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
->default_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
, usec_t usec
) {
84 if (s
->timer_event_source
) {
85 r
= sd_event_source_set_time(s
->timer_event_source
, usec
);
89 return sd_event_source_set_enabled(s
->timer_event_source
, SD_EVENT_ONESHOT
);
92 if (usec
== USEC_INFINITY
)
95 r
= sd_event_add_time(
96 UNIT(s
)->manager
->event
,
97 &s
->timer_event_source
,
100 scope_dispatch_timer
, s
);
104 (void) sd_event_source_set_description(s
->timer_event_source
, "scope-timer");
109 static void scope_set_state(Scope
*s
, ScopeState state
) {
110 ScopeState old_state
;
113 if (s
->state
!= state
)
114 bus_unit_send_pending_change_signal(UNIT(s
), false);
116 old_state
= s
->state
;
119 if (!IN_SET(state
, SCOPE_STOP_SIGTERM
, SCOPE_STOP_SIGKILL
, SCOPE_START_CHOWN
, SCOPE_RUNNING
))
120 s
->timer_event_source
= sd_event_source_disable_unref(s
->timer_event_source
);
122 if (IN_SET(state
, SCOPE_DEAD
, SCOPE_FAILED
)) {
123 unit_unwatch_all_pids(UNIT(s
));
124 unit_dequeue_rewatch_pids(UNIT(s
));
127 if (state
!= old_state
)
128 log_debug("%s changed %s -> %s", UNIT(s
)->id
, scope_state_to_string(old_state
), scope_state_to_string(state
));
130 unit_notify(UNIT(s
), state_translation_table
[old_state
], state_translation_table
[state
], /* reload_success = */ true);
133 static int scope_add_default_dependencies(Scope
*s
) {
138 if (!UNIT(s
)->default_dependencies
)
141 /* Make sure scopes are unloaded on shutdown */
142 r
= unit_add_two_dependencies_by_name(
144 UNIT_BEFORE
, UNIT_CONFLICTS
,
145 SPECIAL_SHUTDOWN_TARGET
, true,
146 UNIT_DEPENDENCY_DEFAULT
);
153 static int scope_verify(Scope
*s
) {
155 assert(UNIT(s
)->load_state
== UNIT_LOADED
);
157 if (set_isempty(UNIT(s
)->pids
) &&
158 !MANAGER_IS_RELOADING(UNIT(s
)->manager
) &&
159 !unit_has_name(UNIT(s
), SPECIAL_INIT_SCOPE
))
160 return log_unit_error_errno(UNIT(s
), SYNTHETIC_ERRNO(ENOENT
), "Scope has no PIDs. Refusing.");
165 static int scope_load_init_scope(Unit
*u
) {
168 if (!unit_has_name(u
, SPECIAL_INIT_SCOPE
))
174 /* init.scope is a bit special, as it has to stick around forever. Because of its special semantics we
175 * synthesize it here, instead of relying on the unit file on disk. */
177 u
->default_dependencies
= false;
179 /* Prettify things, if we can. */
181 u
->description
= strdup("System and Service Manager");
182 if (!u
->documentation
)
183 (void) strv_extend(&u
->documentation
, "man:systemd(1)");
188 static int scope_add_extras(Scope
*s
) {
191 r
= unit_patch_contexts(UNIT(s
));
195 r
= unit_set_default_slice(UNIT(s
));
199 if (s
->oom_policy
< 0)
200 s
->oom_policy
= s
->cgroup_context
.delegate
? OOM_CONTINUE
: UNIT(s
)->manager
->default_oom_policy
;
202 s
->cgroup_context
.memory_oom_group
= s
->oom_policy
== OOM_KILL
;
204 return scope_add_default_dependencies(s
);
207 static int scope_load(Unit
*u
) {
212 assert(u
->load_state
== UNIT_STUB
);
214 if (!u
->transient
&& !MANAGER_IS_RELOADING(u
->manager
))
215 /* Refuse to load non-transient scope units, but allow them while reloading. */
218 r
= scope_load_init_scope(u
);
222 r
= unit_load_fragment_and_dropin(u
, false);
226 if (u
->load_state
!= UNIT_LOADED
)
229 r
= scope_add_extras(s
);
233 return scope_verify(s
);
236 static usec_t
scope_coldplug_timeout(Scope
*s
) {
239 switch (s
->deserialized_state
) {
242 return scope_running_timeout(s
);
244 case SCOPE_STOP_SIGKILL
:
245 case SCOPE_STOP_SIGTERM
:
246 return usec_add(UNIT(s
)->state_change_timestamp
.monotonic
, s
->timeout_stop_usec
);
249 return USEC_INFINITY
;
253 static int scope_coldplug(Unit
*u
) {
258 assert(s
->state
== SCOPE_DEAD
);
260 if (s
->deserialized_state
== s
->state
)
263 r
= scope_arm_timer(s
, scope_coldplug_timeout(s
));
267 if (!IN_SET(s
->deserialized_state
, SCOPE_DEAD
, SCOPE_FAILED
)) {
271 SET_FOREACH(pidp
, u
->pids
) {
272 r
= unit_watch_pid(u
, PTR_TO_PID(pidp
), false);
273 if (r
< 0 && r
!= -EEXIST
)
277 (void) unit_enqueue_rewatch_pids(u
);
280 bus_scope_track_controller(s
);
282 scope_set_state(s
, s
->deserialized_state
);
286 static void scope_dump(Unit
*u
, FILE *f
, const char *prefix
) {
293 "%sScope State: %s\n"
295 "%sRuntimeMaxSec: %s\n"
296 "%sRuntimeRandomizedExtraSec: %s\n"
298 prefix
, scope_state_to_string(s
->state
),
299 prefix
, scope_result_to_string(s
->result
),
300 prefix
, FORMAT_TIMESPAN(s
->runtime_max_usec
, USEC_PER_SEC
),
301 prefix
, FORMAT_TIMESPAN(s
->runtime_rand_extra_usec
, USEC_PER_SEC
),
302 prefix
, oom_policy_to_string(s
->oom_policy
));
304 cgroup_context_dump(UNIT(s
), f
, prefix
);
305 kill_context_dump(&s
->kill_context
, f
, prefix
);
308 static void scope_enter_dead(Scope
*s
, ScopeResult f
) {
311 if (s
->result
== SCOPE_SUCCESS
)
314 unit_log_result(UNIT(s
), s
->result
== SCOPE_SUCCESS
, scope_result_to_string(s
->result
));
315 scope_set_state(s
, s
->result
!= SCOPE_SUCCESS
? SCOPE_FAILED
: SCOPE_DEAD
);
318 static void scope_enter_signal(Scope
*s
, ScopeState state
, ScopeResult f
) {
319 bool skip_signal
= false;
324 if (s
->result
== SCOPE_SUCCESS
)
327 /* Before sending any signal, make sure we track all members of this cgroup */
328 (void) unit_watch_all_pids(UNIT(s
));
330 /* Also, enqueue a job that we recheck all our PIDs a bit later, given that it's likely some processes have
332 (void) unit_enqueue_rewatch_pids(UNIT(s
));
334 /* If we have a controller set let's ask the controller nicely to terminate the scope, instead of us going
335 * directly into SIGTERM berserk mode */
336 if (state
== SCOPE_STOP_SIGTERM
)
337 skip_signal
= bus_scope_send_request_stop(s
) > 0;
342 r
= unit_kill_context(
345 state
!= SCOPE_STOP_SIGTERM
? KILL_KILL
:
346 s
->was_abandoned
? KILL_TERMINATE_AND_LOG
:
354 r
= scope_arm_timer(s
, usec_add(now(CLOCK_MONOTONIC
), s
->timeout_stop_usec
));
358 scope_set_state(s
, state
);
359 } else if (state
== SCOPE_STOP_SIGTERM
)
360 scope_enter_signal(s
, SCOPE_STOP_SIGKILL
, SCOPE_SUCCESS
);
362 scope_enter_dead(s
, SCOPE_SUCCESS
);
367 log_unit_warning_errno(UNIT(s
), r
, "Failed to kill processes: %m");
369 scope_enter_dead(s
, SCOPE_FAILURE_RESOURCES
);
372 static int scope_enter_start_chown(Scope
*s
) {
380 r
= scope_arm_timer(s
, usec_add(now(CLOCK_MONOTONIC
), u
->manager
->default_timeout_start_usec
));
384 r
= unit_fork_helper_process(u
, "(sd-chown-cgroup)", &pid
);
389 uid_t uid
= UID_INVALID
;
390 gid_t gid
= GID_INVALID
;
392 if (!isempty(s
->user
)) {
393 const char *user
= s
->user
;
395 r
= get_user_creds(&user
, &uid
, &gid
, NULL
, NULL
, 0);
397 log_unit_error_errno(UNIT(s
), r
, "Failed to resolve user \"%s\": %m", user
);
402 if (!isempty(s
->group
)) {
403 const char *group
= s
->group
;
405 r
= get_group_creds(&group
, &gid
, 0);
407 log_unit_error_errno(UNIT(s
), r
, "Failed to resolve group \"%s\": %m", group
);
412 r
= cg_set_access(SYSTEMD_CGROUP_CONTROLLER
, u
->cgroup_path
, uid
, gid
);
414 log_unit_error_errno(UNIT(s
), r
, "Failed to adjust control group access: %m");
421 r
= unit_watch_pid(UNIT(s
), pid
, true);
425 scope_set_state(s
, SCOPE_START_CHOWN
);
429 s
->timer_event_source
= sd_event_source_disable_unref(s
->timer_event_source
);
433 static int scope_enter_running(Scope
*s
) {
439 (void) bus_scope_track_controller(s
);
441 r
= unit_acquire_invocation_id(u
);
445 unit_export_state_files(u
);
447 r
= unit_attach_pids_to_cgroup(u
, u
->pids
, NULL
);
449 log_unit_warning_errno(u
, r
, "Failed to add PIDs to scope's control group: %m");
450 scope_enter_dead(s
, SCOPE_FAILURE_RESOURCES
);
454 log_unit_warning(u
, "No PIDs left to attach to the scope's control group, refusing.");
455 scope_enter_dead(s
, SCOPE_FAILURE_RESOURCES
);
458 log_unit_debug(u
, "%i %s added to scope's control group.", r
, r
== 1 ? "process" : "processes");
460 s
->result
= SCOPE_SUCCESS
;
462 scope_set_state(s
, SCOPE_RUNNING
);
464 /* Set the maximum runtime timeout. */
465 scope_arm_timer(s
, scope_running_timeout(s
));
467 /* On unified we use proper notifications hence we can unwatch the PIDs
468 * we just attached to the scope. This can also be done on legacy as
469 * we're going to update the list of the processes we watch with the
470 * PIDs currently in the scope anyway. */
471 unit_unwatch_all_pids(u
);
473 /* Start watching the PIDs currently in the scope (legacy hierarchy only) */
474 (void) unit_enqueue_rewatch_pids(u
);
478 static int scope_start(Unit
*u
) {
483 if (unit_has_name(u
, SPECIAL_INIT_SCOPE
))
486 if (s
->state
== SCOPE_FAILED
)
489 /* We can't fulfill this right now, please try again later */
490 if (IN_SET(s
->state
, SCOPE_STOP_SIGTERM
, SCOPE_STOP_SIGKILL
))
493 assert(s
->state
== SCOPE_DEAD
);
495 if (!u
->transient
&& !MANAGER_IS_RELOADING(u
->manager
))
498 (void) unit_realize_cgroup(u
);
499 (void) unit_reset_accounting(u
);
501 /* We check only for User= option to keep behavior consistent with logic for service units,
502 * i.e. having 'Delegate=true Group=foo' w/o specifying User= has no effect. */
503 if (s
->user
&& unit_cgroup_delegate(u
))
504 return scope_enter_start_chown(s
);
506 return scope_enter_running(s
);
509 static int scope_stop(Unit
*u
) {
514 if (IN_SET(s
->state
, SCOPE_STOP_SIGTERM
, SCOPE_STOP_SIGKILL
))
517 assert(IN_SET(s
->state
, SCOPE_RUNNING
, SCOPE_ABANDONED
));
519 scope_enter_signal(s
, SCOPE_STOP_SIGTERM
, SCOPE_SUCCESS
);
523 static void scope_reset_failed(Unit
*u
) {
528 if (s
->state
== SCOPE_FAILED
)
529 scope_set_state(s
, SCOPE_DEAD
);
531 s
->result
= SCOPE_SUCCESS
;
534 static int scope_kill(Unit
*u
, KillWho who
, int signo
, int code
, int value
, sd_bus_error
*error
) {
535 return unit_kill_common(u
, who
, signo
, code
, value
, -1, -1, error
);
538 static int scope_get_timeout(Unit
*u
, usec_t
*timeout
) {
543 if (!s
->timer_event_source
)
546 r
= sd_event_source_get_time(s
->timer_event_source
, &t
);
549 if (t
== USEC_INFINITY
)
556 static int scope_serialize(Unit
*u
, FILE *f
, FDSet
*fds
) {
564 (void) serialize_item(f
, "state", scope_state_to_string(s
->state
));
565 (void) serialize_bool(f
, "was-abandoned", s
->was_abandoned
);
568 (void) serialize_item(f
, "controller", s
->controller
);
570 SET_FOREACH(pidp
, u
->pids
)
571 serialize_item_format(f
, "pids", PID_FMT
, PTR_TO_PID(pidp
));
576 static int scope_deserialize_item(Unit
*u
, const char *key
, const char *value
, FDSet
*fds
) {
585 if (streq(key
, "state")) {
588 state
= scope_state_from_string(value
);
590 log_unit_debug(u
, "Failed to parse state value: %s", value
);
592 s
->deserialized_state
= state
;
594 } else if (streq(key
, "was-abandoned")) {
597 k
= parse_boolean(value
);
599 log_unit_debug(u
, "Failed to parse boolean value: %s", value
);
601 s
->was_abandoned
= k
;
602 } else if (streq(key
, "controller")) {
604 r
= free_and_strdup(&s
->controller
, value
);
608 } else if (streq(key
, "pids")) {
611 if (parse_pid(value
, &pid
) < 0)
612 log_unit_debug(u
, "Failed to parse pids value: %s", value
);
614 r
= set_ensure_put(&u
->pids
, NULL
, PID_TO_PTR(pid
));
619 log_unit_debug(u
, "Unknown serialization key: %s", key
);
624 static void scope_notify_cgroup_empty_event(Unit
*u
) {
628 log_unit_debug(u
, "cgroup is empty");
630 if (IN_SET(s
->state
, SCOPE_RUNNING
, SCOPE_ABANDONED
, SCOPE_STOP_SIGTERM
, SCOPE_STOP_SIGKILL
))
631 scope_enter_dead(s
, SCOPE_SUCCESS
);
634 static void scope_notify_cgroup_oom_event(Unit
*u
, bool managed_oom
) {
638 log_unit_debug(u
, "Process(es) of control group were killed by systemd-oomd.");
640 log_unit_debug(u
, "Process of control group was killed by the OOM killer.");
642 if (s
->oom_policy
== OOM_CONTINUE
)
647 case SCOPE_START_CHOWN
:
649 scope_enter_signal(s
, SCOPE_STOP_SIGTERM
, SCOPE_FAILURE_OOM_KILL
);
652 case SCOPE_STOP_SIGTERM
:
653 scope_enter_signal(s
, SCOPE_STOP_SIGKILL
, SCOPE_FAILURE_OOM_KILL
);
656 case SCOPE_STOP_SIGKILL
:
657 if (s
->result
== SCOPE_SUCCESS
)
658 s
->result
= SCOPE_FAILURE_OOM_KILL
;
660 /* SCOPE_DEAD, SCOPE_ABANDONED, and SCOPE_FAILED end up in default */
666 static void scope_sigchld_event(Unit
*u
, pid_t pid
, int code
, int status
) {
671 if (s
->state
== SCOPE_START_CHOWN
) {
672 if (!is_clean_exit(code
, status
, EXIT_CLEAN_COMMAND
, NULL
))
673 scope_enter_dead(s
, SCOPE_FAILURE_RESOURCES
);
675 scope_enter_running(s
);
679 /* If we get a SIGCHLD event for one of the processes we were interested in, then we look for others to
680 * watch, under the assumption that we'll sooner or later get a SIGCHLD for them, as the original
681 * process we watched was probably the parent of them, and they are hence now our children. */
683 (void) unit_enqueue_rewatch_pids(u
);
686 static int scope_dispatch_timer(sd_event_source
*source
, usec_t usec
, void *userdata
) {
687 Scope
*s
= SCOPE(userdata
);
690 assert(s
->timer_event_source
== source
);
695 log_unit_warning(UNIT(s
), "Scope reached runtime time limit. Stopping.");
696 scope_enter_signal(s
, SCOPE_STOP_SIGTERM
, SCOPE_FAILURE_TIMEOUT
);
699 case SCOPE_STOP_SIGTERM
:
700 if (s
->kill_context
.send_sigkill
) {
701 log_unit_warning(UNIT(s
), "Stopping timed out. Killing.");
702 scope_enter_signal(s
, SCOPE_STOP_SIGKILL
, SCOPE_FAILURE_TIMEOUT
);
704 log_unit_warning(UNIT(s
), "Stopping timed out. Skipping SIGKILL.");
705 scope_enter_dead(s
, SCOPE_FAILURE_TIMEOUT
);
710 case SCOPE_STOP_SIGKILL
:
711 log_unit_warning(UNIT(s
), "Still around after SIGKILL. Ignoring.");
712 scope_enter_dead(s
, SCOPE_FAILURE_TIMEOUT
);
715 case SCOPE_START_CHOWN
:
716 log_unit_warning(UNIT(s
), "User lookup timed out. Entering failed state.");
717 scope_enter_dead(s
, SCOPE_FAILURE_TIMEOUT
);
721 assert_not_reached();
727 int scope_abandon(Scope
*s
) {
730 if (unit_has_name(UNIT(s
), SPECIAL_INIT_SCOPE
))
733 if (!IN_SET(s
->state
, SCOPE_RUNNING
, SCOPE_ABANDONED
))
736 s
->was_abandoned
= true;
738 s
->controller
= mfree(s
->controller
);
739 s
->controller_track
= sd_bus_track_unref(s
->controller_track
);
741 scope_set_state(s
, SCOPE_ABANDONED
);
743 /* The client is no longer watching the remaining processes, so let's step in here, under the assumption that
744 * the remaining processes will be sooner or later reassigned to us as parent. */
745 (void) unit_enqueue_rewatch_pids(UNIT(s
));
750 _pure_
static UnitActiveState
scope_active_state(Unit
*u
) {
753 return state_translation_table
[SCOPE(u
)->state
];
756 _pure_
static const char *scope_sub_state_to_string(Unit
*u
) {
759 return scope_state_to_string(SCOPE(u
)->state
);
762 static void scope_enumerate_perpetual(Manager
*m
) {
768 /* Let's unconditionally add the "init.scope" special unit
769 * that encapsulates PID 1. Note that PID 1 already is in the
770 * cgroup for this, we hence just need to allocate the object
771 * for it and that's it. */
773 u
= manager_get_unit(m
, SPECIAL_INIT_SCOPE
);
775 r
= unit_new_for_name(m
, sizeof(Scope
), SPECIAL_INIT_SCOPE
, &u
);
777 log_error_errno(r
, "Failed to allocate the special " SPECIAL_INIT_SCOPE
" unit: %m");
784 SCOPE(u
)->deserialized_state
= SCOPE_RUNNING
;
786 unit_add_to_load_queue(u
);
787 unit_add_to_dbus_queue(u
);
788 /* Enqueue an explicit cgroup realization here. Unlike other cgroups this one already exists and is
789 * populated (by us, after all!) already, even when we are not in a reload cycle. Hence we cannot
790 * apply the settings at creation time anymore, but let's at least apply them asynchronously. */
791 unit_add_to_cgroup_realize_queue(u
);
794 static const char* const scope_result_table
[_SCOPE_RESULT_MAX
] = {
795 [SCOPE_SUCCESS
] = "success",
796 [SCOPE_FAILURE_RESOURCES
] = "resources",
797 [SCOPE_FAILURE_TIMEOUT
] = "timeout",
798 [SCOPE_FAILURE_OOM_KILL
] = "oom-kill",
801 DEFINE_STRING_TABLE_LOOKUP(scope_result
, ScopeResult
);
803 const UnitVTable scope_vtable
= {
804 .object_size
= sizeof(Scope
),
805 .cgroup_context_offset
= offsetof(Scope
, cgroup_context
),
806 .kill_context_offset
= offsetof(Scope
, kill_context
),
812 .private_section
= "Scope",
814 .can_transient
= true,
815 .can_delegate
= true,
818 .can_set_managed_oom
= true,
824 .coldplug
= scope_coldplug
,
828 .start
= scope_start
,
833 .freeze
= unit_freeze_vtable_common
,
834 .thaw
= unit_thaw_vtable_common
,
836 .get_timeout
= scope_get_timeout
,
838 .serialize
= scope_serialize
,
839 .deserialize_item
= scope_deserialize_item
,
841 .active_state
= scope_active_state
,
842 .sub_state_to_string
= scope_sub_state_to_string
,
844 .sigchld_event
= scope_sigchld_event
,
846 .reset_failed
= scope_reset_failed
,
848 .notify_cgroup_empty
= scope_notify_cgroup_empty_event
,
849 .notify_cgroup_oom
= scope_notify_cgroup_oom_event
,
851 .bus_set_property
= bus_scope_set_property
,
852 .bus_commit_properties
= bus_scope_commit_properties
,
854 .enumerate_perpetual
= scope_enumerate_perpetual
,