]> git.ipfire.org Git - thirdparty/systemd.git/blob - src/basic/cgroup-util.c
projects / thirdparty / systemd.git / blob
? search:


0ecb8d944e98ea133dd25cb47cce5a17ab6efda0
[thirdparty/systemd.git] / src / basic / cgroup-util.c
1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
2
3 #include <errno.h>
4 #include <ftw.h>
5 #include <limits.h>
6 #include <signal.h>
7 #include <stddef.h>
8 #include <stdlib.h>
9 #include <sys/types.h>
10 #include <sys/utsname.h>
11 #include <sys/xattr.h>
12 #include <unistd.h>
13
14 #include "alloc-util.h"
15 #include "cgroup-util.h"
16 #include "def.h"
17 #include "dirent-util.h"
18 #include "extract-word.h"
19 #include "fd-util.h"
20 #include "fileio.h"
21 #include "format-util.h"
22 #include "fs-util.h"
23 #include "log.h"
24 #include "login-util.h"
25 #include "macro.h"
26 #include "missing_magic.h"
27 #include "mkdir.h"
28 #include "parse-util.h"
29 #include "path-util.h"
30 #include "process-util.h"
31 #include "set.h"
32 #include "special.h"
33 #include "stat-util.h"
34 #include "stdio-util.h"
35 #include "string-table.h"
36 #include "string-util.h"
37 #include "strv.h"
38 #include "unit-name.h"
39 #include "user-util.h"
40 #include "xattr-util.h"
41
42 static int cg_enumerate_items(const char *controller, const char *path, FILE **_f, const char *item) {
43 _cleanup_free_ char *fs = NULL;
44 FILE *f;
45 int r;
46
47 assert(_f);
48
49 r = cg_get_path(controller, path, item, &fs);
50 if (r < 0)
51 return r;
52
53 f = fopen(fs, "re");
54 if (!f)
55 return -errno;
56
57 *_f = f;
58 return 0;
59 }
60
61 int cg_enumerate_processes(const char *controller, const char *path, FILE **_f) {
62 return cg_enumerate_items(controller, path, _f, "cgroup.procs");
63 }
64
65 int cg_read_pid(FILE *f, pid_t *_pid) {
66 unsigned long ul;
67
68 /* Note that the cgroup.procs might contain duplicates! See
69 * cgroups.txt for details. */
70
71 assert(f);
72 assert(_pid);
73
74 errno = 0;
75 if (fscanf(f, "%lu", &ul) != 1) {
76
77 if (feof(f))
78 return 0;
79
80 return errno_or_else(EIO);
81 }
82
83 if (ul <= 0)
84 return -EIO;
85
86 *_pid = (pid_t) ul;
87 return 1;
88 }
89
90 int cg_read_event(
91 const char *controller,
92 const char *path,
93 const char *event,
94 char **ret) {
95
96 _cleanup_free_ char *events = NULL, *content = NULL;
97 int r;
98
99 r = cg_get_path(controller, path, "cgroup.events", &events);
100 if (r < 0)
101 return r;
102
103 r = read_full_file(events, &content, NULL);
104 if (r < 0)
105 return r;
106
107 for (const char *p = content;;) {
108 _cleanup_free_ char *line = NULL, *key = NULL, *val = NULL;
109 const char *q;
110
111 r = extract_first_word(&p, &line, "\n", 0);
112 if (r < 0)
113 return r;
114 if (r == 0)
115 return -ENOENT;
116
117 q = line;
118 r = extract_first_word(&q, &key, " ", 0);
119 if (r < 0)
120 return r;
121 if (r == 0)
122 return -EINVAL;
123
124 if (!streq(key, event))
125 continue;
126
127 val = strdup(q);
128 if (!val)
129 return -ENOMEM;
130
131 *ret = TAKE_PTR(val);
132 return 0;
133 }
134 }
135
136 bool cg_ns_supported(void) {
137 static thread_local int enabled = -1;
138
139 if (enabled >= 0)
140 return enabled;
141
142 if (access("/proc/self/ns/cgroup", F_OK) < 0) {
143 if (errno != ENOENT)
144 log_debug_errno(errno, "Failed to check whether /proc/self/ns/cgroup is available, assuming not: %m");
145 enabled = false;
146 } else
147 enabled = true;
148
149 return enabled;
150 }
151
152 bool cg_freezer_supported(void) {
153 static thread_local int supported = -1;
154
155 if (supported >= 0)
156 return supported;
157
158 supported = cg_all_unified() > 0 && access("/sys/fs/cgroup/init.scope/cgroup.freeze", F_OK) == 0;
159
160 return supported;
161 }
162
163 int cg_enumerate_subgroups(const char *controller, const char *path, DIR **_d) {
164 _cleanup_free_ char *fs = NULL;
165 int r;
166 DIR *d;
167
168 assert(_d);
169
170 /* This is not recursive! */
171
172 r = cg_get_path(controller, path, NULL, &fs);
173 if (r < 0)
174 return r;
175
176 d = opendir(fs);
177 if (!d)
178 return -errno;
179
180 *_d = d;
181 return 0;
182 }
183
184 int cg_read_subgroup(DIR *d, char **fn) {
185 struct dirent *de;
186
187 assert(d);
188 assert(fn);
189
190 FOREACH_DIRENT_ALL(de, d, return -errno) {
191 char *b;
192
193 if (de->d_type != DT_DIR)
194 continue;
195
196 if (dot_or_dot_dot(de->d_name))
197 continue;
198
199 b = strdup(de->d_name);
200 if (!b)
201 return -ENOMEM;
202
203 *fn = b;
204 return 1;
205 }
206
207 return 0;
208 }
209
210 int cg_rmdir(const char *controller, const char *path) {
211 _cleanup_free_ char *p = NULL;
212 int r;
213
214 r = cg_get_path(controller, path, NULL, &p);
215 if (r < 0)
216 return r;
217
218 r = rmdir(p);
219 if (r < 0 && errno != ENOENT)
220 return -errno;
221
222 r = cg_hybrid_unified();
223 if (r <= 0)
224 return r;
225
226 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
227 r = cg_rmdir(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path);
228 if (r < 0)
229 log_warning_errno(r, "Failed to remove compat systemd cgroup %s: %m", path);
230 }
231
232 return 0;
233 }
234
235 static int cg_kill_items(
236 const char *controller,
237 const char *path,
238 int sig,
239 CGroupFlags flags,
240 Set *s,
241 cg_kill_log_func_t log_kill,
242 void *userdata,
243 const char *item) {
244
245 _cleanup_set_free_ Set *allocated_set = NULL;
246 bool done = false;
247 int r, ret = 0, ret_log_kill = 0;
248 pid_t my_pid;
249
250 assert(sig >= 0);
251
252 /* Don't send SIGCONT twice. Also, SIGKILL always works even when process is suspended, hence don't send
253 * SIGCONT on SIGKILL. */
254 if (IN_SET(sig, SIGCONT, SIGKILL))
255 flags &= ~CGROUP_SIGCONT;
256
257 /* This goes through the tasks list and kills them all. This
258 * is repeated until no further processes are added to the
259 * tasks list, to properly handle forking processes */
260
261 if (!s) {
262 s = allocated_set = set_new(NULL);
263 if (!s)
264 return -ENOMEM;
265 }
266
267 my_pid = getpid_cached();
268
269 do {
270 _cleanup_fclose_ FILE *f = NULL;
271 pid_t pid = 0;
272 done = true;
273
274 r = cg_enumerate_items(controller, path, &f, item);
275 if (r < 0) {
276 if (ret >= 0 && r != -ENOENT)
277 return r;
278
279 return ret;
280 }
281
282 while ((r = cg_read_pid(f, &pid)) > 0) {
283
284 if ((flags & CGROUP_IGNORE_SELF) && pid == my_pid)
285 continue;
286
287 if (set_get(s, PID_TO_PTR(pid)) == PID_TO_PTR(pid))
288 continue;
289
290 if (log_kill)
291 ret_log_kill = log_kill(pid, sig, userdata);
292
293 /* If we haven't killed this process yet, kill
294 * it */
295 if (kill(pid, sig) < 0) {
296 if (ret >= 0 && errno != ESRCH)
297 ret = -errno;
298 } else {
299 if (flags & CGROUP_SIGCONT)
300 (void) kill(pid, SIGCONT);
301
302 if (ret == 0) {
303 if (log_kill)
304 ret = ret_log_kill;
305 else
306 ret = 1;
307 }
308 }
309
310 done = false;
311
312 r = set_put(s, PID_TO_PTR(pid));
313 if (r < 0) {
314 if (ret >= 0)
315 return r;
316
317 return ret;
318 }
319 }
320
321 if (r < 0) {
322 if (ret >= 0)
323 return r;
324
325 return ret;
326 }
327
328 /* To avoid racing against processes which fork
329 * quicker than we can kill them we repeat this until
330 * no new pids need to be killed. */
331
332 } while (!done);
333
334 return ret;
335 }
336
337 int cg_kill(
338 const char *controller,
339 const char *path,
340 int sig,
341 CGroupFlags flags,
342 Set *s,
343 cg_kill_log_func_t log_kill,
344 void *userdata) {
345 int r;
346
347 r = cg_kill_items(controller, path, sig, flags, s, log_kill, userdata, "cgroup.procs");
348 if (r < 0 || sig != SIGKILL)
349 return r;
350
351 /* Only in case of killing with SIGKILL and when using cgroupsv2, kill remaining threads manually as
352 a workaround for kernel bug. It was fixed in 5.2-rc5 (c03cd7738a83), backported to 4.19.66
353 (4340d175b898) and 4.14.138 (feb6b123b7dd). */
354 r = cg_unified_controller(controller);
355 if (r <= 0)
356 return r;
357
358 return cg_kill_items(controller, path, sig, flags, s, log_kill, userdata, "cgroup.threads");
359 }
360
361 int cg_kill_recursive(
362 const char *controller,
363 const char *path,
364 int sig,
365 CGroupFlags flags,
366 Set *s,
367 cg_kill_log_func_t log_kill,
368 void *userdata) {
369
370 _cleanup_set_free_ Set *allocated_set = NULL;
371 _cleanup_closedir_ DIR *d = NULL;
372 int r, ret;
373 char *fn;
374
375 assert(path);
376 assert(sig >= 0);
377
378 if (!s) {
379 s = allocated_set = set_new(NULL);
380 if (!s)
381 return -ENOMEM;
382 }
383
384 ret = cg_kill(controller, path, sig, flags, s, log_kill, userdata);
385
386 r = cg_enumerate_subgroups(controller, path, &d);
387 if (r < 0) {
388 if (ret >= 0 && r != -ENOENT)
389 return r;
390
391 return ret;
392 }
393
394 while ((r = cg_read_subgroup(d, &fn)) > 0) {
395 _cleanup_free_ char *p = NULL;
396
397 p = path_join(empty_to_root(path), fn);
398 free(fn);
399 if (!p)
400 return -ENOMEM;
401
402 r = cg_kill_recursive(controller, p, sig, flags, s, log_kill, userdata);
403 if (r != 0 && ret >= 0)
404 ret = r;
405 }
406 if (ret >= 0 && r < 0)
407 ret = r;
408
409 if (flags & CGROUP_REMOVE) {
410 r = cg_rmdir(controller, path);
411 if (r < 0 && ret >= 0 && !IN_SET(r, -ENOENT, -EBUSY))
412 return r;
413 }
414
415 return ret;
416 }
417
418 static const char *controller_to_dirname(const char *controller) {
419 const char *e;
420
421 assert(controller);
422
423 /* Converts a controller name to the directory name below
424 * /sys/fs/cgroup/ we want to mount it to. Effectively, this
425 * just cuts off the name= prefixed used for named
426 * hierarchies, if it is specified. */
427
428 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
429 if (cg_hybrid_unified() > 0)
430 controller = SYSTEMD_CGROUP_CONTROLLER_HYBRID;
431 else
432 controller = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
433 }
434
435 e = startswith(controller, "name=");
436 if (e)
437 return e;
438
439 return controller;
440 }
441
442 static int join_path_legacy(const char *controller, const char *path, const char *suffix, char **fs) {
443 const char *dn;
444 char *t = NULL;
445
446 assert(fs);
447 assert(controller);
448
449 dn = controller_to_dirname(controller);
450
451 if (isempty(path) && isempty(suffix))
452 t = path_join("/sys/fs/cgroup", dn);
453 else if (isempty(path))
454 t = path_join("/sys/fs/cgroup", dn, suffix);
455 else if (isempty(suffix))
456 t = path_join("/sys/fs/cgroup", dn, path);
457 else
458 t = path_join("/sys/fs/cgroup", dn, path, suffix);
459 if (!t)
460 return -ENOMEM;
461
462 *fs = t;
463 return 0;
464 }
465
466 static int join_path_unified(const char *path, const char *suffix, char **fs) {
467 char *t;
468
469 assert(fs);
470
471 if (isempty(path) && isempty(suffix))
472 t = strdup("/sys/fs/cgroup");
473 else if (isempty(path))
474 t = path_join("/sys/fs/cgroup", suffix);
475 else if (isempty(suffix))
476 t = path_join("/sys/fs/cgroup", path);
477 else
478 t = path_join("/sys/fs/cgroup", path, suffix);
479 if (!t)
480 return -ENOMEM;
481
482 *fs = t;
483 return 0;
484 }
485
486 int cg_get_path(const char *controller, const char *path, const char *suffix, char **fs) {
487 int r;
488
489 assert(fs);
490
491 if (!controller) {
492 char *t;
493
494 /* If no controller is specified, we return the path
495 * *below* the controllers, without any prefix. */
496
497 if (!path && !suffix)
498 return -EINVAL;
499
500 if (!suffix)
501 t = strdup(path);
502 else if (!path)
503 t = strdup(suffix);
504 else
505 t = path_join(path, suffix);
506 if (!t)
507 return -ENOMEM;
508
509 *fs = path_simplify(t, false);
510 return 0;
511 }
512
513 if (!cg_controller_is_valid(controller))
514 return -EINVAL;
515
516 r = cg_all_unified();
517 if (r < 0)
518 return r;
519 if (r > 0)
520 r = join_path_unified(path, suffix, fs);
521 else
522 r = join_path_legacy(controller, path, suffix, fs);
523 if (r < 0)
524 return r;
525
526 path_simplify(*fs, false);
527 return 0;
528 }
529
530 static int controller_is_accessible(const char *controller) {
531 int r;
532
533 assert(controller);
534
535 /* Checks whether a specific controller is accessible,
536 * i.e. its hierarchy mounted. In the unified hierarchy all
537 * controllers are considered accessible, except for the named
538 * hierarchies */
539
540 if (!cg_controller_is_valid(controller))
541 return -EINVAL;
542
543 r = cg_all_unified();
544 if (r < 0)
545 return r;
546 if (r > 0) {
547 /* We don't support named hierarchies if we are using
548 * the unified hierarchy. */
549
550 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
551 return 0;
552
553 if (startswith(controller, "name="))
554 return -EOPNOTSUPP;
555
556 } else {
557 const char *cc, *dn;
558
559 dn = controller_to_dirname(controller);
560 cc = strjoina("/sys/fs/cgroup/", dn);
561
562 if (laccess(cc, F_OK) < 0)
563 return -errno;
564 }
565
566 return 0;
567 }
568
569 int cg_get_path_and_check(const char *controller, const char *path, const char *suffix, char **fs) {
570 int r;
571
572 assert(controller);
573 assert(fs);
574
575 /* Check if the specified controller is actually accessible */
576 r = controller_is_accessible(controller);
577 if (r < 0)
578 return r;
579
580 return cg_get_path(controller, path, suffix, fs);
581 }
582
583 int cg_set_xattr(const char *controller, const char *path, const char *name, const void *value, size_t size, int flags) {
584 _cleanup_free_ char *fs = NULL;
585 int r;
586
587 assert(path);
588 assert(name);
589 assert(value || size <= 0);
590
591 r = cg_get_path(controller, path, NULL, &fs);
592 if (r < 0)
593 return r;
594
595 if (setxattr(fs, name, value, size, flags) < 0)
596 return -errno;
597
598 return 0;
599 }
600
601 int cg_get_xattr(const char *controller, const char *path, const char *name, void *value, size_t size) {
602 _cleanup_free_ char *fs = NULL;
603 ssize_t n;
604 int r;
605
606 assert(path);
607 assert(name);
608
609 r = cg_get_path(controller, path, NULL, &fs);
610 if (r < 0)
611 return r;
612
613 n = getxattr(fs, name, value, size);
614 if (n < 0)
615 return -errno;
616
617 return (int) n;
618 }
619
620 int cg_get_xattr_malloc(const char *controller, const char *path, const char *name, char **ret) {
621 _cleanup_free_ char *fs = NULL;
622 int r;
623
624 assert(path);
625 assert(name);
626
627 r = cg_get_path(controller, path, NULL, &fs);
628 if (r < 0)
629 return r;
630
631 r = getxattr_malloc(fs, name, ret, false);
632 if (r < 0)
633 return r;
634
635 return r;
636 }
637
638 int cg_get_xattr_bool(const char *controller, const char *path, const char *name) {
639 _cleanup_free_ char *val = NULL;
640 int r;
641
642 assert(path);
643 assert(name);
644
645 r = cg_get_xattr_malloc(controller, path, name, &val);
646 if (r < 0)
647 return r;
648
649 return parse_boolean(val);
650 }
651
652 int cg_remove_xattr(const char *controller, const char *path, const char *name) {
653 _cleanup_free_ char *fs = NULL;
654 int r;
655
656 assert(path);
657 assert(name);
658
659 r = cg_get_path(controller, path, NULL, &fs);
660 if (r < 0)
661 return r;
662
663 if (removexattr(fs, name) < 0)
664 return -errno;
665
666 return 0;
667 }
668
669 int cg_pid_get_path(const char *controller, pid_t pid, char **ret_path) {
670 _cleanup_fclose_ FILE *f = NULL;
671 const char *fs, *controller_str;
672 int unified, r;
673
674 assert(pid >= 0);
675 assert(ret_path);
676
677 if (controller) {
678 if (!cg_controller_is_valid(controller))
679 return -EINVAL;
680 } else
681 controller = SYSTEMD_CGROUP_CONTROLLER;
682
683 unified = cg_unified_controller(controller);
684 if (unified < 0)
685 return unified;
686 if (unified == 0) {
687 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
688 controller_str = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
689 else
690 controller_str = controller;
691 }
692
693 fs = procfs_file_alloca(pid, "cgroup");
694 r = fopen_unlocked(fs, "re", &f);
695 if (r == -ENOENT)
696 return -ESRCH;
697 if (r < 0)
698 return r;
699
700 for (;;) {
701 _cleanup_free_ char *line = NULL;
702 char *e;
703
704 r = read_line(f, LONG_LINE_MAX, &line);
705 if (r < 0)
706 return r;
707 if (r == 0)
708 return -ENODATA;
709
710 if (unified) {
711 e = startswith(line, "0:");
712 if (!e)
713 continue;
714
715 e = strchr(e, ':');
716 if (!e)
717 continue;
718 } else {
719 char *l;
720
721 l = strchr(line, ':');
722 if (!l)
723 continue;
724
725 l++;
726 e = strchr(l, ':');
727 if (!e)
728 continue;
729 *e = 0;
730
731 r = string_contains_word(l, ",", controller_str);
732 if (r < 0)
733 return r;
734 if (r == 0)
735 continue;
736 }
737
738 char *path = strdup(e + 1);
739 if (!path)
740 return -ENOMEM;
741
742 /* Truncate suffix indicating the process is a zombie */
743 e = endswith(path, " (deleted)");
744 if (e)
745 *e = 0;
746
747 *ret_path = path;
748 return 0;
749 }
750 }
751
752 int cg_install_release_agent(const char *controller, const char *agent) {
753 _cleanup_free_ char *fs = NULL, *contents = NULL;
754 const char *sc;
755 int r;
756
757 assert(agent);
758
759 r = cg_unified_controller(controller);
760 if (r < 0)
761 return r;
762 if (r > 0) /* doesn't apply to unified hierarchy */
763 return -EOPNOTSUPP;
764
765 r = cg_get_path(controller, NULL, "release_agent", &fs);
766 if (r < 0)
767 return r;
768
769 r = read_one_line_file(fs, &contents);
770 if (r < 0)
771 return r;
772
773 sc = strstrip(contents);
774 if (isempty(sc)) {
775 r = write_string_file(fs, agent, WRITE_STRING_FILE_DISABLE_BUFFER);
776 if (r < 0)
777 return r;
778 } else if (!path_equal(sc, agent))
779 return -EEXIST;
780
781 fs = mfree(fs);
782 r = cg_get_path(controller, NULL, "notify_on_release", &fs);
783 if (r < 0)
784 return r;
785
786 contents = mfree(contents);
787 r = read_one_line_file(fs, &contents);
788 if (r < 0)
789 return r;
790
791 sc = strstrip(contents);
792 if (streq(sc, "0")) {
793 r = write_string_file(fs, "1", WRITE_STRING_FILE_DISABLE_BUFFER);
794 if (r < 0)
795 return r;
796
797 return 1;
798 }
799
800 if (!streq(sc, "1"))
801 return -EIO;
802
803 return 0;
804 }
805
806 int cg_uninstall_release_agent(const char *controller) {
807 _cleanup_free_ char *fs = NULL;
808 int r;
809
810 r = cg_unified_controller(controller);
811 if (r < 0)
812 return r;
813 if (r > 0) /* Doesn't apply to unified hierarchy */
814 return -EOPNOTSUPP;
815
816 r = cg_get_path(controller, NULL, "notify_on_release", &fs);
817 if (r < 0)
818 return r;
819
820 r = write_string_file(fs, "0", WRITE_STRING_FILE_DISABLE_BUFFER);
821 if (r < 0)
822 return r;
823
824 fs = mfree(fs);
825
826 r = cg_get_path(controller, NULL, "release_agent", &fs);
827 if (r < 0)
828 return r;
829
830 r = write_string_file(fs, "", WRITE_STRING_FILE_DISABLE_BUFFER);
831 if (r < 0)
832 return r;
833
834 return 0;
835 }
836
837 int cg_is_empty(const char *controller, const char *path) {
838 _cleanup_fclose_ FILE *f = NULL;
839 pid_t pid;
840 int r;
841
842 assert(path);
843
844 r = cg_enumerate_processes(controller, path, &f);
845 if (r == -ENOENT)
846 return true;
847 if (r < 0)
848 return r;
849
850 r = cg_read_pid(f, &pid);
851 if (r < 0)
852 return r;
853
854 return r == 0;
855 }
856
857 int cg_is_empty_recursive(const char *controller, const char *path) {
858 int r;
859
860 assert(path);
861
862 /* The root cgroup is always populated */
863 if (controller && empty_or_root(path))
864 return false;
865
866 r = cg_unified_controller(controller);
867 if (r < 0)
868 return r;
869 if (r > 0) {
870 _cleanup_free_ char *t = NULL;
871
872 /* On the unified hierarchy we can check empty state
873 * via the "populated" attribute of "cgroup.events". */
874
875 r = cg_read_event(controller, path, "populated", &t);
876 if (r == -ENOENT)
877 return true;
878 if (r < 0)
879 return r;
880
881 return streq(t, "0");
882 } else {
883 _cleanup_closedir_ DIR *d = NULL;
884 char *fn;
885
886 r = cg_is_empty(controller, path);
887 if (r <= 0)
888 return r;
889
890 r = cg_enumerate_subgroups(controller, path, &d);
891 if (r == -ENOENT)
892 return true;
893 if (r < 0)
894 return r;
895
896 while ((r = cg_read_subgroup(d, &fn)) > 0) {
897 _cleanup_free_ char *p = NULL;
898
899 p = path_join(path, fn);
900 free(fn);
901 if (!p)
902 return -ENOMEM;
903
904 r = cg_is_empty_recursive(controller, p);
905 if (r <= 0)
906 return r;
907 }
908 if (r < 0)
909 return r;
910
911 return true;
912 }
913 }
914
915 int cg_split_spec(const char *spec, char **ret_controller, char **ret_path) {
916 _cleanup_free_ char *controller = NULL, *path = NULL;
917
918 assert(spec);
919
920 if (*spec == '/') {
921 if (!path_is_normalized(spec))
922 return -EINVAL;
923
924 if (ret_path) {
925 path = strdup(spec);
926 if (!path)
927 return -ENOMEM;
928
929 path_simplify(path, false);
930 }
931
932 } else {
933 const char *e;
934
935 e = strchr(spec, ':');
936 if (e) {
937 controller = strndup(spec, e-spec);
938 if (!controller)
939 return -ENOMEM;
940 if (!cg_controller_is_valid(controller))
941 return -EINVAL;
942
943 if (!isempty(e + 1)) {
944 path = strdup(e+1);
945 if (!path)
946 return -ENOMEM;
947
948 if (!path_is_normalized(path) ||
949 !path_is_absolute(path))
950 return -EINVAL;
951
952 path_simplify(path, false);
953 }
954
955 } else {
956 if (!cg_controller_is_valid(spec))
957 return -EINVAL;
958
959 if (ret_controller) {
960 controller = strdup(spec);
961 if (!controller)
962 return -ENOMEM;
963 }
964 }
965 }
966
967 if (ret_controller)
968 *ret_controller = TAKE_PTR(controller);
969 if (ret_path)
970 *ret_path = TAKE_PTR(path);
971 return 0;
972 }
973
974 int cg_mangle_path(const char *path, char **result) {
975 _cleanup_free_ char *c = NULL, *p = NULL;
976 char *t;
977 int r;
978
979 assert(path);
980 assert(result);
981
982 /* First, check if it already is a filesystem path */
983 if (path_startswith(path, "/sys/fs/cgroup")) {
984
985 t = strdup(path);
986 if (!t)
987 return -ENOMEM;
988
989 *result = path_simplify(t, false);
990 return 0;
991 }
992
993 /* Otherwise, treat it as cg spec */
994 r = cg_split_spec(path, &c, &p);
995 if (r < 0)
996 return r;
997
998 return cg_get_path(c ?: SYSTEMD_CGROUP_CONTROLLER, p ?: "/", NULL, result);
999 }
1000
1001 int cg_get_root_path(char **path) {
1002 char *p, *e;
1003 int r;
1004
1005 assert(path);
1006
1007 r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 1, &p);
1008 if (r < 0)
1009 return r;
1010
1011 e = endswith(p, "/" SPECIAL_INIT_SCOPE);
1012 if (!e)
1013 e = endswith(p, "/" SPECIAL_SYSTEM_SLICE); /* legacy */
1014 if (!e)
1015 e = endswith(p, "/system"); /* even more legacy */
1016 if (e)
1017 *e = 0;
1018
1019 *path = p;
1020 return 0;
1021 }
1022
1023 int cg_shift_path(const char *cgroup, const char *root, const char **shifted) {
1024 _cleanup_free_ char *rt = NULL;
1025 char *p;
1026 int r;
1027
1028 assert(cgroup);
1029 assert(shifted);
1030
1031 if (!root) {
1032 /* If the root was specified let's use that, otherwise
1033 * let's determine it from PID 1 */
1034
1035 r = cg_get_root_path(&rt);
1036 if (r < 0)
1037 return r;
1038
1039 root = rt;
1040 }
1041
1042 p = path_startswith(cgroup, root);
1043 if (p && p > cgroup)
1044 *shifted = p - 1;
1045 else
1046 *shifted = cgroup;
1047
1048 return 0;
1049 }
1050
1051 int cg_pid_get_path_shifted(pid_t pid, const char *root, char **cgroup) {
1052 _cleanup_free_ char *raw = NULL;
1053 const char *c;
1054 int r;
1055
1056 assert(pid >= 0);
1057 assert(cgroup);
1058
1059 r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &raw);
1060 if (r < 0)
1061 return r;
1062
1063 r = cg_shift_path(raw, root, &c);
1064 if (r < 0)
1065 return r;
1066
1067 if (c == raw)
1068 *cgroup = TAKE_PTR(raw);
1069 else {
1070 char *n;
1071
1072 n = strdup(c);
1073 if (!n)
1074 return -ENOMEM;
1075
1076 *cgroup = n;
1077 }
1078
1079 return 0;
1080 }
1081
1082 int cg_path_decode_unit(const char *cgroup, char **unit) {
1083 char *c, *s;
1084 size_t n;
1085
1086 assert(cgroup);
1087 assert(unit);
1088
1089 n = strcspn(cgroup, "/");
1090 if (n < 3)
1091 return -ENXIO;
1092
1093 c = strndupa(cgroup, n);
1094 c = cg_unescape(c);
1095
1096 if (!unit_name_is_valid(c, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1097 return -ENXIO;
1098
1099 s = strdup(c);
1100 if (!s)
1101 return -ENOMEM;
1102
1103 *unit = s;
1104 return 0;
1105 }
1106
1107 static bool valid_slice_name(const char *p, size_t n) {
1108
1109 if (!p)
1110 return false;
1111
1112 if (n < STRLEN("x.slice"))
1113 return false;
1114
1115 if (memcmp(p + n - 6, ".slice", 6) == 0) {
1116 char buf[n+1], *c;
1117
1118 memcpy(buf, p, n);
1119 buf[n] = 0;
1120
1121 c = cg_unescape(buf);
1122
1123 return unit_name_is_valid(c, UNIT_NAME_PLAIN);
1124 }
1125
1126 return false;
1127 }
1128
1129 static const char *skip_slices(const char *p) {
1130 assert(p);
1131
1132 /* Skips over all slice assignments */
1133
1134 for (;;) {
1135 size_t n;
1136
1137 p += strspn(p, "/");
1138
1139 n = strcspn(p, "/");
1140 if (!valid_slice_name(p, n))
1141 return p;
1142
1143 p += n;
1144 }
1145 }
1146
1147 int cg_path_get_unit(const char *path, char **ret) {
1148 _cleanup_free_ char *unit = NULL;
1149 const char *e;
1150 int r;
1151
1152 assert(path);
1153 assert(ret);
1154
1155 e = skip_slices(path);
1156
1157 r = cg_path_decode_unit(e, &unit);
1158 if (r < 0)
1159 return r;
1160
1161 /* We skipped over the slices, don't accept any now */
1162 if (endswith(unit, ".slice"))
1163 return -ENXIO;
1164
1165 *ret = TAKE_PTR(unit);
1166 return 0;
1167 }
1168
1169 int cg_pid_get_unit(pid_t pid, char **unit) {
1170 _cleanup_free_ char *cgroup = NULL;
1171 int r;
1172
1173 assert(unit);
1174
1175 r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1176 if (r < 0)
1177 return r;
1178
1179 return cg_path_get_unit(cgroup, unit);
1180 }
1181
1182 /**
1183 * Skip session-*.scope, but require it to be there.
1184 */
1185 static const char *skip_session(const char *p) {
1186 size_t n;
1187
1188 if (isempty(p))
1189 return NULL;
1190
1191 p += strspn(p, "/");
1192
1193 n = strcspn(p, "/");
1194 if (n < STRLEN("session-x.scope"))
1195 return NULL;
1196
1197 if (memcmp(p, "session-", 8) == 0 && memcmp(p + n - 6, ".scope", 6) == 0) {
1198 char buf[n - 8 - 6 + 1];
1199
1200 memcpy(buf, p + 8, n - 8 - 6);
1201 buf[n - 8 - 6] = 0;
1202
1203 /* Note that session scopes never need unescaping,
1204 * since they cannot conflict with the kernel's own
1205 * names, hence we don't need to call cg_unescape()
1206 * here. */
1207
1208 if (!session_id_valid(buf))
1209 return false;
1210
1211 p += n;
1212 p += strspn(p, "/");
1213 return p;
1214 }
1215
1216 return NULL;
1217 }
1218
1219 /**
1220 * Skip user@*.service, but require it to be there.
1221 */
1222 static const char *skip_user_manager(const char *p) {
1223 size_t n;
1224
1225 if (isempty(p))
1226 return NULL;
1227
1228 p += strspn(p, "/");
1229
1230 n = strcspn(p, "/");
1231 if (n < STRLEN("user@x.service"))
1232 return NULL;
1233
1234 if (memcmp(p, "user@", 5) == 0 && memcmp(p + n - 8, ".service", 8) == 0) {
1235 char buf[n - 5 - 8 + 1];
1236
1237 memcpy(buf, p + 5, n - 5 - 8);
1238 buf[n - 5 - 8] = 0;
1239
1240 /* Note that user manager services never need unescaping,
1241 * since they cannot conflict with the kernel's own
1242 * names, hence we don't need to call cg_unescape()
1243 * here. */
1244
1245 if (parse_uid(buf, NULL) < 0)
1246 return NULL;
1247
1248 p += n;
1249 p += strspn(p, "/");
1250
1251 return p;
1252 }
1253
1254 return NULL;
1255 }
1256
1257 static const char *skip_user_prefix(const char *path) {
1258 const char *e, *t;
1259
1260 assert(path);
1261
1262 /* Skip slices, if there are any */
1263 e = skip_slices(path);
1264
1265 /* Skip the user manager, if it's in the path now... */
1266 t = skip_user_manager(e);
1267 if (t)
1268 return t;
1269
1270 /* Alternatively skip the user session if it is in the path... */
1271 return skip_session(e);
1272 }
1273
1274 int cg_path_get_user_unit(const char *path, char **ret) {
1275 const char *t;
1276
1277 assert(path);
1278 assert(ret);
1279
1280 t = skip_user_prefix(path);
1281 if (!t)
1282 return -ENXIO;
1283
1284 /* And from here on it looks pretty much the same as for a system unit, hence let's use the same
1285 * parser. */
1286 return cg_path_get_unit(t, ret);
1287 }
1288
1289 int cg_pid_get_user_unit(pid_t pid, char **unit) {
1290 _cleanup_free_ char *cgroup = NULL;
1291 int r;
1292
1293 assert(unit);
1294
1295 r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1296 if (r < 0)
1297 return r;
1298
1299 return cg_path_get_user_unit(cgroup, unit);
1300 }
1301
1302 int cg_path_get_machine_name(const char *path, char **machine) {
1303 _cleanup_free_ char *u = NULL;
1304 const char *sl;
1305 int r;
1306
1307 r = cg_path_get_unit(path, &u);
1308 if (r < 0)
1309 return r;
1310
1311 sl = strjoina("/run/systemd/machines/unit:", u);
1312 return readlink_malloc(sl, machine);
1313 }
1314
1315 int cg_pid_get_machine_name(pid_t pid, char **machine) {
1316 _cleanup_free_ char *cgroup = NULL;
1317 int r;
1318
1319 assert(machine);
1320
1321 r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1322 if (r < 0)
1323 return r;
1324
1325 return cg_path_get_machine_name(cgroup, machine);
1326 }
1327
1328 int cg_path_get_session(const char *path, char **session) {
1329 _cleanup_free_ char *unit = NULL;
1330 char *start, *end;
1331 int r;
1332
1333 assert(path);
1334
1335 r = cg_path_get_unit(path, &unit);
1336 if (r < 0)
1337 return r;
1338
1339 start = startswith(unit, "session-");
1340 if (!start)
1341 return -ENXIO;
1342 end = endswith(start, ".scope");
1343 if (!end)
1344 return -ENXIO;
1345
1346 *end = 0;
1347 if (!session_id_valid(start))
1348 return -ENXIO;
1349
1350 if (session) {
1351 char *rr;
1352
1353 rr = strdup(start);
1354 if (!rr)
1355 return -ENOMEM;
1356
1357 *session = rr;
1358 }
1359
1360 return 0;
1361 }
1362
1363 int cg_pid_get_session(pid_t pid, char **session) {
1364 _cleanup_free_ char *cgroup = NULL;
1365 int r;
1366
1367 r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1368 if (r < 0)
1369 return r;
1370
1371 return cg_path_get_session(cgroup, session);
1372 }
1373
1374 int cg_path_get_owner_uid(const char *path, uid_t *uid) {
1375 _cleanup_free_ char *slice = NULL;
1376 char *start, *end;
1377 int r;
1378
1379 assert(path);
1380
1381 r = cg_path_get_slice(path, &slice);
1382 if (r < 0)
1383 return r;
1384
1385 start = startswith(slice, "user-");
1386 if (!start)
1387 return -ENXIO;
1388 end = endswith(start, ".slice");
1389 if (!end)
1390 return -ENXIO;
1391
1392 *end = 0;
1393 if (parse_uid(start, uid) < 0)
1394 return -ENXIO;
1395
1396 return 0;
1397 }
1398
1399 int cg_pid_get_owner_uid(pid_t pid, uid_t *uid) {
1400 _cleanup_free_ char *cgroup = NULL;
1401 int r;
1402
1403 r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1404 if (r < 0)
1405 return r;
1406
1407 return cg_path_get_owner_uid(cgroup, uid);
1408 }
1409
1410 int cg_path_get_slice(const char *p, char **slice) {
1411 const char *e = NULL;
1412
1413 assert(p);
1414 assert(slice);
1415
1416 /* Finds the right-most slice unit from the beginning, but
1417 * stops before we come to the first non-slice unit. */
1418
1419 for (;;) {
1420 size_t n;
1421
1422 p += strspn(p, "/");
1423
1424 n = strcspn(p, "/");
1425 if (!valid_slice_name(p, n)) {
1426
1427 if (!e) {
1428 char *s;
1429
1430 s = strdup(SPECIAL_ROOT_SLICE);
1431 if (!s)
1432 return -ENOMEM;
1433
1434 *slice = s;
1435 return 0;
1436 }
1437
1438 return cg_path_decode_unit(e, slice);
1439 }
1440
1441 e = p;
1442 p += n;
1443 }
1444 }
1445
1446 int cg_pid_get_slice(pid_t pid, char **slice) {
1447 _cleanup_free_ char *cgroup = NULL;
1448 int r;
1449
1450 assert(slice);
1451
1452 r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1453 if (r < 0)
1454 return r;
1455
1456 return cg_path_get_slice(cgroup, slice);
1457 }
1458
1459 int cg_path_get_user_slice(const char *p, char **slice) {
1460 const char *t;
1461 assert(p);
1462 assert(slice);
1463
1464 t = skip_user_prefix(p);
1465 if (!t)
1466 return -ENXIO;
1467
1468 /* And now it looks pretty much the same as for a system
1469 * slice, so let's just use the same parser from here on. */
1470 return cg_path_get_slice(t, slice);
1471 }
1472
1473 int cg_pid_get_user_slice(pid_t pid, char **slice) {
1474 _cleanup_free_ char *cgroup = NULL;
1475 int r;
1476
1477 assert(slice);
1478
1479 r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1480 if (r < 0)
1481 return r;
1482
1483 return cg_path_get_user_slice(cgroup, slice);
1484 }
1485
1486 char *cg_escape(const char *p) {
1487 bool need_prefix = false;
1488
1489 /* This implements very minimal escaping for names to be used
1490 * as file names in the cgroup tree: any name which might
1491 * conflict with a kernel name or is prefixed with '_' is
1492 * prefixed with a '_'. That way, when reading cgroup names it
1493 * is sufficient to remove a single prefixing underscore if
1494 * there is one. */
1495
1496 /* The return value of this function (unlike cg_unescape())
1497 * needs free()! */
1498
1499 if (IN_SET(p[0], 0, '_', '.') ||
1500 STR_IN_SET(p, "notify_on_release", "release_agent", "tasks") ||
1501 startswith(p, "cgroup."))
1502 need_prefix = true;
1503 else {
1504 const char *dot;
1505
1506 dot = strrchr(p, '.');
1507 if (dot) {
1508 CGroupController c;
1509 size_t l = dot - p;
1510
1511 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1512 const char *n;
1513
1514 n = cgroup_controller_to_string(c);
1515
1516 if (l != strlen(n))
1517 continue;
1518
1519 if (memcmp(p, n, l) != 0)
1520 continue;
1521
1522 need_prefix = true;
1523 break;
1524 }
1525 }
1526 }
1527
1528 if (need_prefix)
1529 return strjoin("_", p);
1530
1531 return strdup(p);
1532 }
1533
1534 char *cg_unescape(const char *p) {
1535 assert(p);
1536
1537 /* The return value of this function (unlike cg_escape())
1538 * doesn't need free()! */
1539
1540 if (p[0] == '_')
1541 return (char*) p+1;
1542
1543 return (char*) p;
1544 }
1545
1546 #define CONTROLLER_VALID \
1547 DIGITS LETTERS \
1548 "_"
1549
1550 bool cg_controller_is_valid(const char *p) {
1551 const char *t, *s;
1552
1553 if (!p)
1554 return false;
1555
1556 if (streq(p, SYSTEMD_CGROUP_CONTROLLER))
1557 return true;
1558
1559 s = startswith(p, "name=");
1560 if (s)
1561 p = s;
1562
1563 if (IN_SET(*p, 0, '_'))
1564 return false;
1565
1566 for (t = p; *t; t++)
1567 if (!strchr(CONTROLLER_VALID, *t))
1568 return false;
1569
1570 if (t - p > FILENAME_MAX)
1571 return false;
1572
1573 return true;
1574 }
1575
1576 int cg_slice_to_path(const char *unit, char **ret) {
1577 _cleanup_free_ char *p = NULL, *s = NULL, *e = NULL;
1578 const char *dash;
1579 int r;
1580
1581 assert(unit);
1582 assert(ret);
1583
1584 if (streq(unit, SPECIAL_ROOT_SLICE)) {
1585 char *x;
1586
1587 x = strdup("");
1588 if (!x)
1589 return -ENOMEM;
1590 *ret = x;
1591 return 0;
1592 }
1593
1594 if (!unit_name_is_valid(unit, UNIT_NAME_PLAIN))
1595 return -EINVAL;
1596
1597 if (!endswith(unit, ".slice"))
1598 return -EINVAL;
1599
1600 r = unit_name_to_prefix(unit, &p);
1601 if (r < 0)
1602 return r;
1603
1604 dash = strchr(p, '-');
1605
1606 /* Don't allow initial dashes */
1607 if (dash == p)
1608 return -EINVAL;
1609
1610 while (dash) {
1611 _cleanup_free_ char *escaped = NULL;
1612 char n[dash - p + sizeof(".slice")];
1613
1614 #if HAS_FEATURE_MEMORY_SANITIZER
1615 /* msan doesn't instrument stpncpy, so it thinks
1616 * n is later used uninitialized:
1617 * https://github.com/google/sanitizers/issues/926
1618 */
1619 zero(n);
1620 #endif
1621
1622 /* Don't allow trailing or double dashes */
1623 if (IN_SET(dash[1], 0, '-'))
1624 return -EINVAL;
1625
1626 strcpy(stpncpy(n, p, dash - p), ".slice");
1627 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN))
1628 return -EINVAL;
1629
1630 escaped = cg_escape(n);
1631 if (!escaped)
1632 return -ENOMEM;
1633
1634 if (!strextend(&s, escaped, "/"))
1635 return -ENOMEM;
1636
1637 dash = strchr(dash+1, '-');
1638 }
1639
1640 e = cg_escape(unit);
1641 if (!e)
1642 return -ENOMEM;
1643
1644 if (!strextend(&s, e))
1645 return -ENOMEM;
1646
1647 *ret = TAKE_PTR(s);
1648
1649 return 0;
1650 }
1651
1652 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value) {
1653 _cleanup_free_ char *p = NULL;
1654 int r;
1655
1656 r = cg_get_path(controller, path, attribute, &p);
1657 if (r < 0)
1658 return r;
1659
1660 return write_string_file(p, value, WRITE_STRING_FILE_DISABLE_BUFFER);
1661 }
1662
1663 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret) {
1664 _cleanup_free_ char *p = NULL;
1665 int r;
1666
1667 r = cg_get_path(controller, path, attribute, &p);
1668 if (r < 0)
1669 return r;
1670
1671 return read_one_line_file(p, ret);
1672 }
1673
1674 int cg_get_attribute_as_uint64(const char *controller, const char *path, const char *attribute, uint64_t *ret) {
1675 _cleanup_free_ char *value = NULL;
1676 uint64_t v;
1677 int r;
1678
1679 assert(ret);
1680
1681 r = cg_get_attribute(controller, path, attribute, &value);
1682 if (r == -ENOENT)
1683 return -ENODATA;
1684 if (r < 0)
1685 return r;
1686
1687 if (streq(value, "max")) {
1688 *ret = CGROUP_LIMIT_MAX;
1689 return 0;
1690 }
1691
1692 r = safe_atou64(value, &v);
1693 if (r < 0)
1694 return r;
1695
1696 *ret = v;
1697 return 0;
1698 }
1699
1700 int cg_get_attribute_as_bool(const char *controller, const char *path, const char *attribute, bool *ret) {
1701 _cleanup_free_ char *value = NULL;
1702 int r;
1703
1704 assert(ret);
1705
1706 r = cg_get_attribute(controller, path, attribute, &value);
1707 if (r == -ENOENT)
1708 return -ENODATA;
1709 if (r < 0)
1710 return r;
1711
1712 r = parse_boolean(value);
1713 if (r < 0)
1714 return r;
1715
1716 *ret = r;
1717 return 0;
1718 }
1719
1720 int cg_get_owner(const char *controller, const char *path, uid_t *ret_uid) {
1721 _cleanup_free_ char *f = NULL;
1722 struct stat stats;
1723 int r;
1724
1725 assert(ret_uid);
1726
1727 r = cg_get_path(controller, path, NULL, &f);
1728 if (r < 0)
1729 return r;
1730
1731 r = stat(f, &stats);
1732 if (r < 0)
1733 return -errno;
1734
1735 *ret_uid = stats.st_uid;
1736 return 0;
1737 }
1738
1739 int cg_get_keyed_attribute_full(
1740 const char *controller,
1741 const char *path,
1742 const char *attribute,
1743 char **keys,
1744 char **ret_values,
1745 CGroupKeyMode mode) {
1746
1747 _cleanup_free_ char *filename = NULL, *contents = NULL;
1748 const char *p;
1749 size_t n, i, n_done = 0;
1750 char **v;
1751 int r;
1752
1753 /* Reads one or more fields of a cgroup v2 keyed attribute file. The 'keys' parameter should be an strv with
1754 * all keys to retrieve. The 'ret_values' parameter should be passed as string size with the same number of
1755 * entries as 'keys'. On success each entry will be set to the value of the matching key.
1756 *
1757 * If the attribute file doesn't exist at all returns ENOENT, if any key is not found returns ENXIO. If mode
1758 * is set to GG_KEY_MODE_GRACEFUL we ignore missing keys and return those that were parsed successfully. */
1759
1760 r = cg_get_path(controller, path, attribute, &filename);
1761 if (r < 0)
1762 return r;
1763
1764 r = read_full_file(filename, &contents, NULL);
1765 if (r < 0)
1766 return r;
1767
1768 n = strv_length(keys);
1769 if (n == 0) /* No keys to retrieve? That's easy, we are done then */
1770 return 0;
1771
1772 /* Let's build this up in a temporary array for now in order not to clobber the return parameter on failure */
1773 v = newa0(char*, n);
1774
1775 for (p = contents; *p;) {
1776 const char *w = NULL;
1777
1778 for (i = 0; i < n; i++)
1779 if (!v[i]) {
1780 w = first_word(p, keys[i]);
1781 if (w)
1782 break;
1783 }
1784
1785 if (w) {
1786 size_t l;
1787
1788 l = strcspn(w, NEWLINE);
1789 v[i] = strndup(w, l);
1790 if (!v[i]) {
1791 r = -ENOMEM;
1792 goto fail;
1793 }
1794
1795 n_done++;
1796 if (n_done >= n)
1797 goto done;
1798
1799 p = w + l;
1800 } else
1801 p += strcspn(p, NEWLINE);
1802
1803 p += strspn(p, NEWLINE);
1804 }
1805
1806 if (mode & CG_KEY_MODE_GRACEFUL)
1807 goto done;
1808
1809 r = -ENXIO;
1810
1811 fail:
1812 for (i = 0; i < n; i++)
1813 free(v[i]);
1814
1815 return r;
1816
1817 done:
1818 memcpy(ret_values, v, sizeof(char*) * n);
1819 if (mode & CG_KEY_MODE_GRACEFUL)
1820 return n_done;
1821
1822 return 0;
1823 }
1824
1825 int cg_mask_to_string(CGroupMask mask, char **ret) {
1826 _cleanup_free_ char *s = NULL;
1827 size_t n = 0, allocated = 0;
1828 bool space = false;
1829 CGroupController c;
1830
1831 assert(ret);
1832
1833 if (mask == 0) {
1834 *ret = NULL;
1835 return 0;
1836 }
1837
1838 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1839 const char *k;
1840 size_t l;
1841
1842 if (!FLAGS_SET(mask, CGROUP_CONTROLLER_TO_MASK(c)))
1843 continue;
1844
1845 k = cgroup_controller_to_string(c);
1846 l = strlen(k);
1847
1848 if (!GREEDY_REALLOC(s, allocated, n + space + l + 1))
1849 return -ENOMEM;
1850
1851 if (space)
1852 s[n] = ' ';
1853 memcpy(s + n + space, k, l);
1854 n += space + l;
1855
1856 space = true;
1857 }
1858
1859 assert(s);
1860
1861 s[n] = 0;
1862 *ret = TAKE_PTR(s);
1863
1864 return 0;
1865 }
1866
1867 int cg_mask_from_string(const char *value, CGroupMask *ret) {
1868 CGroupMask m = 0;
1869
1870 assert(ret);
1871 assert(value);
1872
1873 for (;;) {
1874 _cleanup_free_ char *n = NULL;
1875 CGroupController v;
1876 int r;
1877
1878 r = extract_first_word(&value, &n, NULL, 0);
1879 if (r < 0)
1880 return r;
1881 if (r == 0)
1882 break;
1883
1884 v = cgroup_controller_from_string(n);
1885 if (v < 0)
1886 continue;
1887
1888 m |= CGROUP_CONTROLLER_TO_MASK(v);
1889 }
1890
1891 *ret = m;
1892 return 0;
1893 }
1894
1895 int cg_mask_supported(CGroupMask *ret) {
1896 CGroupMask mask;
1897 int r;
1898
1899 /* Determines the mask of supported cgroup controllers. Only includes controllers we can make sense of and that
1900 * are actually accessible. Only covers real controllers, i.e. not the CGROUP_CONTROLLER_BPF_xyz
1901 * pseudo-controllers. */
1902
1903 r = cg_all_unified();
1904 if (r < 0)
1905 return r;
1906 if (r > 0) {
1907 _cleanup_free_ char *root = NULL, *controllers = NULL, *path = NULL;
1908
1909 /* In the unified hierarchy we can read the supported and accessible controllers from
1910 * the top-level cgroup attribute */
1911
1912 r = cg_get_root_path(&root);
1913 if (r < 0)
1914 return r;
1915
1916 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, root, "cgroup.controllers", &path);
1917 if (r < 0)
1918 return r;
1919
1920 r = read_one_line_file(path, &controllers);
1921 if (r < 0)
1922 return r;
1923
1924 r = cg_mask_from_string(controllers, &mask);
1925 if (r < 0)
1926 return r;
1927
1928 /* Mask controllers that are not supported in unified hierarchy. */
1929 mask &= CGROUP_MASK_V2;
1930
1931 } else {
1932 CGroupController c;
1933
1934 /* In the legacy hierarchy, we check which hierarchies are mounted. */
1935
1936 mask = 0;
1937 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1938 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
1939 const char *n;
1940
1941 if (!FLAGS_SET(CGROUP_MASK_V1, bit))
1942 continue;
1943
1944 n = cgroup_controller_to_string(c);
1945 if (controller_is_accessible(n) >= 0)
1946 mask |= bit;
1947 }
1948 }
1949
1950 *ret = mask;
1951 return 0;
1952 }
1953
1954 int cg_kernel_controllers(Set **ret) {
1955 _cleanup_set_free_free_ Set *controllers = NULL;
1956 _cleanup_fclose_ FILE *f = NULL;
1957 int r;
1958
1959 assert(ret);
1960
1961 /* Determines the full list of kernel-known controllers. Might include controllers we don't actually support
1962 * and controllers that aren't currently accessible (because not mounted). This does not include "name="
1963 * pseudo-controllers. */
1964
1965 controllers = set_new(&string_hash_ops);
1966 if (!controllers)
1967 return -ENOMEM;
1968
1969 r = fopen_unlocked("/proc/cgroups", "re", &f);
1970 if (r == -ENOENT) {
1971 *ret = NULL;
1972 return 0;
1973 }
1974 if (r < 0)
1975 return r;
1976
1977 /* Ignore the header line */
1978 (void) read_line(f, SIZE_MAX, NULL);
1979
1980 for (;;) {
1981 char *controller;
1982 int enabled = 0;
1983
1984 errno = 0;
1985 if (fscanf(f, "%ms %*i %*i %i", &controller, &enabled) != 2) {
1986
1987 if (feof(f))
1988 break;
1989
1990 if (ferror(f))
1991 return errno_or_else(EIO);
1992
1993 return -EBADMSG;
1994 }
1995
1996 if (!enabled) {
1997 free(controller);
1998 continue;
1999 }
2000
2001 if (!cg_controller_is_valid(controller)) {
2002 free(controller);
2003 return -EBADMSG;
2004 }
2005
2006 r = set_consume(controllers, controller);
2007 if (r < 0)
2008 return r;
2009 }
2010
2011 *ret = TAKE_PTR(controllers);
2012
2013 return 0;
2014 }
2015
2016 /* The hybrid mode was initially implemented in v232 and simply mounted cgroup2 on
2017 * /sys/fs/cgroup/systemd. This unfortunately broke other tools (such as docker) which expected the v1
2018 * "name=systemd" hierarchy on /sys/fs/cgroup/systemd. From v233 and on, the hybrid mode mounts v2 on
2019 * /sys/fs/cgroup/unified and maintains "name=systemd" hierarchy on /sys/fs/cgroup/systemd for compatibility
2020 * with other tools.
2021 *
2022 * To keep live upgrade working, we detect and support v232 layout. When v232 layout is detected, to keep
2023 * cgroup v2 process management but disable the compat dual layout, we return true on
2024 * cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) and false on cg_hybrid_unified().
2025 */
2026 static thread_local bool unified_systemd_v232;
2027
2028 int cg_unified_cached(bool flush) {
2029 static thread_local CGroupUnified unified_cache = CGROUP_UNIFIED_UNKNOWN;
2030
2031 struct statfs fs;
2032
2033 /* Checks if we support the unified hierarchy. Returns an
2034 * error when the cgroup hierarchies aren't mounted yet or we
2035 * have any other trouble determining if the unified hierarchy
2036 * is supported. */
2037
2038 if (flush)
2039 unified_cache = CGROUP_UNIFIED_UNKNOWN;
2040 else if (unified_cache >= CGROUP_UNIFIED_NONE)
2041 return unified_cache;
2042
2043 if (statfs("/sys/fs/cgroup/", &fs) < 0)
2044 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/\") failed: %m");
2045
2046 if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2047 log_debug("Found cgroup2 on /sys/fs/cgroup/, full unified hierarchy");
2048 unified_cache = CGROUP_UNIFIED_ALL;
2049 } else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC)) {
2050 if (statfs("/sys/fs/cgroup/unified/", &fs) == 0 &&
2051 F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2052 log_debug("Found cgroup2 on /sys/fs/cgroup/unified, unified hierarchy for systemd controller");
2053 unified_cache = CGROUP_UNIFIED_SYSTEMD;
2054 unified_systemd_v232 = false;
2055 } else {
2056 if (statfs("/sys/fs/cgroup/systemd/", &fs) < 0)
2057 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/systemd\" failed: %m");
2058
2059 if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2060 log_debug("Found cgroup2 on /sys/fs/cgroup/systemd, unified hierarchy for systemd controller (v232 variant)");
2061 unified_cache = CGROUP_UNIFIED_SYSTEMD;
2062 unified_systemd_v232 = true;
2063 } else if (F_TYPE_EQUAL(fs.f_type, CGROUP_SUPER_MAGIC)) {
2064 log_debug("Found cgroup on /sys/fs/cgroup/systemd, legacy hierarchy");
2065 unified_cache = CGROUP_UNIFIED_NONE;
2066 } else {
2067 log_debug("Unexpected filesystem type %llx mounted on /sys/fs/cgroup/systemd, assuming legacy hierarchy",
2068 (unsigned long long) fs.f_type);
2069 unified_cache = CGROUP_UNIFIED_NONE;
2070 }
2071 }
2072 } else if (F_TYPE_EQUAL(fs.f_type, SYSFS_MAGIC)) {
2073 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM),
2074 "No filesystem is currently mounted on /sys/fs/cgroup.");
2075 } else
2076 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM),
2077 "Unknown filesystem type %llx mounted on /sys/fs/cgroup.",
2078 (unsigned long long)fs.f_type);
2079
2080 return unified_cache;
2081 }
2082
2083 int cg_unified_controller(const char *controller) {
2084 int r;
2085
2086 r = cg_unified_cached(false);
2087 if (r < 0)
2088 return r;
2089
2090 if (r == CGROUP_UNIFIED_NONE)
2091 return false;
2092
2093 if (r >= CGROUP_UNIFIED_ALL)
2094 return true;
2095
2096 return streq_ptr(controller, SYSTEMD_CGROUP_CONTROLLER);
2097 }
2098
2099 int cg_all_unified(void) {
2100 int r;
2101
2102 r = cg_unified_cached(false);
2103 if (r < 0)
2104 return r;
2105
2106 return r >= CGROUP_UNIFIED_ALL;
2107 }
2108
2109 int cg_hybrid_unified(void) {
2110 int r;
2111
2112 r = cg_unified_cached(false);
2113 if (r < 0)
2114 return r;
2115
2116 return r == CGROUP_UNIFIED_SYSTEMD && !unified_systemd_v232;
2117 }
2118
2119 const uint64_t cgroup_io_limit_defaults[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2120 [CGROUP_IO_RBPS_MAX] = CGROUP_LIMIT_MAX,
2121 [CGROUP_IO_WBPS_MAX] = CGROUP_LIMIT_MAX,
2122 [CGROUP_IO_RIOPS_MAX] = CGROUP_LIMIT_MAX,
2123 [CGROUP_IO_WIOPS_MAX] = CGROUP_LIMIT_MAX,
2124 };
2125
2126 static const char* const cgroup_io_limit_type_table[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2127 [CGROUP_IO_RBPS_MAX] = "IOReadBandwidthMax",
2128 [CGROUP_IO_WBPS_MAX] = "IOWriteBandwidthMax",
2129 [CGROUP_IO_RIOPS_MAX] = "IOReadIOPSMax",
2130 [CGROUP_IO_WIOPS_MAX] = "IOWriteIOPSMax",
2131 };
2132
2133 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type, CGroupIOLimitType);
2134
2135 bool is_cgroup_fs(const struct statfs *s) {
2136 return is_fs_type(s, CGROUP_SUPER_MAGIC) ||
2137 is_fs_type(s, CGROUP2_SUPER_MAGIC);
2138 }
2139
2140 bool fd_is_cgroup_fs(int fd) {
2141 struct statfs s;
2142
2143 if (fstatfs(fd, &s) < 0)
2144 return -errno;
2145
2146 return is_cgroup_fs(&s);
2147 }
2148
2149 static const char *const cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
2150 [CGROUP_CONTROLLER_CPU] = "cpu",
2151 [CGROUP_CONTROLLER_CPUACCT] = "cpuacct",
2152 [CGROUP_CONTROLLER_CPUSET] = "cpuset",
2153 [CGROUP_CONTROLLER_IO] = "io",
2154 [CGROUP_CONTROLLER_BLKIO] = "blkio",
2155 [CGROUP_CONTROLLER_MEMORY] = "memory",
2156 [CGROUP_CONTROLLER_DEVICES] = "devices",
2157 [CGROUP_CONTROLLER_PIDS] = "pids",
2158 [CGROUP_CONTROLLER_BPF_FIREWALL] = "bpf-firewall",
2159 [CGROUP_CONTROLLER_BPF_DEVICES] = "bpf-devices",
2160 };
2161
2162 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller, CGroupController);
2163
2164 CGroupMask get_cpu_accounting_mask(void) {
2165 static CGroupMask needed_mask = (CGroupMask) -1;
2166
2167 /* On kernel ≥4.15 with unified hierarchy, cpu.stat's usage_usec is
2168 * provided externally from the CPU controller, which means we don't
2169 * need to enable the CPU controller just to get metrics. This is good,
2170 * because enabling the CPU controller comes at a minor performance
2171 * hit, especially when it's propagated deep into large hierarchies.
2172 * There's also no separate CPU accounting controller available within
2173 * a unified hierarchy.
2174 *
2175 * This combination of factors results in the desired cgroup mask to
2176 * enable for CPU accounting varying as follows:
2177 *
2178 * ╔═════════════════════╤═════════════════════╗
2179 * ║ Linux ≥4.15 │ Linux <4.15 ║
2180 * ╔═══════════════╬═════════════════════╪═════════════════════╣
2181 * ║ Unified ║ nothing │ CGROUP_MASK_CPU ║
2182 * ╟───────────────╫─────────────────────┼─────────────────────╢
2183 * ║ Hybrid/Legacy ║ CGROUP_MASK_CPUACCT │ CGROUP_MASK_CPUACCT ║
2184 * ╚═══════════════╩═════════════════════╧═════════════════════╝
2185 *
2186 * We check kernel version here instead of manually checking whether
2187 * cpu.stat is present for every cgroup, as that check in itself would
2188 * already be fairly expensive.
2189 *
2190 * Kernels where this patch has been backported will therefore have the
2191 * CPU controller enabled unnecessarily. This is more expensive than
2192 * necessary, but harmless. ☺️
2193 */
2194
2195 if (needed_mask == (CGroupMask) -1) {
2196 if (cg_all_unified()) {
2197 struct utsname u;
2198 assert_se(uname(&u) >= 0);
2199
2200 if (strverscmp_improved(u.release, "4.15") < 0)
2201 needed_mask = CGROUP_MASK_CPU;
2202 else
2203 needed_mask = 0;
2204 } else
2205 needed_mask = CGROUP_MASK_CPUACCT;
2206 }
2207
2208 return needed_mask;
2209 }
2210
2211 bool cpu_accounting_is_cheap(void) {
2212 return get_cpu_accounting_mask() == 0;
2213 }
2214
2215 static const char* const managed_oom_mode_table[_MANAGED_OOM_MODE_MAX] = {
2216 [MANAGED_OOM_AUTO] = "auto",
2217 [MANAGED_OOM_KILL] = "kill",
2218 };
2219
2220 DEFINE_STRING_TABLE_LOOKUP(managed_oom_mode, ManagedOOMMode);
2221
2222 static const char* const managed_oom_preference_table[_MANAGED_OOM_PREFERENCE_MAX] = {
2223 [MANAGED_OOM_PREFERENCE_NONE] = "none",
2224 [MANAGED_OOM_PREFERENCE_AVOID] = "avoid",
2225 [MANAGED_OOM_PREFERENCE_OMIT] = "omit",
2226 };
2227
2228 DEFINE_STRING_TABLE_LOOKUP(managed_oom_preference, ManagedOOMPreference);
Unnamed repository; edit this file 'description' to name the repository.