1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
10 #include <sys/utsname.h>
11 #include <sys/xattr.h>
14 #include "alloc-util.h"
15 #include "cgroup-util.h"
17 #include "dirent-util.h"
18 #include "extract-word.h"
21 #include "format-util.h"
24 #include "login-util.h"
26 #include "missing_magic.h"
28 #include "parse-util.h"
29 #include "path-util.h"
30 #include "process-util.h"
33 #include "stat-util.h"
34 #include "stdio-util.h"
35 #include "string-table.h"
36 #include "string-util.h"
38 #include "unit-name.h"
39 #include "user-util.h"
40 #include "xattr-util.h"
42 static int cg_enumerate_items(const char *controller
, const char *path
, FILE **_f
, const char *item
) {
43 _cleanup_free_
char *fs
= NULL
;
49 r
= cg_get_path(controller
, path
, item
, &fs
);
61 int cg_enumerate_processes(const char *controller
, const char *path
, FILE **_f
) {
62 return cg_enumerate_items(controller
, path
, _f
, "cgroup.procs");
65 int cg_read_pid(FILE *f
, pid_t
*_pid
) {
68 /* Note that the cgroup.procs might contain duplicates! See
69 * cgroups.txt for details. */
75 if (fscanf(f
, "%lu", &ul
) != 1) {
80 return errno_or_else(EIO
);
91 const char *controller
,
96 _cleanup_free_
char *events
= NULL
, *content
= NULL
;
99 r
= cg_get_path(controller
, path
, "cgroup.events", &events
);
103 r
= read_full_file(events
, &content
, NULL
);
107 for (const char *p
= content
;;) {
108 _cleanup_free_
char *line
= NULL
, *key
= NULL
, *val
= NULL
;
111 r
= extract_first_word(&p
, &line
, "\n", 0);
118 r
= extract_first_word(&q
, &key
, " ", 0);
124 if (!streq(key
, event
))
131 *ret
= TAKE_PTR(val
);
136 bool cg_ns_supported(void) {
137 static thread_local
int enabled
= -1;
142 if (access("/proc/self/ns/cgroup", F_OK
) < 0) {
144 log_debug_errno(errno
, "Failed to check whether /proc/self/ns/cgroup is available, assuming not: %m");
152 bool cg_freezer_supported(void) {
153 static thread_local
int supported
= -1;
158 supported
= cg_all_unified() > 0 && access("/sys/fs/cgroup/init.scope/cgroup.freeze", F_OK
) == 0;
163 int cg_enumerate_subgroups(const char *controller
, const char *path
, DIR **_d
) {
164 _cleanup_free_
char *fs
= NULL
;
170 /* This is not recursive! */
172 r
= cg_get_path(controller
, path
, NULL
, &fs
);
184 int cg_read_subgroup(DIR *d
, char **fn
) {
190 FOREACH_DIRENT_ALL(de
, d
, return -errno
) {
193 if (de
->d_type
!= DT_DIR
)
196 if (dot_or_dot_dot(de
->d_name
))
199 b
= strdup(de
->d_name
);
210 int cg_rmdir(const char *controller
, const char *path
) {
211 _cleanup_free_
char *p
= NULL
;
214 r
= cg_get_path(controller
, path
, NULL
, &p
);
219 if (r
< 0 && errno
!= ENOENT
)
222 r
= cg_hybrid_unified();
226 if (streq(controller
, SYSTEMD_CGROUP_CONTROLLER
)) {
227 r
= cg_rmdir(SYSTEMD_CGROUP_CONTROLLER_LEGACY
, path
);
229 log_warning_errno(r
, "Failed to remove compat systemd cgroup %s: %m", path
);
235 static int cg_kill_items(
236 const char *controller
,
241 cg_kill_log_func_t log_kill
,
245 _cleanup_set_free_ Set
*allocated_set
= NULL
;
247 int r
, ret
= 0, ret_log_kill
= 0;
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
;
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 */
262 s
= allocated_set
= set_new(NULL
);
267 my_pid
= getpid_cached();
270 _cleanup_fclose_
FILE *f
= NULL
;
274 r
= cg_enumerate_items(controller
, path
, &f
, item
);
276 if (ret
>= 0 && r
!= -ENOENT
)
282 while ((r
= cg_read_pid(f
, &pid
)) > 0) {
284 if ((flags
& CGROUP_IGNORE_SELF
) && pid
== my_pid
)
287 if (set_get(s
, PID_TO_PTR(pid
)) == PID_TO_PTR(pid
))
291 ret_log_kill
= log_kill(pid
, sig
, userdata
);
293 /* If we haven't killed this process yet, kill
295 if (kill(pid
, sig
) < 0) {
296 if (ret
>= 0 && errno
!= ESRCH
)
299 if (flags
& CGROUP_SIGCONT
)
300 (void) kill(pid
, SIGCONT
);
312 r
= set_put(s
, PID_TO_PTR(pid
));
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. */
338 const char *controller
,
343 cg_kill_log_func_t log_kill
,
347 r
= cg_kill_items(controller
, path
, sig
, flags
, s
, log_kill
, userdata
, "cgroup.procs");
348 if (r
< 0 || sig
!= SIGKILL
)
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
);
358 return cg_kill_items(controller
, path
, sig
, flags
, s
, log_kill
, userdata
, "cgroup.threads");
361 int cg_kill_recursive(
362 const char *controller
,
367 cg_kill_log_func_t log_kill
,
370 _cleanup_set_free_ Set
*allocated_set
= NULL
;
371 _cleanup_closedir_
DIR *d
= NULL
;
379 s
= allocated_set
= set_new(NULL
);
384 ret
= cg_kill(controller
, path
, sig
, flags
, s
, log_kill
, userdata
);
386 r
= cg_enumerate_subgroups(controller
, path
, &d
);
388 if (ret
>= 0 && r
!= -ENOENT
)
394 while ((r
= cg_read_subgroup(d
, &fn
)) > 0) {
395 _cleanup_free_
char *p
= NULL
;
397 p
= path_join(empty_to_root(path
), fn
);
402 r
= cg_kill_recursive(controller
, p
, sig
, flags
, s
, log_kill
, userdata
);
403 if (r
!= 0 && ret
>= 0)
406 if (ret
>= 0 && r
< 0)
409 if (flags
& CGROUP_REMOVE
) {
410 r
= cg_rmdir(controller
, path
);
411 if (r
< 0 && ret
>= 0 && !IN_SET(r
, -ENOENT
, -EBUSY
))
418 static const char *controller_to_dirname(const char *controller
) {
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. */
428 if (streq(controller
, SYSTEMD_CGROUP_CONTROLLER
)) {
429 if (cg_hybrid_unified() > 0)
430 controller
= SYSTEMD_CGROUP_CONTROLLER_HYBRID
;
432 controller
= SYSTEMD_CGROUP_CONTROLLER_LEGACY
;
435 e
= startswith(controller
, "name=");
442 static int join_path_legacy(const char *controller
, const char *path
, const char *suffix
, char **fs
) {
449 dn
= controller_to_dirname(controller
);
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
);
458 t
= path_join("/sys/fs/cgroup", dn
, path
, suffix
);
466 static int join_path_unified(const char *path
, const char *suffix
, char **fs
) {
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
);
478 t
= path_join("/sys/fs/cgroup", path
, suffix
);
486 int cg_get_path(const char *controller
, const char *path
, const char *suffix
, char **fs
) {
494 /* If no controller is specified, we return the path
495 * *below* the controllers, without any prefix. */
497 if (!path
&& !suffix
)
505 t
= path_join(path
, suffix
);
509 *fs
= path_simplify(t
, false);
513 if (!cg_controller_is_valid(controller
))
516 r
= cg_all_unified();
520 r
= join_path_unified(path
, suffix
, fs
);
522 r
= join_path_legacy(controller
, path
, suffix
, fs
);
526 path_simplify(*fs
, false);
530 static int controller_is_v1_accessible(const char *root
, const char *controller
) {
531 const char *cpath
, *dn
;
535 dn
= controller_to_dirname(controller
);
537 /* If root if specified, we check that:
538 * - possible subcgroup is created at root,
539 * - we can modify the hierarchy. */
541 cpath
= strjoina("/sys/fs/cgroup/", dn
, root
, root
? "/cgroup.procs" : NULL
);
542 if (laccess(cpath
, root
? W_OK
: F_OK
) < 0)
548 int cg_get_path_and_check(const char *controller
, const char *path
, const char *suffix
, char **fs
) {
554 if (!cg_controller_is_valid(controller
))
557 r
= cg_all_unified();
561 /* In the unified hierarchy all controllers are considered accessible,
562 * except for the named hierarchies */
563 if (startswith(controller
, "name="))
566 /* Check if the specified controller is actually accessible */
567 r
= controller_is_v1_accessible(NULL
, controller
);
572 return cg_get_path(controller
, path
, suffix
, fs
);
575 int cg_set_xattr(const char *controller
, const char *path
, const char *name
, const void *value
, size_t size
, int flags
) {
576 _cleanup_free_
char *fs
= NULL
;
581 assert(value
|| size
<= 0);
583 r
= cg_get_path(controller
, path
, NULL
, &fs
);
587 if (setxattr(fs
, name
, value
, size
, flags
) < 0)
593 int cg_get_xattr(const char *controller
, const char *path
, const char *name
, void *value
, size_t size
) {
594 _cleanup_free_
char *fs
= NULL
;
601 r
= cg_get_path(controller
, path
, NULL
, &fs
);
605 n
= getxattr(fs
, name
, value
, size
);
612 int cg_get_xattr_malloc(const char *controller
, const char *path
, const char *name
, char **ret
) {
613 _cleanup_free_
char *fs
= NULL
;
619 r
= cg_get_path(controller
, path
, NULL
, &fs
);
623 r
= getxattr_malloc(fs
, name
, ret
, false);
630 int cg_get_xattr_bool(const char *controller
, const char *path
, const char *name
) {
631 _cleanup_free_
char *val
= NULL
;
637 r
= cg_get_xattr_malloc(controller
, path
, name
, &val
);
641 return parse_boolean(val
);
644 int cg_remove_xattr(const char *controller
, const char *path
, const char *name
) {
645 _cleanup_free_
char *fs
= NULL
;
651 r
= cg_get_path(controller
, path
, NULL
, &fs
);
655 if (removexattr(fs
, name
) < 0)
661 int cg_pid_get_path(const char *controller
, pid_t pid
, char **ret_path
) {
662 _cleanup_fclose_
FILE *f
= NULL
;
663 const char *fs
, *controller_str
;
670 if (!cg_controller_is_valid(controller
))
673 controller
= SYSTEMD_CGROUP_CONTROLLER
;
675 unified
= cg_unified_controller(controller
);
679 if (streq(controller
, SYSTEMD_CGROUP_CONTROLLER
))
680 controller_str
= SYSTEMD_CGROUP_CONTROLLER_LEGACY
;
682 controller_str
= controller
;
685 fs
= procfs_file_alloca(pid
, "cgroup");
686 r
= fopen_unlocked(fs
, "re", &f
);
693 _cleanup_free_
char *line
= NULL
;
696 r
= read_line(f
, LONG_LINE_MAX
, &line
);
703 e
= startswith(line
, "0:");
713 l
= strchr(line
, ':');
723 r
= string_contains_word(l
, ",", controller_str
);
730 char *path
= strdup(e
+ 1);
734 /* Truncate suffix indicating the process is a zombie */
735 e
= endswith(path
, " (deleted)");
744 int cg_install_release_agent(const char *controller
, const char *agent
) {
745 _cleanup_free_
char *fs
= NULL
, *contents
= NULL
;
751 r
= cg_unified_controller(controller
);
754 if (r
> 0) /* doesn't apply to unified hierarchy */
757 r
= cg_get_path(controller
, NULL
, "release_agent", &fs
);
761 r
= read_one_line_file(fs
, &contents
);
765 sc
= strstrip(contents
);
767 r
= write_string_file(fs
, agent
, WRITE_STRING_FILE_DISABLE_BUFFER
);
770 } else if (!path_equal(sc
, agent
))
774 r
= cg_get_path(controller
, NULL
, "notify_on_release", &fs
);
778 contents
= mfree(contents
);
779 r
= read_one_line_file(fs
, &contents
);
783 sc
= strstrip(contents
);
784 if (streq(sc
, "0")) {
785 r
= write_string_file(fs
, "1", WRITE_STRING_FILE_DISABLE_BUFFER
);
798 int cg_uninstall_release_agent(const char *controller
) {
799 _cleanup_free_
char *fs
= NULL
;
802 r
= cg_unified_controller(controller
);
805 if (r
> 0) /* Doesn't apply to unified hierarchy */
808 r
= cg_get_path(controller
, NULL
, "notify_on_release", &fs
);
812 r
= write_string_file(fs
, "0", WRITE_STRING_FILE_DISABLE_BUFFER
);
818 r
= cg_get_path(controller
, NULL
, "release_agent", &fs
);
822 r
= write_string_file(fs
, "", WRITE_STRING_FILE_DISABLE_BUFFER
);
829 int cg_is_empty(const char *controller
, const char *path
) {
830 _cleanup_fclose_
FILE *f
= NULL
;
836 r
= cg_enumerate_processes(controller
, path
, &f
);
842 r
= cg_read_pid(f
, &pid
);
849 int cg_is_empty_recursive(const char *controller
, const char *path
) {
854 /* The root cgroup is always populated */
855 if (controller
&& empty_or_root(path
))
858 r
= cg_unified_controller(controller
);
862 _cleanup_free_
char *t
= NULL
;
864 /* On the unified hierarchy we can check empty state
865 * via the "populated" attribute of "cgroup.events". */
867 r
= cg_read_event(controller
, path
, "populated", &t
);
873 return streq(t
, "0");
875 _cleanup_closedir_
DIR *d
= NULL
;
878 r
= cg_is_empty(controller
, path
);
882 r
= cg_enumerate_subgroups(controller
, path
, &d
);
888 while ((r
= cg_read_subgroup(d
, &fn
)) > 0) {
889 _cleanup_free_
char *p
= NULL
;
891 p
= path_join(path
, fn
);
896 r
= cg_is_empty_recursive(controller
, p
);
907 int cg_split_spec(const char *spec
, char **ret_controller
, char **ret_path
) {
908 _cleanup_free_
char *controller
= NULL
, *path
= NULL
;
913 if (!path_is_normalized(spec
))
921 path_simplify(path
, false);
927 e
= strchr(spec
, ':');
929 controller
= strndup(spec
, e
-spec
);
932 if (!cg_controller_is_valid(controller
))
935 if (!isempty(e
+ 1)) {
940 if (!path_is_normalized(path
) ||
941 !path_is_absolute(path
))
944 path_simplify(path
, false);
948 if (!cg_controller_is_valid(spec
))
951 if (ret_controller
) {
952 controller
= strdup(spec
);
960 *ret_controller
= TAKE_PTR(controller
);
962 *ret_path
= TAKE_PTR(path
);
966 int cg_mangle_path(const char *path
, char **result
) {
967 _cleanup_free_
char *c
= NULL
, *p
= NULL
;
974 /* First, check if it already is a filesystem path */
975 if (path_startswith(path
, "/sys/fs/cgroup")) {
981 *result
= path_simplify(t
, false);
985 /* Otherwise, treat it as cg spec */
986 r
= cg_split_spec(path
, &c
, &p
);
990 return cg_get_path(c
?: SYSTEMD_CGROUP_CONTROLLER
, p
?: "/", NULL
, result
);
993 int cg_get_root_path(char **path
) {
999 r
= cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER
, 1, &p
);
1003 e
= endswith(p
, "/" SPECIAL_INIT_SCOPE
);
1005 e
= endswith(p
, "/" SPECIAL_SYSTEM_SLICE
); /* legacy */
1007 e
= endswith(p
, "/system"); /* even more legacy */
1015 int cg_shift_path(const char *cgroup
, const char *root
, const char **shifted
) {
1016 _cleanup_free_
char *rt
= NULL
;
1024 /* If the root was specified let's use that, otherwise
1025 * let's determine it from PID 1 */
1027 r
= cg_get_root_path(&rt
);
1034 p
= path_startswith(cgroup
, root
);
1035 if (p
&& p
> cgroup
)
1043 int cg_pid_get_path_shifted(pid_t pid
, const char *root
, char **cgroup
) {
1044 _cleanup_free_
char *raw
= NULL
;
1051 r
= cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER
, pid
, &raw
);
1055 r
= cg_shift_path(raw
, root
, &c
);
1060 *cgroup
= TAKE_PTR(raw
);
1074 int cg_path_decode_unit(const char *cgroup
, char **unit
) {
1081 n
= strcspn(cgroup
, "/");
1085 c
= strndupa(cgroup
, n
);
1088 if (!unit_name_is_valid(c
, UNIT_NAME_PLAIN
|UNIT_NAME_INSTANCE
))
1099 static bool valid_slice_name(const char *p
, size_t n
) {
1104 if (n
< STRLEN("x.slice"))
1107 if (memcmp(p
+ n
- 6, ".slice", 6) == 0) {
1113 c
= cg_unescape(buf
);
1115 return unit_name_is_valid(c
, UNIT_NAME_PLAIN
);
1121 static const char *skip_slices(const char *p
) {
1124 /* Skips over all slice assignments */
1129 p
+= strspn(p
, "/");
1131 n
= strcspn(p
, "/");
1132 if (!valid_slice_name(p
, n
))
1139 int cg_path_get_unit(const char *path
, char **ret
) {
1140 _cleanup_free_
char *unit
= NULL
;
1147 e
= skip_slices(path
);
1149 r
= cg_path_decode_unit(e
, &unit
);
1153 /* We skipped over the slices, don't accept any now */
1154 if (endswith(unit
, ".slice"))
1157 *ret
= TAKE_PTR(unit
);
1161 int cg_pid_get_unit(pid_t pid
, char **unit
) {
1162 _cleanup_free_
char *cgroup
= NULL
;
1167 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1171 return cg_path_get_unit(cgroup
, unit
);
1175 * Skip session-*.scope, but require it to be there.
1177 static const char *skip_session(const char *p
) {
1183 p
+= strspn(p
, "/");
1185 n
= strcspn(p
, "/");
1186 if (n
< STRLEN("session-x.scope"))
1189 if (memcmp(p
, "session-", 8) == 0 && memcmp(p
+ n
- 6, ".scope", 6) == 0) {
1190 char buf
[n
- 8 - 6 + 1];
1192 memcpy(buf
, p
+ 8, n
- 8 - 6);
1195 /* Note that session scopes never need unescaping,
1196 * since they cannot conflict with the kernel's own
1197 * names, hence we don't need to call cg_unescape()
1200 if (!session_id_valid(buf
))
1204 p
+= strspn(p
, "/");
1212 * Skip user@*.service, but require it to be there.
1214 static const char *skip_user_manager(const char *p
) {
1220 p
+= strspn(p
, "/");
1222 n
= strcspn(p
, "/");
1223 if (n
< STRLEN("user@x.service"))
1226 if (memcmp(p
, "user@", 5) == 0 && memcmp(p
+ n
- 8, ".service", 8) == 0) {
1227 char buf
[n
- 5 - 8 + 1];
1229 memcpy(buf
, p
+ 5, n
- 5 - 8);
1232 /* Note that user manager services never need unescaping,
1233 * since they cannot conflict with the kernel's own
1234 * names, hence we don't need to call cg_unescape()
1237 if (parse_uid(buf
, NULL
) < 0)
1241 p
+= strspn(p
, "/");
1249 static const char *skip_user_prefix(const char *path
) {
1254 /* Skip slices, if there are any */
1255 e
= skip_slices(path
);
1257 /* Skip the user manager, if it's in the path now... */
1258 t
= skip_user_manager(e
);
1262 /* Alternatively skip the user session if it is in the path... */
1263 return skip_session(e
);
1266 int cg_path_get_user_unit(const char *path
, char **ret
) {
1272 t
= skip_user_prefix(path
);
1276 /* And from here on it looks pretty much the same as for a system unit, hence let's use the same
1278 return cg_path_get_unit(t
, ret
);
1281 int cg_pid_get_user_unit(pid_t pid
, char **unit
) {
1282 _cleanup_free_
char *cgroup
= NULL
;
1287 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1291 return cg_path_get_user_unit(cgroup
, unit
);
1294 int cg_path_get_machine_name(const char *path
, char **machine
) {
1295 _cleanup_free_
char *u
= NULL
;
1299 r
= cg_path_get_unit(path
, &u
);
1303 sl
= strjoina("/run/systemd/machines/unit:", u
);
1304 return readlink_malloc(sl
, machine
);
1307 int cg_pid_get_machine_name(pid_t pid
, char **machine
) {
1308 _cleanup_free_
char *cgroup
= NULL
;
1313 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1317 return cg_path_get_machine_name(cgroup
, machine
);
1320 int cg_path_get_session(const char *path
, char **session
) {
1321 _cleanup_free_
char *unit
= NULL
;
1327 r
= cg_path_get_unit(path
, &unit
);
1331 start
= startswith(unit
, "session-");
1334 end
= endswith(start
, ".scope");
1339 if (!session_id_valid(start
))
1355 int cg_pid_get_session(pid_t pid
, char **session
) {
1356 _cleanup_free_
char *cgroup
= NULL
;
1359 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1363 return cg_path_get_session(cgroup
, session
);
1366 int cg_path_get_owner_uid(const char *path
, uid_t
*uid
) {
1367 _cleanup_free_
char *slice
= NULL
;
1373 r
= cg_path_get_slice(path
, &slice
);
1377 start
= startswith(slice
, "user-");
1380 end
= endswith(start
, ".slice");
1385 if (parse_uid(start
, uid
) < 0)
1391 int cg_pid_get_owner_uid(pid_t pid
, uid_t
*uid
) {
1392 _cleanup_free_
char *cgroup
= NULL
;
1395 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1399 return cg_path_get_owner_uid(cgroup
, uid
);
1402 int cg_path_get_slice(const char *p
, char **slice
) {
1403 const char *e
= NULL
;
1408 /* Finds the right-most slice unit from the beginning, but
1409 * stops before we come to the first non-slice unit. */
1414 p
+= strspn(p
, "/");
1416 n
= strcspn(p
, "/");
1417 if (!valid_slice_name(p
, n
)) {
1422 s
= strdup(SPECIAL_ROOT_SLICE
);
1430 return cg_path_decode_unit(e
, slice
);
1438 int cg_pid_get_slice(pid_t pid
, char **slice
) {
1439 _cleanup_free_
char *cgroup
= NULL
;
1444 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1448 return cg_path_get_slice(cgroup
, slice
);
1451 int cg_path_get_user_slice(const char *p
, char **slice
) {
1456 t
= skip_user_prefix(p
);
1460 /* And now it looks pretty much the same as for a system
1461 * slice, so let's just use the same parser from here on. */
1462 return cg_path_get_slice(t
, slice
);
1465 int cg_pid_get_user_slice(pid_t pid
, char **slice
) {
1466 _cleanup_free_
char *cgroup
= NULL
;
1471 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1475 return cg_path_get_user_slice(cgroup
, slice
);
1478 char *cg_escape(const char *p
) {
1479 bool need_prefix
= false;
1481 /* This implements very minimal escaping for names to be used
1482 * as file names in the cgroup tree: any name which might
1483 * conflict with a kernel name or is prefixed with '_' is
1484 * prefixed with a '_'. That way, when reading cgroup names it
1485 * is sufficient to remove a single prefixing underscore if
1488 /* The return value of this function (unlike cg_unescape())
1491 if (IN_SET(p
[0], 0, '_', '.') ||
1492 STR_IN_SET(p
, "notify_on_release", "release_agent", "tasks") ||
1493 startswith(p
, "cgroup."))
1498 dot
= strrchr(p
, '.');
1503 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
1506 n
= cgroup_controller_to_string(c
);
1511 if (memcmp(p
, n
, l
) != 0)
1521 return strjoin("_", p
);
1526 char *cg_unescape(const char *p
) {
1529 /* The return value of this function (unlike cg_escape())
1530 * doesn't need free()! */
1538 #define CONTROLLER_VALID \
1542 bool cg_controller_is_valid(const char *p
) {
1548 if (streq(p
, SYSTEMD_CGROUP_CONTROLLER
))
1551 s
= startswith(p
, "name=");
1555 if (IN_SET(*p
, 0, '_'))
1558 for (t
= p
; *t
; t
++)
1559 if (!strchr(CONTROLLER_VALID
, *t
))
1562 if (t
- p
> NAME_MAX
)
1568 int cg_slice_to_path(const char *unit
, char **ret
) {
1569 _cleanup_free_
char *p
= NULL
, *s
= NULL
, *e
= NULL
;
1576 if (streq(unit
, SPECIAL_ROOT_SLICE
)) {
1586 if (!unit_name_is_valid(unit
, UNIT_NAME_PLAIN
))
1589 if (!endswith(unit
, ".slice"))
1592 r
= unit_name_to_prefix(unit
, &p
);
1596 dash
= strchr(p
, '-');
1598 /* Don't allow initial dashes */
1603 _cleanup_free_
char *escaped
= NULL
;
1604 char n
[dash
- p
+ sizeof(".slice")];
1606 #if HAS_FEATURE_MEMORY_SANITIZER
1607 /* msan doesn't instrument stpncpy, so it thinks
1608 * n is later used uninitialized:
1609 * https://github.com/google/sanitizers/issues/926
1614 /* Don't allow trailing or double dashes */
1615 if (IN_SET(dash
[1], 0, '-'))
1618 strcpy(stpncpy(n
, p
, dash
- p
), ".slice");
1619 if (!unit_name_is_valid(n
, UNIT_NAME_PLAIN
))
1622 escaped
= cg_escape(n
);
1626 if (!strextend(&s
, escaped
, "/"))
1629 dash
= strchr(dash
+1, '-');
1632 e
= cg_escape(unit
);
1636 if (!strextend(&s
, e
))
1644 int cg_set_attribute(const char *controller
, const char *path
, const char *attribute
, const char *value
) {
1645 _cleanup_free_
char *p
= NULL
;
1648 r
= cg_get_path(controller
, path
, attribute
, &p
);
1652 return write_string_file(p
, value
, WRITE_STRING_FILE_DISABLE_BUFFER
);
1655 int cg_get_attribute(const char *controller
, const char *path
, const char *attribute
, char **ret
) {
1656 _cleanup_free_
char *p
= NULL
;
1659 r
= cg_get_path(controller
, path
, attribute
, &p
);
1663 return read_one_line_file(p
, ret
);
1666 int cg_get_attribute_as_uint64(const char *controller
, const char *path
, const char *attribute
, uint64_t *ret
) {
1667 _cleanup_free_
char *value
= NULL
;
1673 r
= cg_get_attribute(controller
, path
, attribute
, &value
);
1679 if (streq(value
, "max")) {
1680 *ret
= CGROUP_LIMIT_MAX
;
1684 r
= safe_atou64(value
, &v
);
1692 int cg_get_attribute_as_bool(const char *controller
, const char *path
, const char *attribute
, bool *ret
) {
1693 _cleanup_free_
char *value
= NULL
;
1698 r
= cg_get_attribute(controller
, path
, attribute
, &value
);
1704 r
= parse_boolean(value
);
1712 int cg_get_owner(const char *controller
, const char *path
, uid_t
*ret_uid
) {
1713 _cleanup_free_
char *f
= NULL
;
1719 r
= cg_get_path(controller
, path
, NULL
, &f
);
1723 r
= stat(f
, &stats
);
1727 *ret_uid
= stats
.st_uid
;
1731 int cg_get_keyed_attribute_full(
1732 const char *controller
,
1734 const char *attribute
,
1737 CGroupKeyMode mode
) {
1739 _cleanup_free_
char *filename
= NULL
, *contents
= NULL
;
1741 size_t n
, i
, n_done
= 0;
1745 /* Reads one or more fields of a cgroup v2 keyed attribute file. The 'keys' parameter should be an strv with
1746 * all keys to retrieve. The 'ret_values' parameter should be passed as string size with the same number of
1747 * entries as 'keys'. On success each entry will be set to the value of the matching key.
1749 * If the attribute file doesn't exist at all returns ENOENT, if any key is not found returns ENXIO. If mode
1750 * is set to GG_KEY_MODE_GRACEFUL we ignore missing keys and return those that were parsed successfully. */
1752 r
= cg_get_path(controller
, path
, attribute
, &filename
);
1756 r
= read_full_file(filename
, &contents
, NULL
);
1760 n
= strv_length(keys
);
1761 if (n
== 0) /* No keys to retrieve? That's easy, we are done then */
1764 /* Let's build this up in a temporary array for now in order not to clobber the return parameter on failure */
1765 v
= newa0(char*, n
);
1767 for (p
= contents
; *p
;) {
1768 const char *w
= NULL
;
1770 for (i
= 0; i
< n
; i
++)
1772 w
= first_word(p
, keys
[i
]);
1780 l
= strcspn(w
, NEWLINE
);
1781 v
[i
] = strndup(w
, l
);
1793 p
+= strcspn(p
, NEWLINE
);
1795 p
+= strspn(p
, NEWLINE
);
1798 if (mode
& CG_KEY_MODE_GRACEFUL
)
1804 for (i
= 0; i
< n
; i
++)
1810 memcpy(ret_values
, v
, sizeof(char*) * n
);
1811 if (mode
& CG_KEY_MODE_GRACEFUL
)
1817 int cg_mask_to_string(CGroupMask mask
, char **ret
) {
1818 _cleanup_free_
char *s
= NULL
;
1819 size_t n
= 0, allocated
= 0;
1830 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
1834 if (!FLAGS_SET(mask
, CGROUP_CONTROLLER_TO_MASK(c
)))
1837 k
= cgroup_controller_to_string(c
);
1840 if (!GREEDY_REALLOC(s
, allocated
, n
+ space
+ l
+ 1))
1845 memcpy(s
+ n
+ space
, k
, l
);
1859 int cg_mask_from_string(const char *value
, CGroupMask
*ret
) {
1866 _cleanup_free_
char *n
= NULL
;
1870 r
= extract_first_word(&value
, &n
, NULL
, 0);
1876 v
= cgroup_controller_from_string(n
);
1880 m
|= CGROUP_CONTROLLER_TO_MASK(v
);
1887 int cg_mask_supported_subtree(const char *root
, CGroupMask
*ret
) {
1891 /* Determines the mask of supported cgroup controllers. Only includes controllers we can make sense of and that
1892 * are actually accessible. Only covers real controllers, i.e. not the CGROUP_CONTROLLER_BPF_xyz
1893 * pseudo-controllers. */
1895 r
= cg_all_unified();
1899 _cleanup_free_
char *controllers
= NULL
, *path
= NULL
;
1901 /* In the unified hierarchy we can read the supported and accessible controllers from
1902 * the top-level cgroup attribute */
1904 r
= cg_get_path(SYSTEMD_CGROUP_CONTROLLER
, root
, "cgroup.controllers", &path
);
1908 r
= read_one_line_file(path
, &controllers
);
1912 r
= cg_mask_from_string(controllers
, &mask
);
1916 /* Mask controllers that are not supported in unified hierarchy. */
1917 mask
&= CGROUP_MASK_V2
;
1922 /* In the legacy hierarchy, we check which hierarchies are accessible. */
1925 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
1926 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
1929 if (!FLAGS_SET(CGROUP_MASK_V1
, bit
))
1932 n
= cgroup_controller_to_string(c
);
1933 if (controller_is_v1_accessible(root
, n
) >= 0)
1942 int cg_mask_supported(CGroupMask
*ret
) {
1943 _cleanup_free_
char *root
= NULL
;
1946 r
= cg_get_root_path(&root
);
1950 return cg_mask_supported_subtree(root
, ret
);
1953 int cg_kernel_controllers(Set
**ret
) {
1954 _cleanup_set_free_free_ Set
*controllers
= NULL
;
1955 _cleanup_fclose_
FILE *f
= NULL
;
1960 /* Determines the full list of kernel-known controllers. Might include controllers we don't actually support
1961 * and controllers that aren't currently accessible (because not mounted). This does not include "name="
1962 * pseudo-controllers. */
1964 controllers
= set_new(&string_hash_ops
);
1968 r
= fopen_unlocked("/proc/cgroups", "re", &f
);
1976 /* Ignore the header line */
1977 (void) read_line(f
, SIZE_MAX
, NULL
);
1984 if (fscanf(f
, "%ms %*i %*i %i", &controller
, &enabled
) != 2) {
1990 return errno_or_else(EIO
);
2000 if (!cg_controller_is_valid(controller
)) {
2005 r
= set_consume(controllers
, controller
);
2010 *ret
= TAKE_PTR(controllers
);
2015 /* The hybrid mode was initially implemented in v232 and simply mounted cgroup2 on
2016 * /sys/fs/cgroup/systemd. This unfortunately broke other tools (such as docker) which expected the v1
2017 * "name=systemd" hierarchy on /sys/fs/cgroup/systemd. From v233 and on, the hybrid mode mounts v2 on
2018 * /sys/fs/cgroup/unified and maintains "name=systemd" hierarchy on /sys/fs/cgroup/systemd for compatibility
2021 * To keep live upgrade working, we detect and support v232 layout. When v232 layout is detected, to keep
2022 * cgroup v2 process management but disable the compat dual layout, we return true on
2023 * cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) and false on cg_hybrid_unified().
2025 static thread_local
bool unified_systemd_v232
;
2027 int cg_unified_cached(bool flush
) {
2028 static thread_local CGroupUnified unified_cache
= CGROUP_UNIFIED_UNKNOWN
;
2032 /* Checks if we support the unified hierarchy. Returns an
2033 * error when the cgroup hierarchies aren't mounted yet or we
2034 * have any other trouble determining if the unified hierarchy
2038 unified_cache
= CGROUP_UNIFIED_UNKNOWN
;
2039 else if (unified_cache
>= CGROUP_UNIFIED_NONE
)
2040 return unified_cache
;
2042 if (statfs("/sys/fs/cgroup/", &fs
) < 0)
2043 return log_debug_errno(errno
, "statfs(\"/sys/fs/cgroup/\") failed: %m");
2045 if (F_TYPE_EQUAL(fs
.f_type
, CGROUP2_SUPER_MAGIC
)) {
2046 log_debug("Found cgroup2 on /sys/fs/cgroup/, full unified hierarchy");
2047 unified_cache
= CGROUP_UNIFIED_ALL
;
2048 } else if (F_TYPE_EQUAL(fs
.f_type
, TMPFS_MAGIC
)) {
2049 if (statfs("/sys/fs/cgroup/unified/", &fs
) == 0 &&
2050 F_TYPE_EQUAL(fs
.f_type
, CGROUP2_SUPER_MAGIC
)) {
2051 log_debug("Found cgroup2 on /sys/fs/cgroup/unified, unified hierarchy for systemd controller");
2052 unified_cache
= CGROUP_UNIFIED_SYSTEMD
;
2053 unified_systemd_v232
= false;
2055 if (statfs("/sys/fs/cgroup/systemd/", &fs
) < 0)
2056 return log_debug_errno(errno
, "statfs(\"/sys/fs/cgroup/systemd\" failed: %m");
2058 if (F_TYPE_EQUAL(fs
.f_type
, CGROUP2_SUPER_MAGIC
)) {
2059 log_debug("Found cgroup2 on /sys/fs/cgroup/systemd, unified hierarchy for systemd controller (v232 variant)");
2060 unified_cache
= CGROUP_UNIFIED_SYSTEMD
;
2061 unified_systemd_v232
= true;
2062 } else if (F_TYPE_EQUAL(fs
.f_type
, CGROUP_SUPER_MAGIC
)) {
2063 log_debug("Found cgroup on /sys/fs/cgroup/systemd, legacy hierarchy");
2064 unified_cache
= CGROUP_UNIFIED_NONE
;
2066 log_debug("Unexpected filesystem type %llx mounted on /sys/fs/cgroup/systemd, assuming legacy hierarchy",
2067 (unsigned long long) fs
.f_type
);
2068 unified_cache
= CGROUP_UNIFIED_NONE
;
2071 } else if (F_TYPE_EQUAL(fs
.f_type
, SYSFS_MAGIC
)) {
2072 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM
),
2073 "No filesystem is currently mounted on /sys/fs/cgroup.");
2075 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM
),
2076 "Unknown filesystem type %llx mounted on /sys/fs/cgroup.",
2077 (unsigned long long)fs
.f_type
);
2079 return unified_cache
;
2082 int cg_unified_controller(const char *controller
) {
2085 r
= cg_unified_cached(false);
2089 if (r
== CGROUP_UNIFIED_NONE
)
2092 if (r
>= CGROUP_UNIFIED_ALL
)
2095 return streq_ptr(controller
, SYSTEMD_CGROUP_CONTROLLER
);
2098 int cg_all_unified(void) {
2101 r
= cg_unified_cached(false);
2105 return r
>= CGROUP_UNIFIED_ALL
;
2108 int cg_hybrid_unified(void) {
2111 r
= cg_unified_cached(false);
2115 return r
== CGROUP_UNIFIED_SYSTEMD
&& !unified_systemd_v232
;
2118 const uint64_t cgroup_io_limit_defaults
[_CGROUP_IO_LIMIT_TYPE_MAX
] = {
2119 [CGROUP_IO_RBPS_MAX
] = CGROUP_LIMIT_MAX
,
2120 [CGROUP_IO_WBPS_MAX
] = CGROUP_LIMIT_MAX
,
2121 [CGROUP_IO_RIOPS_MAX
] = CGROUP_LIMIT_MAX
,
2122 [CGROUP_IO_WIOPS_MAX
] = CGROUP_LIMIT_MAX
,
2125 static const char* const cgroup_io_limit_type_table
[_CGROUP_IO_LIMIT_TYPE_MAX
] = {
2126 [CGROUP_IO_RBPS_MAX
] = "IOReadBandwidthMax",
2127 [CGROUP_IO_WBPS_MAX
] = "IOWriteBandwidthMax",
2128 [CGROUP_IO_RIOPS_MAX
] = "IOReadIOPSMax",
2129 [CGROUP_IO_WIOPS_MAX
] = "IOWriteIOPSMax",
2132 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type
, CGroupIOLimitType
);
2134 bool is_cgroup_fs(const struct statfs
*s
) {
2135 return is_fs_type(s
, CGROUP_SUPER_MAGIC
) ||
2136 is_fs_type(s
, CGROUP2_SUPER_MAGIC
);
2139 bool fd_is_cgroup_fs(int fd
) {
2142 if (fstatfs(fd
, &s
) < 0)
2145 return is_cgroup_fs(&s
);
2148 static const char *const cgroup_controller_table
[_CGROUP_CONTROLLER_MAX
] = {
2149 [CGROUP_CONTROLLER_CPU
] = "cpu",
2150 [CGROUP_CONTROLLER_CPUACCT
] = "cpuacct",
2151 [CGROUP_CONTROLLER_CPUSET
] = "cpuset",
2152 [CGROUP_CONTROLLER_IO
] = "io",
2153 [CGROUP_CONTROLLER_BLKIO
] = "blkio",
2154 [CGROUP_CONTROLLER_MEMORY
] = "memory",
2155 [CGROUP_CONTROLLER_DEVICES
] = "devices",
2156 [CGROUP_CONTROLLER_PIDS
] = "pids",
2157 [CGROUP_CONTROLLER_BPF_FIREWALL
] = "bpf-firewall",
2158 [CGROUP_CONTROLLER_BPF_DEVICES
] = "bpf-devices",
2161 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller
, CGroupController
);
2163 CGroupMask
get_cpu_accounting_mask(void) {
2164 static CGroupMask needed_mask
= (CGroupMask
) -1;
2166 /* On kernel ≥4.15 with unified hierarchy, cpu.stat's usage_usec is
2167 * provided externally from the CPU controller, which means we don't
2168 * need to enable the CPU controller just to get metrics. This is good,
2169 * because enabling the CPU controller comes at a minor performance
2170 * hit, especially when it's propagated deep into large hierarchies.
2171 * There's also no separate CPU accounting controller available within
2172 * a unified hierarchy.
2174 * This combination of factors results in the desired cgroup mask to
2175 * enable for CPU accounting varying as follows:
2177 * ╔═════════════════════╤═════════════════════╗
2178 * ║ Linux ≥4.15 │ Linux <4.15 ║
2179 * ╔═══════════════╬═════════════════════╪═════════════════════╣
2180 * ║ Unified ║ nothing │ CGROUP_MASK_CPU ║
2181 * ╟───────────────╫─────────────────────┼─────────────────────╢
2182 * ║ Hybrid/Legacy ║ CGROUP_MASK_CPUACCT │ CGROUP_MASK_CPUACCT ║
2183 * ╚═══════════════╩═════════════════════╧═════════════════════╝
2185 * We check kernel version here instead of manually checking whether
2186 * cpu.stat is present for every cgroup, as that check in itself would
2187 * already be fairly expensive.
2189 * Kernels where this patch has been backported will therefore have the
2190 * CPU controller enabled unnecessarily. This is more expensive than
2191 * necessary, but harmless. ☺️
2194 if (needed_mask
== (CGroupMask
) -1) {
2195 if (cg_all_unified()) {
2197 assert_se(uname(&u
) >= 0);
2199 if (strverscmp_improved(u
.release
, "4.15") < 0)
2200 needed_mask
= CGROUP_MASK_CPU
;
2204 needed_mask
= CGROUP_MASK_CPUACCT
;
2210 bool cpu_accounting_is_cheap(void) {
2211 return get_cpu_accounting_mask() == 0;
2214 static const char* const managed_oom_mode_table
[_MANAGED_OOM_MODE_MAX
] = {
2215 [MANAGED_OOM_AUTO
] = "auto",
2216 [MANAGED_OOM_KILL
] = "kill",
2219 DEFINE_STRING_TABLE_LOOKUP(managed_oom_mode
, ManagedOOMMode
);
2221 static const char* const managed_oom_preference_table
[_MANAGED_OOM_PREFERENCE_MAX
] = {
2222 [MANAGED_OOM_PREFERENCE_NONE
] = "none",
2223 [MANAGED_OOM_PREFERENCE_AVOID
] = "avoid",
2224 [MANAGED_OOM_PREFERENCE_OMIT
] = "omit",
2227 DEFINE_STRING_TABLE_LOOKUP(managed_oom_preference
, ManagedOOMPreference
);