2 This file is part of systemd.
4 Copyright 2010 Lennart Poettering
6 systemd is free software; you can redistribute it and/or modify it
7 under the terms of the GNU Lesser General Public License as published by
8 the Free Software Foundation; either version 2.1 of the License, or
9 (at your option) any later version.
11 systemd is distributed in the hope that it will be useful, but
12 WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
16 You should have received a copy of the GNU Lesser General Public License
17 along with systemd; If not, see <http://www.gnu.org/licenses/>.
29 #include <sys/statfs.h>
30 #include <sys/types.h>
33 #include "alloc-util.h"
34 #include "cgroup-util.h"
36 #include "dirent-util.h"
37 #include "extract-word.h"
40 #include "formats-util.h"
43 #include "login-util.h"
47 #include "parse-util.h"
48 #include "path-util.h"
49 #include "proc-cmdline.h"
50 #include "process-util.h"
53 #include "stat-util.h"
54 #include "stdio-util.h"
55 #include "string-table.h"
56 #include "string-util.h"
57 #include "unit-name.h"
58 #include "user-util.h"
60 int cg_enumerate_processes(const char *controller
, const char *path
, FILE **_f
) {
61 _cleanup_free_
char *fs
= NULL
;
67 r
= cg_get_path(controller
, path
, "cgroup.procs", &fs
);
79 int cg_read_pid(FILE *f
, pid_t
*_pid
) {
82 /* Note that the cgroup.procs might contain duplicates! See
83 * cgroups.txt for details. */
89 if (fscanf(f
, "%lu", &ul
) != 1) {
94 return errno
> 0 ? -errno
: -EIO
;
104 int cg_read_event(const char *controller
, const char *path
, const char *event
,
107 _cleanup_free_
char *events
= NULL
, *content
= NULL
;
111 r
= cg_get_path(controller
, path
, "cgroup.events", &events
);
115 r
= read_full_file(events
, &content
, NULL
);
120 while ((line
= strsep(&p
, "\n"))) {
123 key
= strsep(&line
, " ");
127 if (strcmp(key
, event
))
137 bool cg_ns_supported(void) {
138 static thread_local
int enabled
= -1;
143 if (access("/proc/self/ns/cgroup", F_OK
) == 0)
151 int cg_enumerate_subgroups(const char *controller
, const char *path
, DIR **_d
) {
152 _cleanup_free_
char *fs
= NULL
;
158 /* This is not recursive! */
160 r
= cg_get_path(controller
, path
, NULL
, &fs
);
172 int cg_read_subgroup(DIR *d
, char **fn
) {
178 FOREACH_DIRENT_ALL(de
, d
, return -errno
) {
181 if (de
->d_type
!= DT_DIR
)
184 if (streq(de
->d_name
, ".") ||
185 streq(de
->d_name
, ".."))
188 b
= strdup(de
->d_name
);
199 int cg_rmdir(const char *controller
, const char *path
) {
200 _cleanup_free_
char *p
= NULL
;
203 r
= cg_get_path(controller
, path
, NULL
, &p
);
208 if (r
< 0 && errno
!= ENOENT
)
215 const char *controller
,
220 cg_kill_log_func_t log_kill
,
223 _cleanup_set_free_ Set
*allocated_set
= NULL
;
230 /* Don't send SIGCONT twice. Also, SIGKILL always works even when process is suspended, hence don't send
231 * SIGCONT on SIGKILL. */
232 if (IN_SET(sig
, SIGCONT
, SIGKILL
))
233 flags
&= ~CGROUP_SIGCONT
;
235 /* This goes through the tasks list and kills them all. This
236 * is repeated until no further processes are added to the
237 * tasks list, to properly handle forking processes */
240 s
= allocated_set
= set_new(NULL
);
248 _cleanup_fclose_
FILE *f
= NULL
;
252 r
= cg_enumerate_processes(controller
, path
, &f
);
254 if (ret
>= 0 && r
!= -ENOENT
)
260 while ((r
= cg_read_pid(f
, &pid
)) > 0) {
262 if ((flags
& CGROUP_IGNORE_SELF
) && pid
== my_pid
)
265 if (set_get(s
, PID_TO_PTR(pid
)) == PID_TO_PTR(pid
))
269 log_kill(pid
, sig
, userdata
);
271 /* If we haven't killed this process yet, kill
273 if (kill(pid
, sig
) < 0) {
274 if (ret
>= 0 && errno
!= ESRCH
)
277 if (flags
& CGROUP_SIGCONT
)
278 (void) kill(pid
, SIGCONT
);
286 r
= set_put(s
, PID_TO_PTR(pid
));
302 /* To avoid racing against processes which fork
303 * quicker than we can kill them we repeat this until
304 * no new pids need to be killed. */
311 int cg_kill_recursive(
312 const char *controller
,
317 cg_kill_log_func_t log_kill
,
320 _cleanup_set_free_ Set
*allocated_set
= NULL
;
321 _cleanup_closedir_
DIR *d
= NULL
;
329 s
= allocated_set
= set_new(NULL
);
334 ret
= cg_kill(controller
, path
, sig
, flags
, s
, log_kill
, userdata
);
336 r
= cg_enumerate_subgroups(controller
, path
, &d
);
338 if (ret
>= 0 && r
!= -ENOENT
)
344 while ((r
= cg_read_subgroup(d
, &fn
)) > 0) {
345 _cleanup_free_
char *p
= NULL
;
347 p
= strjoin(path
, "/", fn
, NULL
);
352 r
= cg_kill_recursive(controller
, p
, sig
, flags
, s
, log_kill
, userdata
);
353 if (r
!= 0 && ret
>= 0)
356 if (ret
>= 0 && r
< 0)
359 if (flags
& CGROUP_REMOVE
) {
360 r
= cg_rmdir(controller
, path
);
361 if (r
< 0 && ret
>= 0 && r
!= -ENOENT
&& r
!= -EBUSY
)
376 _cleanup_set_free_ Set
*s
= NULL
;
392 _cleanup_fclose_
FILE *f
= NULL
;
396 r
= cg_enumerate_processes(cfrom
, pfrom
, &f
);
398 if (ret
>= 0 && r
!= -ENOENT
)
404 while ((r
= cg_read_pid(f
, &pid
)) > 0) {
406 /* This might do weird stuff if we aren't a
407 * single-threaded program. However, we
408 * luckily know we are not */
409 if ((flags
& CGROUP_IGNORE_SELF
) && pid
== my_pid
)
412 if (set_get(s
, PID_TO_PTR(pid
)) == PID_TO_PTR(pid
))
415 /* Ignore kernel threads. Since they can only
416 * exist in the root cgroup, we only check for
419 (isempty(pfrom
) || path_equal(pfrom
, "/")) &&
420 is_kernel_thread(pid
) > 0)
423 r
= cg_attach(cto
, pto
, pid
);
425 if (ret
>= 0 && r
!= -ESRCH
)
432 r
= set_put(s
, PID_TO_PTR(pid
));
452 int cg_migrate_recursive(
459 _cleanup_closedir_
DIR *d
= NULL
;
468 ret
= cg_migrate(cfrom
, pfrom
, cto
, pto
, flags
);
470 r
= cg_enumerate_subgroups(cfrom
, pfrom
, &d
);
472 if (ret
>= 0 && r
!= -ENOENT
)
478 while ((r
= cg_read_subgroup(d
, &fn
)) > 0) {
479 _cleanup_free_
char *p
= NULL
;
481 p
= strjoin(pfrom
, "/", fn
, NULL
);
486 r
= cg_migrate_recursive(cfrom
, p
, cto
, pto
, flags
);
487 if (r
!= 0 && ret
>= 0)
491 if (r
< 0 && ret
>= 0)
494 if (flags
& CGROUP_REMOVE
) {
495 r
= cg_rmdir(cfrom
, pfrom
);
496 if (r
< 0 && ret
>= 0 && r
!= -ENOENT
&& r
!= -EBUSY
)
503 int cg_migrate_recursive_fallback(
517 r
= cg_migrate_recursive(cfrom
, pfrom
, cto
, pto
, flags
);
519 char prefix
[strlen(pto
) + 1];
521 /* This didn't work? Then let's try all prefixes of the destination */
523 PATH_FOREACH_PREFIX(prefix
, pto
) {
526 q
= cg_migrate_recursive(cfrom
, pfrom
, cto
, prefix
, flags
);
535 static const char *controller_to_dirname(const char *controller
) {
540 /* Converts a controller name to the directory name below
541 * /sys/fs/cgroup/ we want to mount it to. Effectively, this
542 * just cuts off the name= prefixed used for named
543 * hierarchies, if it is specified. */
545 e
= startswith(controller
, "name=");
552 static int join_path_legacy(const char *controller
, const char *path
, const char *suffix
, char **fs
) {
559 dn
= controller_to_dirname(controller
);
561 if (isempty(path
) && isempty(suffix
))
562 t
= strappend("/sys/fs/cgroup/", dn
);
563 else if (isempty(path
))
564 t
= strjoin("/sys/fs/cgroup/", dn
, "/", suffix
, NULL
);
565 else if (isempty(suffix
))
566 t
= strjoin("/sys/fs/cgroup/", dn
, "/", path
, NULL
);
568 t
= strjoin("/sys/fs/cgroup/", dn
, "/", path
, "/", suffix
, NULL
);
576 static int join_path_unified(const char *path
, const char *suffix
, char **fs
) {
581 if (isempty(path
) && isempty(suffix
))
582 t
= strdup("/sys/fs/cgroup");
583 else if (isempty(path
))
584 t
= strappend("/sys/fs/cgroup/", suffix
);
585 else if (isempty(suffix
))
586 t
= strappend("/sys/fs/cgroup/", path
);
588 t
= strjoin("/sys/fs/cgroup/", path
, "/", suffix
, NULL
);
596 int cg_get_path(const char *controller
, const char *path
, const char *suffix
, char **fs
) {
604 /* If no controller is specified, we return the path
605 * *below* the controllers, without any prefix. */
607 if (!path
&& !suffix
)
615 t
= strjoin(path
, "/", suffix
, NULL
);
619 *fs
= path_kill_slashes(t
);
623 if (!cg_controller_is_valid(controller
))
626 unified
= cg_unified();
631 r
= join_path_unified(path
, suffix
, fs
);
633 r
= join_path_legacy(controller
, path
, suffix
, fs
);
637 path_kill_slashes(*fs
);
641 static int controller_is_accessible(const char *controller
) {
646 /* Checks whether a specific controller is accessible,
647 * i.e. its hierarchy mounted. In the unified hierarchy all
648 * controllers are considered accessible, except for the named
651 if (!cg_controller_is_valid(controller
))
654 unified
= cg_unified();
658 /* We don't support named hierarchies if we are using
659 * the unified hierarchy. */
661 if (streq(controller
, SYSTEMD_CGROUP_CONTROLLER
))
664 if (startswith(controller
, "name="))
670 dn
= controller_to_dirname(controller
);
671 cc
= strjoina("/sys/fs/cgroup/", dn
);
673 if (laccess(cc
, F_OK
) < 0)
680 int cg_get_path_and_check(const char *controller
, const char *path
, const char *suffix
, char **fs
) {
686 /* Check if the specified controller is actually accessible */
687 r
= controller_is_accessible(controller
);
691 return cg_get_path(controller
, path
, suffix
, fs
);
694 static int trim_cb(const char *path
, const struct stat
*sb
, int typeflag
, struct FTW
*ftwbuf
) {
699 if (typeflag
!= FTW_DP
)
702 if (ftwbuf
->level
< 1)
709 int cg_trim(const char *controller
, const char *path
, bool delete_root
) {
710 _cleanup_free_
char *fs
= NULL
;
715 r
= cg_get_path(controller
, path
, NULL
, &fs
);
720 if (nftw(fs
, trim_cb
, 64, FTW_DEPTH
|FTW_MOUNT
|FTW_PHYS
) != 0) {
730 if (rmdir(fs
) < 0 && errno
!= ENOENT
)
737 int cg_create(const char *controller
, const char *path
) {
738 _cleanup_free_
char *fs
= NULL
;
741 r
= cg_get_path_and_check(controller
, path
, NULL
, &fs
);
745 r
= mkdir_parents(fs
, 0755);
749 if (mkdir(fs
, 0755) < 0) {
760 int cg_create_and_attach(const char *controller
, const char *path
, pid_t pid
) {
765 r
= cg_create(controller
, path
);
769 q
= cg_attach(controller
, path
, pid
);
773 /* This does not remove the cgroup on failure */
777 int cg_attach(const char *controller
, const char *path
, pid_t pid
) {
778 _cleanup_free_
char *fs
= NULL
;
779 char c
[DECIMAL_STR_MAX(pid_t
) + 2];
785 r
= cg_get_path_and_check(controller
, path
, "cgroup.procs", &fs
);
792 xsprintf(c
, PID_FMT
"\n", pid
);
794 return write_string_file(fs
, c
, 0);
797 int cg_attach_fallback(const char *controller
, const char *path
, pid_t pid
) {
804 r
= cg_attach(controller
, path
, pid
);
806 char prefix
[strlen(path
) + 1];
808 /* This didn't work? Then let's try all prefixes of
811 PATH_FOREACH_PREFIX(prefix
, path
) {
814 q
= cg_attach(controller
, prefix
, pid
);
823 int cg_set_group_access(
824 const char *controller
,
830 _cleanup_free_
char *fs
= NULL
;
833 if (mode
== MODE_INVALID
&& uid
== UID_INVALID
&& gid
== GID_INVALID
)
836 if (mode
!= MODE_INVALID
)
839 r
= cg_get_path(controller
, path
, NULL
, &fs
);
843 return chmod_and_chown(fs
, mode
, uid
, gid
);
846 int cg_set_task_access(
847 const char *controller
,
853 _cleanup_free_
char *fs
= NULL
, *procs
= NULL
;
858 if (mode
== MODE_INVALID
&& uid
== UID_INVALID
&& gid
== GID_INVALID
)
861 if (mode
!= MODE_INVALID
)
864 r
= cg_get_path(controller
, path
, "cgroup.procs", &fs
);
868 r
= chmod_and_chown(fs
, mode
, uid
, gid
);
872 unified
= cg_unified();
878 /* Compatibility, Always keep values for "tasks" in sync with
880 if (cg_get_path(controller
, path
, "tasks", &procs
) >= 0)
881 (void) chmod_and_chown(procs
, mode
, uid
, gid
);
886 int cg_pid_get_path(const char *controller
, pid_t pid
, char **path
) {
887 _cleanup_fclose_
FILE *f
= NULL
;
896 unified
= cg_unified();
901 if (!cg_controller_is_valid(controller
))
904 controller
= SYSTEMD_CGROUP_CONTROLLER
;
906 cs
= strlen(controller
);
909 fs
= procfs_file_alloca(pid
, "cgroup");
912 return errno
== ENOENT
? -ESRCH
: -errno
;
914 FOREACH_LINE(line
, f
, return -errno
) {
920 e
= startswith(line
, "0:");
930 const char *word
, *state
;
933 l
= strchr(line
, ':');
943 FOREACH_WORD_SEPARATOR(word
, k
, l
, ",", state
) {
944 if (k
== cs
&& memcmp(word
, controller
, cs
) == 0) {
965 int cg_install_release_agent(const char *controller
, const char *agent
) {
966 _cleanup_free_
char *fs
= NULL
, *contents
= NULL
;
972 unified
= cg_unified();
975 if (unified
) /* doesn't apply to unified hierarchy */
978 r
= cg_get_path(controller
, NULL
, "release_agent", &fs
);
982 r
= read_one_line_file(fs
, &contents
);
986 sc
= strstrip(contents
);
988 r
= write_string_file(fs
, agent
, 0);
991 } else if (!path_equal(sc
, agent
))
995 r
= cg_get_path(controller
, NULL
, "notify_on_release", &fs
);
999 contents
= mfree(contents
);
1000 r
= read_one_line_file(fs
, &contents
);
1004 sc
= strstrip(contents
);
1005 if (streq(sc
, "0")) {
1006 r
= write_string_file(fs
, "1", 0);
1013 if (!streq(sc
, "1"))
1019 int cg_uninstall_release_agent(const char *controller
) {
1020 _cleanup_free_
char *fs
= NULL
;
1023 unified
= cg_unified();
1026 if (unified
) /* Doesn't apply to unified hierarchy */
1029 r
= cg_get_path(controller
, NULL
, "notify_on_release", &fs
);
1033 r
= write_string_file(fs
, "0", 0);
1039 r
= cg_get_path(controller
, NULL
, "release_agent", &fs
);
1043 r
= write_string_file(fs
, "", 0);
1050 int cg_is_empty(const char *controller
, const char *path
) {
1051 _cleanup_fclose_
FILE *f
= NULL
;
1057 r
= cg_enumerate_processes(controller
, path
, &f
);
1063 r
= cg_read_pid(f
, &pid
);
1070 int cg_is_empty_recursive(const char *controller
, const char *path
) {
1075 /* The root cgroup is always populated */
1076 if (controller
&& (isempty(path
) || path_equal(path
, "/")))
1079 unified
= cg_unified();
1084 _cleanup_free_
char *t
= NULL
;
1086 /* On the unified hierarchy we can check empty state
1087 * via the "populated" attribute of "cgroup.events". */
1089 r
= cg_read_event(controller
, path
, "populated", &t
);
1093 return streq(t
, "0");
1095 _cleanup_closedir_
DIR *d
= NULL
;
1098 r
= cg_is_empty(controller
, path
);
1102 r
= cg_enumerate_subgroups(controller
, path
, &d
);
1108 while ((r
= cg_read_subgroup(d
, &fn
)) > 0) {
1109 _cleanup_free_
char *p
= NULL
;
1111 p
= strjoin(path
, "/", fn
, NULL
);
1116 r
= cg_is_empty_recursive(controller
, p
);
1127 int cg_split_spec(const char *spec
, char **controller
, char **path
) {
1128 char *t
= NULL
, *u
= NULL
;
1134 if (!path_is_safe(spec
))
1142 *path
= path_kill_slashes(t
);
1151 e
= strchr(spec
, ':');
1153 if (!cg_controller_is_valid(spec
))
1170 t
= strndup(spec
, e
-spec
);
1173 if (!cg_controller_is_valid(t
)) {
1187 if (!path_is_safe(u
) ||
1188 !path_is_absolute(u
)) {
1194 path_kill_slashes(u
);
1210 int cg_mangle_path(const char *path
, char **result
) {
1211 _cleanup_free_
char *c
= NULL
, *p
= NULL
;
1218 /* First, check if it already is a filesystem path */
1219 if (path_startswith(path
, "/sys/fs/cgroup")) {
1225 *result
= path_kill_slashes(t
);
1229 /* Otherwise, treat it as cg spec */
1230 r
= cg_split_spec(path
, &c
, &p
);
1234 return cg_get_path(c
?: SYSTEMD_CGROUP_CONTROLLER
, p
?: "/", NULL
, result
);
1237 int cg_get_root_path(char **path
) {
1243 r
= cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER
, 1, &p
);
1247 e
= endswith(p
, "/" SPECIAL_INIT_SCOPE
);
1249 e
= endswith(p
, "/" SPECIAL_SYSTEM_SLICE
); /* legacy */
1251 e
= endswith(p
, "/system"); /* even more legacy */
1259 int cg_shift_path(const char *cgroup
, const char *root
, const char **shifted
) {
1260 _cleanup_free_
char *rt
= NULL
;
1268 /* If the root was specified let's use that, otherwise
1269 * let's determine it from PID 1 */
1271 r
= cg_get_root_path(&rt
);
1278 p
= path_startswith(cgroup
, root
);
1279 if (p
&& p
> cgroup
)
1287 int cg_pid_get_path_shifted(pid_t pid
, const char *root
, char **cgroup
) {
1288 _cleanup_free_
char *raw
= NULL
;
1295 r
= cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER
, pid
, &raw
);
1299 r
= cg_shift_path(raw
, root
, &c
);
1319 int cg_path_decode_unit(const char *cgroup
, char **unit
) {
1326 n
= strcspn(cgroup
, "/");
1330 c
= strndupa(cgroup
, n
);
1333 if (!unit_name_is_valid(c
, UNIT_NAME_PLAIN
|UNIT_NAME_INSTANCE
))
1344 static bool valid_slice_name(const char *p
, size_t n
) {
1349 if (n
< strlen("x.slice"))
1352 if (memcmp(p
+ n
- 6, ".slice", 6) == 0) {
1358 c
= cg_unescape(buf
);
1360 return unit_name_is_valid(c
, UNIT_NAME_PLAIN
);
1366 static const char *skip_slices(const char *p
) {
1369 /* Skips over all slice assignments */
1374 p
+= strspn(p
, "/");
1376 n
= strcspn(p
, "/");
1377 if (!valid_slice_name(p
, n
))
1384 int cg_path_get_unit(const char *path
, char **ret
) {
1392 e
= skip_slices(path
);
1394 r
= cg_path_decode_unit(e
, &unit
);
1398 /* We skipped over the slices, don't accept any now */
1399 if (endswith(unit
, ".slice")) {
1408 int cg_pid_get_unit(pid_t pid
, char **unit
) {
1409 _cleanup_free_
char *cgroup
= NULL
;
1414 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1418 return cg_path_get_unit(cgroup
, unit
);
1422 * Skip session-*.scope, but require it to be there.
1424 static const char *skip_session(const char *p
) {
1430 p
+= strspn(p
, "/");
1432 n
= strcspn(p
, "/");
1433 if (n
< strlen("session-x.scope"))
1436 if (memcmp(p
, "session-", 8) == 0 && memcmp(p
+ n
- 6, ".scope", 6) == 0) {
1437 char buf
[n
- 8 - 6 + 1];
1439 memcpy(buf
, p
+ 8, n
- 8 - 6);
1442 /* Note that session scopes never need unescaping,
1443 * since they cannot conflict with the kernel's own
1444 * names, hence we don't need to call cg_unescape()
1447 if (!session_id_valid(buf
))
1451 p
+= strspn(p
, "/");
1459 * Skip user@*.service, but require it to be there.
1461 static const char *skip_user_manager(const char *p
) {
1467 p
+= strspn(p
, "/");
1469 n
= strcspn(p
, "/");
1470 if (n
< strlen("user@x.service"))
1473 if (memcmp(p
, "user@", 5) == 0 && memcmp(p
+ n
- 8, ".service", 8) == 0) {
1474 char buf
[n
- 5 - 8 + 1];
1476 memcpy(buf
, p
+ 5, n
- 5 - 8);
1479 /* Note that user manager services never need unescaping,
1480 * since they cannot conflict with the kernel's own
1481 * names, hence we don't need to call cg_unescape()
1484 if (parse_uid(buf
, NULL
) < 0)
1488 p
+= strspn(p
, "/");
1496 static const char *skip_user_prefix(const char *path
) {
1501 /* Skip slices, if there are any */
1502 e
= skip_slices(path
);
1504 /* Skip the user manager, if it's in the path now... */
1505 t
= skip_user_manager(e
);
1509 /* Alternatively skip the user session if it is in the path... */
1510 return skip_session(e
);
1513 int cg_path_get_user_unit(const char *path
, char **ret
) {
1519 t
= skip_user_prefix(path
);
1523 /* And from here on it looks pretty much the same as for a
1524 * system unit, hence let's use the same parser from here
1526 return cg_path_get_unit(t
, ret
);
1529 int cg_pid_get_user_unit(pid_t pid
, char **unit
) {
1530 _cleanup_free_
char *cgroup
= NULL
;
1535 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1539 return cg_path_get_user_unit(cgroup
, unit
);
1542 int cg_path_get_machine_name(const char *path
, char **machine
) {
1543 _cleanup_free_
char *u
= NULL
;
1547 r
= cg_path_get_unit(path
, &u
);
1551 sl
= strjoina("/run/systemd/machines/unit:", u
);
1552 return readlink_malloc(sl
, machine
);
1555 int cg_pid_get_machine_name(pid_t pid
, char **machine
) {
1556 _cleanup_free_
char *cgroup
= NULL
;
1561 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1565 return cg_path_get_machine_name(cgroup
, machine
);
1568 int cg_path_get_session(const char *path
, char **session
) {
1569 _cleanup_free_
char *unit
= NULL
;
1575 r
= cg_path_get_unit(path
, &unit
);
1579 start
= startswith(unit
, "session-");
1582 end
= endswith(start
, ".scope");
1587 if (!session_id_valid(start
))
1603 int cg_pid_get_session(pid_t pid
, char **session
) {
1604 _cleanup_free_
char *cgroup
= NULL
;
1607 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1611 return cg_path_get_session(cgroup
, session
);
1614 int cg_path_get_owner_uid(const char *path
, uid_t
*uid
) {
1615 _cleanup_free_
char *slice
= NULL
;
1621 r
= cg_path_get_slice(path
, &slice
);
1625 start
= startswith(slice
, "user-");
1628 end
= endswith(start
, ".slice");
1633 if (parse_uid(start
, uid
) < 0)
1639 int cg_pid_get_owner_uid(pid_t pid
, uid_t
*uid
) {
1640 _cleanup_free_
char *cgroup
= NULL
;
1643 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1647 return cg_path_get_owner_uid(cgroup
, uid
);
1650 int cg_path_get_slice(const char *p
, char **slice
) {
1651 const char *e
= NULL
;
1656 /* Finds the right-most slice unit from the beginning, but
1657 * stops before we come to the first non-slice unit. */
1662 p
+= strspn(p
, "/");
1664 n
= strcspn(p
, "/");
1665 if (!valid_slice_name(p
, n
)) {
1670 s
= strdup("-.slice");
1678 return cg_path_decode_unit(e
, slice
);
1686 int cg_pid_get_slice(pid_t pid
, char **slice
) {
1687 _cleanup_free_
char *cgroup
= NULL
;
1692 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1696 return cg_path_get_slice(cgroup
, slice
);
1699 int cg_path_get_user_slice(const char *p
, char **slice
) {
1704 t
= skip_user_prefix(p
);
1708 /* And now it looks pretty much the same as for a system
1709 * slice, so let's just use the same parser from here on. */
1710 return cg_path_get_slice(t
, slice
);
1713 int cg_pid_get_user_slice(pid_t pid
, char **slice
) {
1714 _cleanup_free_
char *cgroup
= NULL
;
1719 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1723 return cg_path_get_user_slice(cgroup
, slice
);
1726 char *cg_escape(const char *p
) {
1727 bool need_prefix
= false;
1729 /* This implements very minimal escaping for names to be used
1730 * as file names in the cgroup tree: any name which might
1731 * conflict with a kernel name or is prefixed with '_' is
1732 * prefixed with a '_'. That way, when reading cgroup names it
1733 * is sufficient to remove a single prefixing underscore if
1736 /* The return value of this function (unlike cg_unescape())
1742 streq(p
, "notify_on_release") ||
1743 streq(p
, "release_agent") ||
1744 streq(p
, "tasks") ||
1745 startswith(p
, "cgroup."))
1750 dot
= strrchr(p
, '.');
1755 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
1758 n
= cgroup_controller_to_string(c
);
1763 if (memcmp(p
, n
, l
) != 0)
1773 return strappend("_", p
);
1778 char *cg_unescape(const char *p
) {
1781 /* The return value of this function (unlike cg_escape())
1782 * doesn't need free()! */
1790 #define CONTROLLER_VALID \
1794 bool cg_controller_is_valid(const char *p
) {
1800 s
= startswith(p
, "name=");
1804 if (*p
== 0 || *p
== '_')
1807 for (t
= p
; *t
; t
++)
1808 if (!strchr(CONTROLLER_VALID
, *t
))
1811 if (t
- p
> FILENAME_MAX
)
1817 int cg_slice_to_path(const char *unit
, char **ret
) {
1818 _cleanup_free_
char *p
= NULL
, *s
= NULL
, *e
= NULL
;
1825 if (streq(unit
, "-.slice")) {
1835 if (!unit_name_is_valid(unit
, UNIT_NAME_PLAIN
))
1838 if (!endswith(unit
, ".slice"))
1841 r
= unit_name_to_prefix(unit
, &p
);
1845 dash
= strchr(p
, '-');
1847 /* Don't allow initial dashes */
1852 _cleanup_free_
char *escaped
= NULL
;
1853 char n
[dash
- p
+ sizeof(".slice")];
1855 /* Don't allow trailing or double dashes */
1856 if (dash
[1] == 0 || dash
[1] == '-')
1859 strcpy(stpncpy(n
, p
, dash
- p
), ".slice");
1860 if (!unit_name_is_valid(n
, UNIT_NAME_PLAIN
))
1863 escaped
= cg_escape(n
);
1867 if (!strextend(&s
, escaped
, "/", NULL
))
1870 dash
= strchr(dash
+1, '-');
1873 e
= cg_escape(unit
);
1877 if (!strextend(&s
, e
, NULL
))
1886 int cg_set_attribute(const char *controller
, const char *path
, const char *attribute
, const char *value
) {
1887 _cleanup_free_
char *p
= NULL
;
1890 r
= cg_get_path(controller
, path
, attribute
, &p
);
1894 return write_string_file(p
, value
, 0);
1897 int cg_get_attribute(const char *controller
, const char *path
, const char *attribute
, char **ret
) {
1898 _cleanup_free_
char *p
= NULL
;
1901 r
= cg_get_path(controller
, path
, attribute
, &p
);
1905 return read_one_line_file(p
, ret
);
1908 int cg_get_keyed_attribute(const char *controller
, const char *path
, const char *attribute
, const char **keys
, char **values
) {
1909 _cleanup_free_
char *filename
= NULL
, *content
= NULL
;
1913 for (i
= 0; keys
[i
]; i
++)
1916 r
= cg_get_path(controller
, path
, attribute
, &filename
);
1920 r
= read_full_file(filename
, &content
, NULL
);
1925 while ((line
= strsep(&p
, "\n"))) {
1928 key
= strsep(&line
, " ");
1930 for (i
= 0; keys
[i
]; i
++) {
1931 if (streq(key
, keys
[i
])) {
1932 values
[i
] = strdup(line
);
1938 for (i
= 0; keys
[i
]; i
++) {
1940 for (i
= 0; keys
[i
]; i
++) {
1951 int cg_create_everywhere(CGroupMask supported
, CGroupMask mask
, const char *path
) {
1955 /* This one will create a cgroup in our private tree, but also
1956 * duplicate it in the trees specified in mask, and remove it
1959 /* First create the cgroup in our own hierarchy. */
1960 r
= cg_create(SYSTEMD_CGROUP_CONTROLLER
, path
);
1964 /* If we are in the unified hierarchy, we are done now */
1965 unified
= cg_unified();
1971 /* Otherwise, do the same in the other hierarchies */
1972 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
1973 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
1976 n
= cgroup_controller_to_string(c
);
1979 (void) cg_create(n
, path
);
1980 else if (supported
& bit
)
1981 (void) cg_trim(n
, path
, true);
1987 int cg_attach_everywhere(CGroupMask supported
, const char *path
, pid_t pid
, cg_migrate_callback_t path_callback
, void *userdata
) {
1991 r
= cg_attach(SYSTEMD_CGROUP_CONTROLLER
, path
, pid
);
1995 unified
= cg_unified();
2001 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
2002 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
2003 const char *p
= NULL
;
2005 if (!(supported
& bit
))
2009 p
= path_callback(bit
, userdata
);
2014 (void) cg_attach_fallback(cgroup_controller_to_string(c
), p
, pid
);
2020 int cg_attach_many_everywhere(CGroupMask supported
, const char *path
, Set
* pids
, cg_migrate_callback_t path_callback
, void *userdata
) {
2025 SET_FOREACH(pidp
, pids
, i
) {
2026 pid_t pid
= PTR_TO_PID(pidp
);
2029 q
= cg_attach_everywhere(supported
, path
, pid
, path_callback
, userdata
);
2030 if (q
< 0 && r
>= 0)
2037 int cg_migrate_everywhere(CGroupMask supported
, const char *from
, const char *to
, cg_migrate_callback_t to_callback
, void *userdata
) {
2041 if (!path_equal(from
, to
)) {
2042 r
= cg_migrate_recursive(SYSTEMD_CGROUP_CONTROLLER
, from
, SYSTEMD_CGROUP_CONTROLLER
, to
, CGROUP_REMOVE
);
2047 unified
= cg_unified();
2053 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
2054 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
2055 const char *p
= NULL
;
2057 if (!(supported
& bit
))
2061 p
= to_callback(bit
, userdata
);
2066 (void) cg_migrate_recursive_fallback(SYSTEMD_CGROUP_CONTROLLER
, to
, cgroup_controller_to_string(c
), p
, 0);
2072 int cg_trim_everywhere(CGroupMask supported
, const char *path
, bool delete_root
) {
2076 r
= cg_trim(SYSTEMD_CGROUP_CONTROLLER
, path
, delete_root
);
2080 unified
= cg_unified();
2086 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
2087 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
2089 if (!(supported
& bit
))
2092 (void) cg_trim(cgroup_controller_to_string(c
), path
, delete_root
);
2098 int cg_mask_supported(CGroupMask
*ret
) {
2099 CGroupMask mask
= 0;
2102 /* Determines the mask of supported cgroup controllers. Only
2103 * includes controllers we can make sense of and that are
2104 * actually accessible. */
2106 unified
= cg_unified();
2110 _cleanup_free_
char *root
= NULL
, *controllers
= NULL
, *path
= NULL
;
2113 /* In the unified hierarchy we can read the supported
2114 * and accessible controllers from a the top-level
2115 * cgroup attribute */
2117 r
= cg_get_root_path(&root
);
2121 r
= cg_get_path(SYSTEMD_CGROUP_CONTROLLER
, root
, "cgroup.controllers", &path
);
2125 r
= read_one_line_file(path
, &controllers
);
2131 _cleanup_free_
char *n
= NULL
;
2134 r
= extract_first_word(&c
, &n
, NULL
, 0);
2140 v
= cgroup_controller_from_string(n
);
2144 mask
|= CGROUP_CONTROLLER_TO_MASK(v
);
2147 /* Currently, we support the cpu, memory, io and pids
2148 * controller in the unified hierarchy, mask
2149 * everything else off. */
2150 mask
&= CGROUP_MASK_CPU
| CGROUP_MASK_MEMORY
| CGROUP_MASK_IO
| CGROUP_MASK_PIDS
;
2155 /* In the legacy hierarchy, we check whether which
2156 * hierarchies are mounted. */
2158 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
2161 n
= cgroup_controller_to_string(c
);
2162 if (controller_is_accessible(n
) >= 0)
2163 mask
|= CGROUP_CONTROLLER_TO_MASK(c
);
2171 int cg_kernel_controllers(Set
*controllers
) {
2172 _cleanup_fclose_
FILE *f
= NULL
;
2176 assert(controllers
);
2178 /* Determines the full list of kernel-known controllers. Might
2179 * include controllers we don't actually support, arbitrary
2180 * named hierarchies and controllers that aren't currently
2181 * accessible (because not mounted). */
2183 f
= fopen("/proc/cgroups", "re");
2185 if (errno
== ENOENT
)
2190 /* Ignore the header line */
2191 (void) fgets(buf
, sizeof(buf
), f
);
2198 if (fscanf(f
, "%ms %*i %*i %i", &controller
, &enabled
) != 2) {
2203 if (ferror(f
) && errno
> 0)
2214 if (!cg_controller_is_valid(controller
)) {
2219 r
= set_consume(controllers
, controller
);
2227 static thread_local
int unified_cache
= -1;
2229 int cg_unified(void) {
2232 /* Checks if we support the unified hierarchy. Returns an
2233 * error when the cgroup hierarchies aren't mounted yet or we
2234 * have any other trouble determining if the unified hierarchy
2237 if (unified_cache
>= 0)
2238 return unified_cache
;
2240 if (statfs("/sys/fs/cgroup/", &fs
) < 0)
2243 if (F_TYPE_EQUAL(fs
.f_type
, CGROUP2_SUPER_MAGIC
))
2244 unified_cache
= true;
2245 else if (F_TYPE_EQUAL(fs
.f_type
, TMPFS_MAGIC
))
2246 unified_cache
= false;
2250 return unified_cache
;
2253 void cg_unified_flush(void) {
2257 int cg_enable_everywhere(CGroupMask supported
, CGroupMask mask
, const char *p
) {
2258 _cleanup_free_
char *fs
= NULL
;
2267 unified
= cg_unified();
2270 if (!unified
) /* on the legacy hiearchy there's no joining of controllers defined */
2273 r
= cg_get_path(SYSTEMD_CGROUP_CONTROLLER
, p
, "cgroup.subtree_control", &fs
);
2277 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
2278 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
2281 if (!(supported
& bit
))
2284 n
= cgroup_controller_to_string(c
);
2286 char s
[1 + strlen(n
) + 1];
2288 s
[0] = mask
& bit
? '+' : '-';
2291 r
= write_string_file(fs
, s
, 0);
2293 log_debug_errno(r
, "Failed to enable controller %s for %s (%s): %m", n
, p
, fs
);
2300 bool cg_is_unified_wanted(void) {
2301 static thread_local
int wanted
= -1;
2304 /* If the hierarchy is already mounted, then follow whatever
2305 * was chosen for it. */
2306 unified
= cg_unified();
2310 /* Otherwise, let's see what the kernel command line has to
2311 * say. Since checking that is expensive, let's cache the
2316 r
= get_proc_cmdline_key("systemd.unified_cgroup_hierarchy", NULL
);
2318 return (wanted
= true);
2320 _cleanup_free_
char *value
= NULL
;
2322 r
= get_proc_cmdline_key("systemd.unified_cgroup_hierarchy=", &value
);
2326 return (wanted
= false);
2328 return (wanted
= parse_boolean(value
) > 0);
2332 bool cg_is_legacy_wanted(void) {
2333 return !cg_is_unified_wanted();
2336 int cg_weight_parse(const char *s
, uint64_t *ret
) {
2341 *ret
= CGROUP_WEIGHT_INVALID
;
2345 r
= safe_atou64(s
, &u
);
2349 if (u
< CGROUP_WEIGHT_MIN
|| u
> CGROUP_WEIGHT_MAX
)
2356 const uint64_t cgroup_io_limit_defaults
[_CGROUP_IO_LIMIT_TYPE_MAX
] = {
2357 [CGROUP_IO_RBPS_MAX
] = CGROUP_LIMIT_MAX
,
2358 [CGROUP_IO_WBPS_MAX
] = CGROUP_LIMIT_MAX
,
2359 [CGROUP_IO_RIOPS_MAX
] = CGROUP_LIMIT_MAX
,
2360 [CGROUP_IO_WIOPS_MAX
] = CGROUP_LIMIT_MAX
,
2363 static const char* const cgroup_io_limit_type_table
[_CGROUP_IO_LIMIT_TYPE_MAX
] = {
2364 [CGROUP_IO_RBPS_MAX
] = "IOReadBandwidthMax",
2365 [CGROUP_IO_WBPS_MAX
] = "IOWriteBandwidthMax",
2366 [CGROUP_IO_RIOPS_MAX
] = "IOReadIOPSMax",
2367 [CGROUP_IO_WIOPS_MAX
] = "IOWriteIOPSMax",
2370 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type
, CGroupIOLimitType
);
2372 int cg_cpu_shares_parse(const char *s
, uint64_t *ret
) {
2377 *ret
= CGROUP_CPU_SHARES_INVALID
;
2381 r
= safe_atou64(s
, &u
);
2385 if (u
< CGROUP_CPU_SHARES_MIN
|| u
> CGROUP_CPU_SHARES_MAX
)
2392 int cg_blkio_weight_parse(const char *s
, uint64_t *ret
) {
2397 *ret
= CGROUP_BLKIO_WEIGHT_INVALID
;
2401 r
= safe_atou64(s
, &u
);
2405 if (u
< CGROUP_BLKIO_WEIGHT_MIN
|| u
> CGROUP_BLKIO_WEIGHT_MAX
)
2412 static const char *cgroup_controller_table
[_CGROUP_CONTROLLER_MAX
] = {
2413 [CGROUP_CONTROLLER_CPU
] = "cpu",
2414 [CGROUP_CONTROLLER_CPUACCT
] = "cpuacct",
2415 [CGROUP_CONTROLLER_IO
] = "io",
2416 [CGROUP_CONTROLLER_BLKIO
] = "blkio",
2417 [CGROUP_CONTROLLER_MEMORY
] = "memory",
2418 [CGROUP_CONTROLLER_DEVICES
] = "devices",
2419 [CGROUP_CONTROLLER_PIDS
] = "pids",
2422 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller
, CGroupController
);