1 /*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
4 This file is part of systemd.
6 Copyright 2010 Lennart Poettering
8 systemd is free software; you can redistribute it and/or modify it
9 under the terms of the GNU Lesser General Public License as published by
10 the Free Software Foundation; either version 2.1 of the License, or
11 (at your option) any later version.
13 systemd is distributed in the hope that it will be useful, but
14 WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 Lesser General Public License for more details.
18 You should have received a copy of the GNU Lesser General Public License
19 along with systemd; If not, see <http://www.gnu.org/licenses/>.
29 #include <sys/types.h>
35 #include "formats-util.h"
36 #include "process-util.h"
37 #include "path-util.h"
38 #include "unit-name.h"
42 #include "login-util.h"
43 #include "cgroup-util.h"
45 int cg_enumerate_processes(const char *controller
, const char *path
, FILE **_f
) {
46 _cleanup_free_
char *fs
= NULL
;
52 r
= cg_get_path(controller
, path
, "cgroup.procs", &fs
);
64 int cg_read_pid(FILE *f
, pid_t
*_pid
) {
67 /* Note that the cgroup.procs might contain duplicates! See
68 * cgroups.txt for details. */
74 if (fscanf(f
, "%lu", &ul
) != 1) {
79 return errno
? -errno
: -EIO
;
89 int cg_enumerate_subgroups(const char *controller
, const char *path
, DIR **_d
) {
90 _cleanup_free_
char *fs
= NULL
;
96 /* This is not recursive! */
98 r
= cg_get_path(controller
, path
, NULL
, &fs
);
110 int cg_read_subgroup(DIR *d
, char **fn
) {
116 FOREACH_DIRENT_ALL(de
, d
, return -errno
) {
119 if (de
->d_type
!= DT_DIR
)
122 if (streq(de
->d_name
, ".") ||
123 streq(de
->d_name
, ".."))
126 b
= strdup(de
->d_name
);
137 int cg_rmdir(const char *controller
, const char *path
) {
138 _cleanup_free_
char *p
= NULL
;
141 r
= cg_get_path(controller
, path
, NULL
, &p
);
146 if (r
< 0 && errno
!= ENOENT
)
152 int cg_kill(const char *controller
, const char *path
, int sig
, bool sigcont
, bool ignore_self
, Set
*s
) {
153 _cleanup_set_free_ Set
*allocated_set
= NULL
;
160 /* This goes through the tasks list and kills them all. This
161 * is repeated until no further processes are added to the
162 * tasks list, to properly handle forking processes */
165 s
= allocated_set
= set_new(NULL
);
173 _cleanup_fclose_
FILE *f
= NULL
;
177 r
= cg_enumerate_processes(controller
, path
, &f
);
179 if (ret
>= 0 && r
!= -ENOENT
)
185 while ((r
= cg_read_pid(f
, &pid
)) > 0) {
187 if (ignore_self
&& pid
== my_pid
)
190 if (set_get(s
, PID_TO_PTR(pid
)) == PID_TO_PTR(pid
))
193 /* If we haven't killed this process yet, kill
195 if (kill(pid
, sig
) < 0) {
196 if (ret
>= 0 && errno
!= ESRCH
)
199 if (sigcont
&& sig
!= SIGKILL
)
200 (void) kill(pid
, SIGCONT
);
208 r
= set_put(s
, PID_TO_PTR(pid
));
224 /* To avoid racing against processes which fork
225 * quicker than we can kill them we repeat this until
226 * no new pids need to be killed. */
233 int cg_kill_recursive(const char *controller
, const char *path
, int sig
, bool sigcont
, bool ignore_self
, bool rem
, Set
*s
) {
234 _cleanup_set_free_ Set
*allocated_set
= NULL
;
235 _cleanup_closedir_
DIR *d
= NULL
;
243 s
= allocated_set
= set_new(NULL
);
248 ret
= cg_kill(controller
, path
, sig
, sigcont
, ignore_self
, s
);
250 r
= cg_enumerate_subgroups(controller
, path
, &d
);
252 if (ret
>= 0 && r
!= -ENOENT
)
258 while ((r
= cg_read_subgroup(d
, &fn
)) > 0) {
259 _cleanup_free_
char *p
= NULL
;
261 p
= strjoin(path
, "/", fn
, NULL
);
266 r
= cg_kill_recursive(controller
, p
, sig
, sigcont
, ignore_self
, rem
, s
);
267 if (r
!= 0 && ret
>= 0)
271 if (ret
>= 0 && r
< 0)
275 r
= cg_rmdir(controller
, path
);
276 if (r
< 0 && ret
>= 0 && r
!= -ENOENT
&& r
!= -EBUSY
)
283 int cg_migrate(const char *cfrom
, const char *pfrom
, const char *cto
, const char *pto
, bool ignore_self
) {
285 _cleanup_set_free_ Set
*s
= NULL
;
301 _cleanup_fclose_
FILE *f
= NULL
;
305 r
= cg_enumerate_processes(cfrom
, pfrom
, &f
);
307 if (ret
>= 0 && r
!= -ENOENT
)
313 while ((r
= cg_read_pid(f
, &pid
)) > 0) {
315 /* This might do weird stuff if we aren't a
316 * single-threaded program. However, we
317 * luckily know we are not */
318 if (ignore_self
&& pid
== my_pid
)
321 if (set_get(s
, PID_TO_PTR(pid
)) == PID_TO_PTR(pid
))
324 /* Ignore kernel threads. Since they can only
325 * exist in the root cgroup, we only check for
328 (isempty(pfrom
) || path_equal(pfrom
, "/")) &&
329 is_kernel_thread(pid
) > 0)
332 r
= cg_attach(cto
, pto
, pid
);
334 if (ret
>= 0 && r
!= -ESRCH
)
341 r
= set_put(s
, PID_TO_PTR(pid
));
361 int cg_migrate_recursive(
369 _cleanup_closedir_
DIR *d
= NULL
;
378 ret
= cg_migrate(cfrom
, pfrom
, cto
, pto
, ignore_self
);
380 r
= cg_enumerate_subgroups(cfrom
, pfrom
, &d
);
382 if (ret
>= 0 && r
!= -ENOENT
)
388 while ((r
= cg_read_subgroup(d
, &fn
)) > 0) {
389 _cleanup_free_
char *p
= NULL
;
391 p
= strjoin(pfrom
, "/", fn
, NULL
);
396 r
= cg_migrate_recursive(cfrom
, p
, cto
, pto
, ignore_self
, rem
);
397 if (r
!= 0 && ret
>= 0)
401 if (r
< 0 && ret
>= 0)
405 r
= cg_rmdir(cfrom
, pfrom
);
406 if (r
< 0 && ret
>= 0 && r
!= -ENOENT
&& r
!= -EBUSY
)
413 int cg_migrate_recursive_fallback(
428 r
= cg_migrate_recursive(cfrom
, pfrom
, cto
, pto
, ignore_self
, rem
);
430 char prefix
[strlen(pto
) + 1];
432 /* This didn't work? Then let's try all prefixes of the destination */
434 PATH_FOREACH_PREFIX(prefix
, pto
) {
437 q
= cg_migrate_recursive(cfrom
, pfrom
, cto
, prefix
, ignore_self
, rem
);
446 static const char *controller_to_dirname(const char *controller
) {
451 /* Converts a controller name to the directory name below
452 * /sys/fs/cgroup/ we want to mount it to. Effectively, this
453 * just cuts off the name= prefixed used for named
454 * hierarchies, if it is specified. */
456 e
= startswith(controller
, "name=");
463 static int join_path_legacy(const char *controller
, const char *path
, const char *suffix
, char **fs
) {
470 dn
= controller_to_dirname(controller
);
472 if (isempty(path
) && isempty(suffix
))
473 t
= strappend("/sys/fs/cgroup/", dn
);
474 else if (isempty(path
))
475 t
= strjoin("/sys/fs/cgroup/", dn
, "/", suffix
, NULL
);
476 else if (isempty(suffix
))
477 t
= strjoin("/sys/fs/cgroup/", dn
, "/", path
, NULL
);
479 t
= strjoin("/sys/fs/cgroup/", dn
, "/", path
, "/", suffix
, NULL
);
487 static int join_path_unified(const char *path
, const char *suffix
, char **fs
) {
492 if (isempty(path
) && isempty(suffix
))
493 t
= strdup("/sys/fs/cgroup");
494 else if (isempty(path
))
495 t
= strappend("/sys/fs/cgroup/", suffix
);
496 else if (isempty(suffix
))
497 t
= strappend("/sys/fs/cgroup/", path
);
499 t
= strjoin("/sys/fs/cgroup/", path
, "/", suffix
, NULL
);
507 int cg_get_path(const char *controller
, const char *path
, const char *suffix
, char **fs
) {
515 /* If no controller is specified, we return the path
516 * *below* the controllers, without any prefix. */
518 if (!path
&& !suffix
)
523 else if (isempty(path
))
526 t
= strjoin(path
, "/", suffix
, NULL
);
530 *fs
= path_kill_slashes(t
);
534 if (!cg_controller_is_valid(controller
))
537 unified
= cg_unified();
542 r
= join_path_unified(path
, suffix
, fs
);
544 r
= join_path_legacy(controller
, path
, suffix
, fs
);
548 path_kill_slashes(*fs
);
552 static int controller_is_accessible(const char *controller
) {
557 /* Checks whether a specific controller is accessible,
558 * i.e. its hierarchy mounted. In the unified hierarchy all
559 * controllers are considered accessible, except for the named
562 if (!cg_controller_is_valid(controller
))
565 unified
= cg_unified();
569 /* We don't support named hierarchies if we are using
570 * the unified hierarchy. */
572 if (streq(controller
, SYSTEMD_CGROUP_CONTROLLER
))
575 if (startswith(controller
, "name="))
581 dn
= controller_to_dirname(controller
);
582 cc
= strjoina("/sys/fs/cgroup/", dn
);
584 if (laccess(cc
, F_OK
) < 0)
591 int cg_get_path_and_check(const char *controller
, const char *path
, const char *suffix
, char **fs
) {
597 /* Check if the specified controller is actually accessible */
598 r
= controller_is_accessible(controller
);
602 return cg_get_path(controller
, path
, suffix
, fs
);
605 static int trim_cb(const char *path
, const struct stat
*sb
, int typeflag
, struct FTW
*ftwbuf
) {
610 if (typeflag
!= FTW_DP
)
613 if (ftwbuf
->level
< 1)
620 int cg_trim(const char *controller
, const char *path
, bool delete_root
) {
621 _cleanup_free_
char *fs
= NULL
;
626 r
= cg_get_path(controller
, path
, NULL
, &fs
);
631 if (nftw(fs
, trim_cb
, 64, FTW_DEPTH
|FTW_MOUNT
|FTW_PHYS
) != 0) {
641 if (rmdir(fs
) < 0 && errno
!= ENOENT
)
648 int cg_create(const char *controller
, const char *path
) {
649 _cleanup_free_
char *fs
= NULL
;
652 r
= cg_get_path_and_check(controller
, path
, NULL
, &fs
);
656 r
= mkdir_parents(fs
, 0755);
660 if (mkdir(fs
, 0755) < 0) {
671 int cg_create_and_attach(const char *controller
, const char *path
, pid_t pid
) {
676 r
= cg_create(controller
, path
);
680 q
= cg_attach(controller
, path
, pid
);
684 /* This does not remove the cgroup on failure */
688 int cg_attach(const char *controller
, const char *path
, pid_t pid
) {
689 _cleanup_free_
char *fs
= NULL
;
690 char c
[DECIMAL_STR_MAX(pid_t
) + 2];
696 r
= cg_get_path_and_check(controller
, path
, "cgroup.procs", &fs
);
703 snprintf(c
, sizeof(c
), PID_FMT
"\n", pid
);
705 return write_string_file(fs
, c
, 0);
708 int cg_attach_fallback(const char *controller
, const char *path
, pid_t pid
) {
715 r
= cg_attach(controller
, path
, pid
);
717 char prefix
[strlen(path
) + 1];
719 /* This didn't work? Then let's try all prefixes of
722 PATH_FOREACH_PREFIX(prefix
, path
) {
725 q
= cg_attach(controller
, prefix
, pid
);
734 int cg_set_group_access(
735 const char *controller
,
741 _cleanup_free_
char *fs
= NULL
;
744 if (mode
== MODE_INVALID
&& uid
== UID_INVALID
&& gid
== GID_INVALID
)
747 if (mode
!= MODE_INVALID
)
750 r
= cg_get_path(controller
, path
, NULL
, &fs
);
754 return chmod_and_chown(fs
, mode
, uid
, gid
);
757 int cg_set_task_access(
758 const char *controller
,
764 _cleanup_free_
char *fs
= NULL
, *procs
= NULL
;
769 if (mode
== MODE_INVALID
&& uid
== UID_INVALID
&& gid
== GID_INVALID
)
772 if (mode
!= MODE_INVALID
)
775 r
= cg_get_path(controller
, path
, "cgroup.procs", &fs
);
779 r
= chmod_and_chown(fs
, mode
, uid
, gid
);
783 unified
= cg_unified();
789 /* Compatibility, Always keep values for "tasks" in sync with
791 if (cg_get_path(controller
, path
, "tasks", &procs
) >= 0)
792 (void) chmod_and_chown(procs
, mode
, uid
, gid
);
797 int cg_pid_get_path(const char *controller
, pid_t pid
, char **path
) {
798 _cleanup_fclose_
FILE *f
= NULL
;
807 unified
= cg_unified();
812 if (!cg_controller_is_valid(controller
))
815 controller
= SYSTEMD_CGROUP_CONTROLLER
;
817 cs
= strlen(controller
);
820 fs
= procfs_file_alloca(pid
, "cgroup");
823 return errno
== ENOENT
? -ESRCH
: -errno
;
825 FOREACH_LINE(line
, f
, return -errno
) {
831 e
= startswith(line
, "0:");
841 const char *word
, *state
;
844 l
= strchr(line
, ':');
854 FOREACH_WORD_SEPARATOR(word
, k
, l
, ",", state
) {
855 if (k
== cs
&& memcmp(word
, controller
, cs
) == 0) {
876 int cg_install_release_agent(const char *controller
, const char *agent
) {
877 _cleanup_free_
char *fs
= NULL
, *contents
= NULL
;
883 unified
= cg_unified();
886 if (unified
) /* doesn't apply to unified hierarchy */
889 r
= cg_get_path(controller
, NULL
, "release_agent", &fs
);
893 r
= read_one_line_file(fs
, &contents
);
897 sc
= strstrip(contents
);
899 r
= write_string_file(fs
, agent
, 0);
902 } else if (!path_equal(sc
, agent
))
906 r
= cg_get_path(controller
, NULL
, "notify_on_release", &fs
);
910 contents
= mfree(contents
);
911 r
= read_one_line_file(fs
, &contents
);
915 sc
= strstrip(contents
);
916 if (streq(sc
, "0")) {
917 r
= write_string_file(fs
, "1", 0);
930 int cg_uninstall_release_agent(const char *controller
) {
931 _cleanup_free_
char *fs
= NULL
;
934 unified
= cg_unified();
937 if (unified
) /* Doesn't apply to unified hierarchy */
940 r
= cg_get_path(controller
, NULL
, "notify_on_release", &fs
);
944 r
= write_string_file(fs
, "0", 0);
950 r
= cg_get_path(controller
, NULL
, "release_agent", &fs
);
954 r
= write_string_file(fs
, "", 0);
961 int cg_is_empty(const char *controller
, const char *path
) {
962 _cleanup_fclose_
FILE *f
= NULL
;
968 r
= cg_enumerate_processes(controller
, path
, &f
);
974 r
= cg_read_pid(f
, &pid
);
981 int cg_is_empty_recursive(const char *controller
, const char *path
) {
986 /* The root cgroup is always populated */
987 if (controller
&& (isempty(path
) || path_equal(path
, "/")))
990 unified
= cg_unified();
995 _cleanup_free_
char *populated
= NULL
, *t
= NULL
;
997 /* On the unified hierarchy we can check empty state
998 * via the "cgroup.populated" attribute. */
1000 r
= cg_get_path(controller
, path
, "cgroup.populated", &populated
);
1004 r
= read_one_line_file(populated
, &t
);
1008 return streq(t
, "0");
1010 _cleanup_closedir_
DIR *d
= NULL
;
1013 r
= cg_is_empty(controller
, path
);
1017 r
= cg_enumerate_subgroups(controller
, path
, &d
);
1023 while ((r
= cg_read_subgroup(d
, &fn
)) > 0) {
1024 _cleanup_free_
char *p
= NULL
;
1026 p
= strjoin(path
, "/", fn
, NULL
);
1031 r
= cg_is_empty_recursive(controller
, p
);
1042 int cg_split_spec(const char *spec
, char **controller
, char **path
) {
1043 char *t
= NULL
, *u
= NULL
;
1049 if (!path_is_safe(spec
))
1057 *path
= path_kill_slashes(t
);
1066 e
= strchr(spec
, ':');
1068 if (!cg_controller_is_valid(spec
))
1085 t
= strndup(spec
, e
-spec
);
1088 if (!cg_controller_is_valid(t
)) {
1102 if (!path_is_safe(u
) ||
1103 !path_is_absolute(u
)) {
1109 path_kill_slashes(u
);
1125 int cg_mangle_path(const char *path
, char **result
) {
1126 _cleanup_free_
char *c
= NULL
, *p
= NULL
;
1133 /* First, check if it already is a filesystem path */
1134 if (path_startswith(path
, "/sys/fs/cgroup")) {
1140 *result
= path_kill_slashes(t
);
1144 /* Otherwise, treat it as cg spec */
1145 r
= cg_split_spec(path
, &c
, &p
);
1149 return cg_get_path(c
?: SYSTEMD_CGROUP_CONTROLLER
, p
?: "/", NULL
, result
);
1152 int cg_get_root_path(char **path
) {
1158 r
= cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER
, 1, &p
);
1162 e
= endswith(p
, "/" SPECIAL_INIT_SCOPE
);
1164 e
= endswith(p
, "/" SPECIAL_SYSTEM_SLICE
); /* legacy */
1166 e
= endswith(p
, "/system"); /* even more legacy */
1174 int cg_shift_path(const char *cgroup
, const char *root
, const char **shifted
) {
1175 _cleanup_free_
char *rt
= NULL
;
1183 /* If the root was specified let's use that, otherwise
1184 * let's determine it from PID 1 */
1186 r
= cg_get_root_path(&rt
);
1193 p
= path_startswith(cgroup
, root
);
1194 if (p
&& p
> cgroup
)
1202 int cg_pid_get_path_shifted(pid_t pid
, const char *root
, char **cgroup
) {
1203 _cleanup_free_
char *raw
= NULL
;
1210 r
= cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER
, pid
, &raw
);
1214 r
= cg_shift_path(raw
, root
, &c
);
1234 int cg_path_decode_unit(const char *cgroup
, char **unit
){
1241 n
= strcspn(cgroup
, "/");
1245 c
= strndupa(cgroup
, n
);
1248 if (!unit_name_is_valid(c
, UNIT_NAME_PLAIN
|UNIT_NAME_INSTANCE
))
1259 static bool valid_slice_name(const char *p
, size_t n
) {
1264 if (n
< strlen("x.slice"))
1267 if (memcmp(p
+ n
- 6, ".slice", 6) == 0) {
1273 c
= cg_unescape(buf
);
1275 return unit_name_is_valid(c
, UNIT_NAME_PLAIN
);
1281 static const char *skip_slices(const char *p
) {
1284 /* Skips over all slice assignments */
1289 p
+= strspn(p
, "/");
1291 n
= strcspn(p
, "/");
1292 if (!valid_slice_name(p
, n
))
1299 int cg_path_get_unit(const char *path
, char **ret
) {
1307 e
= skip_slices(path
);
1309 r
= cg_path_decode_unit(e
, &unit
);
1313 /* We skipped over the slices, don't accept any now */
1314 if (endswith(unit
, ".slice")) {
1323 int cg_pid_get_unit(pid_t pid
, char **unit
) {
1324 _cleanup_free_
char *cgroup
= NULL
;
1329 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1333 return cg_path_get_unit(cgroup
, unit
);
1337 * Skip session-*.scope, but require it to be there.
1339 static const char *skip_session(const char *p
) {
1345 p
+= strspn(p
, "/");
1347 n
= strcspn(p
, "/");
1348 if (n
< strlen("session-x.scope"))
1351 if (memcmp(p
, "session-", 8) == 0 && memcmp(p
+ n
- 6, ".scope", 6) == 0) {
1352 char buf
[n
- 8 - 6 + 1];
1354 memcpy(buf
, p
+ 8, n
- 8 - 6);
1357 /* Note that session scopes never need unescaping,
1358 * since they cannot conflict with the kernel's own
1359 * names, hence we don't need to call cg_unescape()
1362 if (!session_id_valid(buf
))
1366 p
+= strspn(p
, "/");
1374 * Skip user@*.service, but require it to be there.
1376 static const char *skip_user_manager(const char *p
) {
1382 p
+= strspn(p
, "/");
1384 n
= strcspn(p
, "/");
1385 if (n
< strlen("user@x.service"))
1388 if (memcmp(p
, "user@", 5) == 0 && memcmp(p
+ n
- 8, ".service", 8) == 0) {
1389 char buf
[n
- 5 - 8 + 1];
1391 memcpy(buf
, p
+ 5, n
- 5 - 8);
1394 /* Note that user manager services never need unescaping,
1395 * since they cannot conflict with the kernel's own
1396 * names, hence we don't need to call cg_unescape()
1399 if (parse_uid(buf
, NULL
) < 0)
1403 p
+= strspn(p
, "/");
1411 static const char *skip_user_prefix(const char *path
) {
1416 /* Skip slices, if there are any */
1417 e
= skip_slices(path
);
1419 /* Skip the user manager, if it's in the path now... */
1420 t
= skip_user_manager(e
);
1424 /* Alternatively skip the user session if it is in the path... */
1425 return skip_session(e
);
1428 int cg_path_get_user_unit(const char *path
, char **ret
) {
1434 t
= skip_user_prefix(path
);
1438 /* And from here on it looks pretty much the same as for a
1439 * system unit, hence let's use the same parser from here
1441 return cg_path_get_unit(t
, ret
);
1444 int cg_pid_get_user_unit(pid_t pid
, char **unit
) {
1445 _cleanup_free_
char *cgroup
= NULL
;
1450 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1454 return cg_path_get_user_unit(cgroup
, unit
);
1457 int cg_path_get_machine_name(const char *path
, char **machine
) {
1458 _cleanup_free_
char *u
= NULL
;
1462 r
= cg_path_get_unit(path
, &u
);
1466 sl
= strjoina("/run/systemd/machines/unit:", u
);
1467 return readlink_malloc(sl
, machine
);
1470 int cg_pid_get_machine_name(pid_t pid
, char **machine
) {
1471 _cleanup_free_
char *cgroup
= NULL
;
1476 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1480 return cg_path_get_machine_name(cgroup
, machine
);
1483 int cg_path_get_session(const char *path
, char **session
) {
1484 _cleanup_free_
char *unit
= NULL
;
1490 r
= cg_path_get_unit(path
, &unit
);
1494 start
= startswith(unit
, "session-");
1497 end
= endswith(start
, ".scope");
1502 if (!session_id_valid(start
))
1518 int cg_pid_get_session(pid_t pid
, char **session
) {
1519 _cleanup_free_
char *cgroup
= NULL
;
1522 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1526 return cg_path_get_session(cgroup
, session
);
1529 int cg_path_get_owner_uid(const char *path
, uid_t
*uid
) {
1530 _cleanup_free_
char *slice
= NULL
;
1536 r
= cg_path_get_slice(path
, &slice
);
1540 start
= startswith(slice
, "user-");
1543 end
= endswith(start
, ".slice");
1548 if (parse_uid(start
, uid
) < 0)
1554 int cg_pid_get_owner_uid(pid_t pid
, uid_t
*uid
) {
1555 _cleanup_free_
char *cgroup
= NULL
;
1558 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1562 return cg_path_get_owner_uid(cgroup
, uid
);
1565 int cg_path_get_slice(const char *p
, char **slice
) {
1566 const char *e
= NULL
;
1571 /* Finds the right-most slice unit from the beginning, but
1572 * stops before we come to the first non-slice unit. */
1577 p
+= strspn(p
, "/");
1579 n
= strcspn(p
, "/");
1580 if (!valid_slice_name(p
, n
)) {
1585 s
= strdup("-.slice");
1593 return cg_path_decode_unit(e
, slice
);
1601 int cg_pid_get_slice(pid_t pid
, char **slice
) {
1602 _cleanup_free_
char *cgroup
= NULL
;
1607 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1611 return cg_path_get_slice(cgroup
, slice
);
1614 int cg_path_get_user_slice(const char *p
, char **slice
) {
1619 t
= skip_user_prefix(p
);
1623 /* And now it looks pretty much the same as for a system
1624 * slice, so let's just use the same parser from here on. */
1625 return cg_path_get_slice(t
, slice
);
1628 int cg_pid_get_user_slice(pid_t pid
, char **slice
) {
1629 _cleanup_free_
char *cgroup
= NULL
;
1634 r
= cg_pid_get_path_shifted(pid
, NULL
, &cgroup
);
1638 return cg_path_get_user_slice(cgroup
, slice
);
1641 char *cg_escape(const char *p
) {
1642 bool need_prefix
= false;
1644 /* This implements very minimal escaping for names to be used
1645 * as file names in the cgroup tree: any name which might
1646 * conflict with a kernel name or is prefixed with '_' is
1647 * prefixed with a '_'. That way, when reading cgroup names it
1648 * is sufficient to remove a single prefixing underscore if
1651 /* The return value of this function (unlike cg_unescape())
1657 streq(p
, "notify_on_release") ||
1658 streq(p
, "release_agent") ||
1659 streq(p
, "tasks") ||
1660 startswith(p
, "cgroup."))
1665 dot
= strrchr(p
, '.');
1670 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
1673 n
= cgroup_controller_to_string(c
);
1678 if (memcmp(p
, n
, l
) != 0)
1688 return strappend("_", p
);
1693 char *cg_unescape(const char *p
) {
1696 /* The return value of this function (unlike cg_escape())
1697 * doesn't need free()! */
1705 #define CONTROLLER_VALID \
1709 bool cg_controller_is_valid(const char *p
) {
1715 s
= startswith(p
, "name=");
1719 if (*p
== 0 || *p
== '_')
1722 for (t
= p
; *t
; t
++)
1723 if (!strchr(CONTROLLER_VALID
, *t
))
1726 if (t
- p
> FILENAME_MAX
)
1732 int cg_slice_to_path(const char *unit
, char **ret
) {
1733 _cleanup_free_
char *p
= NULL
, *s
= NULL
, *e
= NULL
;
1740 if (streq(unit
, "-.slice")) {
1750 if (!unit_name_is_valid(unit
, UNIT_NAME_PLAIN
))
1753 if (!endswith(unit
, ".slice"))
1756 r
= unit_name_to_prefix(unit
, &p
);
1760 dash
= strchr(p
, '-');
1762 /* Don't allow initial dashes */
1767 _cleanup_free_
char *escaped
= NULL
;
1768 char n
[dash
- p
+ sizeof(".slice")];
1770 /* Don't allow trailing or double dashes */
1771 if (dash
[1] == 0 || dash
[1] == '-')
1774 strcpy(stpncpy(n
, p
, dash
- p
), ".slice");
1775 if (!unit_name_is_valid(n
, UNIT_NAME_PLAIN
))
1778 escaped
= cg_escape(n
);
1782 if (!strextend(&s
, escaped
, "/", NULL
))
1785 dash
= strchr(dash
+1, '-');
1788 e
= cg_escape(unit
);
1792 if (!strextend(&s
, e
, NULL
))
1801 int cg_set_attribute(const char *controller
, const char *path
, const char *attribute
, const char *value
) {
1802 _cleanup_free_
char *p
= NULL
;
1805 r
= cg_get_path(controller
, path
, attribute
, &p
);
1809 return write_string_file(p
, value
, 0);
1812 int cg_get_attribute(const char *controller
, const char *path
, const char *attribute
, char **ret
) {
1813 _cleanup_free_
char *p
= NULL
;
1816 r
= cg_get_path(controller
, path
, attribute
, &p
);
1820 return read_one_line_file(p
, ret
);
1823 int cg_create_everywhere(CGroupMask supported
, CGroupMask mask
, const char *path
) {
1827 /* This one will create a cgroup in our private tree, but also
1828 * duplicate it in the trees specified in mask, and remove it
1831 /* First create the cgroup in our own hierarchy. */
1832 r
= cg_create(SYSTEMD_CGROUP_CONTROLLER
, path
);
1836 /* If we are in the unified hierarchy, we are done now */
1837 unified
= cg_unified();
1843 /* Otherwise, do the same in the other hierarchies */
1844 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
1845 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
1848 n
= cgroup_controller_to_string(c
);
1851 (void) cg_create(n
, path
);
1852 else if (supported
& bit
)
1853 (void) cg_trim(n
, path
, true);
1859 int cg_attach_everywhere(CGroupMask supported
, const char *path
, pid_t pid
, cg_migrate_callback_t path_callback
, void *userdata
) {
1863 r
= cg_attach(SYSTEMD_CGROUP_CONTROLLER
, path
, pid
);
1867 unified
= cg_unified();
1873 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
1874 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
1875 const char *p
= NULL
;
1877 if (!(supported
& bit
))
1881 p
= path_callback(bit
, userdata
);
1886 (void) cg_attach_fallback(cgroup_controller_to_string(c
), p
, pid
);
1892 int cg_attach_many_everywhere(CGroupMask supported
, const char *path
, Set
* pids
, cg_migrate_callback_t path_callback
, void *userdata
) {
1897 SET_FOREACH(pidp
, pids
, i
) {
1898 pid_t pid
= PTR_TO_PID(pidp
);
1901 q
= cg_attach_everywhere(supported
, path
, pid
, path_callback
, userdata
);
1902 if (q
< 0 && r
>= 0)
1909 int cg_migrate_everywhere(CGroupMask supported
, const char *from
, const char *to
, cg_migrate_callback_t to_callback
, void *userdata
) {
1913 if (!path_equal(from
, to
)) {
1914 r
= cg_migrate_recursive(SYSTEMD_CGROUP_CONTROLLER
, from
, SYSTEMD_CGROUP_CONTROLLER
, to
, false, true);
1919 unified
= cg_unified();
1925 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
1926 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
1927 const char *p
= NULL
;
1929 if (!(supported
& bit
))
1933 p
= to_callback(bit
, userdata
);
1938 (void) cg_migrate_recursive_fallback(SYSTEMD_CGROUP_CONTROLLER
, to
, cgroup_controller_to_string(c
), p
, false, false);
1944 int cg_trim_everywhere(CGroupMask supported
, const char *path
, bool delete_root
) {
1948 r
= cg_trim(SYSTEMD_CGROUP_CONTROLLER
, path
, delete_root
);
1952 unified
= cg_unified();
1958 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
1959 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
1961 if (!(supported
& bit
))
1964 (void) cg_trim(cgroup_controller_to_string(c
), path
, delete_root
);
1970 int cg_mask_supported(CGroupMask
*ret
) {
1971 CGroupMask mask
= 0;
1974 /* Determines the mask of supported cgroup controllers. Only
1975 * includes controllers we can make sense of and that are
1976 * actually accessible. */
1978 unified
= cg_unified();
1982 _cleanup_free_
char *controllers
= NULL
;
1985 /* In the unified hierarchy we can read the supported
1986 * and accessible controllers from a the top-level
1987 * cgroup attribute */
1989 r
= read_one_line_file("/sys/fs/cgroup/cgroup.controllers", &controllers
);
1995 _cleanup_free_
char *n
= NULL
;
1998 r
= extract_first_word(&c
, &n
, NULL
, 0);
2004 v
= cgroup_controller_from_string(n
);
2008 mask
|= CGROUP_CONTROLLER_TO_MASK(v
);
2011 /* Currently, we only support the memory controller in
2012 * the unified hierarchy, mask everything else off. */
2013 mask
&= CGROUP_MASK_MEMORY
;
2018 /* In the legacy hierarchy, we check whether which
2019 * hierarchies are mounted. */
2021 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
2024 n
= cgroup_controller_to_string(c
);
2025 if (controller_is_accessible(n
) >= 0)
2026 mask
|= CGROUP_CONTROLLER_TO_MASK(c
);
2034 int cg_kernel_controllers(Set
*controllers
) {
2035 _cleanup_fclose_
FILE *f
= NULL
;
2039 assert(controllers
);
2041 /* Determines the full list of kernel-known controllers. Might
2042 * include controllers we don't actually support, arbitrary
2043 * named hierarchies and controllers that aren't currently
2044 * accessible (because not mounted). */
2046 f
= fopen("/proc/cgroups", "re");
2048 if (errno
== ENOENT
)
2053 /* Ignore the header line */
2054 (void) fgets(buf
, sizeof(buf
), f
);
2061 if (fscanf(f
, "%ms %*i %*i %i", &controller
, &enabled
) != 2) {
2066 if (ferror(f
) && errno
!= 0)
2077 if (!cg_controller_is_valid(controller
)) {
2082 r
= set_consume(controllers
, controller
);
2090 static thread_local
int unified_cache
= -1;
2092 int cg_unified(void) {
2095 /* Checks if we support the unified hierarchy. Returns an
2096 * error when the cgroup hierarchies aren't mounted yet or we
2097 * have any other trouble determining if the unified hierarchy
2100 if (unified_cache
>= 0)
2101 return unified_cache
;
2103 if (statfs("/sys/fs/cgroup/", &fs
) < 0)
2106 if (F_TYPE_EQUAL(fs
.f_type
, CGROUP_SUPER_MAGIC
))
2107 unified_cache
= true;
2108 else if (F_TYPE_EQUAL(fs
.f_type
, TMPFS_MAGIC
))
2109 unified_cache
= false;
2113 return unified_cache
;
2116 void cg_unified_flush(void) {
2120 int cg_enable_everywhere(CGroupMask supported
, CGroupMask mask
, const char *p
) {
2121 _cleanup_free_
char *fs
= NULL
;
2130 unified
= cg_unified();
2133 if (!unified
) /* on the legacy hiearchy there's no joining of controllers defined */
2136 r
= cg_get_path(SYSTEMD_CGROUP_CONTROLLER
, p
, "cgroup.subtree_control", &fs
);
2140 for (c
= 0; c
< _CGROUP_CONTROLLER_MAX
; c
++) {
2141 CGroupMask bit
= CGROUP_CONTROLLER_TO_MASK(c
);
2144 if (!(supported
& bit
))
2147 n
= cgroup_controller_to_string(c
);
2149 char s
[1 + strlen(n
) + 1];
2151 s
[0] = mask
& bit
? '+' : '-';
2154 r
= write_string_file(fs
, s
, 0);
2156 log_warning_errno(r
, "Failed to enable controller %s for %s (%s): %m", n
, p
, fs
);
2163 bool cg_is_unified_wanted(void) {
2164 static thread_local
int wanted
= -1;
2167 /* If the hierarchy is already mounted, then follow whatever
2168 * was chosen for it. */
2169 unified
= cg_unified();
2173 /* Otherwise, let's see what the kernel command line has to
2174 * say. Since checking that is expensive, let's cache the
2179 r
= get_proc_cmdline_key("systemd.unified_cgroup_hierarchy", NULL
);
2181 return (wanted
= true);
2183 _cleanup_free_
char *value
= NULL
;
2185 r
= get_proc_cmdline_key("systemd.unified_cgroup_hierarchy=", &value
);
2189 return (wanted
= false);
2191 return (wanted
= parse_boolean(value
) > 0);
2195 bool cg_is_legacy_wanted(void) {
2196 return !cg_is_unified_wanted();
2199 static const char *cgroup_controller_table
[_CGROUP_CONTROLLER_MAX
] = {
2200 [CGROUP_CONTROLLER_CPU
] = "cpu",
2201 [CGROUP_CONTROLLER_CPUACCT
] = "cpuacct",
2202 [CGROUP_CONTROLLER_BLKIO
] = "blkio",
2203 [CGROUP_CONTROLLER_MEMORY
] = "memory",
2204 [CGROUP_CONTROLLER_DEVICE
] = "device",
2207 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller
, CGroupController
);