD_EVENT_RUNNING;
2645 r = source_dispatch(p);
2646 e->state = SD_EVENT_INITIAL;
2653 e->state = SD_EVENT_INITIAL;
2658 static void event_log_delays(sd_event *e) {
2659 char b[ELEMENTSOF(e->delays) * DECIMAL_STR_MAX(unsigned) + 1];
2663 for (i = o = 0; i < ELEMENTSOF(e->delays); i++) {
2664 o += snprintf(&b[o], sizeof(b) - o, "%u ", e->delays[i]);
2667 log_debug("Event loop iterations: %.*s", o, b);
2670 _public_ int sd_event_run(sd_event *e, uint64_t timeout) {
2673 assert_return(e, -EINVAL);
2674 assert_return(!event_pid_changed(e), -ECHILD);
2675 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
2676 assert_return(e->state == SD_EVENT_INITIAL, -EBUSY);
2678 if (e->profile_delays && e->last_run) {
2682 this_run = now(CLOCK_MONOTONIC);
2684 l = u64log2(this_run - e->last_run);
2685 assert(l < sizeof(e->delays));
2688 if (this_run - e->last_log >= 5*USEC_PER_SEC) {
2689 event_log_delays(e);
2690 e->last_log = this_run;
2694 r = sd_event_prepare(e);
2696 /* There was nothing? Then wait... */
2697 r = sd_event_wait(e, timeout);
2699 if (e->profile_delays)
2700 e->last_run = now(CLOCK_MONOTONIC);
2703 /* There's something now, then let's dispatch it */
2704 r = sd_event_dispatch(e);
2714 #if 0 /// UNNEEDED by elogind
2715 _public_ int sd_event_loop(sd_event *e) {
2718 assert_return(e, -EINVAL);
2719 assert_return(!event_pid_changed(e), -ECHILD);
2720 assert_return(e->state == SD_EVENT_INITIAL, -EBUSY);
2724 while (e->state != SD_EVENT_FINISHED) {
2725 r = sd_event_run(e, (uint64_t) -1);
2737 _public_ int sd_event_get_fd(sd_event *e) {
2739 assert_return(e, -EINVAL);
2740 assert_return(!event_pid_changed(e), -ECHILD);
2746 _public_ int sd_event_get_state(sd_event *e) {
2747 assert_return(e, -EINVAL);
2748 assert_return(!event_pid_changed(e), -ECHILD);
2753 #if 0 /// UNNEEDED by elogind
2754 _public_ int sd_event_get_exit_code(sd_event *e, int *code) {
2755 assert_return(e, -EINVAL);
2756 assert_return(code, -EINVAL);
2757 assert_return(!event_pid_changed(e), -ECHILD);
2759 if (!e->exit_requested)
2762 *code = e->exit_code;
2767 _public_ int sd_event_exit(sd_event *e, int code) {
2768 assert_return(e, -EINVAL);
2769 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
2770 assert_return(!event_pid_changed(e), -ECHILD);
2772 e->exit_requested = true;
2773 e->exit_code = code;
2778 #if 0 /// UNNEEDED by elogind
2779 _public_ int sd_event_now(sd_event *e, clockid_t clock, uint64_t *usec) {
2780 assert_return(e, -EINVAL);
2781 assert_return(usec, -EINVAL);
2782 assert_return(!event_pid_changed(e), -ECHILD);
2783 assert_return(IN_SET(clock,
2785 CLOCK_REALTIME_ALARM,
2788 CLOCK_BOOTTIME_ALARM), -EOPNOTSUPP);
2790 if (!dual_timestamp_is_set(&e->timestamp)) {
2791 /* Implicitly fall back to now() if we never ran
2792 * before and thus have no cached time. */
2799 case CLOCK_REALTIME:
2800 case CLOCK_REALTIME_ALARM:
2801 *usec = e->timestamp.realtime;
2804 case CLOCK_MONOTONIC:
2805 *usec = e->timestamp.monotonic;
2809 *usec = e->timestamp_boottime;
2817 _public_ int sd_event_default(sd_event **ret) {
2819 static thread_local sd_event *default_event = NULL;
2824 return !!default_event;
2826 if (default_event) {
2827 *ret = sd_event_ref(default_event);
2831 r = sd_event_new(&e);
2835 e->default_event_ptr = &default_event;
2843 #if 0 /// UNNEEDED by elogind
2844 _public_ int sd_event_get_tid(sd_event *e, pid_t *tid) {
2845 assert_return(e, -EINVAL);
2846 assert_return(tid, -EINVAL);
2847 assert_return(!event_pid_changed(e), -ECHILD);
2858 _public_ int sd_event_set_watchdog(sd_event *e, int b) {
2861 assert_return(e, -EINVAL);
2862 assert_return(!event_pid_changed(e), -ECHILD);
2864 if (e->watchdog == !!b)
2868 struct epoll_event ev = {};
2870 r = sd_watchdog_enabled(false, &e->watchdog_period);
2874 /* Issue first ping immediately */
2875 sd_notify(false, "WATCHDOG=1");
2876 e->watchdog_last = now(CLOCK_MONOTONIC);
2878 e->watchdog_fd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK|TFD_CLOEXEC);
2879 if (e->watchdog_fd < 0)
2882 r = arm_watchdog(e);
2886 ev.events = EPOLLIN;
2887 ev.data.ptr = INT_TO_PTR(SOURCE_WATCHDOG);
2889 r = epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, e->watchdog_fd, &ev);
2896 if (e->watchdog_fd >= 0) {
2897 epoll_ctl(e->epoll_fd, EPOLL_CTL_DEL, e->watchdog_fd, NULL);
2898 e->watchdog_fd = safe_close(e->watchdog_fd);
2906 e->watchdog_fd = safe_close(e->watchdog_fd);
2910 #if 0 /// UNNEEDED by elogind
2911 _public_ int sd_event_get_watchdog(sd_event *e) {
2912 assert_return(e, -EINVAL);
2913 assert_return(!event_pid_changed(e), -ECHILD);