-/*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
-
+/* SPDX-License-Identifier: LGPL-2.1+ */
/***
- This file is part of systemd.
-
Copyright 2013 Lennart Poettering
-
- systemd is free software; you can redistribute it and/or modify it
- under the terms of the GNU Lesser General Public License as published by
- the Free Software Foundation; either version 2.1 of the License, or
- (at your option) any later version.
-
- systemd is distributed in the hope that it will be useful, but
- WITHOUT ANY WARRANTY; without even the implied warranty of
- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- Lesser General Public License for more details.
-
- You should have received a copy of the GNU Lesser General Public License
- along with systemd; If not, see <http://www.gnu.org/licenses/>.
***/
#include <sys/epoll.h>
#include <sys/timerfd.h>
#include <sys/wait.h>
-#include "sd-id128.h"
#include "sd-daemon.h"
-#include "macro.h"
-#include "prioq.h"
+#include "sd-event.h"
+#include "sd-id128.h"
+
+#include "alloc-util.h"
+#include "fd-util.h"
+//#include "fs-util.h"
#include "hashmap.h"
-#include "util.h"
-#include "time-util.h"
+#include "list.h"
+#include "macro.h"
#include "missing.h"
+#include "prioq.h"
+#include "process-util.h"
#include "set.h"
-#include "list.h"
#include "signal-util.h"
-
-#include "sd-event.h"
+#include "string-table.h"
+#include "string-util.h"
+#include "time-util.h"
+#include "util.h"
#define DEFAULT_ACCURACY_USEC (250 * USEC_PER_MSEC)
SOURCE_POST,
SOURCE_EXIT,
SOURCE_WATCHDOG,
+ SOURCE_INOTIFY,
_SOURCE_EVENT_SOURCE_TYPE_MAX,
_SOURCE_EVENT_SOURCE_TYPE_INVALID = -1
} EventSourceType;
+static const char* const event_source_type_table[_SOURCE_EVENT_SOURCE_TYPE_MAX] = {
+ [SOURCE_IO] = "io",
+ [SOURCE_TIME_REALTIME] = "realtime",
+ [SOURCE_TIME_BOOTTIME] = "bootime",
+ [SOURCE_TIME_MONOTONIC] = "monotonic",
+ [SOURCE_TIME_REALTIME_ALARM] = "realtime-alarm",
+ [SOURCE_TIME_BOOTTIME_ALARM] = "boottime-alarm",
+ [SOURCE_SIGNAL] = "signal",
+ [SOURCE_CHILD] = "child",
+ [SOURCE_DEFER] = "defer",
+ [SOURCE_POST] = "post",
+ [SOURCE_EXIT] = "exit",
+ [SOURCE_WATCHDOG] = "watchdog",
+ [SOURCE_INOTIFY] = "inotify",
+};
+
+DEFINE_PRIVATE_STRING_TABLE_LOOKUP_TO_STRING(event_source_type, int);
+
/* All objects we use in epoll events start with this value, so that
* we know how to dispatch it */
typedef enum WakeupType {
WAKEUP_EVENT_SOURCE,
WAKEUP_CLOCK_DATA,
WAKEUP_SIGNAL_DATA,
+ WAKEUP_INOTIFY_DATA,
_WAKEUP_TYPE_MAX,
_WAKEUP_TYPE_INVALID = -1,
} WakeupType;
#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)
+struct inode_data;
+
struct sd_event_source {
WakeupType wakeup;
int64_t priority;
unsigned pending_index;
unsigned prepare_index;
- unsigned pending_iteration;
- unsigned prepare_iteration;
+ uint64_t pending_iteration;
+ uint64_t prepare_iteration;
+
+ sd_event_destroy_t destroy_callback;
LIST_FIELDS(sd_event_source, sources);
uint32_t events;
uint32_t revents;
bool registered:1;
+ bool owned:1;
} io;
struct {
sd_event_time_handler_t callback;
sd_event_handler_t callback;
unsigned prioq_index;
} exit;
+ struct {
+ sd_event_inotify_handler_t callback;
+ uint32_t mask;
+ struct inode_data *inode_data;
+ LIST_FIELDS(sd_event_source, by_inode_data);
+ } inotify;
};
};
sd_event_source *current;
};
+/* A structure listing all event sources currently watching a specific inode */
+struct inode_data {
+ /* The identifier for the inode, the combination of the .st_dev + .st_ino fields of the file */
+ ino_t ino;
+ dev_t dev;
+
+ /* An fd of the inode to watch. The fd is kept open until the next iteration of the loop, so that we can
+ * rearrange the priority still until then, as we need the original inode to change the priority as we need to
+ * add a watch descriptor to the right inotify for the priority which we can only do if we have a handle to the
+ * original inode. We keep a list of all inode_data objects with an open fd in the to_close list (see below) of
+ * the sd-event object, so that it is efficient to close everything, before entering the next event loop
+ * iteration. */
+ int fd;
+
+ /* The inotify "watch descriptor" */
+ int wd;
+
+ /* The combination of the mask of all inotify watches on this inode we manage. This is also the mask that has
+ * most recently been set on the watch descriptor. */
+ uint32_t combined_mask;
+
+ /* All event sources subscribed to this inode */
+ LIST_HEAD(sd_event_source, event_sources);
+
+ /* The inotify object we watch this inode with */
+ struct inotify_data *inotify_data;
+
+ /* A linked list of all inode data objects with fds to close (see above) */
+ LIST_FIELDS(struct inode_data, to_close);
+};
+
+/* A structure encapsulating an inotify fd */
+struct inotify_data {
+ WakeupType wakeup;
+
+ /* For each priority we maintain one inotify fd, so that we only have to dequeue a single event per priority at
+ * a time */
+
+ int fd;
+ int64_t priority;
+
+ Hashmap *inodes; /* The inode_data structures keyed by dev+ino */
+ Hashmap *wd; /* The inode_data structures keyed by the watch descriptor for each */
+
+ /* The buffer we read inotify events into */
+ union inotify_event_buffer buffer;
+ size_t buffer_filled; /* fill level of the buffer */
+
+ /* How many event sources are currently marked pending for this inotify. We won't read new events off the
+ * inotify fd as long as there are still pending events on the inotify (because we have no strategy of queuing
+ * the events locally if they can't be coalesced). */
+ unsigned n_pending;
+
+ /* A linked list of all inotify objects with data already read, that still need processing. We keep this list
+ * to make it efficient to figure out what inotify objects to process data on next. */
+ LIST_FIELDS(struct inotify_data, buffered);
+};
+
struct sd_event {
unsigned n_ref;
Prioq *exit;
+ Hashmap *inotify_data; /* indexed by priority */
+
+ /* A list of inode structures that still have an fd open, that we need to close before the next loop iteration */
+ LIST_HEAD(struct inode_data, inode_data_to_close);
+
+ /* A list of inotify objects that already have events buffered which aren't processed yet */
+ LIST_HEAD(struct inotify_data, inotify_data_buffered);
+
pid_t original_pid;
- unsigned iteration;
- dual_timestamp timestamp;
- usec_t timestamp_boottime;
+ uint64_t iteration;
+ triple_timestamp timestamp;
int state;
bool exit_requested:1;
bool need_process_child:1;
bool watchdog:1;
+ bool profile_delays:1;
int exit_code;
unsigned n_sources;
LIST_HEAD(sd_event_source, sources);
+
+ usec_t last_run, last_log;
+ unsigned delays[sizeof(usec_t) * 8];
};
+static thread_local sd_event *default_event = NULL;
+
static void source_disconnect(sd_event_source *s);
+static void event_gc_inode_data(sd_event *e, struct inode_data *d);
+
+static sd_event *event_resolve(sd_event *e) {
+ return e == SD_EVENT_DEFAULT ? default_event : e;
+}
static int pending_prioq_compare(const void *a, const void *b) {
const sd_event_source *x = a, *y = b;
if (x->pending_iteration > y->pending_iteration)
return 1;
- /* Stability for the rest */
- if (x < y)
- return -1;
- if (x > y)
- return 1;
-
return 0;
}
assert(x->prepare);
assert(y->prepare);
+ /* Enabled ones first */
+ if (x->enabled != SD_EVENT_OFF && y->enabled == SD_EVENT_OFF)
+ return -1;
+ if (x->enabled == SD_EVENT_OFF && y->enabled != SD_EVENT_OFF)
+ return 1;
+
/* Move most recently prepared ones last, so that we can stop
* preparing as soon as we hit one that has already been
* prepared in the current iteration */
if (x->prepare_iteration > y->prepare_iteration)
return 1;
- /* Enabled ones first */
- if (x->enabled != SD_EVENT_OFF && y->enabled == SD_EVENT_OFF)
- return -1;
- if (x->enabled == SD_EVENT_OFF && y->enabled != SD_EVENT_OFF)
- return 1;
-
/* Lower priority values first */
if (x->priority < y->priority)
return -1;
if (x->priority > y->priority)
return 1;
- /* Stability for the rest */
- if (x < y)
- return -1;
- if (x > y)
- return 1;
-
return 0;
}
if (x->time.next > y->time.next)
return 1;
- /* Stability for the rest */
- if (x < y)
- return -1;
- if (x > y)
- return 1;
-
return 0;
}
+static usec_t time_event_source_latest(const sd_event_source *s) {
+ return usec_add(s->time.next, s->time.accuracy);
+}
+
static int latest_time_prioq_compare(const void *a, const void *b) {
const sd_event_source *x = a, *y = b;
return 1;
/* Order by time */
- if (x->time.next + x->time.accuracy < y->time.next + y->time.accuracy)
+ if (time_event_source_latest(x) < time_event_source_latest(y))
return -1;
- if (x->time.next + x->time.accuracy > y->time.next + y->time.accuracy)
- return 1;
-
- /* Stability for the rest */
- if (x < y)
- return -1;
- if (x > y)
+ if (time_event_source_latest(x) > time_event_source_latest(y))
return 1;
return 0;
if (x->priority > y->priority)
return 1;
- /* Stability for the rest */
- if (x < y)
- return -1;
- if (x > y)
- return 1;
-
return 0;
}
free(e->signal_sources);
hashmap_free(e->signal_data);
+ hashmap_free(e->inotify_data);
+
hashmap_free(e->child_sources);
set_free(e->post_sources);
free(e);
assert_return(ret, -EINVAL);
- e = new0(sd_event, 1);
+ e = new(sd_event, 1);
if (!e)
return -ENOMEM;
- e->n_ref = 1;
- e->watchdog_fd = e->epoll_fd = e->realtime.fd = e->boottime.fd = e->monotonic.fd = e->realtime_alarm.fd = e->boottime_alarm.fd = -1;
- e->realtime.next = e->boottime.next = e->monotonic.next = e->realtime_alarm.next = e->boottime_alarm.next = USEC_INFINITY;
- e->realtime.wakeup = e->boottime.wakeup = e->monotonic.wakeup = e->realtime_alarm.wakeup = e->boottime_alarm.wakeup = WAKEUP_CLOCK_DATA;
- e->original_pid = getpid();
- e->perturb = USEC_INFINITY;
+ *e = (sd_event) {
+ .n_ref = 1,
+ .epoll_fd = -1,
+ .watchdog_fd = -1,
+ .realtime.wakeup = WAKEUP_CLOCK_DATA,
+ .realtime.fd = -1,
+ .realtime.next = USEC_INFINITY,
+ .boottime.wakeup = WAKEUP_CLOCK_DATA,
+ .boottime.fd = -1,
+ .boottime.next = USEC_INFINITY,
+ .monotonic.wakeup = WAKEUP_CLOCK_DATA,
+ .monotonic.fd = -1,
+ .monotonic.next = USEC_INFINITY,
+ .realtime_alarm.wakeup = WAKEUP_CLOCK_DATA,
+ .realtime_alarm.fd = -1,
+ .realtime_alarm.next = USEC_INFINITY,
+ .boottime_alarm.wakeup = WAKEUP_CLOCK_DATA,
+ .boottime_alarm.fd = -1,
+ .boottime_alarm.next = USEC_INFINITY,
+ .perturb = USEC_INFINITY,
+ .original_pid = getpid_cached(),
+ };
- e->pending = prioq_new(pending_prioq_compare);
- if (!e->pending) {
- r = -ENOMEM;
+ r = prioq_ensure_allocated(&e->pending, pending_prioq_compare);
+ if (r < 0)
goto fail;
- }
e->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
if (e->epoll_fd < 0) {
goto fail;
}
+ e->epoll_fd = fd_move_above_stdio(e->epoll_fd);
+
+ if (secure_getenv("SD_EVENT_PROFILE_DELAYS")) {
+ 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.");
+ e->profile_delays = true;
+ }
+
*ret = e;
return 0;
}
_public_ sd_event* sd_event_ref(sd_event *e) {
- assert_return(e, NULL);
+
+ if (!e)
+ return NULL;
assert(e->n_ref >= 1);
e->n_ref++;
/* We don't support people creating an event loop and keeping
* it around over a fork(). Let's complain. */
- return e->original_pid != getpid();
+ return e->original_pid != getpid_cached();
}
static void source_io_unregister(sd_event_source *s) {
r = epoll_ctl(s->event->epoll_fd, EPOLL_CTL_DEL, s->io.fd, NULL);
if (r < 0)
- log_debug_errno(errno, "Failed to remove source %s from epoll: %m", strna(s->description));
+ log_debug_errno(errno, "Failed to remove source %s (type %s) from epoll: %m",
+ strna(s->description), event_source_type_to_string(s->type));
s->io.registered = false;
}
int enabled,
uint32_t events) {
- struct epoll_event ev = {};
+ struct epoll_event ev;
int r;
assert(s);
assert(s->type == SOURCE_IO);
assert(enabled != SD_EVENT_OFF);
- ev.events = events;
- ev.data.ptr = s;
-
- if (enabled == SD_EVENT_ONESHOT)
- ev.events |= EPOLLONESHOT;
+ ev = (struct epoll_event) {
+ .events = events | (enabled == SD_EVENT_ONESHOT ? EPOLLONESHOT : 0),
+ .data.ptr = s,
+ };
if (s->io.registered)
r = epoll_ctl(s->event->epoll_fd, EPOLL_CTL_MOD, s->io.fd, &ev);
return 0;
}
-/// UNNEEDED by elogind
-#if 0
static clockid_t event_source_type_to_clock(EventSourceType t) {
switch (t) {
return (clockid_t) -1;
}
}
-#endif // 0
static EventSourceType clock_to_event_source_type(clockid_t clock) {
int sig,
struct signal_data **ret) {
- struct epoll_event ev = {};
+ struct epoll_event ev;
struct signal_data *d;
bool added = false;
sigset_t ss_copy;
if (e->signal_sources && e->signal_sources[sig])
priority = e->signal_sources[sig]->priority;
else
- priority = 0;
+ priority = SD_EVENT_PRIORITY_NORMAL;
d = hashmap_get(e->signal_data, &priority);
if (d) {
if (sigismember(&d->sigset, sig) > 0) {
if (ret)
*ret = d;
- return 0;
+ return 0;
}
} else {
r = hashmap_ensure_allocated(&e->signal_data, &uint64_hash_ops);
if (r < 0)
return r;
- d = new0(struct signal_data, 1);
+ d = new(struct signal_data, 1);
if (!d)
return -ENOMEM;
- d->wakeup = WAKEUP_SIGNAL_DATA;
- d->fd = -1;
- d->priority = priority;
+ *d = (struct signal_data) {
+ .wakeup = WAKEUP_SIGNAL_DATA,
+ .fd = -1,
+ .priority = priority,
+ };
r = hashmap_put(e->signal_data, &d->priority, d);
- if (r < 0)
+ if (r < 0) {
+ free(d);
return r;
+ }
added = true;
}
return 0;
}
- d->fd = r;
+ d->fd = fd_move_above_stdio(r);
- ev.events = EPOLLIN;
- ev.data.ptr = d;
+ ev = (struct epoll_event) {
+ .events = EPOLLIN,
+ .data.ptr = d,
+ };
r = epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, d->fd, &ev);
- if (r < 0) {
+ if (r < 0) {
r = -errno;
goto fail;
}
/* If all the mask is all-zero we can get rid of the structure */
hashmap_remove(e->signal_data, &d->priority);
- assert(!d->current);
safe_close(d->fd);
free(d);
return;
s->event->n_enabled_child_sources--;
}
- (void) hashmap_remove(s->event->child_sources, INT_TO_PTR(s->child.pid));
+ (void) hashmap_remove(s->event->child_sources, PID_TO_PTR(s->child.pid));
event_gc_signal_data(s->event, &s->priority, SIGCHLD);
}
prioq_remove(s->event->exit, s, &s->exit.prioq_index);
break;
+ case SOURCE_INOTIFY: {
+ struct inode_data *inode_data;
+
+ inode_data = s->inotify.inode_data;
+ if (inode_data) {
+ struct inotify_data *inotify_data;
+ assert_se(inotify_data = inode_data->inotify_data);
+
+ /* Detach this event source from the inode object */
+ LIST_REMOVE(inotify.by_inode_data, inode_data->event_sources, s);
+ s->inotify.inode_data = NULL;
+
+ if (s->pending) {
+ assert(inotify_data->n_pending > 0);
+ inotify_data->n_pending--;
+ }
+
+ /* Note that we don't reduce the inotify mask for the watch descriptor here if the inode is
+ * continued to being watched. That's because inotify doesn't really have an API for that: we
+ * can only change watch masks with access to the original inode either by fd or by path. But
+ * paths aren't stable, and keeping an O_PATH fd open all the time would mean wasting an fd
+ * continously and keeping the mount busy which we can't really do. We could reconstruct the
+ * original inode from /proc/self/fdinfo/$INOTIFY_FD (as all watch descriptors are listed
+ * there), but given the need for open_by_handle_at() which is privileged and not universally
+ * available this would be quite an incomplete solution. Hence we go the other way, leave the
+ * mask set, even if it is not minimized now, and ignore all events we aren't interested in
+ * anymore after reception. Yes, this sucks, but … Linux … */
+
+ /* Maybe release the inode data (and its inotify) */
+ event_gc_inode_data(s->event, inode_data);
+ }
+
+ break;
+ }
+
default:
assert_not_reached("Wut? I shouldn't exist.");
}
assert(s);
source_disconnect(s);
+
+ if (s->type == SOURCE_IO && s->io.owned)
+ s->io.fd = safe_close(s->io.fd);
+
+ if (s->destroy_callback)
+ s->destroy_callback(s->userdata);
+
free(s->description);
free(s);
}
d->current = NULL;
}
+ if (s->type == SOURCE_INOTIFY) {
+
+ assert(s->inotify.inode_data);
+ assert(s->inotify.inode_data->inotify_data);
+
+ if (b)
+ s->inotify.inode_data->inotify_data->n_pending ++;
+ else {
+ assert(s->inotify.inode_data->inotify_data->n_pending > 0);
+ s->inotify.inode_data->inotify_data->n_pending --;
+ }
+ }
+
return 0;
}
assert(e);
- s = new0(sd_event_source, 1);
+ s = new(sd_event_source, 1);
if (!s)
return NULL;
- s->n_ref = 1;
- s->event = e;
- s->floating = floating;
- s->type = type;
- s->pending_index = s->prepare_index = PRIOQ_IDX_NULL;
+ *s = (struct sd_event_source) {
+ .n_ref = 1,
+ .event = e,
+ .floating = floating,
+ .type = type,
+ .pending_index = PRIOQ_IDX_NULL,
+ .prepare_index = PRIOQ_IDX_NULL,
+ };
if (!floating)
sd_event_ref(e);
LIST_PREPEND(sources, e->sources, s);
- e->n_sources ++;
+ e->n_sources++;
return s;
}
int r;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(fd >= 0, -EBADF);
assert_return(!(events & ~(EPOLLIN|EPOLLOUT|EPOLLRDHUP|EPOLLPRI|EPOLLERR|EPOLLHUP|EPOLLET)), -EINVAL);
assert_return(callback, -EINVAL);
struct clock_data *d,
clockid_t clock) {
- struct epoll_event ev = {};
+ struct epoll_event ev;
int r, fd;
assert(e);
if (fd < 0)
return -errno;
- ev.events = EPOLLIN;
- ev.data.ptr = d;
+ fd = fd_move_above_stdio(fd);
+
+ ev = (struct epoll_event) {
+ .events = EPOLLIN,
+ .data.ptr = d,
+ };
r = epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, fd, &ev);
if (r < 0) {
int r;
assert_return(e, -EINVAL);
- assert_return(usec != (uint64_t) -1, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(accuracy != (uint64_t) -1, -EINVAL);
assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(!event_pid_changed(e), -ECHILD);
+ if (!clock_supported(clock)) /* Checks whether the kernel supports the clock */
+ return -EOPNOTSUPP;
+
+ type = clock_to_event_source_type(clock); /* checks whether sd-event supports this clock */
+ if (type < 0)
+ return -EOPNOTSUPP;
+
if (!callback)
callback = time_exit_callback;
- type = clock_to_event_source_type(clock);
- assert_return(type >= 0, -EOPNOTSUPP);
-
d = event_get_clock_data(e, type);
assert(d);
- if (!d->earliest) {
- d->earliest = prioq_new(earliest_time_prioq_compare);
- if (!d->earliest)
- return -ENOMEM;
- }
+ r = prioq_ensure_allocated(&d->earliest, earliest_time_prioq_compare);
+ if (r < 0)
+ return r;
- if (!d->latest) {
- d->latest = prioq_new(latest_time_prioq_compare);
- if (!d->latest)
- return -ENOMEM;
- }
+ r = prioq_ensure_allocated(&d->latest, latest_time_prioq_compare);
+ if (r < 0)
+ return r;
if (d->fd < 0) {
r = event_setup_timer_fd(e, d, clock);
int r;
assert_return(e, -EINVAL);
- assert_return(sig > 0, -EINVAL);
- assert_return(sig < _NSIG, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
+ assert_return(SIGNAL_VALID(sig), -EINVAL);
assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(!event_pid_changed(e), -ECHILD);
callback = signal_exit_callback;
r = pthread_sigmask(SIG_SETMASK, NULL, &ss);
- if (r < 0)
- return -errno;
+ if (r != 0)
+ return -r;
if (!sigismember(&ss, sig))
return -EBUSY;
e->signal_sources[sig] = s;
r = event_make_signal_data(e, sig, &d);
- if (r < 0) {
- source_free(s);
- return r;
- }
+ if (r < 0) {
+ source_free(s);
+ return r;
+ }
/* Use the signal name as description for the event source by default */
(void) sd_event_source_set_description(s, signal_to_string(sig));
int r;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(pid > 1, -EINVAL);
assert_return(!(options & ~(WEXITED|WSTOPPED|WCONTINUED)), -EINVAL);
assert_return(options != 0, -EINVAL);
if (r < 0)
return r;
- if (hashmap_contains(e->child_sources, INT_TO_PTR(pid)))
+ if (hashmap_contains(e->child_sources, PID_TO_PTR(pid)))
return -EBUSY;
s = source_new(e, !ret, SOURCE_CHILD);
s->userdata = userdata;
s->enabled = SD_EVENT_ONESHOT;
- r = hashmap_put(e->child_sources, INT_TO_PTR(pid), s);
+ r = hashmap_put(e->child_sources, PID_TO_PTR(pid), s);
if (r < 0) {
source_free(s);
return r;
}
- e->n_enabled_child_sources ++;
+ e->n_enabled_child_sources++;
r = event_make_signal_data(e, SIGCHLD, NULL);
- if (r < 0) {
+ if (r < 0) {
e->n_enabled_child_sources--;
- source_free(s);
- return r;
- }
+ source_free(s);
+ return r;
+ }
e->need_process_child = true;
int r;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(callback, -EINVAL);
assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(!event_pid_changed(e), -ECHILD);
int r;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(callback, -EINVAL);
assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(!event_pid_changed(e), -ECHILD);
int r;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(callback, -EINVAL);
assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(!event_pid_changed(e), -ECHILD);
- if (!e->exit) {
- e->exit = prioq_new(exit_prioq_compare);
- if (!e->exit)
- return -ENOMEM;
- }
+ r = prioq_ensure_allocated(&e->exit, exit_prioq_compare);
+ if (r < 0)
+ return r;
s = source_new(e, !ret, SOURCE_EXIT);
if (!s)
return 0;
}
-/// UNNEEDED by elogind
-#if 0
+static void event_free_inotify_data(sd_event *e, struct inotify_data *d) {
+ assert(e);
+
+ if (!d)
+ return;
+
+ assert(hashmap_isempty(d->inodes));
+ assert(hashmap_isempty(d->wd));
+
+ if (d->buffer_filled > 0)
+ LIST_REMOVE(buffered, e->inotify_data_buffered, d);
+
+ hashmap_free(d->inodes);
+ hashmap_free(d->wd);
+
+ assert_se(hashmap_remove(e->inotify_data, &d->priority) == d);
+
+ if (d->fd >= 0) {
+ if (epoll_ctl(e->epoll_fd, EPOLL_CTL_DEL, d->fd, NULL) < 0)
+ log_debug_errno(errno, "Failed to remove inotify fd from epoll, ignoring: %m");
+
+ safe_close(d->fd);
+ }
+ free(d);
+}
+
+static int event_make_inotify_data(
+ sd_event *e,
+ int64_t priority,
+ struct inotify_data **ret) {
+
+ _cleanup_close_ int fd = -1;
+ struct inotify_data *d;
+ struct epoll_event ev;
+ int r;
+
+ assert(e);
+
+ d = hashmap_get(e->inotify_data, &priority);
+ if (d) {
+ if (ret)
+ *ret = d;
+ return 0;
+ }
+
+ fd = inotify_init1(IN_NONBLOCK|O_CLOEXEC);
+ if (fd < 0)
+ return -errno;
+
+ fd = fd_move_above_stdio(fd);
+
+ r = hashmap_ensure_allocated(&e->inotify_data, &uint64_hash_ops);
+ if (r < 0)
+ return r;
+
+ d = new(struct inotify_data, 1);
+ if (!d)
+ return -ENOMEM;
+
+ *d = (struct inotify_data) {
+ .wakeup = WAKEUP_INOTIFY_DATA,
+ .fd = TAKE_FD(fd),
+ .priority = priority,
+ };
+
+ r = hashmap_put(e->inotify_data, &d->priority, d);
+ if (r < 0) {
+ d->fd = safe_close(d->fd);
+ free(d);
+ return r;
+ }
+
+ ev = (struct epoll_event) {
+ .events = EPOLLIN,
+ .data.ptr = d,
+ };
+
+ if (epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, d->fd, &ev) < 0) {
+ r = -errno;
+ d->fd = safe_close(d->fd); /* let's close this ourselves, as event_free_inotify_data() would otherwise
+ * remove the fd from the epoll first, which we don't want as we couldn't
+ * add it in the first place. */
+ event_free_inotify_data(e, d);
+ return r;
+ }
+
+ if (ret)
+ *ret = d;
+
+ return 1;
+}
+
+static int inode_data_compare(const void *a, const void *b) {
+ const struct inode_data *x = a, *y = b;
+
+ assert(x);
+ assert(y);
+
+ if (x->dev < y->dev)
+ return -1;
+ if (x->dev > y->dev)
+ return 1;
+
+ if (x->ino < y->ino)
+ return -1;
+ if (x->ino > y->ino)
+ return 1;
+
+ return 0;
+}
+
+static void inode_data_hash_func(const void *p, struct siphash *state) {
+ const struct inode_data *d = p;
+
+ assert(p);
+
+ siphash24_compress(&d->dev, sizeof(d->dev), state);
+ siphash24_compress(&d->ino, sizeof(d->ino), state);
+}
+
+const struct hash_ops inode_data_hash_ops = {
+ .hash = inode_data_hash_func,
+ .compare = inode_data_compare
+};
+
+static void event_free_inode_data(
+ sd_event *e,
+ struct inode_data *d) {
+
+ assert(e);
+
+ if (!d)
+ return;
+
+ assert(!d->event_sources);
+
+ if (d->fd >= 0) {
+ LIST_REMOVE(to_close, e->inode_data_to_close, d);
+ safe_close(d->fd);
+ }
+
+ if (d->inotify_data) {
+
+ if (d->wd >= 0) {
+ if (d->inotify_data->fd >= 0) {
+ /* So here's a problem. At the time this runs the watch descriptor might already be
+ * invalidated, because an IN_IGNORED event might be queued right the moment we enter
+ * the syscall. Hence, whenever we get EINVAL, ignore it entirely, since it's a very
+ * likely case to happen. */
+
+ if (inotify_rm_watch(d->inotify_data->fd, d->wd) < 0 && errno != EINVAL)
+ log_debug_errno(errno, "Failed to remove watch descriptor %i from inotify, ignoring: %m", d->wd);
+ }
+
+ assert_se(hashmap_remove(d->inotify_data->wd, INT_TO_PTR(d->wd)) == d);
+ }
+
+ assert_se(hashmap_remove(d->inotify_data->inodes, d) == d);
+ }
+
+ free(d);
+}
+
+static void event_gc_inode_data(
+ sd_event *e,
+ struct inode_data *d) {
+
+ struct inotify_data *inotify_data;
+
+ assert(e);
+
+ if (!d)
+ return;
+
+ if (d->event_sources)
+ return;
+
+ inotify_data = d->inotify_data;
+ event_free_inode_data(e, d);
+
+ if (inotify_data && hashmap_isempty(inotify_data->inodes))
+ event_free_inotify_data(e, inotify_data);
+}
+
+static int event_make_inode_data(
+ sd_event *e,
+ struct inotify_data *inotify_data,
+ dev_t dev,
+ ino_t ino,
+ struct inode_data **ret) {
+
+ struct inode_data *d, key;
+ int r;
+
+ assert(e);
+ assert(inotify_data);
+
+ key = (struct inode_data) {
+ .ino = ino,
+ .dev = dev,
+ };
+
+ d = hashmap_get(inotify_data->inodes, &key);
+ if (d) {
+ if (ret)
+ *ret = d;
+
+ return 0;
+ }
+
+ r = hashmap_ensure_allocated(&inotify_data->inodes, &inode_data_hash_ops);
+ if (r < 0)
+ return r;
+
+ d = new(struct inode_data, 1);
+ if (!d)
+ return -ENOMEM;
+
+ *d = (struct inode_data) {
+ .dev = dev,
+ .ino = ino,
+ .wd = -1,
+ .fd = -1,
+ .inotify_data = inotify_data,
+ };
+
+ r = hashmap_put(inotify_data->inodes, d, d);
+ if (r < 0) {
+ free(d);
+ return r;
+ }
+
+ if (ret)
+ *ret = d;
+
+ return 1;
+}
+
+static uint32_t inode_data_determine_mask(struct inode_data *d) {
+ bool excl_unlink = true;
+ uint32_t combined = 0;
+ sd_event_source *s;
+
+ assert(d);
+
+ /* Combines the watch masks of all event sources watching this inode. We generally just OR them together, but
+ * the IN_EXCL_UNLINK flag is ANDed instead.
+ *
+ * Note that we add all sources to the mask here, regardless whether enabled, disabled or oneshot. That's
+ * because we cannot change the mask anymore after the event source was created once, since the kernel has no
+ * API for that. Hence we need to subscribe to the maximum mask we ever might be interested in, and supress
+ * events we don't care for client-side. */
+
+ LIST_FOREACH(inotify.by_inode_data, s, d->event_sources) {
+
+ if ((s->inotify.mask & IN_EXCL_UNLINK) == 0)
+ excl_unlink = false;
+
+ combined |= s->inotify.mask;
+ }
+
+ return (combined & ~(IN_ONESHOT|IN_DONT_FOLLOW|IN_ONLYDIR|IN_EXCL_UNLINK)) | (excl_unlink ? IN_EXCL_UNLINK : 0);
+}
+
+static int inode_data_realize_watch(sd_event *e, struct inode_data *d) {
+ uint32_t combined_mask;
+ int wd, r;
+
+ assert(d);
+ assert(d->fd >= 0);
+
+ combined_mask = inode_data_determine_mask(d);
+
+ if (d->wd >= 0 && combined_mask == d->combined_mask)
+ return 0;
+
+ r = hashmap_ensure_allocated(&d->inotify_data->wd, NULL);
+ if (r < 0)
+ return r;
+
+ wd = inotify_add_watch_fd(d->inotify_data->fd, d->fd, combined_mask);
+ if (wd < 0)
+ return -errno;
+
+ if (d->wd < 0) {
+ r = hashmap_put(d->inotify_data->wd, INT_TO_PTR(wd), d);
+ if (r < 0) {
+ (void) inotify_rm_watch(d->inotify_data->fd, wd);
+ return r;
+ }
+
+ d->wd = wd;
+
+ } else if (d->wd != wd) {
+
+ log_debug("Weird, the watch descriptor we already knew for this inode changed?");
+ (void) inotify_rm_watch(d->fd, wd);
+ return -EINVAL;
+ }
+
+ d->combined_mask = combined_mask;
+ return 1;
+}
+
+_public_ int sd_event_add_inotify(
+ sd_event *e,
+ sd_event_source **ret,
+ const char *path,
+ uint32_t mask,
+ sd_event_inotify_handler_t callback,
+ void *userdata) {
+
+ bool rm_inotify = false, rm_inode = false;
+ struct inotify_data *inotify_data = NULL;
+ struct inode_data *inode_data = NULL;
+ _cleanup_close_ int fd = -1;
+ sd_event_source *s;
+ struct stat st;
+ int r;
+
+ assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
+ assert_return(path, -EINVAL);
+ assert_return(callback, -EINVAL);
+ assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
+ assert_return(!event_pid_changed(e), -ECHILD);
+
+ /* Refuse IN_MASK_ADD since we coalesce watches on the same inode, and hence really don't want to merge
+ * masks. Or in other words, this whole code exists only to manage IN_MASK_ADD type operations for you, hence
+ * the user can't use them for us. */
+ if (mask & IN_MASK_ADD)
+ return -EINVAL;
+
+ fd = open(path, O_PATH|O_CLOEXEC|
+ (mask & IN_ONLYDIR ? O_DIRECTORY : 0)|
+ (mask & IN_DONT_FOLLOW ? O_NOFOLLOW : 0));
+ if (fd < 0)
+ return -errno;
+
+ if (fstat(fd, &st) < 0)
+ return -errno;
+
+ s = source_new(e, !ret, SOURCE_INOTIFY);
+ if (!s)
+ return -ENOMEM;
+
+ s->enabled = mask & IN_ONESHOT ? SD_EVENT_ONESHOT : SD_EVENT_ON;
+ s->inotify.mask = mask;
+ s->inotify.callback = callback;
+ s->userdata = userdata;
+
+ /* Allocate an inotify object for this priority, and an inode object within it */
+ r = event_make_inotify_data(e, SD_EVENT_PRIORITY_NORMAL, &inotify_data);
+ if (r < 0)
+ goto fail;
+ rm_inotify = r > 0;
+
+ r = event_make_inode_data(e, inotify_data, st.st_dev, st.st_ino, &inode_data);
+ if (r < 0)
+ goto fail;
+ rm_inode = r > 0;
+
+ /* Keep the O_PATH fd around until the first iteration of the loop, so that we can still change the priority of
+ * the event source, until then, for which we need the original inode. */
+ if (inode_data->fd < 0) {
+ inode_data->fd = TAKE_FD(fd);
+ LIST_PREPEND(to_close, e->inode_data_to_close, inode_data);
+ }
+
+ /* Link our event source to the inode data object */
+ LIST_PREPEND(inotify.by_inode_data, inode_data->event_sources, s);
+ s->inotify.inode_data = inode_data;
+
+ rm_inode = rm_inotify = false;
+
+ /* Actually realize the watch now */
+ r = inode_data_realize_watch(e, inode_data);
+ if (r < 0)
+ goto fail;
+
+ (void) sd_event_source_set_description(s, path);
+
+ if (ret)
+ *ret = s;
+
+ return 0;
+
+fail:
+ source_free(s);
+
+ if (rm_inode)
+ event_free_inode_data(e, inode_data);
+
+ if (rm_inotify)
+ event_free_inotify_data(e, inotify_data);
+
+ return r;
+}
+
_public_ sd_event_source* sd_event_source_ref(sd_event_source *s) {
- assert_return(s, NULL);
+
+ if (!s)
+ return NULL;
assert(s->n_ref >= 1);
s->n_ref++;
return s;
}
-#endif // 0
_public_ sd_event_source* sd_event_source_unref(sd_event_source *s) {
return free_and_strdup(&s->description, description);
}
-/// UNNEEDED by elogind
-#if 0
_public_ int sd_event_source_get_description(sd_event_source *s, const char **description) {
assert_return(s, -EINVAL);
assert_return(description, -EINVAL);
*description = s->description;
return 0;
}
-#endif // 0
_public_ sd_event *sd_event_source_get_event(sd_event_source *s) {
assert_return(s, NULL);
return s->event;
}
-/// UNNEEDED by elogind
-#if 0
_public_ int sd_event_source_get_pending(sd_event_source *s) {
assert_return(s, -EINVAL);
assert_return(s->type != SOURCE_EXIT, -EDOM);
return s->io.fd;
}
-#endif // 0
_public_ int sd_event_source_set_io_fd(sd_event_source *s, int fd) {
int r;
return 0;
}
-/// UNNEEDED by elogind
-#if 0
+_public_ int sd_event_source_get_io_fd_own(sd_event_source *s) {
+ assert_return(s, -EINVAL);
+ assert_return(s->type == SOURCE_IO, -EDOM);
+
+ return s->io.owned;
+}
+
+_public_ int sd_event_source_set_io_fd_own(sd_event_source *s, int own) {
+ assert_return(s, -EINVAL);
+ assert_return(s->type == SOURCE_IO, -EDOM);
+
+ s->io.owned = own;
+ return 0;
+}
+
_public_ int sd_event_source_get_io_events(sd_event_source *s, uint32_t* events) {
assert_return(s, -EINVAL);
assert_return(events, -EINVAL);
*events = s->io.events;
return 0;
}
-#endif // 0
_public_ int sd_event_source_set_io_events(sd_event_source *s, uint32_t events) {
int r;
if (s->io.events == events && !(events & EPOLLET))
return 0;
+ r = source_set_pending(s, false);
+ if (r < 0)
+ return r;
+
if (s->enabled != SD_EVENT_OFF) {
r = source_io_register(s, s->enabled, events);
if (r < 0)
}
s->io.events = events;
- source_set_pending(s, false);
return 0;
}
-/// UNNEEDED by elogind
-#if 0
_public_ int sd_event_source_get_io_revents(sd_event_source *s, uint32_t* revents) {
assert_return(s, -EINVAL);
assert_return(revents, -EINVAL);
assert_return(s, -EINVAL);
assert_return(!event_pid_changed(s->event), -ECHILD);
- return s->priority;
+ *priority = s->priority;
+ return 0;
}
-#endif // 0
_public_ int sd_event_source_set_priority(sd_event_source *s, int64_t priority) {
+ bool rm_inotify = false, rm_inode = false;
+ struct inotify_data *new_inotify_data = NULL;
+ struct inode_data *new_inode_data = NULL;
int r;
assert_return(s, -EINVAL);
if (s->priority == priority)
return 0;
- if (s->type == SOURCE_SIGNAL && s->enabled != SD_EVENT_OFF) {
+ if (s->type == SOURCE_INOTIFY) {
+ struct inode_data *old_inode_data;
+
+ assert(s->inotify.inode_data);
+ old_inode_data = s->inotify.inode_data;
+
+ /* We need the original fd to change the priority. If we don't have it we can't change the priority,
+ * anymore. Note that we close any fds when entering the next event loop iteration, i.e. for inotify
+ * events we allow priority changes only until the first following iteration. */
+ if (old_inode_data->fd < 0)
+ return -EOPNOTSUPP;
+
+ r = event_make_inotify_data(s->event, priority, &new_inotify_data);
+ if (r < 0)
+ return r;
+ rm_inotify = r > 0;
+
+ r = event_make_inode_data(s->event, new_inotify_data, old_inode_data->dev, old_inode_data->ino, &new_inode_data);
+ if (r < 0)
+ goto fail;
+ rm_inode = r > 0;
+
+ if (new_inode_data->fd < 0) {
+ /* Duplicate the fd for the new inode object if we don't have any yet */
+ new_inode_data->fd = fcntl(old_inode_data->fd, F_DUPFD_CLOEXEC, 3);
+ if (new_inode_data->fd < 0) {
+ r = -errno;
+ goto fail;
+ }
+
+ LIST_PREPEND(to_close, s->event->inode_data_to_close, new_inode_data);
+ }
+
+ /* Move the event source to the new inode data structure */
+ LIST_REMOVE(inotify.by_inode_data, old_inode_data->event_sources, s);
+ LIST_PREPEND(inotify.by_inode_data, new_inode_data->event_sources, s);
+ s->inotify.inode_data = new_inode_data;
+
+ /* Now create the new watch */
+ r = inode_data_realize_watch(s->event, new_inode_data);
+ if (r < 0) {
+ /* Move it back */
+ LIST_REMOVE(inotify.by_inode_data, new_inode_data->event_sources, s);
+ LIST_PREPEND(inotify.by_inode_data, old_inode_data->event_sources, s);
+ s->inotify.inode_data = old_inode_data;
+ goto fail;
+ }
+
+ s->priority = priority;
+
+ event_gc_inode_data(s->event, old_inode_data);
+
+ } else if (s->type == SOURCE_SIGNAL && s->enabled != SD_EVENT_OFF) {
struct signal_data *old, *d;
/* Move us from the signalfd belonging to the old
event_unmask_signal_data(s->event, old, s->signal.sig);
} else
- s->priority = priority;
+ s->priority = priority;
if (s->pending)
prioq_reshuffle(s->event->pending, s, &s->pending_index);
prioq_reshuffle(s->event->exit, s, &s->exit.prioq_index);
return 0;
+
+fail:
+ if (rm_inode)
+ event_free_inode_data(s->event, new_inode_data);
+
+ if (rm_inotify)
+ event_free_inotify_data(s->event, new_inotify_data);
+
+ return r;
}
-/// UNNEEDED by elogind
-#if 0
_public_ int sd_event_source_get_enabled(sd_event_source *s, int *m) {
assert_return(s, -EINVAL);
assert_return(m, -EINVAL);
*m = s->enabled;
return 0;
}
-#endif // 0
_public_ int sd_event_source_set_enabled(sd_event_source *s, int m) {
int r;
assert_return(s, -EINVAL);
- assert_return(m == SD_EVENT_OFF || m == SD_EVENT_ON || m == SD_EVENT_ONESHOT, -EINVAL);
+ assert_return(IN_SET(m, SD_EVENT_OFF, SD_EVENT_ON, SD_EVENT_ONESHOT), -EINVAL);
assert_return(!event_pid_changed(s->event), -ECHILD);
/* If we are dead anyway, we are fine with turning off
if (m == SD_EVENT_OFF) {
+ /* Unset the pending flag when this event source is disabled */
+ if (!IN_SET(s->type, SOURCE_DEFER, SOURCE_EXIT)) {
+ r = source_set_pending(s, false);
+ if (r < 0)
+ return r;
+ }
+
switch (s->type) {
case SOURCE_IO:
case SOURCE_DEFER:
case SOURCE_POST:
+ case SOURCE_INOTIFY:
s->enabled = m;
break;
}
} else {
+
+ /* Unset the pending flag when this event source is enabled */
+ if (s->enabled == SD_EVENT_OFF && !IN_SET(s->type, SOURCE_DEFER, SOURCE_EXIT)) {
+ r = source_set_pending(s, false);
+ if (r < 0)
+ return r;
+ }
+
switch (s->type) {
case SOURCE_IO:
s->enabled = m;
r = event_make_signal_data(s->event, s->signal.sig, NULL);
- if (r < 0) {
- s->enabled = SD_EVENT_OFF;
+ if (r < 0) {
+ s->enabled = SD_EVENT_OFF;
event_gc_signal_data(s->event, &s->priority, s->signal.sig);
- return r;
- }
+ return r;
+ }
break;
s->enabled = m;
r = event_make_signal_data(s->event, SIGCHLD, NULL);
- if (r < 0) {
- s->enabled = SD_EVENT_OFF;
+ if (r < 0) {
+ s->enabled = SD_EVENT_OFF;
s->event->n_enabled_child_sources--;
event_gc_signal_data(s->event, &s->priority, SIGCHLD);
- return r;
- }
+ return r;
+ }
break;
case SOURCE_DEFER:
case SOURCE_POST:
+ case SOURCE_INOTIFY:
s->enabled = m;
break;
_public_ int sd_event_source_set_time(sd_event_source *s, uint64_t usec) {
struct clock_data *d;
+ int r;
assert_return(s, -EINVAL);
- assert_return(usec != (uint64_t) -1, -EINVAL);
assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM);
assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(!event_pid_changed(s->event), -ECHILD);
- s->time.next = usec;
+ r = source_set_pending(s, false);
+ if (r < 0)
+ return r;
- source_set_pending(s, false);
+ s->time.next = usec;
d = event_get_clock_data(s->event, s->type);
assert(d);
return 0;
}
-/// UNNEEDED by elogind
-#if 0
_public_ int sd_event_source_get_time_accuracy(sd_event_source *s, uint64_t *usec) {
assert_return(s, -EINVAL);
assert_return(usec, -EINVAL);
_public_ int sd_event_source_set_time_accuracy(sd_event_source *s, uint64_t usec) {
struct clock_data *d;
+ int r;
assert_return(s, -EINVAL);
assert_return(usec != (uint64_t) -1, -EINVAL);
assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(!event_pid_changed(s->event), -ECHILD);
+ r = source_set_pending(s, false);
+ if (r < 0)
+ return r;
+
if (usec == 0)
usec = DEFAULT_ACCURACY_USEC;
s->time.accuracy = usec;
- source_set_pending(s, false);
-
d = event_get_clock_data(s->event, s->type);
assert(d);
*pid = s->child.pid;
return 0;
}
-#endif // 0
+
+_public_ int sd_event_source_get_inotify_mask(sd_event_source *s, uint32_t *mask) {
+ assert_return(s, -EINVAL);
+ assert_return(mask, -EINVAL);
+ assert_return(s->type == SOURCE_INOTIFY, -EDOM);
+ assert_return(!event_pid_changed(s->event), -ECHILD);
+
+ *mask = s->inotify.mask;
+ return 0;
+}
_public_ int sd_event_source_set_prepare(sd_event_source *s, sd_event_handler_t callback) {
int r;
return 0;
}
-/// UNNEEDED by elogind
-#if 0
_public_ void* sd_event_source_get_userdata(sd_event_source *s) {
assert_return(s, NULL);
return ret;
}
-#endif // 0
static usec_t sleep_between(sd_event *e, usec_t a, usec_t b) {
usec_t c;
if (a <= 0)
return 0;
+ if (a >= USEC_INFINITY)
+ return USEC_INFINITY;
if (b <= a + 1)
return a;
d->needs_rearm = false;
a = prioq_peek(d->earliest);
- if (!a || a->enabled == SD_EVENT_OFF) {
+ if (!a || a->enabled == SD_EVENT_OFF || a->time.next == USEC_INFINITY) {
if (d->fd < 0)
return 0;
b = prioq_peek(d->latest);
assert_se(b && b->enabled != SD_EVENT_OFF);
- t = sleep_between(e, a->time.next, b->time.next + b->time.accuracy);
+ t = sleep_between(e, a->time.next, time_event_source_latest(b));
if (d->next == t)
return 0;
ss = read(fd, &x, sizeof(x));
if (ss < 0) {
- if (errno == EAGAIN || errno == EINTR)
+ if (IN_SET(errno, EAGAIN, EINTR))
return 0;
return -errno;
return -errno;
if (s->child.siginfo.si_pid != 0) {
- bool zombie =
- s->child.siginfo.si_code == CLD_EXITED ||
- s->child.siginfo.si_code == CLD_KILLED ||
- s->child.siginfo.si_code == CLD_DUMPED;
+ bool zombie = IN_SET(s->child.siginfo.si_code, CLD_EXITED, CLD_KILLED, CLD_DUMPED);
if (!zombie && (s->child.options & WEXITED)) {
/* If the child isn't dead then let's
int r;
assert(e);
+ assert(d);
assert_return(events == EPOLLIN, -EIO);
/* If there's a signal queued on this priority and SIGCHLD is
n = read(d->fd, &si, sizeof(si));
if (n < 0) {
- if (errno == EAGAIN || errno == EINTR)
+ if (IN_SET(errno, EAGAIN, EINTR))
return read_one;
return -errno;
if (_unlikely_(n != sizeof(si)))
return -EIO;
- assert(si.ssi_signo < _NSIG);
+ assert(SIGNAL_VALID(si.ssi_signo));
read_one = true;
}
}
+static int event_inotify_data_read(sd_event *e, struct inotify_data *d, uint32_t revents) {
+ ssize_t n;
+
+ assert(e);
+ assert(d);
+
+ assert_return(revents == EPOLLIN, -EIO);
+
+ /* If there's already an event source pending for this priority, don't read another */
+ if (d->n_pending > 0)
+ return 0;
+
+ /* Is the read buffer non-empty? If so, let's not read more */
+ if (d->buffer_filled > 0)
+ return 0;
+
+ n = read(d->fd, &d->buffer, sizeof(d->buffer));
+ if (n < 0) {
+ if (IN_SET(errno, EAGAIN, EINTR))
+ return 0;
+
+ return -errno;
+ }
+
+ assert(n > 0);
+ d->buffer_filled = (size_t) n;
+ LIST_PREPEND(buffered, e->inotify_data_buffered, d);
+
+ return 1;
+}
+
+static void event_inotify_data_drop(sd_event *e, struct inotify_data *d, size_t sz) {
+ assert(e);
+ assert(d);
+ assert(sz <= d->buffer_filled);
+
+ if (sz == 0)
+ return;
+
+ /* Move the rest to the buffer to the front, in order to get things properly aligned again */
+ memmove(d->buffer.raw, d->buffer.raw + sz, d->buffer_filled - sz);
+ d->buffer_filled -= sz;
+
+ if (d->buffer_filled == 0)
+ LIST_REMOVE(buffered, e->inotify_data_buffered, d);
+}
+
+static int event_inotify_data_process(sd_event *e, struct inotify_data *d) {
+ int r;
+
+ assert(e);
+ assert(d);
+
+ /* If there's already an event source pending for this priority, don't read another */
+ if (d->n_pending > 0)
+ return 0;
+
+ while (d->buffer_filled > 0) {
+ size_t sz;
+
+ /* Let's validate that the event structures are complete */
+ if (d->buffer_filled < offsetof(struct inotify_event, name))
+ return -EIO;
+
+ sz = offsetof(struct inotify_event, name) + d->buffer.ev.len;
+ if (d->buffer_filled < sz)
+ return -EIO;
+
+ if (d->buffer.ev.mask & IN_Q_OVERFLOW) {
+ struct inode_data *inode_data;
+ Iterator i;
+
+ /* The queue overran, let's pass this event to all event sources connected to this inotify
+ * object */
+
+ HASHMAP_FOREACH(inode_data, d->inodes, i) {
+ sd_event_source *s;
+
+ LIST_FOREACH(inotify.by_inode_data, s, inode_data->event_sources) {
+
+ if (s->enabled == SD_EVENT_OFF)
+ continue;
+
+ r = source_set_pending(s, true);
+ if (r < 0)
+ return r;
+ }
+ }
+ } else {
+ struct inode_data *inode_data;
+ sd_event_source *s;
+
+ /* Find the inode object for this watch descriptor. If IN_IGNORED is set we also remove it from
+ * our watch descriptor table. */
+ if (d->buffer.ev.mask & IN_IGNORED) {
+
+ inode_data = hashmap_remove(d->wd, INT_TO_PTR(d->buffer.ev.wd));
+ if (!inode_data) {
+ event_inotify_data_drop(e, d, sz);
+ continue;
+ }
+
+ /* The watch descriptor was removed by the kernel, let's drop it here too */
+ inode_data->wd = -1;
+ } else {
+ inode_data = hashmap_get(d->wd, INT_TO_PTR(d->buffer.ev.wd));
+ if (!inode_data) {
+ event_inotify_data_drop(e, d, sz);
+ continue;
+ }
+ }
+
+ /* Trigger all event sources that are interested in these events. Also trigger all event
+ * sources if IN_IGNORED or IN_UNMOUNT is set. */
+ LIST_FOREACH(inotify.by_inode_data, s, inode_data->event_sources) {
+
+ if (s->enabled == SD_EVENT_OFF)
+ continue;
+
+ if ((d->buffer.ev.mask & (IN_IGNORED|IN_UNMOUNT)) == 0 &&
+ (s->inotify.mask & d->buffer.ev.mask & IN_ALL_EVENTS) == 0)
+ continue;
+
+ r = source_set_pending(s, true);
+ if (r < 0)
+ return r;
+ }
+ }
+
+ /* Something pending now? If so, let's finish, otherwise let's read more. */
+ if (d->n_pending > 0)
+ return 1;
+ }
+
+ return 0;
+}
+
+static int process_inotify(sd_event *e) {
+ struct inotify_data *d;
+ int r, done = 0;
+
+ assert(e);
+
+ LIST_FOREACH(buffered, d, e->inotify_data_buffered) {
+ r = event_inotify_data_process(e, d);
+ if (r < 0)
+ return r;
+ if (r > 0)
+ done ++;
+ }
+
+ return done;
+}
+
static int source_dispatch(sd_event_source *s) {
+ EventSourceType saved_type;
int r = 0;
assert(s);
assert(s->pending || s->type == SOURCE_EXIT);
- if (s->type != SOURCE_DEFER && s->type != SOURCE_EXIT) {
+ /* Save the event source type, here, so that we still know it after the event callback which might invalidate
+ * the event. */
+ saved_type = s->type;
+
+ if (!IN_SET(s->type, SOURCE_DEFER, SOURCE_EXIT)) {
r = source_set_pending(s, false);
if (r < 0)
return r;
case SOURCE_CHILD: {
bool zombie;
- zombie = s->child.siginfo.si_code == CLD_EXITED ||
- s->child.siginfo.si_code == CLD_KILLED ||
- s->child.siginfo.si_code == CLD_DUMPED;
+ zombie = IN_SET(s->child.siginfo.si_code, CLD_EXITED, CLD_KILLED, CLD_DUMPED);
r = s->child.callback(s, &s->child.siginfo, s->userdata);
/* Now, reap the PID for good. */
if (zombie)
- waitid(P_PID, s->child.pid, &s->child.siginfo, WNOHANG|WEXITED);
+ (void) waitid(P_PID, s->child.pid, &s->child.siginfo, WNOHANG|WEXITED);
break;
}
r = s->exit.callback(s, s->userdata);
break;
+ case SOURCE_INOTIFY: {
+ struct sd_event *e = s->event;
+ struct inotify_data *d;
+ size_t sz;
+
+ assert(s->inotify.inode_data);
+ assert_se(d = s->inotify.inode_data->inotify_data);
+
+ assert(d->buffer_filled >= offsetof(struct inotify_event, name));
+ sz = offsetof(struct inotify_event, name) + d->buffer.ev.len;
+ assert(d->buffer_filled >= sz);
+
+ r = s->inotify.callback(s, &d->buffer.ev, s->userdata);
+
+ /* When no event is pending anymore on this inotify object, then let's drop the event from the
+ * buffer. */
+ if (d->n_pending == 0)
+ event_inotify_data_drop(e, d, sz);
+
+ break;
+ }
+
case SOURCE_WATCHDOG:
case _SOURCE_EVENT_SOURCE_TYPE_MAX:
case _SOURCE_EVENT_SOURCE_TYPE_INVALID:
s->dispatching = false;
- if (r < 0) {
- if (s->description)
- log_debug_errno(r, "Event source '%s' returned error, disabling: %m", s->description);
- else
- log_debug_errno(r, "Event source %p returned error, disabling: %m", s);
- }
+ if (r < 0)
+ log_debug_errno(r, "Event source %s (type %s) returned error, disabling: %m",
+ strna(s->description), event_source_type_to_string(saved_type));
if (s->n_ref == 0)
source_free(s);
r = s->prepare(s, s->userdata);
s->dispatching = false;
- if (r < 0) {
- if (s->description)
- log_debug_errno(r, "Prepare callback of event source '%s' returned error, disabling: %m", s->description);
- else
- log_debug_errno(r, "Prepare callback of event source %p returned error, disabling: %m", s);
- }
+ if (r < 0)
+ log_debug_errno(r, "Prepare callback of event source %s (type %s) returned error, disabling: %m",
+ strna(s->description), event_source_type_to_string(s->type));
if (s->n_ref == 0)
source_free(s);
static int dispatch_exit(sd_event *e) {
sd_event_source *p;
+ _cleanup_(sd_event_unrefp) sd_event *ref = NULL;
int r;
assert(e);
return 0;
}
- sd_event_ref(e);
+ ref = sd_event_ref(e);
e->iteration++;
e->state = SD_EVENT_EXITING;
-
r = source_dispatch(p);
-
e->state = SD_EVENT_INITIAL;
- sd_event_unref(e);
-
return r;
}
return arm_watchdog(e);
}
+static void event_close_inode_data_fds(sd_event *e) {
+ struct inode_data *d;
+
+ assert(e);
+
+ /* Close the fds pointing to the inodes to watch now. We need to close them as they might otherwise pin
+ * filesystems. But we can't close them right-away as we need them as long as the user still wants to make
+ * adjustments to the even source, such as changing the priority (which requires us to remove and readd a watch
+ * for the inode). Hence, let's close them when entering the first iteration after they were added, as a
+ * compromise. */
+
+ while ((d = e->inode_data_to_close)) {
+ assert(d->fd >= 0);
+ d->fd = safe_close(d->fd);
+
+ LIST_REMOVE(to_close, e->inode_data_to_close, d);
+ }
+}
+
_public_ int sd_event_prepare(sd_event *e) {
int r;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(!event_pid_changed(e), -ECHILD);
assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(e->state == SD_EVENT_INITIAL, -EBUSY);
e->iteration++;
+ e->state = SD_EVENT_PREPARING;
r = event_prepare(e);
+ e->state = SD_EVENT_INITIAL;
if (r < 0)
return r;
if (r < 0)
return r;
+ event_close_inode_data_fds(e);
+
if (event_next_pending(e) || e->need_process_child)
goto pending;
int r, m, i;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(!event_pid_changed(e), -ECHILD);
assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(e->state == SD_EVENT_ARMED, -EBUSY);
ev_queue_max = MAX(e->n_sources, 1u);
ev_queue = newa(struct epoll_event, ev_queue_max);
+ /* If we still have inotify data buffered, then query the other fds, but don't wait on it */
+ if (e->inotify_data_buffered)
+ timeout = 0;
+
m = epoll_wait(e->epoll_fd, ev_queue, ev_queue_max,
timeout == (uint64_t) -1 ? -1 : (int) ((timeout + USEC_PER_MSEC - 1) / USEC_PER_MSEC));
if (m < 0) {
goto finish;
}
- dual_timestamp_get(&e->timestamp);
- e->timestamp_boottime = now(CLOCK_BOOTTIME);
+ triple_timestamp_get(&e->timestamp);
for (i = 0; i < m; i++) {
switch (*t) {
case WAKEUP_EVENT_SOURCE:
- r = process_io(e, ev_queue[i].data.ptr, ev_queue[i].events);
+ r = process_io(e, ev_queue[i].data.ptr, ev_queue[i].events);
break;
case WAKEUP_CLOCK_DATA: {
r = process_signal(e, ev_queue[i].data.ptr, ev_queue[i].events);
break;
+ case WAKEUP_INOTIFY_DATA:
+ r = event_inotify_data_read(e, ev_queue[i].data.ptr, ev_queue[i].events);
+ break;
+
default:
assert_not_reached("Invalid wake-up pointer");
}
if (r < 0)
goto finish;
- r = process_timer(e, e->timestamp_boottime, &e->boottime);
+ r = process_timer(e, e->timestamp.boottime, &e->boottime);
if (r < 0)
goto finish;
if (r < 0)
goto finish;
- r = process_timer(e, e->timestamp_boottime, &e->boottime_alarm);
+ r = process_timer(e, e->timestamp.boottime, &e->boottime_alarm);
if (r < 0)
goto finish;
goto finish;
}
+ r = process_inotify(e);
+ if (r < 0)
+ goto finish;
+
if (event_next_pending(e)) {
e->state = SD_EVENT_PENDING;
int r;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(!event_pid_changed(e), -ECHILD);
assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(e->state == SD_EVENT_PENDING, -EBUSY);
p = event_next_pending(e);
if (p) {
- sd_event_ref(e);
+ _cleanup_(sd_event_unrefp) sd_event *ref = NULL;
+ ref = sd_event_ref(e);
e->state = SD_EVENT_RUNNING;
r = source_dispatch(p);
e->state = SD_EVENT_INITIAL;
-
- sd_event_unref(e);
-
return r;
}
return 1;
}
+static void event_log_delays(sd_event *e) {
+ char b[ELEMENTSOF(e->delays) * DECIMAL_STR_MAX(unsigned) + 1];
+ unsigned i;
+ int o;
+
+ for (i = o = 0; i < ELEMENTSOF(e->delays); i++) {
+ o += snprintf(&b[o], sizeof(b) - o, "%u ", e->delays[i]);
+ e->delays[i] = 0;
+ }
+ log_debug("Event loop iterations: %.*s", o, b);
+}
+
_public_ int sd_event_run(sd_event *e, uint64_t timeout) {
int r;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(!event_pid_changed(e), -ECHILD);
assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(e->state == SD_EVENT_INITIAL, -EBUSY);
+ if (e->profile_delays && e->last_run) {
+ usec_t this_run;
+ unsigned l;
+
+ this_run = now(CLOCK_MONOTONIC);
+
+ l = u64log2(this_run - e->last_run);
+ assert(l < sizeof(e->delays));
+ e->delays[l]++;
+
+ if (this_run - e->last_log >= 5*USEC_PER_SEC) {
+ event_log_delays(e);
+ e->last_log = this_run;
+ }
+ }
+
r = sd_event_prepare(e);
if (r == 0)
/* There was nothing? Then wait... */
r = sd_event_wait(e, timeout);
+ if (e->profile_delays)
+ e->last_run = now(CLOCK_MONOTONIC);
+
if (r > 0) {
/* There's something now, then let's dispatch it */
r = sd_event_dispatch(e);
return r;
}
-/// UNNEEDED by elogind
-#if 0
_public_ int sd_event_loop(sd_event *e) {
+ _cleanup_(sd_event_unrefp) sd_event *ref = NULL;
int r;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(!event_pid_changed(e), -ECHILD);
assert_return(e->state == SD_EVENT_INITIAL, -EBUSY);
- sd_event_ref(e);
+ ref = sd_event_ref(e);
while (e->state != SD_EVENT_FINISHED) {
r = sd_event_run(e, (uint64_t) -1);
if (r < 0)
- goto finish;
+ return r;
}
- r = e->exit_code;
-
-finish:
- sd_event_unref(e);
- return r;
+ return e->exit_code;
}
_public_ int sd_event_get_fd(sd_event *e) {
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(!event_pid_changed(e), -ECHILD);
return e->epoll_fd;
}
-#endif // 0
_public_ int sd_event_get_state(sd_event *e) {
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(!event_pid_changed(e), -ECHILD);
return e->state;
_public_ int sd_event_get_exit_code(sd_event *e, int *code) {
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(code, -EINVAL);
assert_return(!event_pid_changed(e), -ECHILD);
_public_ int sd_event_exit(sd_event *e, int code) {
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(e->state != SD_EVENT_FINISHED, -ESTALE);
assert_return(!event_pid_changed(e), -ECHILD);
return 0;
}
-/// UNNEEDED by elogind
-#if 0
_public_ int sd_event_now(sd_event *e, clockid_t clock, uint64_t *usec) {
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(usec, -EINVAL);
assert_return(!event_pid_changed(e), -ECHILD);
- if (!dual_timestamp_is_set(&e->timestamp)) {
+ if (!TRIPLE_TIMESTAMP_HAS_CLOCK(clock))
+ return -EOPNOTSUPP;
+
+ /* Generate a clean error in case CLOCK_BOOTTIME is not available. Note that don't use clock_supported() here,
+ * for a reason: there are systems where CLOCK_BOOTTIME is supported, but CLOCK_BOOTTIME_ALARM is not, but for
+ * the purpose of getting the time this doesn't matter. */
+ if (IN_SET(clock, CLOCK_BOOTTIME, CLOCK_BOOTTIME_ALARM) && !clock_boottime_supported())
+ return -EOPNOTSUPP;
+
+ if (!triple_timestamp_is_set(&e->timestamp)) {
/* Implicitly fall back to now() if we never ran
* before and thus have no cached time. */
*usec = now(clock);
return 1;
}
- switch (clock) {
-
- case CLOCK_REALTIME:
- case CLOCK_REALTIME_ALARM:
- *usec = e->timestamp.realtime;
- break;
-
- case CLOCK_MONOTONIC:
- *usec = e->timestamp.monotonic;
- break;
-
- case CLOCK_BOOTTIME:
- case CLOCK_BOOTTIME_ALARM:
- *usec = e->timestamp_boottime;
- break;
- }
-
+ *usec = triple_timestamp_by_clock(&e->timestamp, clock);
return 0;
}
-#endif // 0
_public_ int sd_event_default(sd_event **ret) {
-
- static thread_local sd_event *default_event = NULL;
sd_event *e = NULL;
int r;
_public_ int sd_event_get_tid(sd_event *e, pid_t *tid) {
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(tid, -EINVAL);
assert_return(!event_pid_changed(e), -ECHILD);
int r;
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(!event_pid_changed(e), -ECHILD);
if (e->watchdog == !!b)
return e->watchdog;
if (b) {
- struct epoll_event ev = {};
+ struct epoll_event ev;
r = sd_watchdog_enabled(false, &e->watchdog_period);
if (r <= 0)
if (r < 0)
goto fail;
- ev.events = EPOLLIN;
- ev.data.ptr = INT_TO_PTR(SOURCE_WATCHDOG);
+ ev = (struct epoll_event) {
+ .events = EPOLLIN,
+ .data.ptr = INT_TO_PTR(SOURCE_WATCHDOG),
+ };
r = epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, e->watchdog_fd, &ev);
if (r < 0) {
return r;
}
-/// UNNEEDED by elogind
-#if 0
_public_ int sd_event_get_watchdog(sd_event *e) {
assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
assert_return(!event_pid_changed(e), -ECHILD);
return e->watchdog;
}
-#endif // 0
+
+_public_ int sd_event_get_iteration(sd_event *e, uint64_t *ret) {
+ assert_return(e, -EINVAL);
+ assert_return(e = event_resolve(e), -ENOPKG);
+ assert_return(!event_pid_changed(e), -ECHILD);
+
+ *ret = e->iteration;
+ return 0;
+}
+
+_public_ int sd_event_source_set_destroy_callback(sd_event_source *s, sd_event_destroy_t callback) {
+ assert_return(s, -EINVAL);
+
+ s->destroy_callback = callback;
+ return 0;
+}
+
+_public_ int sd_event_source_get_destroy_callback(sd_event_source *s, sd_event_destroy_t *ret) {
+ assert_return(s, -EINVAL);
+
+ if (ret)
+ *ret = s->destroy_callback;
+
+ return !!s->destroy_callback;
+}
~ianmdlvl / elogind.git / blobdiff