const DynamicCreds *dcreds,
int user_lookup_fd,
int socket_fd,
+ int exec_fd,
int *fds, size_t n_fds) {
size_t n_dont_close = 0;
if (socket_fd >= 0)
dont_close[n_dont_close++] = socket_fd;
+ if (exec_fd >= 0)
+ dont_close[n_dont_close++] = exec_fd;
if (n_fds > 0) {
memcpy(dont_close + n_dont_close, fds, sizeof(int) * n_fds);
n_dont_close += n_fds;
int *exit_status) {
_cleanup_strv_free_ char **our_env = NULL, **pass_env = NULL, **accum_env = NULL, **final_argv = NULL;
- _cleanup_free_ char *home_buffer = NULL;
+ int *fds_with_exec_fd, n_fds_with_exec_fd, r, ngids = 0, exec_fd = -1;
_cleanup_free_ gid_t *supplementary_gids = NULL;
const char *username = NULL, *groupname = NULL;
+ _cleanup_free_ char *home_buffer = NULL;
const char *home = NULL, *shell = NULL;
dev_t journal_stream_dev = 0;
ino_t journal_stream_ino = 0;
#endif
uid_t uid = UID_INVALID;
gid_t gid = GID_INVALID;
- int r, ngids = 0;
size_t n_fds;
ExecDirectoryType dt;
int secure_bits;
/* In case anything used libc syslog(), close this here, too */
closelog();
- n_fds = n_storage_fds + n_socket_fds;
- r = close_remaining_fds(params, runtime, dcreds, user_lookup_fd, socket_fd, fds, n_fds);
+ n_fds = n_socket_fds + n_storage_fds;
+ r = close_remaining_fds(params, runtime, dcreds, user_lookup_fd, socket_fd, params->exec_fd, fds, n_fds);
if (r < 0) {
*exit_status = EXIT_FDS;
return log_unit_error_errno(unit, r, "Failed to close unwanted file descriptors: %m");
}
/* We repeat the fd closing here, to make sure that nothing is leaked from the PAM modules. Note that we are
- * more aggressive this time since socket_fd and the netns fds we don't need anymore. The custom endpoint fd
- * was needed to upload the policy and can now be closed as well. */
- r = close_all_fds(fds, n_fds);
+ * more aggressive this time since socket_fd and the netns fds we don't need anymore. We do keep the exec_fd
+ * however if we have it as we want to keep it open until the final execve(). */
+
+ if (params->exec_fd >= 0) {
+ exec_fd = params->exec_fd;
+
+ if (exec_fd < 3 + (int) n_fds) {
+ int moved_fd;
+
+ /* Let's move the exec fd far up, so that it's outside of the fd range we want to pass to the
+ * process we are about to execute. */
+
+ moved_fd = fcntl(exec_fd, F_DUPFD_CLOEXEC, 3 + (int) n_fds);
+ if (moved_fd < 0) {
+ *exit_status = EXIT_FDS;
+ return log_unit_error_errno(unit, errno, "Couldn't move exec fd up: %m");
+ }
+
+ safe_close(exec_fd);
+ exec_fd = moved_fd;
+ } else {
+ /* This fd should be FD_CLOEXEC already, but let's make sure. */
+ r = fd_cloexec(exec_fd, true);
+ if (r < 0) {
+ *exit_status = EXIT_FDS;
+ return log_unit_error_errno(unit, r, "Failed to make exec fd FD_CLOEXEC: %m");
+ }
+ }
+
+ fds_with_exec_fd = newa(int, n_fds + 1);
+ memcpy(fds_with_exec_fd, fds, n_fds * sizeof(int));
+ fds_with_exec_fd[n_fds] = exec_fd;
+ n_fds_with_exec_fd = n_fds + 1;
+ } else {
+ fds_with_exec_fd = fds;
+ n_fds_with_exec_fd = n_fds;
+ }
+
+ r = close_all_fds(fds_with_exec_fd, n_fds_with_exec_fd);
if (r >= 0)
r = shift_fds(fds, n_fds);
if (r >= 0)
return log_unit_error_errno(unit, r, "Failed to adjust passed file descriptors: %m");
}
+ /* At this point, the fds we want to pass to the program are all ready and set up, with O_CLOEXEC turned off
+ * and at the right fd numbers. The are no other fds open, with one exception: the exec_fd if it is defined,
+ * and it has O_CLOEXEC set, after all we want it to be closed by the execve(), so that our parent knows we
+ * came this far. */
+
secure_bits = context->secure_bits;
if (needs_sandboxing) {
LOG_UNIT_INVOCATION_ID(unit));
}
+ if (exec_fd >= 0) {
+ uint8_t hot = 1;
+
+ /* We have finished with all our initializations. Let's now let the manager know that. From this point
+ * on, if the manager sees POLLHUP on the exec_fd, then execve() was successful. */
+
+ if (write(exec_fd, &hot, sizeof(hot)) < 0) {
+ *exit_status = EXIT_EXEC;
+ return log_unit_error_errno(unit, errno, "Failed to enable exec_fd: %m");
+ }
+ }
+
execve(command->path, final_argv, accum_env);
+ r = -errno;
+
+ if (exec_fd >= 0) {
+ uint8_t hot = 0;
+
+ /* The execve() failed. This means the exec_fd is still open. Which means we need to tell the manager
+ * that POLLHUP on it no longer means execve() succeeded. */
+
+ if (write(exec_fd, &hot, sizeof(hot)) < 0) {
+ *exit_status = EXIT_EXEC;
+ return log_unit_error_errno(unit, errno, "Failed to disable exec_fd: %m");
+ }
+ }
- if (errno == ENOENT && (command->flags & EXEC_COMMAND_IGNORE_FAILURE)) {
- log_struct_errno(LOG_INFO, errno,
+ if (r == -ENOENT && (command->flags & EXEC_COMMAND_IGNORE_FAILURE)) {
+ log_struct_errno(LOG_INFO, r,
"MESSAGE_ID=" SD_MESSAGE_SPAWN_FAILED_STR,
LOG_UNIT_ID(unit),
LOG_UNIT_INVOCATION_ID(unit),
}
*exit_status = EXIT_EXEC;
- return log_unit_error_errno(unit, errno, "Failed to execute command: %m");
+ return log_unit_error_errno(unit, r, "Failed to execute command: %m");
}
static int exec_context_load_environment(const Unit *unit, const ExecContext *c, char ***l);
[SERVICE_AUTO_RESTART] = UNIT_ACTIVATING
};
-static int service_dispatch_io(sd_event_source *source, int fd, uint32_t events, void *userdata);
+static int service_dispatch_inotify_io(sd_event_source *source, int fd, uint32_t events, void *userdata);
static int service_dispatch_timer(sd_event_source *source, usec_t usec, void *userdata);
static int service_dispatch_watchdog(sd_event_source *source, usec_t usec, void *userdata);
+static int service_dispatch_exec_io(sd_event_source *source, int fd, uint32_t events, void *userdata);
static void service_enter_signal(Service *s, ServiceState state, ServiceResult f);
static void service_enter_reload_by_notify(Service *s);
service_stop_watchdog(s);
s->timer_event_source = sd_event_source_unref(s->timer_event_source);
+ s->exec_fd_event_source = sd_event_source_unref(s->exec_fd_event_source);
service_release_resources(u);
}
!(state == SERVICE_DEAD && UNIT(s)->job))
service_close_socket_fd(s);
+ if (state != SERVICE_START)
+ s->exec_fd_event_source = sd_event_source_unref(s->exec_fd_event_source);
+
if (!IN_SET(state, SERVICE_START_POST, SERVICE_RUNNING, SERVICE_RELOAD))
service_stop_watchdog(s);
return 0;
}
+static int service_allocate_exec_fd_event_source(
+ Service *s,
+ int fd,
+ sd_event_source **ret_event_source) {
+
+ _cleanup_(sd_event_source_unrefp) sd_event_source *source = NULL;
+ int r;
+
+ assert(s);
+ assert(fd >= 0);
+ assert(ret_event_source);
+
+ r = sd_event_add_io(UNIT(s)->manager->event, &source, fd, 0, service_dispatch_exec_io, s);
+ if (r < 0)
+ return log_unit_error_errno(UNIT(s), r, "Failed to allocate exec_fd event source: %m");
+
+ /* This is a bit lower priority than SIGCHLD, as that carries a lot more interesting failure information */
+
+ r = sd_event_source_set_priority(source, SD_EVENT_PRIORITY_NORMAL-3);
+ if (r < 0)
+ return log_unit_error_errno(UNIT(s), r, "Failed to adjust priority of exec_fd event source: %m");
+
+ (void) sd_event_source_set_description(source, "service event_fd");
+
+ r = sd_event_source_set_io_fd_own(source, true);
+ if (r < 0)
+ return log_unit_error_errno(UNIT(s), r, "Failed to pass ownership of fd to event source: %m");
+
+ *ret_event_source = TAKE_PTR(source);
+ return 0;
+}
+
+static int service_allocate_exec_fd(
+ Service *s,
+ sd_event_source **ret_event_source,
+ int* ret_exec_fd) {
+
+ _cleanup_close_pair_ int p[2] = { -1, -1 };
+ int r;
+
+ assert(s);
+ assert(ret_event_source);
+ assert(ret_exec_fd);
+
+ if (pipe2(p, O_CLOEXEC|O_NONBLOCK) < 0)
+ return log_unit_error_errno(UNIT(s), errno, "Failed to allocate exec_fd pipe: %m");
+
+ r = service_allocate_exec_fd_event_source(s, p[0], ret_event_source);
+ if (r < 0)
+ return r;
+
+ p[0] = -1;
+ *ret_exec_fd = TAKE_FD(p[1]);
+
+ return 0;
+}
+
static bool service_exec_needs_notify_socket(Service *s, ExecFlags flags) {
assert(s);
.exec_fd = -1,
};
_cleanup_strv_free_ char **final_env = NULL, **our_env = NULL, **fd_names = NULL;
+ _cleanup_(sd_event_source_unrefp) sd_event_source *exec_fd_source = NULL;
size_t n_socket_fds = 0, n_storage_fds = 0, n_env = 0;
+ _cleanup_close_ int exec_fd = -1;
_cleanup_free_ int *fds = NULL;
pid_t pid;
int r;
log_unit_debug(UNIT(s), "Passing %zu fds to service", n_socket_fds + n_storage_fds);
}
+ if (!FLAGS_SET(flags, EXEC_IS_CONTROL) && s->type == SERVICE_EXEC) {
+ assert(!s->exec_fd_event_source);
+
+ r = service_allocate_exec_fd(s, &exec_fd_source, &exec_fd);
+ if (r < 0)
+ return r;
+ }
+
r = service_arm_timer(s, usec_add(now(CLOCK_MONOTONIC), timeout));
if (r < 0)
return r;
exec_params.stdin_fd = s->stdin_fd;
exec_params.stdout_fd = s->stdout_fd;
exec_params.stderr_fd = s->stderr_fd;
+ exec_params.exec_fd = exec_fd;
r = exec_spawn(UNIT(s),
c,
if (r < 0)
return r;
+ s->exec_fd_event_source = TAKE_PTR(exec_fd_source);
+ s->exec_fd_hot = false;
+
r = unit_watch_pid(UNIT(s), pid);
if (r < 0) /* FIXME: we need to do something here */
return r;
s->control_pid = pid;
service_set_state(s, SERVICE_START);
- } else if (IN_SET(s->type, SERVICE_ONESHOT, SERVICE_DBUS, SERVICE_NOTIFY)) {
+ } else if (IN_SET(s->type, SERVICE_ONESHOT, SERVICE_DBUS, SERVICE_NOTIFY, SERVICE_EXEC)) {
- /* For oneshot services we wait until the start
- * process exited, too, but it is our main process. */
+ /* For oneshot services we wait until the start process exited, too, but it is our main process. */
- /* For D-Bus services we know the main pid right away,
- * but wait for the bus name to appear on the
- * bus. Notify services are similar. */
+ /* For D-Bus services we know the main pid right away, but wait for the bus name to appear on the
+ * bus. 'notify' and 'exec' services are similar. */
service_set_main_pid(s, pid);
service_set_state(s, SERVICE_START);
if (r < 0)
return r;
+ if (s->exec_fd_event_source) {
+ r = unit_serialize_item_fd(u, f, fds, "exec-fd", sd_event_source_get_io_fd(s->exec_fd_event_source));
+ if (r < 0)
+ return r;
+ unit_serialize_item(u, f, "exec-fd-hot", yes_no(s->exec_fd_hot));
+ }
+
if (UNIT_ISSET(s->accept_socket)) {
r = unit_serialize_item(u, f, "accept-socket", UNIT_DEREF(s->accept_socket)->id);
if (r < 0)
s->stderr_fd = fdset_remove(fds, fd);
s->exec_context.stdio_as_fds = true;
}
+ } else if (streq(key, "exec-fd")) {
+ int fd;
+
+ if (safe_atoi(value, &fd) < 0 || fd < 0 || !fdset_contains(fds, fd))
+ log_unit_debug(u, "Failed to parse exec-fd value: %s", value);
+ else {
+ s->exec_fd_event_source = sd_event_source_unref(s->exec_fd_event_source);
+
+ fd = fdset_remove(fds, fd);
+ if (service_allocate_exec_fd_event_source(s, fd, &s->exec_fd_event_source) < 0)
+ safe_close(fd);
+ }
} else if (streq(key, "watchdog-override-usec")) {
usec_t watchdog_override_usec;
if (timestamp_deserialize(value, &watchdog_override_usec) < 0)
log_unit_debug(UNIT(s), "Setting watch for PID file %s", s->pid_file_pathspec->path);
- r = path_spec_watch(s->pid_file_pathspec, service_dispatch_io);
+ r = path_spec_watch(s->pid_file_pathspec, service_dispatch_inotify_io);
if (r < 0)
goto fail;
return service_watch_pid_file(s);
}
-static int service_dispatch_io(sd_event_source *source, int fd, uint32_t events, void *userdata) {
+static int service_dispatch_inotify_io(sd_event_source *source, int fd, uint32_t events, void *userdata) {
PathSpec *p = userdata;
Service *s;
return 0;
}
+static int service_dispatch_exec_io(sd_event_source *source, int fd, uint32_t events, void *userdata) {
+ Service *s = SERVICE(userdata);
+
+ assert(s);
+
+ log_unit_debug(UNIT(s), "got exec-fd event");
+
+ /* If Type=exec is set, we'll consider a service started successfully the instant we invoked execve()
+ * successfully for it. We implement this through a pipe() towards the child, which the kernel automatically
+ * closes for us due to O_CLOEXEC on execve() in the child, which then triggers EOF on the pipe in the
+ * parent. We need to be careful however, as there are other reasons that we might cause the child's side of
+ * the pipe to be closed (for example, a simple exit()). To deal with that we'll ignore EOFs on the pipe unless
+ * the child signalled us first that it is about to call the execve(). It does so by sending us a simple
+ * non-zero byte via the pipe. We also provide the child with a way to inform us in case execve() failed: if it
+ * sends a zero byte we'll ignore POLLHUP on the fd again. */
+
+ for (;;) {
+ uint8_t x;
+ ssize_t n;
+
+ n = read(fd, &x, sizeof(x));
+ if (n < 0) {
+ if (errno == EAGAIN) /* O_NONBLOCK in effect → everything queued has now been processed. */
+ return 0;
+
+ return log_unit_error_errno(UNIT(s), errno, "Failed to read from exec_fd: %m");
+ }
+ if (n == 0) { /* EOF → the event we are waiting for */
+
+ s->exec_fd_event_source = sd_event_source_unref(s->exec_fd_event_source);
+
+ if (s->exec_fd_hot) { /* Did the child tell us to expect EOF now? */
+ log_unit_debug(UNIT(s), "Got EOF on exec-fd");
+
+ s->exec_fd_hot = false;
+
+ /* Nice! This is what we have been waiting for. Transition to next state. */
+ if (s->type == SERVICE_EXEC && s->state == SERVICE_START)
+ service_enter_start_post(s);
+ } else
+ log_unit_debug(UNIT(s), "Got EOF on exec-fd while it was disabled, ignoring.");
+
+ return 0;
+ }
+
+ /* A byte was read → this turns on/off the exec fd logic */
+ assert(n == sizeof(x));
+ s->exec_fd_hot = x;
+ }
+
+ return 0;
+}
+
static void service_notify_cgroup_empty_event(Unit *u) {
Service *s = SERVICE(u);
[SERVICE_ONESHOT] = "oneshot",
[SERVICE_DBUS] = "dbus",
[SERVICE_NOTIFY] = "notify",
- [SERVICE_IDLE] = "idle"
+ [SERVICE_IDLE] = "idle",
+ [SERVICE_EXEC] = "exec",
};
DEFINE_STRING_TABLE_LOOKUP(service_type, ServiceType);
projects / platform / upstream / systemd.git / commitdiff