2 * lxc: linux Container library
4 * Copyright © 2016 Canonical Ltd.
7 * Serge Hallyn <serge.hallyn@ubuntu.com>
8 * Christian Brauner <christian.brauner@ubuntu.com>
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 * cgfs-ng.c: this is a new, simplified implementation of a filesystem
27 * cgroup backend. The original cgfs.c was designed to be as flexible
28 * as possible. It would try to find cgroup filesystems no matter where
29 * or how you had them mounted, and deduce the most usable mount for
32 * This new implementation assumes that cgroup filesystems are mounted
33 * under /sys/fs/cgroup/clist where clist is either the controller, or
34 * a comman-separated list of controllers.
48 #include <linux/kdev_t.h>
49 #include <linux/types.h>
50 #include <sys/types.h>
54 #include "cgroup_utils.h"
58 #include "storage/storage.h"
62 #include "include/strlcpy.h"
66 #include "include/strlcat.h"
69 lxc_log_define(cgfsng
, cgroup
);
71 static void free_string_list(char **clist
)
78 for (i
= 0; clist
[i
]; i
++)
84 /* Allocate a pointer, do not fail. */
85 static void *must_alloc(size_t sz
)
87 return must_realloc(NULL
, sz
);
90 /* Given a pointer to a null-terminated array of pointers, realloc to add one
91 * entry, and point the new entry to NULL. Do not fail. Return the index to the
92 * second-to-last entry - that is, the one which is now available for use
93 * (keeping the list null-terminated).
95 static int append_null_to_list(void ***list
)
100 for (; (*list
)[newentry
]; newentry
++)
103 *list
= must_realloc(*list
, (newentry
+ 2) * sizeof(void **));
104 (*list
)[newentry
+ 1] = NULL
;
108 /* Given a null-terminated array of strings, check whether @entry is one of the
111 static bool string_in_list(char **list
, const char *entry
)
118 for (i
= 0; list
[i
]; i
++)
119 if (strcmp(list
[i
], entry
) == 0)
125 /* Return a copy of @entry prepending "name=", i.e. turn "systemd" into
126 * "name=systemd". Do not fail.
128 static char *cg_legacy_must_prefix_named(char *entry
)
134 prefixed
= must_alloc(len
+ 6);
136 memcpy(prefixed
, "name=", sizeof("name=") - 1);
137 memcpy(prefixed
+ sizeof("name=") - 1, entry
, len
);
138 prefixed
[len
+ 5] = '\0';
142 /* Append an entry to the clist. Do not fail. @clist must be NULL the first time
145 * We also handle named subsystems here. Any controller which is not a kernel
146 * subsystem, we prefix "name=". Any which is both a kernel and named subsystem,
147 * we refuse to use because we're not sure which we have here.
148 * (TODO: We could work around this in some cases by just remounting to be
149 * unambiguous, or by comparing mountpoint contents with current cgroup.)
151 * The last entry will always be NULL.
153 static void must_append_controller(char **klist
, char **nlist
, char ***clist
,
159 if (string_in_list(klist
, entry
) && string_in_list(nlist
, entry
)) {
160 ERROR("Refusing to use ambiguous controller \"%s\"", entry
);
161 ERROR("It is both a named and kernel subsystem");
165 newentry
= append_null_to_list((void ***)clist
);
167 if (strncmp(entry
, "name=", 5) == 0)
168 copy
= must_copy_string(entry
);
169 else if (string_in_list(klist
, entry
))
170 copy
= must_copy_string(entry
);
172 copy
= cg_legacy_must_prefix_named(entry
);
174 (*clist
)[newentry
] = copy
;
177 /* Given a handler's cgroup data, return the struct hierarchy for the controller
178 * @c, or NULL if there is none.
180 struct hierarchy
*get_hierarchy(struct cgroup_ops
*ops
, const char *c
)
184 if (!ops
->hierarchies
)
187 for (i
= 0; ops
->hierarchies
[i
]; i
++) {
189 /* This is the empty unified hierarchy. */
190 if (ops
->hierarchies
[i
]->controllers
&&
191 !ops
->hierarchies
[i
]->controllers
[0])
192 return ops
->hierarchies
[i
];
197 if (string_in_list(ops
->hierarchies
[i
]->controllers
, c
))
198 return ops
->hierarchies
[i
];
204 #define BATCH_SIZE 50
205 static void batch_realloc(char **mem
, size_t oldlen
, size_t newlen
)
207 int newbatches
= (newlen
/ BATCH_SIZE
) + 1;
208 int oldbatches
= (oldlen
/ BATCH_SIZE
) + 1;
210 if (!*mem
|| newbatches
> oldbatches
) {
211 *mem
= must_realloc(*mem
, newbatches
* BATCH_SIZE
);
215 static void append_line(char **dest
, size_t oldlen
, char *new, size_t newlen
)
217 size_t full
= oldlen
+ newlen
;
219 batch_realloc(dest
, oldlen
, full
+ 1);
221 memcpy(*dest
+ oldlen
, new, newlen
+ 1);
224 /* Slurp in a whole file */
225 static char *read_file(const char *fnam
)
228 char *line
= NULL
, *buf
= NULL
;
229 size_t len
= 0, fulllen
= 0;
232 f
= fopen(fnam
, "r");
235 while ((linelen
= getline(&line
, &len
, f
)) != -1) {
236 append_line(&buf
, fulllen
, line
, linelen
);
244 /* Taken over modified from the kernel sources. */
245 #define NBITS 32 /* bits in uint32_t */
246 #define DIV_ROUND_UP(n, d) (((n) + (d)-1) / (d))
247 #define BITS_TO_LONGS(nr) DIV_ROUND_UP(nr, NBITS)
249 static void set_bit(unsigned bit
, uint32_t *bitarr
)
251 bitarr
[bit
/ NBITS
] |= (1 << (bit
% NBITS
));
254 static void clear_bit(unsigned bit
, uint32_t *bitarr
)
256 bitarr
[bit
/ NBITS
] &= ~(1 << (bit
% NBITS
));
259 static bool is_set(unsigned bit
, uint32_t *bitarr
)
261 return (bitarr
[bit
/ NBITS
] & (1 << (bit
% NBITS
))) != 0;
264 /* Create cpumask from cpulist aka turn:
272 static uint32_t *lxc_cpumask(char *buf
, size_t nbits
)
277 char *saveptr
= NULL
;
279 arrlen
= BITS_TO_LONGS(nbits
);
280 bitarr
= calloc(arrlen
, sizeof(uint32_t));
284 for (; (token
= strtok_r(buf
, ",", &saveptr
)); buf
= NULL
) {
289 start
= strtoul(token
, NULL
, 0);
291 range
= strchr(token
, '-');
293 end
= strtoul(range
+ 1, NULL
, 0);
295 if (!(start
<= end
)) {
306 set_bit(start
++, bitarr
);
312 /* Turn cpumask into simple, comma-separated cpulist. */
313 static char *lxc_cpumask_to_cpulist(uint32_t *bitarr
, size_t nbits
)
317 char **cpulist
= NULL
;
318 char numstr
[LXC_NUMSTRLEN64
] = {0};
320 for (i
= 0; i
<= nbits
; i
++) {
321 if (!is_set(i
, bitarr
))
324 ret
= snprintf(numstr
, LXC_NUMSTRLEN64
, "%zu", i
);
325 if (ret
< 0 || (size_t)ret
>= LXC_NUMSTRLEN64
) {
326 lxc_free_array((void **)cpulist
, free
);
330 ret
= lxc_append_string(&cpulist
, numstr
);
332 lxc_free_array((void **)cpulist
, free
);
340 return lxc_string_join(",", (const char **)cpulist
, false);
343 static ssize_t
get_max_cpus(char *cpulist
)
346 char *maxcpus
= cpulist
;
349 c1
= strrchr(maxcpus
, ',');
353 c2
= strrchr(maxcpus
, '-');
367 cpus
= strtoul(c1
, NULL
, 0);
374 #define __ISOL_CPUS "/sys/devices/system/cpu/isolated"
375 static bool cg_legacy_filter_and_set_cpus(char *path
, bool am_initialized
)
379 char *lastslash
, *fpath
, oldv
;
380 ssize_t maxisol
= 0, maxposs
= 0;
381 char *cpulist
= NULL
, *isolcpus
= NULL
, *posscpus
= NULL
;
382 uint32_t *isolmask
= NULL
, *possmask
= NULL
;
383 bool bret
= false, flipped_bit
= false;
385 lastslash
= strrchr(path
, '/');
387 ERROR("Failed to detect \"/\" in \"%s\"", path
);
392 fpath
= must_make_path(path
, "cpuset.cpus", NULL
);
393 posscpus
= read_file(fpath
);
395 SYSERROR("Failed to read file \"%s\"", fpath
);
399 /* Get maximum number of cpus found in possible cpuset. */
400 maxposs
= get_max_cpus(posscpus
);
404 if (!file_exists(__ISOL_CPUS
)) {
405 /* This system doesn't expose isolated cpus. */
406 DEBUG("The path \""__ISOL_CPUS
"\" to read isolated cpus from does not exist");
408 /* No isolated cpus but we weren't already initialized by
409 * someone. We should simply copy the parents cpuset.cpus
412 if (!am_initialized
) {
413 DEBUG("Copying cpu settings of parent cgroup");
416 /* No isolated cpus but we were already initialized by someone.
417 * Nothing more to do for us.
422 isolcpus
= read_file(__ISOL_CPUS
);
424 SYSERROR("Failed to read file \""__ISOL_CPUS
"\"");
427 if (!isdigit(isolcpus
[0])) {
428 TRACE("No isolated cpus detected");
430 /* No isolated cpus but we weren't already initialized by
431 * someone. We should simply copy the parents cpuset.cpus
434 if (!am_initialized
) {
435 DEBUG("Copying cpu settings of parent cgroup");
438 /* No isolated cpus but we were already initialized by someone.
439 * Nothing more to do for us.
444 /* Get maximum number of cpus found in isolated cpuset. */
445 maxisol
= get_max_cpus(isolcpus
);
449 if (maxposs
< maxisol
)
453 possmask
= lxc_cpumask(posscpus
, maxposs
);
455 ERROR("Failed to create cpumask for possible cpus");
459 isolmask
= lxc_cpumask(isolcpus
, maxposs
);
461 ERROR("Failed to create cpumask for isolated cpus");
465 for (i
= 0; i
<= maxposs
; i
++) {
466 if (!is_set(i
, isolmask
) || !is_set(i
, possmask
))
470 clear_bit(i
, possmask
);
474 DEBUG("No isolated cpus present in cpuset");
477 DEBUG("Removed isolated cpus from cpuset");
479 cpulist
= lxc_cpumask_to_cpulist(possmask
, maxposs
);
481 ERROR("Failed to create cpu list");
488 fpath
= must_make_path(path
, "cpuset.cpus", NULL
);
489 ret
= lxc_write_to_file(fpath
, cpulist
, strlen(cpulist
), false, 0666);
491 SYSERROR("Failed to write cpu list to \"%s\"", fpath
);
504 if (posscpus
!= cpulist
)
512 /* Copy contents of parent(@path)/@file to @path/@file */
513 static bool copy_parent_file(char *path
, char *file
)
516 char *fpath
, *lastslash
, oldv
;
520 lastslash
= strrchr(path
, '/');
522 ERROR("Failed to detect \"/\" in \"%s\"", path
);
527 fpath
= must_make_path(path
, file
, NULL
);
528 len
= lxc_read_from_file(fpath
, NULL
, 0);
532 value
= must_alloc(len
+ 1);
533 ret
= lxc_read_from_file(fpath
, value
, len
);
539 fpath
= must_make_path(path
, file
, NULL
);
540 ret
= lxc_write_to_file(fpath
, value
, len
, false, 0666);
542 SYSERROR("Failed to write \"%s\" to file \"%s\"", value
, fpath
);
548 SYSERROR("Failed to read file \"%s\"", fpath
);
554 /* Initialize the cpuset hierarchy in first directory of @gname and set
555 * cgroup.clone_children so that children inherit settings. Since the
556 * h->base_path is populated by init or ourselves, we know it is already
559 static bool cg_legacy_handle_cpuset_hierarchy(struct hierarchy
*h
, char *cgname
)
563 char *cgpath
, *clonechildrenpath
, *slash
;
565 if (!string_in_list(h
->controllers
, "cpuset"))
570 slash
= strchr(cgname
, '/');
574 cgpath
= must_make_path(h
->mountpoint
, h
->base_cgroup
, cgname
, NULL
);
578 ret
= mkdir(cgpath
, 0755);
580 if (errno
!= EEXIST
) {
581 SYSERROR("Failed to create directory \"%s\"", cgpath
);
588 must_make_path(cgpath
, "cgroup.clone_children", NULL
);
589 /* unified hierarchy doesn't have clone_children */
590 if (!file_exists(clonechildrenpath
)) {
591 free(clonechildrenpath
);
596 ret
= lxc_read_from_file(clonechildrenpath
, &v
, 1);
598 SYSERROR("Failed to read file \"%s\"", clonechildrenpath
);
599 free(clonechildrenpath
);
604 /* Make sure any isolated cpus are removed from cpuset.cpus. */
605 if (!cg_legacy_filter_and_set_cpus(cgpath
, v
== '1')) {
606 SYSERROR("Failed to remove isolated cpus");
607 free(clonechildrenpath
);
612 /* Already set for us by someone else. */
614 DEBUG("\"cgroup.clone_children\" was already set to \"1\"");
615 free(clonechildrenpath
);
620 /* copy parent's settings */
621 if (!copy_parent_file(cgpath
, "cpuset.mems")) {
622 SYSERROR("Failed to copy \"cpuset.mems\" settings");
624 free(clonechildrenpath
);
629 ret
= lxc_write_to_file(clonechildrenpath
, "1", 1, false, 0666);
631 /* Set clone_children so children inherit our settings */
632 SYSERROR("Failed to write 1 to \"%s\"", clonechildrenpath
);
633 free(clonechildrenpath
);
636 free(clonechildrenpath
);
640 /* Given two null-terminated lists of strings, return true if any string is in
643 static bool controller_lists_intersect(char **l1
, char **l2
)
650 for (i
= 0; l1
[i
]; i
++) {
651 if (string_in_list(l2
, l1
[i
]))
658 /* For a null-terminated list of controllers @clist, return true if any of those
659 * controllers is already listed the null-terminated list of hierarchies @hlist.
660 * Realistically, if one is present, all must be present.
662 static bool controller_list_is_dup(struct hierarchy
**hlist
, char **clist
)
669 for (i
= 0; hlist
[i
]; i
++)
670 if (controller_lists_intersect(hlist
[i
]->controllers
, clist
))
676 /* Return true if the controller @entry is found in the null-terminated list of
677 * hierarchies @hlist.
679 static bool controller_found(struct hierarchy
**hlist
, char *entry
)
686 for (i
= 0; hlist
[i
]; i
++)
687 if (string_in_list(hlist
[i
]->controllers
, entry
))
693 /* Return true if all of the controllers which we require have been found. The
694 * required list is freezer and anything in lxc.cgroup.use.
696 static bool all_controllers_found(struct cgroup_ops
*ops
)
699 char *saveptr
= NULL
;
700 struct hierarchy
**hlist
= ops
->hierarchies
;
702 if (!controller_found(hlist
, "freezer")) {
703 ERROR("No freezer controller mountpoint found");
707 if (!ops
->cgroup_use
)
710 for (; (p
= strtok_r(ops
->cgroup_use
, ",", &saveptr
)); ops
->cgroup_use
= NULL
)
711 if (!controller_found(hlist
, p
)) {
712 ERROR("No %s controller mountpoint found", p
);
719 /* Get the controllers from a mountinfo line There are other ways we could get
720 * this info. For lxcfs, field 3 is /cgroup/controller-list. For cgroupfs, we
721 * could parse the mount options. But we simply assume that the mountpoint must
722 * be /sys/fs/cgroup/controller-list
724 static char **cg_hybrid_get_controllers(char **klist
, char **nlist
, char *line
,
727 /* The fourth field is /sys/fs/cgroup/comma-delimited-controller-list
728 * for legacy hierarchies.
731 char *dup
, *p2
, *tok
;
732 char *p
= line
, *saveptr
= NULL
, *sep
= ",";
735 for (i
= 0; i
< 4; i
++) {
742 /* Note, if we change how mountinfo works, then our caller will need to
743 * verify /sys/fs/cgroup/ in this field.
745 if (strncmp(p
, "/sys/fs/cgroup/", 15) != 0) {
746 ERROR("Found hierarchy not under /sys/fs/cgroup: \"%s\"", p
);
753 ERROR("Corrupt mountinfo");
758 if (type
== CGROUP_SUPER_MAGIC
) {
759 /* strdup() here for v1 hierarchies. Otherwise strtok_r() will
760 * destroy mountpoints such as "/sys/fs/cgroup/cpu,cpuacct".
766 for (tok
= strtok_r(dup
, sep
, &saveptr
); tok
;
767 tok
= strtok_r(NULL
, sep
, &saveptr
))
768 must_append_controller(klist
, nlist
, &aret
, tok
);
777 static char **cg_unified_make_empty_controller(void)
782 newentry
= append_null_to_list((void ***)&aret
);
783 aret
[newentry
] = NULL
;
787 static char **cg_unified_get_controllers(const char *file
)
790 char *saveptr
= NULL
, *sep
= " \t\n";
793 buf
= read_file(file
);
797 for (tok
= strtok_r(buf
, sep
, &saveptr
); tok
;
798 tok
= strtok_r(NULL
, sep
, &saveptr
)) {
802 newentry
= append_null_to_list((void ***)&aret
);
803 copy
= must_copy_string(tok
);
804 aret
[newentry
] = copy
;
811 static struct hierarchy
*add_hierarchy(struct hierarchy
***h
, char **clist
, char *mountpoint
,
812 char *base_cgroup
, int type
)
814 struct hierarchy
*new;
817 new = must_alloc(sizeof(*new));
818 new->controllers
= clist
;
819 new->mountpoint
= mountpoint
;
820 new->base_cgroup
= base_cgroup
;
821 new->fullcgpath
= NULL
;
824 newentry
= append_null_to_list((void ***)h
);
825 (*h
)[newentry
] = new;
829 /* Get a copy of the mountpoint from @line, which is a line from
830 * /proc/self/mountinfo.
832 static char *cg_hybrid_get_mountpoint(char *line
)
837 char *p
= line
, *sret
= NULL
;
839 for (i
= 0; i
< 4; i
++) {
846 if (strncmp(p
, "/sys/fs/cgroup/", 15) != 0)
849 p2
= strchr(p
+ 15, ' ');
855 sret
= must_alloc(len
+ 1);
856 memcpy(sret
, p
, len
);
861 /* Given a multi-line string, return a null-terminated copy of the current line. */
862 static char *copy_to_eol(char *p
)
864 char *p2
= strchr(p
, '\n'), *sret
;
871 sret
= must_alloc(len
+ 1);
872 memcpy(sret
, p
, len
);
877 /* cgline: pointer to character after the first ':' in a line in a \n-terminated
878 * /proc/self/cgroup file. Check whether controller c is present.
880 static bool controller_in_clist(char *cgline
, char *c
)
882 char *tok
, *saveptr
= NULL
, *eol
, *tmp
;
885 eol
= strchr(cgline
, ':');
890 tmp
= alloca(len
+ 1);
891 memcpy(tmp
, cgline
, len
);
894 for (tok
= strtok_r(tmp
, ",", &saveptr
); tok
;
895 tok
= strtok_r(NULL
, ",", &saveptr
)) {
896 if (strcmp(tok
, c
) == 0)
903 /* @basecginfo is a copy of /proc/$$/cgroup. Return the current cgroup for
906 static char *cg_hybrid_get_current_cgroup(char *basecginfo
, char *controller
,
909 char *p
= basecginfo
;
912 bool is_cgv2_base_cgroup
= false;
914 /* cgroup v2 entry in "/proc/<pid>/cgroup": "0::/some/path" */
915 if ((type
== CGROUP2_SUPER_MAGIC
) && (*p
== '0'))
916 is_cgv2_base_cgroup
= true;
923 if (is_cgv2_base_cgroup
|| (controller
&& controller_in_clist(p
, controller
))) {
928 return copy_to_eol(p
);
938 static void must_append_string(char ***list
, char *entry
)
943 newentry
= append_null_to_list((void ***)list
);
944 copy
= must_copy_string(entry
);
945 (*list
)[newentry
] = copy
;
948 static int get_existing_subsystems(char ***klist
, char ***nlist
)
954 f
= fopen("/proc/self/cgroup", "r");
958 while (getline(&line
, &len
, f
) != -1) {
959 char *p
, *p2
, *tok
, *saveptr
= NULL
;
960 p
= strchr(line
, ':');
969 /* If the kernel has cgroup v2 support, then /proc/self/cgroup
970 * contains an entry of the form:
974 * In this case we use "cgroup2" as controller name.
977 must_append_string(klist
, "cgroup2");
981 for (tok
= strtok_r(p
, ",", &saveptr
); tok
;
982 tok
= strtok_r(NULL
, ",", &saveptr
)) {
983 if (strncmp(tok
, "name=", 5) == 0)
984 must_append_string(nlist
, tok
);
986 must_append_string(klist
, tok
);
995 static void trim(char *s
)
1000 while ((len
> 1) && (s
[len
- 1] == '\n'))
1004 static void lxc_cgfsng_print_hierarchies(struct cgroup_ops
*ops
)
1007 struct hierarchy
**it
;
1009 if (!ops
->hierarchies
) {
1010 TRACE(" No hierarchies found");
1014 TRACE(" Hierarchies:");
1015 for (i
= 0, it
= ops
->hierarchies
; it
&& *it
; it
++, i
++) {
1019 TRACE(" %d: base_cgroup: %s", i
, (*it
)->base_cgroup
? (*it
)->base_cgroup
: "(null)");
1020 TRACE(" mountpoint: %s", (*it
)->mountpoint
? (*it
)->mountpoint
: "(null)");
1021 TRACE(" controllers:");
1022 for (j
= 0, cit
= (*it
)->controllers
; cit
&& *cit
; cit
++, j
++)
1023 TRACE(" %d: %s", j
, *cit
);
1027 static void lxc_cgfsng_print_basecg_debuginfo(char *basecginfo
, char **klist
,
1033 TRACE("basecginfo is:");
1034 TRACE("%s", basecginfo
);
1036 for (k
= 0, it
= klist
; it
&& *it
; it
++, k
++)
1037 TRACE("kernel subsystem %d: %s", k
, *it
);
1039 for (k
= 0, it
= nlist
; it
&& *it
; it
++, k
++)
1040 TRACE("named subsystem %d: %s", k
, *it
);
1043 static int cgroup_rmdir(struct hierarchy
**hierarchies
,
1044 const char *container_cgroup
)
1048 if (!container_cgroup
|| !hierarchies
)
1051 for (i
= 0; hierarchies
[i
]; i
++) {
1053 struct hierarchy
*h
= hierarchies
[i
];
1058 ret
= recursive_destroy(h
->fullcgpath
);
1060 WARN("Failed to destroy \"%s\"", h
->fullcgpath
);
1062 free(h
->fullcgpath
);
1063 h
->fullcgpath
= NULL
;
1069 struct generic_userns_exec_data
{
1070 struct hierarchy
**hierarchies
;
1071 const char *container_cgroup
;
1072 struct lxc_conf
*conf
;
1073 uid_t origuid
; /* target uid in parent namespace */
1077 static int cgroup_rmdir_wrapper(void *data
)
1080 struct generic_userns_exec_data
*arg
= data
;
1081 uid_t nsuid
= (arg
->conf
->root_nsuid_map
!= NULL
) ? 0 : arg
->conf
->init_uid
;
1082 gid_t nsgid
= (arg
->conf
->root_nsgid_map
!= NULL
) ? 0 : arg
->conf
->init_gid
;
1084 ret
= setresgid(nsgid
, nsgid
, nsgid
);
1086 SYSERROR("Failed to setresgid(%d, %d, %d)", (int)nsgid
,
1087 (int)nsgid
, (int)nsgid
);
1091 ret
= setresuid(nsuid
, nsuid
, nsuid
);
1093 SYSERROR("Failed to setresuid(%d, %d, %d)", (int)nsuid
,
1094 (int)nsuid
, (int)nsuid
);
1098 ret
= setgroups(0, NULL
);
1099 if (ret
< 0 && errno
!= EPERM
) {
1100 SYSERROR("Failed to setgroups(0, NULL)");
1104 return cgroup_rmdir(arg
->hierarchies
, arg
->container_cgroup
);
1107 static void cgfsng_destroy(struct cgroup_ops
*ops
, struct lxc_handler
*handler
)
1110 struct generic_userns_exec_data wrap
;
1113 wrap
.container_cgroup
= ops
->container_cgroup
;
1114 wrap
.hierarchies
= ops
->hierarchies
;
1115 wrap
.conf
= handler
->conf
;
1117 if (handler
->conf
&& !lxc_list_empty(&handler
->conf
->id_map
))
1118 ret
= userns_exec_1(handler
->conf
, cgroup_rmdir_wrapper
, &wrap
,
1119 "cgroup_rmdir_wrapper");
1121 ret
= cgroup_rmdir(ops
->hierarchies
, ops
->container_cgroup
);
1123 WARN("Failed to destroy cgroups");
1128 static bool cg_unified_create_cgroup(struct hierarchy
*h
, char *cgname
)
1130 size_t i
, parts_len
;
1132 size_t full_len
= 0;
1133 char *add_controllers
= NULL
, *cgroup
= NULL
;
1134 char **parts
= NULL
;
1137 if (h
->version
!= CGROUP2_SUPER_MAGIC
)
1140 if (!h
->controllers
)
1143 /* For now we simply enable all controllers that we have detected by
1144 * creating a string like "+memory +pids +cpu +io".
1145 * TODO: In the near future we might want to support "-<controller>"
1146 * etc. but whether supporting semantics like this make sense will need
1149 for (it
= h
->controllers
; it
&& *it
; it
++) {
1150 full_len
+= strlen(*it
) + 2;
1151 add_controllers
= must_realloc(add_controllers
, full_len
+ 1);
1153 if (h
->controllers
[0] == *it
)
1154 add_controllers
[0] = '\0';
1156 (void)strlcat(add_controllers
, "+", full_len
+ 1);
1157 (void)strlcat(add_controllers
, *it
, full_len
+ 1);
1159 if ((it
+ 1) && *(it
+ 1))
1160 (void)strlcat(add_controllers
, " ", full_len
+ 1);
1163 parts
= lxc_string_split(cgname
, '/');
1167 parts_len
= lxc_array_len((void **)parts
);
1171 cgroup
= must_make_path(h
->mountpoint
, h
->base_cgroup
, NULL
);
1172 for (i
= 0; i
< parts_len
; i
++) {
1176 cgroup
= must_append_path(cgroup
, parts
[i
], NULL
);
1177 target
= must_make_path(cgroup
, "cgroup.subtree_control", NULL
);
1178 ret
= lxc_write_to_file(target
, add_controllers
, full_len
, false, 0666);
1181 SYSERROR("Could not enable \"%s\" controllers in the "
1182 "unified cgroup \"%s\"", add_controllers
, cgroup
);
1190 lxc_free_array((void **)parts
, free
);
1191 free(add_controllers
);
1196 static bool create_path_for_hierarchy(struct hierarchy
*h
, char *cgname
)
1200 h
->fullcgpath
= must_make_path(h
->mountpoint
, h
->base_cgroup
, cgname
, NULL
);
1201 if (dir_exists(h
->fullcgpath
)) {
1202 ERROR("The cgroup \"%s\" already existed", h
->fullcgpath
);
1206 if (!cg_legacy_handle_cpuset_hierarchy(h
, cgname
)) {
1207 ERROR("Failed to handle legacy cpuset controller");
1211 ret
= mkdir_p(h
->fullcgpath
, 0755);
1213 ERROR("Failed to create cgroup \"%s\"", h
->fullcgpath
);
1217 return cg_unified_create_cgroup(h
, cgname
);
1220 static void remove_path_for_hierarchy(struct hierarchy
*h
, char *cgname
)
1224 ret
= rmdir(h
->fullcgpath
);
1226 SYSERROR("Failed to rmdir(\"%s\") from failed creation attempt", h
->fullcgpath
);
1228 free(h
->fullcgpath
);
1229 h
->fullcgpath
= NULL
;
1232 /* Try to create the same cgroup in all hierarchies. Start with cgroup_pattern;
1233 * next cgroup_pattern-1, -2, ..., -999.
1235 static inline bool cgfsng_create(struct cgroup_ops
*ops
,
1236 struct lxc_handler
*handler
)
1240 char *container_cgroup
, *offset
, *tmp
;
1242 struct lxc_conf
*conf
= handler
->conf
;
1244 if (ops
->container_cgroup
) {
1245 WARN("cgfsng_create called a second time: %s", ops
->container_cgroup
);
1252 if (conf
->cgroup_meta
.dir
)
1253 tmp
= lxc_string_join("/", (const char *[]){conf
->cgroup_meta
.dir
, handler
->name
, NULL
}, false);
1255 tmp
= lxc_string_replace("%n", handler
->name
, ops
->cgroup_pattern
);
1257 ERROR("Failed expanding cgroup name pattern");
1261 len
= strlen(tmp
) + 5; /* leave room for -NNN\0 */
1262 container_cgroup
= must_alloc(len
);
1263 (void)strlcpy(container_cgroup
, tmp
, len
);
1265 offset
= container_cgroup
+ len
- 5;
1269 ERROR("Too many conflicting cgroup names");
1276 ret
= snprintf(offset
, 5, "-%d", idx
);
1277 if (ret
< 0 || (size_t)ret
>= 5) {
1278 FILE *f
= fopen("/dev/null", "w");
1280 fprintf(f
, "Workaround for GCC7 bug: "
1281 "https://gcc.gnu.org/bugzilla/"
1282 "show_bug.cgi?id=78969");
1288 for (i
= 0; ops
->hierarchies
[i
]; i
++) {
1289 if (!create_path_for_hierarchy(ops
->hierarchies
[i
], container_cgroup
)) {
1291 ERROR("Failed to create cgroup \"%s\"", ops
->hierarchies
[i
]->fullcgpath
);
1292 free(ops
->hierarchies
[i
]->fullcgpath
);
1293 ops
->hierarchies
[i
]->fullcgpath
= NULL
;
1294 for (j
= 0; j
< i
; j
++)
1295 remove_path_for_hierarchy(ops
->hierarchies
[j
], container_cgroup
);
1301 ops
->container_cgroup
= container_cgroup
;
1306 free(container_cgroup
);
1311 static bool cgfsng_enter(struct cgroup_ops
*ops
, pid_t pid
)
1316 len
= snprintf(pidstr
, 25, "%d", pid
);
1317 if (len
< 0 || len
>= 25)
1320 for (i
= 0; ops
->hierarchies
[i
]; i
++) {
1324 fullpath
= must_make_path(ops
->hierarchies
[i
]->fullcgpath
,
1325 "cgroup.procs", NULL
);
1326 ret
= lxc_write_to_file(fullpath
, pidstr
, len
, false, 0666);
1328 SYSERROR("Failed to enter cgroup \"%s\"", fullpath
);
1338 static int chowmod(char *path
, uid_t chown_uid
, gid_t chown_gid
,
1343 ret
= chown(path
, chown_uid
, chown_gid
);
1345 SYSWARN("Failed to chown(%s, %d, %d)", path
, (int)chown_uid
, (int)chown_gid
);
1349 ret
= chmod(path
, chmod_mode
);
1351 SYSWARN("Failed to chmod(%s, %d)", path
, (int)chmod_mode
);
1358 /* chgrp the container cgroups to container group. We leave
1359 * the container owner as cgroup owner. So we must make the
1360 * directories 775 so that the container can create sub-cgroups.
1362 * Also chown the tasks and cgroup.procs files. Those may not
1363 * exist depending on kernel version.
1365 static int chown_cgroup_wrapper(void *data
)
1369 struct generic_userns_exec_data
*arg
= data
;
1370 uid_t nsuid
= (arg
->conf
->root_nsuid_map
!= NULL
) ? 0 : arg
->conf
->init_uid
;
1371 gid_t nsgid
= (arg
->conf
->root_nsgid_map
!= NULL
) ? 0 : arg
->conf
->init_gid
;
1373 ret
= setresgid(nsgid
, nsgid
, nsgid
);
1375 SYSERROR("Failed to setresgid(%d, %d, %d)",
1376 (int)nsgid
, (int)nsgid
, (int)nsgid
);
1380 ret
= setresuid(nsuid
, nsuid
, nsuid
);
1382 SYSERROR("Failed to setresuid(%d, %d, %d)",
1383 (int)nsuid
, (int)nsuid
, (int)nsuid
);
1387 ret
= setgroups(0, NULL
);
1388 if (ret
< 0 && errno
!= EPERM
) {
1389 SYSERROR("Failed to setgroups(0, NULL)");
1393 destuid
= get_ns_uid(arg
->origuid
);
1395 for (i
= 0; arg
->hierarchies
[i
]; i
++) {
1397 char *path
= arg
->hierarchies
[i
]->fullcgpath
;
1399 ret
= chowmod(path
, destuid
, nsgid
, 0775);
1403 /* Failures to chown() these are inconvenient but not
1404 * detrimental We leave these owned by the container launcher,
1405 * so that container root can write to the files to attach. We
1406 * chmod() them 664 so that container systemd can write to the
1407 * files (which systemd in wily insists on doing).
1410 if (arg
->hierarchies
[i
]->version
== CGROUP_SUPER_MAGIC
) {
1411 fullpath
= must_make_path(path
, "tasks", NULL
);
1412 (void)chowmod(fullpath
, destuid
, nsgid
, 0664);
1416 fullpath
= must_make_path(path
, "cgroup.procs", NULL
);
1417 (void)chowmod(fullpath
, destuid
, nsgid
, 0664);
1420 if (arg
->hierarchies
[i
]->version
!= CGROUP2_SUPER_MAGIC
)
1423 fullpath
= must_make_path(path
, "cgroup.subtree_control", NULL
);
1424 (void)chowmod(fullpath
, destuid
, nsgid
, 0664);
1427 fullpath
= must_make_path(path
, "cgroup.threads", NULL
);
1428 (void)chowmod(fullpath
, destuid
, nsgid
, 0664);
1435 static bool cgfsng_chown(struct cgroup_ops
*ops
, struct lxc_conf
*conf
)
1437 struct generic_userns_exec_data wrap
;
1439 if (lxc_list_empty(&conf
->id_map
))
1442 wrap
.origuid
= geteuid();
1444 wrap
.hierarchies
= ops
->hierarchies
;
1447 if (userns_exec_1(conf
, chown_cgroup_wrapper
, &wrap
,
1448 "chown_cgroup_wrapper") < 0) {
1449 ERROR("Error requesting cgroup chown in new user namespace");
1456 /* cgroup-full:* is done, no need to create subdirs */
1457 static bool cg_mount_needs_subdirs(int type
)
1459 if (type
>= LXC_AUTO_CGROUP_FULL_RO
)
1465 /* After $rootfs/sys/fs/container/controller/the/cg/path has been created,
1466 * remount controller ro if needed and bindmount the cgroupfs onto
1467 * controll/the/cg/path.
1469 static int cg_legacy_mount_controllers(int type
, struct hierarchy
*h
,
1470 char *controllerpath
, char *cgpath
,
1471 const char *container_cgroup
)
1473 int ret
, remount_flags
;
1475 int flags
= MS_BIND
;
1477 if (type
== LXC_AUTO_CGROUP_RO
|| type
== LXC_AUTO_CGROUP_MIXED
) {
1478 ret
= mount(controllerpath
, controllerpath
, "cgroup", MS_BIND
, NULL
);
1480 SYSERROR("Failed to bind mount \"%s\" onto \"%s\"",
1481 controllerpath
, controllerpath
);
1485 remount_flags
= add_required_remount_flags(controllerpath
,
1487 flags
| MS_REMOUNT
);
1488 ret
= mount(controllerpath
, controllerpath
, "cgroup",
1489 remount_flags
| MS_REMOUNT
| MS_BIND
| MS_RDONLY
,
1492 SYSERROR("Failed to remount \"%s\" ro", controllerpath
);
1496 INFO("Remounted %s read-only", controllerpath
);
1499 sourcepath
= must_make_path(h
->mountpoint
, h
->base_cgroup
,
1500 container_cgroup
, NULL
);
1501 if (type
== LXC_AUTO_CGROUP_RO
)
1504 ret
= mount(sourcepath
, cgpath
, "cgroup", flags
, NULL
);
1506 SYSERROR("Failed to mount \"%s\" onto \"%s\"", h
->controllers
[0], cgpath
);
1510 INFO("Mounted \"%s\" onto \"%s\"", h
->controllers
[0], cgpath
);
1512 if (flags
& MS_RDONLY
) {
1513 remount_flags
= add_required_remount_flags(sourcepath
, cgpath
,
1514 flags
| MS_REMOUNT
);
1515 ret
= mount(sourcepath
, cgpath
, "cgroup", remount_flags
, NULL
);
1517 SYSERROR("Failed to remount \"%s\" ro", cgpath
);
1521 INFO("Remounted %s read-only", cgpath
);
1525 INFO("Completed second stage cgroup automounts for \"%s\"", cgpath
);
1529 /* __cg_mount_direct
1531 * Mount cgroup hierarchies directly without using bind-mounts. The main
1532 * uses-cases are mounting cgroup hierarchies in cgroup namespaces and mounting
1533 * cgroups for the LXC_AUTO_CGROUP_FULL option.
1535 static int __cg_mount_direct(int type
, struct hierarchy
*h
,
1536 const char *controllerpath
)
1539 char *controllers
= NULL
;
1540 char *fstype
= "cgroup2";
1541 unsigned long flags
= 0;
1546 flags
|= MS_RELATIME
;
1548 if (type
== LXC_AUTO_CGROUP_RO
|| type
== LXC_AUTO_CGROUP_FULL_RO
)
1551 if (h
->version
!= CGROUP2_SUPER_MAGIC
) {
1552 controllers
= lxc_string_join(",", (const char **)h
->controllers
, false);
1558 ret
= mount("cgroup", controllerpath
, fstype
, flags
, controllers
);
1561 SYSERROR("Failed to mount \"%s\" with cgroup filesystem type %s", controllerpath
, fstype
);
1565 DEBUG("Mounted \"%s\" with cgroup filesystem type %s", controllerpath
, fstype
);
1569 static inline int cg_mount_in_cgroup_namespace(int type
, struct hierarchy
*h
,
1570 const char *controllerpath
)
1572 return __cg_mount_direct(type
, h
, controllerpath
);
1575 static inline int cg_mount_cgroup_full(int type
, struct hierarchy
*h
,
1576 const char *controllerpath
)
1578 if (type
< LXC_AUTO_CGROUP_FULL_RO
|| type
> LXC_AUTO_CGROUP_FULL_MIXED
)
1581 return __cg_mount_direct(type
, h
, controllerpath
);
1584 static bool cgfsng_mount(struct cgroup_ops
*ops
, struct lxc_handler
*handler
,
1585 const char *root
, int type
)
1588 char *tmpfspath
= NULL
;
1589 bool has_cgns
= false, retval
= false, wants_force_mount
= false;
1591 if ((type
& LXC_AUTO_CGROUP_MASK
) == 0)
1594 if (type
& LXC_AUTO_CGROUP_FORCE
) {
1595 type
&= ~LXC_AUTO_CGROUP_FORCE
;
1596 wants_force_mount
= true;
1599 if (!wants_force_mount
){
1600 if (!lxc_list_empty(&handler
->conf
->keepcaps
))
1601 wants_force_mount
= !in_caplist(CAP_SYS_ADMIN
, &handler
->conf
->keepcaps
);
1603 wants_force_mount
= in_caplist(CAP_SYS_ADMIN
, &handler
->conf
->caps
);
1606 has_cgns
= cgns_supported();
1607 if (has_cgns
&& !wants_force_mount
)
1610 if (type
== LXC_AUTO_CGROUP_NOSPEC
)
1611 type
= LXC_AUTO_CGROUP_MIXED
;
1612 else if (type
== LXC_AUTO_CGROUP_FULL_NOSPEC
)
1613 type
= LXC_AUTO_CGROUP_FULL_MIXED
;
1616 tmpfspath
= must_make_path(root
, "/sys/fs/cgroup", NULL
);
1617 ret
= safe_mount(NULL
, tmpfspath
, "tmpfs",
1618 MS_NOSUID
| MS_NODEV
| MS_NOEXEC
| MS_RELATIME
,
1619 "size=10240k,mode=755", root
);
1623 for (i
= 0; ops
->hierarchies
[i
]; i
++) {
1624 char *controllerpath
, *path2
;
1625 struct hierarchy
*h
= ops
->hierarchies
[i
];
1626 char *controller
= strrchr(h
->mountpoint
, '/');
1632 controllerpath
= must_make_path(tmpfspath
, controller
, NULL
);
1633 if (dir_exists(controllerpath
)) {
1634 free(controllerpath
);
1638 ret
= mkdir(controllerpath
, 0755);
1640 SYSERROR("Error creating cgroup path: %s", controllerpath
);
1641 free(controllerpath
);
1645 if (has_cgns
&& wants_force_mount
) {
1646 /* If cgroup namespaces are supported but the container
1647 * will not have CAP_SYS_ADMIN after it has started we
1648 * need to mount the cgroups manually.
1650 ret
= cg_mount_in_cgroup_namespace(type
, h
, controllerpath
);
1651 free(controllerpath
);
1658 ret
= cg_mount_cgroup_full(type
, h
, controllerpath
);
1660 free(controllerpath
);
1664 if (!cg_mount_needs_subdirs(type
)) {
1665 free(controllerpath
);
1669 path2
= must_make_path(controllerpath
, h
->base_cgroup
,
1670 ops
->container_cgroup
, NULL
);
1671 ret
= mkdir_p(path2
, 0755);
1673 free(controllerpath
);
1678 ret
= cg_legacy_mount_controllers(type
, h
, controllerpath
,
1679 path2
, ops
->container_cgroup
);
1680 free(controllerpath
);
1692 static int recursive_count_nrtasks(char *dirname
)
1694 struct dirent
*direntp
;
1699 dir
= opendir(dirname
);
1703 while ((direntp
= readdir(dir
))) {
1706 if (!strcmp(direntp
->d_name
, ".") ||
1707 !strcmp(direntp
->d_name
, ".."))
1710 path
= must_make_path(dirname
, direntp
->d_name
, NULL
);
1712 if (lstat(path
, &mystat
))
1715 if (!S_ISDIR(mystat
.st_mode
))
1718 count
+= recursive_count_nrtasks(path
);
1723 path
= must_make_path(dirname
, "cgroup.procs", NULL
);
1724 ret
= lxc_count_file_lines(path
);
1729 (void)closedir(dir
);
1734 static int cgfsng_nrtasks(struct cgroup_ops
*ops
)
1739 if (!ops
->container_cgroup
|| !ops
->hierarchies
)
1742 path
= must_make_path(ops
->hierarchies
[0]->fullcgpath
, NULL
);
1743 count
= recursive_count_nrtasks(path
);
1748 /* Only root needs to escape to the cgroup of its init. */
1749 static bool cgfsng_escape(const struct cgroup_ops
*ops
)
1756 for (i
= 0; ops
->hierarchies
[i
]; i
++) {
1760 fullpath
= must_make_path(ops
->hierarchies
[i
]->mountpoint
,
1761 ops
->hierarchies
[i
]->base_cgroup
,
1762 "cgroup.procs", NULL
);
1763 ret
= lxc_write_to_file(fullpath
, "0", 2, false, 0666);
1765 SYSERROR("Failed to escape to cgroup \"%s\"", fullpath
);
1775 static int cgfsng_num_hierarchies(struct cgroup_ops
*ops
)
1779 for (i
= 0; ops
->hierarchies
[i
]; i
++)
1785 static bool cgfsng_get_hierarchies(struct cgroup_ops
*ops
, int n
, char ***out
)
1789 /* sanity check n */
1790 for (i
= 0; i
< n
; i
++)
1791 if (!ops
->hierarchies
[i
])
1794 *out
= ops
->hierarchies
[i
]->controllers
;
1799 #define THAWED "THAWED"
1800 #define THAWED_LEN (strlen(THAWED))
1802 /* TODO: If the unified cgroup hierarchy grows a freezer controller this needs
1805 static bool cgfsng_unfreeze(struct cgroup_ops
*ops
)
1809 struct hierarchy
*h
;
1811 h
= get_hierarchy(ops
, "freezer");
1815 fullpath
= must_make_path(h
->fullcgpath
, "freezer.state", NULL
);
1816 ret
= lxc_write_to_file(fullpath
, THAWED
, THAWED_LEN
, false, 0666);
1824 static const char *cgfsng_get_cgroup(struct cgroup_ops
*ops
,
1825 const char *controller
)
1827 struct hierarchy
*h
;
1829 h
= get_hierarchy(ops
, controller
);
1831 WARN("Failed to find hierarchy for controller \"%s\"",
1832 controller
? controller
: "(null)");
1836 return h
->fullcgpath
? h
->fullcgpath
+ strlen(h
->mountpoint
) : NULL
;
1839 /* Given a cgroup path returned from lxc_cmd_get_cgroup_path, build a full path,
1840 * which must be freed by the caller.
1842 static inline char *build_full_cgpath_from_monitorpath(struct hierarchy
*h
,
1844 const char *filename
)
1846 return must_make_path(h
->mountpoint
, inpath
, filename
, NULL
);
1849 /* Technically, we're always at a delegation boundary here (This is especially
1850 * true when cgroup namespaces are available.). The reasoning is that in order
1851 * for us to have been able to start a container in the first place the root
1852 * cgroup must have been a leaf node. Now, either the container's init system
1853 * has populated the cgroup and kept it as a leaf node or it has created
1854 * subtrees. In the former case we will simply attach to the leaf node we
1855 * created when we started the container in the latter case we create our own
1856 * cgroup for the attaching process.
1858 static int __cg_unified_attach(const struct hierarchy
*h
, const char *name
,
1859 const char *lxcpath
, const char *pidstr
,
1860 size_t pidstr_len
, const char *controller
)
1864 int fret
= -1, idx
= 0;
1865 char *base_path
= NULL
, *container_cgroup
= NULL
, *full_path
= NULL
;
1867 container_cgroup
= lxc_cmd_get_cgroup_path(name
, lxcpath
, controller
);
1869 if (!container_cgroup
)
1872 base_path
= must_make_path(h
->mountpoint
, container_cgroup
, NULL
);
1873 full_path
= must_make_path(base_path
, "cgroup.procs", NULL
);
1874 /* cgroup is populated */
1875 ret
= lxc_write_to_file(full_path
, pidstr
, pidstr_len
, false, 0666);
1876 if (ret
< 0 && errno
!= EBUSY
)
1884 len
= strlen(base_path
) + sizeof("/lxc-1000") - 1 +
1885 sizeof("/cgroup-procs") - 1;
1886 full_path
= must_alloc(len
+ 1);
1889 ret
= snprintf(full_path
, len
+ 1, "%s/lxc-%d",
1892 ret
= snprintf(full_path
, len
+ 1, "%s/lxc", base_path
);
1893 if (ret
< 0 || (size_t)ret
>= len
+ 1)
1896 ret
= mkdir_p(full_path
, 0755);
1897 if (ret
< 0 && errno
!= EEXIST
)
1900 (void)strlcat(full_path
, "/cgroup.procs", len
+ 1);
1901 ret
= lxc_write_to_file(full_path
, pidstr
, len
, false, 0666);
1905 /* this is a non-leaf node */
1909 } while (++idx
> 0 && idx
< 1000);
1917 free(container_cgroup
);
1923 static bool cgfsng_attach(struct cgroup_ops
*ops
, const char *name
,
1924 const char *lxcpath
, pid_t pid
)
1929 len
= snprintf(pidstr
, 25, "%d", pid
);
1930 if (len
< 0 || len
>= 25)
1933 for (i
= 0; ops
->hierarchies
[i
]; i
++) {
1935 char *fullpath
= NULL
;
1936 struct hierarchy
*h
= ops
->hierarchies
[i
];
1938 if (h
->version
== CGROUP2_SUPER_MAGIC
) {
1939 ret
= __cg_unified_attach(h
, name
, lxcpath
, pidstr
, len
,
1947 path
= lxc_cmd_get_cgroup_path(name
, lxcpath
, h
->controllers
[0]);
1952 fullpath
= build_full_cgpath_from_monitorpath(h
, path
, "cgroup.procs");
1954 ret
= lxc_write_to_file(fullpath
, pidstr
, len
, false, 0666);
1956 SYSERROR("Failed to attach %d to %s", (int)pid
, fullpath
);
1966 /* Called externally (i.e. from 'lxc-cgroup') to query cgroup limits. Here we
1967 * don't have a cgroup_data set up, so we ask the running container through the
1968 * commands API for the cgroup path.
1970 static int cgfsng_get(struct cgroup_ops
*ops
, const char *filename
, char *value
,
1971 size_t len
, const char *name
, const char *lxcpath
)
1974 size_t controller_len
;
1975 char *controller
, *p
, *path
;
1976 struct hierarchy
*h
;
1978 controller_len
= strlen(filename
);
1979 controller
= alloca(controller_len
+ 1);
1980 (void)strlcpy(controller
, filename
, controller_len
+ 1);
1982 p
= strchr(controller
, '.');
1986 path
= lxc_cmd_get_cgroup_path(name
, lxcpath
, controller
);
1991 h
= get_hierarchy(ops
, controller
);
1995 fullpath
= build_full_cgpath_from_monitorpath(h
, path
, filename
);
1996 ret
= lxc_read_from_file(fullpath
, value
, len
);
2004 /* Called externally (i.e. from 'lxc-cgroup') to set new cgroup limits. Here we
2005 * don't have a cgroup_data set up, so we ask the running container through the
2006 * commands API for the cgroup path.
2008 static int cgfsng_set(struct cgroup_ops
*ops
, const char *filename
,
2009 const char *value
, const char *name
, const char *lxcpath
)
2012 size_t controller_len
;
2013 char *controller
, *p
, *path
;
2014 struct hierarchy
*h
;
2016 controller_len
= strlen(filename
);
2017 controller
= alloca(controller_len
+ 1);
2018 (void)strlcpy(controller
, filename
, controller_len
+ 1);
2020 p
= strchr(controller
, '.');
2024 path
= lxc_cmd_get_cgroup_path(name
, lxcpath
, controller
);
2029 h
= get_hierarchy(ops
, controller
);
2033 fullpath
= build_full_cgpath_from_monitorpath(h
, path
, filename
);
2034 ret
= lxc_write_to_file(fullpath
, value
, strlen(value
), false, 0666);
2042 /* take devices cgroup line
2044 * and convert it to a valid
2045 * type major:minor mode
2046 * line. Return <0 on error. Dest is a preallocated buffer long enough to hold
2049 static int convert_devpath(const char *invalue
, char *dest
)
2052 char *p
, *path
, type
;
2053 unsigned long minor
, major
;
2058 path
= must_copy_string(invalue
);
2060 /* Read path followed by mode. Ignore any trailing text.
2061 * A ' # comment' would be legal. Technically other text is not
2062 * legal, we could check for that if we cared to.
2064 for (n_parts
= 1, p
= path
; *p
&& n_parts
< 3; p
++) {
2086 ret
= stat(path
, &sb
);
2090 mode_t m
= sb
.st_mode
& S_IFMT
;
2099 ERROR("Unsupported device type %i for \"%s\"", m
, path
);
2104 major
= MAJOR(sb
.st_rdev
);
2105 minor
= MINOR(sb
.st_rdev
);
2106 ret
= snprintf(dest
, 50, "%c %lu:%lu %s", type
, major
, minor
, mode
);
2107 if (ret
< 0 || ret
>= 50) {
2108 ERROR("Error on configuration value \"%c %lu:%lu %s\" (max 50 "
2109 "chars)", type
, major
, minor
, mode
);
2110 ret
= -ENAMETOOLONG
;
2120 /* Called from setup_limits - here we have the container's cgroup_data because
2121 * we created the cgroups.
2123 static int cg_legacy_set_data(struct cgroup_ops
*ops
, const char *filename
,
2128 /* "b|c <2^64-1>:<2^64-1> r|w|m" = 47 chars max */
2129 char converted_value
[50];
2130 struct hierarchy
*h
;
2132 char *controller
= NULL
;
2134 len
= strlen(filename
);
2135 controller
= alloca(len
+ 1);
2136 (void)strlcpy(controller
, filename
, len
+ 1);
2138 p
= strchr(controller
, '.');
2142 if (strcmp("devices.allow", filename
) == 0 && value
[0] == '/') {
2143 ret
= convert_devpath(value
, converted_value
);
2146 value
= converted_value
;
2149 h
= get_hierarchy(ops
, controller
);
2151 ERROR("Failed to setup limits for the \"%s\" controller. "
2152 "The controller seems to be unused by \"cgfsng\" cgroup "
2153 "driver or not enabled on the cgroup hierarchy",
2159 fullpath
= must_make_path(h
->fullcgpath
, filename
, NULL
);
2160 ret
= lxc_write_to_file(fullpath
, value
, strlen(value
), false, 0666);
2165 static bool __cg_legacy_setup_limits(struct cgroup_ops
*ops
,
2166 struct lxc_list
*cgroup_settings
,
2169 struct lxc_list
*iterator
, *next
, *sorted_cgroup_settings
;
2170 struct lxc_cgroup
*cg
;
2173 if (lxc_list_empty(cgroup_settings
))
2176 sorted_cgroup_settings
= sort_cgroup_settings(cgroup_settings
);
2177 if (!sorted_cgroup_settings
)
2180 lxc_list_for_each(iterator
, sorted_cgroup_settings
) {
2181 cg
= iterator
->elem
;
2183 if (do_devices
== !strncmp("devices", cg
->subsystem
, 7)) {
2184 if (cg_legacy_set_data(ops
, cg
->subsystem
, cg
->value
)) {
2185 if (do_devices
&& (errno
== EACCES
|| errno
== EPERM
)) {
2186 WARN("Failed to set \"%s\" to \"%s\"",
2187 cg
->subsystem
, cg
->value
);
2190 WARN("Failed to set \"%s\" to \"%s\"",
2191 cg
->subsystem
, cg
->value
);
2194 DEBUG("Set controller \"%s\" set to \"%s\"",
2195 cg
->subsystem
, cg
->value
);
2200 INFO("Limits for the legacy cgroup hierarchies have been setup");
2202 lxc_list_for_each_safe(iterator
, sorted_cgroup_settings
, next
) {
2203 lxc_list_del(iterator
);
2206 free(sorted_cgroup_settings
);
2210 static bool __cg_unified_setup_limits(struct cgroup_ops
*ops
,
2211 struct lxc_list
*cgroup_settings
)
2213 struct lxc_list
*iterator
;
2214 struct hierarchy
*h
= ops
->unified
;
2216 if (lxc_list_empty(cgroup_settings
))
2222 lxc_list_for_each(iterator
, cgroup_settings
) {
2225 struct lxc_cgroup
*cg
= iterator
->elem
;
2227 fullpath
= must_make_path(h
->fullcgpath
, cg
->subsystem
, NULL
);
2228 ret
= lxc_write_to_file(fullpath
, cg
->value
, strlen(cg
->value
), false, 0666);
2231 SYSERROR("Failed to set \"%s\" to \"%s\"",
2232 cg
->subsystem
, cg
->value
);
2235 TRACE("Set \"%s\" to \"%s\"", cg
->subsystem
, cg
->value
);
2238 INFO("Limits for the unified cgroup hierarchy have been setup");
2242 static bool cgfsng_setup_limits(struct cgroup_ops
*ops
, struct lxc_conf
*conf
,
2247 bret
= __cg_legacy_setup_limits(ops
, &conf
->cgroup
, do_devices
);
2251 return __cg_unified_setup_limits(ops
, &conf
->cgroup2
);
2254 /* At startup, parse_hierarchies finds all the info we need about cgroup
2255 * mountpoints and current cgroups, and stores it in @d.
2257 static bool cg_hybrid_init(struct cgroup_ops
*ops
)
2265 char **klist
= NULL
, **nlist
= NULL
;
2267 /* Root spawned containers escape the current cgroup, so use init's
2268 * cgroups as our base in that case.
2270 will_escape
= (geteuid() == 0);
2272 basecginfo
= read_file("/proc/1/cgroup");
2274 basecginfo
= read_file("/proc/self/cgroup");
2278 ret
= get_existing_subsystems(&klist
, &nlist
);
2280 ERROR("Failed to retrieve available legacy cgroup controllers");
2285 f
= fopen("/proc/self/mountinfo", "r");
2287 ERROR("Failed to open \"/proc/self/mountinfo\"");
2292 lxc_cgfsng_print_basecg_debuginfo(basecginfo
, klist
, nlist
);
2294 while (getline(&line
, &len
, f
) != -1) {
2297 struct hierarchy
*new;
2298 char *base_cgroup
= NULL
, *mountpoint
= NULL
;
2299 char **controller_list
= NULL
;
2301 type
= get_cgroup_version(line
);
2305 if (type
== CGROUP2_SUPER_MAGIC
&& ops
->unified
)
2308 if (ops
->cgroup_layout
== CGROUP_LAYOUT_UNKNOWN
) {
2309 if (type
== CGROUP2_SUPER_MAGIC
)
2310 ops
->cgroup_layout
= CGROUP_LAYOUT_UNIFIED
;
2311 else if (type
== CGROUP_SUPER_MAGIC
)
2312 ops
->cgroup_layout
= CGROUP_LAYOUT_LEGACY
;
2313 } else if (ops
->cgroup_layout
== CGROUP_LAYOUT_UNIFIED
) {
2314 if (type
== CGROUP_SUPER_MAGIC
)
2315 ops
->cgroup_layout
= CGROUP_LAYOUT_HYBRID
;
2316 } else if (ops
->cgroup_layout
== CGROUP_LAYOUT_LEGACY
) {
2317 if (type
== CGROUP2_SUPER_MAGIC
)
2318 ops
->cgroup_layout
= CGROUP_LAYOUT_HYBRID
;
2321 controller_list
= cg_hybrid_get_controllers(klist
, nlist
, line
, type
);
2322 if (!controller_list
&& type
== CGROUP_SUPER_MAGIC
)
2325 if (type
== CGROUP_SUPER_MAGIC
)
2326 if (controller_list_is_dup(ops
->hierarchies
, controller_list
))
2329 mountpoint
= cg_hybrid_get_mountpoint(line
);
2331 ERROR("Failed parsing mountpoint from \"%s\"", line
);
2335 if (type
== CGROUP_SUPER_MAGIC
)
2336 base_cgroup
= cg_hybrid_get_current_cgroup(basecginfo
, controller_list
[0], CGROUP_SUPER_MAGIC
);
2338 base_cgroup
= cg_hybrid_get_current_cgroup(basecginfo
, NULL
, CGROUP2_SUPER_MAGIC
);
2340 ERROR("Failed to find current cgroup");
2345 prune_init_scope(base_cgroup
);
2346 if (type
== CGROUP2_SUPER_MAGIC
)
2347 writeable
= test_writeable_v2(mountpoint
, base_cgroup
);
2349 writeable
= test_writeable_v1(mountpoint
, base_cgroup
);
2353 if (type
== CGROUP2_SUPER_MAGIC
) {
2354 char *cgv2_ctrl_path
;
2356 cgv2_ctrl_path
= must_make_path(mountpoint
, base_cgroup
,
2357 "cgroup.controllers",
2360 controller_list
= cg_unified_get_controllers(cgv2_ctrl_path
);
2361 free(cgv2_ctrl_path
);
2362 if (!controller_list
) {
2363 controller_list
= cg_unified_make_empty_controller();
2364 TRACE("No controllers are enabled for "
2365 "delegation in the unified hierarchy");
2369 new = add_hierarchy(&ops
->hierarchies
, controller_list
, mountpoint
, base_cgroup
, type
);
2370 if (type
== CGROUP2_SUPER_MAGIC
&& !ops
->unified
)
2376 free_string_list(controller_list
);
2381 free_string_list(klist
);
2382 free_string_list(nlist
);
2389 TRACE("Writable cgroup hierarchies:");
2390 lxc_cgfsng_print_hierarchies(ops
);
2392 /* verify that all controllers in cgroup.use and all crucial
2393 * controllers are accounted for
2395 if (!all_controllers_found(ops
))
2401 static int cg_is_pure_unified(void)
2407 ret
= statfs("/sys/fs/cgroup", &fs
);
2411 if (is_fs_type(&fs
, CGROUP2_SUPER_MAGIC
))
2412 return CGROUP2_SUPER_MAGIC
;
2417 /* Get current cgroup from /proc/self/cgroup for the cgroupfs v2 hierarchy. */
2418 static char *cg_unified_get_current_cgroup(void)
2420 char *basecginfo
, *base_cgroup
;
2424 will_escape
= (geteuid() == 0);
2426 basecginfo
= read_file("/proc/1/cgroup");
2428 basecginfo
= read_file("/proc/self/cgroup");
2432 base_cgroup
= strstr(basecginfo
, "0::/");
2434 goto cleanup_on_err
;
2436 base_cgroup
= base_cgroup
+ 3;
2437 copy
= copy_to_eol(base_cgroup
);
2439 goto cleanup_on_err
;
2449 static int cg_unified_init(struct cgroup_ops
*ops
)
2452 char *mountpoint
, *subtree_path
;
2454 char *base_cgroup
= NULL
;
2456 ret
= cg_is_pure_unified();
2457 if (ret
== -ENOMEDIUM
)
2460 if (ret
!= CGROUP2_SUPER_MAGIC
)
2463 base_cgroup
= cg_unified_get_current_cgroup();
2466 prune_init_scope(base_cgroup
);
2468 /* We assume that we have already been given controllers to delegate
2469 * further down the hierarchy. If not it is up to the user to delegate
2472 mountpoint
= must_copy_string("/sys/fs/cgroup");
2473 subtree_path
= must_make_path(mountpoint
, base_cgroup
,
2474 "cgroup.subtree_control", NULL
);
2475 delegatable
= cg_unified_get_controllers(subtree_path
);
2478 delegatable
= cg_unified_make_empty_controller();
2479 if (!delegatable
[0])
2480 TRACE("No controllers are enabled for delegation");
2482 /* TODO: If the user requested specific controllers via lxc.cgroup.use
2483 * we should verify here. The reason I'm not doing it right is that I'm
2484 * not convinced that lxc.cgroup.use will be the future since it is a
2485 * global property. I much rather have an option that lets you request
2486 * controllers per container.
2489 add_hierarchy(&ops
->hierarchies
, delegatable
, mountpoint
, base_cgroup
, CGROUP2_SUPER_MAGIC
);
2491 ops
->cgroup_layout
= CGROUP_LAYOUT_UNIFIED
;
2492 return CGROUP2_SUPER_MAGIC
;
2495 static bool cg_init(struct cgroup_ops
*ops
)
2500 tmp
= lxc_global_config_value("lxc.cgroup.use");
2502 ops
->cgroup_use
= must_copy_string(tmp
);
2504 ret
= cg_unified_init(ops
);
2508 if (ret
== CGROUP2_SUPER_MAGIC
)
2511 return cg_hybrid_init(ops
);
2514 static bool cgfsng_data_init(struct cgroup_ops
*ops
)
2516 const char *cgroup_pattern
;
2518 /* copy system-wide cgroup information */
2519 cgroup_pattern
= lxc_global_config_value("lxc.cgroup.pattern");
2520 if (!cgroup_pattern
) {
2521 /* lxc.cgroup.pattern is only NULL on error. */
2522 ERROR("Failed to retrieve cgroup pattern");
2525 ops
->cgroup_pattern
= must_copy_string(cgroup_pattern
);
2530 struct cgroup_ops
*cgfsng_ops_init(void)
2532 struct cgroup_ops
*cgfsng_ops
;
2534 cgfsng_ops
= malloc(sizeof(struct cgroup_ops
));
2538 memset(cgfsng_ops
, 0, sizeof(struct cgroup_ops
));
2539 cgfsng_ops
->cgroup_layout
= CGROUP_LAYOUT_UNKNOWN
;
2541 if (!cg_init(cgfsng_ops
)) {
2546 cgfsng_ops
->data_init
= cgfsng_data_init
;
2547 cgfsng_ops
->destroy
= cgfsng_destroy
;
2548 cgfsng_ops
->create
= cgfsng_create
;
2549 cgfsng_ops
->enter
= cgfsng_enter
;
2550 cgfsng_ops
->escape
= cgfsng_escape
;
2551 cgfsng_ops
->num_hierarchies
= cgfsng_num_hierarchies
;
2552 cgfsng_ops
->get_hierarchies
= cgfsng_get_hierarchies
;
2553 cgfsng_ops
->get_cgroup
= cgfsng_get_cgroup
;
2554 cgfsng_ops
->get
= cgfsng_get
;
2555 cgfsng_ops
->set
= cgfsng_set
;
2556 cgfsng_ops
->unfreeze
= cgfsng_unfreeze
;
2557 cgfsng_ops
->setup_limits
= cgfsng_setup_limits
;
2558 cgfsng_ops
->driver
= "cgfsng";
2559 cgfsng_ops
->version
= "1.0.0";
2560 cgfsng_ops
->attach
= cgfsng_attach
;
2561 cgfsng_ops
->chown
= cgfsng_chown
;
2562 cgfsng_ops
->mount
= cgfsng_mount
;
2563 cgfsng_ops
->nrtasks
= cgfsng_nrtasks
;