1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
6 #include "alloc-util.h"
7 #include "cgroup-setup.h"
8 #include "dbus-scope.h"
10 #include "exit-status.h"
11 #include "load-dropin.h"
13 #include "process-util.h"
14 #include "random-util.h"
16 #include "serialize.h"
18 #include "string-table.h"
19 #include "string-util.h"
21 #include "unit-name.h"
23 #include "user-util.h"
25 static const UnitActiveState state_translation_table
[_SCOPE_STATE_MAX
] = {
26 [SCOPE_DEAD
] = UNIT_INACTIVE
,
27 [SCOPE_START_CHOWN
] = UNIT_ACTIVATING
,
28 [SCOPE_RUNNING
] = UNIT_ACTIVE
,
29 [SCOPE_ABANDONED
] = UNIT_ACTIVE
,
30 [SCOPE_STOP_SIGTERM
] = UNIT_DEACTIVATING
,
31 [SCOPE_STOP_SIGKILL
] = UNIT_DEACTIVATING
,
32 [SCOPE_FAILED
] = UNIT_FAILED
,
35 static int scope_dispatch_timer(sd_event_source
*source
, usec_t usec
, void *userdata
);
37 static void scope_init(Unit
*u
) {
41 assert(u
->load_state
== UNIT_STUB
);
43 s
->runtime_max_usec
= USEC_INFINITY
;
44 s
->timeout_stop_usec
= u
->manager
->defaults
.timeout_stop_usec
;
45 u
->ignore_on_isolate
= true;
46 s
->user
= s
->group
= NULL
;
47 s
->oom_policy
= _OOM_POLICY_INVALID
;
50 static void scope_done(Unit
*u
) {
55 s
->controller
= mfree(s
->controller
);
56 s
->controller_track
= sd_bus_track_unref(s
->controller_track
);
58 s
->timer_event_source
= sd_event_source_disable_unref(s
->timer_event_source
);
60 s
->user
= mfree(s
->user
);
61 s
->group
= mfree(s
->group
);
64 static usec_t
scope_running_timeout(Scope
*s
) {
69 if (s
->runtime_rand_extra_usec
!= 0) {
70 delta
= random_u64_range(s
->runtime_rand_extra_usec
);
71 log_unit_debug(UNIT(s
), "Adding delta of %s sec to timeout", FORMAT_TIMESPAN(delta
, USEC_PER_SEC
));
74 return usec_add(usec_add(UNIT(s
)->active_enter_timestamp
.monotonic
,
79 static int scope_arm_timer(Scope
*s
, bool relative
, usec_t usec
) {
82 return unit_arm_timer(UNIT(s
), &s
->timer_event_source
, relative
, usec
, scope_dispatch_timer
);
85 static void scope_set_state(Scope
*s
, ScopeState state
) {
89 if (s
->state
!= state
)
90 bus_unit_send_pending_change_signal(UNIT(s
), false);
95 if (!IN_SET(state
, SCOPE_STOP_SIGTERM
, SCOPE_STOP_SIGKILL
, SCOPE_START_CHOWN
, SCOPE_RUNNING
))
96 s
->timer_event_source
= sd_event_source_disable_unref(s
->timer_event_source
);
98 if (!IN_SET(old_state
, SCOPE_DEAD
, SCOPE_FAILED
) && IN_SET(state
, SCOPE_DEAD
, SCOPE_FAILED
)) {
99 unit_unwatch_all_pids(UNIT(s
));
100 unit_dequeue_rewatch_pids(UNIT(s
));
103 if (state
!= old_state
)
104 log_debug("%s changed %s -> %s", UNIT(s
)->id
, scope_state_to_string(old_state
), scope_state_to_string(state
));
106 unit_notify(UNIT(s
), state_translation_table
[old_state
], state_translation_table
[state
], /* reload_success = */ true);
109 static int scope_add_default_dependencies(Scope
*s
) {
114 if (!UNIT(s
)->default_dependencies
)
117 /* Make sure scopes are unloaded on shutdown */
118 r
= unit_add_two_dependencies_by_name(
120 UNIT_BEFORE
, UNIT_CONFLICTS
,
121 SPECIAL_SHUTDOWN_TARGET
, true,
122 UNIT_DEPENDENCY_DEFAULT
);
129 static int scope_verify(Scope
*s
) {
131 assert(UNIT(s
)->load_state
== UNIT_LOADED
);
133 if (set_isempty(UNIT(s
)->pids
) &&
134 !MANAGER_IS_RELOADING(UNIT(s
)->manager
) &&
135 !unit_has_name(UNIT(s
), SPECIAL_INIT_SCOPE
))
136 return log_unit_error_errno(UNIT(s
), SYNTHETIC_ERRNO(ENOENT
), "Scope has no PIDs. Refusing.");
141 static int scope_load_init_scope(Unit
*u
) {
144 if (!unit_has_name(u
, SPECIAL_INIT_SCOPE
))
150 /* init.scope is a bit special, as it has to stick around forever. Because of its special semantics we
151 * synthesize it here, instead of relying on the unit file on disk. */
153 u
->default_dependencies
= false;
155 /* Prettify things, if we can. */
157 u
->description
= strdup("System and Service Manager");
158 if (!u
->documentation
)
159 (void) strv_extend(&u
->documentation
, "man:systemd(1)");
164 static int scope_add_extras(Scope
*s
) {
167 r
= unit_patch_contexts(UNIT(s
));
171 r
= unit_set_default_slice(UNIT(s
));
175 if (s
->oom_policy
< 0)
176 s
->oom_policy
= s
->cgroup_context
.delegate
? OOM_CONTINUE
: UNIT(s
)->manager
->defaults
.oom_policy
;
178 s
->cgroup_context
.memory_oom_group
= s
->oom_policy
== OOM_KILL
;
180 return scope_add_default_dependencies(s
);
183 static int scope_load(Unit
*u
) {
188 assert(u
->load_state
== UNIT_STUB
);
190 if (!u
->transient
&& !MANAGER_IS_RELOADING(u
->manager
))
191 /* Refuse to load non-transient scope units, but allow them while reloading. */
194 r
= scope_load_init_scope(u
);
198 r
= unit_load_fragment_and_dropin(u
, false);
202 if (u
->load_state
!= UNIT_LOADED
)
205 r
= scope_add_extras(s
);
209 return scope_verify(s
);
212 static usec_t
scope_coldplug_timeout(Scope
*s
) {
215 switch (s
->deserialized_state
) {
218 return scope_running_timeout(s
);
220 case SCOPE_STOP_SIGKILL
:
221 case SCOPE_STOP_SIGTERM
:
222 return usec_add(UNIT(s
)->state_change_timestamp
.monotonic
, s
->timeout_stop_usec
);
225 return USEC_INFINITY
;
229 static int scope_coldplug(Unit
*u
) {
234 assert(s
->state
== SCOPE_DEAD
);
236 if (s
->deserialized_state
== s
->state
)
239 r
= scope_arm_timer(s
, /* relative= */ false, scope_coldplug_timeout(s
));
243 if (!IN_SET(s
->deserialized_state
, SCOPE_DEAD
, SCOPE_FAILED
)) {
247 SET_FOREACH(pid
, u
->pids
) {
248 r
= unit_watch_pidref(u
, pid
, /* exclusive= */ false);
249 if (r
< 0 && r
!= -EEXIST
)
253 (void) unit_enqueue_rewatch_pids(u
);
256 bus_scope_track_controller(s
);
258 scope_set_state(s
, s
->deserialized_state
);
262 static void scope_dump(Unit
*u
, FILE *f
, const char *prefix
) {
269 "%sScope State: %s\n"
271 "%sRuntimeMaxSec: %s\n"
272 "%sRuntimeRandomizedExtraSec: %s\n"
274 prefix
, scope_state_to_string(s
->state
),
275 prefix
, scope_result_to_string(s
->result
),
276 prefix
, FORMAT_TIMESPAN(s
->runtime_max_usec
, USEC_PER_SEC
),
277 prefix
, FORMAT_TIMESPAN(s
->runtime_rand_extra_usec
, USEC_PER_SEC
),
278 prefix
, oom_policy_to_string(s
->oom_policy
));
280 cgroup_context_dump(UNIT(s
), f
, prefix
);
281 kill_context_dump(&s
->kill_context
, f
, prefix
);
284 static void scope_enter_dead(Scope
*s
, ScopeResult f
) {
287 if (s
->result
== SCOPE_SUCCESS
)
290 unit_log_result(UNIT(s
), s
->result
== SCOPE_SUCCESS
, scope_result_to_string(s
->result
));
291 scope_set_state(s
, s
->result
!= SCOPE_SUCCESS
? SCOPE_FAILED
: SCOPE_DEAD
);
294 static void scope_enter_signal(Scope
*s
, ScopeState state
, ScopeResult f
) {
295 bool skip_signal
= false;
300 if (s
->result
== SCOPE_SUCCESS
)
303 /* Before sending any signal, make sure we track all members of this cgroup */
304 (void) unit_watch_all_pids(UNIT(s
));
306 /* Also, enqueue a job that we recheck all our PIDs a bit later, given that it's likely some processes have
308 (void) unit_enqueue_rewatch_pids(UNIT(s
));
310 /* If we have a controller set let's ask the controller nicely to terminate the scope, instead of us going
311 * directly into SIGTERM berserk mode */
312 if (state
== SCOPE_STOP_SIGTERM
)
313 skip_signal
= bus_scope_send_request_stop(s
) > 0;
318 r
= unit_kill_context(
320 state
!= SCOPE_STOP_SIGTERM
? KILL_KILL
:
321 s
->was_abandoned
? KILL_TERMINATE_AND_LOG
:
324 log_unit_warning_errno(UNIT(s
), r
, "Failed to kill processes: %m");
330 r
= scope_arm_timer(s
, /* relative= */ true, s
->timeout_stop_usec
);
332 log_unit_warning_errno(UNIT(s
), r
, "Failed to install timer: %m");
336 scope_set_state(s
, state
);
337 } else if (state
== SCOPE_STOP_SIGTERM
)
338 scope_enter_signal(s
, SCOPE_STOP_SIGKILL
, SCOPE_SUCCESS
);
340 scope_enter_dead(s
, SCOPE_SUCCESS
);
345 scope_enter_dead(s
, SCOPE_FAILURE_RESOURCES
);
348 static int scope_enter_start_chown(Scope
*s
) {
349 _cleanup_(pidref_done
) PidRef pidref
= PIDREF_NULL
;
356 r
= scope_arm_timer(s
, /* relative= */ true, u
->manager
->defaults
.timeout_start_usec
);
360 r
= unit_fork_helper_process(u
, "(sd-chown-cgroup)", &pidref
);
365 uid_t uid
= UID_INVALID
;
366 gid_t gid
= GID_INVALID
;
368 if (!isempty(s
->user
)) {
369 const char *user
= s
->user
;
371 r
= get_user_creds(&user
, &uid
, &gid
, NULL
, NULL
, 0);
373 log_unit_error_errno(UNIT(s
), r
, "Failed to resolve user \"%s\": %m", user
);
378 if (!isempty(s
->group
)) {
379 const char *group
= s
->group
;
381 r
= get_group_creds(&group
, &gid
, 0);
383 log_unit_error_errno(UNIT(s
), r
, "Failed to resolve group \"%s\": %m", group
);
388 r
= cg_set_access(SYSTEMD_CGROUP_CONTROLLER
, u
->cgroup_path
, uid
, gid
);
390 log_unit_error_errno(UNIT(s
), r
, "Failed to adjust control group access: %m");
397 r
= unit_watch_pidref(UNIT(s
), &pidref
, /* exclusive= */ true);
401 scope_set_state(s
, SCOPE_START_CHOWN
);
405 s
->timer_event_source
= sd_event_source_disable_unref(s
->timer_event_source
);
409 static int scope_enter_running(Scope
*s
) {
415 (void) bus_scope_track_controller(s
);
417 r
= unit_acquire_invocation_id(u
);
421 unit_export_state_files(u
);
423 r
= unit_attach_pids_to_cgroup(u
, u
->pids
, NULL
);
425 log_unit_warning_errno(u
, r
, "Failed to add PIDs to scope's control group: %m");
429 r
= log_unit_warning_errno(u
, SYNTHETIC_ERRNO(ECHILD
), "No PIDs left to attach to the scope's control group, refusing.");
432 log_unit_debug(u
, "%i %s added to scope's control group.", r
, r
== 1 ? "process" : "processes");
434 s
->result
= SCOPE_SUCCESS
;
436 scope_set_state(s
, SCOPE_RUNNING
);
438 /* Set the maximum runtime timeout. */
439 scope_arm_timer(s
, /* relative= */ false, scope_running_timeout(s
));
441 /* On unified we use proper notifications hence we can unwatch the PIDs
442 * we just attached to the scope. This can also be done on legacy as
443 * we're going to update the list of the processes we watch with the
444 * PIDs currently in the scope anyway. */
445 unit_unwatch_all_pids(u
);
447 /* Start watching the PIDs currently in the scope (legacy hierarchy only) */
448 (void) unit_enqueue_rewatch_pids(u
);
452 scope_enter_dead(s
, SCOPE_FAILURE_RESOURCES
);
456 static int scope_start(Unit
*u
) {
461 if (unit_has_name(u
, SPECIAL_INIT_SCOPE
))
464 if (s
->state
== SCOPE_FAILED
)
467 /* We can't fulfill this right now, please try again later */
468 if (IN_SET(s
->state
, SCOPE_STOP_SIGTERM
, SCOPE_STOP_SIGKILL
))
471 assert(s
->state
== SCOPE_DEAD
);
473 if (!u
->transient
&& !MANAGER_IS_RELOADING(u
->manager
))
476 (void) unit_realize_cgroup(u
);
477 (void) unit_reset_accounting(u
);
479 /* We check only for User= option to keep behavior consistent with logic for service units,
480 * i.e. having 'Delegate=true Group=foo' w/o specifying User= has no effect. */
481 if (s
->user
&& unit_cgroup_delegate(u
))
482 return scope_enter_start_chown(s
);
484 return scope_enter_running(s
);
487 static int scope_stop(Unit
*u
) {
492 if (IN_SET(s
->state
, SCOPE_STOP_SIGTERM
, SCOPE_STOP_SIGKILL
))
495 assert(IN_SET(s
->state
, SCOPE_RUNNING
, SCOPE_ABANDONED
));
497 scope_enter_signal(s
, SCOPE_STOP_SIGTERM
, SCOPE_SUCCESS
);
501 static void scope_reset_failed(Unit
*u
) {
506 if (s
->state
== SCOPE_FAILED
)
507 scope_set_state(s
, SCOPE_DEAD
);
509 s
->result
= SCOPE_SUCCESS
;
512 static int scope_get_timeout(Unit
*u
, usec_t
*timeout
) {
517 if (!s
->timer_event_source
)
520 r
= sd_event_source_get_time(s
->timer_event_source
, &t
);
523 if (t
== USEC_INFINITY
)
530 static int scope_serialize(Unit
*u
, FILE *f
, FDSet
*fds
) {
538 (void) serialize_item(f
, "state", scope_state_to_string(s
->state
));
539 (void) serialize_bool(f
, "was-abandoned", s
->was_abandoned
);
542 (void) serialize_item(f
, "controller", s
->controller
);
544 SET_FOREACH(pid
, u
->pids
)
545 serialize_pidref(f
, fds
, "pids", pid
);
550 static int scope_deserialize_item(Unit
*u
, const char *key
, const char *value
, FDSet
*fds
) {
559 if (streq(key
, "state")) {
562 state
= scope_state_from_string(value
);
564 log_unit_debug(u
, "Failed to parse state value: %s", value
);
566 s
->deserialized_state
= state
;
568 } else if (streq(key
, "was-abandoned")) {
571 k
= parse_boolean(value
);
573 log_unit_debug(u
, "Failed to parse boolean value: %s", value
);
575 s
->was_abandoned
= k
;
576 } else if (streq(key
, "controller")) {
578 r
= free_and_strdup(&s
->controller
, value
);
582 } else if (streq(key
, "pids")) {
583 _cleanup_(pidref_done
) PidRef pidref
= PIDREF_NULL
;
585 if (deserialize_pidref(fds
, value
, &pidref
) >= 0) {
586 r
= unit_watch_pidref(u
, &pidref
, /* exclusive= */ false);
588 log_unit_debug(u
, "Failed to watch PID, ignoring: %s", value
);
591 log_unit_debug(u
, "Unknown serialization key: %s", key
);
596 static void scope_notify_cgroup_empty_event(Unit
*u
) {
600 log_unit_debug(u
, "cgroup is empty");
602 if (IN_SET(s
->state
, SCOPE_RUNNING
, SCOPE_ABANDONED
, SCOPE_STOP_SIGTERM
, SCOPE_STOP_SIGKILL
))
603 scope_enter_dead(s
, SCOPE_SUCCESS
);
606 static void scope_notify_cgroup_oom_event(Unit
*u
, bool managed_oom
) {
610 log_unit_debug(u
, "Process(es) of control group were killed by systemd-oomd.");
612 log_unit_debug(u
, "Process of control group was killed by the OOM killer.");
614 if (s
->oom_policy
== OOM_CONTINUE
)
619 case SCOPE_START_CHOWN
:
621 scope_enter_signal(s
, SCOPE_STOP_SIGTERM
, SCOPE_FAILURE_OOM_KILL
);
624 case SCOPE_STOP_SIGTERM
:
625 scope_enter_signal(s
, SCOPE_STOP_SIGKILL
, SCOPE_FAILURE_OOM_KILL
);
628 case SCOPE_STOP_SIGKILL
:
629 if (s
->result
== SCOPE_SUCCESS
)
630 s
->result
= SCOPE_FAILURE_OOM_KILL
;
632 /* SCOPE_DEAD, SCOPE_ABANDONED, and SCOPE_FAILED end up in default */
638 static void scope_sigchld_event(Unit
*u
, pid_t pid
, int code
, int status
) {
643 if (s
->state
== SCOPE_START_CHOWN
) {
644 if (!is_clean_exit(code
, status
, EXIT_CLEAN_COMMAND
, NULL
))
645 scope_enter_dead(s
, SCOPE_FAILURE_RESOURCES
);
647 scope_enter_running(s
);
651 /* If we get a SIGCHLD event for one of the processes we were interested in, then we look for others to
652 * watch, under the assumption that we'll sooner or later get a SIGCHLD for them, as the original
653 * process we watched was probably the parent of them, and they are hence now our children. */
655 (void) unit_enqueue_rewatch_pids(u
);
658 static int scope_dispatch_timer(sd_event_source
*source
, usec_t usec
, void *userdata
) {
659 Scope
*s
= SCOPE(userdata
);
662 assert(s
->timer_event_source
== source
);
667 log_unit_warning(UNIT(s
), "Scope reached runtime time limit. Stopping.");
668 scope_enter_signal(s
, SCOPE_STOP_SIGTERM
, SCOPE_FAILURE_TIMEOUT
);
671 case SCOPE_STOP_SIGTERM
:
672 if (s
->kill_context
.send_sigkill
) {
673 log_unit_warning(UNIT(s
), "Stopping timed out. Killing.");
674 scope_enter_signal(s
, SCOPE_STOP_SIGKILL
, SCOPE_FAILURE_TIMEOUT
);
676 log_unit_warning(UNIT(s
), "Stopping timed out. Skipping SIGKILL.");
677 scope_enter_dead(s
, SCOPE_FAILURE_TIMEOUT
);
682 case SCOPE_STOP_SIGKILL
:
683 log_unit_warning(UNIT(s
), "Still around after SIGKILL. Ignoring.");
684 scope_enter_dead(s
, SCOPE_FAILURE_TIMEOUT
);
687 case SCOPE_START_CHOWN
:
688 log_unit_warning(UNIT(s
), "User lookup timed out. Entering failed state.");
689 scope_enter_dead(s
, SCOPE_FAILURE_TIMEOUT
);
693 assert_not_reached();
699 int scope_abandon(Scope
*s
) {
702 if (unit_has_name(UNIT(s
), SPECIAL_INIT_SCOPE
))
705 if (!IN_SET(s
->state
, SCOPE_RUNNING
, SCOPE_ABANDONED
))
708 s
->was_abandoned
= true;
710 s
->controller
= mfree(s
->controller
);
711 s
->controller_track
= sd_bus_track_unref(s
->controller_track
);
713 scope_set_state(s
, SCOPE_ABANDONED
);
715 /* The client is no longer watching the remaining processes, so let's step in here, under the assumption that
716 * the remaining processes will be sooner or later reassigned to us as parent. */
717 (void) unit_enqueue_rewatch_pids(UNIT(s
));
722 static UnitActiveState
scope_active_state(Unit
*u
) {
725 return state_translation_table
[SCOPE(u
)->state
];
728 static const char *scope_sub_state_to_string(Unit
*u
) {
731 return scope_state_to_string(SCOPE(u
)->state
);
734 static void scope_enumerate_perpetual(Manager
*m
) {
740 /* Let's unconditionally add the "init.scope" special unit
741 * that encapsulates PID 1. Note that PID 1 already is in the
742 * cgroup for this, we hence just need to allocate the object
743 * for it and that's it. */
745 u
= manager_get_unit(m
, SPECIAL_INIT_SCOPE
);
747 r
= unit_new_for_name(m
, sizeof(Scope
), SPECIAL_INIT_SCOPE
, &u
);
749 log_error_errno(r
, "Failed to allocate the special " SPECIAL_INIT_SCOPE
" unit: %m");
756 SCOPE(u
)->deserialized_state
= SCOPE_RUNNING
;
758 unit_add_to_load_queue(u
);
759 unit_add_to_dbus_queue(u
);
760 /* Enqueue an explicit cgroup realization here. Unlike other cgroups this one already exists and is
761 * populated (by us, after all!) already, even when we are not in a reload cycle. Hence we cannot
762 * apply the settings at creation time anymore, but let's at least apply them asynchronously. */
763 unit_add_to_cgroup_realize_queue(u
);
766 static const char* const scope_result_table
[_SCOPE_RESULT_MAX
] = {
767 [SCOPE_SUCCESS
] = "success",
768 [SCOPE_FAILURE_RESOURCES
] = "resources",
769 [SCOPE_FAILURE_TIMEOUT
] = "timeout",
770 [SCOPE_FAILURE_OOM_KILL
] = "oom-kill",
773 DEFINE_STRING_TABLE_LOOKUP(scope_result
, ScopeResult
);
775 const UnitVTable scope_vtable
= {
776 .object_size
= sizeof(Scope
),
777 .cgroup_context_offset
= offsetof(Scope
, cgroup_context
),
778 .kill_context_offset
= offsetof(Scope
, kill_context
),
784 .private_section
= "Scope",
786 .can_transient
= true,
787 .can_delegate
= true,
790 .can_set_managed_oom
= true,
796 .coldplug
= scope_coldplug
,
800 .start
= scope_start
,
803 .freeze
= unit_freeze_vtable_common
,
804 .thaw
= unit_thaw_vtable_common
,
806 .get_timeout
= scope_get_timeout
,
808 .serialize
= scope_serialize
,
809 .deserialize_item
= scope_deserialize_item
,
811 .active_state
= scope_active_state
,
812 .sub_state_to_string
= scope_sub_state_to_string
,
814 .sigchld_event
= scope_sigchld_event
,
816 .reset_failed
= scope_reset_failed
,
818 .notify_cgroup_empty
= scope_notify_cgroup_empty_event
,
819 .notify_cgroup_oom
= scope_notify_cgroup_oom_event
,
821 .bus_set_property
= bus_scope_set_property
,
822 .bus_commit_properties
= bus_scope_commit_properties
,
824 .enumerate_perpetual
= scope_enumerate_perpetual
,