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);
1077 if (IN_SET(clock, CLOCK_BOOTTIME, CLOCK_BOOTTIME_ALARM) &&
1078 !clock_boottime_supported())
1082 callback = time_exit_callback;
1084 type = clock_to_event_source_type(clock);
1085 assert_return(type >= 0, -EOPNOTSUPP);
1087 d = event_get_clock_data(e, type);
1090 r = prioq_ensure_allocated(&d->earliest, earliest_time_prioq_compare);
1094 r = prioq_ensure_allocated(&d->latest, latest_time_prioq_compare);
1099 r = event_setup_timer_fd(e, d, clock);
1104 s = source_new(e, !ret, type);
1108 s->time.next = usec;
1109 s->time.accuracy = accuracy == 0 ? DEFAULT_ACCURACY_USEC : accuracy;
1110 s->time.callback = callback;
1111 s->time.earliest_index = s->time.latest_index = PRIOQ_IDX_NULL;
1112 s->userdata = userdata;
1113 s->enabled = SD_EVENT_ONESHOT;
1115 d->needs_rearm = true;
1117 r = prioq_put(d->earliest, s, &s->time.earliest_index);
1121 r = prioq_put(d->latest, s, &s->time.latest_index);
1135 static int signal_exit_callback(sd_event_source *s, const struct signalfd_siginfo *si, void *userdata) {
1138 return sd_event_exit(sd_event_source_get_event(s), PTR_TO_INT(userdata));
1141 _public_ int sd_event_add_signal(
1143 sd_event_source **ret,
1145 sd_event_signal_handler_t callback,
1149 struct signal_data *d;
1153 assert_return(e, -EINVAL);
1154 assert_return(SIGNAL_VALID(sig), -EINVAL);
1155 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1156 assert_return(!event_pid_changed(e), -ECHILD);
1159 callback = signal_exit_callback;
1161 r = pthread_sigmask(SIG_SETMASK, NULL, &ss);
1165 if (!sigismember(&ss, sig))
1168 if (!e->signal_sources) {
1169 e->signal_sources = new0(sd_event_source*, _NSIG);
1170 if (!e->signal_sources)
1172 } else if (e->signal_sources[sig])
1175 s = source_new(e, !ret, SOURCE_SIGNAL);
1179 s->signal.sig = sig;
1180 s->signal.callback = callback;
1181 s->userdata = userdata;
1182 s->enabled = SD_EVENT_ON;
1184 e->signal_sources[sig] = s;
1186 r = event_make_signal_data(e, sig, &d);
1192 /* Use the signal name as description for the event source by default */
1193 (void) sd_event_source_set_description(s, signal_to_string(sig));
1201 #if 0 /// UNNEEDED by elogind
1202 _public_ int sd_event_add_child(
1204 sd_event_source **ret,
1207 sd_event_child_handler_t callback,
1213 assert_return(e, -EINVAL);
1214 assert_return(pid > 1, -EINVAL);
1215 assert_return(!(options & ~(WEXITED|WSTOPPED|WCONTINUED)), -EINVAL);
1216 assert_return(options != 0, -EINVAL);
1217 assert_return(callback, -EINVAL);
1218 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1219 assert_return(!event_pid_changed(e), -ECHILD);
1221 r = hashmap_ensure_allocated(&e->child_sources, NULL);
1225 if (hashmap_contains(e->child_sources, PID_TO_PTR(pid)))
1228 s = source_new(e, !ret, SOURCE_CHILD);
1233 s->child.options = options;
1234 s->child.callback = callback;
1235 s->userdata = userdata;
1236 s->enabled = SD_EVENT_ONESHOT;
1238 r = hashmap_put(e->child_sources, PID_TO_PTR(pid), s);
1244 e->n_enabled_child_sources++;
1246 r = event_make_signal_data(e, SIGCHLD, NULL);
1248 e->n_enabled_child_sources--;
1253 e->need_process_child = true;
1261 _public_ int sd_event_add_defer(
1263 sd_event_source **ret,
1264 sd_event_handler_t callback,
1270 assert_return(e, -EINVAL);
1271 assert_return(callback, -EINVAL);
1272 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1273 assert_return(!event_pid_changed(e), -ECHILD);
1275 s = source_new(e, !ret, SOURCE_DEFER);
1279 s->defer.callback = callback;
1280 s->userdata = userdata;
1281 s->enabled = SD_EVENT_ONESHOT;
1283 r = source_set_pending(s, true);
1296 _public_ int sd_event_add_post(
1298 sd_event_source **ret,
1299 sd_event_handler_t callback,
1305 assert_return(e, -EINVAL);
1306 assert_return(callback, -EINVAL);
1307 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1308 assert_return(!event_pid_changed(e), -ECHILD);
1310 r = set_ensure_allocated(&e->post_sources, NULL);
1314 s = source_new(e, !ret, SOURCE_POST);
1318 s->post.callback = callback;
1319 s->userdata = userdata;
1320 s->enabled = SD_EVENT_ON;
1322 r = set_put(e->post_sources, s);
1334 _public_ int sd_event_add_exit(
1336 sd_event_source **ret,
1337 sd_event_handler_t callback,
1343 assert_return(e, -EINVAL);
1344 assert_return(callback, -EINVAL);
1345 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
1346 assert_return(!event_pid_changed(e), -ECHILD);
1348 r = prioq_ensure_allocated(&e->exit, exit_prioq_compare);
1352 s = source_new(e, !ret, SOURCE_EXIT);
1356 s->exit.callback = callback;
1357 s->userdata = userdata;
1358 s->exit.prioq_index = PRIOQ_IDX_NULL;
1359 s->enabled = SD_EVENT_ONESHOT;
1361 r = prioq_put(s->event->exit, s, &s->exit.prioq_index);
1373 #if 0 /// UNNEEDED by elogind
1374 _public_ sd_event_source* sd_event_source_ref(sd_event_source *s) {
1379 assert(s->n_ref >= 1);
1386 _public_ sd_event_source* sd_event_source_unref(sd_event_source *s) {
1391 assert(s->n_ref >= 1);
1394 if (s->n_ref <= 0) {
1395 /* Here's a special hack: when we are called from a
1396 * dispatch handler we won't free the event source
1397 * immediately, but we will detach the fd from the
1398 * epoll. This way it is safe for the caller to unref
1399 * the event source and immediately close the fd, but
1400 * we still retain a valid event source object after
1403 if (s->dispatching) {
1404 if (s->type == SOURCE_IO)
1405 source_io_unregister(s);
1407 source_disconnect(s);
1415 _public_ int sd_event_source_set_description(sd_event_source *s, const char *description) {
1416 assert_return(s, -EINVAL);
1417 assert_return(!event_pid_changed(s->event), -ECHILD);
1419 return free_and_strdup(&s->description, description);
1422 #if 0 /// UNNEEDED by elogind
1423 _public_ int sd_event_source_get_description(sd_event_source *s, const char **description) {
1424 assert_return(s, -EINVAL);
1425 assert_return(description, -EINVAL);
1426 assert_return(s->description, -ENXIO);
1427 assert_return(!event_pid_changed(s->event), -ECHILD);
1429 *description = s->description;
1434 _public_ sd_event *sd_event_source_get_event(sd_event_source *s) {
1435 assert_return(s, NULL);
1440 #if 0 /// UNNEEDED by elogind
1441 _public_ int sd_event_source_get_pending(sd_event_source *s) {
1442 assert_return(s, -EINVAL);
1443 assert_return(s->type != SOURCE_EXIT, -EDOM);
1444 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1445 assert_return(!event_pid_changed(s->event), -ECHILD);
1450 _public_ int sd_event_source_get_io_fd(sd_event_source *s) {
1451 assert_return(s, -EINVAL);
1452 assert_return(s->type == SOURCE_IO, -EDOM);
1453 assert_return(!event_pid_changed(s->event), -ECHILD);
1459 _public_ int sd_event_source_set_io_fd(sd_event_source *s, int fd) {
1462 assert_return(s, -EINVAL);
1463 assert_return(fd >= 0, -EBADF);
1464 assert_return(s->type == SOURCE_IO, -EDOM);
1465 assert_return(!event_pid_changed(s->event), -ECHILD);
1470 if (s->enabled == SD_EVENT_OFF) {
1472 s->io.registered = false;
1476 saved_fd = s->io.fd;
1477 assert(s->io.registered);
1480 s->io.registered = false;
1482 r = source_io_register(s, s->enabled, s->io.events);
1484 s->io.fd = saved_fd;
1485 s->io.registered = true;
1489 epoll_ctl(s->event->epoll_fd, EPOLL_CTL_DEL, saved_fd, NULL);
1495 #if 0 /// UNNEEDED by elogind
1496 _public_ int sd_event_source_get_io_events(sd_event_source *s, uint32_t* events) {
1497 assert_return(s, -EINVAL);
1498 assert_return(events, -EINVAL);
1499 assert_return(s->type == SOURCE_IO, -EDOM);
1500 assert_return(!event_pid_changed(s->event), -ECHILD);
1502 *events = s->io.events;
1507 _public_ int sd_event_source_set_io_events(sd_event_source *s, uint32_t events) {
1510 assert_return(s, -EINVAL);
1511 assert_return(s->type == SOURCE_IO, -EDOM);
1512 assert_return(!(events & ~(EPOLLIN|EPOLLOUT|EPOLLRDHUP|EPOLLPRI|EPOLLERR|EPOLLHUP|EPOLLET)), -EINVAL);
1513 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1514 assert_return(!event_pid_changed(s->event), -ECHILD);
1516 /* edge-triggered updates are never skipped, so we can reset edges */
1517 if (s->io.events == events && !(events & EPOLLET))
1520 if (s->enabled != SD_EVENT_OFF) {
1521 r = source_io_register(s, s->enabled, events);
1526 s->io.events = events;
1527 source_set_pending(s, false);
1532 #if 0 /// UNNEEDED by elogind
1533 _public_ int sd_event_source_get_io_revents(sd_event_source *s, uint32_t* revents) {
1534 assert_return(s, -EINVAL);
1535 assert_return(revents, -EINVAL);
1536 assert_return(s->type == SOURCE_IO, -EDOM);
1537 assert_return(s->pending, -ENODATA);
1538 assert_return(!event_pid_changed(s->event), -ECHILD);
1540 *revents = s->io.revents;
1544 _public_ int sd_event_source_get_signal(sd_event_source *s) {
1545 assert_return(s, -EINVAL);
1546 assert_return(s->type == SOURCE_SIGNAL, -EDOM);
1547 assert_return(!event_pid_changed(s->event), -ECHILD);
1549 return s->signal.sig;
1552 _public_ int sd_event_source_get_priority(sd_event_source *s, int64_t *priority) {
1553 assert_return(s, -EINVAL);
1554 assert_return(!event_pid_changed(s->event), -ECHILD);
1560 _public_ int sd_event_source_set_priority(sd_event_source *s, int64_t priority) {
1563 assert_return(s, -EINVAL);
1564 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1565 assert_return(!event_pid_changed(s->event), -ECHILD);
1567 if (s->priority == priority)
1570 if (s->type == SOURCE_SIGNAL && s->enabled != SD_EVENT_OFF) {
1571 struct signal_data *old, *d;
1573 /* Move us from the signalfd belonging to the old
1574 * priority to the signalfd of the new priority */
1576 assert_se(old = hashmap_get(s->event->signal_data, &s->priority));
1578 s->priority = priority;
1580 r = event_make_signal_data(s->event, s->signal.sig, &d);
1582 s->priority = old->priority;
1586 event_unmask_signal_data(s->event, old, s->signal.sig);
1588 s->priority = priority;
1591 prioq_reshuffle(s->event->pending, s, &s->pending_index);
1594 prioq_reshuffle(s->event->prepare, s, &s->prepare_index);
1596 if (s->type == SOURCE_EXIT)
1597 prioq_reshuffle(s->event->exit, s, &s->exit.prioq_index);
1602 #if 0 /// UNNEEDED by elogind
1603 _public_ int sd_event_source_get_enabled(sd_event_source *s, int *m) {
1604 assert_return(s, -EINVAL);
1605 assert_return(m, -EINVAL);
1606 assert_return(!event_pid_changed(s->event), -ECHILD);
1613 _public_ int sd_event_source_set_enabled(sd_event_source *s, int m) {
1616 assert_return(s, -EINVAL);
1617 assert_return(m == SD_EVENT_OFF || m == SD_EVENT_ON || m == SD_EVENT_ONESHOT, -EINVAL);
1618 assert_return(!event_pid_changed(s->event), -ECHILD);
1620 /* If we are dead anyway, we are fine with turning off
1621 * sources, but everything else needs to fail. */
1622 if (s->event->state == SD_EVENT_FINISHED)
1623 return m == SD_EVENT_OFF ? 0 : -ESTALE;
1625 if (s->enabled == m)
1628 if (m == SD_EVENT_OFF) {
1633 source_io_unregister(s);
1637 case SOURCE_TIME_REALTIME:
1638 case SOURCE_TIME_BOOTTIME:
1639 case SOURCE_TIME_MONOTONIC:
1640 case SOURCE_TIME_REALTIME_ALARM:
1641 case SOURCE_TIME_BOOTTIME_ALARM: {
1642 struct clock_data *d;
1645 d = event_get_clock_data(s->event, s->type);
1648 prioq_reshuffle(d->earliest, s, &s->time.earliest_index);
1649 prioq_reshuffle(d->latest, s, &s->time.latest_index);
1650 d->needs_rearm = true;
1657 event_gc_signal_data(s->event, &s->priority, s->signal.sig);
1663 assert(s->event->n_enabled_child_sources > 0);
1664 s->event->n_enabled_child_sources--;
1666 event_gc_signal_data(s->event, &s->priority, SIGCHLD);
1671 prioq_reshuffle(s->event->exit, s, &s->exit.prioq_index);
1680 assert_not_reached("Wut? I shouldn't exist.");
1687 r = source_io_register(s, m, s->io.events);
1694 case SOURCE_TIME_REALTIME:
1695 case SOURCE_TIME_BOOTTIME:
1696 case SOURCE_TIME_MONOTONIC:
1697 case SOURCE_TIME_REALTIME_ALARM:
1698 case SOURCE_TIME_BOOTTIME_ALARM: {
1699 struct clock_data *d;
1702 d = event_get_clock_data(s->event, s->type);
1705 prioq_reshuffle(d->earliest, s, &s->time.earliest_index);
1706 prioq_reshuffle(d->latest, s, &s->time.latest_index);
1707 d->needs_rearm = true;
1715 r = event_make_signal_data(s->event, s->signal.sig, NULL);
1717 s->enabled = SD_EVENT_OFF;
1718 event_gc_signal_data(s->event, &s->priority, s->signal.sig);
1726 if (s->enabled == SD_EVENT_OFF)
1727 s->event->n_enabled_child_sources++;
1731 r = event_make_signal_data(s->event, SIGCHLD, NULL);
1733 s->enabled = SD_EVENT_OFF;
1734 s->event->n_enabled_child_sources--;
1735 event_gc_signal_data(s->event, &s->priority, SIGCHLD);
1743 prioq_reshuffle(s->event->exit, s, &s->exit.prioq_index);
1752 assert_not_reached("Wut? I shouldn't exist.");
1757 prioq_reshuffle(s->event->pending, s, &s->pending_index);
1760 prioq_reshuffle(s->event->prepare, s, &s->prepare_index);
1765 _public_ int sd_event_source_get_time(sd_event_source *s, uint64_t *usec) {
1766 assert_return(s, -EINVAL);
1767 assert_return(usec, -EINVAL);
1768 assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
1769 assert_return(!event_pid_changed(s->event), -ECHILD);
1771 *usec = s->time.next;
1775 _public_ int sd_event_source_set_time(sd_event_source *s, uint64_t usec) {
1776 struct clock_data *d;
1778 assert_return(s, -EINVAL);
1779 assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
1780 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1781 assert_return(!event_pid_changed(s->event), -ECHILD);
1783 s->time.next = usec;
1785 source_set_pending(s, false);
1787 d = event_get_clock_data(s->event, s->type);
1790 prioq_reshuffle(d->earliest, s, &s->time.earliest_index);
1791 prioq_reshuffle(d->latest, s, &s->time.latest_index);
1792 d->needs_rearm = true;
1797 #if 0 /// UNNEEDED by elogind
1798 _public_ int sd_event_source_get_time_accuracy(sd_event_source *s, uint64_t *usec) {
1799 assert_return(s, -EINVAL);
1800 assert_return(usec, -EINVAL);
1801 assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
1802 assert_return(!event_pid_changed(s->event), -ECHILD);
1804 *usec = s->time.accuracy;
1808 _public_ int sd_event_source_set_time_accuracy(sd_event_source *s, uint64_t usec) {
1809 struct clock_data *d;
1811 assert_return(s, -EINVAL);
1812 assert_return(usec != (uint64_t) -1, -EINVAL);
1813 assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
1814 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1815 assert_return(!event_pid_changed(s->event), -ECHILD);
1818 usec = DEFAULT_ACCURACY_USEC;
1820 s->time.accuracy = usec;
1822 source_set_pending(s, false);
1824 d = event_get_clock_data(s->event, s->type);
1827 prioq_reshuffle(d->latest, s, &s->time.latest_index);
1828 d->needs_rearm = true;
1833 _public_ int sd_event_source_get_time_clock(sd_event_source *s, clockid_t *clock) {
1834 assert_return(s, -EINVAL);
1835 assert_return(clock, -EINVAL);
1836 assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
1837 assert_return(!event_pid_changed(s->event), -ECHILD);
1839 *clock = event_source_type_to_clock(s->type);
1843 _public_ int sd_event_source_get_child_pid(sd_event_source *s, pid_t *pid) {
1844 assert_return(s, -EINVAL);
1845 assert_return(pid, -EINVAL);
1846 assert_return(s->type == SOURCE_CHILD, -EDOM);
1847 assert_return(!event_pid_changed(s->event), -ECHILD);
1849 *pid = s->child.pid;
1854 _public_ int sd_event_source_set_prepare(sd_event_source *s, sd_event_handler_t callback) {
1857 assert_return(s, -EINVAL);
1858 assert_return(s->type != SOURCE_EXIT, -EDOM);
1859 assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
1860 assert_return(!event_pid_changed(s->event), -ECHILD);
1862 if (s->prepare == callback)
1865 if (callback && s->prepare) {
1866 s->prepare = callback;
1870 r = prioq_ensure_allocated(&s->event->prepare, prepare_prioq_compare);
1874 s->prepare = callback;
1877 r = prioq_put(s->event->prepare, s, &s->prepare_index);
1881 prioq_remove(s->event->prepare, s, &s->prepare_index);
1886 #if 0 /// UNNEEDED by elogind
1887 _public_ void* sd_event_source_get_userdata(sd_event_source *s) {
1888 assert_return(s, NULL);
1893 _public_ void *sd_event_source_set_userdata(sd_event_source *s, void *userdata) {
1896 assert_return(s, NULL);
1899 s->userdata = userdata;
1905 static usec_t sleep_between(sd_event *e, usec_t a, usec_t b) {
1912 if (a >= USEC_INFINITY)
1913 return USEC_INFINITY;
1918 initialize_perturb(e);
1921 Find a good time to wake up again between times a and b. We
1922 have two goals here:
1924 a) We want to wake up as seldom as possible, hence prefer
1925 later times over earlier times.
1927 b) But if we have to wake up, then let's make sure to
1928 dispatch as much as possible on the entire system.
1930 We implement this by waking up everywhere at the same time
1931 within any given minute if we can, synchronised via the
1932 perturbation value determined from the boot ID. If we can't,
1933 then we try to find the same spot in every 10s, then 1s and
1934 then 250ms step. Otherwise, we pick the last possible time
1938 c = (b / USEC_PER_MINUTE) * USEC_PER_MINUTE + e->perturb;
1940 if (_unlikely_(c < USEC_PER_MINUTE))
1943 c -= USEC_PER_MINUTE;
1949 c = (b / (USEC_PER_SEC*10)) * (USEC_PER_SEC*10) + (e->perturb % (USEC_PER_SEC*10));
1951 if (_unlikely_(c < USEC_PER_SEC*10))
1954 c -= USEC_PER_SEC*10;
1960 c = (b / USEC_PER_SEC) * USEC_PER_SEC + (e->perturb % USEC_PER_SEC);
1962 if (_unlikely_(c < USEC_PER_SEC))
1971 c = (b / (USEC_PER_MSEC*250)) * (USEC_PER_MSEC*250) + (e->perturb % (USEC_PER_MSEC*250));
1973 if (_unlikely_(c < USEC_PER_MSEC*250))
1976 c -= USEC_PER_MSEC*250;
1985 static int event_arm_timer(
1987 struct clock_data *d) {
1989 struct itimerspec its = {};
1990 sd_event_source *a, *b;
1997 if (!d->needs_rearm)
2000 d->needs_rearm = false;
2002 a = prioq_peek(d->earliest);
2003 if (!a || a->enabled == SD_EVENT_OFF || a->time.next == USEC_INFINITY) {
2008 if (d->next == USEC_INFINITY)
2012 r = timerfd_settime(d->fd, TFD_TIMER_ABSTIME, &its, NULL);
2016 d->next = USEC_INFINITY;
2020 b = prioq_peek(d->latest);
2021 assert_se(b && b->enabled != SD_EVENT_OFF);
2023 t = sleep_between(e, a->time.next, time_event_source_latest(b));
2027 assert_se(d->fd >= 0);
2030 /* We don' want to disarm here, just mean some time looooong ago. */
2031 its.it_value.tv_sec = 0;
2032 its.it_value.tv_nsec = 1;
2034 timespec_store(&its.it_value, t);
2036 r = timerfd_settime(d->fd, TFD_TIMER_ABSTIME, &its, NULL);
2044 static int process_io(sd_event *e, sd_event_source *s, uint32_t revents) {
2047 assert(s->type == SOURCE_IO);
2049 /* If the event source was already pending, we just OR in the
2050 * new revents, otherwise we reset the value. The ORing is
2051 * necessary to handle EPOLLONESHOT events properly where
2052 * readability might happen independently of writability, and
2053 * we need to keep track of both */
2056 s->io.revents |= revents;
2058 s->io.revents = revents;
2060 return source_set_pending(s, true);
2063 static int flush_timer(sd_event *e, int fd, uint32_t events, usec_t *next) {
2070 assert_return(events == EPOLLIN, -EIO);
2072 ss = read(fd, &x, sizeof(x));
2074 if (errno == EAGAIN || errno == EINTR)
2080 if (_unlikely_(ss != sizeof(x)))
2084 *next = USEC_INFINITY;
2089 static int process_timer(
2092 struct clock_data *d) {
2101 s = prioq_peek(d->earliest);
2104 s->enabled == SD_EVENT_OFF ||
2108 r = source_set_pending(s, true);
2112 prioq_reshuffle(d->earliest, s, &s->time.earliest_index);
2113 prioq_reshuffle(d->latest, s, &s->time.latest_index);
2114 d->needs_rearm = true;
2120 static int process_child(sd_event *e) {
2127 e->need_process_child = false;
2130 So, this is ugly. We iteratively invoke waitid() with P_PID
2131 + WNOHANG for each PID we wait for, instead of using
2132 P_ALL. This is because we only want to get child
2133 information of very specific child processes, and not all
2134 of them. We might not have processed the SIGCHLD even of a
2135 previous invocation and we don't want to maintain a
2136 unbounded *per-child* event queue, hence we really don't
2137 want anything flushed out of the kernel's queue that we
2138 don't care about. Since this is O(n) this means that if you
2139 have a lot of processes you probably want to handle SIGCHLD
2142 We do not reap the children here (by using WNOWAIT), this
2143 is only done after the event source is dispatched so that
2144 the callback still sees the process as a zombie.
2147 HASHMAP_FOREACH(s, e->child_sources, i) {
2148 assert(s->type == SOURCE_CHILD);
2153 if (s->enabled == SD_EVENT_OFF)
2156 zero(s->child.siginfo);
2157 r = waitid(P_PID, s->child.pid, &s->child.siginfo,
2158 WNOHANG | (s->child.options & WEXITED ? WNOWAIT : 0) | s->child.options);
2162 if (s->child.siginfo.si_pid != 0) {
2164 s->child.siginfo.si_code == CLD_EXITED ||
2165 s->child.siginfo.si_code == CLD_KILLED ||
2166 s->child.siginfo.si_code == CLD_DUMPED;
2168 if (!zombie && (s->child.options & WEXITED)) {
2169 /* If the child isn't dead then let's
2170 * immediately remove the state change
2171 * from the queue, since there's no
2172 * benefit in leaving it queued */
2174 assert(s->child.options & (WSTOPPED|WCONTINUED));
2175 waitid(P_PID, s->child.pid, &s->child.siginfo, WNOHANG|(s->child.options & (WSTOPPED|WCONTINUED)));
2178 r = source_set_pending(s, true);
2187 static int process_signal(sd_event *e, struct signal_data *d, uint32_t events) {
2188 bool read_one = false;
2192 assert_return(events == EPOLLIN, -EIO);
2194 /* If there's a signal queued on this priority and SIGCHLD is
2195 on this priority too, then make sure to recheck the
2196 children we watch. This is because we only ever dequeue
2197 the first signal per priority, and if we dequeue one, and
2198 SIGCHLD might be enqueued later we wouldn't know, but we
2199 might have higher priority children we care about hence we
2200 need to check that explicitly. */
2202 if (sigismember(&d->sigset, SIGCHLD))
2203 e->need_process_child = true;
2205 /* If there's already an event source pending for this
2206 * priority we don't read another */
2211 struct signalfd_siginfo si;
2213 sd_event_source *s = NULL;
2215 n = read(d->fd, &si, sizeof(si));
2217 if (errno == EAGAIN || errno == EINTR)
2223 if (_unlikely_(n != sizeof(si)))
2226 assert(SIGNAL_VALID(si.ssi_signo));
2230 if (e->signal_sources)
2231 s = e->signal_sources[si.ssi_signo];
2237 s->signal.siginfo = si;
2240 r = source_set_pending(s, true);
2248 static int source_dispatch(sd_event_source *s) {
2252 assert(s->pending || s->type == SOURCE_EXIT);
2254 if (s->type != SOURCE_DEFER && s->type != SOURCE_EXIT) {
2255 r = source_set_pending(s, false);
2260 if (s->type != SOURCE_POST) {
2264 /* If we execute a non-post source, let's mark all
2265 * post sources as pending */
2267 SET_FOREACH(z, s->event->post_sources, i) {
2268 if (z->enabled == SD_EVENT_OFF)
2271 r = source_set_pending(z, true);
2277 if (s->enabled == SD_EVENT_ONESHOT) {
2278 r = sd_event_source_set_enabled(s, SD_EVENT_OFF);
2283 s->dispatching = true;
2288 r = s->io.callback(s, s->io.fd, s->io.revents, s->userdata);
2291 case SOURCE_TIME_REALTIME:
2292 case SOURCE_TIME_BOOTTIME:
2293 case SOURCE_TIME_MONOTONIC:
2294 case SOURCE_TIME_REALTIME_ALARM:
2295 case SOURCE_TIME_BOOTTIME_ALARM:
2296 r = s->time.callback(s, s->time.next, s->userdata);
2300 r = s->signal.callback(s, &s->signal.siginfo, s->userdata);
2303 case SOURCE_CHILD: {
2306 zombie = s->child.siginfo.si_code == CLD_EXITED ||
2307 s->child.siginfo.si_code == CLD_KILLED ||
2308 s->child.siginfo.si_code == CLD_DUMPED;
2310 r = s->child.callback(s, &s->child.siginfo, s->userdata);
2312 /* Now, reap the PID for good. */
2314 waitid(P_PID, s->child.pid, &s->child.siginfo, WNOHANG|WEXITED);
2320 r = s->defer.callback(s, s->userdata);
2324 r = s->post.callback(s, s->userdata);
2328 r = s->exit.callback(s, s->userdata);
2331 case SOURCE_WATCHDOG:
2332 case _SOURCE_EVENT_SOURCE_TYPE_MAX:
2333 case _SOURCE_EVENT_SOURCE_TYPE_INVALID:
2334 assert_not_reached("Wut? I shouldn't exist.");
2337 s->dispatching = false;
2340 log_debug_errno(r, "Event source %s (type %s) returned error, disabling: %m",
2341 strna(s->description), event_source_type_to_string(s->type));
2346 sd_event_source_set_enabled(s, SD_EVENT_OFF);
2351 static int event_prepare(sd_event *e) {
2359 s = prioq_peek(e->prepare);
2360 if (!s || s->prepare_iteration == e->iteration || s->enabled == SD_EVENT_OFF)
2363 s->prepare_iteration = e->iteration;
2364 r = prioq_reshuffle(e->prepare, s, &s->prepare_index);
2370 s->dispatching = true;
2371 r = s->prepare(s, s->userdata);
2372 s->dispatching = false;
2375 log_debug_errno(r, "Prepare callback of event source %s (type %s) returned error, disabling: %m",
2376 strna(s->description), event_source_type_to_string(s->type));
2381 sd_event_source_set_enabled(s, SD_EVENT_OFF);
2387 static int dispatch_exit(sd_event *e) {
2393 p = prioq_peek(e->exit);
2394 if (!p || p->enabled == SD_EVENT_OFF) {
2395 e->state = SD_EVENT_FINISHED;
2401 e->state = SD_EVENT_EXITING;
2403 r = source_dispatch(p);
2405 e->state = SD_EVENT_INITIAL;
2411 static sd_event_source* event_next_pending(sd_event *e) {
2416 p = prioq_peek(e->pending);
2420 if (p->enabled == SD_EVENT_OFF)
2426 static int arm_watchdog(sd_event *e) {
2427 struct itimerspec its = {};
2432 assert(e->watchdog_fd >= 0);
2434 t = sleep_between(e,
2435 e->watchdog_last + (e->watchdog_period / 2),
2436 e->watchdog_last + (e->watchdog_period * 3 / 4));
2438 timespec_store(&its.it_value, t);
2440 /* Make sure we never set the watchdog to 0, which tells the
2441 * kernel to disable it. */
2442 if (its.it_value.tv_sec == 0 && its.it_value.tv_nsec == 0)
2443 its.it_value.tv_nsec = 1;
2445 r = timerfd_settime(e->watchdog_fd, TFD_TIMER_ABSTIME, &its, NULL);
2452 static int process_watchdog(sd_event *e) {
2458 /* Don't notify watchdog too often */
2459 if (e->watchdog_last + e->watchdog_period / 4 > e->timestamp.monotonic)
2462 sd_notify(false, "WATCHDOG=1");
2463 e->watchdog_last = e->timestamp.monotonic;
2465 return arm_watchdog(e);
2468 _public_ int sd_event_prepare(sd_event *e) {
2471 assert_return(e, -EINVAL);
2472 assert_return(!event_pid_changed(e), -ECHILD);
2473 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
2474 assert_return(e->state == SD_EVENT_INITIAL, -EBUSY);
2476 if (e->exit_requested)
2481 e->state = SD_EVENT_PREPARING;
2482 r = event_prepare(e);
2483 e->state = SD_EVENT_INITIAL;
2487 r = event_arm_timer(e, &e->realtime);
2491 r = event_arm_timer(e, &e->boottime);
2495 r = event_arm_timer(e, &e->monotonic);
2499 r = event_arm_timer(e, &e->realtime_alarm);
2503 r = event_arm_timer(e, &e->boottime_alarm);
2507 if (event_next_pending(e) || e->need_process_child)
2510 e->state = SD_EVENT_ARMED;
2515 e->state = SD_EVENT_ARMED;
2516 r = sd_event_wait(e, 0);
2518 e->state = SD_EVENT_ARMED;
2523 _public_ int sd_event_wait(sd_event *e, uint64_t timeout) {
2524 struct epoll_event *ev_queue;
2525 unsigned ev_queue_max;
2528 assert_return(e, -EINVAL);
2529 assert_return(!event_pid_changed(e), -ECHILD);
2530 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
2531 assert_return(e->state == SD_EVENT_ARMED, -EBUSY);
2533 if (e->exit_requested) {
2534 e->state = SD_EVENT_PENDING;
2538 ev_queue_max = MAX(e->n_sources, 1u);
2539 ev_queue = newa(struct epoll_event, ev_queue_max);
2541 m = epoll_wait(e->epoll_fd, ev_queue, ev_queue_max,
2542 timeout == (uint64_t) -1 ? -1 : (int) ((timeout + USEC_PER_MSEC - 1) / USEC_PER_MSEC));
2544 if (errno == EINTR) {
2545 e->state = SD_EVENT_PENDING;
2553 dual_timestamp_get(&e->timestamp);
2554 if (clock_boottime_supported())
2555 e->timestamp_boottime = now(CLOCK_BOOTTIME);
2557 for (i = 0; i < m; i++) {
2559 if (ev_queue[i].data.ptr == INT_TO_PTR(SOURCE_WATCHDOG))
2560 r = flush_timer(e, e->watchdog_fd, ev_queue[i].events, NULL);
2562 WakeupType *t = ev_queue[i].data.ptr;
2566 case WAKEUP_EVENT_SOURCE:
2567 r = process_io(e, ev_queue[i].data.ptr, ev_queue[i].events);
2570 case WAKEUP_CLOCK_DATA: {
2571 struct clock_data *d = ev_queue[i].data.ptr;
2572 r = flush_timer(e, d->fd, ev_queue[i].events, &d->next);
2576 case WAKEUP_SIGNAL_DATA:
2577 r = process_signal(e, ev_queue[i].data.ptr, ev_queue[i].events);
2581 assert_not_reached("Invalid wake-up pointer");
2588 r = process_watchdog(e);
2592 r = process_timer(e, e->timestamp.realtime, &e->realtime);
2596 r = process_timer(e, e->timestamp_boottime, &e->boottime);
2600 r = process_timer(e, e->timestamp.monotonic, &e->monotonic);
2604 r = process_timer(e, e->timestamp.realtime, &e->realtime_alarm);
2608 r = process_timer(e, e->timestamp_boottime, &e->boottime_alarm);
2612 if (e->need_process_child) {
2613 r = process_child(e);
2618 if (event_next_pending(e)) {
2619 e->state = SD_EVENT_PENDING;
2627 e->state = SD_EVENT_INITIAL;
2632 _public_ int sd_event_dispatch(sd_event *e) {
2636 assert_return(e, -EINVAL);
2637 assert_return(!event_pid_changed(e), -ECHILD);
2638 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
2639 assert_return(e->state == SD_EVENT_PENDING, -EBUSY);
2641 if (e->exit_requested)
2642 return dispatch_exit(e);
2644 p = event_next_pending(e);
2648 e->state = SD_EVENT_RUNNING;
2649 r = source_dispatch(p);
2650 e->state = SD_EVENT_INITIAL;
2657 e->state = SD_EVENT_INITIAL;
2662 static void event_log_delays(sd_event *e) {
2663 char b[ELEMENTSOF(e->delays) * DECIMAL_STR_MAX(unsigned) + 1];
2667 for (i = o = 0; i < ELEMENTSOF(e->delays); i++) {
2668 o += snprintf(&b[o], sizeof(b) - o, "%u ", e->delays[i]);
2671 log_debug("Event loop iterations: %.*s", o, b);
2674 _public_ int sd_event_run(sd_event *e, uint64_t timeout) {
2677 assert_return(e, -EINVAL);
2678 assert_return(!event_pid_changed(e), -ECHILD);
2679 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
2680 assert_return(e->state == SD_EVENT_INITIAL, -EBUSY);
2682 if (e->profile_delays && e->last_run) {
2686 this_run = now(CLOCK_MONOTONIC);
2688 l = u64log2(this_run - e->last_run);
2689 assert(l < sizeof(e->delays));
2692 if (this_run - e->last_log >= 5*USEC_PER_SEC) {
2693 event_log_delays(e);
2694 e->last_log = this_run;
2698 r = sd_event_prepare(e);
2700 /* There was nothing? Then wait... */
2701 r = sd_event_wait(e, timeout);
2703 if (e->profile_delays)
2704 e->last_run = now(CLOCK_MONOTONIC);
2707 /* There's something now, then let's dispatch it */
2708 r = sd_event_dispatch(e);
2718 #if 0 /// UNNEEDED by elogind
2719 _public_ int sd_event_loop(sd_event *e) {
2722 assert_return(e, -EINVAL);
2723 assert_return(!event_pid_changed(e), -ECHILD);
2724 assert_return(e->state == SD_EVENT_INITIAL, -EBUSY);
2728 while (e->state != SD_EVENT_FINISHED) {
2729 r = sd_event_run(e, (uint64_t) -1);
2741 _public_ int sd_event_get_fd(sd_event *e) {
2743 assert_return(e, -EINVAL);
2744 assert_return(!event_pid_changed(e), -ECHILD);
2750 _public_ int sd_event_get_state(sd_event *e) {
2751 assert_return(e, -EINVAL);
2752 assert_return(!event_pid_changed(e), -ECHILD);
2757 #if 0 /// UNNEEDED by elogind
2758 _public_ int sd_event_get_exit_code(sd_event *e, int *code) {
2759 assert_return(e, -EINVAL);
2760 assert_return(code, -EINVAL);
2761 assert_return(!event_pid_changed(e), -ECHILD);
2763 if (!e->exit_requested)
2766 *code = e->exit_code;
2771 _public_ int sd_event_exit(sd_event *e, int code) {
2772 assert_return(e, -EINVAL);
2773 assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
2774 assert_return(!event_pid_changed(e), -ECHILD);
2776 e->exit_requested = true;
2777 e->exit_code = code;
2782 #if 0 /// UNNEEDED by elogind
2783 _public_ int sd_event_now(sd_event *e, clockid_t clock, uint64_t *usec) {
2784 assert_return(e, -EINVAL);
2785 assert_return(usec, -EINVAL);
2786 assert_return(!event_pid_changed(e), -ECHILD);
2787 assert_return(IN_SET(clock,
2789 CLOCK_REALTIME_ALARM,
2792 CLOCK_BOOTTIME_ALARM), -EOPNOTSUPP);
2794 if (IN_SET(clock, CLOCK_BOOTTIME, CLOCK_BOOTTIME_ALARM) && !clock_boottime_supported())
2797 if (!dual_timestamp_is_set(&e->timestamp)) {
2798 /* Implicitly fall back to now() if we never ran
2799 * before and thus have no cached time. */
2806 case CLOCK_REALTIME:
2807 case CLOCK_REALTIME_ALARM:
2808 *usec = e->timestamp.realtime;
2811 case CLOCK_MONOTONIC:
2812 *usec = e->timestamp.monotonic;
2815 case CLOCK_BOOTTIME:
2816 case CLOCK_BOOTTIME_ALARM:
2817 *usec = e->timestamp_boottime;
2821 assert_not_reached("Unknown clock?");
2828 _public_ int sd_event_default(sd_event **ret) {
2830 static thread_local sd_event *default_event = NULL;
2835 return !!default_event;
2837 if (default_event) {
2838 *ret = sd_event_ref(default_event);
2842 r = sd_event_new(&e);
2846 e->default_event_ptr = &default_event;
2854 #if 0 /// UNNEEDED by elogind
2855 _public_ int sd_event_get_tid(sd_event *e, pid_t *tid) {
2856 assert_return(e, -EINVAL);
2857 assert_return(tid, -EINVAL);
2858 assert_return(!event_pid_changed(e), -ECHILD);
2869 _public_ int sd_event_set_watchdog(sd_event *e, int b) {
2872 assert_return(e, -EINVAL);
2873 assert_return(!event_pid_changed(e), -ECHILD);
2875 if (e->watchdog == !!b)
2879 struct epoll_event ev = {};
2881 r = sd_watchdog_enabled(false, &e->watchdog_period);
2885 /* Issue first ping immediately */
2886 sd_notify(false, "WATCHDOG=1");
2887 e->watchdog_last = now(CLOCK_MONOTONIC);
2889 e->watchdog_fd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK|TFD_CLOEXEC);
2890 if (e->watchdog_fd < 0)
2893 r = arm_watchdog(e);
2897 ev.events = EPOLLIN;
2898 ev.data.ptr = INT_TO_PTR(SOURCE_WATCHDOG);
2900 r = epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, e->watchdog_fd, &ev);
2907 if (e->watchdog_fd >= 0) {
2908 epoll_ctl(e->epoll_fd, EPOLL_CTL_DEL, e->watchdog_fd, NULL);
2909 e->watchdog_fd = safe_close(e->watchdog_fd);
2917 e->watchdog_fd = safe_close(e->watchdog_fd);
2921 #if 0 /// UNNEEDED by elogind
2922 _public_ int sd_event_get_watchdog(sd_event *e) {
2923 assert_return(e, -EINVAL);
2924 assert_return(!event_pid_changed(e), -ECHILD);