From 2cb7e0aaedad73b076ea18c6900b0e86c5760d79 Mon Sep 17 00:00:00 2001 From: Daniel Baumann Date: Sat, 27 Apr 2024 15:00:47 +0200 Subject: Adding upstream version 247.3. Signed-off-by: Daniel Baumann --- src/libsystemd/sd-event/event-source.h | 213 ++ src/libsystemd/sd-event/event-util.c | 99 + src/libsystemd/sd-event/event-util.h | 13 + src/libsystemd/sd-event/sd-event.c | 4010 ++++++++++++++++++++++++++++++++ src/libsystemd/sd-event/test-event.c | 607 +++++ 5 files changed, 4942 insertions(+) create mode 100644 src/libsystemd/sd-event/event-source.h create mode 100644 src/libsystemd/sd-event/event-util.c create mode 100644 src/libsystemd/sd-event/event-util.h create mode 100644 src/libsystemd/sd-event/sd-event.c create mode 100644 src/libsystemd/sd-event/test-event.c (limited to 'src/libsystemd/sd-event') diff --git a/src/libsystemd/sd-event/event-source.h b/src/libsystemd/sd-event/event-source.h new file mode 100644 index 0000000..62d0718 --- /dev/null +++ b/src/libsystemd/sd-event/event-source.h @@ -0,0 +1,213 @@ +#pragma once +/* SPDX-License-Identifier: LGPL-2.1-or-later */ + +#include +#include +#include + +#include "sd-event.h" + +#include "fs-util.h" +#include "hashmap.h" +#include "list.h" +#include "prioq.h" + +typedef enum EventSourceType { + SOURCE_IO, + SOURCE_TIME_REALTIME, + SOURCE_TIME_BOOTTIME, + SOURCE_TIME_MONOTONIC, + SOURCE_TIME_REALTIME_ALARM, + SOURCE_TIME_BOOTTIME_ALARM, + SOURCE_SIGNAL, + SOURCE_CHILD, + SOURCE_DEFER, + SOURCE_POST, + SOURCE_EXIT, + SOURCE_WATCHDOG, + SOURCE_INOTIFY, + _SOURCE_EVENT_SOURCE_TYPE_MAX, + _SOURCE_EVENT_SOURCE_TYPE_INVALID = -1 +} EventSourceType; + +/* All objects we use in epoll events start with this value, so that + * we know how to dispatch it */ +typedef enum WakeupType { + WAKEUP_NONE, + WAKEUP_EVENT_SOURCE, /* either I/O or pidfd wakeup */ + WAKEUP_CLOCK_DATA, + WAKEUP_SIGNAL_DATA, + WAKEUP_INOTIFY_DATA, + _WAKEUP_TYPE_MAX, + _WAKEUP_TYPE_INVALID = -1, +} WakeupType; + +struct inode_data; + +struct sd_event_source { + WakeupType wakeup; + + unsigned n_ref; + + sd_event *event; + void *userdata; + sd_event_handler_t prepare; + + char *description; + + EventSourceType type:5; + signed int enabled:3; + bool pending:1; + bool dispatching:1; + bool floating:1; + bool exit_on_failure:1; + + int64_t priority; + unsigned pending_index; + unsigned prepare_index; + uint64_t pending_iteration; + uint64_t prepare_iteration; + + sd_event_destroy_t destroy_callback; + + LIST_FIELDS(sd_event_source, sources); + + union { + struct { + sd_event_io_handler_t callback; + int fd; + uint32_t events; + uint32_t revents; + bool registered:1; + bool owned:1; + } io; + struct { + sd_event_time_handler_t callback; + usec_t next, accuracy; + unsigned earliest_index; + unsigned latest_index; + } time; + struct { + sd_event_signal_handler_t callback; + struct signalfd_siginfo siginfo; + int sig; + } signal; + struct { + sd_event_child_handler_t callback; + siginfo_t siginfo; + pid_t pid; + int options; + int pidfd; + bool registered:1; /* whether the pidfd is registered in the epoll */ + bool pidfd_owned:1; /* close pidfd when event source is freed */ + bool process_owned:1; /* kill+reap process when event source is freed */ + bool exited:1; /* true if process exited (i.e. if there's value in SIGKILLing it if we want to get rid of it) */ + bool waited:1; /* true if process was waited for (i.e. if there's value in waitid(P_PID)'ing it if we want to get rid of it) */ + } child; + struct { + sd_event_handler_t callback; + } defer; + struct { + sd_event_handler_t callback; + } post; + struct { + 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; + }; +}; + +struct clock_data { + WakeupType wakeup; + int fd; + + /* For all clocks we maintain two priority queues each, one + * ordered for the earliest times the events may be + * dispatched, and one ordered by the latest times they must + * have been dispatched. The range between the top entries in + * the two prioqs is the time window we can freely schedule + * wakeups in */ + + Prioq *earliest; + Prioq *latest; + usec_t next; + + bool needs_rearm:1; +}; + +struct signal_data { + WakeupType wakeup; + + /* For each priority we maintain one signal fd, so that we + * only have to dequeue a single event per priority at a + * time. */ + + int fd; + int64_t priority; + sigset_t sigset; + 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); +}; diff --git a/src/libsystemd/sd-event/event-util.c b/src/libsystemd/sd-event/event-util.c new file mode 100644 index 0000000..132796f --- /dev/null +++ b/src/libsystemd/sd-event/event-util.c @@ -0,0 +1,99 @@ +/* SPDX-License-Identifier: LGPL-2.1-or-later */ + +#include + +#include "event-source.h" +#include "event-util.h" +#include "log.h" +#include "string-util.h" + +int event_reset_time( + sd_event *e, + sd_event_source **s, + clockid_t clock, + uint64_t usec, + uint64_t accuracy, + sd_event_time_handler_t callback, + void *userdata, + int64_t priority, + const char *description, + bool force_reset) { + + bool created = false; + int enabled, r; + clockid_t c; + + assert(e); + assert(s); + + if (*s) { + if (!force_reset) { + r = sd_event_source_get_enabled(*s, &enabled); + if (r < 0) + return log_debug_errno(r, "sd-event: Failed to query whether event source \"%s\" is enabled or not: %m", + strna((*s)->description ?: description)); + + if (enabled != SD_EVENT_OFF) + return 0; + } + + r = sd_event_source_get_time_clock(*s, &c); + if (r < 0) + return log_debug_errno(r, "sd-event: Failed to get clock id of event source \"%s\": %m", strna((*s)->description ?: description)); + + if (c != clock) + return log_debug_errno(SYNTHETIC_ERRNO(EINVAL), + "sd-event: Current clock id %i of event source \"%s\" is different from specified one %i.", + (int)c, + strna((*s)->description ? : description), + (int)clock); + + r = sd_event_source_set_time(*s, usec); + if (r < 0) + return log_debug_errno(r, "sd-event: Failed to set time for event source \"%s\": %m", strna((*s)->description ?: description)); + + r = sd_event_source_set_time_accuracy(*s, accuracy); + if (r < 0) + return log_debug_errno(r, "sd-event: Failed to set accuracy for event source \"%s\": %m", strna((*s)->description ?: description)); + + /* callback function is not updated, as we do not have sd_event_source_set_time_callback(). */ + + (void) sd_event_source_set_userdata(*s, userdata); + + r = sd_event_source_set_enabled(*s, SD_EVENT_ONESHOT); + if (r < 0) + return log_debug_errno(r, "sd-event: Failed to enable event source \"%s\": %m", strna((*s)->description ?: description)); + } else { + r = sd_event_add_time(e, s, clock, usec, accuracy, callback, userdata); + if (r < 0) + return log_debug_errno(r, "sd-event: Failed to create timer event \"%s\": %m", strna(description)); + + created = true; + } + + r = sd_event_source_set_priority(*s, priority); + if (r < 0) + return log_debug_errno(r, "sd-event: Failed to set priority for event source \"%s\": %m", strna((*s)->description ?: description)); + + if (description) { + r = sd_event_source_set_description(*s, description); + if (r < 0) + return log_debug_errno(r, "sd-event: Failed to set description for event source \"%s\": %m", description); + } + + return created; +} + +int event_source_disable(sd_event_source *s) { + if (!s) + return 0; + + return sd_event_source_set_enabled(s, SD_EVENT_OFF); +} + +int event_source_is_enabled(sd_event_source *s) { + if (!s) + return false; + + return sd_event_source_get_enabled(s, NULL); +} diff --git a/src/libsystemd/sd-event/event-util.h b/src/libsystemd/sd-event/event-util.h new file mode 100644 index 0000000..c8f97bc --- /dev/null +++ b/src/libsystemd/sd-event/event-util.h @@ -0,0 +1,13 @@ +/* SPDX-License-Identifier: LGPL-2.1-or-later */ +#pragma once + +#include + +#include "sd-event.h" + +int event_reset_time(sd_event *e, sd_event_source **s, + clockid_t clock, uint64_t usec, uint64_t accuracy, + sd_event_time_handler_t callback, void *userdata, + int64_t priority, const char *description, bool force_reset); +int event_source_disable(sd_event_source *s); +int event_source_is_enabled(sd_event_source *s); diff --git a/src/libsystemd/sd-event/sd-event.c b/src/libsystemd/sd-event/sd-event.c new file mode 100644 index 0000000..789a8c7 --- /dev/null +++ b/src/libsystemd/sd-event/sd-event.c @@ -0,0 +1,4010 @@ +/* SPDX-License-Identifier: LGPL-2.1-or-later */ + +#include +#include +#include + +#include "sd-daemon.h" +#include "sd-event.h" +#include "sd-id128.h" + +#include "alloc-util.h" +#include "env-util.h" +#include "event-source.h" +#include "fd-util.h" +#include "fs-util.h" +#include "hashmap.h" +#include "list.h" +#include "macro.h" +#include "memory-util.h" +#include "missing_syscall.h" +#include "prioq.h" +#include "process-util.h" +#include "set.h" +#include "signal-util.h" +#include "string-table.h" +#include "string-util.h" +#include "strxcpyx.h" +#include "time-util.h" + +#define DEFAULT_ACCURACY_USEC (250 * USEC_PER_MSEC) + +static bool EVENT_SOURCE_WATCH_PIDFD(sd_event_source *s) { + /* Returns true if this is a PID event source and can be implemented by watching EPOLLIN */ + return s && + s->type == SOURCE_CHILD && + s->child.pidfd >= 0 && + s->child.options == WEXITED; +} + +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); + +#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 sd_event { + unsigned n_ref; + + int epoll_fd; + int watchdog_fd; + + Prioq *pending; + Prioq *prepare; + + /* timerfd_create() only supports these five clocks so far. We + * can add support for more clocks when the kernel learns to + * deal with them, too. */ + struct clock_data realtime; + struct clock_data boottime; + struct clock_data monotonic; + struct clock_data realtime_alarm; + struct clock_data boottime_alarm; + + usec_t perturb; + + sd_event_source **signal_sources; /* indexed by signal number */ + Hashmap *signal_data; /* indexed by priority */ + + Hashmap *child_sources; + unsigned n_enabled_child_sources; + + Set *post_sources; + + 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; + + 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; + + pid_t tid; + sd_event **default_event_ptr; + + usec_t watchdog_last, watchdog_period; + + unsigned n_sources; + + struct epoll_event *event_queue; + size_t event_queue_allocated; + + 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; + int r; + + assert(x->pending); + assert(y->pending); + + /* 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 */ + r = CMP(x->priority, y->priority); + if (r != 0) + return r; + + /* Older entries first */ + return CMP(x->pending_iteration, y->pending_iteration); +} + +static int prepare_prioq_compare(const void *a, const void *b) { + const sd_event_source *x = a, *y = b; + int r; + + 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 */ + r = CMP(x->prepare_iteration, y->prepare_iteration); + if (r != 0) + return r; + + /* Lower priority values first */ + return CMP(x->priority, y->priority); +} + +static int earliest_time_prioq_compare(const void *a, const void *b) { + const sd_event_source *x = a, *y = b; + + assert(EVENT_SOURCE_IS_TIME(x->type)); + assert(x->type == y->type); + + /* 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 the pending ones to the end */ + if (!x->pending && y->pending) + return -1; + if (x->pending && !y->pending) + return 1; + + /* Order by time */ + return CMP(x->time.next, y->time.next); +} + +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; + + assert(EVENT_SOURCE_IS_TIME(x->type)); + assert(x->type == y->type); + + /* 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 the pending ones to the end */ + if (!x->pending && y->pending) + return -1; + if (x->pending && !y->pending) + return 1; + + /* Order by time */ + return CMP(time_event_source_latest(x), time_event_source_latest(y)); +} + +static int exit_prioq_compare(const void *a, const void *b) { + const sd_event_source *x = a, *y = b; + + assert(x->type == SOURCE_EXIT); + assert(y->type == SOURCE_EXIT); + + /* 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 */ + return CMP(x->priority, y->priority); +} + +static void free_clock_data(struct clock_data *d) { + assert(d); + assert(d->wakeup == WAKEUP_CLOCK_DATA); + + safe_close(d->fd); + prioq_free(d->earliest); + prioq_free(d->latest); +} + +static sd_event *event_free(sd_event *e) { + sd_event_source *s; + + assert(e); + + while ((s = e->sources)) { + assert(s->floating); + source_disconnect(s); + sd_event_source_unref(s); + } + + assert(e->n_sources == 0); + + if (e->default_event_ptr) + *(e->default_event_ptr) = NULL; + + safe_close(e->epoll_fd); + safe_close(e->watchdog_fd); + + free_clock_data(&e->realtime); + free_clock_data(&e->boottime); + free_clock_data(&e->monotonic); + free_clock_data(&e->realtime_alarm); + free_clock_data(&e->boottime_alarm); + + prioq_free(e->pending); + prioq_free(e->prepare); + prioq_free(e->exit); + + 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->event_queue); + + return mfree(e); +} + +_public_ int sd_event_new(sd_event** ret) { + sd_event *e; + int r; + + assert_return(ret, -EINVAL); + + e = new(sd_event, 1); + if (!e) + return -ENOMEM; + + *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(), + }; + + 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) { + r = -errno; + 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; + +fail: + event_free(e); + return r; +} + +DEFINE_PUBLIC_TRIVIAL_REF_UNREF_FUNC(sd_event, sd_event, event_free); + +_public_ sd_event_source* sd_event_source_disable_unref(sd_event_source *s) { + if (s) + (void) sd_event_source_set_enabled(s, SD_EVENT_OFF); + return sd_event_source_unref(s); +} + +static bool event_pid_changed(sd_event *e) { + assert(e); + + /* We don't support people creating an event loop and keeping + * it around over a fork(). Let's complain. */ + + return e->original_pid != getpid_cached(); +} + +static void source_io_unregister(sd_event_source *s) { + assert(s); + assert(s->type == SOURCE_IO); + + if (event_pid_changed(s->event)) + return; + + if (!s->io.registered) + return; + + if (epoll_ctl(s->event->epoll_fd, EPOLL_CTL_DEL, s->io.fd, NULL) < 0) + 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; +} + +static int source_io_register( + sd_event_source *s, + int enabled, + uint32_t events) { + + assert(s); + assert(s->type == SOURCE_IO); + assert(enabled != SD_EVENT_OFF); + + struct epoll_event ev = { + .events = events | (enabled == SD_EVENT_ONESHOT ? EPOLLONESHOT : 0), + .data.ptr = s, + }; + + if (epoll_ctl(s->event->epoll_fd, + s->io.registered ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, + s->io.fd, &ev) < 0) + return -errno; + + s->io.registered = true; + + return 0; +} + +static void source_child_pidfd_unregister(sd_event_source *s) { + assert(s); + assert(s->type == SOURCE_CHILD); + + if (event_pid_changed(s->event)) + return; + + if (!s->child.registered) + return; + + if (EVENT_SOURCE_WATCH_PIDFD(s)) + if (epoll_ctl(s->event->epoll_fd, EPOLL_CTL_DEL, s->child.pidfd, NULL) < 0) + 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->child.registered = false; +} + +static int source_child_pidfd_register(sd_event_source *s, int enabled) { + assert(s); + assert(s->type == SOURCE_CHILD); + assert(enabled != SD_EVENT_OFF); + + if (EVENT_SOURCE_WATCH_PIDFD(s)) { + struct epoll_event ev = { + .events = EPOLLIN | (enabled == SD_EVENT_ONESHOT ? EPOLLONESHOT : 0), + .data.ptr = s, + }; + + if (epoll_ctl(s->event->epoll_fd, + s->child.registered ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, + s->child.pidfd, &ev) < 0) + return -errno; + } + + s->child.registered = true; + return 0; +} + +static clockid_t event_source_type_to_clock(EventSourceType t) { + + switch (t) { + + case SOURCE_TIME_REALTIME: + return CLOCK_REALTIME; + + case SOURCE_TIME_BOOTTIME: + return CLOCK_BOOTTIME; + + case SOURCE_TIME_MONOTONIC: + return CLOCK_MONOTONIC; + + case SOURCE_TIME_REALTIME_ALARM: + return CLOCK_REALTIME_ALARM; + + case SOURCE_TIME_BOOTTIME_ALARM: + return CLOCK_BOOTTIME_ALARM; + + default: + return (clockid_t) -1; + } +} + +static EventSourceType clock_to_event_source_type(clockid_t clock) { + + switch (clock) { + + case CLOCK_REALTIME: + return SOURCE_TIME_REALTIME; + + case CLOCK_BOOTTIME: + return SOURCE_TIME_BOOTTIME; + + case CLOCK_MONOTONIC: + return SOURCE_TIME_MONOTONIC; + + case CLOCK_REALTIME_ALARM: + return SOURCE_TIME_REALTIME_ALARM; + + case CLOCK_BOOTTIME_ALARM: + return SOURCE_TIME_BOOTTIME_ALARM; + + default: + return _SOURCE_EVENT_SOURCE_TYPE_INVALID; + } +} + +static struct clock_data* event_get_clock_data(sd_event *e, EventSourceType t) { + assert(e); + + switch (t) { + + case SOURCE_TIME_REALTIME: + return &e->realtime; + + case SOURCE_TIME_BOOTTIME: + return &e->boottime; + + case SOURCE_TIME_MONOTONIC: + return &e->monotonic; + + case SOURCE_TIME_REALTIME_ALARM: + return &e->realtime_alarm; + + case SOURCE_TIME_BOOTTIME_ALARM: + return &e->boottime_alarm; + + default: + return NULL; + } +} + +static void event_free_signal_data(sd_event *e, struct signal_data *d) { + assert(e); + + if (!d) + return; + + hashmap_remove(e->signal_data, &d->priority); + safe_close(d->fd); + free(d); +} + +static int event_make_signal_data( + sd_event *e, + int sig, + struct signal_data **ret) { + + struct signal_data *d; + bool added = false; + sigset_t ss_copy; + int64_t priority; + int r; + + assert(e); + + if (event_pid_changed(e)) + return -ECHILD; + + if (e->signal_sources && e->signal_sources[sig]) + priority = e->signal_sources[sig]->priority; + else + 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; + } + } else { + r = hashmap_ensure_allocated(&e->signal_data, &uint64_hash_ops); + if (r < 0) + return r; + + d = new(struct signal_data, 1); + if (!d) + return -ENOMEM; + + *d = (struct signal_data) { + .wakeup = WAKEUP_SIGNAL_DATA, + .fd = -1, + .priority = priority, + }; + + r = hashmap_put(e->signal_data, &d->priority, d); + if (r < 0) { + free(d); + return r; + } + + added = true; + } + + ss_copy = d->sigset; + assert_se(sigaddset(&ss_copy, sig) >= 0); + + r = signalfd(d->fd, &ss_copy, SFD_NONBLOCK|SFD_CLOEXEC); + if (r < 0) { + r = -errno; + goto fail; + } + + d->sigset = ss_copy; + + if (d->fd >= 0) { + if (ret) + *ret = d; + return 0; + } + + d->fd = fd_move_above_stdio(r); + + struct epoll_event ev = { + .events = EPOLLIN, + .data.ptr = d, + }; + + if (epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, d->fd, &ev) < 0) { + r = -errno; + goto fail; + } + + if (ret) + *ret = d; + + return 0; + +fail: + if (added) + event_free_signal_data(e, d); + + return r; +} + +static void event_unmask_signal_data(sd_event *e, struct signal_data *d, int sig) { + assert(e); + assert(d); + + /* Turns off the specified signal in the signal data + * object. If the signal mask of the object becomes empty that + * way removes it. */ + + if (sigismember(&d->sigset, sig) == 0) + return; + + assert_se(sigdelset(&d->sigset, sig) >= 0); + + if (sigisemptyset(&d->sigset)) { + /* If all the mask is all-zero we can get rid of the structure */ + event_free_signal_data(e, d); + return; + } + + assert(d->fd >= 0); + + if (signalfd(d->fd, &d->sigset, SFD_NONBLOCK|SFD_CLOEXEC) < 0) + log_debug_errno(errno, "Failed to unset signal bit, ignoring: %m"); +} + +static void event_gc_signal_data(sd_event *e, const int64_t *priority, int sig) { + struct signal_data *d; + static const int64_t zero_priority = 0; + + assert(e); + + /* Rechecks if the specified signal is still something we are interested in. If not, we'll unmask it, + * and possibly drop the signalfd for it. */ + + if (sig == SIGCHLD && + e->n_enabled_child_sources > 0) + return; + + if (e->signal_sources && + e->signal_sources[sig] && + e->signal_sources[sig]->enabled != SD_EVENT_OFF) + return; + + /* + * The specified signal might be enabled in three different queues: + * + * 1) the one that belongs to the priority passed (if it is non-NULL) + * 2) the one that belongs to the priority of the event source of the signal (if there is one) + * 3) the 0 priority (to cover the SIGCHLD case) + * + * Hence, let's remove it from all three here. + */ + + if (priority) { + d = hashmap_get(e->signal_data, priority); + if (d) + event_unmask_signal_data(e, d, sig); + } + + if (e->signal_sources && e->signal_sources[sig]) { + d = hashmap_get(e->signal_data, &e->signal_sources[sig]->priority); + if (d) + event_unmask_signal_data(e, d, sig); + } + + d = hashmap_get(e->signal_data, &zero_priority); + if (d) + event_unmask_signal_data(e, d, sig); +} + +static void event_source_pp_prioq_reshuffle(sd_event_source *s) { + assert(s); + + /* Reshuffles the pending + prepare prioqs. Called whenever the dispatch order changes, i.e. when + * they are enabled/disabled or marked pending and such. */ + + if (s->pending) + prioq_reshuffle(s->event->pending, s, &s->pending_index); + + if (s->prepare) + prioq_reshuffle(s->event->prepare, s, &s->prepare_index); +} + +static void event_source_time_prioq_reshuffle(sd_event_source *s) { + struct clock_data *d; + + assert(s); + assert(EVENT_SOURCE_IS_TIME(s->type)); + + /* Called whenever the event source's timer ordering properties changed, i.e. time, accuracy, + * pending, enable state. Makes sure the two prioq's are ordered properly again. */ + assert_se(d = event_get_clock_data(s->event, s->type)); + prioq_reshuffle(d->earliest, s, &s->time.earliest_index); + prioq_reshuffle(d->latest, s, &s->time.latest_index); + d->needs_rearm = true; +} + +static void source_disconnect(sd_event_source *s) { + sd_event *event; + + assert(s); + + if (!s->event) + return; + + assert(s->event->n_sources > 0); + + switch (s->type) { + + case SOURCE_IO: + if (s->io.fd >= 0) + source_io_unregister(s); + + break; + + case SOURCE_TIME_REALTIME: + case SOURCE_TIME_BOOTTIME: + case SOURCE_TIME_MONOTONIC: + case SOURCE_TIME_REALTIME_ALARM: + case SOURCE_TIME_BOOTTIME_ALARM: { + struct clock_data *d; + + d = event_get_clock_data(s->event, s->type); + assert(d); + + prioq_remove(d->earliest, s, &s->time.earliest_index); + prioq_remove(d->latest, s, &s->time.latest_index); + d->needs_rearm = true; + break; + } + + case SOURCE_SIGNAL: + if (s->signal.sig > 0) { + + if (s->event->signal_sources) + s->event->signal_sources[s->signal.sig] = NULL; + + event_gc_signal_data(s->event, &s->priority, s->signal.sig); + } + + break; + + case SOURCE_CHILD: + if (s->child.pid > 0) { + if (s->enabled != SD_EVENT_OFF) { + assert(s->event->n_enabled_child_sources > 0); + s->event->n_enabled_child_sources--; + } + + (void) hashmap_remove(s->event->child_sources, PID_TO_PTR(s->child.pid)); + } + + if (EVENT_SOURCE_WATCH_PIDFD(s)) + source_child_pidfd_unregister(s); + else + event_gc_signal_data(s->event, &s->priority, SIGCHLD); + + break; + + case SOURCE_DEFER: + /* nothing */ + break; + + case SOURCE_POST: + set_remove(s->event->post_sources, s); + break; + + case SOURCE_EXIT: + 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 + * continuously 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."); + } + + if (s->pending) + prioq_remove(s->event->pending, s, &s->pending_index); + + if (s->prepare) + prioq_remove(s->event->prepare, s, &s->prepare_index); + + event = TAKE_PTR(s->event); + LIST_REMOVE(sources, event->sources, s); + event->n_sources--; + + /* Note that we don't invalidate the type here, since we still need it in order to close the fd or + * pidfd associated with this event source, which we'll do only on source_free(). */ + + if (!s->floating) + sd_event_unref(event); +} + +static void source_free(sd_event_source *s) { + assert(s); + + source_disconnect(s); + + if (s->type == SOURCE_IO && s->io.owned) + s->io.fd = safe_close(s->io.fd); + + if (s->type == SOURCE_CHILD) { + /* Eventually the kernel will do this automatically for us, but for now let's emulate this (unreliably) in userspace. */ + + if (s->child.process_owned) { + + if (!s->child.exited) { + bool sent = false; + + if (s->child.pidfd >= 0) { + if (pidfd_send_signal(s->child.pidfd, SIGKILL, NULL, 0) < 0) { + if (errno == ESRCH) /* Already dead */ + sent = true; + else if (!ERRNO_IS_NOT_SUPPORTED(errno)) + log_debug_errno(errno, "Failed to kill process " PID_FMT " via pidfd_send_signal(), re-trying via kill(): %m", + s->child.pid); + } else + sent = true; + } + + if (!sent) + if (kill(s->child.pid, SIGKILL) < 0) + if (errno != ESRCH) /* Already dead */ + log_debug_errno(errno, "Failed to kill process " PID_FMT " via kill(), ignoring: %m", + s->child.pid); + } + + if (!s->child.waited) { + siginfo_t si = {}; + + /* Reap the child if we can */ + (void) waitid(P_PID, s->child.pid, &si, WEXITED); + } + } + + if (s->child.pidfd_owned) + s->child.pidfd = safe_close(s->child.pidfd); + } + + if (s->destroy_callback) + s->destroy_callback(s->userdata); + + free(s->description); + free(s); +} +DEFINE_TRIVIAL_CLEANUP_FUNC(sd_event_source*, source_free); + +static int source_set_pending(sd_event_source *s, bool b) { + int r; + + assert(s); + assert(s->type != SOURCE_EXIT); + + if (s->pending == b) + return 0; + + s->pending = b; + + if (b) { + s->pending_iteration = s->event->iteration; + + r = prioq_put(s->event->pending, s, &s->pending_index); + if (r < 0) { + s->pending = false; + return r; + } + } else + assert_se(prioq_remove(s->event->pending, s, &s->pending_index)); + + if (EVENT_SOURCE_IS_TIME(s->type)) + event_source_time_prioq_reshuffle(s); + + if (s->type == SOURCE_SIGNAL && !b) { + struct signal_data *d; + + d = hashmap_get(s->event->signal_data, &s->priority); + if (d && d->current == 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; +} + +static sd_event_source *source_new(sd_event *e, bool floating, EventSourceType type) { + sd_event_source *s; + + assert(e); + + s = new(sd_event_source, 1); + if (!s) + return 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++; + + return s; +} + +static int io_exit_callback(sd_event_source *s, int fd, uint32_t revents, void *userdata) { + assert(s); + + return sd_event_exit(sd_event_source_get_event(s), PTR_TO_INT(userdata)); +} + +_public_ int sd_event_add_io( + sd_event *e, + sd_event_source **ret, + int fd, + uint32_t events, + sd_event_io_handler_t callback, + void *userdata) { + + _cleanup_(source_freep) sd_event_source *s = NULL; + 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(e->state != SD_EVENT_FINISHED, -ESTALE); + assert_return(!event_pid_changed(e), -ECHILD); + + if (!callback) + callback = io_exit_callback; + + s = source_new(e, !ret, SOURCE_IO); + if (!s) + return -ENOMEM; + + s->wakeup = WAKEUP_EVENT_SOURCE; + s->io.fd = fd; + s->io.events = events; + s->io.callback = callback; + s->userdata = userdata; + s->enabled = SD_EVENT_ON; + + r = source_io_register(s, s->enabled, events); + if (r < 0) + return r; + + if (ret) + *ret = s; + TAKE_PTR(s); + + return 0; +} + +static void initialize_perturb(sd_event *e) { + sd_id128_t bootid = {}; + + /* When we sleep for longer, we try to realign the wakeup to + the same time within each minute/second/250ms, so that + events all across the system can be coalesced into a single + CPU wakeup. However, let's take some system-specific + randomness for this value, so that in a network of systems + with synced clocks timer events are distributed a + bit. Here, we calculate a perturbation usec offset from the + boot ID. */ + + if (_likely_(e->perturb != USEC_INFINITY)) + return; + + if (sd_id128_get_boot(&bootid) >= 0) + e->perturb = (bootid.qwords[0] ^ bootid.qwords[1]) % USEC_PER_MINUTE; +} + +static int event_setup_timer_fd( + sd_event *e, + struct clock_data *d, + clockid_t clock) { + + assert(e); + assert(d); + + if (_likely_(d->fd >= 0)) + return 0; + + _cleanup_close_ int fd = -1; + + fd = timerfd_create(clock, TFD_NONBLOCK|TFD_CLOEXEC); + if (fd < 0) + return -errno; + + fd = fd_move_above_stdio(fd); + + struct epoll_event ev = { + .events = EPOLLIN, + .data.ptr = d, + }; + + if (epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, fd, &ev) < 0) + return -errno; + + d->fd = TAKE_FD(fd); + return 0; +} + +static int time_exit_callback(sd_event_source *s, uint64_t usec, void *userdata) { + assert(s); + + return sd_event_exit(sd_event_source_get_event(s), PTR_TO_INT(userdata)); +} + +_public_ int sd_event_add_time( + sd_event *e, + sd_event_source **ret, + clockid_t clock, + uint64_t usec, + uint64_t accuracy, + sd_event_time_handler_t callback, + void *userdata) { + + EventSourceType type; + _cleanup_(source_freep) sd_event_source *s = NULL; + struct clock_data *d; + int r; + + assert_return(e, -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; + + d = event_get_clock_data(e, type); + assert(d); + + r = prioq_ensure_allocated(&d->earliest, earliest_time_prioq_compare); + if (r < 0) + return r; + + 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); + if (r < 0) + return r; + } + + s = source_new(e, !ret, type); + if (!s) + return -ENOMEM; + + s->time.next = usec; + s->time.accuracy = accuracy == 0 ? DEFAULT_ACCURACY_USEC : accuracy; + s->time.callback = callback; + s->time.earliest_index = s->time.latest_index = PRIOQ_IDX_NULL; + s->userdata = userdata; + s->enabled = SD_EVENT_ONESHOT; + + d->needs_rearm = true; + + r = prioq_put(d->earliest, s, &s->time.earliest_index); + if (r < 0) + return r; + + r = prioq_put(d->latest, s, &s->time.latest_index); + if (r < 0) + return r; + + if (ret) + *ret = s; + TAKE_PTR(s); + + return 0; +} + +_public_ int sd_event_add_time_relative( + sd_event *e, + sd_event_source **ret, + clockid_t clock, + uint64_t usec, + uint64_t accuracy, + sd_event_time_handler_t callback, + void *userdata) { + + usec_t t; + int r; + + /* Same as sd_event_add_time() but operates relative to the event loop's current point in time, and + * checks for overflow. */ + + r = sd_event_now(e, clock, &t); + if (r < 0) + return r; + + if (usec >= USEC_INFINITY - t) + return -EOVERFLOW; + + return sd_event_add_time(e, ret, clock, t + usec, accuracy, callback, userdata); +} + +static int signal_exit_callback(sd_event_source *s, const struct signalfd_siginfo *si, void *userdata) { + assert(s); + + return sd_event_exit(sd_event_source_get_event(s), PTR_TO_INT(userdata)); +} + +_public_ int sd_event_add_signal( + sd_event *e, + sd_event_source **ret, + int sig, + sd_event_signal_handler_t callback, + void *userdata) { + + _cleanup_(source_freep) sd_event_source *s = NULL; + struct signal_data *d; + int r; + + assert_return(e, -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); + + if (!callback) + callback = signal_exit_callback; + + r = signal_is_blocked(sig); + if (r < 0) + return r; + if (r == 0) + return -EBUSY; + + if (!e->signal_sources) { + e->signal_sources = new0(sd_event_source*, _NSIG); + if (!e->signal_sources) + return -ENOMEM; + } else if (e->signal_sources[sig]) + return -EBUSY; + + s = source_new(e, !ret, SOURCE_SIGNAL); + if (!s) + return -ENOMEM; + + s->signal.sig = sig; + s->signal.callback = callback; + s->userdata = userdata; + s->enabled = SD_EVENT_ON; + + e->signal_sources[sig] = s; + + r = event_make_signal_data(e, sig, &d); + if (r < 0) + 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)); + + if (ret) + *ret = s; + TAKE_PTR(s); + + return 0; +} + +static int child_exit_callback(sd_event_source *s, const siginfo_t *si, void *userdata) { + assert(s); + + return sd_event_exit(sd_event_source_get_event(s), PTR_TO_INT(userdata)); +} + +static bool shall_use_pidfd(void) { + /* Mostly relevant for debugging, i.e. this is used in test-event.c to test the event loop once with and once without pidfd */ + return getenv_bool_secure("SYSTEMD_PIDFD") != 0; +} + +_public_ int sd_event_add_child( + sd_event *e, + sd_event_source **ret, + pid_t pid, + int options, + sd_event_child_handler_t callback, + void *userdata) { + + _cleanup_(source_freep) sd_event_source *s = NULL; + 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); + assert_return(e->state != SD_EVENT_FINISHED, -ESTALE); + assert_return(!event_pid_changed(e), -ECHILD); + + if (!callback) + callback = child_exit_callback; + + if (e->n_enabled_child_sources == 0) { + /* Caller must block SIGCHLD before using us to watch children, even if pidfd is available, + * for compatibility with pre-pidfd and because we don't want the reap the child processes + * ourselves, i.e. call waitid(), and don't want Linux' default internal logic for that to + * take effect. + * + * (As an optimization we only do this check on the first child event source created.) */ + r = signal_is_blocked(SIGCHLD); + if (r < 0) + return r; + if (r == 0) + return -EBUSY; + } + + r = hashmap_ensure_allocated(&e->child_sources, NULL); + if (r < 0) + return r; + + if (hashmap_contains(e->child_sources, PID_TO_PTR(pid))) + return -EBUSY; + + s = source_new(e, !ret, SOURCE_CHILD); + if (!s) + return -ENOMEM; + + s->wakeup = WAKEUP_EVENT_SOURCE; + s->child.pid = pid; + s->child.options = options; + s->child.callback = callback; + s->userdata = userdata; + s->enabled = SD_EVENT_ONESHOT; + + /* We always take a pidfd here if we can, even if we wait for anything else than WEXITED, so that we + * pin the PID, and make regular waitid() handling race-free. */ + + if (shall_use_pidfd()) { + s->child.pidfd = pidfd_open(s->child.pid, 0); + if (s->child.pidfd < 0) { + /* Propagate errors unless the syscall is not supported or blocked */ + if (!ERRNO_IS_NOT_SUPPORTED(errno) && !ERRNO_IS_PRIVILEGE(errno)) + return -errno; + } else + s->child.pidfd_owned = true; /* If we allocate the pidfd we own it by default */ + } else + s->child.pidfd = -1; + + r = hashmap_put(e->child_sources, PID_TO_PTR(pid), s); + if (r < 0) + return r; + + if (EVENT_SOURCE_WATCH_PIDFD(s)) { + /* We have a pidfd and we only want to watch for exit */ + r = source_child_pidfd_register(s, s->enabled); + if (r < 0) + return r; + + } else { + /* We have no pidfd or we shall wait for some other event than WEXITED */ + r = event_make_signal_data(e, SIGCHLD, NULL); + if (r < 0) + return r; + + e->need_process_child = true; + } + + e->n_enabled_child_sources++; + + if (ret) + *ret = s; + TAKE_PTR(s); + return 0; +} + +_public_ int sd_event_add_child_pidfd( + sd_event *e, + sd_event_source **ret, + int pidfd, + int options, + sd_event_child_handler_t callback, + void *userdata) { + + + _cleanup_(source_freep) sd_event_source *s = NULL; + pid_t pid; + int r; + + assert_return(e, -EINVAL); + assert_return(e = event_resolve(e), -ENOPKG); + assert_return(pidfd >= 0, -EBADF); + assert_return(!(options & ~(WEXITED|WSTOPPED|WCONTINUED)), -EINVAL); + assert_return(options != 0, -EINVAL); + assert_return(e->state != SD_EVENT_FINISHED, -ESTALE); + assert_return(!event_pid_changed(e), -ECHILD); + + if (!callback) + callback = child_exit_callback; + + if (e->n_enabled_child_sources == 0) { + r = signal_is_blocked(SIGCHLD); + if (r < 0) + return r; + if (r == 0) + return -EBUSY; + } + + r = hashmap_ensure_allocated(&e->child_sources, NULL); + if (r < 0) + return r; + + r = pidfd_get_pid(pidfd, &pid); + if (r < 0) + return r; + + if (hashmap_contains(e->child_sources, PID_TO_PTR(pid))) + return -EBUSY; + + s = source_new(e, !ret, SOURCE_CHILD); + if (!s) + return -ENOMEM; + + s->wakeup = WAKEUP_EVENT_SOURCE; + s->child.pidfd = pidfd; + s->child.pid = pid; + s->child.options = options; + s->child.callback = callback; + 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) */ + s->userdata = userdata; + s->enabled = SD_EVENT_ONESHOT; + + r = hashmap_put(e->child_sources, PID_TO_PTR(pid), s); + if (r < 0) + return r; + + if (EVENT_SOURCE_WATCH_PIDFD(s)) { + /* We only want to watch for WEXITED */ + r = source_child_pidfd_register(s, s->enabled); + if (r < 0) + return r; + } else { + /* We shall wait for some other event than WEXITED */ + r = event_make_signal_data(e, SIGCHLD, NULL); + if (r < 0) + return r; + + e->need_process_child = true; + } + + e->n_enabled_child_sources++; + + if (ret) + *ret = s; + TAKE_PTR(s); + return 0; +} + +static int generic_exit_callback(sd_event_source *s, void *userdata) { + assert(s); + + return sd_event_exit(sd_event_source_get_event(s), PTR_TO_INT(userdata)); +} + +_public_ int sd_event_add_defer( + sd_event *e, + sd_event_source **ret, + sd_event_handler_t callback, + void *userdata) { + + _cleanup_(source_freep) sd_event_source *s = NULL; + int r; + + 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); + + if (!callback) + callback = generic_exit_callback; + + s = source_new(e, !ret, SOURCE_DEFER); + if (!s) + return -ENOMEM; + + s->defer.callback = callback; + s->userdata = userdata; + s->enabled = SD_EVENT_ONESHOT; + + r = source_set_pending(s, true); + if (r < 0) + return r; + + if (ret) + *ret = s; + TAKE_PTR(s); + + return 0; +} + +_public_ int sd_event_add_post( + sd_event *e, + sd_event_source **ret, + sd_event_handler_t callback, + void *userdata) { + + _cleanup_(source_freep) sd_event_source *s = NULL; + int r; + + 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); + + if (!callback) + callback = generic_exit_callback; + + s = source_new(e, !ret, SOURCE_POST); + if (!s) + return -ENOMEM; + + s->post.callback = callback; + s->userdata = userdata; + s->enabled = SD_EVENT_ON; + + r = set_ensure_put(&e->post_sources, NULL, s); + if (r < 0) + return r; + assert(r > 0); + + if (ret) + *ret = s; + TAKE_PTR(s); + + return 0; +} + +_public_ int sd_event_add_exit( + sd_event *e, + sd_event_source **ret, + sd_event_handler_t callback, + void *userdata) { + + _cleanup_(source_freep) sd_event_source *s = NULL; + 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); + + r = prioq_ensure_allocated(&e->exit, exit_prioq_compare); + if (r < 0) + return r; + + s = source_new(e, !ret, SOURCE_EXIT); + if (!s) + return -ENOMEM; + + s->exit.callback = callback; + s->userdata = userdata; + s->exit.prioq_index = PRIOQ_IDX_NULL; + s->enabled = SD_EVENT_ONESHOT; + + r = prioq_put(s->event->exit, s, &s->exit.prioq_index); + if (r < 0) + return r; + + if (ret) + *ret = s; + TAKE_PTR(s); + + return 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; + 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; + } + + struct epoll_event ev = { + .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 struct inode_data *x, const struct inode_data *y) { + int r; + + assert(x); + assert(y); + + r = CMP(x->dev, y->dev); + if (r != 0) + return r; + + return CMP(x->ino, y->ino); +} + +static void inode_data_hash_func(const struct inode_data *d, struct siphash *state) { + assert(d); + + siphash24_compress(&d->dev, sizeof(d->dev), state); + siphash24_compress(&d->ino, sizeof(d->ino), state); +} + +DEFINE_PRIVATE_HASH_OPS(inode_data_hash_ops, struct inode_data, inode_data_hash_func, 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 suppress + * 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; +} + +static int inotify_exit_callback(sd_event_source *s, const struct inotify_event *event, void *userdata) { + assert(s); + + return sd_event_exit(sd_event_source_get_event(s), PTR_TO_INT(userdata)); +} + +_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) { + + struct inotify_data *inotify_data = NULL; + struct inode_data *inode_data = NULL; + _cleanup_close_ int fd = -1; + _cleanup_(source_freep) sd_event_source *s = NULL; + struct stat st; + int r; + + assert_return(e, -EINVAL); + assert_return(e = event_resolve(e), -ENOPKG); + assert_return(path, -EINVAL); + assert_return(e->state != SD_EVENT_FINISHED, -ESTALE); + assert_return(!event_pid_changed(e), -ECHILD); + + if (!callback) + callback = inotify_exit_callback; + + /* 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) + return r; + + r = event_make_inode_data(e, inotify_data, st.st_dev, st.st_ino, &inode_data); + if (r < 0) { + event_free_inotify_data(e, inotify_data); + return r; + } + + /* 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; + + /* Actually realize the watch now */ + r = inode_data_realize_watch(e, inode_data); + if (r < 0) + return r; + + (void) sd_event_source_set_description(s, path); + + if (ret) + *ret = s; + TAKE_PTR(s); + + return 0; +} + +static sd_event_source* event_source_free(sd_event_source *s) { + if (!s) + return NULL; + + /* Here's a special hack: when we are called from a + * dispatch handler we won't free the event source + * immediately, but we will detach the fd from the + * epoll. This way it is safe for the caller to unref + * the event source and immediately close the fd, but + * we still retain a valid event source object after + * the callback. */ + + if (s->dispatching) { + if (s->type == SOURCE_IO) + source_io_unregister(s); + + source_disconnect(s); + } else + source_free(s); + + return NULL; +} + +DEFINE_PUBLIC_TRIVIAL_REF_UNREF_FUNC(sd_event_source, sd_event_source, event_source_free); + +_public_ int sd_event_source_set_description(sd_event_source *s, const char *description) { + assert_return(s, -EINVAL); + assert_return(!event_pid_changed(s->event), -ECHILD); + + return free_and_strdup(&s->description, description); +} + +_public_ int sd_event_source_get_description(sd_event_source *s, const char **description) { + assert_return(s, -EINVAL); + assert_return(description, -EINVAL); + assert_return(!event_pid_changed(s->event), -ECHILD); + + if (!s->description) + return -ENXIO; + + *description = s->description; + return 0; +} + +_public_ sd_event *sd_event_source_get_event(sd_event_source *s) { + assert_return(s, NULL); + + return s->event; +} + +_public_ int sd_event_source_get_pending(sd_event_source *s) { + assert_return(s, -EINVAL); + assert_return(s->type != SOURCE_EXIT, -EDOM); + assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE); + assert_return(!event_pid_changed(s->event), -ECHILD); + + return s->pending; +} + +_public_ int sd_event_source_get_io_fd(sd_event_source *s) { + assert_return(s, -EINVAL); + assert_return(s->type == SOURCE_IO, -EDOM); + assert_return(!event_pid_changed(s->event), -ECHILD); + + return s->io.fd; +} + +_public_ int sd_event_source_set_io_fd(sd_event_source *s, int fd) { + int r; + + assert_return(s, -EINVAL); + assert_return(fd >= 0, -EBADF); + assert_return(s->type == SOURCE_IO, -EDOM); + assert_return(!event_pid_changed(s->event), -ECHILD); + + if (s->io.fd == fd) + return 0; + + if (s->enabled == SD_EVENT_OFF) { + s->io.fd = fd; + s->io.registered = false; + } else { + int saved_fd; + + saved_fd = s->io.fd; + assert(s->io.registered); + + s->io.fd = fd; + s->io.registered = false; + + r = source_io_register(s, s->enabled, s->io.events); + if (r < 0) { + s->io.fd = saved_fd; + s->io.registered = true; + return r; + } + + (void) epoll_ctl(s->event->epoll_fd, EPOLL_CTL_DEL, saved_fd, NULL); + } + + return 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); + assert_return(s->type == SOURCE_IO, -EDOM); + assert_return(!event_pid_changed(s->event), -ECHILD); + + *events = s->io.events; + return 0; +} + +_public_ int sd_event_source_set_io_events(sd_event_source *s, uint32_t events) { + int r; + + assert_return(s, -EINVAL); + assert_return(s->type == SOURCE_IO, -EDOM); + assert_return(!(events & ~(EPOLLIN|EPOLLOUT|EPOLLRDHUP|EPOLLPRI|EPOLLERR|EPOLLHUP|EPOLLET)), -EINVAL); + assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE); + assert_return(!event_pid_changed(s->event), -ECHILD); + + /* edge-triggered updates are never skipped, so we can reset edges */ + 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) + return r; + } + + s->io.events = events; + + return 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->type == SOURCE_IO, -EDOM); + assert_return(s->pending, -ENODATA); + assert_return(!event_pid_changed(s->event), -ECHILD); + + *revents = s->io.revents; + return 0; +} + +_public_ int sd_event_source_get_signal(sd_event_source *s) { + assert_return(s, -EINVAL); + assert_return(s->type == SOURCE_SIGNAL, -EDOM); + assert_return(!event_pid_changed(s->event), -ECHILD); + + return s->signal.sig; +} + +_public_ int sd_event_source_get_priority(sd_event_source *s, int64_t *priority) { + assert_return(s, -EINVAL); + assert_return(!event_pid_changed(s->event), -ECHILD); + + *priority = s->priority; + return 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); + assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE); + assert_return(!event_pid_changed(s->event), -ECHILD); + + if (s->priority == priority) + return 0; + + 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 + * priority to the signalfd of the new priority */ + + assert_se(old = hashmap_get(s->event->signal_data, &s->priority)); + + s->priority = priority; + + r = event_make_signal_data(s->event, s->signal.sig, &d); + if (r < 0) { + s->priority = old->priority; + return r; + } + + event_unmask_signal_data(s->event, old, s->signal.sig); + } else + s->priority = priority; + + event_source_pp_prioq_reshuffle(s); + + if (s->type == SOURCE_EXIT) + 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; +} + +_public_ int sd_event_source_get_enabled(sd_event_source *s, int *m) { + assert_return(s, -EINVAL); + assert_return(!event_pid_changed(s->event), -ECHILD); + + if (m) + *m = s->enabled; + return s->enabled != SD_EVENT_OFF; +} + +static int event_source_disable(sd_event_source *s) { + int r; + + assert(s); + assert(s->enabled != 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; + } + + s->enabled = SD_EVENT_OFF; + + switch (s->type) { + + case SOURCE_IO: + source_io_unregister(s); + break; + + case SOURCE_TIME_REALTIME: + case SOURCE_TIME_BOOTTIME: + case SOURCE_TIME_MONOTONIC: + case SOURCE_TIME_REALTIME_ALARM: + case SOURCE_TIME_BOOTTIME_ALARM: + event_source_time_prioq_reshuffle(s); + break; + + case SOURCE_SIGNAL: + event_gc_signal_data(s->event, &s->priority, s->signal.sig); + break; + + case SOURCE_CHILD: + assert(s->event->n_enabled_child_sources > 0); + s->event->n_enabled_child_sources--; + + if (EVENT_SOURCE_WATCH_PIDFD(s)) + source_child_pidfd_unregister(s); + else + event_gc_signal_data(s->event, &s->priority, SIGCHLD); + break; + + case SOURCE_EXIT: + prioq_reshuffle(s->event->exit, s, &s->exit.prioq_index); + break; + + case SOURCE_DEFER: + case SOURCE_POST: + case SOURCE_INOTIFY: + break; + + default: + assert_not_reached("Wut? I shouldn't exist."); + } + + return 0; +} + +static int event_source_enable(sd_event_source *s, int enable) { + int r; + + assert(s); + assert(IN_SET(enable, SD_EVENT_ON, SD_EVENT_ONESHOT)); + assert(s->enabled == SD_EVENT_OFF); + + /* Unset the pending flag when this event source is enabled */ + 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: + r = source_io_register(s, enable, s->io.events); + if (r < 0) + return r; + break; + + case SOURCE_SIGNAL: + r = event_make_signal_data(s->event, s->signal.sig, NULL); + if (r < 0) { + event_gc_signal_data(s->event, &s->priority, s->signal.sig); + return r; + } + + break; + + case SOURCE_CHILD: + if (EVENT_SOURCE_WATCH_PIDFD(s)) { + /* yes, we have pidfd */ + + r = source_child_pidfd_register(s, enable); + if (r < 0) + return r; + } else { + /* no pidfd, or something other to watch for than WEXITED */ + + r = event_make_signal_data(s->event, SIGCHLD, NULL); + if (r < 0) { + event_gc_signal_data(s->event, &s->priority, SIGCHLD); + return r; + } + } + + s->event->n_enabled_child_sources++; + + break; + + case SOURCE_TIME_REALTIME: + case SOURCE_TIME_BOOTTIME: + case SOURCE_TIME_MONOTONIC: + case SOURCE_TIME_REALTIME_ALARM: + case SOURCE_TIME_BOOTTIME_ALARM: + case SOURCE_EXIT: + case SOURCE_DEFER: + case SOURCE_POST: + case SOURCE_INOTIFY: + break; + + default: + assert_not_reached("Wut? I shouldn't exist."); + } + + s->enabled = enable; + + /* Non-failing operations below */ + switch (s->type) { + case SOURCE_TIME_REALTIME: + case SOURCE_TIME_BOOTTIME: + case SOURCE_TIME_MONOTONIC: + case SOURCE_TIME_REALTIME_ALARM: + case SOURCE_TIME_BOOTTIME_ALARM: + event_source_time_prioq_reshuffle(s); + break; + + case SOURCE_EXIT: + prioq_reshuffle(s->event->exit, s, &s->exit.prioq_index); + break; + + default: + break; + } + + return 0; +} + +_public_ int sd_event_source_set_enabled(sd_event_source *s, int m) { + int r; + + assert_return(s, -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 sources, but everything else needs to fail. */ + if (s->event->state == SD_EVENT_FINISHED) + return m == SD_EVENT_OFF ? 0 : -ESTALE; + + if (s->enabled == m) /* No change? */ + return 0; + + if (m == SD_EVENT_OFF) + r = event_source_disable(s); + else { + if (s->enabled != SD_EVENT_OFF) { + /* Switching from "on" to "oneshot" or back? If that's the case, we can take a shortcut, the + * event source is already enabled after all. */ + s->enabled = m; + return 0; + } + + r = event_source_enable(s, m); + } + if (r < 0) + return r; + + event_source_pp_prioq_reshuffle(s); + return 0; +} + +_public_ int sd_event_source_get_time(sd_event_source *s, uint64_t *usec) { + assert_return(s, -EINVAL); + assert_return(usec, -EINVAL); + assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM); + assert_return(!event_pid_changed(s->event), -ECHILD); + + *usec = s->time.next; + return 0; +} + +_public_ int sd_event_source_set_time(sd_event_source *s, uint64_t usec) { + int r; + + assert_return(s, -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); + + r = source_set_pending(s, false); + if (r < 0) + return r; + + s->time.next = usec; + + event_source_time_prioq_reshuffle(s); + return 0; +} + +_public_ int sd_event_source_set_time_relative(sd_event_source *s, uint64_t usec) { + usec_t t; + int r; + + assert_return(s, -EINVAL); + assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM); + + r = sd_event_now(s->event, event_source_type_to_clock(s->type), &t); + if (r < 0) + return r; + + if (usec >= USEC_INFINITY - t) + return -EOVERFLOW; + + return sd_event_source_set_time(s, t + usec); +} + +_public_ int sd_event_source_get_time_accuracy(sd_event_source *s, uint64_t *usec) { + assert_return(s, -EINVAL); + assert_return(usec, -EINVAL); + assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM); + assert_return(!event_pid_changed(s->event), -ECHILD); + + *usec = s->time.accuracy; + return 0; +} + +_public_ int sd_event_source_set_time_accuracy(sd_event_source *s, uint64_t usec) { + 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); + + r = source_set_pending(s, false); + if (r < 0) + return r; + + if (usec == 0) + usec = DEFAULT_ACCURACY_USEC; + + s->time.accuracy = usec; + + event_source_time_prioq_reshuffle(s); + return 0; +} + +_public_ int sd_event_source_get_time_clock(sd_event_source *s, clockid_t *clock) { + assert_return(s, -EINVAL); + assert_return(clock, -EINVAL); + assert_return(EVENT_SOURCE_IS_TIME(s->type), -EDOM); + assert_return(!event_pid_changed(s->event), -ECHILD); + + *clock = event_source_type_to_clock(s->type); + return 0; +} + +_public_ int sd_event_source_get_child_pid(sd_event_source *s, pid_t *pid) { + assert_return(s, -EINVAL); + assert_return(pid, -EINVAL); + assert_return(s->type == SOURCE_CHILD, -EDOM); + assert_return(!event_pid_changed(s->event), -ECHILD); + + *pid = s->child.pid; + return 0; +} + +_public_ int sd_event_source_get_child_pidfd(sd_event_source *s) { + assert_return(s, -EINVAL); + assert_return(s->type == SOURCE_CHILD, -EDOM); + assert_return(!event_pid_changed(s->event), -ECHILD); + + if (s->child.pidfd < 0) + return -EOPNOTSUPP; + + return s->child.pidfd; +} + +_public_ int sd_event_source_send_child_signal(sd_event_source *s, int sig, const siginfo_t *si, unsigned flags) { + assert_return(s, -EINVAL); + assert_return(s->type == SOURCE_CHILD, -EDOM); + assert_return(!event_pid_changed(s->event), -ECHILD); + assert_return(SIGNAL_VALID(sig), -EINVAL); + + /* If we already have seen indication the process exited refuse sending a signal early. This way we + * can be sure we don't accidentally kill the wrong process on PID reuse when pidfds are not + * available. */ + if (s->child.exited) + return -ESRCH; + + if (s->child.pidfd >= 0) { + siginfo_t copy; + + /* pidfd_send_signal() changes the siginfo_t argument. This is weird, let's hence copy the + * structure here */ + if (si) + copy = *si; + + if (pidfd_send_signal(s->child.pidfd, sig, si ? © : NULL, 0) < 0) { + /* Let's propagate the error only if the system call is not implemented or prohibited */ + if (!ERRNO_IS_NOT_SUPPORTED(errno) && !ERRNO_IS_PRIVILEGE(errno)) + return -errno; + } else + return 0; + } + + /* Flags are only supported for pidfd_send_signal(), not for rt_sigqueueinfo(), hence let's refuse + * this here. */ + if (flags != 0) + return -EOPNOTSUPP; + + if (si) { + /* We use rt_sigqueueinfo() only if siginfo_t is specified. */ + siginfo_t copy = *si; + + if (rt_sigqueueinfo(s->child.pid, sig, ©) < 0) + return -errno; + } else if (kill(s->child.pid, sig) < 0) + return -errno; + + return 0; +} + +_public_ int sd_event_source_get_child_pidfd_own(sd_event_source *s) { + assert_return(s, -EINVAL); + assert_return(s->type == SOURCE_CHILD, -EDOM); + + if (s->child.pidfd < 0) + return -EOPNOTSUPP; + + return s->child.pidfd_owned; +} + +_public_ int sd_event_source_set_child_pidfd_own(sd_event_source *s, int own) { + assert_return(s, -EINVAL); + assert_return(s->type == SOURCE_CHILD, -EDOM); + + if (s->child.pidfd < 0) + return -EOPNOTSUPP; + + s->child.pidfd_owned = own; + return 0; +} + +_public_ int sd_event_source_get_child_process_own(sd_event_source *s) { + assert_return(s, -EINVAL); + assert_return(s->type == SOURCE_CHILD, -EDOM); + + return s->child.process_owned; +} + +_public_ int sd_event_source_set_child_process_own(sd_event_source *s, int own) { + assert_return(s, -EINVAL); + assert_return(s->type == SOURCE_CHILD, -EDOM); + + s->child.process_owned = own; + return 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; + + assert_return(s, -EINVAL); + assert_return(s->type != SOURCE_EXIT, -EDOM); + assert_return(s->event->state != SD_EVENT_FINISHED, -ESTALE); + assert_return(!event_pid_changed(s->event), -ECHILD); + + if (s->prepare == callback) + return 0; + + if (callback && s->prepare) { + s->prepare = callback; + return 0; + } + + r = prioq_ensure_allocated(&s->event->prepare, prepare_prioq_compare); + if (r < 0) + return r; + + s->prepare = callback; + + if (callback) { + r = prioq_put(s->event->prepare, s, &s->prepare_index); + if (r < 0) + return r; + } else + prioq_remove(s->event->prepare, s, &s->prepare_index); + + return 0; +} + +_public_ void* sd_event_source_get_userdata(sd_event_source *s) { + assert_return(s, NULL); + + return s->userdata; +} + +_public_ void *sd_event_source_set_userdata(sd_event_source *s, void *userdata) { + void *ret; + + assert_return(s, NULL); + + ret = s->userdata; + s->userdata = userdata; + + return ret; +} + +static usec_t sleep_between(sd_event *e, usec_t a, usec_t b) { + usec_t c; + assert(e); + assert(a <= b); + + if (a <= 0) + return 0; + if (a >= USEC_INFINITY) + return USEC_INFINITY; + + if (b <= a + 1) + return a; + + initialize_perturb(e); + + /* + Find a good time to wake up again between times a and b. We + have two goals here: + + a) We want to wake up as seldom as possible, hence prefer + later times over earlier times. + + b) But if we have to wake up, then let's make sure to + dispatch as much as possible on the entire system. + + We implement this by waking up everywhere at the same time + within any given minute if we can, synchronised via the + perturbation value determined from the boot ID. If we can't, + then we try to find the same spot in every 10s, then 1s and + then 250ms step. Otherwise, we pick the last possible time + to wake up. + */ + + c = (b / USEC_PER_MINUTE) * USEC_PER_MINUTE + e->perturb; + if (c >= b) { + if (_unlikely_(c < USEC_PER_MINUTE)) + return b; + + c -= USEC_PER_MINUTE; + } + + if (c >= a) + return c; + + c = (b / (USEC_PER_SEC*10)) * (USEC_PER_SEC*10) + (e->perturb % (USEC_PER_SEC*10)); + if (c >= b) { + if (_unlikely_(c < USEC_PER_SEC*10)) + return b; + + c -= USEC_PER_SEC*10; + } + + if (c >= a) + return c; + + c = (b / USEC_PER_SEC) * USEC_PER_SEC + (e->perturb % USEC_PER_SEC); + if (c >= b) { + if (_unlikely_(c < USEC_PER_SEC)) + return b; + + c -= USEC_PER_SEC; + } + + if (c >= a) + return c; + + c = (b / (USEC_PER_MSEC*250)) * (USEC_PER_MSEC*250) + (e->perturb % (USEC_PER_MSEC*250)); + if (c >= b) { + if (_unlikely_(c < USEC_PER_MSEC*250)) + return b; + + c -= USEC_PER_MSEC*250; + } + + if (c >= a) + return c; + + return b; +} + +static int event_arm_timer( + sd_event *e, + struct clock_data *d) { + + struct itimerspec its = {}; + sd_event_source *a, *b; + usec_t t; + + assert(e); + assert(d); + + if (!d->needs_rearm) + return 0; + else + d->needs_rearm = false; + + a = prioq_peek(d->earliest); + if (!a || a->enabled == SD_EVENT_OFF || a->time.next == USEC_INFINITY) { + + if (d->fd < 0) + return 0; + + if (d->next == USEC_INFINITY) + return 0; + + /* disarm */ + if (timerfd_settime(d->fd, TFD_TIMER_ABSTIME, &its, NULL) < 0) + return -errno; + + d->next = USEC_INFINITY; + return 0; + } + + b = prioq_peek(d->latest); + assert_se(b && b->enabled != SD_EVENT_OFF); + + t = sleep_between(e, a->time.next, time_event_source_latest(b)); + if (d->next == t) + return 0; + + assert_se(d->fd >= 0); + + if (t == 0) { + /* We don' want to disarm here, just mean some time looooong ago. */ + its.it_value.tv_sec = 0; + its.it_value.tv_nsec = 1; + } else + timespec_store(&its.it_value, t); + + if (timerfd_settime(d->fd, TFD_TIMER_ABSTIME, &its, NULL) < 0) + return -errno; + + d->next = t; + return 0; +} + +static int process_io(sd_event *e, sd_event_source *s, uint32_t revents) { + assert(e); + assert(s); + assert(s->type == SOURCE_IO); + + /* If the event source was already pending, we just OR in the + * new revents, otherwise we reset the value. The ORing is + * necessary to handle EPOLLONESHOT events properly where + * readability might happen independently of writability, and + * we need to keep track of both */ + + if (s->pending) + s->io.revents |= revents; + else + s->io.revents = revents; + + return source_set_pending(s, true); +} + +static int flush_timer(sd_event *e, int fd, uint32_t events, usec_t *next) { + uint64_t x; + ssize_t ss; + + assert(e); + assert(fd >= 0); + + assert_return(events == EPOLLIN, -EIO); + + ss = read(fd, &x, sizeof(x)); + if (ss < 0) { + if (IN_SET(errno, EAGAIN, EINTR)) + return 0; + + return -errno; + } + + if (_unlikely_(ss != sizeof(x))) + return -EIO; + + if (next) + *next = USEC_INFINITY; + + return 0; +} + +static int process_timer( + sd_event *e, + usec_t n, + struct clock_data *d) { + + sd_event_source *s; + int r; + + assert(e); + assert(d); + + for (;;) { + s = prioq_peek(d->earliest); + if (!s || + s->time.next > n || + s->enabled == SD_EVENT_OFF || + s->pending) + break; + + r = source_set_pending(s, true); + if (r < 0) + return r; + + event_source_time_prioq_reshuffle(s); + } + + return 0; +} + +static int process_child(sd_event *e) { + sd_event_source *s; + int r; + + assert(e); + + e->need_process_child = false; + + /* + So, this is ugly. We iteratively invoke waitid() with P_PID + + WNOHANG for each PID we wait for, instead of using + P_ALL. This is because we only want to get child + information of very specific child processes, and not all + of them. We might not have processed the SIGCHLD even of a + previous invocation and we don't want to maintain a + unbounded *per-child* event queue, hence we really don't + want anything flushed out of the kernel's queue that we + don't care about. Since this is O(n) this means that if you + have a lot of processes you probably want to handle SIGCHLD + yourself. + + We do not reap the children here (by using WNOWAIT), this + is only done after the event source is dispatched so that + the callback still sees the process as a zombie. + */ + + HASHMAP_FOREACH(s, e->child_sources) { + assert(s->type == SOURCE_CHILD); + + if (s->pending) + continue; + + if (s->enabled == SD_EVENT_OFF) + continue; + + if (s->child.exited) + continue; + + if (EVENT_SOURCE_WATCH_PIDFD(s)) /* There's a usable pidfd known for this event source? then don't waitid() for it here */ + continue; + + zero(s->child.siginfo); + if (waitid(P_PID, s->child.pid, &s->child.siginfo, + WNOHANG | (s->child.options & WEXITED ? WNOWAIT : 0) | s->child.options) < 0) + return -errno; + + if (s->child.siginfo.si_pid != 0) { + bool zombie = IN_SET(s->child.siginfo.si_code, CLD_EXITED, CLD_KILLED, CLD_DUMPED); + + if (zombie) + s->child.exited = true; + + if (!zombie && (s->child.options & WEXITED)) { + /* If the child isn't dead then let's + * immediately remove the state change + * from the queue, since there's no + * benefit in leaving it queued */ + + assert(s->child.options & (WSTOPPED|WCONTINUED)); + (void) waitid(P_PID, s->child.pid, &s->child.siginfo, WNOHANG|(s->child.options & (WSTOPPED|WCONTINUED))); + } + + r = source_set_pending(s, true); + if (r < 0) + return r; + } + } + + return 0; +} + +static int process_pidfd(sd_event *e, sd_event_source *s, uint32_t revents) { + assert(e); + assert(s); + assert(s->type == SOURCE_CHILD); + + if (s->pending) + return 0; + + if (s->enabled == SD_EVENT_OFF) + return 0; + + if (!EVENT_SOURCE_WATCH_PIDFD(s)) + return 0; + + zero(s->child.siginfo); + if (waitid(P_PID, s->child.pid, &s->child.siginfo, WNOHANG | WNOWAIT | s->child.options) < 0) + return -errno; + + if (s->child.siginfo.si_pid == 0) + return 0; + + if (IN_SET(s->child.siginfo.si_code, CLD_EXITED, CLD_KILLED, CLD_DUMPED)) + s->child.exited = true; + + return source_set_pending(s, true); +} + +static int process_signal(sd_event *e, struct signal_data *d, uint32_t events) { + bool read_one = false; + int r; + + assert(e); + assert(d); + assert_return(events == EPOLLIN, -EIO); + + /* If there's a signal queued on this priority and SIGCHLD is + on this priority too, then make sure to recheck the + children we watch. This is because we only ever dequeue + the first signal per priority, and if we dequeue one, and + SIGCHLD might be enqueued later we wouldn't know, but we + might have higher priority children we care about hence we + need to check that explicitly. */ + + if (sigismember(&d->sigset, SIGCHLD)) + e->need_process_child = true; + + /* If there's already an event source pending for this + * priority we don't read another */ + if (d->current) + return 0; + + for (;;) { + struct signalfd_siginfo si; + ssize_t n; + sd_event_source *s = NULL; + + n = read(d->fd, &si, sizeof(si)); + if (n < 0) { + if (IN_SET(errno, EAGAIN, EINTR)) + return read_one; + + return -errno; + } + + if (_unlikely_(n != sizeof(si))) + return -EIO; + + assert(SIGNAL_VALID(si.ssi_signo)); + + read_one = true; + + if (e->signal_sources) + s = e->signal_sources[si.ssi_signo]; + if (!s) + continue; + if (s->pending) + continue; + + s->signal.siginfo = si; + d->current = s; + + r = source_set_pending(s, true); + if (r < 0) + return r; + + return 1; + } +} + +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; + + /* The queue overran, let's pass this event to all event sources connected to this inotify + * object */ + + HASHMAP_FOREACH(inode_data, d->inodes) { + 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) { + _cleanup_(sd_event_unrefp) sd_event *saved_event = NULL; + EventSourceType saved_type; + int r = 0; + + assert(s); + assert(s->pending || 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; + + /* Similar, store a reference to the event loop object, so that we can still access it after the + * callback might have invalidated/disconnected the event source. */ + saved_event = sd_event_ref(s->event); + + if (!IN_SET(s->type, SOURCE_DEFER, SOURCE_EXIT)) { + r = source_set_pending(s, false); + if (r < 0) + return r; + } + + if (s->type != SOURCE_POST) { + sd_event_source *z; + + /* If we execute a non-post source, let's mark all + * post sources as pending */ + + SET_FOREACH(z, s->event->post_sources) { + if (z->enabled == SD_EVENT_OFF) + continue; + + r = source_set_pending(z, true); + if (r < 0) + return r; + } + } + + if (s->enabled == SD_EVENT_ONESHOT) { + r = sd_event_source_set_enabled(s, SD_EVENT_OFF); + if (r < 0) + return r; + } + + s->dispatching = true; + + switch (s->type) { + + case SOURCE_IO: + r = s->io.callback(s, s->io.fd, s->io.revents, s->userdata); + break; + + case SOURCE_TIME_REALTIME: + case SOURCE_TIME_BOOTTIME: + case SOURCE_TIME_MONOTONIC: + case SOURCE_TIME_REALTIME_ALARM: + case SOURCE_TIME_BOOTTIME_ALARM: + r = s->time.callback(s, s->time.next, s->userdata); + break; + + case SOURCE_SIGNAL: + r = s->signal.callback(s, &s->signal.siginfo, s->userdata); + break; + + case SOURCE_CHILD: { + bool zombie; + + 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) { + (void) waitid(P_PID, s->child.pid, &s->child.siginfo, WNOHANG|WEXITED); + s->child.waited = true; + } + + break; + } + + case SOURCE_DEFER: + r = s->defer.callback(s, s->userdata); + break; + + case SOURCE_POST: + r = s->post.callback(s, s->userdata); + break; + + case SOURCE_EXIT: + 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: + assert_not_reached("Wut? I shouldn't exist."); + } + + s->dispatching = false; + + if (r < 0) { + log_debug_errno(r, "Event source %s (type %s) returned error, %s: %m", + strna(s->description), + event_source_type_to_string(saved_type), + s->exit_on_failure ? "exiting" : "disabling"); + + if (s->exit_on_failure) + (void) sd_event_exit(saved_event, r); + } + + if (s->n_ref == 0) + source_free(s); + else if (r < 0) + sd_event_source_set_enabled(s, SD_EVENT_OFF); + + return 1; +} + +static int event_prepare(sd_event *e) { + int r; + + assert(e); + + for (;;) { + sd_event_source *s; + + s = prioq_peek(e->prepare); + if (!s || s->prepare_iteration == e->iteration || s->enabled == SD_EVENT_OFF) + break; + + s->prepare_iteration = e->iteration; + r = prioq_reshuffle(e->prepare, s, &s->prepare_index); + if (r < 0) + return r; + + assert(s->prepare); + + s->dispatching = true; + r = s->prepare(s, s->userdata); + s->dispatching = false; + + if (r < 0) { + log_debug_errno(r, "Prepare callback of event source %s (type %s) returned error, %s: %m", + strna(s->description), + event_source_type_to_string(s->type), + s->exit_on_failure ? "exiting" : "disabling"); + + if (s->exit_on_failure) + (void) sd_event_exit(e, r); + } + + if (s->n_ref == 0) + source_free(s); + else if (r < 0) + sd_event_source_set_enabled(s, SD_EVENT_OFF); + } + + return 0; +} + +static int dispatch_exit(sd_event *e) { + sd_event_source *p; + _cleanup_(sd_event_unrefp) sd_event *ref = NULL; + int r; + + assert(e); + + p = prioq_peek(e->exit); + if (!p || p->enabled == SD_EVENT_OFF) { + e->state = SD_EVENT_FINISHED; + return 0; + } + + ref = sd_event_ref(e); + e->iteration++; + e->state = SD_EVENT_EXITING; + r = source_dispatch(p); + e->state = SD_EVENT_INITIAL; + return r; +} + +static sd_event_source* event_next_pending(sd_event *e) { + sd_event_source *p; + + assert(e); + + p = prioq_peek(e->pending); + if (!p) + return NULL; + + if (p->enabled == SD_EVENT_OFF) + return NULL; + + return p; +} + +static int arm_watchdog(sd_event *e) { + struct itimerspec its = {}; + usec_t t; + + assert(e); + assert(e->watchdog_fd >= 0); + + t = sleep_between(e, + e->watchdog_last + (e->watchdog_period / 2), + e->watchdog_last + (e->watchdog_period * 3 / 4)); + + timespec_store(&its.it_value, t); + + /* Make sure we never set the watchdog to 0, which tells the + * kernel to disable it. */ + if (its.it_value.tv_sec == 0 && its.it_value.tv_nsec == 0) + its.it_value.tv_nsec = 1; + + if (timerfd_settime(e->watchdog_fd, TFD_TIMER_ABSTIME, &its, NULL) < 0) + return -errno; + + return 0; +} + +static int process_watchdog(sd_event *e) { + assert(e); + + if (!e->watchdog) + return 0; + + /* Don't notify watchdog too often */ + if (e->watchdog_last + e->watchdog_period / 4 > e->timestamp.monotonic) + return 0; + + sd_notify(false, "WATCHDOG=1"); + e->watchdog_last = e->timestamp.monotonic; + + 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 re-add 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); + + /* Let's check that if we are a default event loop we are executed in the correct thread. We only do + * this check here once, since gettid() is typically not cached, and thus want to minimize + * syscalls */ + assert_return(!e->default_event_ptr || e->tid == gettid(), -EREMOTEIO); + + if (e->exit_requested) + goto pending; + + e->iteration++; + + e->state = SD_EVENT_PREPARING; + r = event_prepare(e); + e->state = SD_EVENT_INITIAL; + if (r < 0) + return r; + + r = event_arm_timer(e, &e->realtime); + if (r < 0) + return r; + + r = event_arm_timer(e, &e->boottime); + if (r < 0) + return r; + + r = event_arm_timer(e, &e->monotonic); + if (r < 0) + return r; + + r = event_arm_timer(e, &e->realtime_alarm); + if (r < 0) + return r; + + r = event_arm_timer(e, &e->boottime_alarm); + if (r < 0) + return r; + + event_close_inode_data_fds(e); + + if (event_next_pending(e) || e->need_process_child) + goto pending; + + e->state = SD_EVENT_ARMED; + + return 0; + +pending: + e->state = SD_EVENT_ARMED; + r = sd_event_wait(e, 0); + if (r == 0) + e->state = SD_EVENT_ARMED; + + return r; +} + +_public_ int sd_event_wait(sd_event *e, uint64_t timeout) { + size_t event_queue_max; + 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); + + if (e->exit_requested) { + e->state = SD_EVENT_PENDING; + return 1; + } + + event_queue_max = MAX(e->n_sources, 1u); + if (!GREEDY_REALLOC(e->event_queue, e->event_queue_allocated, event_queue_max)) + return -ENOMEM; + + /* 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, e->event_queue, event_queue_max, + timeout == (uint64_t) -1 ? -1 : (int) DIV_ROUND_UP(timeout, USEC_PER_MSEC)); + if (m < 0) { + if (errno == EINTR) { + e->state = SD_EVENT_PENDING; + return 1; + } + + r = -errno; + goto finish; + } + + triple_timestamp_get(&e->timestamp); + + for (i = 0; i < m; i++) { + + if (e->event_queue[i].data.ptr == INT_TO_PTR(SOURCE_WATCHDOG)) + r = flush_timer(e, e->watchdog_fd, e->event_queue[i].events, NULL); + else { + WakeupType *t = e->event_queue[i].data.ptr; + + switch (*t) { + + case WAKEUP_EVENT_SOURCE: { + sd_event_source *s = e->event_queue[i].data.ptr; + + assert(s); + + switch (s->type) { + + case SOURCE_IO: + r = process_io(e, s, e->event_queue[i].events); + break; + + case SOURCE_CHILD: + r = process_pidfd(e, s, e->event_queue[i].events); + break; + + default: + assert_not_reached("Unexpected event source type"); + } + + break; + } + + case WAKEUP_CLOCK_DATA: { + struct clock_data *d = e->event_queue[i].data.ptr; + + assert(d); + + r = flush_timer(e, d->fd, e->event_queue[i].events, &d->next); + break; + } + + case WAKEUP_SIGNAL_DATA: + r = process_signal(e, e->event_queue[i].data.ptr, e->event_queue[i].events); + break; + + case WAKEUP_INOTIFY_DATA: + r = event_inotify_data_read(e, e->event_queue[i].data.ptr, e->event_queue[i].events); + break; + + default: + assert_not_reached("Invalid wake-up pointer"); + } + } + if (r < 0) + goto finish; + } + + r = process_watchdog(e); + if (r < 0) + goto finish; + + r = process_timer(e, e->timestamp.realtime, &e->realtime); + if (r < 0) + goto finish; + + r = process_timer(e, e->timestamp.boottime, &e->boottime); + if (r < 0) + goto finish; + + r = process_timer(e, e->timestamp.monotonic, &e->monotonic); + if (r < 0) + goto finish; + + r = process_timer(e, e->timestamp.realtime, &e->realtime_alarm); + if (r < 0) + goto finish; + + r = process_timer(e, e->timestamp.boottime, &e->boottime_alarm); + if (r < 0) + goto finish; + + if (e->need_process_child) { + r = process_child(e); + if (r < 0) + goto finish; + } + + r = process_inotify(e); + if (r < 0) + goto finish; + + if (event_next_pending(e)) { + e->state = SD_EVENT_PENDING; + + return 1; + } + + r = 0; + +finish: + e->state = SD_EVENT_INITIAL; + + return r; +} + +_public_ int sd_event_dispatch(sd_event *e) { + sd_event_source *p; + 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); + + if (e->exit_requested) + return dispatch_exit(e); + + p = event_next_pending(e); + if (p) { + _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; + return r; + } + + e->state = SD_EVENT_INITIAL; + + return 1; +} + +static void event_log_delays(sd_event *e) { + char b[ELEMENTSOF(e->delays) * DECIMAL_STR_MAX(unsigned) + 1], *p; + size_t l, i; + + p = b; + l = sizeof(b); + for (i = 0; i < ELEMENTSOF(e->delays); i++) { + l = strpcpyf(&p, l, "%u ", e->delays[i]); + e->delays[i] = 0; + } + log_debug("Event loop iterations: %s", 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 < ELEMENTSOF(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); + if (r < 0) + return r; + + return 1; + } + + return r; +} + +_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); + + ref = sd_event_ref(e); + + while (e->state != SD_EVENT_FINISHED) { + r = sd_event_run(e, (uint64_t) -1); + if (r < 0) + 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; +} + +_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); + + if (!e->exit_requested) + return -ENODATA; + + *code = e->exit_code; + return 0; +} + +_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); + + e->exit_requested = true; + e->exit_code = code; + + return 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 (!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; + } + + *usec = triple_timestamp_by_clock(&e->timestamp, clock); + return 0; +} + +_public_ int sd_event_default(sd_event **ret) { + sd_event *e = NULL; + int r; + + if (!ret) + return !!default_event; + + if (default_event) { + *ret = sd_event_ref(default_event); + return 0; + } + + r = sd_event_new(&e); + if (r < 0) + return r; + + e->default_event_ptr = &default_event; + e->tid = gettid(); + default_event = e; + + *ret = e; + return 1; +} + +_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); + + if (e->tid != 0) { + *tid = e->tid; + return 0; + } + + return -ENXIO; +} + +_public_ int sd_event_set_watchdog(sd_event *e, int b) { + 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) { + r = sd_watchdog_enabled(false, &e->watchdog_period); + if (r <= 0) + return r; + + /* Issue first ping immediately */ + sd_notify(false, "WATCHDOG=1"); + e->watchdog_last = now(CLOCK_MONOTONIC); + + e->watchdog_fd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK|TFD_CLOEXEC); + if (e->watchdog_fd < 0) + return -errno; + + r = arm_watchdog(e); + if (r < 0) + goto fail; + + struct epoll_event ev = { + .events = EPOLLIN, + .data.ptr = INT_TO_PTR(SOURCE_WATCHDOG), + }; + + if (epoll_ctl(e->epoll_fd, EPOLL_CTL_ADD, e->watchdog_fd, &ev) < 0) { + r = -errno; + goto fail; + } + + } else { + if (e->watchdog_fd >= 0) { + (void) epoll_ctl(e->epoll_fd, EPOLL_CTL_DEL, e->watchdog_fd, NULL); + e->watchdog_fd = safe_close(e->watchdog_fd); + } + } + + e->watchdog = !!b; + return e->watchdog; + +fail: + e->watchdog_fd = safe_close(e->watchdog_fd); + return r; +} + +_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; +} + +_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; +} + +_public_ int sd_event_source_get_floating(sd_event_source *s) { + assert_return(s, -EINVAL); + + return s->floating; +} + +_public_ int sd_event_source_set_floating(sd_event_source *s, int b) { + assert_return(s, -EINVAL); + + if (s->floating == !!b) + return 0; + + if (!s->event) /* Already disconnected */ + return -ESTALE; + + s->floating = b; + + if (b) { + sd_event_source_ref(s); + sd_event_unref(s->event); + } else { + sd_event_ref(s->event); + sd_event_source_unref(s); + } + + return 1; +} + +_public_ int sd_event_source_get_exit_on_failure(sd_event_source *s) { + assert_return(s, -EINVAL); + assert_return(s->type != SOURCE_EXIT, -EDOM); + + return s->exit_on_failure; +} + +_public_ int sd_event_source_set_exit_on_failure(sd_event_source *s, int b) { + assert_return(s, -EINVAL); + assert_return(s->type != SOURCE_EXIT, -EDOM); + + if (s->exit_on_failure == !!b) + return 0; + + s->exit_on_failure = b; + return 1; +} diff --git a/src/libsystemd/sd-event/test-event.c b/src/libsystemd/sd-event/test-event.c new file mode 100644 index 0000000..1c4d0e2 --- /dev/null +++ b/src/libsystemd/sd-event/test-event.c @@ -0,0 +1,607 @@ +/* SPDX-License-Identifier: LGPL-2.1-or-later */ + +#include + +#include "sd-event.h" + +#include "alloc-util.h" +#include "fd-util.h" +#include "fs-util.h" +#include "log.h" +#include "macro.h" +#include "missing_syscall.h" +#include "parse-util.h" +#include "path-util.h" +#include "process-util.h" +#include "rm-rf.h" +#include "signal-util.h" +#include "stdio-util.h" +#include "string-util.h" +#include "tests.h" +#include "tmpfile-util.h" +#include "util.h" + +static int prepare_handler(sd_event_source *s, void *userdata) { + log_info("preparing %c", PTR_TO_INT(userdata)); + return 1; +} + +static bool got_a, got_b, got_c, got_unref; +static unsigned got_d; + +static int unref_handler(sd_event_source *s, int fd, uint32_t revents, void *userdata) { + sd_event_source_unref(s); + got_unref = true; + return 0; +} + +static int io_handler(sd_event_source *s, int fd, uint32_t revents, void *userdata) { + + log_info("got IO on %c", PTR_TO_INT(userdata)); + + if (userdata == INT_TO_PTR('a')) { + assert_se(sd_event_source_set_enabled(s, SD_EVENT_OFF) >= 0); + assert_se(!got_a); + got_a = true; + } else if (userdata == INT_TO_PTR('b')) { + assert_se(!got_b); + got_b = true; + } else if (userdata == INT_TO_PTR('d')) { + got_d++; + if (got_d < 2) + assert_se(sd_event_source_set_enabled(s, SD_EVENT_ONESHOT) >= 0); + else + assert_se(sd_event_source_set_enabled(s, SD_EVENT_OFF) >= 0); + } else + assert_not_reached("Yuck!"); + + return 1; +} + +static int child_handler(sd_event_source *s, const siginfo_t *si, void *userdata) { + + assert_se(s); + assert_se(si); + + assert_se(si->si_uid == getuid()); + assert_se(si->si_signo == SIGCHLD); + assert_se(si->si_code == CLD_EXITED); + assert_se(si->si_status == 78); + + log_info("got child on %c", PTR_TO_INT(userdata)); + + assert_se(userdata == INT_TO_PTR('f')); + + assert_se(sd_event_exit(sd_event_source_get_event(s), 0) >= 0); + sd_event_source_unref(s); + + return 1; +} + +static int signal_handler(sd_event_source *s, const struct signalfd_siginfo *si, void *userdata) { + sd_event_source *p = NULL; + pid_t pid; + siginfo_t plain_si; + + assert_se(s); + assert_se(si); + + log_info("got signal on %c", PTR_TO_INT(userdata)); + + assert_se(userdata == INT_TO_PTR('e')); + + assert_se(sigprocmask_many(SIG_BLOCK, NULL, SIGCHLD, SIGUSR2, -1) >= 0); + + pid = fork(); + assert_se(pid >= 0); + + if (pid == 0) { + sigset_t ss; + + assert_se(sigemptyset(&ss) >= 0); + assert_se(sigaddset(&ss, SIGUSR2) >= 0); + + zero(plain_si); + assert_se(sigwaitinfo(&ss, &plain_si) >= 0); + + assert_se(plain_si.si_signo == SIGUSR2); + assert_se(plain_si.si_value.sival_int == 4711); + + _exit(78); + } + + assert_se(sd_event_add_child(sd_event_source_get_event(s), &p, pid, WEXITED, child_handler, INT_TO_PTR('f')) >= 0); + assert_se(sd_event_source_set_enabled(p, SD_EVENT_ONESHOT) >= 0); + assert_se(sd_event_source_set_child_process_own(p, true) >= 0); + + /* We can't use structured initialization here, since the structure contains various unions and these + * fields lie in overlapping (carefully aligned) unions that LLVM is allergic to allow assignments + * to */ + zero(plain_si); + plain_si.si_signo = SIGUSR2; + plain_si.si_code = SI_QUEUE; + plain_si.si_pid = getpid(); + plain_si.si_uid = getuid(); + plain_si.si_value.sival_int = 4711; + + assert_se(sd_event_source_send_child_signal(p, SIGUSR2, &plain_si, 0) >= 0); + + sd_event_source_unref(s); + + return 1; +} + +static int defer_handler(sd_event_source *s, void *userdata) { + sd_event_source *p = NULL; + + assert_se(s); + + log_info("got defer on %c", PTR_TO_INT(userdata)); + + assert_se(userdata == INT_TO_PTR('d')); + + assert_se(sigprocmask_many(SIG_BLOCK, NULL, SIGUSR1, -1) >= 0); + + assert_se(sd_event_add_signal(sd_event_source_get_event(s), &p, SIGUSR1, signal_handler, INT_TO_PTR('e')) >= 0); + assert_se(sd_event_source_set_enabled(p, SD_EVENT_ONESHOT) >= 0); + raise(SIGUSR1); + + sd_event_source_unref(s); + + return 1; +} + +static bool do_quit; + +static int time_handler(sd_event_source *s, uint64_t usec, void *userdata) { + log_info("got timer on %c", PTR_TO_INT(userdata)); + + if (userdata == INT_TO_PTR('c')) { + + if (do_quit) { + sd_event_source *p; + + assert_se(sd_event_add_defer(sd_event_source_get_event(s), &p, defer_handler, INT_TO_PTR('d')) >= 0); + assert_se(sd_event_source_set_enabled(p, SD_EVENT_ONESHOT) >= 0); + } else { + assert_se(!got_c); + got_c = true; + } + } else + assert_not_reached("Huh?"); + + return 2; +} + +static bool got_exit = false; + +static int exit_handler(sd_event_source *s, void *userdata) { + log_info("got quit handler on %c", PTR_TO_INT(userdata)); + + got_exit = true; + + return 3; +} + +static bool got_post = false; + +static int post_handler(sd_event_source *s, void *userdata) { + log_info("got post handler"); + + got_post = true; + + return 2; +} + +static void test_basic(bool with_pidfd) { + sd_event *e = NULL; + sd_event_source *w = NULL, *x = NULL, *y = NULL, *z = NULL, *q = NULL, *t = NULL; + static const char ch = 'x'; + int a[2] = { -1, -1 }, b[2] = { -1, -1}, d[2] = { -1, -1}, k[2] = { -1, -1 }; + uint64_t event_now; + int64_t priority; + + assert_se(setenv("SYSTEMD_PIDFD", yes_no(with_pidfd), 1) >= 0); + + assert_se(pipe(a) >= 0); + assert_se(pipe(b) >= 0); + assert_se(pipe(d) >= 0); + assert_se(pipe(k) >= 0); + + assert_se(sd_event_default(&e) >= 0); + assert_se(sd_event_now(e, CLOCK_MONOTONIC, &event_now) > 0); + + assert_se(sd_event_set_watchdog(e, true) >= 0); + + /* Test whether we cleanly can destroy an io event source from its own handler */ + got_unref = false; + assert_se(sd_event_add_io(e, &t, k[0], EPOLLIN, unref_handler, NULL) >= 0); + assert_se(write(k[1], &ch, 1) == 1); + assert_se(sd_event_run(e, (uint64_t) -1) >= 1); + assert_se(got_unref); + + got_a = false, got_b = false, got_c = false, got_d = 0; + + /* Add a oneshot handler, trigger it, reenable it, and trigger + * it again. */ + assert_se(sd_event_add_io(e, &w, d[0], EPOLLIN, io_handler, INT_TO_PTR('d')) >= 0); + assert_se(sd_event_source_set_enabled(w, SD_EVENT_ONESHOT) >= 0); + assert_se(write(d[1], &ch, 1) >= 0); + assert_se(sd_event_run(e, (uint64_t) -1) >= 1); + assert_se(got_d == 1); + assert_se(write(d[1], &ch, 1) >= 0); + assert_se(sd_event_run(e, (uint64_t) -1) >= 1); + assert_se(got_d == 2); + + assert_se(sd_event_add_io(e, &x, a[0], EPOLLIN, io_handler, INT_TO_PTR('a')) >= 0); + assert_se(sd_event_add_io(e, &y, b[0], EPOLLIN, io_handler, INT_TO_PTR('b')) >= 0); + + do_quit = false; + assert_se(sd_event_add_time(e, &z, CLOCK_MONOTONIC, 0, 0, time_handler, INT_TO_PTR('c')) >= 0); + assert_se(sd_event_add_exit(e, &q, exit_handler, INT_TO_PTR('g')) >= 0); + + assert_se(sd_event_source_set_priority(x, 99) >= 0); + assert_se(sd_event_source_get_priority(x, &priority) >= 0); + assert_se(priority == 99); + assert_se(sd_event_source_set_enabled(y, SD_EVENT_ONESHOT) >= 0); + assert_se(sd_event_source_set_prepare(x, prepare_handler) >= 0); + assert_se(sd_event_source_set_priority(z, 50) >= 0); + assert_se(sd_event_source_set_enabled(z, SD_EVENT_ONESHOT) >= 0); + assert_se(sd_event_source_set_prepare(z, prepare_handler) >= 0); + + /* Test for floating event sources */ + assert_se(sigprocmask_many(SIG_BLOCK, NULL, SIGRTMIN+1, -1) >= 0); + assert_se(sd_event_add_signal(e, NULL, SIGRTMIN+1, NULL, NULL) >= 0); + + assert_se(write(a[1], &ch, 1) >= 0); + assert_se(write(b[1], &ch, 1) >= 0); + + assert_se(!got_a && !got_b && !got_c); + + assert_se(sd_event_run(e, (uint64_t) -1) >= 1); + + assert_se(!got_a && got_b && !got_c); + + assert_se(sd_event_run(e, (uint64_t) -1) >= 1); + + assert_se(!got_a && got_b && got_c); + + assert_se(sd_event_run(e, (uint64_t) -1) >= 1); + + assert_se(got_a && got_b && got_c); + + sd_event_source_unref(x); + sd_event_source_unref(y); + + do_quit = true; + assert_se(sd_event_add_post(e, NULL, post_handler, NULL) >= 0); + assert_se(sd_event_now(e, CLOCK_MONOTONIC, &event_now) == 0); + assert_se(sd_event_source_set_time(z, event_now + 200 * USEC_PER_MSEC) >= 0); + assert_se(sd_event_source_set_enabled(z, SD_EVENT_ONESHOT) >= 0); + + assert_se(sd_event_loop(e) >= 0); + assert_se(got_post); + assert_se(got_exit); + + sd_event_source_unref(z); + sd_event_source_unref(q); + + sd_event_source_unref(w); + + sd_event_unref(e); + + safe_close_pair(a); + safe_close_pair(b); + safe_close_pair(d); + safe_close_pair(k); + + assert_se(unsetenv("SYSTEMD_PIDFD") >= 0); +} + +static void test_sd_event_now(void) { + _cleanup_(sd_event_unrefp) sd_event *e = NULL; + uint64_t event_now; + + assert_se(sd_event_new(&e) >= 0); + assert_se(sd_event_now(e, CLOCK_MONOTONIC, &event_now) > 0); + assert_se(sd_event_now(e, CLOCK_REALTIME, &event_now) > 0); + assert_se(sd_event_now(e, CLOCK_REALTIME_ALARM, &event_now) > 0); + if (clock_boottime_supported()) { + assert_se(sd_event_now(e, CLOCK_BOOTTIME, &event_now) > 0); + assert_se(sd_event_now(e, CLOCK_BOOTTIME_ALARM, &event_now) > 0); + } + assert_se(sd_event_now(e, -1, &event_now) == -EOPNOTSUPP); + assert_se(sd_event_now(e, 900 /* arbitrary big number */, &event_now) == -EOPNOTSUPP); + + assert_se(sd_event_run(e, 0) == 0); + + assert_se(sd_event_now(e, CLOCK_MONOTONIC, &event_now) == 0); + assert_se(sd_event_now(e, CLOCK_REALTIME, &event_now) == 0); + assert_se(sd_event_now(e, CLOCK_REALTIME_ALARM, &event_now) == 0); + if (clock_boottime_supported()) { + assert_se(sd_event_now(e, CLOCK_BOOTTIME, &event_now) == 0); + assert_se(sd_event_now(e, CLOCK_BOOTTIME_ALARM, &event_now) == 0); + } + assert_se(sd_event_now(e, -1, &event_now) == -EOPNOTSUPP); + assert_se(sd_event_now(e, 900 /* arbitrary big number */, &event_now) == -EOPNOTSUPP); +} + +static int last_rtqueue_sigval = 0; +static int n_rtqueue = 0; + +static int rtqueue_handler(sd_event_source *s, const struct signalfd_siginfo *si, void *userdata) { + last_rtqueue_sigval = si->ssi_int; + n_rtqueue++; + return 0; +} + +static void test_rtqueue(void) { + sd_event_source *u = NULL, *v = NULL, *s = NULL; + sd_event *e = NULL; + + assert_se(sd_event_default(&e) >= 0); + + assert_se(sigprocmask_many(SIG_BLOCK, NULL, SIGRTMIN+2, SIGRTMIN+3, SIGUSR2, -1) >= 0); + assert_se(sd_event_add_signal(e, &u, SIGRTMIN+2, rtqueue_handler, NULL) >= 0); + assert_se(sd_event_add_signal(e, &v, SIGRTMIN+3, rtqueue_handler, NULL) >= 0); + assert_se(sd_event_add_signal(e, &s, SIGUSR2, rtqueue_handler, NULL) >= 0); + + assert_se(sd_event_source_set_priority(v, -10) >= 0); + + assert_se(sigqueue(getpid_cached(), SIGRTMIN+2, (union sigval) { .sival_int = 1 }) >= 0); + assert_se(sigqueue(getpid_cached(), SIGRTMIN+3, (union sigval) { .sival_int = 2 }) >= 0); + assert_se(sigqueue(getpid_cached(), SIGUSR2, (union sigval) { .sival_int = 3 }) >= 0); + assert_se(sigqueue(getpid_cached(), SIGRTMIN+3, (union sigval) { .sival_int = 4 }) >= 0); + assert_se(sigqueue(getpid_cached(), SIGUSR2, (union sigval) { .sival_int = 5 }) >= 0); + + assert_se(n_rtqueue == 0); + assert_se(last_rtqueue_sigval == 0); + + assert_se(sd_event_run(e, (uint64_t) -1) >= 1); + assert_se(n_rtqueue == 1); + assert_se(last_rtqueue_sigval == 2); /* first SIGRTMIN+3 */ + + assert_se(sd_event_run(e, (uint64_t) -1) >= 1); + assert_se(n_rtqueue == 2); + assert_se(last_rtqueue_sigval == 4); /* second SIGRTMIN+3 */ + + assert_se(sd_event_run(e, (uint64_t) -1) >= 1); + assert_se(n_rtqueue == 3); + assert_se(last_rtqueue_sigval == 3); /* first SIGUSR2 */ + + assert_se(sd_event_run(e, (uint64_t) -1) >= 1); + assert_se(n_rtqueue == 4); + assert_se(last_rtqueue_sigval == 1); /* SIGRTMIN+2 */ + + assert_se(sd_event_run(e, 0) == 0); /* the other SIGUSR2 is dropped, because the first one was still queued */ + assert_se(n_rtqueue == 4); + assert_se(last_rtqueue_sigval == 1); + + sd_event_source_unref(u); + sd_event_source_unref(v); + sd_event_source_unref(s); + + sd_event_unref(e); +} + +#define CREATE_EVENTS_MAX (70000U) + +struct inotify_context { + bool delete_self_handler_called; + unsigned create_called[CREATE_EVENTS_MAX]; + unsigned create_overflow; + unsigned n_create_events; +}; + +static void maybe_exit(sd_event_source *s, struct inotify_context *c) { + unsigned n; + + assert(s); + assert(c); + + if (!c->delete_self_handler_called) + return; + + for (n = 0; n < 3; n++) { + unsigned i; + + if (c->create_overflow & (1U << n)) + continue; + + for (i = 0; i < c->n_create_events; i++) + if (!(c->create_called[i] & (1U << n))) + return; + } + + sd_event_exit(sd_event_source_get_event(s), 0); +} + +static int inotify_handler(sd_event_source *s, const struct inotify_event *ev, void *userdata) { + struct inotify_context *c = userdata; + const char *description; + unsigned bit, n; + + assert_se(sd_event_source_get_description(s, &description) >= 0); + assert_se(safe_atou(description, &n) >= 0); + + assert_se(n <= 3); + bit = 1U << n; + + if (ev->mask & IN_Q_OVERFLOW) { + log_info("inotify-handler <%s>: overflow", description); + c->create_overflow |= bit; + } else if (ev->mask & IN_CREATE) { + unsigned i; + + log_debug("inotify-handler <%s>: create on %s", description, ev->name); + + if (!streq(ev->name, "sub")) { + assert_se(safe_atou(ev->name, &i) >= 0); + + assert_se(i < c->n_create_events); + c->create_called[i] |= bit; + } + } else if (ev->mask & IN_DELETE) { + log_info("inotify-handler <%s>: delete of %s", description, ev->name); + assert_se(streq(ev->name, "sub")); + } else + assert_not_reached("unexpected inotify event"); + + maybe_exit(s, c); + return 1; +} + +static int delete_self_handler(sd_event_source *s, const struct inotify_event *ev, void *userdata) { + struct inotify_context *c = userdata; + + if (ev->mask & IN_Q_OVERFLOW) { + log_info("delete-self-handler: overflow"); + c->delete_self_handler_called = true; + } else if (ev->mask & IN_DELETE_SELF) { + log_info("delete-self-handler: delete-self"); + c->delete_self_handler_called = true; + } else if (ev->mask & IN_IGNORED) { + log_info("delete-self-handler: ignore"); + } else + assert_not_reached("unexpected inotify event (delete-self)"); + + maybe_exit(s, c); + return 1; +} + +static void test_inotify(unsigned n_create_events) { + _cleanup_(rm_rf_physical_and_freep) char *p = NULL; + sd_event_source *a = NULL, *b = NULL, *c = NULL, *d = NULL; + struct inotify_context context = { + .n_create_events = n_create_events, + }; + sd_event *e = NULL; + const char *q; + unsigned i; + + assert_se(sd_event_default(&e) >= 0); + + assert_se(mkdtemp_malloc("/tmp/test-inotify-XXXXXX", &p) >= 0); + + assert_se(sd_event_add_inotify(e, &a, p, IN_CREATE|IN_ONLYDIR, inotify_handler, &context) >= 0); + assert_se(sd_event_add_inotify(e, &b, p, IN_CREATE|IN_DELETE|IN_DONT_FOLLOW, inotify_handler, &context) >= 0); + assert_se(sd_event_source_set_priority(b, SD_EVENT_PRIORITY_IDLE) >= 0); + assert_se(sd_event_source_set_priority(b, SD_EVENT_PRIORITY_NORMAL) >= 0); + assert_se(sd_event_add_inotify(e, &c, p, IN_CREATE|IN_DELETE|IN_EXCL_UNLINK, inotify_handler, &context) >= 0); + assert_se(sd_event_source_set_priority(c, SD_EVENT_PRIORITY_IDLE) >= 0); + + assert_se(sd_event_source_set_description(a, "0") >= 0); + assert_se(sd_event_source_set_description(b, "1") >= 0); + assert_se(sd_event_source_set_description(c, "2") >= 0); + + q = strjoina(p, "/sub"); + assert_se(touch(q) >= 0); + assert_se(sd_event_add_inotify(e, &d, q, IN_DELETE_SELF, delete_self_handler, &context) >= 0); + + for (i = 0; i < n_create_events; i++) { + char buf[DECIMAL_STR_MAX(unsigned)+1]; + _cleanup_free_ char *z; + + xsprintf(buf, "%u", i); + assert_se(z = path_join(p, buf)); + + assert_se(touch(z) >= 0); + } + + assert_se(unlink(q) >= 0); + + assert_se(sd_event_loop(e) >= 0); + + sd_event_source_unref(a); + sd_event_source_unref(b); + sd_event_source_unref(c); + sd_event_source_unref(d); + + sd_event_unref(e); +} + +static int pidfd_handler(sd_event_source *s, const siginfo_t *si, void *userdata) { + assert_se(s); + assert_se(si); + + assert_se(si->si_uid == getuid()); + assert_se(si->si_signo == SIGCHLD); + assert_se(si->si_code == CLD_EXITED); + assert_se(si->si_status == 66); + + log_info("got pidfd on %c", PTR_TO_INT(userdata)); + + assert_se(userdata == INT_TO_PTR('p')); + + assert_se(sd_event_exit(sd_event_source_get_event(s), 0) >= 0); + sd_event_source_unref(s); + + return 0; +} + +static void test_pidfd(void) { + sd_event_source *s = NULL, *t = NULL; + sd_event *e = NULL; + int pidfd; + pid_t pid, pid2; + + assert_se(sigprocmask_many(SIG_BLOCK, NULL, SIGCHLD, -1) >= 0); + + pid = fork(); + if (pid == 0) + /* child */ + _exit(66); + + assert_se(pid > 1); + + pidfd = pidfd_open(pid, 0); + if (pidfd < 0) { + /* No pidfd_open() supported or blocked? */ + assert_se(ERRNO_IS_NOT_SUPPORTED(errno) || ERRNO_IS_PRIVILEGE(errno)); + (void) wait_for_terminate(pid, NULL); + return; + } + + pid2 = fork(); + if (pid2 == 0) + freeze(); + + assert_se(pid > 2); + + assert_se(sd_event_default(&e) >= 0); + assert_se(sd_event_add_child_pidfd(e, &s, pidfd, WEXITED, pidfd_handler, INT_TO_PTR('p')) >= 0); + assert_se(sd_event_source_set_child_pidfd_own(s, true) >= 0); + + /* This one should never trigger, since our second child lives forever */ + assert_se(sd_event_add_child(e, &t, pid2, WEXITED, pidfd_handler, INT_TO_PTR('q')) >= 0); + assert_se(sd_event_source_set_child_process_own(t, true) >= 0); + + assert_se(sd_event_loop(e) >= 0); + + /* Child should still be alive */ + assert_se(kill(pid2, 0) >= 0); + + t = sd_event_source_unref(t); + + /* Child should now be dead, since we dropped the ref */ + assert_se(kill(pid2, 0) < 0 && errno == ESRCH); + + sd_event_unref(e); +} + +int main(int argc, char *argv[]) { + test_setup_logging(LOG_INFO); + + test_basic(true); /* test with pidfd */ + test_basic(false); /* test without pidfd */ + + test_sd_event_now(); + test_rtqueue(); + + test_inotify(100); /* should work without overflow */ + test_inotify(33000); /* should trigger a q overflow */ + + test_pidfd(); + + return 0; +} -- cgit v1.2.3