1 /* SPDX-License-Identifier: LGPL-2.1+ */
4 #include <sys/timerfd.h>
11 #include "alloc-util.h"
13 #include "event-source.h"
19 #include "memory-util.h"
20 #include "missing_syscall.h"
22 #include "process-util.h"
24 #include "signal-util.h"
25 #include "string-table.h"
26 #include "string-util.h"
28 #include "time-util.h"
30 #define DEFAULT_ACCURACY_USEC (250 * USEC_PER_MSEC)
32 static bool EVENT_SOURCE_WATCH_PIDFD(sd_event_source
*s
) {
33 /* Returns true if this is a PID event source and can be implemented by watching EPOLLIN */
35 s
->type
== SOURCE_CHILD
&&
36 s
->child
.pidfd
>= 0 &&
37 s
->child
.options
== WEXITED
;
40 static const char* const event_source_type_table
[_SOURCE_EVENT_SOURCE_TYPE_MAX
] = {
42 [SOURCE_TIME_REALTIME
] = "realtime",
43 [SOURCE_TIME_BOOTTIME
] = "bootime",
44 [SOURCE_TIME_MONOTONIC
] = "monotonic",
45 [SOURCE_TIME_REALTIME_ALARM
] = "realtime-alarm",
46 [SOURCE_TIME_BOOTTIME_ALARM
] = "boottime-alarm",
47 [SOURCE_SIGNAL
] = "signal",
48 [SOURCE_CHILD
] = "child",
49 [SOURCE_DEFER
] = "defer",
50 [SOURCE_POST
] = "post",
51 [SOURCE_EXIT
] = "exit",
52 [SOURCE_WATCHDOG
] = "watchdog",
53 [SOURCE_INOTIFY
] = "inotify",
56 DEFINE_PRIVATE_STRING_TABLE_LOOKUP_TO_STRING(event_source_type
, int);
58 #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)
69 /* timerfd_create() only supports these five clocks so far. We
70 * can add support for more clocks when the kernel learns to
71 * deal with them, too. */
72 struct clock_data realtime
;
73 struct clock_data boottime
;
74 struct clock_data monotonic
;
75 struct clock_data realtime_alarm
;
76 struct clock_data boottime_alarm
;
80 sd_event_source
**signal_sources
; /* indexed by signal number */
81 Hashmap
*signal_data
; /* indexed by priority */
83 Hashmap
*child_sources
;
84 unsigned n_enabled_child_sources
;
90 Hashmap
*inotify_data
; /* indexed by priority */
92 /* A list of inode structures that still have an fd open, that we need to close before the next loop iteration */
93 LIST_HEAD(struct inode_data
, inode_data_to_close
);
95 /* A list of inotify objects that already have events buffered which aren't processed yet */
96 LIST_HEAD(struct inotify_data
, inotify_data_buffered
);
101 triple_timestamp timestamp
;
104 bool exit_requested
:1;
105 bool need_process_child
:1;
107 bool profile_delays
:1;
112 sd_event
**default_event_ptr
;
114 usec_t watchdog_last
, watchdog_period
;
118 struct epoll_event
*event_queue
;
119 size_t event_queue_allocated
;
121 LIST_HEAD(sd_event_source
, sources
);
123 usec_t last_run
, last_log
;
124 unsigned delays
[sizeof(usec_t
) * 8];
127 static thread_local sd_event
*default_event
= NULL
;
129 static void source_disconnect(sd_event_source
*s
);
130 static void event_gc_inode_data(sd_event
*e
, struct inode_data
*d
);
132 static sd_event
*event_resolve(sd_event
*e
) {
133 return e
== SD_EVENT_DEFAULT
? default_event
: e
;
136 static int pending_prioq_compare(const void *a
, const void *b
) {
137 const sd_event_source
*x
= a
, *y
= b
;
143 /* Enabled ones first */
144 if (x
->enabled
!= SD_EVENT_OFF
&& y
->enabled
== SD_EVENT_OFF
)
146 if (x
->enabled
== SD_EVENT_OFF
&& y
->enabled
!= SD_EVENT_OFF
)
149 /* Lower priority values first */
150 r
= CMP(x
->priority
, y
->priority
);
154 /* Older entries first */
155 return CMP(x
->pending_iteration
, y
->pending_iteration
);
158 static int prepare_prioq_compare(const void *a
, const void *b
) {
159 const sd_event_source
*x
= a
, *y
= b
;
165 /* Enabled ones first */
166 if (x
->enabled
!= SD_EVENT_OFF
&& y
->enabled
== SD_EVENT_OFF
)
168 if (x
->enabled
== SD_EVENT_OFF
&& y
->enabled
!= SD_EVENT_OFF
)
171 /* Move most recently prepared ones last, so that we can stop
172 * preparing as soon as we hit one that has already been
173 * prepared in the current iteration */
174 r
= CMP(x
->prepare_iteration
, y
->prepare_iteration
);
178 /* Lower priority values first */
179 return CMP(x
->priority
, y
->priority
);
182 static int earliest_time_prioq_compare(const void *a
, const void *b
) {
183 const sd_event_source
*x
= a
, *y
= b
;
185 assert(EVENT_SOURCE_IS_TIME(x
->type
));
186 assert(x
->type
== y
->type
);
188 /* Enabled ones first */
189 if (x
->enabled
!= SD_EVENT_OFF
&& y
->enabled
== SD_EVENT_OFF
)
191 if (x
->enabled
== SD_EVENT_OFF
&& y
->enabled
!= SD_EVENT_OFF
)
194 /* Move the pending ones to the end */
195 if (!x
->pending
&& y
->pending
)
197 if (x
->pending
&& !y
->pending
)
201 return CMP(x
->time
.next
, y
->time
.next
);
204 static usec_t
time_event_source_latest(const sd_event_source
*s
) {
205 return usec_add(s
->time
.next
, s
->time
.accuracy
);
208 static int latest_time_prioq_compare(const void *a
, const void *b
) {
209 const sd_event_source
*x
= a
, *y
= b
;
211 assert(EVENT_SOURCE_IS_TIME(x
->type
));
212 assert(x
->type
== y
->type
);
214 /* Enabled ones first */
215 if (x
->enabled
!= SD_EVENT_OFF
&& y
->enabled
== SD_EVENT_OFF
)
217 if (x
->enabled
== SD_EVENT_OFF
&& y
->enabled
!= SD_EVENT_OFF
)
220 /* Move the pending ones to the end */
221 if (!x
->pending
&& y
->pending
)
223 if (x
->pending
&& !y
->pending
)
227 return CMP(time_event_source_latest(x
), time_event_source_latest(y
));
230 static int exit_prioq_compare(const void *a
, const void *b
) {
231 const sd_event_source
*x
= a
, *y
= b
;
233 assert(x
->type
== SOURCE_EXIT
);
234 assert(y
->type
== SOURCE_EXIT
);
236 /* Enabled ones first */
237 if (x
->enabled
!= SD_EVENT_OFF
&& y
->enabled
== SD_EVENT_OFF
)
239 if (x
->enabled
== SD_EVENT_OFF
&& y
->enabled
!= SD_EVENT_OFF
)
242 /* Lower priority values first */
243 return CMP(x
->priority
, y
->priority
);
246 static void free_clock_data(struct clock_data
*d
) {
248 assert(d
->wakeup
== WAKEUP_CLOCK_DATA
);
251 prioq_free(d
->earliest
);
252 prioq_free(d
->latest
);
255 static sd_event
*event_free(sd_event
*e
) {
260 while ((s
= e
->sources
)) {
262 source_disconnect(s
);
263 sd_event_source_unref(s
);
266 assert(e
->n_sources
== 0);
268 if (e
->default_event_ptr
)
269 *(e
->default_event_ptr
) = NULL
;
271 safe_close(e
->epoll_fd
);
272 safe_close(e
->watchdog_fd
);
274 free_clock_data(&e
->realtime
);
275 free_clock_data(&e
->boottime
);
276 free_clock_data(&e
->monotonic
);
277 free_clock_data(&e
->realtime_alarm
);
278 free_clock_data(&e
->boottime_alarm
);
280 prioq_free(e
->pending
);
281 prioq_free(e
->prepare
);
284 free(e
->signal_sources
);
285 hashmap_free(e
->signal_data
);
287 hashmap_free(e
->inotify_data
);
289 hashmap_free(e
->child_sources
);
290 set_free(e
->post_sources
);
292 free(e
->event_queue
);
297 _public_
int sd_event_new(sd_event
** ret
) {
301 assert_return(ret
, -EINVAL
);
303 e
= new(sd_event
, 1);
311 .realtime
.wakeup
= WAKEUP_CLOCK_DATA
,
313 .realtime
.next
= USEC_INFINITY
,
314 .boottime
.wakeup
= WAKEUP_CLOCK_DATA
,
316 .boottime
.next
= USEC_INFINITY
,
317 .monotonic
.wakeup
= WAKEUP_CLOCK_DATA
,
319 .monotonic
.next
= USEC_INFINITY
,
320 .realtime_alarm
.wakeup
= WAKEUP_CLOCK_DATA
,
321 .realtime_alarm
.fd
= -1,
322 .realtime_alarm
.next
= USEC_INFINITY
,
323 .boottime_alarm
.wakeup
= WAKEUP_CLOCK_DATA
,
324 .boottime_alarm
.fd
= -1,
325 .boottime_alarm
.next
= USEC_INFINITY
,
326 .perturb
= USEC_INFINITY
,
327 .original_pid
= getpid_cached(),
330 r
= prioq_ensure_allocated(&e
->pending
, pending_prioq_compare
);
334 e
->epoll_fd
= epoll_create1(EPOLL_CLOEXEC
);
335 if (e
->epoll_fd
< 0) {
340 e
->epoll_fd
= fd_move_above_stdio(e
->epoll_fd
);
342 if (secure_getenv("SD_EVENT_PROFILE_DELAYS")) {
343 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.");
344 e
->profile_delays
= true;
355 DEFINE_PUBLIC_TRIVIAL_REF_UNREF_FUNC(sd_event
, sd_event
, event_free
);
357 _public_ sd_event_source
* sd_event_source_disable_unref(sd_event_source
*s
) {
359 (void) sd_event_source_set_enabled(s
, SD_EVENT_OFF
);
360 return sd_event_source_unref(s
);
363 static bool event_pid_changed(sd_event
*e
) {
366 /* We don't support people creating an event loop and keeping
367 * it around over a fork(). Let's complain. */
369 return e
->original_pid
!= getpid_cached();
372 static void source_io_unregister(sd_event_source
*s
) {
374 assert(s
->type
== SOURCE_IO
);
376 if (event_pid_changed(s
->event
))
379 if (!s
->io
.registered
)
382 if (epoll_ctl(s
->event
->epoll_fd
, EPOLL_CTL_DEL
, s
->io
.fd
, NULL
) < 0)
383 log_debug_errno(errno
, "Failed to remove source %s (type %s) from epoll: %m",
384 strna(s
->description
), event_source_type_to_string(s
->type
));
386 s
->io
.registered
= false;
389 static int source_io_register(
395 assert(s
->type
== SOURCE_IO
);
396 assert(enabled
!= SD_EVENT_OFF
);
398 struct epoll_event ev
= {
399 .events
= events
| (enabled
== SD_EVENT_ONESHOT
? EPOLLONESHOT
: 0),
404 r
= epoll_ctl(s
->event
->epoll_fd
,
405 s
->io
.registered
? EPOLL_CTL_MOD
: EPOLL_CTL_ADD
,
411 s
->io
.registered
= true;
416 static void source_child_pidfd_unregister(sd_event_source
*s
) {
418 assert(s
->type
== SOURCE_CHILD
);
420 if (event_pid_changed(s
->event
))
423 if (!s
->child
.registered
)
426 if (EVENT_SOURCE_WATCH_PIDFD(s
))
427 if (epoll_ctl(s
->event
->epoll_fd
, EPOLL_CTL_DEL
, s
->child
.pidfd
, NULL
) < 0)
428 log_debug_errno(errno
, "Failed to remove source %s (type %s) from epoll: %m",
429 strna(s
->description
), event_source_type_to_string(s
->type
));
431 s
->child
.registered
= false;
434 static int source_child_pidfd_register(sd_event_source
*s
, int enabled
) {
438 assert(s
->type
== SOURCE_CHILD
);
439 assert(enabled
!= SD_EVENT_OFF
);
441 if (EVENT_SOURCE_WATCH_PIDFD(s
)) {
442 struct epoll_event ev
= {
443 .events
= EPOLLIN
| (enabled
== SD_EVENT_ONESHOT
? EPOLLONESHOT
: 0),
447 if (s
->child
.registered
)
448 r
= epoll_ctl(s
->event
->epoll_fd
, EPOLL_CTL_MOD
, s
->child
.pidfd
, &ev
);
450 r
= epoll_ctl(s
->event
->epoll_fd
, EPOLL_CTL_ADD
, s
->child
.pidfd
, &ev
);
455 s
->child
.registered
= true;
459 static clockid_t
event_source_type_to_clock(EventSourceType t
) {
463 case SOURCE_TIME_REALTIME
:
464 return CLOCK_REALTIME
;
466 case SOURCE_TIME_BOOTTIME
:
467 return CLOCK_BOOTTIME
;
469 case SOURCE_TIME_MONOTONIC
:
470 return CLOCK_MONOTONIC
;
472 case SOURCE_TIME_REALTIME_ALARM
:
473 return CLOCK_REALTIME_ALARM
;
475 case SOURCE_TIME_BOOTTIME_ALARM
:
476 return CLOCK_BOOTTIME_ALARM
;
479 return (clockid_t
) -1;
483 static EventSourceType
clock_to_event_source_type(clockid_t clock
) {
488 return SOURCE_TIME_REALTIME
;
491 return SOURCE_TIME_BOOTTIME
;
493 case CLOCK_MONOTONIC
:
494 return SOURCE_TIME_MONOTONIC
;
496 case CLOCK_REALTIME_ALARM
:
497 return SOURCE_TIME_REALTIME_ALARM
;
499 case CLOCK_BOOTTIME_ALARM
:
500 return SOURCE_TIME_BOOTTIME_ALARM
;
503 return _SOURCE_EVENT_SOURCE_TYPE_INVALID
;
507 static struct clock_data
* event_get_clock_data(sd_event
*e
, EventSourceType t
) {
512 case SOURCE_TIME_REALTIME
:
515 case SOURCE_TIME_BOOTTIME
:
518 case SOURCE_TIME_MONOTONIC
:
519 return &e
->monotonic
;
521 case SOURCE_TIME_REALTIME_ALARM
:
522 return &e
->realtime_alarm
;
524 case SOURCE_TIME_BOOTTIME_ALARM
:
525 return &e
->boottime_alarm
;
532 static void event_free_signal_data(sd_event
*e
, struct signal_data
*d
) {
538 hashmap_remove(e
->signal_data
, &d
->priority
);
543 static int event_make_signal_data(
546 struct signal_data
**ret
) {
548 struct signal_data
*d
;
556 if (event_pid_changed(e
))
559 if (e
->signal_sources
&& e
->signal_sources
[sig
])
560 priority
= e
->signal_sources
[sig
]->priority
;
562 priority
= SD_EVENT_PRIORITY_NORMAL
;
564 d
= hashmap_get(e
->signal_data
, &priority
);
566 if (sigismember(&d
->sigset
, sig
) > 0) {
572 r
= hashmap_ensure_allocated(&e
->signal_data
, &uint64_hash_ops
);
576 d
= new(struct signal_data
, 1);
580 *d
= (struct signal_data
) {
581 .wakeup
= WAKEUP_SIGNAL_DATA
,
583 .priority
= priority
,
586 r
= hashmap_put(e
->signal_data
, &d
->priority
, d
);
596 assert_se(sigaddset(&ss_copy
, sig
) >= 0);
598 r
= signalfd(d
->fd
, &ss_copy
, SFD_NONBLOCK
|SFD_CLOEXEC
);
612 d
->fd
= fd_move_above_stdio(r
);
614 struct epoll_event ev
= {
619 r
= epoll_ctl(e
->epoll_fd
, EPOLL_CTL_ADD
, d
->fd
, &ev
);
632 event_free_signal_data(e
, d
);
637 static void event_unmask_signal_data(sd_event
*e
, struct signal_data
*d
, int sig
) {
641 /* Turns off the specified signal in the signal data
642 * object. If the signal mask of the object becomes empty that
645 if (sigismember(&d
->sigset
, sig
) == 0)
648 assert_se(sigdelset(&d
->sigset
, sig
) >= 0);
650 if (sigisemptyset(&d
->sigset
)) {
651 /* If all the mask is all-zero we can get rid of the structure */
652 event_free_signal_data(e
, d
);
658 if (signalfd(d
->fd
, &d
->sigset
, SFD_NONBLOCK
|SFD_CLOEXEC
) < 0)
659 log_debug_errno(errno
, "Failed to unset signal bit, ignoring: %m");
662 static void event_gc_signal_data(sd_event
*e
, const int64_t *priority
, int sig
) {
663 struct signal_data
*d
;
664 static const int64_t zero_priority
= 0;
668 /* Rechecks if the specified signal is still something we are interested in. If not, we'll unmask it,
669 * and possibly drop the signalfd for it. */
671 if (sig
== SIGCHLD
&&
672 e
->n_enabled_child_sources
> 0)
675 if (e
->signal_sources
&&
676 e
->signal_sources
[sig
] &&
677 e
->signal_sources
[sig
]->enabled
!= SD_EVENT_OFF
)
681 * The specified signal might be enabled in three different queues:
683 * 1) the one that belongs to the priority passed (if it is non-NULL)
684 * 2) the one that belongs to the priority of the event source of the signal (if there is one)
685 * 3) the 0 priority (to cover the SIGCHLD case)
687 * Hence, let's remove it from all three here.
691 d
= hashmap_get(e
->signal_data
, priority
);
693 event_unmask_signal_data(e
, d
, sig
);
696 if (e
->signal_sources
&& e
->signal_sources
[sig
]) {
697 d
= hashmap_get(e
->signal_data
, &e
->signal_sources
[sig
]->priority
);
699 event_unmask_signal_data(e
, d
, sig
);
702 d
= hashmap_get(e
->signal_data
, &zero_priority
);
704 event_unmask_signal_data(e
, d
, sig
);
707 static void source_disconnect(sd_event_source
*s
) {
715 assert(s
->event
->n_sources
> 0);
721 source_io_unregister(s
);
725 case SOURCE_TIME_REALTIME
:
726 case SOURCE_TIME_BOOTTIME
:
727 case SOURCE_TIME_MONOTONIC
:
728 case SOURCE_TIME_REALTIME_ALARM
:
729 case SOURCE_TIME_BOOTTIME_ALARM
: {
730 struct clock_data
*d
;
732 d
= event_get_clock_data(s
->event
, s
->type
);
735 prioq_remove(d
->earliest
, s
, &s
->time
.earliest_index
);
736 prioq_remove(d
->latest
, s
, &s
->time
.latest_index
);
737 d
->needs_rearm
= true;
742 if (s
->signal
.sig
> 0) {
744 if (s
->event
->signal_sources
)
745 s
->event
->signal_sources
[s
->signal
.sig
] = NULL
;
747 event_gc_signal_data(s
->event
, &s
->priority
, s
->signal
.sig
);
753 if (s
->child
.pid
> 0) {
754 if (s
->enabled
!= SD_EVENT_OFF
) {
755 assert(s
->event
->n_enabled_child_sources
> 0);
756 s
->event
->n_enabled_child_sources
--;
759 (void) hashmap_remove(s
->event
->child_sources
, PID_TO_PTR(s
->child
.pid
));
762 if (EVENT_SOURCE_WATCH_PIDFD(s
))
763 source_child_pidfd_unregister(s
);
765 event_gc_signal_data(s
->event
, &s
->priority
, SIGCHLD
);
774 set_remove(s
->event
->post_sources
, s
);
778 prioq_remove(s
->event
->exit
, s
, &s
->exit
.prioq_index
);
781 case SOURCE_INOTIFY
: {
782 struct inode_data
*inode_data
;
784 inode_data
= s
->inotify
.inode_data
;
786 struct inotify_data
*inotify_data
;
787 assert_se(inotify_data
= inode_data
->inotify_data
);
789 /* Detach this event source from the inode object */
790 LIST_REMOVE(inotify
.by_inode_data
, inode_data
->event_sources
, s
);
791 s
->inotify
.inode_data
= NULL
;
794 assert(inotify_data
->n_pending
> 0);
795 inotify_data
->n_pending
--;
798 /* Note that we don't reduce the inotify mask for the watch descriptor here if the inode is
799 * continued to being watched. That's because inotify doesn't really have an API for that: we
800 * can only change watch masks with access to the original inode either by fd or by path. But
801 * paths aren't stable, and keeping an O_PATH fd open all the time would mean wasting an fd
802 * continuously and keeping the mount busy which we can't really do. We could reconstruct the
803 * original inode from /proc/self/fdinfo/$INOTIFY_FD (as all watch descriptors are listed
804 * there), but given the need for open_by_handle_at() which is privileged and not universally
805 * available this would be quite an incomplete solution. Hence we go the other way, leave the
806 * mask set, even if it is not minimized now, and ignore all events we aren't interested in
807 * anymore after reception. Yes, this sucks, but … Linux … */
809 /* Maybe release the inode data (and its inotify) */
810 event_gc_inode_data(s
->event
, inode_data
);
817 assert_not_reached("Wut? I shouldn't exist.");
821 prioq_remove(s
->event
->pending
, s
, &s
->pending_index
);
824 prioq_remove(s
->event
->prepare
, s
, &s
->prepare_index
);
826 event
= TAKE_PTR(s
->event
);
827 LIST_REMOVE(sources
, event
->sources
, s
);
830 /* Note that we don't invalidate the type here, since we still need it in order to close the fd or
831 * pidfd associated with this event source, which we'll do only on source_free(). */
834 sd_event_unref(event
);
837 static void source_free(sd_event_source
*s
) {
840 source_disconnect(s
);
842 if (s
->type
== SOURCE_IO
&& s
->io
.owned
)
843 s
->io
.fd
= safe_close(s
->io
.fd
);
845 if (s
->type
== SOURCE_CHILD
) {
846 /* Eventually the kernel will do this automatically for us, but for now let's emulate this (unreliably) in userspace. */
848 if (s
->child
.process_owned
) {
850 if (!s
->child
.exited
) {
853 if (s
->child
.pidfd
>= 0) {
854 if (pidfd_send_signal(s
->child
.pidfd
, SIGKILL
, NULL
, 0) < 0) {
855 if (errno
== ESRCH
) /* Already dead */
857 else if (!ERRNO_IS_NOT_SUPPORTED(errno
))
858 log_debug_errno(errno
, "Failed to kill process " PID_FMT
" via pidfd_send_signal(), re-trying via kill(): %m",
865 if (kill(s
->child
.pid
, SIGKILL
) < 0)
866 if (errno
!= ESRCH
) /* Already dead */
867 log_debug_errno(errno
, "Failed to kill process " PID_FMT
" via kill(), ignoring: %m",
871 if (!s
->child
.waited
) {
874 /* Reap the child if we can */
875 (void) waitid(P_PID
, s
->child
.pid
, &si
, WEXITED
);
879 if (s
->child
.pidfd_owned
)
880 s
->child
.pidfd
= safe_close(s
->child
.pidfd
);
883 if (s
->destroy_callback
)
884 s
->destroy_callback(s
->userdata
);
886 free(s
->description
);
889 DEFINE_TRIVIAL_CLEANUP_FUNC(sd_event_source
*, source_free
);
891 static int source_set_pending(sd_event_source
*s
, bool b
) {
895 assert(s
->type
!= SOURCE_EXIT
);
903 s
->pending_iteration
= s
->event
->iteration
;
905 r
= prioq_put(s
->event
->pending
, s
, &s
->pending_index
);
911 assert_se(prioq_remove(s
->event
->pending
, s
, &s
->pending_index
));
913 if (EVENT_SOURCE_IS_TIME(s
->type
)) {
914 struct clock_data
*d
;
916 d
= event_get_clock_data(s
->event
, s
->type
);
919 prioq_reshuffle(d
->earliest
, s
, &s
->time
.earliest_index
);
920 prioq_reshuffle(d
->latest
, s
, &s
->time
.latest_index
);
921 d
->needs_rearm
= true;
924 if (s
->type
== SOURCE_SIGNAL
&& !b
) {
925 struct signal_data
*d
;
927 d
= hashmap_get(s
->event
->signal_data
, &s
->priority
);
928 if (d
&& d
->current
== s
)
932 if (s
->type
== SOURCE_INOTIFY
) {
934 assert(s
->inotify
.inode_data
);
935 assert(s
->inotify
.inode_data
->inotify_data
);
938 s
->inotify
.inode_data
->inotify_data
->n_pending
++;
940 assert(s
->inotify
.inode_data
->inotify_data
->n_pending
> 0);
941 s
->inotify
.inode_data
->inotify_data
->n_pending
--;
948 static sd_event_source
*source_new(sd_event
*e
, bool floating
, EventSourceType type
) {
953 s
= new(sd_event_source
, 1);
957 *s
= (struct sd_event_source
) {
960 .floating
= floating
,
962 .pending_index
= PRIOQ_IDX_NULL
,
963 .prepare_index
= PRIOQ_IDX_NULL
,
969 LIST_PREPEND(sources
, e
->sources
, s
);
975 _public_
int sd_event_add_io(
977 sd_event_source
**ret
,
980 sd_event_io_handler_t callback
,
983 _cleanup_(source_freep
) sd_event_source
*s
= NULL
;
986 assert_return(e
, -EINVAL
);
987 assert_return(e
= event_resolve(e
), -ENOPKG
);
988 assert_return(fd
>= 0, -EBADF
);
989 assert_return(!(events
& ~(EPOLLIN
|EPOLLOUT
|EPOLLRDHUP
|EPOLLPRI
|EPOLLERR
|EPOLLHUP
|EPOLLET
)), -EINVAL
);
990 assert_return(callback
, -EINVAL
);
991 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
992 assert_return(!event_pid_changed(e
), -ECHILD
);
994 s
= source_new(e
, !ret
, SOURCE_IO
);
998 s
->wakeup
= WAKEUP_EVENT_SOURCE
;
1000 s
->io
.events
= events
;
1001 s
->io
.callback
= callback
;
1002 s
->userdata
= userdata
;
1003 s
->enabled
= SD_EVENT_ON
;
1005 r
= source_io_register(s
, s
->enabled
, events
);
1016 static void initialize_perturb(sd_event
*e
) {
1017 sd_id128_t bootid
= {};
1019 /* When we sleep for longer, we try to realign the wakeup to
1020 the same time within each minute/second/250ms, so that
1021 events all across the system can be coalesced into a single
1022 CPU wakeup. However, let's take some system-specific
1023 randomness for this value, so that in a network of systems
1024 with synced clocks timer events are distributed a
1025 bit. Here, we calculate a perturbation usec offset from the
1028 if (_likely_(e
->perturb
!= USEC_INFINITY
))
1031 if (sd_id128_get_boot(&bootid
) >= 0)
1032 e
->perturb
= (bootid
.qwords
[0] ^ bootid
.qwords
[1]) % USEC_PER_MINUTE
;
1035 static int event_setup_timer_fd(
1037 struct clock_data
*d
,
1043 if (_likely_(d
->fd
>= 0))
1046 _cleanup_close_
int fd
= -1;
1049 fd
= timerfd_create(clock
, TFD_NONBLOCK
|TFD_CLOEXEC
);
1053 fd
= fd_move_above_stdio(fd
);
1055 struct epoll_event ev
= {
1060 r
= epoll_ctl(e
->epoll_fd
, EPOLL_CTL_ADD
, fd
, &ev
);
1064 d
->fd
= TAKE_FD(fd
);
1068 static int time_exit_callback(sd_event_source
*s
, uint64_t usec
, void *userdata
) {
1071 return sd_event_exit(sd_event_source_get_event(s
), PTR_TO_INT(userdata
));
1074 _public_
int sd_event_add_time(
1076 sd_event_source
**ret
,
1080 sd_event_time_handler_t callback
,
1083 EventSourceType type
;
1084 _cleanup_(source_freep
) sd_event_source
*s
= NULL
;
1085 struct clock_data
*d
;
1088 assert_return(e
, -EINVAL
);
1089 assert_return(e
= event_resolve(e
), -ENOPKG
);
1090 assert_return(accuracy
!= (uint64_t) -1, -EINVAL
);
1091 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
1092 assert_return(!event_pid_changed(e
), -ECHILD
);
1094 if (!clock_supported(clock
)) /* Checks whether the kernel supports the clock */
1097 type
= clock_to_event_source_type(clock
); /* checks whether sd-event supports this clock */
1102 callback
= time_exit_callback
;
1104 d
= event_get_clock_data(e
, type
);
1107 r
= prioq_ensure_allocated(&d
->earliest
, earliest_time_prioq_compare
);
1111 r
= prioq_ensure_allocated(&d
->latest
, latest_time_prioq_compare
);
1116 r
= event_setup_timer_fd(e
, d
, clock
);
1121 s
= source_new(e
, !ret
, type
);
1125 s
->time
.next
= usec
;
1126 s
->time
.accuracy
= accuracy
== 0 ? DEFAULT_ACCURACY_USEC
: accuracy
;
1127 s
->time
.callback
= callback
;
1128 s
->time
.earliest_index
= s
->time
.latest_index
= PRIOQ_IDX_NULL
;
1129 s
->userdata
= userdata
;
1130 s
->enabled
= SD_EVENT_ONESHOT
;
1132 d
->needs_rearm
= true;
1134 r
= prioq_put(d
->earliest
, s
, &s
->time
.earliest_index
);
1138 r
= prioq_put(d
->latest
, s
, &s
->time
.latest_index
);
1149 static int signal_exit_callback(sd_event_source
*s
, const struct signalfd_siginfo
*si
, void *userdata
) {
1152 return sd_event_exit(sd_event_source_get_event(s
), PTR_TO_INT(userdata
));
1155 _public_
int sd_event_add_signal(
1157 sd_event_source
**ret
,
1159 sd_event_signal_handler_t callback
,
1162 _cleanup_(source_freep
) sd_event_source
*s
= NULL
;
1163 struct signal_data
*d
;
1166 assert_return(e
, -EINVAL
);
1167 assert_return(e
= event_resolve(e
), -ENOPKG
);
1168 assert_return(SIGNAL_VALID(sig
), -EINVAL
);
1169 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
1170 assert_return(!event_pid_changed(e
), -ECHILD
);
1173 callback
= signal_exit_callback
;
1175 r
= signal_is_blocked(sig
);
1181 if (!e
->signal_sources
) {
1182 e
->signal_sources
= new0(sd_event_source
*, _NSIG
);
1183 if (!e
->signal_sources
)
1185 } else if (e
->signal_sources
[sig
])
1188 s
= source_new(e
, !ret
, SOURCE_SIGNAL
);
1192 s
->signal
.sig
= sig
;
1193 s
->signal
.callback
= callback
;
1194 s
->userdata
= userdata
;
1195 s
->enabled
= SD_EVENT_ON
;
1197 e
->signal_sources
[sig
] = s
;
1199 r
= event_make_signal_data(e
, sig
, &d
);
1203 /* Use the signal name as description for the event source by default */
1204 (void) sd_event_source_set_description(s
, signal_to_string(sig
));
1213 static bool shall_use_pidfd(void) {
1214 /* Mostly relevant for debugging, i.e. this is used in test-event.c to test the event loop once with and once without pidfd */
1215 return getenv_bool_secure("SYSTEMD_PIDFD") != 0;
1218 _public_
int sd_event_add_child(
1220 sd_event_source
**ret
,
1223 sd_event_child_handler_t callback
,
1226 _cleanup_(source_freep
) sd_event_source
*s
= NULL
;
1229 assert_return(e
, -EINVAL
);
1230 assert_return(e
= event_resolve(e
), -ENOPKG
);
1231 assert_return(pid
> 1, -EINVAL
);
1232 assert_return(!(options
& ~(WEXITED
|WSTOPPED
|WCONTINUED
)), -EINVAL
);
1233 assert_return(options
!= 0, -EINVAL
);
1234 assert_return(callback
, -EINVAL
);
1235 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
1236 assert_return(!event_pid_changed(e
), -ECHILD
);
1238 if (e
->n_enabled_child_sources
== 0) {
1239 /* Caller must block SIGCHLD before using us to watch children, even if pidfd is available,
1240 * for compatibility with pre-pidfd and because we don't want the reap the child processes
1241 * ourselves, i.e. call waitid(), and don't want Linux' default internal logic for that to
1244 * (As an optimization we only do this check on the first child event source created.) */
1245 r
= signal_is_blocked(SIGCHLD
);
1252 r
= hashmap_ensure_allocated(&e
->child_sources
, NULL
);
1256 if (hashmap_contains(e
->child_sources
, PID_TO_PTR(pid
)))
1259 s
= source_new(e
, !ret
, SOURCE_CHILD
);
1263 s
->wakeup
= WAKEUP_EVENT_SOURCE
;
1265 s
->child
.options
= options
;
1266 s
->child
.callback
= callback
;
1267 s
->userdata
= userdata
;
1268 s
->enabled
= SD_EVENT_ONESHOT
;
1270 /* We always take a pidfd here if we can, even if we wait for anything else than WEXITED, so that we
1271 * pin the PID, and make regular waitid() handling race-free. */
1273 if (shall_use_pidfd()) {
1274 s
->child
.pidfd
= pidfd_open(s
->child
.pid
, 0);
1275 if (s
->child
.pidfd
< 0) {
1276 /* Propagate errors unless the syscall is not supported or blocked */
1277 if (!ERRNO_IS_NOT_SUPPORTED(errno
) && !ERRNO_IS_PRIVILEGE(errno
))
1280 s
->child
.pidfd_owned
= true; /* If we allocate the pidfd we own it by default */
1282 s
->child
.pidfd
= -1;
1284 r
= hashmap_put(e
->child_sources
, PID_TO_PTR(pid
), s
);
1288 e
->n_enabled_child_sources
++;
1290 if (EVENT_SOURCE_WATCH_PIDFD(s
)) {
1291 /* We have a pidfd and we only want to watch for exit */
1293 r
= source_child_pidfd_register(s
, s
->enabled
);
1295 e
->n_enabled_child_sources
--;
1299 /* We have no pidfd or we shall wait for some other event than WEXITED */
1301 r
= event_make_signal_data(e
, SIGCHLD
, NULL
);
1303 e
->n_enabled_child_sources
--;
1307 e
->need_process_child
= true;
1317 _public_
int sd_event_add_child_pidfd(
1319 sd_event_source
**ret
,
1322 sd_event_child_handler_t callback
,
1326 _cleanup_(source_freep
) sd_event_source
*s
= NULL
;
1330 assert_return(e
, -EINVAL
);
1331 assert_return(e
= event_resolve(e
), -ENOPKG
);
1332 assert_return(pidfd
>= 0, -EBADF
);
1333 assert_return(!(options
& ~(WEXITED
|WSTOPPED
|WCONTINUED
)), -EINVAL
);
1334 assert_return(options
!= 0, -EINVAL
);
1335 assert_return(callback
, -EINVAL
);
1336 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
1337 assert_return(!event_pid_changed(e
), -ECHILD
);
1339 if (e
->n_enabled_child_sources
== 0) {
1340 r
= signal_is_blocked(SIGCHLD
);
1347 r
= hashmap_ensure_allocated(&e
->child_sources
, NULL
);
1351 r
= pidfd_get_pid(pidfd
, &pid
);
1355 if (hashmap_contains(e
->child_sources
, PID_TO_PTR(pid
)))
1358 s
= source_new(e
, !ret
, SOURCE_CHILD
);
1362 s
->wakeup
= WAKEUP_EVENT_SOURCE
;
1363 s
->child
.pidfd
= pidfd
;
1365 s
->child
.options
= options
;
1366 s
->child
.callback
= callback
;
1367 s
->child
.pidfd_owned
= false; /* If we got the pidfd passed in we don't own it by default (similar to the IO fd case) */
1368 s
->userdata
= userdata
;
1369 s
->enabled
= SD_EVENT_ONESHOT
;
1371 r
= hashmap_put(e
->child_sources
, PID_TO_PTR(pid
), s
);
1375 e
->n_enabled_child_sources
++;
1377 if (EVENT_SOURCE_WATCH_PIDFD(s
)) {
1378 /* We only want to watch for WEXITED */
1380 r
= source_child_pidfd_register(s
, s
->enabled
);
1382 e
->n_enabled_child_sources
--;
1386 /* We shall wait for some other event than WEXITED */
1388 r
= event_make_signal_data(e
, SIGCHLD
, NULL
);
1390 e
->n_enabled_child_sources
--;
1394 e
->need_process_child
= true;
1404 _public_
int sd_event_add_defer(
1406 sd_event_source
**ret
,
1407 sd_event_handler_t callback
,
1410 _cleanup_(source_freep
) sd_event_source
*s
= NULL
;
1413 assert_return(e
, -EINVAL
);
1414 assert_return(e
= event_resolve(e
), -ENOPKG
);
1415 assert_return(callback
, -EINVAL
);
1416 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
1417 assert_return(!event_pid_changed(e
), -ECHILD
);
1419 s
= source_new(e
, !ret
, SOURCE_DEFER
);
1423 s
->defer
.callback
= callback
;
1424 s
->userdata
= userdata
;
1425 s
->enabled
= SD_EVENT_ONESHOT
;
1427 r
= source_set_pending(s
, true);
1438 _public_
int sd_event_add_post(
1440 sd_event_source
**ret
,
1441 sd_event_handler_t callback
,
1444 _cleanup_(source_freep
) sd_event_source
*s
= NULL
;
1447 assert_return(e
, -EINVAL
);
1448 assert_return(e
= event_resolve(e
), -ENOPKG
);
1449 assert_return(callback
, -EINVAL
);
1450 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
1451 assert_return(!event_pid_changed(e
), -ECHILD
);
1453 s
= source_new(e
, !ret
, SOURCE_POST
);
1457 s
->post
.callback
= callback
;
1458 s
->userdata
= userdata
;
1459 s
->enabled
= SD_EVENT_ON
;
1461 r
= set_ensure_put(&e
->post_sources
, NULL
, s
);
1473 _public_
int sd_event_add_exit(
1475 sd_event_source
**ret
,
1476 sd_event_handler_t callback
,
1479 _cleanup_(source_freep
) sd_event_source
*s
= NULL
;
1482 assert_return(e
, -EINVAL
);
1483 assert_return(e
= event_resolve(e
), -ENOPKG
);
1484 assert_return(callback
, -EINVAL
);
1485 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
1486 assert_return(!event_pid_changed(e
), -ECHILD
);
1488 r
= prioq_ensure_allocated(&e
->exit
, exit_prioq_compare
);
1492 s
= source_new(e
, !ret
, SOURCE_EXIT
);
1496 s
->exit
.callback
= callback
;
1497 s
->userdata
= userdata
;
1498 s
->exit
.prioq_index
= PRIOQ_IDX_NULL
;
1499 s
->enabled
= SD_EVENT_ONESHOT
;
1501 r
= prioq_put(s
->event
->exit
, s
, &s
->exit
.prioq_index
);
1512 static void event_free_inotify_data(sd_event
*e
, struct inotify_data
*d
) {
1518 assert(hashmap_isempty(d
->inodes
));
1519 assert(hashmap_isempty(d
->wd
));
1521 if (d
->buffer_filled
> 0)
1522 LIST_REMOVE(buffered
, e
->inotify_data_buffered
, d
);
1524 hashmap_free(d
->inodes
);
1525 hashmap_free(d
->wd
);
1527 assert_se(hashmap_remove(e
->inotify_data
, &d
->priority
) == d
);
1530 if (epoll_ctl(e
->epoll_fd
, EPOLL_CTL_DEL
, d
->fd
, NULL
) < 0)
1531 log_debug_errno(errno
, "Failed to remove inotify fd from epoll, ignoring: %m");
1538 static int event_make_inotify_data(
1541 struct inotify_data
**ret
) {
1543 _cleanup_close_
int fd
= -1;
1544 struct inotify_data
*d
;
1549 d
= hashmap_get(e
->inotify_data
, &priority
);
1556 fd
= inotify_init1(IN_NONBLOCK
|O_CLOEXEC
);
1560 fd
= fd_move_above_stdio(fd
);
1562 r
= hashmap_ensure_allocated(&e
->inotify_data
, &uint64_hash_ops
);
1566 d
= new(struct inotify_data
, 1);
1570 *d
= (struct inotify_data
) {
1571 .wakeup
= WAKEUP_INOTIFY_DATA
,
1573 .priority
= priority
,
1576 r
= hashmap_put(e
->inotify_data
, &d
->priority
, d
);
1578 d
->fd
= safe_close(d
->fd
);
1583 struct epoll_event ev
= {
1588 if (epoll_ctl(e
->epoll_fd
, EPOLL_CTL_ADD
, d
->fd
, &ev
) < 0) {
1590 d
->fd
= safe_close(d
->fd
); /* let's close this ourselves, as event_free_inotify_data() would otherwise
1591 * remove the fd from the epoll first, which we don't want as we couldn't
1592 * add it in the first place. */
1593 event_free_inotify_data(e
, d
);
1603 static int inode_data_compare(const struct inode_data
*x
, const struct inode_data
*y
) {
1609 r
= CMP(x
->dev
, y
->dev
);
1613 return CMP(x
->ino
, y
->ino
);
1616 static void inode_data_hash_func(const struct inode_data
*d
, struct siphash
*state
) {
1619 siphash24_compress(&d
->dev
, sizeof(d
->dev
), state
);
1620 siphash24_compress(&d
->ino
, sizeof(d
->ino
), state
);
1623 DEFINE_PRIVATE_HASH_OPS(inode_data_hash_ops
, struct inode_data
, inode_data_hash_func
, inode_data_compare
);
1625 static void event_free_inode_data(
1627 struct inode_data
*d
) {
1634 assert(!d
->event_sources
);
1637 LIST_REMOVE(to_close
, e
->inode_data_to_close
, d
);
1641 if (d
->inotify_data
) {
1644 if (d
->inotify_data
->fd
>= 0) {
1645 /* So here's a problem. At the time this runs the watch descriptor might already be
1646 * invalidated, because an IN_IGNORED event might be queued right the moment we enter
1647 * the syscall. Hence, whenever we get EINVAL, ignore it entirely, since it's a very
1648 * likely case to happen. */
1650 if (inotify_rm_watch(d
->inotify_data
->fd
, d
->wd
) < 0 && errno
!= EINVAL
)
1651 log_debug_errno(errno
, "Failed to remove watch descriptor %i from inotify, ignoring: %m", d
->wd
);
1654 assert_se(hashmap_remove(d
->inotify_data
->wd
, INT_TO_PTR(d
->wd
)) == d
);
1657 assert_se(hashmap_remove(d
->inotify_data
->inodes
, d
) == d
);
1663 static void event_gc_inode_data(
1665 struct inode_data
*d
) {
1667 struct inotify_data
*inotify_data
;
1674 if (d
->event_sources
)
1677 inotify_data
= d
->inotify_data
;
1678 event_free_inode_data(e
, d
);
1680 if (inotify_data
&& hashmap_isempty(inotify_data
->inodes
))
1681 event_free_inotify_data(e
, inotify_data
);
1684 static int event_make_inode_data(
1686 struct inotify_data
*inotify_data
,
1689 struct inode_data
**ret
) {
1691 struct inode_data
*d
, key
;
1695 assert(inotify_data
);
1697 key
= (struct inode_data
) {
1702 d
= hashmap_get(inotify_data
->inodes
, &key
);
1710 r
= hashmap_ensure_allocated(&inotify_data
->inodes
, &inode_data_hash_ops
);
1714 d
= new(struct inode_data
, 1);
1718 *d
= (struct inode_data
) {
1723 .inotify_data
= inotify_data
,
1726 r
= hashmap_put(inotify_data
->inodes
, d
, d
);
1738 static uint32_t inode_data_determine_mask(struct inode_data
*d
) {
1739 bool excl_unlink
= true;
1740 uint32_t combined
= 0;
1745 /* Combines the watch masks of all event sources watching this inode. We generally just OR them together, but
1746 * the IN_EXCL_UNLINK flag is ANDed instead.
1748 * Note that we add all sources to the mask here, regardless whether enabled, disabled or oneshot. That's
1749 * because we cannot change the mask anymore after the event source was created once, since the kernel has no
1750 * API for that. Hence we need to subscribe to the maximum mask we ever might be interested in, and suppress
1751 * events we don't care for client-side. */
1753 LIST_FOREACH(inotify
.by_inode_data
, s
, d
->event_sources
) {
1755 if ((s
->inotify
.mask
& IN_EXCL_UNLINK
) == 0)
1756 excl_unlink
= false;
1758 combined
|= s
->inotify
.mask
;
1761 return (combined
& ~(IN_ONESHOT
|IN_DONT_FOLLOW
|IN_ONLYDIR
|IN_EXCL_UNLINK
)) | (excl_unlink
? IN_EXCL_UNLINK
: 0);
1764 static int inode_data_realize_watch(sd_event
*e
, struct inode_data
*d
) {
1765 uint32_t combined_mask
;
1771 combined_mask
= inode_data_determine_mask(d
);
1773 if (d
->wd
>= 0 && combined_mask
== d
->combined_mask
)
1776 r
= hashmap_ensure_allocated(&d
->inotify_data
->wd
, NULL
);
1780 wd
= inotify_add_watch_fd(d
->inotify_data
->fd
, d
->fd
, combined_mask
);
1785 r
= hashmap_put(d
->inotify_data
->wd
, INT_TO_PTR(wd
), d
);
1787 (void) inotify_rm_watch(d
->inotify_data
->fd
, wd
);
1793 } else if (d
->wd
!= wd
) {
1795 log_debug("Weird, the watch descriptor we already knew for this inode changed?");
1796 (void) inotify_rm_watch(d
->fd
, wd
);
1800 d
->combined_mask
= combined_mask
;
1804 _public_
int sd_event_add_inotify(
1806 sd_event_source
**ret
,
1809 sd_event_inotify_handler_t callback
,
1812 struct inotify_data
*inotify_data
= NULL
;
1813 struct inode_data
*inode_data
= NULL
;
1814 _cleanup_close_
int fd
= -1;
1815 _cleanup_(source_freep
) sd_event_source
*s
= NULL
;
1819 assert_return(e
, -EINVAL
);
1820 assert_return(e
= event_resolve(e
), -ENOPKG
);
1821 assert_return(path
, -EINVAL
);
1822 assert_return(callback
, -EINVAL
);
1823 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
1824 assert_return(!event_pid_changed(e
), -ECHILD
);
1826 /* Refuse IN_MASK_ADD since we coalesce watches on the same inode, and hence really don't want to merge
1827 * masks. Or in other words, this whole code exists only to manage IN_MASK_ADD type operations for you, hence
1828 * the user can't use them for us. */
1829 if (mask
& IN_MASK_ADD
)
1832 fd
= open(path
, O_PATH
|O_CLOEXEC
|
1833 (mask
& IN_ONLYDIR
? O_DIRECTORY
: 0)|
1834 (mask
& IN_DONT_FOLLOW
? O_NOFOLLOW
: 0));
1838 if (fstat(fd
, &st
) < 0)
1841 s
= source_new(e
, !ret
, SOURCE_INOTIFY
);
1845 s
->enabled
= mask
& IN_ONESHOT
? SD_EVENT_ONESHOT
: SD_EVENT_ON
;
1846 s
->inotify
.mask
= mask
;
1847 s
->inotify
.callback
= callback
;
1848 s
->userdata
= userdata
;
1850 /* Allocate an inotify object for this priority, and an inode object within it */
1851 r
= event_make_inotify_data(e
, SD_EVENT_PRIORITY_NORMAL
, &inotify_data
);
1855 r
= event_make_inode_data(e
, inotify_data
, st
.st_dev
, st
.st_ino
, &inode_data
);
1857 event_free_inotify_data(e
, inotify_data
);
1861 /* Keep the O_PATH fd around until the first iteration of the loop, so that we can still change the priority of
1862 * the event source, until then, for which we need the original inode. */
1863 if (inode_data
->fd
< 0) {
1864 inode_data
->fd
= TAKE_FD(fd
);
1865 LIST_PREPEND(to_close
, e
->inode_data_to_close
, inode_data
);
1868 /* Link our event source to the inode data object */
1869 LIST_PREPEND(inotify
.by_inode_data
, inode_data
->event_sources
, s
);
1870 s
->inotify
.inode_data
= inode_data
;
1872 /* Actually realize the watch now */
1873 r
= inode_data_realize_watch(e
, inode_data
);
1877 (void) sd_event_source_set_description(s
, path
);
1886 static sd_event_source
* event_source_free(sd_event_source
*s
) {
1890 /* Here's a special hack: when we are called from a
1891 * dispatch handler we won't free the event source
1892 * immediately, but we will detach the fd from the
1893 * epoll. This way it is safe for the caller to unref
1894 * the event source and immediately close the fd, but
1895 * we still retain a valid event source object after
1898 if (s
->dispatching
) {
1899 if (s
->type
== SOURCE_IO
)
1900 source_io_unregister(s
);
1902 source_disconnect(s
);
1909 DEFINE_PUBLIC_TRIVIAL_REF_UNREF_FUNC(sd_event_source
, sd_event_source
, event_source_free
);
1911 _public_
int sd_event_source_set_description(sd_event_source
*s
, const char *description
) {
1912 assert_return(s
, -EINVAL
);
1913 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
1915 return free_and_strdup(&s
->description
, description
);
1918 _public_
int sd_event_source_get_description(sd_event_source
*s
, const char **description
) {
1919 assert_return(s
, -EINVAL
);
1920 assert_return(description
, -EINVAL
);
1921 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
1923 if (!s
->description
)
1926 *description
= s
->description
;
1930 _public_ sd_event
*sd_event_source_get_event(sd_event_source
*s
) {
1931 assert_return(s
, NULL
);
1936 _public_
int sd_event_source_get_pending(sd_event_source
*s
) {
1937 assert_return(s
, -EINVAL
);
1938 assert_return(s
->type
!= SOURCE_EXIT
, -EDOM
);
1939 assert_return(s
->event
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
1940 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
1945 _public_
int sd_event_source_get_io_fd(sd_event_source
*s
) {
1946 assert_return(s
, -EINVAL
);
1947 assert_return(s
->type
== SOURCE_IO
, -EDOM
);
1948 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
1953 _public_
int sd_event_source_set_io_fd(sd_event_source
*s
, int fd
) {
1956 assert_return(s
, -EINVAL
);
1957 assert_return(fd
>= 0, -EBADF
);
1958 assert_return(s
->type
== SOURCE_IO
, -EDOM
);
1959 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
1964 if (s
->enabled
== SD_EVENT_OFF
) {
1966 s
->io
.registered
= false;
1970 saved_fd
= s
->io
.fd
;
1971 assert(s
->io
.registered
);
1974 s
->io
.registered
= false;
1976 r
= source_io_register(s
, s
->enabled
, s
->io
.events
);
1978 s
->io
.fd
= saved_fd
;
1979 s
->io
.registered
= true;
1983 (void) epoll_ctl(s
->event
->epoll_fd
, EPOLL_CTL_DEL
, saved_fd
, NULL
);
1989 _public_
int sd_event_source_get_io_fd_own(sd_event_source
*s
) {
1990 assert_return(s
, -EINVAL
);
1991 assert_return(s
->type
== SOURCE_IO
, -EDOM
);
1996 _public_
int sd_event_source_set_io_fd_own(sd_event_source
*s
, int own
) {
1997 assert_return(s
, -EINVAL
);
1998 assert_return(s
->type
== SOURCE_IO
, -EDOM
);
2004 _public_
int sd_event_source_get_io_events(sd_event_source
*s
, uint32_t* events
) {
2005 assert_return(s
, -EINVAL
);
2006 assert_return(events
, -EINVAL
);
2007 assert_return(s
->type
== SOURCE_IO
, -EDOM
);
2008 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2010 *events
= s
->io
.events
;
2014 _public_
int sd_event_source_set_io_events(sd_event_source
*s
, uint32_t events
) {
2017 assert_return(s
, -EINVAL
);
2018 assert_return(s
->type
== SOURCE_IO
, -EDOM
);
2019 assert_return(!(events
& ~(EPOLLIN
|EPOLLOUT
|EPOLLRDHUP
|EPOLLPRI
|EPOLLERR
|EPOLLHUP
|EPOLLET
)), -EINVAL
);
2020 assert_return(s
->event
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
2021 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2023 /* edge-triggered updates are never skipped, so we can reset edges */
2024 if (s
->io
.events
== events
&& !(events
& EPOLLET
))
2027 r
= source_set_pending(s
, false);
2031 if (s
->enabled
!= SD_EVENT_OFF
) {
2032 r
= source_io_register(s
, s
->enabled
, events
);
2037 s
->io
.events
= events
;
2042 _public_
int sd_event_source_get_io_revents(sd_event_source
*s
, uint32_t* revents
) {
2043 assert_return(s
, -EINVAL
);
2044 assert_return(revents
, -EINVAL
);
2045 assert_return(s
->type
== SOURCE_IO
, -EDOM
);
2046 assert_return(s
->pending
, -ENODATA
);
2047 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2049 *revents
= s
->io
.revents
;
2053 _public_
int sd_event_source_get_signal(sd_event_source
*s
) {
2054 assert_return(s
, -EINVAL
);
2055 assert_return(s
->type
== SOURCE_SIGNAL
, -EDOM
);
2056 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2058 return s
->signal
.sig
;
2061 _public_
int sd_event_source_get_priority(sd_event_source
*s
, int64_t *priority
) {
2062 assert_return(s
, -EINVAL
);
2063 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2065 *priority
= s
->priority
;
2069 _public_
int sd_event_source_set_priority(sd_event_source
*s
, int64_t priority
) {
2070 bool rm_inotify
= false, rm_inode
= false;
2071 struct inotify_data
*new_inotify_data
= NULL
;
2072 struct inode_data
*new_inode_data
= NULL
;
2075 assert_return(s
, -EINVAL
);
2076 assert_return(s
->event
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
2077 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2079 if (s
->priority
== priority
)
2082 if (s
->type
== SOURCE_INOTIFY
) {
2083 struct inode_data
*old_inode_data
;
2085 assert(s
->inotify
.inode_data
);
2086 old_inode_data
= s
->inotify
.inode_data
;
2088 /* We need the original fd to change the priority. If we don't have it we can't change the priority,
2089 * anymore. Note that we close any fds when entering the next event loop iteration, i.e. for inotify
2090 * events we allow priority changes only until the first following iteration. */
2091 if (old_inode_data
->fd
< 0)
2094 r
= event_make_inotify_data(s
->event
, priority
, &new_inotify_data
);
2099 r
= event_make_inode_data(s
->event
, new_inotify_data
, old_inode_data
->dev
, old_inode_data
->ino
, &new_inode_data
);
2104 if (new_inode_data
->fd
< 0) {
2105 /* Duplicate the fd for the new inode object if we don't have any yet */
2106 new_inode_data
->fd
= fcntl(old_inode_data
->fd
, F_DUPFD_CLOEXEC
, 3);
2107 if (new_inode_data
->fd
< 0) {
2112 LIST_PREPEND(to_close
, s
->event
->inode_data_to_close
, new_inode_data
);
2115 /* Move the event source to the new inode data structure */
2116 LIST_REMOVE(inotify
.by_inode_data
, old_inode_data
->event_sources
, s
);
2117 LIST_PREPEND(inotify
.by_inode_data
, new_inode_data
->event_sources
, s
);
2118 s
->inotify
.inode_data
= new_inode_data
;
2120 /* Now create the new watch */
2121 r
= inode_data_realize_watch(s
->event
, new_inode_data
);
2124 LIST_REMOVE(inotify
.by_inode_data
, new_inode_data
->event_sources
, s
);
2125 LIST_PREPEND(inotify
.by_inode_data
, old_inode_data
->event_sources
, s
);
2126 s
->inotify
.inode_data
= old_inode_data
;
2130 s
->priority
= priority
;
2132 event_gc_inode_data(s
->event
, old_inode_data
);
2134 } else if (s
->type
== SOURCE_SIGNAL
&& s
->enabled
!= SD_EVENT_OFF
) {
2135 struct signal_data
*old
, *d
;
2137 /* Move us from the signalfd belonging to the old
2138 * priority to the signalfd of the new priority */
2140 assert_se(old
= hashmap_get(s
->event
->signal_data
, &s
->priority
));
2142 s
->priority
= priority
;
2144 r
= event_make_signal_data(s
->event
, s
->signal
.sig
, &d
);
2146 s
->priority
= old
->priority
;
2150 event_unmask_signal_data(s
->event
, old
, s
->signal
.sig
);
2152 s
->priority
= priority
;
2155 prioq_reshuffle(s
->event
->pending
, s
, &s
->pending_index
);
2158 prioq_reshuffle(s
->event
->prepare
, s
, &s
->prepare_index
);
2160 if (s
->type
== SOURCE_EXIT
)
2161 prioq_reshuffle(s
->event
->exit
, s
, &s
->exit
.prioq_index
);
2167 event_free_inode_data(s
->event
, new_inode_data
);
2170 event_free_inotify_data(s
->event
, new_inotify_data
);
2175 _public_
int sd_event_source_get_enabled(sd_event_source
*s
, int *m
) {
2176 assert_return(s
, -EINVAL
);
2177 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2181 return s
->enabled
!= SD_EVENT_OFF
;
2184 _public_
int sd_event_source_set_enabled(sd_event_source
*s
, int m
) {
2187 assert_return(s
, -EINVAL
);
2188 assert_return(IN_SET(m
, SD_EVENT_OFF
, SD_EVENT_ON
, SD_EVENT_ONESHOT
), -EINVAL
);
2189 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2191 /* If we are dead anyway, we are fine with turning off
2192 * sources, but everything else needs to fail. */
2193 if (s
->event
->state
== SD_EVENT_FINISHED
)
2194 return m
== SD_EVENT_OFF
? 0 : -ESTALE
;
2196 if (s
->enabled
== m
)
2199 if (m
== SD_EVENT_OFF
) {
2201 /* Unset the pending flag when this event source is disabled */
2202 if (!IN_SET(s
->type
, SOURCE_DEFER
, SOURCE_EXIT
)) {
2203 r
= source_set_pending(s
, false);
2211 source_io_unregister(s
);
2215 case SOURCE_TIME_REALTIME
:
2216 case SOURCE_TIME_BOOTTIME
:
2217 case SOURCE_TIME_MONOTONIC
:
2218 case SOURCE_TIME_REALTIME_ALARM
:
2219 case SOURCE_TIME_BOOTTIME_ALARM
: {
2220 struct clock_data
*d
;
2223 d
= event_get_clock_data(s
->event
, s
->type
);
2226 prioq_reshuffle(d
->earliest
, s
, &s
->time
.earliest_index
);
2227 prioq_reshuffle(d
->latest
, s
, &s
->time
.latest_index
);
2228 d
->needs_rearm
= true;
2235 event_gc_signal_data(s
->event
, &s
->priority
, s
->signal
.sig
);
2241 assert(s
->event
->n_enabled_child_sources
> 0);
2242 s
->event
->n_enabled_child_sources
--;
2244 if (EVENT_SOURCE_WATCH_PIDFD(s
))
2245 source_child_pidfd_unregister(s
);
2247 event_gc_signal_data(s
->event
, &s
->priority
, SIGCHLD
);
2253 prioq_reshuffle(s
->event
->exit
, s
, &s
->exit
.prioq_index
);
2258 case SOURCE_INOTIFY
:
2263 assert_not_reached("Wut? I shouldn't exist.");
2268 /* Unset the pending flag when this event source is enabled */
2269 if (s
->enabled
== SD_EVENT_OFF
&& !IN_SET(s
->type
, SOURCE_DEFER
, SOURCE_EXIT
)) {
2270 r
= source_set_pending(s
, false);
2278 r
= source_io_register(s
, m
, s
->io
.events
);
2285 case SOURCE_TIME_REALTIME
:
2286 case SOURCE_TIME_BOOTTIME
:
2287 case SOURCE_TIME_MONOTONIC
:
2288 case SOURCE_TIME_REALTIME_ALARM
:
2289 case SOURCE_TIME_BOOTTIME_ALARM
: {
2290 struct clock_data
*d
;
2293 d
= event_get_clock_data(s
->event
, s
->type
);
2296 prioq_reshuffle(d
->earliest
, s
, &s
->time
.earliest_index
);
2297 prioq_reshuffle(d
->latest
, s
, &s
->time
.latest_index
);
2298 d
->needs_rearm
= true;
2306 r
= event_make_signal_data(s
->event
, s
->signal
.sig
, NULL
);
2308 s
->enabled
= SD_EVENT_OFF
;
2309 event_gc_signal_data(s
->event
, &s
->priority
, s
->signal
.sig
);
2317 if (s
->enabled
== SD_EVENT_OFF
)
2318 s
->event
->n_enabled_child_sources
++;
2322 if (EVENT_SOURCE_WATCH_PIDFD(s
)) {
2323 /* yes, we have pidfd */
2325 r
= source_child_pidfd_register(s
, s
->enabled
);
2327 s
->enabled
= SD_EVENT_OFF
;
2328 s
->event
->n_enabled_child_sources
--;
2332 /* no pidfd, or something other to watch for than WEXITED */
2334 r
= event_make_signal_data(s
->event
, SIGCHLD
, NULL
);
2336 s
->enabled
= SD_EVENT_OFF
;
2337 s
->event
->n_enabled_child_sources
--;
2338 event_gc_signal_data(s
->event
, &s
->priority
, SIGCHLD
);
2347 prioq_reshuffle(s
->event
->exit
, s
, &s
->exit
.prioq_index
);
2352 case SOURCE_INOTIFY
:
2357 assert_not_reached("Wut? I shouldn't exist.");
2362 prioq_reshuffle(s
->event
->pending
, s
, &s
->pending_index
);
2365 prioq_reshuffle(s
->event
->prepare
, s
, &s
->prepare_index
);
2370 _public_
int sd_event_source_get_time(sd_event_source
*s
, uint64_t *usec
) {
2371 assert_return(s
, -EINVAL
);
2372 assert_return(usec
, -EINVAL
);
2373 assert_return(EVENT_SOURCE_IS_TIME(s
->type
), -EDOM
);
2374 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2376 *usec
= s
->time
.next
;
2380 _public_
int sd_event_source_set_time(sd_event_source
*s
, uint64_t usec
) {
2381 struct clock_data
*d
;
2384 assert_return(s
, -EINVAL
);
2385 assert_return(EVENT_SOURCE_IS_TIME(s
->type
), -EDOM
);
2386 assert_return(s
->event
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
2387 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2389 r
= source_set_pending(s
, false);
2393 s
->time
.next
= usec
;
2395 d
= event_get_clock_data(s
->event
, s
->type
);
2398 prioq_reshuffle(d
->earliest
, s
, &s
->time
.earliest_index
);
2399 prioq_reshuffle(d
->latest
, s
, &s
->time
.latest_index
);
2400 d
->needs_rearm
= true;
2405 _public_
int sd_event_source_get_time_accuracy(sd_event_source
*s
, uint64_t *usec
) {
2406 assert_return(s
, -EINVAL
);
2407 assert_return(usec
, -EINVAL
);
2408 assert_return(EVENT_SOURCE_IS_TIME(s
->type
), -EDOM
);
2409 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2411 *usec
= s
->time
.accuracy
;
2415 _public_
int sd_event_source_set_time_accuracy(sd_event_source
*s
, uint64_t usec
) {
2416 struct clock_data
*d
;
2419 assert_return(s
, -EINVAL
);
2420 assert_return(usec
!= (uint64_t) -1, -EINVAL
);
2421 assert_return(EVENT_SOURCE_IS_TIME(s
->type
), -EDOM
);
2422 assert_return(s
->event
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
2423 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2425 r
= source_set_pending(s
, false);
2430 usec
= DEFAULT_ACCURACY_USEC
;
2432 s
->time
.accuracy
= usec
;
2434 d
= event_get_clock_data(s
->event
, s
->type
);
2437 prioq_reshuffle(d
->latest
, s
, &s
->time
.latest_index
);
2438 d
->needs_rearm
= true;
2443 _public_
int sd_event_source_get_time_clock(sd_event_source
*s
, clockid_t
*clock
) {
2444 assert_return(s
, -EINVAL
);
2445 assert_return(clock
, -EINVAL
);
2446 assert_return(EVENT_SOURCE_IS_TIME(s
->type
), -EDOM
);
2447 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2449 *clock
= event_source_type_to_clock(s
->type
);
2453 _public_
int sd_event_source_get_child_pid(sd_event_source
*s
, pid_t
*pid
) {
2454 assert_return(s
, -EINVAL
);
2455 assert_return(pid
, -EINVAL
);
2456 assert_return(s
->type
== SOURCE_CHILD
, -EDOM
);
2457 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2459 *pid
= s
->child
.pid
;
2463 _public_
int sd_event_source_get_child_pidfd(sd_event_source
*s
) {
2464 assert_return(s
, -EINVAL
);
2465 assert_return(s
->type
== SOURCE_CHILD
, -EDOM
);
2466 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2468 if (s
->child
.pidfd
< 0)
2471 return s
->child
.pidfd
;
2474 _public_
int sd_event_source_send_child_signal(sd_event_source
*s
, int sig
, const siginfo_t
*si
, unsigned flags
) {
2475 assert_return(s
, -EINVAL
);
2476 assert_return(s
->type
== SOURCE_CHILD
, -EDOM
);
2477 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2478 assert_return(SIGNAL_VALID(sig
), -EINVAL
);
2480 /* If we already have seen indication the process exited refuse sending a signal early. This way we
2481 * can be sure we don't accidentally kill the wrong process on PID reuse when pidfds are not
2483 if (s
->child
.exited
)
2486 if (s
->child
.pidfd
>= 0) {
2489 /* pidfd_send_signal() changes the siginfo_t argument. This is weird, let's hence copy the
2494 if (pidfd_send_signal(s
->child
.pidfd
, sig
, si
? ©
: NULL
, 0) < 0) {
2495 /* Let's propagate the error only if the system call is not implemented or prohibited */
2496 if (!ERRNO_IS_NOT_SUPPORTED(errno
) && !ERRNO_IS_PRIVILEGE(errno
))
2502 /* Flags are only supported for pidfd_send_signal(), not for rt_sigqueueinfo(), hence let's refuse
2508 /* We use rt_sigqueueinfo() only if siginfo_t is specified. */
2509 siginfo_t copy
= *si
;
2511 if (rt_sigqueueinfo(s
->child
.pid
, sig
, ©
) < 0)
2513 } else if (kill(s
->child
.pid
, sig
) < 0)
2519 _public_
int sd_event_source_get_child_pidfd_own(sd_event_source
*s
) {
2520 assert_return(s
, -EINVAL
);
2521 assert_return(s
->type
== SOURCE_CHILD
, -EDOM
);
2523 if (s
->child
.pidfd
< 0)
2526 return s
->child
.pidfd_owned
;
2529 _public_
int sd_event_source_set_child_pidfd_own(sd_event_source
*s
, int own
) {
2530 assert_return(s
, -EINVAL
);
2531 assert_return(s
->type
== SOURCE_CHILD
, -EDOM
);
2533 if (s
->child
.pidfd
< 0)
2536 s
->child
.pidfd_owned
= own
;
2540 _public_
int sd_event_source_get_child_process_own(sd_event_source
*s
) {
2541 assert_return(s
, -EINVAL
);
2542 assert_return(s
->type
== SOURCE_CHILD
, -EDOM
);
2544 return s
->child
.process_owned
;
2547 _public_
int sd_event_source_set_child_process_own(sd_event_source
*s
, int own
) {
2548 assert_return(s
, -EINVAL
);
2549 assert_return(s
->type
== SOURCE_CHILD
, -EDOM
);
2551 s
->child
.process_owned
= own
;
2555 _public_
int sd_event_source_get_inotify_mask(sd_event_source
*s
, uint32_t *mask
) {
2556 assert_return(s
, -EINVAL
);
2557 assert_return(mask
, -EINVAL
);
2558 assert_return(s
->type
== SOURCE_INOTIFY
, -EDOM
);
2559 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2561 *mask
= s
->inotify
.mask
;
2565 _public_
int sd_event_source_set_prepare(sd_event_source
*s
, sd_event_handler_t callback
) {
2568 assert_return(s
, -EINVAL
);
2569 assert_return(s
->type
!= SOURCE_EXIT
, -EDOM
);
2570 assert_return(s
->event
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
2571 assert_return(!event_pid_changed(s
->event
), -ECHILD
);
2573 if (s
->prepare
== callback
)
2576 if (callback
&& s
->prepare
) {
2577 s
->prepare
= callback
;
2581 r
= prioq_ensure_allocated(&s
->event
->prepare
, prepare_prioq_compare
);
2585 s
->prepare
= callback
;
2588 r
= prioq_put(s
->event
->prepare
, s
, &s
->prepare_index
);
2592 prioq_remove(s
->event
->prepare
, s
, &s
->prepare_index
);
2597 _public_
void* sd_event_source_get_userdata(sd_event_source
*s
) {
2598 assert_return(s
, NULL
);
2603 _public_
void *sd_event_source_set_userdata(sd_event_source
*s
, void *userdata
) {
2606 assert_return(s
, NULL
);
2609 s
->userdata
= userdata
;
2614 static usec_t
sleep_between(sd_event
*e
, usec_t a
, usec_t b
) {
2621 if (a
>= USEC_INFINITY
)
2622 return USEC_INFINITY
;
2627 initialize_perturb(e
);
2630 Find a good time to wake up again between times a and b. We
2631 have two goals here:
2633 a) We want to wake up as seldom as possible, hence prefer
2634 later times over earlier times.
2636 b) But if we have to wake up, then let's make sure to
2637 dispatch as much as possible on the entire system.
2639 We implement this by waking up everywhere at the same time
2640 within any given minute if we can, synchronised via the
2641 perturbation value determined from the boot ID. If we can't,
2642 then we try to find the same spot in every 10s, then 1s and
2643 then 250ms step. Otherwise, we pick the last possible time
2647 c
= (b
/ USEC_PER_MINUTE
) * USEC_PER_MINUTE
+ e
->perturb
;
2649 if (_unlikely_(c
< USEC_PER_MINUTE
))
2652 c
-= USEC_PER_MINUTE
;
2658 c
= (b
/ (USEC_PER_SEC
*10)) * (USEC_PER_SEC
*10) + (e
->perturb
% (USEC_PER_SEC
*10));
2660 if (_unlikely_(c
< USEC_PER_SEC
*10))
2663 c
-= USEC_PER_SEC
*10;
2669 c
= (b
/ USEC_PER_SEC
) * USEC_PER_SEC
+ (e
->perturb
% USEC_PER_SEC
);
2671 if (_unlikely_(c
< USEC_PER_SEC
))
2680 c
= (b
/ (USEC_PER_MSEC
*250)) * (USEC_PER_MSEC
*250) + (e
->perturb
% (USEC_PER_MSEC
*250));
2682 if (_unlikely_(c
< USEC_PER_MSEC
*250))
2685 c
-= USEC_PER_MSEC
*250;
2694 static int event_arm_timer(
2696 struct clock_data
*d
) {
2698 struct itimerspec its
= {};
2699 sd_event_source
*a
, *b
;
2706 if (!d
->needs_rearm
)
2709 d
->needs_rearm
= false;
2711 a
= prioq_peek(d
->earliest
);
2712 if (!a
|| a
->enabled
== SD_EVENT_OFF
|| a
->time
.next
== USEC_INFINITY
) {
2717 if (d
->next
== USEC_INFINITY
)
2721 r
= timerfd_settime(d
->fd
, TFD_TIMER_ABSTIME
, &its
, NULL
);
2725 d
->next
= USEC_INFINITY
;
2729 b
= prioq_peek(d
->latest
);
2730 assert_se(b
&& b
->enabled
!= SD_EVENT_OFF
);
2732 t
= sleep_between(e
, a
->time
.next
, time_event_source_latest(b
));
2736 assert_se(d
->fd
>= 0);
2739 /* We don' want to disarm here, just mean some time looooong ago. */
2740 its
.it_value
.tv_sec
= 0;
2741 its
.it_value
.tv_nsec
= 1;
2743 timespec_store(&its
.it_value
, t
);
2745 r
= timerfd_settime(d
->fd
, TFD_TIMER_ABSTIME
, &its
, NULL
);
2753 static int process_io(sd_event
*e
, sd_event_source
*s
, uint32_t revents
) {
2756 assert(s
->type
== SOURCE_IO
);
2758 /* If the event source was already pending, we just OR in the
2759 * new revents, otherwise we reset the value. The ORing is
2760 * necessary to handle EPOLLONESHOT events properly where
2761 * readability might happen independently of writability, and
2762 * we need to keep track of both */
2765 s
->io
.revents
|= revents
;
2767 s
->io
.revents
= revents
;
2769 return source_set_pending(s
, true);
2772 static int flush_timer(sd_event
*e
, int fd
, uint32_t events
, usec_t
*next
) {
2779 assert_return(events
== EPOLLIN
, -EIO
);
2781 ss
= read(fd
, &x
, sizeof(x
));
2783 if (IN_SET(errno
, EAGAIN
, EINTR
))
2789 if (_unlikely_(ss
!= sizeof(x
)))
2793 *next
= USEC_INFINITY
;
2798 static int process_timer(
2801 struct clock_data
*d
) {
2810 s
= prioq_peek(d
->earliest
);
2813 s
->enabled
== SD_EVENT_OFF
||
2817 r
= source_set_pending(s
, true);
2821 prioq_reshuffle(d
->earliest
, s
, &s
->time
.earliest_index
);
2822 prioq_reshuffle(d
->latest
, s
, &s
->time
.latest_index
);
2823 d
->needs_rearm
= true;
2829 static int process_child(sd_event
*e
) {
2836 e
->need_process_child
= false;
2839 So, this is ugly. We iteratively invoke waitid() with P_PID
2840 + WNOHANG for each PID we wait for, instead of using
2841 P_ALL. This is because we only want to get child
2842 information of very specific child processes, and not all
2843 of them. We might not have processed the SIGCHLD even of a
2844 previous invocation and we don't want to maintain a
2845 unbounded *per-child* event queue, hence we really don't
2846 want anything flushed out of the kernel's queue that we
2847 don't care about. Since this is O(n) this means that if you
2848 have a lot of processes you probably want to handle SIGCHLD
2851 We do not reap the children here (by using WNOWAIT), this
2852 is only done after the event source is dispatched so that
2853 the callback still sees the process as a zombie.
2856 HASHMAP_FOREACH(s
, e
->child_sources
, i
) {
2857 assert(s
->type
== SOURCE_CHILD
);
2862 if (s
->enabled
== SD_EVENT_OFF
)
2865 if (s
->child
.exited
)
2868 if (EVENT_SOURCE_WATCH_PIDFD(s
)) /* There's a usable pidfd known for this event source? then don't waitid() for it here */
2871 zero(s
->child
.siginfo
);
2872 r
= waitid(P_PID
, s
->child
.pid
, &s
->child
.siginfo
,
2873 WNOHANG
| (s
->child
.options
& WEXITED
? WNOWAIT
: 0) | s
->child
.options
);
2877 if (s
->child
.siginfo
.si_pid
!= 0) {
2878 bool zombie
= IN_SET(s
->child
.siginfo
.si_code
, CLD_EXITED
, CLD_KILLED
, CLD_DUMPED
);
2881 s
->child
.exited
= true;
2883 if (!zombie
&& (s
->child
.options
& WEXITED
)) {
2884 /* If the child isn't dead then let's
2885 * immediately remove the state change
2886 * from the queue, since there's no
2887 * benefit in leaving it queued */
2889 assert(s
->child
.options
& (WSTOPPED
|WCONTINUED
));
2890 (void) waitid(P_PID
, s
->child
.pid
, &s
->child
.siginfo
, WNOHANG
|(s
->child
.options
& (WSTOPPED
|WCONTINUED
)));
2893 r
= source_set_pending(s
, true);
2902 static int process_pidfd(sd_event
*e
, sd_event_source
*s
, uint32_t revents
) {
2905 assert(s
->type
== SOURCE_CHILD
);
2910 if (s
->enabled
== SD_EVENT_OFF
)
2913 if (!EVENT_SOURCE_WATCH_PIDFD(s
))
2916 zero(s
->child
.siginfo
);
2917 if (waitid(P_PID
, s
->child
.pid
, &s
->child
.siginfo
, WNOHANG
| WNOWAIT
| s
->child
.options
) < 0)
2920 if (s
->child
.siginfo
.si_pid
== 0)
2923 if (IN_SET(s
->child
.siginfo
.si_code
, CLD_EXITED
, CLD_KILLED
, CLD_DUMPED
))
2924 s
->child
.exited
= true;
2926 return source_set_pending(s
, true);
2929 static int process_signal(sd_event
*e
, struct signal_data
*d
, uint32_t events
) {
2930 bool read_one
= false;
2935 assert_return(events
== EPOLLIN
, -EIO
);
2937 /* If there's a signal queued on this priority and SIGCHLD is
2938 on this priority too, then make sure to recheck the
2939 children we watch. This is because we only ever dequeue
2940 the first signal per priority, and if we dequeue one, and
2941 SIGCHLD might be enqueued later we wouldn't know, but we
2942 might have higher priority children we care about hence we
2943 need to check that explicitly. */
2945 if (sigismember(&d
->sigset
, SIGCHLD
))
2946 e
->need_process_child
= true;
2948 /* If there's already an event source pending for this
2949 * priority we don't read another */
2954 struct signalfd_siginfo si
;
2956 sd_event_source
*s
= NULL
;
2958 n
= read(d
->fd
, &si
, sizeof(si
));
2960 if (IN_SET(errno
, EAGAIN
, EINTR
))
2966 if (_unlikely_(n
!= sizeof(si
)))
2969 assert(SIGNAL_VALID(si
.ssi_signo
));
2973 if (e
->signal_sources
)
2974 s
= e
->signal_sources
[si
.ssi_signo
];
2980 s
->signal
.siginfo
= si
;
2983 r
= source_set_pending(s
, true);
2991 static int event_inotify_data_read(sd_event
*e
, struct inotify_data
*d
, uint32_t revents
) {
2997 assert_return(revents
== EPOLLIN
, -EIO
);
2999 /* If there's already an event source pending for this priority, don't read another */
3000 if (d
->n_pending
> 0)
3003 /* Is the read buffer non-empty? If so, let's not read more */
3004 if (d
->buffer_filled
> 0)
3007 n
= read(d
->fd
, &d
->buffer
, sizeof(d
->buffer
));
3009 if (IN_SET(errno
, EAGAIN
, EINTR
))
3016 d
->buffer_filled
= (size_t) n
;
3017 LIST_PREPEND(buffered
, e
->inotify_data_buffered
, d
);
3022 static void event_inotify_data_drop(sd_event
*e
, struct inotify_data
*d
, size_t sz
) {
3025 assert(sz
<= d
->buffer_filled
);
3030 /* Move the rest to the buffer to the front, in order to get things properly aligned again */
3031 memmove(d
->buffer
.raw
, d
->buffer
.raw
+ sz
, d
->buffer_filled
- sz
);
3032 d
->buffer_filled
-= sz
;
3034 if (d
->buffer_filled
== 0)
3035 LIST_REMOVE(buffered
, e
->inotify_data_buffered
, d
);
3038 static int event_inotify_data_process(sd_event
*e
, struct inotify_data
*d
) {
3044 /* If there's already an event source pending for this priority, don't read another */
3045 if (d
->n_pending
> 0)
3048 while (d
->buffer_filled
> 0) {
3051 /* Let's validate that the event structures are complete */
3052 if (d
->buffer_filled
< offsetof(struct inotify_event
, name
))
3055 sz
= offsetof(struct inotify_event
, name
) + d
->buffer
.ev
.len
;
3056 if (d
->buffer_filled
< sz
)
3059 if (d
->buffer
.ev
.mask
& IN_Q_OVERFLOW
) {
3060 struct inode_data
*inode_data
;
3063 /* The queue overran, let's pass this event to all event sources connected to this inotify
3066 HASHMAP_FOREACH(inode_data
, d
->inodes
, i
) {
3069 LIST_FOREACH(inotify
.by_inode_data
, s
, inode_data
->event_sources
) {
3071 if (s
->enabled
== SD_EVENT_OFF
)
3074 r
= source_set_pending(s
, true);
3080 struct inode_data
*inode_data
;
3083 /* Find the inode object for this watch descriptor. If IN_IGNORED is set we also remove it from
3084 * our watch descriptor table. */
3085 if (d
->buffer
.ev
.mask
& IN_IGNORED
) {
3087 inode_data
= hashmap_remove(d
->wd
, INT_TO_PTR(d
->buffer
.ev
.wd
));
3089 event_inotify_data_drop(e
, d
, sz
);
3093 /* The watch descriptor was removed by the kernel, let's drop it here too */
3094 inode_data
->wd
= -1;
3096 inode_data
= hashmap_get(d
->wd
, INT_TO_PTR(d
->buffer
.ev
.wd
));
3098 event_inotify_data_drop(e
, d
, sz
);
3103 /* Trigger all event sources that are interested in these events. Also trigger all event
3104 * sources if IN_IGNORED or IN_UNMOUNT is set. */
3105 LIST_FOREACH(inotify
.by_inode_data
, s
, inode_data
->event_sources
) {
3107 if (s
->enabled
== SD_EVENT_OFF
)
3110 if ((d
->buffer
.ev
.mask
& (IN_IGNORED
|IN_UNMOUNT
)) == 0 &&
3111 (s
->inotify
.mask
& d
->buffer
.ev
.mask
& IN_ALL_EVENTS
) == 0)
3114 r
= source_set_pending(s
, true);
3120 /* Something pending now? If so, let's finish, otherwise let's read more. */
3121 if (d
->n_pending
> 0)
3128 static int process_inotify(sd_event
*e
) {
3129 struct inotify_data
*d
;
3134 LIST_FOREACH(buffered
, d
, e
->inotify_data_buffered
) {
3135 r
= event_inotify_data_process(e
, d
);
3145 static int source_dispatch(sd_event_source
*s
) {
3146 EventSourceType saved_type
;
3150 assert(s
->pending
|| s
->type
== SOURCE_EXIT
);
3152 /* Save the event source type, here, so that we still know it after the event callback which might invalidate
3154 saved_type
= s
->type
;
3156 if (!IN_SET(s
->type
, SOURCE_DEFER
, SOURCE_EXIT
)) {
3157 r
= source_set_pending(s
, false);
3162 if (s
->type
!= SOURCE_POST
) {
3166 /* If we execute a non-post source, let's mark all
3167 * post sources as pending */
3169 SET_FOREACH(z
, s
->event
->post_sources
, i
) {
3170 if (z
->enabled
== SD_EVENT_OFF
)
3173 r
= source_set_pending(z
, true);
3179 if (s
->enabled
== SD_EVENT_ONESHOT
) {
3180 r
= sd_event_source_set_enabled(s
, SD_EVENT_OFF
);
3185 s
->dispatching
= true;
3190 r
= s
->io
.callback(s
, s
->io
.fd
, s
->io
.revents
, s
->userdata
);
3193 case SOURCE_TIME_REALTIME
:
3194 case SOURCE_TIME_BOOTTIME
:
3195 case SOURCE_TIME_MONOTONIC
:
3196 case SOURCE_TIME_REALTIME_ALARM
:
3197 case SOURCE_TIME_BOOTTIME_ALARM
:
3198 r
= s
->time
.callback(s
, s
->time
.next
, s
->userdata
);
3202 r
= s
->signal
.callback(s
, &s
->signal
.siginfo
, s
->userdata
);
3205 case SOURCE_CHILD
: {
3208 zombie
= IN_SET(s
->child
.siginfo
.si_code
, CLD_EXITED
, CLD_KILLED
, CLD_DUMPED
);
3210 r
= s
->child
.callback(s
, &s
->child
.siginfo
, s
->userdata
);
3212 /* Now, reap the PID for good. */
3214 (void) waitid(P_PID
, s
->child
.pid
, &s
->child
.siginfo
, WNOHANG
|WEXITED
);
3215 s
->child
.waited
= true;
3222 r
= s
->defer
.callback(s
, s
->userdata
);
3226 r
= s
->post
.callback(s
, s
->userdata
);
3230 r
= s
->exit
.callback(s
, s
->userdata
);
3233 case SOURCE_INOTIFY
: {
3234 struct sd_event
*e
= s
->event
;
3235 struct inotify_data
*d
;
3238 assert(s
->inotify
.inode_data
);
3239 assert_se(d
= s
->inotify
.inode_data
->inotify_data
);
3241 assert(d
->buffer_filled
>= offsetof(struct inotify_event
, name
));
3242 sz
= offsetof(struct inotify_event
, name
) + d
->buffer
.ev
.len
;
3243 assert(d
->buffer_filled
>= sz
);
3245 r
= s
->inotify
.callback(s
, &d
->buffer
.ev
, s
->userdata
);
3247 /* When no event is pending anymore on this inotify object, then let's drop the event from the
3249 if (d
->n_pending
== 0)
3250 event_inotify_data_drop(e
, d
, sz
);
3255 case SOURCE_WATCHDOG
:
3256 case _SOURCE_EVENT_SOURCE_TYPE_MAX
:
3257 case _SOURCE_EVENT_SOURCE_TYPE_INVALID
:
3258 assert_not_reached("Wut? I shouldn't exist.");
3261 s
->dispatching
= false;
3264 log_debug_errno(r
, "Event source %s (type %s) returned error, disabling: %m",
3265 strna(s
->description
), event_source_type_to_string(saved_type
));
3270 sd_event_source_set_enabled(s
, SD_EVENT_OFF
);
3275 static int event_prepare(sd_event
*e
) {
3283 s
= prioq_peek(e
->prepare
);
3284 if (!s
|| s
->prepare_iteration
== e
->iteration
|| s
->enabled
== SD_EVENT_OFF
)
3287 s
->prepare_iteration
= e
->iteration
;
3288 r
= prioq_reshuffle(e
->prepare
, s
, &s
->prepare_index
);
3294 s
->dispatching
= true;
3295 r
= s
->prepare(s
, s
->userdata
);
3296 s
->dispatching
= false;
3299 log_debug_errno(r
, "Prepare callback of event source %s (type %s) returned error, disabling: %m",
3300 strna(s
->description
), event_source_type_to_string(s
->type
));
3305 sd_event_source_set_enabled(s
, SD_EVENT_OFF
);
3311 static int dispatch_exit(sd_event
*e
) {
3313 _cleanup_(sd_event_unrefp
) sd_event
*ref
= NULL
;
3318 p
= prioq_peek(e
->exit
);
3319 if (!p
|| p
->enabled
== SD_EVENT_OFF
) {
3320 e
->state
= SD_EVENT_FINISHED
;
3324 ref
= sd_event_ref(e
);
3326 e
->state
= SD_EVENT_EXITING
;
3327 r
= source_dispatch(p
);
3328 e
->state
= SD_EVENT_INITIAL
;
3332 static sd_event_source
* event_next_pending(sd_event
*e
) {
3337 p
= prioq_peek(e
->pending
);
3341 if (p
->enabled
== SD_EVENT_OFF
)
3347 static int arm_watchdog(sd_event
*e
) {
3348 struct itimerspec its
= {};
3353 assert(e
->watchdog_fd
>= 0);
3355 t
= sleep_between(e
,
3356 e
->watchdog_last
+ (e
->watchdog_period
/ 2),
3357 e
->watchdog_last
+ (e
->watchdog_period
* 3 / 4));
3359 timespec_store(&its
.it_value
, t
);
3361 /* Make sure we never set the watchdog to 0, which tells the
3362 * kernel to disable it. */
3363 if (its
.it_value
.tv_sec
== 0 && its
.it_value
.tv_nsec
== 0)
3364 its
.it_value
.tv_nsec
= 1;
3366 r
= timerfd_settime(e
->watchdog_fd
, TFD_TIMER_ABSTIME
, &its
, NULL
);
3373 static int process_watchdog(sd_event
*e
) {
3379 /* Don't notify watchdog too often */
3380 if (e
->watchdog_last
+ e
->watchdog_period
/ 4 > e
->timestamp
.monotonic
)
3383 sd_notify(false, "WATCHDOG=1");
3384 e
->watchdog_last
= e
->timestamp
.monotonic
;
3386 return arm_watchdog(e
);
3389 static void event_close_inode_data_fds(sd_event
*e
) {
3390 struct inode_data
*d
;
3394 /* Close the fds pointing to the inodes to watch now. We need to close them as they might otherwise pin
3395 * filesystems. But we can't close them right-away as we need them as long as the user still wants to make
3396 * adjustments to the even source, such as changing the priority (which requires us to remove and re-add a watch
3397 * for the inode). Hence, let's close them when entering the first iteration after they were added, as a
3400 while ((d
= e
->inode_data_to_close
)) {
3402 d
->fd
= safe_close(d
->fd
);
3404 LIST_REMOVE(to_close
, e
->inode_data_to_close
, d
);
3408 _public_
int sd_event_prepare(sd_event
*e
) {
3411 assert_return(e
, -EINVAL
);
3412 assert_return(e
= event_resolve(e
), -ENOPKG
);
3413 assert_return(!event_pid_changed(e
), -ECHILD
);
3414 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
3415 assert_return(e
->state
== SD_EVENT_INITIAL
, -EBUSY
);
3417 /* Let's check that if we are a default event loop we are executed in the correct thread. We only do
3418 * this check here once, since gettid() is typically not cached, and thus want to minimize
3420 assert_return(!e
->default_event_ptr
|| e
->tid
== gettid(), -EREMOTEIO
);
3422 if (e
->exit_requested
)
3427 e
->state
= SD_EVENT_PREPARING
;
3428 r
= event_prepare(e
);
3429 e
->state
= SD_EVENT_INITIAL
;
3433 r
= event_arm_timer(e
, &e
->realtime
);
3437 r
= event_arm_timer(e
, &e
->boottime
);
3441 r
= event_arm_timer(e
, &e
->monotonic
);
3445 r
= event_arm_timer(e
, &e
->realtime_alarm
);
3449 r
= event_arm_timer(e
, &e
->boottime_alarm
);
3453 event_close_inode_data_fds(e
);
3455 if (event_next_pending(e
) || e
->need_process_child
)
3458 e
->state
= SD_EVENT_ARMED
;
3463 e
->state
= SD_EVENT_ARMED
;
3464 r
= sd_event_wait(e
, 0);
3466 e
->state
= SD_EVENT_ARMED
;
3471 _public_
int sd_event_wait(sd_event
*e
, uint64_t timeout
) {
3472 size_t event_queue_max
;
3475 assert_return(e
, -EINVAL
);
3476 assert_return(e
= event_resolve(e
), -ENOPKG
);
3477 assert_return(!event_pid_changed(e
), -ECHILD
);
3478 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
3479 assert_return(e
->state
== SD_EVENT_ARMED
, -EBUSY
);
3481 if (e
->exit_requested
) {
3482 e
->state
= SD_EVENT_PENDING
;
3486 event_queue_max
= MAX(e
->n_sources
, 1u);
3487 if (!GREEDY_REALLOC(e
->event_queue
, e
->event_queue_allocated
, event_queue_max
))
3490 /* If we still have inotify data buffered, then query the other fds, but don't wait on it */
3491 if (e
->inotify_data_buffered
)
3494 m
= epoll_wait(e
->epoll_fd
, e
->event_queue
, event_queue_max
,
3495 timeout
== (uint64_t) -1 ? -1 : (int) DIV_ROUND_UP(timeout
, USEC_PER_MSEC
));
3497 if (errno
== EINTR
) {
3498 e
->state
= SD_EVENT_PENDING
;
3506 triple_timestamp_get(&e
->timestamp
);
3508 for (i
= 0; i
< m
; i
++) {
3510 if (e
->event_queue
[i
].data
.ptr
== INT_TO_PTR(SOURCE_WATCHDOG
))
3511 r
= flush_timer(e
, e
->watchdog_fd
, e
->event_queue
[i
].events
, NULL
);
3513 WakeupType
*t
= e
->event_queue
[i
].data
.ptr
;
3517 case WAKEUP_EVENT_SOURCE
: {
3518 sd_event_source
*s
= e
->event_queue
[i
].data
.ptr
;
3525 r
= process_io(e
, s
, e
->event_queue
[i
].events
);
3529 r
= process_pidfd(e
, s
, e
->event_queue
[i
].events
);
3533 assert_not_reached("Unexpected event source type");
3539 case WAKEUP_CLOCK_DATA
: {
3540 struct clock_data
*d
= e
->event_queue
[i
].data
.ptr
;
3544 r
= flush_timer(e
, d
->fd
, e
->event_queue
[i
].events
, &d
->next
);
3548 case WAKEUP_SIGNAL_DATA
:
3549 r
= process_signal(e
, e
->event_queue
[i
].data
.ptr
, e
->event_queue
[i
].events
);
3552 case WAKEUP_INOTIFY_DATA
:
3553 r
= event_inotify_data_read(e
, e
->event_queue
[i
].data
.ptr
, e
->event_queue
[i
].events
);
3557 assert_not_reached("Invalid wake-up pointer");
3564 r
= process_watchdog(e
);
3568 r
= process_timer(e
, e
->timestamp
.realtime
, &e
->realtime
);
3572 r
= process_timer(e
, e
->timestamp
.boottime
, &e
->boottime
);
3576 r
= process_timer(e
, e
->timestamp
.monotonic
, &e
->monotonic
);
3580 r
= process_timer(e
, e
->timestamp
.realtime
, &e
->realtime_alarm
);
3584 r
= process_timer(e
, e
->timestamp
.boottime
, &e
->boottime_alarm
);
3588 if (e
->need_process_child
) {
3589 r
= process_child(e
);
3594 r
= process_inotify(e
);
3598 if (event_next_pending(e
)) {
3599 e
->state
= SD_EVENT_PENDING
;
3607 e
->state
= SD_EVENT_INITIAL
;
3612 _public_
int sd_event_dispatch(sd_event
*e
) {
3616 assert_return(e
, -EINVAL
);
3617 assert_return(e
= event_resolve(e
), -ENOPKG
);
3618 assert_return(!event_pid_changed(e
), -ECHILD
);
3619 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
3620 assert_return(e
->state
== SD_EVENT_PENDING
, -EBUSY
);
3622 if (e
->exit_requested
)
3623 return dispatch_exit(e
);
3625 p
= event_next_pending(e
);
3627 _cleanup_(sd_event_unrefp
) sd_event
*ref
= NULL
;
3629 ref
= sd_event_ref(e
);
3630 e
->state
= SD_EVENT_RUNNING
;
3631 r
= source_dispatch(p
);
3632 e
->state
= SD_EVENT_INITIAL
;
3636 e
->state
= SD_EVENT_INITIAL
;
3641 static void event_log_delays(sd_event
*e
) {
3642 char b
[ELEMENTSOF(e
->delays
) * DECIMAL_STR_MAX(unsigned) + 1], *p
;
3647 for (i
= 0; i
< ELEMENTSOF(e
->delays
); i
++) {
3648 l
= strpcpyf(&p
, l
, "%u ", e
->delays
[i
]);
3651 log_debug("Event loop iterations: %s", b
);
3654 _public_
int sd_event_run(sd_event
*e
, uint64_t timeout
) {
3657 assert_return(e
, -EINVAL
);
3658 assert_return(e
= event_resolve(e
), -ENOPKG
);
3659 assert_return(!event_pid_changed(e
), -ECHILD
);
3660 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
3661 assert_return(e
->state
== SD_EVENT_INITIAL
, -EBUSY
);
3663 if (e
->profile_delays
&& e
->last_run
) {
3667 this_run
= now(CLOCK_MONOTONIC
);
3669 l
= u64log2(this_run
- e
->last_run
);
3670 assert(l
< sizeof(e
->delays
));
3673 if (this_run
- e
->last_log
>= 5*USEC_PER_SEC
) {
3674 event_log_delays(e
);
3675 e
->last_log
= this_run
;
3679 r
= sd_event_prepare(e
);
3681 /* There was nothing? Then wait... */
3682 r
= sd_event_wait(e
, timeout
);
3684 if (e
->profile_delays
)
3685 e
->last_run
= now(CLOCK_MONOTONIC
);
3688 /* There's something now, then let's dispatch it */
3689 r
= sd_event_dispatch(e
);
3699 _public_
int sd_event_loop(sd_event
*e
) {
3700 _cleanup_(sd_event_unrefp
) sd_event
*ref
= NULL
;
3703 assert_return(e
, -EINVAL
);
3704 assert_return(e
= event_resolve(e
), -ENOPKG
);
3705 assert_return(!event_pid_changed(e
), -ECHILD
);
3706 assert_return(e
->state
== SD_EVENT_INITIAL
, -EBUSY
);
3708 ref
= sd_event_ref(e
);
3710 while (e
->state
!= SD_EVENT_FINISHED
) {
3711 r
= sd_event_run(e
, (uint64_t) -1);
3716 return e
->exit_code
;
3719 _public_
int sd_event_get_fd(sd_event
*e
) {
3721 assert_return(e
, -EINVAL
);
3722 assert_return(e
= event_resolve(e
), -ENOPKG
);
3723 assert_return(!event_pid_changed(e
), -ECHILD
);
3728 _public_
int sd_event_get_state(sd_event
*e
) {
3729 assert_return(e
, -EINVAL
);
3730 assert_return(e
= event_resolve(e
), -ENOPKG
);
3731 assert_return(!event_pid_changed(e
), -ECHILD
);
3736 _public_
int sd_event_get_exit_code(sd_event
*e
, int *code
) {
3737 assert_return(e
, -EINVAL
);
3738 assert_return(e
= event_resolve(e
), -ENOPKG
);
3739 assert_return(code
, -EINVAL
);
3740 assert_return(!event_pid_changed(e
), -ECHILD
);
3742 if (!e
->exit_requested
)
3745 *code
= e
->exit_code
;
3749 _public_
int sd_event_exit(sd_event
*e
, int code
) {
3750 assert_return(e
, -EINVAL
);
3751 assert_return(e
= event_resolve(e
), -ENOPKG
);
3752 assert_return(e
->state
!= SD_EVENT_FINISHED
, -ESTALE
);
3753 assert_return(!event_pid_changed(e
), -ECHILD
);
3755 e
->exit_requested
= true;
3756 e
->exit_code
= code
;
3761 _public_
int sd_event_now(sd_event
*e
, clockid_t clock
, uint64_t *usec
) {
3762 assert_return(e
, -EINVAL
);
3763 assert_return(e
= event_resolve(e
), -ENOPKG
);
3764 assert_return(usec
, -EINVAL
);
3765 assert_return(!event_pid_changed(e
), -ECHILD
);
3767 if (!TRIPLE_TIMESTAMP_HAS_CLOCK(clock
))
3770 /* Generate a clean error in case CLOCK_BOOTTIME is not available. Note that don't use clock_supported() here,
3771 * for a reason: there are systems where CLOCK_BOOTTIME is supported, but CLOCK_BOOTTIME_ALARM is not, but for
3772 * the purpose of getting the time this doesn't matter. */
3773 if (IN_SET(clock
, CLOCK_BOOTTIME
, CLOCK_BOOTTIME_ALARM
) && !clock_boottime_supported())
3776 if (!triple_timestamp_is_set(&e
->timestamp
)) {
3777 /* Implicitly fall back to now() if we never ran
3778 * before and thus have no cached time. */
3783 *usec
= triple_timestamp_by_clock(&e
->timestamp
, clock
);
3787 _public_
int sd_event_default(sd_event
**ret
) {
3792 return !!default_event
;
3794 if (default_event
) {
3795 *ret
= sd_event_ref(default_event
);
3799 r
= sd_event_new(&e
);
3803 e
->default_event_ptr
= &default_event
;
3811 _public_
int sd_event_get_tid(sd_event
*e
, pid_t
*tid
) {
3812 assert_return(e
, -EINVAL
);
3813 assert_return(e
= event_resolve(e
), -ENOPKG
);
3814 assert_return(tid
, -EINVAL
);
3815 assert_return(!event_pid_changed(e
), -ECHILD
);
3825 _public_
int sd_event_set_watchdog(sd_event
*e
, int b
) {
3828 assert_return(e
, -EINVAL
);
3829 assert_return(e
= event_resolve(e
), -ENOPKG
);
3830 assert_return(!event_pid_changed(e
), -ECHILD
);
3832 if (e
->watchdog
== !!b
)
3836 r
= sd_watchdog_enabled(false, &e
->watchdog_period
);
3840 /* Issue first ping immediately */
3841 sd_notify(false, "WATCHDOG=1");
3842 e
->watchdog_last
= now(CLOCK_MONOTONIC
);
3844 e
->watchdog_fd
= timerfd_create(CLOCK_MONOTONIC
, TFD_NONBLOCK
|TFD_CLOEXEC
);
3845 if (e
->watchdog_fd
< 0)
3848 r
= arm_watchdog(e
);
3852 struct epoll_event ev
= {
3854 .data
.ptr
= INT_TO_PTR(SOURCE_WATCHDOG
),
3857 r
= epoll_ctl(e
->epoll_fd
, EPOLL_CTL_ADD
, e
->watchdog_fd
, &ev
);
3864 if (e
->watchdog_fd
>= 0) {
3865 (void) epoll_ctl(e
->epoll_fd
, EPOLL_CTL_DEL
, e
->watchdog_fd
, NULL
);
3866 e
->watchdog_fd
= safe_close(e
->watchdog_fd
);
3874 e
->watchdog_fd
= safe_close(e
->watchdog_fd
);
3878 _public_
int sd_event_get_watchdog(sd_event
*e
) {
3879 assert_return(e
, -EINVAL
);
3880 assert_return(e
= event_resolve(e
), -ENOPKG
);
3881 assert_return(!event_pid_changed(e
), -ECHILD
);
3886 _public_
int sd_event_get_iteration(sd_event
*e
, uint64_t *ret
) {
3887 assert_return(e
, -EINVAL
);
3888 assert_return(e
= event_resolve(e
), -ENOPKG
);
3889 assert_return(!event_pid_changed(e
), -ECHILD
);
3891 *ret
= e
->iteration
;
3895 _public_
int sd_event_source_set_destroy_callback(sd_event_source
*s
, sd_event_destroy_t callback
) {
3896 assert_return(s
, -EINVAL
);
3898 s
->destroy_callback
= callback
;
3902 _public_
int sd_event_source_get_destroy_callback(sd_event_source
*s
, sd_event_destroy_t
*ret
) {
3903 assert_return(s
, -EINVAL
);
3906 *ret
= s
->destroy_callback
;
3908 return !!s
->destroy_callback
;
3911 _public_
int sd_event_source_get_floating(sd_event_source
*s
) {
3912 assert_return(s
, -EINVAL
);
3917 _public_
int sd_event_source_set_floating(sd_event_source
*s
, int b
) {
3918 assert_return(s
, -EINVAL
);
3920 if (s
->floating
== !!b
)
3923 if (!s
->event
) /* Already disconnected */
3929 sd_event_source_ref(s
);
3930 sd_event_unref(s
->event
);
3932 sd_event_ref(s
->event
);
3933 sd_event_source_unref(s
);