]> git.ipfire.org Git - thirdparty/systemd.git/blob - src/core/scope.c
Merge pull request #9213 from poettering/copy-mount
[thirdparty/systemd.git] / src / core / scope.c
1 /* SPDX-License-Identifier: LGPL-2.1+ */
2 /***
3 This file is part of systemd.
4
5 Copyright 2013 Lennart Poettering
6 ***/
7
8 #include <errno.h>
9 #include <unistd.h>
10
11 #include "alloc-util.h"
12 #include "dbus-scope.h"
13 #include "load-dropin.h"
14 #include "log.h"
15 #include "scope.h"
16 #include "special.h"
17 #include "string-table.h"
18 #include "string-util.h"
19 #include "strv.h"
20 #include "unit-name.h"
21 #include "unit.h"
22
23 static const UnitActiveState state_translation_table[_SCOPE_STATE_MAX] = {
24 [SCOPE_DEAD] = UNIT_INACTIVE,
25 [SCOPE_RUNNING] = UNIT_ACTIVE,
26 [SCOPE_ABANDONED] = UNIT_ACTIVE,
27 [SCOPE_STOP_SIGTERM] = UNIT_DEACTIVATING,
28 [SCOPE_STOP_SIGKILL] = UNIT_DEACTIVATING,
29 [SCOPE_FAILED] = UNIT_FAILED
30 };
31
32 static int scope_dispatch_timer(sd_event_source *source, usec_t usec, void *userdata);
33
34 static void scope_init(Unit *u) {
35 Scope *s = SCOPE(u);
36
37 assert(u);
38 assert(u->load_state == UNIT_STUB);
39
40 s->timeout_stop_usec = u->manager->default_timeout_stop_usec;
41 u->ignore_on_isolate = true;
42 }
43
44 static void scope_done(Unit *u) {
45 Scope *s = SCOPE(u);
46
47 assert(u);
48
49 s->controller = mfree(s->controller);
50 s->controller_track = sd_bus_track_unref(s->controller_track);
51
52 s->timer_event_source = sd_event_source_unref(s->timer_event_source);
53 }
54
55 static int scope_arm_timer(Scope *s, usec_t usec) {
56 int r;
57
58 assert(s);
59
60 if (s->timer_event_source) {
61 r = sd_event_source_set_time(s->timer_event_source, usec);
62 if (r < 0)
63 return r;
64
65 return sd_event_source_set_enabled(s->timer_event_source, SD_EVENT_ONESHOT);
66 }
67
68 if (usec == USEC_INFINITY)
69 return 0;
70
71 r = sd_event_add_time(
72 UNIT(s)->manager->event,
73 &s->timer_event_source,
74 CLOCK_MONOTONIC,
75 usec, 0,
76 scope_dispatch_timer, s);
77 if (r < 0)
78 return r;
79
80 (void) sd_event_source_set_description(s->timer_event_source, "scope-timer");
81
82 return 0;
83 }
84
85 static void scope_set_state(Scope *s, ScopeState state) {
86 ScopeState old_state;
87 assert(s);
88
89 old_state = s->state;
90 s->state = state;
91
92 if (!IN_SET(state, SCOPE_STOP_SIGTERM, SCOPE_STOP_SIGKILL))
93 s->timer_event_source = sd_event_source_unref(s->timer_event_source);
94
95 if (IN_SET(state, SCOPE_DEAD, SCOPE_FAILED)) {
96 unit_unwatch_all_pids(UNIT(s));
97 unit_dequeue_rewatch_pids(UNIT(s));
98 }
99
100 if (state != old_state)
101 log_debug("%s changed %s -> %s", UNIT(s)->id, scope_state_to_string(old_state), scope_state_to_string(state));
102
103 unit_notify(UNIT(s), state_translation_table[old_state], state_translation_table[state], 0);
104 }
105
106 static int scope_add_default_dependencies(Scope *s) {
107 int r;
108
109 assert(s);
110
111 if (!UNIT(s)->default_dependencies)
112 return 0;
113
114 /* Make sure scopes are unloaded on shutdown */
115 r = unit_add_two_dependencies_by_name(
116 UNIT(s),
117 UNIT_BEFORE, UNIT_CONFLICTS,
118 SPECIAL_SHUTDOWN_TARGET, NULL, true,
119 UNIT_DEPENDENCY_DEFAULT);
120 if (r < 0)
121 return r;
122
123 return 0;
124 }
125
126 static int scope_verify(Scope *s) {
127 assert(s);
128
129 if (UNIT(s)->load_state != UNIT_LOADED)
130 return 0;
131
132 if (set_isempty(UNIT(s)->pids) &&
133 !MANAGER_IS_RELOADING(UNIT(s)->manager) &&
134 !unit_has_name(UNIT(s), SPECIAL_INIT_SCOPE)) {
135 log_unit_error(UNIT(s), "Scope has no PIDs. Refusing.");
136 return -EINVAL;
137 }
138
139 return 0;
140 }
141
142 static int scope_load_init_scope(Unit *u) {
143 assert(u);
144
145 if (!unit_has_name(u, SPECIAL_INIT_SCOPE))
146 return 0;
147
148 u->transient = true;
149 u->perpetual = true;
150
151 /* init.scope is a bit special, as it has to stick around forever. Because of its special semantics we
152 * synthesize it here, instead of relying on the unit file on disk. */
153
154 u->default_dependencies = false;
155
156 /* Prettify things, if we can. */
157 if (!u->description)
158 u->description = strdup("System and Service Manager");
159 if (!u->documentation)
160 (void) strv_extend(&u->documentation, "man:systemd(1)");
161
162 return 1;
163 }
164
165 static int scope_load(Unit *u) {
166 Scope *s = SCOPE(u);
167 int r;
168
169 assert(s);
170 assert(u->load_state == UNIT_STUB);
171
172 if (!u->transient && !MANAGER_IS_RELOADING(u->manager))
173 /* Refuse to load non-transient scope units, but allow them while reloading. */
174 return -ENOENT;
175
176 r = scope_load_init_scope(u);
177 if (r < 0)
178 return r;
179 r = unit_load_fragment_and_dropin_optional(u);
180 if (r < 0)
181 return r;
182
183 if (u->load_state == UNIT_LOADED) {
184 r = unit_patch_contexts(u);
185 if (r < 0)
186 return r;
187
188 r = unit_set_default_slice(u);
189 if (r < 0)
190 return r;
191
192 r = scope_add_default_dependencies(s);
193 if (r < 0)
194 return r;
195 }
196
197 return scope_verify(s);
198 }
199
200 static int scope_coldplug(Unit *u) {
201 Scope *s = SCOPE(u);
202 int r;
203
204 assert(s);
205 assert(s->state == SCOPE_DEAD);
206
207 if (s->deserialized_state == s->state)
208 return 0;
209
210 if (IN_SET(s->deserialized_state, SCOPE_STOP_SIGKILL, SCOPE_STOP_SIGTERM)) {
211 r = scope_arm_timer(s, usec_add(u->state_change_timestamp.monotonic, s->timeout_stop_usec));
212 if (r < 0)
213 return r;
214 }
215
216 if (!IN_SET(s->deserialized_state, SCOPE_DEAD, SCOPE_FAILED))
217 (void) unit_enqueue_rewatch_pids(u);
218
219 bus_scope_track_controller(s);
220
221 scope_set_state(s, s->deserialized_state);
222 return 0;
223 }
224
225 static void scope_dump(Unit *u, FILE *f, const char *prefix) {
226 Scope *s = SCOPE(u);
227
228 assert(s);
229 assert(f);
230
231 fprintf(f,
232 "%sScope State: %s\n"
233 "%sResult: %s\n",
234 prefix, scope_state_to_string(s->state),
235 prefix, scope_result_to_string(s->result));
236
237 cgroup_context_dump(&s->cgroup_context, f, prefix);
238 kill_context_dump(&s->kill_context, f, prefix);
239 }
240
241 static void scope_enter_dead(Scope *s, ScopeResult f) {
242 assert(s);
243
244 if (s->result == SCOPE_SUCCESS)
245 s->result = f;
246
247 if (s->result != SCOPE_SUCCESS)
248 log_unit_warning(UNIT(s), "Failed with result '%s'.", scope_result_to_string(s->result));
249
250 scope_set_state(s, s->result != SCOPE_SUCCESS ? SCOPE_FAILED : SCOPE_DEAD);
251 }
252
253 static void scope_enter_signal(Scope *s, ScopeState state, ScopeResult f) {
254 bool skip_signal = false;
255 int r;
256
257 assert(s);
258
259 if (s->result == SCOPE_SUCCESS)
260 s->result = f;
261
262 /* Before sending any signal, make sure we track all members of this cgroup */
263 (void) unit_watch_all_pids(UNIT(s));
264
265 /* Also, enqueue a job that we recheck all our PIDs a bit later, given that it's likely some processes have
266 * died now */
267 (void) unit_enqueue_rewatch_pids(UNIT(s));
268
269 /* If we have a controller set let's ask the controller nicely to terminate the scope, instead of us going
270 * directly into SIGTERM berserk mode */
271 if (state == SCOPE_STOP_SIGTERM)
272 skip_signal = bus_scope_send_request_stop(s) > 0;
273
274 if (skip_signal)
275 r = 1; /* wait */
276 else {
277 r = unit_kill_context(
278 UNIT(s),
279 &s->kill_context,
280 state != SCOPE_STOP_SIGTERM ? KILL_KILL :
281 s->was_abandoned ? KILL_TERMINATE_AND_LOG :
282 KILL_TERMINATE,
283 -1, -1, false);
284 if (r < 0)
285 goto fail;
286 }
287
288 if (r > 0) {
289 r = scope_arm_timer(s, usec_add(now(CLOCK_MONOTONIC), s->timeout_stop_usec));
290 if (r < 0)
291 goto fail;
292
293 scope_set_state(s, state);
294 } else if (state == SCOPE_STOP_SIGTERM)
295 scope_enter_signal(s, SCOPE_STOP_SIGKILL, SCOPE_SUCCESS);
296 else
297 scope_enter_dead(s, SCOPE_SUCCESS);
298
299 return;
300
301 fail:
302 log_unit_warning_errno(UNIT(s), r, "Failed to kill processes: %m");
303
304 scope_enter_dead(s, SCOPE_FAILURE_RESOURCES);
305 }
306
307 static int scope_start(Unit *u) {
308 Scope *s = SCOPE(u);
309 int r;
310
311 assert(s);
312
313 if (unit_has_name(u, SPECIAL_INIT_SCOPE))
314 return -EPERM;
315
316 if (s->state == SCOPE_FAILED)
317 return -EPERM;
318
319 /* We can't fulfill this right now, please try again later */
320 if (IN_SET(s->state, SCOPE_STOP_SIGTERM, SCOPE_STOP_SIGKILL))
321 return -EAGAIN;
322
323 assert(s->state == SCOPE_DEAD);
324
325 if (!u->transient && !MANAGER_IS_RELOADING(u->manager))
326 return -ENOENT;
327
328 (void) bus_scope_track_controller(s);
329
330 r = unit_acquire_invocation_id(u);
331 if (r < 0)
332 return r;
333
334 (void) unit_realize_cgroup(u);
335 (void) unit_reset_cpu_accounting(u);
336 (void) unit_reset_ip_accounting(u);
337
338 unit_export_state_files(UNIT(s));
339
340 r = unit_attach_pids_to_cgroup(u, UNIT(s)->pids, NULL);
341 if (r < 0) {
342 log_unit_warning_errno(UNIT(s), r, "Failed to add PIDs to scope's control group: %m");
343 scope_enter_dead(s, SCOPE_FAILURE_RESOURCES);
344 return r;
345 }
346
347 s->result = SCOPE_SUCCESS;
348
349 scope_set_state(s, SCOPE_RUNNING);
350
351 /* Start watching the PIDs currently in the scope */
352 (void) unit_enqueue_rewatch_pids(UNIT(s));
353 return 1;
354 }
355
356 static int scope_stop(Unit *u) {
357 Scope *s = SCOPE(u);
358
359 assert(s);
360
361 if (IN_SET(s->state, SCOPE_STOP_SIGTERM, SCOPE_STOP_SIGKILL))
362 return 0;
363
364 assert(IN_SET(s->state, SCOPE_RUNNING, SCOPE_ABANDONED));
365
366 scope_enter_signal(s, SCOPE_STOP_SIGTERM, SCOPE_SUCCESS);
367 return 1;
368 }
369
370 static void scope_reset_failed(Unit *u) {
371 Scope *s = SCOPE(u);
372
373 assert(s);
374
375 if (s->state == SCOPE_FAILED)
376 scope_set_state(s, SCOPE_DEAD);
377
378 s->result = SCOPE_SUCCESS;
379 }
380
381 static int scope_kill(Unit *u, KillWho who, int signo, sd_bus_error *error) {
382 return unit_kill_common(u, who, signo, -1, -1, error);
383 }
384
385 static int scope_get_timeout(Unit *u, usec_t *timeout) {
386 Scope *s = SCOPE(u);
387 usec_t t;
388 int r;
389
390 if (!s->timer_event_source)
391 return 0;
392
393 r = sd_event_source_get_time(s->timer_event_source, &t);
394 if (r < 0)
395 return r;
396 if (t == USEC_INFINITY)
397 return 0;
398
399 *timeout = t;
400 return 1;
401 }
402
403 static int scope_serialize(Unit *u, FILE *f, FDSet *fds) {
404 Scope *s = SCOPE(u);
405
406 assert(s);
407 assert(f);
408 assert(fds);
409
410 unit_serialize_item(u, f, "state", scope_state_to_string(s->state));
411 unit_serialize_item(u, f, "was-abandoned", yes_no(s->was_abandoned));
412
413 if (s->controller)
414 unit_serialize_item(u, f, "controller", s->controller);
415
416 return 0;
417 }
418
419 static int scope_deserialize_item(Unit *u, const char *key, const char *value, FDSet *fds) {
420 Scope *s = SCOPE(u);
421 int r;
422
423 assert(u);
424 assert(key);
425 assert(value);
426 assert(fds);
427
428 if (streq(key, "state")) {
429 ScopeState state;
430
431 state = scope_state_from_string(value);
432 if (state < 0)
433 log_unit_debug(u, "Failed to parse state value: %s", value);
434 else
435 s->deserialized_state = state;
436
437 } else if (streq(key, "was-abandoned")) {
438 int k;
439
440 k = parse_boolean(value);
441 if (k < 0)
442 log_unit_debug(u, "Failed to parse boolean value: %s", value);
443 else
444 s->was_abandoned = k;
445 } else if (streq(key, "controller")) {
446
447 r = free_and_strdup(&s->controller, value);
448 if (r < 0)
449 log_oom();
450
451 } else
452 log_unit_debug(u, "Unknown serialization key: %s", key);
453
454 return 0;
455 }
456
457 static void scope_notify_cgroup_empty_event(Unit *u) {
458 Scope *s = SCOPE(u);
459 assert(u);
460
461 log_unit_debug(u, "cgroup is empty");
462
463 if (IN_SET(s->state, SCOPE_RUNNING, SCOPE_ABANDONED, SCOPE_STOP_SIGTERM, SCOPE_STOP_SIGKILL))
464 scope_enter_dead(s, SCOPE_SUCCESS);
465 }
466
467 static void scope_sigchld_event(Unit *u, pid_t pid, int code, int status) {
468 assert(u);
469
470 /* If we get a SIGCHLD event for one of the processes we were interested in, then we look for others to
471 * watch, under the assumption that we'll sooner or later get a SIGCHLD for them, as the original
472 * process we watched was probably the parent of them, and they are hence now our children. */
473
474 (void) unit_enqueue_rewatch_pids(u);
475 }
476
477 static int scope_dispatch_timer(sd_event_source *source, usec_t usec, void *userdata) {
478 Scope *s = SCOPE(userdata);
479
480 assert(s);
481 assert(s->timer_event_source == source);
482
483 switch (s->state) {
484
485 case SCOPE_STOP_SIGTERM:
486 if (s->kill_context.send_sigkill) {
487 log_unit_warning(UNIT(s), "Stopping timed out. Killing.");
488 scope_enter_signal(s, SCOPE_STOP_SIGKILL, SCOPE_FAILURE_TIMEOUT);
489 } else {
490 log_unit_warning(UNIT(s), "Stopping timed out. Skipping SIGKILL.");
491 scope_enter_dead(s, SCOPE_FAILURE_TIMEOUT);
492 }
493
494 break;
495
496 case SCOPE_STOP_SIGKILL:
497 log_unit_warning(UNIT(s), "Still around after SIGKILL. Ignoring.");
498 scope_enter_dead(s, SCOPE_FAILURE_TIMEOUT);
499 break;
500
501 default:
502 assert_not_reached("Timeout at wrong time.");
503 }
504
505 return 0;
506 }
507
508 int scope_abandon(Scope *s) {
509 assert(s);
510
511 if (unit_has_name(UNIT(s), SPECIAL_INIT_SCOPE))
512 return -EPERM;
513
514 if (!IN_SET(s->state, SCOPE_RUNNING, SCOPE_ABANDONED))
515 return -ESTALE;
516
517 s->was_abandoned = true;
518
519 s->controller = mfree(s->controller);
520 s->controller_track = sd_bus_track_unref(s->controller_track);
521
522 scope_set_state(s, SCOPE_ABANDONED);
523
524 /* The client is no longer watching the remaining processes, so let's step in here, under the assumption that
525 * the remaining processes will be sooner or later reassigned to us as parent. */
526 (void) unit_enqueue_rewatch_pids(UNIT(s));
527
528 return 0;
529 }
530
531 _pure_ static UnitActiveState scope_active_state(Unit *u) {
532 assert(u);
533
534 return state_translation_table[SCOPE(u)->state];
535 }
536
537 _pure_ static const char *scope_sub_state_to_string(Unit *u) {
538 assert(u);
539
540 return scope_state_to_string(SCOPE(u)->state);
541 }
542
543 static void scope_enumerate_perpetual(Manager *m) {
544 Unit *u;
545 int r;
546
547 assert(m);
548
549 /* Let's unconditionally add the "init.scope" special unit
550 * that encapsulates PID 1. Note that PID 1 already is in the
551 * cgroup for this, we hence just need to allocate the object
552 * for it and that's it. */
553
554 u = manager_get_unit(m, SPECIAL_INIT_SCOPE);
555 if (!u) {
556 r = unit_new_for_name(m, sizeof(Scope), SPECIAL_INIT_SCOPE, &u);
557 if (r < 0) {
558 log_error_errno(r, "Failed to allocate the special " SPECIAL_INIT_SCOPE " unit: %m");
559 return;
560 }
561 }
562
563 u->transient = true;
564 u->perpetual = true;
565 SCOPE(u)->deserialized_state = SCOPE_RUNNING;
566
567 unit_add_to_load_queue(u);
568 unit_add_to_dbus_queue(u);
569 }
570
571 static const char* const scope_result_table[_SCOPE_RESULT_MAX] = {
572 [SCOPE_SUCCESS] = "success",
573 [SCOPE_FAILURE_RESOURCES] = "resources",
574 [SCOPE_FAILURE_TIMEOUT] = "timeout",
575 };
576
577 DEFINE_STRING_TABLE_LOOKUP(scope_result, ScopeResult);
578
579 const UnitVTable scope_vtable = {
580 .object_size = sizeof(Scope),
581 .cgroup_context_offset = offsetof(Scope, cgroup_context),
582 .kill_context_offset = offsetof(Scope, kill_context),
583
584 .sections =
585 "Unit\0"
586 "Scope\0"
587 "Install\0",
588 .private_section = "Scope",
589
590 .can_transient = true,
591 .can_delegate = true,
592 .once_only = true,
593
594 .init = scope_init,
595 .load = scope_load,
596 .done = scope_done,
597
598 .coldplug = scope_coldplug,
599
600 .dump = scope_dump,
601
602 .start = scope_start,
603 .stop = scope_stop,
604
605 .kill = scope_kill,
606
607 .get_timeout = scope_get_timeout,
608
609 .serialize = scope_serialize,
610 .deserialize_item = scope_deserialize_item,
611
612 .active_state = scope_active_state,
613 .sub_state_to_string = scope_sub_state_to_string,
614
615 .sigchld_event = scope_sigchld_event,
616
617 .reset_failed = scope_reset_failed,
618
619 .notify_cgroup_empty = scope_notify_cgroup_empty_event,
620
621 .bus_vtable = bus_scope_vtable,
622 .bus_set_property = bus_scope_set_property,
623 .bus_commit_properties = bus_scope_commit_properties,
624
625 .enumerate_perpetual = scope_enumerate_perpetual,
626 };