2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/percpu-rwsem.h>
49 #include <linux/string.h>
50 #include <linux/sort.h>
51 #include <linux/kmod.h>
52 #include <linux/delayacct.h>
53 #include <linux/cgroupstats.h>
54 #include <linux/hashtable.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/kthread.h>
59 #include <linux/delay.h>
60 #include <linux/atomic.h>
61 #include <linux/cpuset.h>
62 #include <linux/proc_ns.h>
63 #include <linux/nsproxy.h>
64 #include <linux/file.h>
68 * pidlists linger the following amount before being destroyed. The goal
69 * is avoiding frequent destruction in the middle of consecutive read calls
70 * Expiring in the middle is a performance problem not a correctness one.
71 * 1 sec should be enough.
73 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
75 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
79 * cgroup_mutex is the master lock. Any modification to cgroup or its
80 * hierarchy must be performed while holding it.
82 * css_set_lock protects task->cgroups pointer, the list of css_set
83 * objects, and the chain of tasks off each css_set.
85 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
86 * cgroup.h can use them for lockdep annotations.
88 #ifdef CONFIG_PROVE_RCU
89 DEFINE_MUTEX(cgroup_mutex);
90 DEFINE_SPINLOCK(css_set_lock);
91 EXPORT_SYMBOL_GPL(cgroup_mutex);
92 EXPORT_SYMBOL_GPL(css_set_lock);
94 static DEFINE_MUTEX(cgroup_mutex);
95 static DEFINE_SPINLOCK(css_set_lock);
99 * Protects cgroup_idr and css_idr so that IDs can be released without
100 * grabbing cgroup_mutex.
102 static DEFINE_SPINLOCK(cgroup_idr_lock);
105 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
106 * against file removal/re-creation across css hiding.
108 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
111 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
112 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
114 static DEFINE_SPINLOCK(release_agent_path_lock);
116 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
118 #define cgroup_assert_mutex_or_rcu_locked() \
119 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
120 !lockdep_is_held(&cgroup_mutex), \
121 "cgroup_mutex or RCU read lock required");
124 * cgroup destruction makes heavy use of work items and there can be a lot
125 * of concurrent destructions. Use a separate workqueue so that cgroup
126 * destruction work items don't end up filling up max_active of system_wq
127 * which may lead to deadlock.
129 static struct workqueue_struct *cgroup_destroy_wq;
132 * pidlist destructions need to be flushed on cgroup destruction. Use a
133 * separate workqueue as flush domain.
135 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
137 /* generate an array of cgroup subsystem pointers */
138 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
139 static struct cgroup_subsys *cgroup_subsys[] = {
140 #include <linux/cgroup_subsys.h>
144 /* array of cgroup subsystem names */
145 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
146 static const char *cgroup_subsys_name[] = {
147 #include <linux/cgroup_subsys.h>
151 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
153 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
154 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
155 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
156 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
157 #include <linux/cgroup_subsys.h>
160 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
161 static struct static_key_true *cgroup_subsys_enabled_key[] = {
162 #include <linux/cgroup_subsys.h>
166 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
167 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
168 #include <linux/cgroup_subsys.h>
173 * The default hierarchy, reserved for the subsystems that are otherwise
174 * unattached - it never has more than a single cgroup, and all tasks are
175 * part of that cgroup.
177 struct cgroup_root cgrp_dfl_root;
178 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
181 * The default hierarchy always exists but is hidden until mounted for the
182 * first time. This is for backward compatibility.
184 static bool cgrp_dfl_visible;
186 /* Controllers blocked by the commandline in v1 */
187 static u16 cgroup_no_v1_mask;
189 /* some controllers are not supported in the default hierarchy */
190 static u16 cgrp_dfl_inhibit_ss_mask;
192 /* some controllers are implicitly enabled on the default hierarchy */
193 static unsigned long cgrp_dfl_implicit_ss_mask;
195 /* The list of hierarchy roots */
197 static LIST_HEAD(cgroup_roots);
198 static int cgroup_root_count;
200 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
201 static DEFINE_IDR(cgroup_hierarchy_idr);
204 * Assign a monotonically increasing serial number to csses. It guarantees
205 * cgroups with bigger numbers are newer than those with smaller numbers.
206 * Also, as csses are always appended to the parent's ->children list, it
207 * guarantees that sibling csses are always sorted in the ascending serial
208 * number order on the list. Protected by cgroup_mutex.
210 static u64 css_serial_nr_next = 1;
213 * These bitmask flags indicate whether tasks in the fork and exit paths have
214 * fork/exit handlers to call. This avoids us having to do extra work in the
215 * fork/exit path to check which subsystems have fork/exit callbacks.
217 static u16 have_fork_callback __read_mostly;
218 static u16 have_exit_callback __read_mostly;
219 static u16 have_free_callback __read_mostly;
221 /* cgroup namespace for init task */
222 struct cgroup_namespace init_cgroup_ns = {
223 .count = { .counter = 2, },
224 .user_ns = &init_user_ns,
225 .ns.ops = &cgroupns_operations,
226 .ns.inum = PROC_CGROUP_INIT_INO,
227 .root_cset = &init_css_set,
230 /* Ditto for the can_fork callback. */
231 static u16 have_canfork_callback __read_mostly;
233 static struct file_system_type cgroup2_fs_type;
234 static struct cftype cgroup_dfl_base_files[];
235 static struct cftype cgroup_legacy_base_files[];
237 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask);
238 static void cgroup_lock_and_drain_offline(struct cgroup *cgrp);
239 static int cgroup_apply_control(struct cgroup *cgrp);
240 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
241 static void css_task_iter_advance(struct css_task_iter *it);
242 static int cgroup_destroy_locked(struct cgroup *cgrp);
243 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
244 struct cgroup_subsys *ss);
245 static void css_release(struct percpu_ref *ref);
246 static void kill_css(struct cgroup_subsys_state *css);
247 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
248 struct cgroup *cgrp, struct cftype cfts[],
252 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
253 * @ssid: subsys ID of interest
255 * cgroup_subsys_enabled() can only be used with literal subsys names which
256 * is fine for individual subsystems but unsuitable for cgroup core. This
257 * is slower static_key_enabled() based test indexed by @ssid.
259 static bool cgroup_ssid_enabled(int ssid)
261 if (CGROUP_SUBSYS_COUNT == 0)
264 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
267 static bool cgroup_ssid_no_v1(int ssid)
269 return cgroup_no_v1_mask & (1 << ssid);
273 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
274 * @cgrp: the cgroup of interest
276 * The default hierarchy is the v2 interface of cgroup and this function
277 * can be used to test whether a cgroup is on the default hierarchy for
278 * cases where a subsystem should behave differnetly depending on the
281 * The set of behaviors which change on the default hierarchy are still
282 * being determined and the mount option is prefixed with __DEVEL__.
284 * List of changed behaviors:
286 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
287 * and "name" are disallowed.
289 * - When mounting an existing superblock, mount options should match.
291 * - Remount is disallowed.
293 * - rename(2) is disallowed.
295 * - "tasks" is removed. Everything should be at process granularity. Use
296 * "cgroup.procs" instead.
298 * - "cgroup.procs" is not sorted. pids will be unique unless they got
299 * recycled inbetween reads.
301 * - "release_agent" and "notify_on_release" are removed. Replacement
302 * notification mechanism will be implemented.
304 * - "cgroup.clone_children" is removed.
306 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
307 * and its descendants contain no task; otherwise, 1. The file also
308 * generates kernfs notification which can be monitored through poll and
309 * [di]notify when the value of the file changes.
311 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
312 * take masks of ancestors with non-empty cpus/mems, instead of being
313 * moved to an ancestor.
315 * - cpuset: a task can be moved into an empty cpuset, and again it takes
316 * masks of ancestors.
318 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
321 * - blkcg: blk-throttle becomes properly hierarchical.
323 * - debug: disallowed on the default hierarchy.
325 static bool cgroup_on_dfl(const struct cgroup *cgrp)
327 return cgrp->root == &cgrp_dfl_root;
330 /* IDR wrappers which synchronize using cgroup_idr_lock */
331 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
336 idr_preload(gfp_mask);
337 spin_lock_bh(&cgroup_idr_lock);
338 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
339 spin_unlock_bh(&cgroup_idr_lock);
344 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
348 spin_lock_bh(&cgroup_idr_lock);
349 ret = idr_replace(idr, ptr, id);
350 spin_unlock_bh(&cgroup_idr_lock);
354 static void cgroup_idr_remove(struct idr *idr, int id)
356 spin_lock_bh(&cgroup_idr_lock);
358 spin_unlock_bh(&cgroup_idr_lock);
361 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
363 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
366 return container_of(parent_css, struct cgroup, self);
370 /* subsystems visibly enabled on a cgroup */
371 static u16 cgroup_control(struct cgroup *cgrp)
373 struct cgroup *parent = cgroup_parent(cgrp);
374 u16 root_ss_mask = cgrp->root->subsys_mask;
377 return parent->subtree_control;
379 if (cgroup_on_dfl(cgrp))
380 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
381 cgrp_dfl_implicit_ss_mask);
385 /* subsystems enabled on a cgroup */
386 static u16 cgroup_ss_mask(struct cgroup *cgrp)
388 struct cgroup *parent = cgroup_parent(cgrp);
391 return parent->subtree_ss_mask;
393 return cgrp->root->subsys_mask;
397 * cgroup_css - obtain a cgroup's css for the specified subsystem
398 * @cgrp: the cgroup of interest
399 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
401 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
402 * function must be called either under cgroup_mutex or rcu_read_lock() and
403 * the caller is responsible for pinning the returned css if it wants to
404 * keep accessing it outside the said locks. This function may return
405 * %NULL if @cgrp doesn't have @subsys_id enabled.
407 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
408 struct cgroup_subsys *ss)
411 return rcu_dereference_check(cgrp->subsys[ss->id],
412 lockdep_is_held(&cgroup_mutex));
418 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
419 * @cgrp: the cgroup of interest
420 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
422 * Similar to cgroup_css() but returns the effective css, which is defined
423 * as the matching css of the nearest ancestor including self which has @ss
424 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
425 * function is guaranteed to return non-NULL css.
427 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
428 struct cgroup_subsys *ss)
430 lockdep_assert_held(&cgroup_mutex);
436 * This function is used while updating css associations and thus
437 * can't test the csses directly. Test ss_mask.
439 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
440 cgrp = cgroup_parent(cgrp);
445 return cgroup_css(cgrp, ss);
449 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
450 * @cgrp: the cgroup of interest
451 * @ss: the subsystem of interest
453 * Find and get the effective css of @cgrp for @ss. The effective css is
454 * defined as the matching css of the nearest ancestor including self which
455 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
456 * the root css is returned, so this function always returns a valid css.
457 * The returned css must be put using css_put().
459 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
460 struct cgroup_subsys *ss)
462 struct cgroup_subsys_state *css;
467 css = cgroup_css(cgrp, ss);
469 if (css && css_tryget_online(css))
471 cgrp = cgroup_parent(cgrp);
474 css = init_css_set.subsys[ss->id];
481 /* convenient tests for these bits */
482 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
484 return !(cgrp->self.flags & CSS_ONLINE);
487 static void cgroup_get(struct cgroup *cgrp)
489 WARN_ON_ONCE(cgroup_is_dead(cgrp));
490 css_get(&cgrp->self);
493 static bool cgroup_tryget(struct cgroup *cgrp)
495 return css_tryget(&cgrp->self);
498 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
500 struct cgroup *cgrp = of->kn->parent->priv;
501 struct cftype *cft = of_cft(of);
504 * This is open and unprotected implementation of cgroup_css().
505 * seq_css() is only called from a kernfs file operation which has
506 * an active reference on the file. Because all the subsystem
507 * files are drained before a css is disassociated with a cgroup,
508 * the matching css from the cgroup's subsys table is guaranteed to
509 * be and stay valid until the enclosing operation is complete.
512 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
516 EXPORT_SYMBOL_GPL(of_css);
518 static int notify_on_release(const struct cgroup *cgrp)
520 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
524 * for_each_css - iterate all css's of a cgroup
525 * @css: the iteration cursor
526 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
527 * @cgrp: the target cgroup to iterate css's of
529 * Should be called under cgroup_[tree_]mutex.
531 #define for_each_css(css, ssid, cgrp) \
532 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
533 if (!((css) = rcu_dereference_check( \
534 (cgrp)->subsys[(ssid)], \
535 lockdep_is_held(&cgroup_mutex)))) { } \
539 * for_each_e_css - iterate all effective css's of a cgroup
540 * @css: the iteration cursor
541 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
542 * @cgrp: the target cgroup to iterate css's of
544 * Should be called under cgroup_[tree_]mutex.
546 #define for_each_e_css(css, ssid, cgrp) \
547 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
548 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
553 * for_each_subsys - iterate all enabled cgroup subsystems
554 * @ss: the iteration cursor
555 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
557 #define for_each_subsys(ss, ssid) \
558 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
559 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
562 * do_each_subsys_mask - filter for_each_subsys with a bitmask
563 * @ss: the iteration cursor
564 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
565 * @ss_mask: the bitmask
567 * The block will only run for cases where the ssid-th bit (1 << ssid) of
570 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
571 unsigned long __ss_mask = (ss_mask); \
572 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
576 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
577 (ss) = cgroup_subsys[ssid]; \
580 #define while_each_subsys_mask() \
585 /* iterate across the hierarchies */
586 #define for_each_root(root) \
587 list_for_each_entry((root), &cgroup_roots, root_list)
589 /* iterate over child cgrps, lock should be held throughout iteration */
590 #define cgroup_for_each_live_child(child, cgrp) \
591 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
592 if (({ lockdep_assert_held(&cgroup_mutex); \
593 cgroup_is_dead(child); })) \
597 /* walk live descendants in preorder */
598 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
599 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
600 if (({ lockdep_assert_held(&cgroup_mutex); \
601 (dsct) = (d_css)->cgroup; \
602 cgroup_is_dead(dsct); })) \
606 /* walk live descendants in postorder */
607 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
608 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
609 if (({ lockdep_assert_held(&cgroup_mutex); \
610 (dsct) = (d_css)->cgroup; \
611 cgroup_is_dead(dsct); })) \
615 static void cgroup_release_agent(struct work_struct *work);
616 static void check_for_release(struct cgroup *cgrp);
619 * A cgroup can be associated with multiple css_sets as different tasks may
620 * belong to different cgroups on different hierarchies. In the other
621 * direction, a css_set is naturally associated with multiple cgroups.
622 * This M:N relationship is represented by the following link structure
623 * which exists for each association and allows traversing the associations
626 struct cgrp_cset_link {
627 /* the cgroup and css_set this link associates */
629 struct css_set *cset;
631 /* list of cgrp_cset_links anchored at cgrp->cset_links */
632 struct list_head cset_link;
634 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
635 struct list_head cgrp_link;
639 * The default css_set - used by init and its children prior to any
640 * hierarchies being mounted. It contains a pointer to the root state
641 * for each subsystem. Also used to anchor the list of css_sets. Not
642 * reference-counted, to improve performance when child cgroups
643 * haven't been created.
645 struct css_set init_css_set = {
646 .refcount = ATOMIC_INIT(1),
647 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
648 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
649 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
650 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
651 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
652 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
655 static int css_set_count = 1; /* 1 for init_css_set */
658 * css_set_populated - does a css_set contain any tasks?
659 * @cset: target css_set
661 static bool css_set_populated(struct css_set *cset)
663 lockdep_assert_held(&css_set_lock);
665 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
669 * cgroup_update_populated - updated populated count of a cgroup
670 * @cgrp: the target cgroup
671 * @populated: inc or dec populated count
673 * One of the css_sets associated with @cgrp is either getting its first
674 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
675 * count is propagated towards root so that a given cgroup's populated_cnt
676 * is zero iff the cgroup and all its descendants don't contain any tasks.
678 * @cgrp's interface file "cgroup.populated" is zero if
679 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
680 * changes from or to zero, userland is notified that the content of the
681 * interface file has changed. This can be used to detect when @cgrp and
682 * its descendants become populated or empty.
684 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
686 lockdep_assert_held(&css_set_lock);
692 trigger = !cgrp->populated_cnt++;
694 trigger = !--cgrp->populated_cnt;
699 check_for_release(cgrp);
700 cgroup_file_notify(&cgrp->events_file);
702 cgrp = cgroup_parent(cgrp);
707 * css_set_update_populated - update populated state of a css_set
708 * @cset: target css_set
709 * @populated: whether @cset is populated or depopulated
711 * @cset is either getting the first task or losing the last. Update the
712 * ->populated_cnt of all associated cgroups accordingly.
714 static void css_set_update_populated(struct css_set *cset, bool populated)
716 struct cgrp_cset_link *link;
718 lockdep_assert_held(&css_set_lock);
720 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
721 cgroup_update_populated(link->cgrp, populated);
725 * css_set_move_task - move a task from one css_set to another
726 * @task: task being moved
727 * @from_cset: css_set @task currently belongs to (may be NULL)
728 * @to_cset: new css_set @task is being moved to (may be NULL)
729 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
731 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
732 * css_set, @from_cset can be NULL. If @task is being disassociated
733 * instead of moved, @to_cset can be NULL.
735 * This function automatically handles populated_cnt updates and
736 * css_task_iter adjustments but the caller is responsible for managing
737 * @from_cset and @to_cset's reference counts.
739 static void css_set_move_task(struct task_struct *task,
740 struct css_set *from_cset, struct css_set *to_cset,
743 lockdep_assert_held(&css_set_lock);
745 if (to_cset && !css_set_populated(to_cset))
746 css_set_update_populated(to_cset, true);
749 struct css_task_iter *it, *pos;
751 WARN_ON_ONCE(list_empty(&task->cg_list));
754 * @task is leaving, advance task iterators which are
755 * pointing to it so that they can resume at the next
756 * position. Advancing an iterator might remove it from
757 * the list, use safe walk. See css_task_iter_advance*()
760 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
762 if (it->task_pos == &task->cg_list)
763 css_task_iter_advance(it);
765 list_del_init(&task->cg_list);
766 if (!css_set_populated(from_cset))
767 css_set_update_populated(from_cset, false);
769 WARN_ON_ONCE(!list_empty(&task->cg_list));
774 * We are synchronized through cgroup_threadgroup_rwsem
775 * against PF_EXITING setting such that we can't race
776 * against cgroup_exit() changing the css_set to
777 * init_css_set and dropping the old one.
779 WARN_ON_ONCE(task->flags & PF_EXITING);
781 rcu_assign_pointer(task->cgroups, to_cset);
782 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
788 * hash table for cgroup groups. This improves the performance to find
789 * an existing css_set. This hash doesn't (currently) take into
790 * account cgroups in empty hierarchies.
792 #define CSS_SET_HASH_BITS 7
793 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
795 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
797 unsigned long key = 0UL;
798 struct cgroup_subsys *ss;
801 for_each_subsys(ss, i)
802 key += (unsigned long)css[i];
803 key = (key >> 16) ^ key;
808 static void put_css_set_locked(struct css_set *cset)
810 struct cgrp_cset_link *link, *tmp_link;
811 struct cgroup_subsys *ss;
814 lockdep_assert_held(&css_set_lock);
816 if (!atomic_dec_and_test(&cset->refcount))
819 /* This css_set is dead. unlink it and release cgroup and css refs */
820 for_each_subsys(ss, ssid) {
821 list_del(&cset->e_cset_node[ssid]);
822 css_put(cset->subsys[ssid]);
824 hash_del(&cset->hlist);
827 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
828 list_del(&link->cset_link);
829 list_del(&link->cgrp_link);
830 if (cgroup_parent(link->cgrp))
831 cgroup_put(link->cgrp);
835 kfree_rcu(cset, rcu_head);
838 static void put_css_set(struct css_set *cset)
843 * Ensure that the refcount doesn't hit zero while any readers
844 * can see it. Similar to atomic_dec_and_lock(), but for an
847 if (atomic_add_unless(&cset->refcount, -1, 1))
850 spin_lock_irqsave(&css_set_lock, flags);
851 put_css_set_locked(cset);
852 spin_unlock_irqrestore(&css_set_lock, flags);
856 * refcounted get/put for css_set objects
858 static inline void get_css_set(struct css_set *cset)
860 atomic_inc(&cset->refcount);
864 * compare_css_sets - helper function for find_existing_css_set().
865 * @cset: candidate css_set being tested
866 * @old_cset: existing css_set for a task
867 * @new_cgrp: cgroup that's being entered by the task
868 * @template: desired set of css pointers in css_set (pre-calculated)
870 * Returns true if "cset" matches "old_cset" except for the hierarchy
871 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
873 static bool compare_css_sets(struct css_set *cset,
874 struct css_set *old_cset,
875 struct cgroup *new_cgrp,
876 struct cgroup_subsys_state *template[])
878 struct list_head *l1, *l2;
881 * On the default hierarchy, there can be csets which are
882 * associated with the same set of cgroups but different csses.
883 * Let's first ensure that csses match.
885 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
889 * Compare cgroup pointers in order to distinguish between
890 * different cgroups in hierarchies. As different cgroups may
891 * share the same effective css, this comparison is always
894 l1 = &cset->cgrp_links;
895 l2 = &old_cset->cgrp_links;
897 struct cgrp_cset_link *link1, *link2;
898 struct cgroup *cgrp1, *cgrp2;
902 /* See if we reached the end - both lists are equal length. */
903 if (l1 == &cset->cgrp_links) {
904 BUG_ON(l2 != &old_cset->cgrp_links);
907 BUG_ON(l2 == &old_cset->cgrp_links);
909 /* Locate the cgroups associated with these links. */
910 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
911 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
914 /* Hierarchies should be linked in the same order. */
915 BUG_ON(cgrp1->root != cgrp2->root);
918 * If this hierarchy is the hierarchy of the cgroup
919 * that's changing, then we need to check that this
920 * css_set points to the new cgroup; if it's any other
921 * hierarchy, then this css_set should point to the
922 * same cgroup as the old css_set.
924 if (cgrp1->root == new_cgrp->root) {
925 if (cgrp1 != new_cgrp)
936 * find_existing_css_set - init css array and find the matching css_set
937 * @old_cset: the css_set that we're using before the cgroup transition
938 * @cgrp: the cgroup that we're moving into
939 * @template: out param for the new set of csses, should be clear on entry
941 static struct css_set *find_existing_css_set(struct css_set *old_cset,
943 struct cgroup_subsys_state *template[])
945 struct cgroup_root *root = cgrp->root;
946 struct cgroup_subsys *ss;
947 struct css_set *cset;
952 * Build the set of subsystem state objects that we want to see in the
953 * new css_set. while subsystems can change globally, the entries here
954 * won't change, so no need for locking.
956 for_each_subsys(ss, i) {
957 if (root->subsys_mask & (1UL << i)) {
959 * @ss is in this hierarchy, so we want the
960 * effective css from @cgrp.
962 template[i] = cgroup_e_css(cgrp, ss);
965 * @ss is not in this hierarchy, so we don't want
968 template[i] = old_cset->subsys[i];
972 key = css_set_hash(template);
973 hash_for_each_possible(css_set_table, cset, hlist, key) {
974 if (!compare_css_sets(cset, old_cset, cgrp, template))
977 /* This css_set matches what we need */
981 /* No existing cgroup group matched */
985 static void free_cgrp_cset_links(struct list_head *links_to_free)
987 struct cgrp_cset_link *link, *tmp_link;
989 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
990 list_del(&link->cset_link);
996 * allocate_cgrp_cset_links - allocate cgrp_cset_links
997 * @count: the number of links to allocate
998 * @tmp_links: list_head the allocated links are put on
1000 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1001 * through ->cset_link. Returns 0 on success or -errno.
1003 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1005 struct cgrp_cset_link *link;
1008 INIT_LIST_HEAD(tmp_links);
1010 for (i = 0; i < count; i++) {
1011 link = kzalloc(sizeof(*link), GFP_KERNEL);
1013 free_cgrp_cset_links(tmp_links);
1016 list_add(&link->cset_link, tmp_links);
1022 * link_css_set - a helper function to link a css_set to a cgroup
1023 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1024 * @cset: the css_set to be linked
1025 * @cgrp: the destination cgroup
1027 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1028 struct cgroup *cgrp)
1030 struct cgrp_cset_link *link;
1032 BUG_ON(list_empty(tmp_links));
1034 if (cgroup_on_dfl(cgrp))
1035 cset->dfl_cgrp = cgrp;
1037 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1042 * Always add links to the tail of the lists so that the lists are
1043 * in choronological order.
1045 list_move_tail(&link->cset_link, &cgrp->cset_links);
1046 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1048 if (cgroup_parent(cgrp))
1053 * find_css_set - return a new css_set with one cgroup updated
1054 * @old_cset: the baseline css_set
1055 * @cgrp: the cgroup to be updated
1057 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1058 * substituted into the appropriate hierarchy.
1060 static struct css_set *find_css_set(struct css_set *old_cset,
1061 struct cgroup *cgrp)
1063 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1064 struct css_set *cset;
1065 struct list_head tmp_links;
1066 struct cgrp_cset_link *link;
1067 struct cgroup_subsys *ss;
1071 lockdep_assert_held(&cgroup_mutex);
1073 /* First see if we already have a cgroup group that matches
1074 * the desired set */
1075 spin_lock_irq(&css_set_lock);
1076 cset = find_existing_css_set(old_cset, cgrp, template);
1079 spin_unlock_irq(&css_set_lock);
1084 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1088 /* Allocate all the cgrp_cset_link objects that we'll need */
1089 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1094 atomic_set(&cset->refcount, 1);
1095 INIT_LIST_HEAD(&cset->cgrp_links);
1096 INIT_LIST_HEAD(&cset->tasks);
1097 INIT_LIST_HEAD(&cset->mg_tasks);
1098 INIT_LIST_HEAD(&cset->mg_preload_node);
1099 INIT_LIST_HEAD(&cset->mg_node);
1100 INIT_LIST_HEAD(&cset->task_iters);
1101 INIT_HLIST_NODE(&cset->hlist);
1103 /* Copy the set of subsystem state objects generated in
1104 * find_existing_css_set() */
1105 memcpy(cset->subsys, template, sizeof(cset->subsys));
1107 spin_lock_irq(&css_set_lock);
1108 /* Add reference counts and links from the new css_set. */
1109 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1110 struct cgroup *c = link->cgrp;
1112 if (c->root == cgrp->root)
1114 link_css_set(&tmp_links, cset, c);
1117 BUG_ON(!list_empty(&tmp_links));
1121 /* Add @cset to the hash table */
1122 key = css_set_hash(cset->subsys);
1123 hash_add(css_set_table, &cset->hlist, key);
1125 for_each_subsys(ss, ssid) {
1126 struct cgroup_subsys_state *css = cset->subsys[ssid];
1128 list_add_tail(&cset->e_cset_node[ssid],
1129 &css->cgroup->e_csets[ssid]);
1133 spin_unlock_irq(&css_set_lock);
1138 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1140 struct cgroup *root_cgrp = kf_root->kn->priv;
1142 return root_cgrp->root;
1145 static int cgroup_init_root_id(struct cgroup_root *root)
1149 lockdep_assert_held(&cgroup_mutex);
1151 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1155 root->hierarchy_id = id;
1159 static void cgroup_exit_root_id(struct cgroup_root *root)
1161 lockdep_assert_held(&cgroup_mutex);
1163 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1166 static void cgroup_free_root(struct cgroup_root *root)
1169 idr_destroy(&root->cgroup_idr);
1174 static void cgroup_destroy_root(struct cgroup_root *root)
1176 struct cgroup *cgrp = &root->cgrp;
1177 struct cgrp_cset_link *link, *tmp_link;
1179 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1181 BUG_ON(atomic_read(&root->nr_cgrps));
1182 BUG_ON(!list_empty(&cgrp->self.children));
1184 /* Rebind all subsystems back to the default hierarchy */
1185 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1188 * Release all the links from cset_links to this hierarchy's
1191 spin_lock_irq(&css_set_lock);
1193 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1194 list_del(&link->cset_link);
1195 list_del(&link->cgrp_link);
1199 spin_unlock_irq(&css_set_lock);
1201 if (!list_empty(&root->root_list)) {
1202 list_del(&root->root_list);
1203 cgroup_root_count--;
1206 cgroup_exit_root_id(root);
1208 mutex_unlock(&cgroup_mutex);
1210 kernfs_destroy_root(root->kf_root);
1211 cgroup_free_root(root);
1215 * look up cgroup associated with current task's cgroup namespace on the
1216 * specified hierarchy
1218 static struct cgroup *
1219 current_cgns_cgroup_from_root(struct cgroup_root *root)
1221 struct cgroup *res = NULL;
1222 struct css_set *cset;
1224 lockdep_assert_held(&css_set_lock);
1228 cset = current->nsproxy->cgroup_ns->root_cset;
1229 if (cset == &init_css_set) {
1232 struct cgrp_cset_link *link;
1234 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1235 struct cgroup *c = link->cgrp;
1237 if (c->root == root) {
1249 /* look up cgroup associated with given css_set on the specified hierarchy */
1250 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1251 struct cgroup_root *root)
1253 struct cgroup *res = NULL;
1255 lockdep_assert_held(&cgroup_mutex);
1256 lockdep_assert_held(&css_set_lock);
1258 if (cset == &init_css_set) {
1261 struct cgrp_cset_link *link;
1263 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1264 struct cgroup *c = link->cgrp;
1266 if (c->root == root) {
1278 * Return the cgroup for "task" from the given hierarchy. Must be
1279 * called with cgroup_mutex and css_set_lock held.
1281 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1282 struct cgroup_root *root)
1285 * No need to lock the task - since we hold cgroup_mutex the
1286 * task can't change groups, so the only thing that can happen
1287 * is that it exits and its css is set back to init_css_set.
1289 return cset_cgroup_from_root(task_css_set(task), root);
1293 * A task must hold cgroup_mutex to modify cgroups.
1295 * Any task can increment and decrement the count field without lock.
1296 * So in general, code holding cgroup_mutex can't rely on the count
1297 * field not changing. However, if the count goes to zero, then only
1298 * cgroup_attach_task() can increment it again. Because a count of zero
1299 * means that no tasks are currently attached, therefore there is no
1300 * way a task attached to that cgroup can fork (the other way to
1301 * increment the count). So code holding cgroup_mutex can safely
1302 * assume that if the count is zero, it will stay zero. Similarly, if
1303 * a task holds cgroup_mutex on a cgroup with zero count, it
1304 * knows that the cgroup won't be removed, as cgroup_rmdir()
1307 * A cgroup can only be deleted if both its 'count' of using tasks
1308 * is zero, and its list of 'children' cgroups is empty. Since all
1309 * tasks in the system use _some_ cgroup, and since there is always at
1310 * least one task in the system (init, pid == 1), therefore, root cgroup
1311 * always has either children cgroups and/or using tasks. So we don't
1312 * need a special hack to ensure that root cgroup cannot be deleted.
1314 * P.S. One more locking exception. RCU is used to guard the
1315 * update of a tasks cgroup pointer by cgroup_attach_task()
1318 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1319 static const struct file_operations proc_cgroupstats_operations;
1321 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1324 struct cgroup_subsys *ss = cft->ss;
1326 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1327 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1328 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1329 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1332 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1337 * cgroup_file_mode - deduce file mode of a control file
1338 * @cft: the control file in question
1340 * S_IRUGO for read, S_IWUSR for write.
1342 static umode_t cgroup_file_mode(const struct cftype *cft)
1346 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1349 if (cft->write_u64 || cft->write_s64 || cft->write) {
1350 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1360 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1361 * @subtree_control: the new subtree_control mask to consider
1362 * @this_ss_mask: available subsystems
1364 * On the default hierarchy, a subsystem may request other subsystems to be
1365 * enabled together through its ->depends_on mask. In such cases, more
1366 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1368 * This function calculates which subsystems need to be enabled if
1369 * @subtree_control is to be applied while restricted to @this_ss_mask.
1371 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1373 u16 cur_ss_mask = subtree_control;
1374 struct cgroup_subsys *ss;
1377 lockdep_assert_held(&cgroup_mutex);
1379 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1382 u16 new_ss_mask = cur_ss_mask;
1384 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1385 new_ss_mask |= ss->depends_on;
1386 } while_each_subsys_mask();
1389 * Mask out subsystems which aren't available. This can
1390 * happen only if some depended-upon subsystems were bound
1391 * to non-default hierarchies.
1393 new_ss_mask &= this_ss_mask;
1395 if (new_ss_mask == cur_ss_mask)
1397 cur_ss_mask = new_ss_mask;
1404 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1405 * @kn: the kernfs_node being serviced
1407 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1408 * the method finishes if locking succeeded. Note that once this function
1409 * returns the cgroup returned by cgroup_kn_lock_live() may become
1410 * inaccessible any time. If the caller intends to continue to access the
1411 * cgroup, it should pin it before invoking this function.
1413 static void cgroup_kn_unlock(struct kernfs_node *kn)
1415 struct cgroup *cgrp;
1417 if (kernfs_type(kn) == KERNFS_DIR)
1420 cgrp = kn->parent->priv;
1422 mutex_unlock(&cgroup_mutex);
1424 kernfs_unbreak_active_protection(kn);
1429 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1430 * @kn: the kernfs_node being serviced
1431 * @drain_offline: perform offline draining on the cgroup
1433 * This helper is to be used by a cgroup kernfs method currently servicing
1434 * @kn. It breaks the active protection, performs cgroup locking and
1435 * verifies that the associated cgroup is alive. Returns the cgroup if
1436 * alive; otherwise, %NULL. A successful return should be undone by a
1437 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1438 * cgroup is drained of offlining csses before return.
1440 * Any cgroup kernfs method implementation which requires locking the
1441 * associated cgroup should use this helper. It avoids nesting cgroup
1442 * locking under kernfs active protection and allows all kernfs operations
1443 * including self-removal.
1445 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn,
1448 struct cgroup *cgrp;
1450 if (kernfs_type(kn) == KERNFS_DIR)
1453 cgrp = kn->parent->priv;
1456 * We're gonna grab cgroup_mutex which nests outside kernfs
1457 * active_ref. cgroup liveliness check alone provides enough
1458 * protection against removal. Ensure @cgrp stays accessible and
1459 * break the active_ref protection.
1461 if (!cgroup_tryget(cgrp))
1463 kernfs_break_active_protection(kn);
1466 cgroup_lock_and_drain_offline(cgrp);
1468 mutex_lock(&cgroup_mutex);
1470 if (!cgroup_is_dead(cgrp))
1473 cgroup_kn_unlock(kn);
1477 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1479 char name[CGROUP_FILE_NAME_MAX];
1481 lockdep_assert_held(&cgroup_mutex);
1483 if (cft->file_offset) {
1484 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1485 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1487 spin_lock_irq(&cgroup_file_kn_lock);
1489 spin_unlock_irq(&cgroup_file_kn_lock);
1492 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1496 * css_clear_dir - remove subsys files in a cgroup directory
1499 static void css_clear_dir(struct cgroup_subsys_state *css)
1501 struct cgroup *cgrp = css->cgroup;
1502 struct cftype *cfts;
1504 if (!(css->flags & CSS_VISIBLE))
1507 css->flags &= ~CSS_VISIBLE;
1509 list_for_each_entry(cfts, &css->ss->cfts, node)
1510 cgroup_addrm_files(css, cgrp, cfts, false);
1514 * css_populate_dir - create subsys files in a cgroup directory
1517 * On failure, no file is added.
1519 static int css_populate_dir(struct cgroup_subsys_state *css)
1521 struct cgroup *cgrp = css->cgroup;
1522 struct cftype *cfts, *failed_cfts;
1525 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1529 if (cgroup_on_dfl(cgrp))
1530 cfts = cgroup_dfl_base_files;
1532 cfts = cgroup_legacy_base_files;
1534 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1537 list_for_each_entry(cfts, &css->ss->cfts, node) {
1538 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1545 css->flags |= CSS_VISIBLE;
1549 list_for_each_entry(cfts, &css->ss->cfts, node) {
1550 if (cfts == failed_cfts)
1552 cgroup_addrm_files(css, cgrp, cfts, false);
1557 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1559 struct cgroup *dcgrp = &dst_root->cgrp;
1560 struct cgroup_subsys *ss;
1563 lockdep_assert_held(&cgroup_mutex);
1565 do_each_subsys_mask(ss, ssid, ss_mask) {
1567 * If @ss has non-root csses attached to it, can't move.
1568 * If @ss is an implicit controller, it is exempt from this
1569 * rule and can be stolen.
1571 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1572 !ss->implicit_on_dfl)
1575 /* can't move between two non-dummy roots either */
1576 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1578 } while_each_subsys_mask();
1580 do_each_subsys_mask(ss, ssid, ss_mask) {
1581 struct cgroup_root *src_root = ss->root;
1582 struct cgroup *scgrp = &src_root->cgrp;
1583 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1584 struct css_set *cset;
1586 WARN_ON(!css || cgroup_css(dcgrp, ss));
1588 /* disable from the source */
1589 src_root->subsys_mask &= ~(1 << ssid);
1590 WARN_ON(cgroup_apply_control(scgrp));
1591 cgroup_finalize_control(scgrp, 0);
1594 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1595 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1596 ss->root = dst_root;
1597 css->cgroup = dcgrp;
1599 spin_lock_irq(&css_set_lock);
1600 hash_for_each(css_set_table, i, cset, hlist)
1601 list_move_tail(&cset->e_cset_node[ss->id],
1602 &dcgrp->e_csets[ss->id]);
1603 spin_unlock_irq(&css_set_lock);
1605 /* default hierarchy doesn't enable controllers by default */
1606 dst_root->subsys_mask |= 1 << ssid;
1607 if (dst_root == &cgrp_dfl_root) {
1608 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1610 dcgrp->subtree_control |= 1 << ssid;
1611 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1614 ret = cgroup_apply_control(dcgrp);
1616 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1621 } while_each_subsys_mask();
1623 kernfs_activate(dcgrp->kn);
1627 static int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1628 struct kernfs_root *kf_root)
1632 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1633 struct cgroup *ns_cgroup;
1635 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1639 spin_lock_irq(&css_set_lock);
1640 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1641 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1642 spin_unlock_irq(&css_set_lock);
1644 if (len >= PATH_MAX)
1647 seq_escape(sf, buf, " \t\n\\");
1654 static int cgroup_show_options(struct seq_file *seq,
1655 struct kernfs_root *kf_root)
1657 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1658 struct cgroup_subsys *ss;
1661 if (root != &cgrp_dfl_root)
1662 for_each_subsys(ss, ssid)
1663 if (root->subsys_mask & (1 << ssid))
1664 seq_show_option(seq, ss->legacy_name, NULL);
1665 if (root->flags & CGRP_ROOT_NOPREFIX)
1666 seq_puts(seq, ",noprefix");
1667 if (root->flags & CGRP_ROOT_XATTR)
1668 seq_puts(seq, ",xattr");
1670 spin_lock(&release_agent_path_lock);
1671 if (strlen(root->release_agent_path))
1672 seq_show_option(seq, "release_agent",
1673 root->release_agent_path);
1674 spin_unlock(&release_agent_path_lock);
1676 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1677 seq_puts(seq, ",clone_children");
1678 if (strlen(root->name))
1679 seq_show_option(seq, "name", root->name);
1683 struct cgroup_sb_opts {
1686 char *release_agent;
1687 bool cpuset_clone_children;
1689 /* User explicitly requested empty subsystem */
1693 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1695 char *token, *o = data;
1696 bool all_ss = false, one_ss = false;
1698 struct cgroup_subsys *ss;
1702 #ifdef CONFIG_CPUSETS
1703 mask = ~((u16)1 << cpuset_cgrp_id);
1706 memset(opts, 0, sizeof(*opts));
1708 while ((token = strsep(&o, ",")) != NULL) {
1713 if (!strcmp(token, "none")) {
1714 /* Explicitly have no subsystems */
1718 if (!strcmp(token, "all")) {
1719 /* Mutually exclusive option 'all' + subsystem name */
1725 if (!strcmp(token, "noprefix")) {
1726 opts->flags |= CGRP_ROOT_NOPREFIX;
1729 if (!strcmp(token, "clone_children")) {
1730 opts->cpuset_clone_children = true;
1733 if (!strcmp(token, "xattr")) {
1734 opts->flags |= CGRP_ROOT_XATTR;
1737 if (!strncmp(token, "release_agent=", 14)) {
1738 /* Specifying two release agents is forbidden */
1739 if (opts->release_agent)
1741 opts->release_agent =
1742 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1743 if (!opts->release_agent)
1747 if (!strncmp(token, "name=", 5)) {
1748 const char *name = token + 5;
1749 /* Can't specify an empty name */
1752 /* Must match [\w.-]+ */
1753 for (i = 0; i < strlen(name); i++) {
1757 if ((c == '.') || (c == '-') || (c == '_'))
1761 /* Specifying two names is forbidden */
1764 opts->name = kstrndup(name,
1765 MAX_CGROUP_ROOT_NAMELEN - 1,
1773 for_each_subsys(ss, i) {
1774 if (strcmp(token, ss->legacy_name))
1776 if (!cgroup_ssid_enabled(i))
1778 if (cgroup_ssid_no_v1(i))
1781 /* Mutually exclusive option 'all' + subsystem name */
1784 opts->subsys_mask |= (1 << i);
1789 if (i == CGROUP_SUBSYS_COUNT)
1794 * If the 'all' option was specified select all the subsystems,
1795 * otherwise if 'none', 'name=' and a subsystem name options were
1796 * not specified, let's default to 'all'
1798 if (all_ss || (!one_ss && !opts->none && !opts->name))
1799 for_each_subsys(ss, i)
1800 if (cgroup_ssid_enabled(i) && !cgroup_ssid_no_v1(i))
1801 opts->subsys_mask |= (1 << i);
1804 * We either have to specify by name or by subsystems. (So all
1805 * empty hierarchies must have a name).
1807 if (!opts->subsys_mask && !opts->name)
1811 * Option noprefix was introduced just for backward compatibility
1812 * with the old cpuset, so we allow noprefix only if mounting just
1813 * the cpuset subsystem.
1815 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1818 /* Can't specify "none" and some subsystems */
1819 if (opts->subsys_mask && opts->none)
1825 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1828 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1829 struct cgroup_sb_opts opts;
1830 u16 added_mask, removed_mask;
1832 if (root == &cgrp_dfl_root) {
1833 pr_err("remount is not allowed\n");
1837 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1839 /* See what subsystems are wanted */
1840 ret = parse_cgroupfs_options(data, &opts);
1844 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1845 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1846 task_tgid_nr(current), current->comm);
1848 added_mask = opts.subsys_mask & ~root->subsys_mask;
1849 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1851 /* Don't allow flags or name to change at remount */
1852 if ((opts.flags ^ root->flags) ||
1853 (opts.name && strcmp(opts.name, root->name))) {
1854 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1855 opts.flags, opts.name ?: "", root->flags, root->name);
1860 /* remounting is not allowed for populated hierarchies */
1861 if (!list_empty(&root->cgrp.self.children)) {
1866 ret = rebind_subsystems(root, added_mask);
1870 WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask));
1872 if (opts.release_agent) {
1873 spin_lock(&release_agent_path_lock);
1874 strcpy(root->release_agent_path, opts.release_agent);
1875 spin_unlock(&release_agent_path_lock);
1878 kfree(opts.release_agent);
1880 mutex_unlock(&cgroup_mutex);
1885 * To reduce the fork() overhead for systems that are not actually using
1886 * their cgroups capability, we don't maintain the lists running through
1887 * each css_set to its tasks until we see the list actually used - in other
1888 * words after the first mount.
1890 static bool use_task_css_set_links __read_mostly;
1892 static void cgroup_enable_task_cg_lists(void)
1894 struct task_struct *p, *g;
1896 spin_lock_irq(&css_set_lock);
1898 if (use_task_css_set_links)
1901 use_task_css_set_links = true;
1904 * We need tasklist_lock because RCU is not safe against
1905 * while_each_thread(). Besides, a forking task that has passed
1906 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1907 * is not guaranteed to have its child immediately visible in the
1908 * tasklist if we walk through it with RCU.
1910 read_lock(&tasklist_lock);
1911 do_each_thread(g, p) {
1912 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1913 task_css_set(p) != &init_css_set);
1916 * We should check if the process is exiting, otherwise
1917 * it will race with cgroup_exit() in that the list
1918 * entry won't be deleted though the process has exited.
1919 * Do it while holding siglock so that we don't end up
1920 * racing against cgroup_exit().
1922 * Interrupts were already disabled while acquiring
1923 * the css_set_lock, so we do not need to disable it
1924 * again when acquiring the sighand->siglock here.
1926 spin_lock(&p->sighand->siglock);
1927 if (!(p->flags & PF_EXITING)) {
1928 struct css_set *cset = task_css_set(p);
1930 if (!css_set_populated(cset))
1931 css_set_update_populated(cset, true);
1932 list_add_tail(&p->cg_list, &cset->tasks);
1935 spin_unlock(&p->sighand->siglock);
1936 } while_each_thread(g, p);
1937 read_unlock(&tasklist_lock);
1939 spin_unlock_irq(&css_set_lock);
1942 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1944 struct cgroup_subsys *ss;
1947 INIT_LIST_HEAD(&cgrp->self.sibling);
1948 INIT_LIST_HEAD(&cgrp->self.children);
1949 INIT_LIST_HEAD(&cgrp->cset_links);
1950 INIT_LIST_HEAD(&cgrp->pidlists);
1951 mutex_init(&cgrp->pidlist_mutex);
1952 cgrp->self.cgroup = cgrp;
1953 cgrp->self.flags |= CSS_ONLINE;
1955 for_each_subsys(ss, ssid)
1956 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1958 init_waitqueue_head(&cgrp->offline_waitq);
1959 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1962 static void init_cgroup_root(struct cgroup_root *root,
1963 struct cgroup_sb_opts *opts)
1965 struct cgroup *cgrp = &root->cgrp;
1967 INIT_LIST_HEAD(&root->root_list);
1968 atomic_set(&root->nr_cgrps, 1);
1970 init_cgroup_housekeeping(cgrp);
1971 idr_init(&root->cgroup_idr);
1973 root->flags = opts->flags;
1974 if (opts->release_agent)
1975 strcpy(root->release_agent_path, opts->release_agent);
1977 strcpy(root->name, opts->name);
1978 if (opts->cpuset_clone_children)
1979 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1982 static int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1984 LIST_HEAD(tmp_links);
1985 struct cgroup *root_cgrp = &root->cgrp;
1986 struct css_set *cset;
1989 lockdep_assert_held(&cgroup_mutex);
1991 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1994 root_cgrp->id = ret;
1995 root_cgrp->ancestor_ids[0] = ret;
1997 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
2003 * We're accessing css_set_count without locking css_set_lock here,
2004 * but that's OK - it can only be increased by someone holding
2005 * cgroup_lock, and that's us. Later rebinding may disable
2006 * controllers on the default hierarchy and thus create new csets,
2007 * which can't be more than the existing ones. Allocate 2x.
2009 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2013 ret = cgroup_init_root_id(root);
2017 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
2018 KERNFS_ROOT_CREATE_DEACTIVATED,
2020 if (IS_ERR(root->kf_root)) {
2021 ret = PTR_ERR(root->kf_root);
2024 root_cgrp->kn = root->kf_root->kn;
2026 ret = css_populate_dir(&root_cgrp->self);
2030 ret = rebind_subsystems(root, ss_mask);
2035 * There must be no failure case after here, since rebinding takes
2036 * care of subsystems' refcounts, which are explicitly dropped in
2037 * the failure exit path.
2039 list_add(&root->root_list, &cgroup_roots);
2040 cgroup_root_count++;
2043 * Link the root cgroup in this hierarchy into all the css_set
2046 spin_lock_irq(&css_set_lock);
2047 hash_for_each(css_set_table, i, cset, hlist) {
2048 link_css_set(&tmp_links, cset, root_cgrp);
2049 if (css_set_populated(cset))
2050 cgroup_update_populated(root_cgrp, true);
2052 spin_unlock_irq(&css_set_lock);
2054 BUG_ON(!list_empty(&root_cgrp->self.children));
2055 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2057 kernfs_activate(root_cgrp->kn);
2062 kernfs_destroy_root(root->kf_root);
2063 root->kf_root = NULL;
2065 cgroup_exit_root_id(root);
2067 percpu_ref_exit(&root_cgrp->self.refcnt);
2069 free_cgrp_cset_links(&tmp_links);
2073 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2074 int flags, const char *unused_dev_name,
2077 bool is_v2 = fs_type == &cgroup2_fs_type;
2078 struct super_block *pinned_sb = NULL;
2079 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
2080 struct cgroup_subsys *ss;
2081 struct cgroup_root *root;
2082 struct cgroup_sb_opts opts;
2083 struct dentry *dentry;
2090 /* Check if the caller has permission to mount. */
2091 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2093 return ERR_PTR(-EPERM);
2097 * The first time anyone tries to mount a cgroup, enable the list
2098 * linking each css_set to its tasks and fix up all existing tasks.
2100 if (!use_task_css_set_links)
2101 cgroup_enable_task_cg_lists();
2105 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data);
2107 return ERR_PTR(-EINVAL);
2109 cgrp_dfl_visible = true;
2110 root = &cgrp_dfl_root;
2111 cgroup_get(&root->cgrp);
2115 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
2117 /* First find the desired set of subsystems */
2118 ret = parse_cgroupfs_options(data, &opts);
2123 * Destruction of cgroup root is asynchronous, so subsystems may
2124 * still be dying after the previous unmount. Let's drain the
2125 * dying subsystems. We just need to ensure that the ones
2126 * unmounted previously finish dying and don't care about new ones
2127 * starting. Testing ref liveliness is good enough.
2129 for_each_subsys(ss, i) {
2130 if (!(opts.subsys_mask & (1 << i)) ||
2131 ss->root == &cgrp_dfl_root)
2134 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2135 mutex_unlock(&cgroup_mutex);
2137 ret = restart_syscall();
2140 cgroup_put(&ss->root->cgrp);
2143 for_each_root(root) {
2144 bool name_match = false;
2146 if (root == &cgrp_dfl_root)
2150 * If we asked for a name then it must match. Also, if
2151 * name matches but sybsys_mask doesn't, we should fail.
2152 * Remember whether name matched.
2155 if (strcmp(opts.name, root->name))
2161 * If we asked for subsystems (or explicitly for no
2162 * subsystems) then they must match.
2164 if ((opts.subsys_mask || opts.none) &&
2165 (opts.subsys_mask != root->subsys_mask)) {
2172 if (root->flags ^ opts.flags)
2173 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2176 * We want to reuse @root whose lifetime is governed by its
2177 * ->cgrp. Let's check whether @root is alive and keep it
2178 * that way. As cgroup_kill_sb() can happen anytime, we
2179 * want to block it by pinning the sb so that @root doesn't
2180 * get killed before mount is complete.
2182 * With the sb pinned, tryget_live can reliably indicate
2183 * whether @root can be reused. If it's being killed,
2184 * drain it. We can use wait_queue for the wait but this
2185 * path is super cold. Let's just sleep a bit and retry.
2187 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2188 if (IS_ERR(pinned_sb) ||
2189 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2190 mutex_unlock(&cgroup_mutex);
2191 if (!IS_ERR_OR_NULL(pinned_sb))
2192 deactivate_super(pinned_sb);
2194 ret = restart_syscall();
2203 * No such thing, create a new one. name= matching without subsys
2204 * specification is allowed for already existing hierarchies but we
2205 * can't create new one without subsys specification.
2207 if (!opts.subsys_mask && !opts.none) {
2212 /* Hierarchies may only be created in the initial cgroup namespace. */
2213 if (ns != &init_cgroup_ns) {
2218 root = kzalloc(sizeof(*root), GFP_KERNEL);
2224 init_cgroup_root(root, &opts);
2226 ret = cgroup_setup_root(root, opts.subsys_mask);
2228 cgroup_free_root(root);
2231 mutex_unlock(&cgroup_mutex);
2233 kfree(opts.release_agent);
2238 return ERR_PTR(ret);
2241 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2242 is_v2 ? CGROUP2_SUPER_MAGIC : CGROUP_SUPER_MAGIC,
2246 * In non-init cgroup namespace, instead of root cgroup's
2247 * dentry, we return the dentry corresponding to the
2248 * cgroupns->root_cgrp.
2250 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
2251 struct dentry *nsdentry;
2252 struct cgroup *cgrp;
2254 mutex_lock(&cgroup_mutex);
2255 spin_lock_irq(&css_set_lock);
2257 cgrp = cset_cgroup_from_root(ns->root_cset, root);
2259 spin_unlock_irq(&css_set_lock);
2260 mutex_unlock(&cgroup_mutex);
2262 nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
2267 if (IS_ERR(dentry) || !new_sb)
2268 cgroup_put(&root->cgrp);
2271 * If @pinned_sb, we're reusing an existing root and holding an
2272 * extra ref on its sb. Mount is complete. Put the extra ref.
2276 deactivate_super(pinned_sb);
2283 static void cgroup_kill_sb(struct super_block *sb)
2285 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2286 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2289 * If @root doesn't have any mounts or children, start killing it.
2290 * This prevents new mounts by disabling percpu_ref_tryget_live().
2291 * cgroup_mount() may wait for @root's release.
2293 * And don't kill the default root.
2295 if (!list_empty(&root->cgrp.self.children) ||
2296 root == &cgrp_dfl_root)
2297 cgroup_put(&root->cgrp);
2299 percpu_ref_kill(&root->cgrp.self.refcnt);
2304 static struct file_system_type cgroup_fs_type = {
2306 .mount = cgroup_mount,
2307 .kill_sb = cgroup_kill_sb,
2308 .fs_flags = FS_USERNS_MOUNT,
2311 static struct file_system_type cgroup2_fs_type = {
2313 .mount = cgroup_mount,
2314 .kill_sb = cgroup_kill_sb,
2315 .fs_flags = FS_USERNS_MOUNT,
2318 static char *cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2319 struct cgroup_namespace *ns)
2321 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2324 ret = kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2325 if (ret < 0 || ret >= buflen)
2330 char *cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2331 struct cgroup_namespace *ns)
2335 mutex_lock(&cgroup_mutex);
2336 spin_lock_irq(&css_set_lock);
2338 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2340 spin_unlock_irq(&css_set_lock);
2341 mutex_unlock(&cgroup_mutex);
2345 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2348 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2349 * @task: target task
2350 * @buf: the buffer to write the path into
2351 * @buflen: the length of the buffer
2353 * Determine @task's cgroup on the first (the one with the lowest non-zero
2354 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2355 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2356 * cgroup controller callbacks.
2358 * Return value is the same as kernfs_path().
2360 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2362 struct cgroup_root *root;
2363 struct cgroup *cgrp;
2364 int hierarchy_id = 1;
2367 mutex_lock(&cgroup_mutex);
2368 spin_lock_irq(&css_set_lock);
2370 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2373 cgrp = task_cgroup_from_root(task, root);
2374 path = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2376 /* if no hierarchy exists, everyone is in "/" */
2377 if (strlcpy(buf, "/", buflen) < buflen)
2381 spin_unlock_irq(&css_set_lock);
2382 mutex_unlock(&cgroup_mutex);
2385 EXPORT_SYMBOL_GPL(task_cgroup_path);
2387 /* used to track tasks and other necessary states during migration */
2388 struct cgroup_taskset {
2389 /* the src and dst cset list running through cset->mg_node */
2390 struct list_head src_csets;
2391 struct list_head dst_csets;
2393 /* the subsys currently being processed */
2397 * Fields for cgroup_taskset_*() iteration.
2399 * Before migration is committed, the target migration tasks are on
2400 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2401 * the csets on ->dst_csets. ->csets point to either ->src_csets
2402 * or ->dst_csets depending on whether migration is committed.
2404 * ->cur_csets and ->cur_task point to the current task position
2407 struct list_head *csets;
2408 struct css_set *cur_cset;
2409 struct task_struct *cur_task;
2412 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2413 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2414 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2415 .csets = &tset.src_csets, \
2419 * cgroup_taskset_add - try to add a migration target task to a taskset
2420 * @task: target task
2421 * @tset: target taskset
2423 * Add @task, which is a migration target, to @tset. This function becomes
2424 * noop if @task doesn't need to be migrated. @task's css_set should have
2425 * been added as a migration source and @task->cg_list will be moved from
2426 * the css_set's tasks list to mg_tasks one.
2428 static void cgroup_taskset_add(struct task_struct *task,
2429 struct cgroup_taskset *tset)
2431 struct css_set *cset;
2433 lockdep_assert_held(&css_set_lock);
2435 /* @task either already exited or can't exit until the end */
2436 if (task->flags & PF_EXITING)
2439 /* leave @task alone if post_fork() hasn't linked it yet */
2440 if (list_empty(&task->cg_list))
2443 cset = task_css_set(task);
2444 if (!cset->mg_src_cgrp)
2447 list_move_tail(&task->cg_list, &cset->mg_tasks);
2448 if (list_empty(&cset->mg_node))
2449 list_add_tail(&cset->mg_node, &tset->src_csets);
2450 if (list_empty(&cset->mg_dst_cset->mg_node))
2451 list_move_tail(&cset->mg_dst_cset->mg_node,
2456 * cgroup_taskset_first - reset taskset and return the first task
2457 * @tset: taskset of interest
2458 * @dst_cssp: output variable for the destination css
2460 * @tset iteration is initialized and the first task is returned.
2462 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2463 struct cgroup_subsys_state **dst_cssp)
2465 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2466 tset->cur_task = NULL;
2468 return cgroup_taskset_next(tset, dst_cssp);
2472 * cgroup_taskset_next - iterate to the next task in taskset
2473 * @tset: taskset of interest
2474 * @dst_cssp: output variable for the destination css
2476 * Return the next task in @tset. Iteration must have been initialized
2477 * with cgroup_taskset_first().
2479 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2480 struct cgroup_subsys_state **dst_cssp)
2482 struct css_set *cset = tset->cur_cset;
2483 struct task_struct *task = tset->cur_task;
2485 while (&cset->mg_node != tset->csets) {
2487 task = list_first_entry(&cset->mg_tasks,
2488 struct task_struct, cg_list);
2490 task = list_next_entry(task, cg_list);
2492 if (&task->cg_list != &cset->mg_tasks) {
2493 tset->cur_cset = cset;
2494 tset->cur_task = task;
2497 * This function may be called both before and
2498 * after cgroup_taskset_migrate(). The two cases
2499 * can be distinguished by looking at whether @cset
2500 * has its ->mg_dst_cset set.
2502 if (cset->mg_dst_cset)
2503 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2505 *dst_cssp = cset->subsys[tset->ssid];
2510 cset = list_next_entry(cset, mg_node);
2518 * cgroup_taskset_migrate - migrate a taskset
2519 * @tset: taget taskset
2520 * @root: cgroup root the migration is taking place on
2522 * Migrate tasks in @tset as setup by migration preparation functions.
2523 * This function fails iff one of the ->can_attach callbacks fails and
2524 * guarantees that either all or none of the tasks in @tset are migrated.
2525 * @tset is consumed regardless of success.
2527 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2528 struct cgroup_root *root)
2530 struct cgroup_subsys *ss;
2531 struct task_struct *task, *tmp_task;
2532 struct css_set *cset, *tmp_cset;
2533 int ssid, failed_ssid, ret;
2535 /* methods shouldn't be called if no task is actually migrating */
2536 if (list_empty(&tset->src_csets))
2539 /* check that we can legitimately attach to the cgroup */
2540 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2541 if (ss->can_attach) {
2543 ret = ss->can_attach(tset);
2546 goto out_cancel_attach;
2549 } while_each_subsys_mask();
2552 * Now that we're guaranteed success, proceed to move all tasks to
2553 * the new cgroup. There are no failure cases after here, so this
2554 * is the commit point.
2556 spin_lock_irq(&css_set_lock);
2557 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2558 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2559 struct css_set *from_cset = task_css_set(task);
2560 struct css_set *to_cset = cset->mg_dst_cset;
2562 get_css_set(to_cset);
2563 css_set_move_task(task, from_cset, to_cset, true);
2564 put_css_set_locked(from_cset);
2567 spin_unlock_irq(&css_set_lock);
2570 * Migration is committed, all target tasks are now on dst_csets.
2571 * Nothing is sensitive to fork() after this point. Notify
2572 * controllers that migration is complete.
2574 tset->csets = &tset->dst_csets;
2576 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2581 } while_each_subsys_mask();
2584 goto out_release_tset;
2587 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2588 if (ssid == failed_ssid)
2590 if (ss->cancel_attach) {
2592 ss->cancel_attach(tset);
2594 } while_each_subsys_mask();
2596 spin_lock_irq(&css_set_lock);
2597 list_splice_init(&tset->dst_csets, &tset->src_csets);
2598 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2599 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2600 list_del_init(&cset->mg_node);
2602 spin_unlock_irq(&css_set_lock);
2607 * cgroup_may_migrate_to - verify whether a cgroup can be migration destination
2608 * @dst_cgrp: destination cgroup to test
2610 * On the default hierarchy, except for the root, subtree_control must be
2611 * zero for migration destination cgroups with tasks so that child cgroups
2612 * don't compete against tasks.
2614 static bool cgroup_may_migrate_to(struct cgroup *dst_cgrp)
2616 return !cgroup_on_dfl(dst_cgrp) || !cgroup_parent(dst_cgrp) ||
2617 !dst_cgrp->subtree_control;
2621 * cgroup_migrate_finish - cleanup after attach
2622 * @preloaded_csets: list of preloaded css_sets
2624 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2625 * those functions for details.
2627 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2629 struct css_set *cset, *tmp_cset;
2631 lockdep_assert_held(&cgroup_mutex);
2633 spin_lock_irq(&css_set_lock);
2634 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2635 cset->mg_src_cgrp = NULL;
2636 cset->mg_dst_cgrp = NULL;
2637 cset->mg_dst_cset = NULL;
2638 list_del_init(&cset->mg_preload_node);
2639 put_css_set_locked(cset);
2641 spin_unlock_irq(&css_set_lock);
2645 * cgroup_migrate_add_src - add a migration source css_set
2646 * @src_cset: the source css_set to add
2647 * @dst_cgrp: the destination cgroup
2648 * @preloaded_csets: list of preloaded css_sets
2650 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2651 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2652 * up by cgroup_migrate_finish().
2654 * This function may be called without holding cgroup_threadgroup_rwsem
2655 * even if the target is a process. Threads may be created and destroyed
2656 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2657 * into play and the preloaded css_sets are guaranteed to cover all
2660 static void cgroup_migrate_add_src(struct css_set *src_cset,
2661 struct cgroup *dst_cgrp,
2662 struct list_head *preloaded_csets)
2664 struct cgroup *src_cgrp;
2666 lockdep_assert_held(&cgroup_mutex);
2667 lockdep_assert_held(&css_set_lock);
2670 * If ->dead, @src_set is associated with one or more dead cgroups
2671 * and doesn't contain any migratable tasks. Ignore it early so
2672 * that the rest of migration path doesn't get confused by it.
2677 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2679 if (!list_empty(&src_cset->mg_preload_node))
2682 WARN_ON(src_cset->mg_src_cgrp);
2683 WARN_ON(src_cset->mg_dst_cgrp);
2684 WARN_ON(!list_empty(&src_cset->mg_tasks));
2685 WARN_ON(!list_empty(&src_cset->mg_node));
2687 src_cset->mg_src_cgrp = src_cgrp;
2688 src_cset->mg_dst_cgrp = dst_cgrp;
2689 get_css_set(src_cset);
2690 list_add(&src_cset->mg_preload_node, preloaded_csets);
2694 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2695 * @preloaded_csets: list of preloaded source css_sets
2697 * Tasks are about to be moved and all the source css_sets have been
2698 * preloaded to @preloaded_csets. This function looks up and pins all
2699 * destination css_sets, links each to its source, and append them to
2702 * This function must be called after cgroup_migrate_add_src() has been
2703 * called on each migration source css_set. After migration is performed
2704 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2707 static int cgroup_migrate_prepare_dst(struct list_head *preloaded_csets)
2710 struct css_set *src_cset, *tmp_cset;
2712 lockdep_assert_held(&cgroup_mutex);
2714 /* look up the dst cset for each src cset and link it to src */
2715 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2716 struct css_set *dst_cset;
2718 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2722 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2725 * If src cset equals dst, it's noop. Drop the src.
2726 * cgroup_migrate() will skip the cset too. Note that we
2727 * can't handle src == dst as some nodes are used by both.
2729 if (src_cset == dst_cset) {
2730 src_cset->mg_src_cgrp = NULL;
2731 src_cset->mg_dst_cgrp = NULL;
2732 list_del_init(&src_cset->mg_preload_node);
2733 put_css_set(src_cset);
2734 put_css_set(dst_cset);
2738 src_cset->mg_dst_cset = dst_cset;
2740 if (list_empty(&dst_cset->mg_preload_node))
2741 list_add(&dst_cset->mg_preload_node, &csets);
2743 put_css_set(dst_cset);
2746 list_splice_tail(&csets, preloaded_csets);
2749 cgroup_migrate_finish(&csets);
2754 * cgroup_migrate - migrate a process or task to a cgroup
2755 * @leader: the leader of the process or the task to migrate
2756 * @threadgroup: whether @leader points to the whole process or a single task
2757 * @root: cgroup root migration is taking place on
2759 * Migrate a process or task denoted by @leader. If migrating a process,
2760 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2761 * responsible for invoking cgroup_migrate_add_src() and
2762 * cgroup_migrate_prepare_dst() on the targets before invoking this
2763 * function and following up with cgroup_migrate_finish().
2765 * As long as a controller's ->can_attach() doesn't fail, this function is
2766 * guaranteed to succeed. This means that, excluding ->can_attach()
2767 * failure, when migrating multiple targets, the success or failure can be
2768 * decided for all targets by invoking group_migrate_prepare_dst() before
2769 * actually starting migrating.
2771 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2772 struct cgroup_root *root)
2774 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2775 struct task_struct *task;
2778 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2779 * already PF_EXITING could be freed from underneath us unless we
2780 * take an rcu_read_lock.
2782 spin_lock_irq(&css_set_lock);
2786 cgroup_taskset_add(task, &tset);
2789 } while_each_thread(leader, task);
2791 spin_unlock_irq(&css_set_lock);
2793 return cgroup_taskset_migrate(&tset, root);
2797 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2798 * @dst_cgrp: the cgroup to attach to
2799 * @leader: the task or the leader of the threadgroup to be attached
2800 * @threadgroup: attach the whole threadgroup?
2802 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2804 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2805 struct task_struct *leader, bool threadgroup)
2807 LIST_HEAD(preloaded_csets);
2808 struct task_struct *task;
2811 if (!cgroup_may_migrate_to(dst_cgrp))
2814 /* look up all src csets */
2815 spin_lock_irq(&css_set_lock);
2819 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2823 } while_each_thread(leader, task);
2825 spin_unlock_irq(&css_set_lock);
2827 /* prepare dst csets and commit */
2828 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
2830 ret = cgroup_migrate(leader, threadgroup, dst_cgrp->root);
2832 cgroup_migrate_finish(&preloaded_csets);
2836 static int cgroup_procs_write_permission(struct task_struct *task,
2837 struct cgroup *dst_cgrp,
2838 struct kernfs_open_file *of)
2840 const struct cred *cred = current_cred();
2841 const struct cred *tcred = get_task_cred(task);
2845 * even if we're attaching all tasks in the thread group, we only
2846 * need to check permissions on one of them.
2848 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2849 !uid_eq(cred->euid, tcred->uid) &&
2850 !uid_eq(cred->euid, tcred->suid))
2853 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2854 struct super_block *sb = of->file->f_path.dentry->d_sb;
2855 struct cgroup *cgrp;
2856 struct inode *inode;
2858 spin_lock_irq(&css_set_lock);
2859 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2860 spin_unlock_irq(&css_set_lock);
2862 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2863 cgrp = cgroup_parent(cgrp);
2866 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2868 ret = inode_permission(inode, MAY_WRITE);
2878 * Find the task_struct of the task to attach by vpid and pass it along to the
2879 * function to attach either it or all tasks in its threadgroup. Will lock
2880 * cgroup_mutex and threadgroup.
2882 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2883 size_t nbytes, loff_t off, bool threadgroup)
2885 struct task_struct *tsk;
2886 struct cgroup_subsys *ss;
2887 struct cgroup *cgrp;
2891 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2894 cgrp = cgroup_kn_lock_live(of->kn, false);
2898 percpu_down_write(&cgroup_threadgroup_rwsem);
2901 tsk = find_task_by_vpid(pid);
2904 goto out_unlock_rcu;
2911 tsk = tsk->group_leader;
2914 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2915 * trapped in a cpuset, or RT worker may be born in a cgroup
2916 * with no rt_runtime allocated. Just say no.
2918 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2920 goto out_unlock_rcu;
2923 get_task_struct(tsk);
2926 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2928 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2930 put_task_struct(tsk);
2931 goto out_unlock_threadgroup;
2935 out_unlock_threadgroup:
2936 percpu_up_write(&cgroup_threadgroup_rwsem);
2937 for_each_subsys(ss, ssid)
2938 if (ss->post_attach)
2940 cgroup_kn_unlock(of->kn);
2941 return ret ?: nbytes;
2945 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2946 * @from: attach to all cgroups of a given task
2947 * @tsk: the task to be attached
2949 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2951 struct cgroup_root *root;
2954 mutex_lock(&cgroup_mutex);
2955 percpu_down_write(&cgroup_threadgroup_rwsem);
2956 for_each_root(root) {
2957 struct cgroup *from_cgrp;
2959 if (root == &cgrp_dfl_root)
2962 spin_lock_irq(&css_set_lock);
2963 from_cgrp = task_cgroup_from_root(from, root);
2964 spin_unlock_irq(&css_set_lock);
2966 retval = cgroup_attach_task(from_cgrp, tsk, false);
2970 percpu_up_write(&cgroup_threadgroup_rwsem);
2971 mutex_unlock(&cgroup_mutex);
2975 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2977 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2978 char *buf, size_t nbytes, loff_t off)
2980 return __cgroup_procs_write(of, buf, nbytes, off, false);
2983 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2984 char *buf, size_t nbytes, loff_t off)
2986 return __cgroup_procs_write(of, buf, nbytes, off, true);
2989 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2990 char *buf, size_t nbytes, loff_t off)
2992 struct cgroup *cgrp;
2994 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2996 cgrp = cgroup_kn_lock_live(of->kn, false);
2999 spin_lock(&release_agent_path_lock);
3000 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
3001 sizeof(cgrp->root->release_agent_path));
3002 spin_unlock(&release_agent_path_lock);
3003 cgroup_kn_unlock(of->kn);
3007 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
3009 struct cgroup *cgrp = seq_css(seq)->cgroup;
3011 spin_lock(&release_agent_path_lock);
3012 seq_puts(seq, cgrp->root->release_agent_path);
3013 spin_unlock(&release_agent_path_lock);
3014 seq_putc(seq, '\n');
3018 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
3020 seq_puts(seq, "0\n");
3024 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
3026 struct cgroup_subsys *ss;
3027 bool printed = false;
3030 do_each_subsys_mask(ss, ssid, ss_mask) {
3033 seq_printf(seq, "%s", ss->name);
3035 } while_each_subsys_mask();
3037 seq_putc(seq, '\n');
3040 /* show controllers which are enabled from the parent */
3041 static int cgroup_controllers_show(struct seq_file *seq, void *v)
3043 struct cgroup *cgrp = seq_css(seq)->cgroup;
3045 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
3049 /* show controllers which are enabled for a given cgroup's children */
3050 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
3052 struct cgroup *cgrp = seq_css(seq)->cgroup;
3054 cgroup_print_ss_mask(seq, cgrp->subtree_control);
3059 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3060 * @cgrp: root of the subtree to update csses for
3062 * @cgrp's control masks have changed and its subtree's css associations
3063 * need to be updated accordingly. This function looks up all css_sets
3064 * which are attached to the subtree, creates the matching updated css_sets
3065 * and migrates the tasks to the new ones.
3067 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3069 LIST_HEAD(preloaded_csets);
3070 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
3071 struct cgroup_subsys_state *d_css;
3072 struct cgroup *dsct;
3073 struct css_set *src_cset;
3076 lockdep_assert_held(&cgroup_mutex);
3078 percpu_down_write(&cgroup_threadgroup_rwsem);
3080 /* look up all csses currently attached to @cgrp's subtree */
3081 spin_lock_irq(&css_set_lock);
3082 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3083 struct cgrp_cset_link *link;
3085 list_for_each_entry(link, &dsct->cset_links, cset_link)
3086 cgroup_migrate_add_src(link->cset, dsct,
3089 spin_unlock_irq(&css_set_lock);
3091 /* NULL dst indicates self on default hierarchy */
3092 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
3096 spin_lock_irq(&css_set_lock);
3097 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
3098 struct task_struct *task, *ntask;
3100 /* src_csets precede dst_csets, break on the first dst_cset */
3101 if (!src_cset->mg_src_cgrp)
3104 /* all tasks in src_csets need to be migrated */
3105 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3106 cgroup_taskset_add(task, &tset);
3108 spin_unlock_irq(&css_set_lock);
3110 ret = cgroup_taskset_migrate(&tset, cgrp->root);
3112 cgroup_migrate_finish(&preloaded_csets);
3113 percpu_up_write(&cgroup_threadgroup_rwsem);
3118 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3119 * @cgrp: root of the target subtree
3121 * Because css offlining is asynchronous, userland may try to re-enable a
3122 * controller while the previous css is still around. This function grabs
3123 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3125 static void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3126 __acquires(&cgroup_mutex)
3128 struct cgroup *dsct;
3129 struct cgroup_subsys_state *d_css;
3130 struct cgroup_subsys *ss;
3134 mutex_lock(&cgroup_mutex);
3136 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3137 for_each_subsys(ss, ssid) {
3138 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3141 if (!css || !percpu_ref_is_dying(&css->refcnt))
3145 prepare_to_wait(&dsct->offline_waitq, &wait,
3146 TASK_UNINTERRUPTIBLE);
3148 mutex_unlock(&cgroup_mutex);
3150 finish_wait(&dsct->offline_waitq, &wait);
3159 * cgroup_save_control - save control masks of a subtree
3160 * @cgrp: root of the target subtree
3162 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
3163 * prefixed fields for @cgrp's subtree including @cgrp itself.
3165 static void cgroup_save_control(struct cgroup *cgrp)
3167 struct cgroup *dsct;
3168 struct cgroup_subsys_state *d_css;
3170 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3171 dsct->old_subtree_control = dsct->subtree_control;
3172 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3177 * cgroup_propagate_control - refresh control masks of a subtree
3178 * @cgrp: root of the target subtree
3180 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3181 * ->subtree_control and propagate controller availability through the
3182 * subtree so that descendants don't have unavailable controllers enabled.
3184 static void cgroup_propagate_control(struct cgroup *cgrp)
3186 struct cgroup *dsct;
3187 struct cgroup_subsys_state *d_css;
3189 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3190 dsct->subtree_control &= cgroup_control(dsct);
3191 dsct->subtree_ss_mask =
3192 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3193 cgroup_ss_mask(dsct));
3198 * cgroup_restore_control - restore control masks of a subtree
3199 * @cgrp: root of the target subtree
3201 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
3202 * prefixed fields for @cgrp's subtree including @cgrp itself.
3204 static void cgroup_restore_control(struct cgroup *cgrp)
3206 struct cgroup *dsct;
3207 struct cgroup_subsys_state *d_css;
3209 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3210 dsct->subtree_control = dsct->old_subtree_control;
3211 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3215 static bool css_visible(struct cgroup_subsys_state *css)
3217 struct cgroup_subsys *ss = css->ss;
3218 struct cgroup *cgrp = css->cgroup;
3220 if (cgroup_control(cgrp) & (1 << ss->id))
3222 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3224 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3228 * cgroup_apply_control_enable - enable or show csses according to control
3229 * @cgrp: root of the target subtree
3231 * Walk @cgrp's subtree and create new csses or make the existing ones
3232 * visible. A css is created invisible if it's being implicitly enabled
3233 * through dependency. An invisible css is made visible when the userland
3234 * explicitly enables it.
3236 * Returns 0 on success, -errno on failure. On failure, csses which have
3237 * been processed already aren't cleaned up. The caller is responsible for
3238 * cleaning up with cgroup_apply_control_disble().
3240 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3242 struct cgroup *dsct;
3243 struct cgroup_subsys_state *d_css;
3244 struct cgroup_subsys *ss;
3247 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3248 for_each_subsys(ss, ssid) {
3249 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3251 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3253 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3257 css = css_create(dsct, ss);
3259 return PTR_ERR(css);
3262 if (css_visible(css)) {
3263 ret = css_populate_dir(css);
3274 * cgroup_apply_control_disable - kill or hide csses according to control
3275 * @cgrp: root of the target subtree
3277 * Walk @cgrp's subtree and kill and hide csses so that they match
3278 * cgroup_ss_mask() and cgroup_visible_mask().
3280 * A css is hidden when the userland requests it to be disabled while other
3281 * subsystems are still depending on it. The css must not actively control
3282 * resources and be in the vanilla state if it's made visible again later.
3283 * Controllers which may be depended upon should provide ->css_reset() for
3286 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3288 struct cgroup *dsct;
3289 struct cgroup_subsys_state *d_css;
3290 struct cgroup_subsys *ss;
3293 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3294 for_each_subsys(ss, ssid) {
3295 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3297 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3303 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3305 } else if (!css_visible(css)) {
3315 * cgroup_apply_control - apply control mask updates to the subtree
3316 * @cgrp: root of the target subtree
3318 * subsystems can be enabled and disabled in a subtree using the following
3321 * 1. Call cgroup_save_control() to stash the current state.
3322 * 2. Update ->subtree_control masks in the subtree as desired.
3323 * 3. Call cgroup_apply_control() to apply the changes.
3324 * 4. Optionally perform other related operations.
3325 * 5. Call cgroup_finalize_control() to finish up.
3327 * This function implements step 3 and propagates the mask changes
3328 * throughout @cgrp's subtree, updates csses accordingly and perform
3329 * process migrations.
3331 static int cgroup_apply_control(struct cgroup *cgrp)
3335 cgroup_propagate_control(cgrp);
3337 ret = cgroup_apply_control_enable(cgrp);
3342 * At this point, cgroup_e_css() results reflect the new csses
3343 * making the following cgroup_update_dfl_csses() properly update
3344 * css associations of all tasks in the subtree.
3346 ret = cgroup_update_dfl_csses(cgrp);
3354 * cgroup_finalize_control - finalize control mask update
3355 * @cgrp: root of the target subtree
3356 * @ret: the result of the update
3358 * Finalize control mask update. See cgroup_apply_control() for more info.
3360 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3363 cgroup_restore_control(cgrp);
3364 cgroup_propagate_control(cgrp);
3367 cgroup_apply_control_disable(cgrp);
3370 /* change the enabled child controllers for a cgroup in the default hierarchy */
3371 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3372 char *buf, size_t nbytes,
3375 u16 enable = 0, disable = 0;
3376 struct cgroup *cgrp, *child;
3377 struct cgroup_subsys *ss;
3382 * Parse input - space separated list of subsystem names prefixed
3383 * with either + or -.
3385 buf = strstrip(buf);
3386 while ((tok = strsep(&buf, " "))) {
3389 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3390 if (!cgroup_ssid_enabled(ssid) ||
3391 strcmp(tok + 1, ss->name))
3395 enable |= 1 << ssid;
3396 disable &= ~(1 << ssid);
3397 } else if (*tok == '-') {
3398 disable |= 1 << ssid;
3399 enable &= ~(1 << ssid);
3404 } while_each_subsys_mask();
3405 if (ssid == CGROUP_SUBSYS_COUNT)
3409 cgrp = cgroup_kn_lock_live(of->kn, true);
3413 for_each_subsys(ss, ssid) {
3414 if (enable & (1 << ssid)) {
3415 if (cgrp->subtree_control & (1 << ssid)) {
3416 enable &= ~(1 << ssid);
3420 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3424 } else if (disable & (1 << ssid)) {
3425 if (!(cgrp->subtree_control & (1 << ssid))) {
3426 disable &= ~(1 << ssid);
3430 /* a child has it enabled? */
3431 cgroup_for_each_live_child(child, cgrp) {
3432 if (child->subtree_control & (1 << ssid)) {
3440 if (!enable && !disable) {
3446 * Except for the root, subtree_control must be zero for a cgroup
3447 * with tasks so that child cgroups don't compete against tasks.
3449 if (enable && cgroup_parent(cgrp)) {
3450 struct cgrp_cset_link *link;
3453 * Because namespaces pin csets too, @cgrp->cset_links
3454 * might not be empty even when @cgrp is empty. Walk and
3457 spin_lock_irq(&css_set_lock);
3460 list_for_each_entry(link, &cgrp->cset_links, cset_link) {
3461 if (css_set_populated(link->cset)) {
3467 spin_unlock_irq(&css_set_lock);
3473 /* save and update control masks and prepare csses */
3474 cgroup_save_control(cgrp);
3476 cgrp->subtree_control |= enable;
3477 cgrp->subtree_control &= ~disable;
3479 ret = cgroup_apply_control(cgrp);
3481 cgroup_finalize_control(cgrp, ret);
3483 kernfs_activate(cgrp->kn);
3486 cgroup_kn_unlock(of->kn);
3487 return ret ?: nbytes;
3490 static int cgroup_events_show(struct seq_file *seq, void *v)
3492 seq_printf(seq, "populated %d\n",
3493 cgroup_is_populated(seq_css(seq)->cgroup));
3497 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3498 size_t nbytes, loff_t off)
3500 struct cgroup *cgrp = of->kn->parent->priv;
3501 struct cftype *cft = of->kn->priv;
3502 struct cgroup_subsys_state *css;
3506 return cft->write(of, buf, nbytes, off);
3509 * kernfs guarantees that a file isn't deleted with operations in
3510 * flight, which means that the matching css is and stays alive and
3511 * doesn't need to be pinned. The RCU locking is not necessary
3512 * either. It's just for the convenience of using cgroup_css().
3515 css = cgroup_css(cgrp, cft->ss);
3518 if (cft->write_u64) {
3519 unsigned long long v;
3520 ret = kstrtoull(buf, 0, &v);
3522 ret = cft->write_u64(css, cft, v);
3523 } else if (cft->write_s64) {
3525 ret = kstrtoll(buf, 0, &v);
3527 ret = cft->write_s64(css, cft, v);
3532 return ret ?: nbytes;
3535 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3537 return seq_cft(seq)->seq_start(seq, ppos);
3540 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3542 return seq_cft(seq)->seq_next(seq, v, ppos);
3545 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3547 seq_cft(seq)->seq_stop(seq, v);
3550 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3552 struct cftype *cft = seq_cft(m);
3553 struct cgroup_subsys_state *css = seq_css(m);
3556 return cft->seq_show(m, arg);
3559 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3560 else if (cft->read_s64)
3561 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3567 static struct kernfs_ops cgroup_kf_single_ops = {
3568 .atomic_write_len = PAGE_SIZE,
3569 .write = cgroup_file_write,
3570 .seq_show = cgroup_seqfile_show,
3573 static struct kernfs_ops cgroup_kf_ops = {
3574 .atomic_write_len = PAGE_SIZE,
3575 .write = cgroup_file_write,
3576 .seq_start = cgroup_seqfile_start,
3577 .seq_next = cgroup_seqfile_next,
3578 .seq_stop = cgroup_seqfile_stop,
3579 .seq_show = cgroup_seqfile_show,
3583 * cgroup_rename - Only allow simple rename of directories in place.
3585 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3586 const char *new_name_str)
3588 struct cgroup *cgrp = kn->priv;
3591 if (kernfs_type(kn) != KERNFS_DIR)
3593 if (kn->parent != new_parent)
3597 * This isn't a proper migration and its usefulness is very
3598 * limited. Disallow on the default hierarchy.
3600 if (cgroup_on_dfl(cgrp))
3604 * We're gonna grab cgroup_mutex which nests outside kernfs
3605 * active_ref. kernfs_rename() doesn't require active_ref
3606 * protection. Break them before grabbing cgroup_mutex.
3608 kernfs_break_active_protection(new_parent);
3609 kernfs_break_active_protection(kn);
3611 mutex_lock(&cgroup_mutex);
3613 ret = kernfs_rename(kn, new_parent, new_name_str);
3615 mutex_unlock(&cgroup_mutex);
3617 kernfs_unbreak_active_protection(kn);
3618 kernfs_unbreak_active_protection(new_parent);
3622 /* set uid and gid of cgroup dirs and files to that of the creator */
3623 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3625 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3626 .ia_uid = current_fsuid(),
3627 .ia_gid = current_fsgid(), };
3629 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3630 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3633 return kernfs_setattr(kn, &iattr);
3636 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3639 char name[CGROUP_FILE_NAME_MAX];
3640 struct kernfs_node *kn;
3641 struct lock_class_key *key = NULL;
3644 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3645 key = &cft->lockdep_key;
3647 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3648 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3653 ret = cgroup_kn_set_ugid(kn);
3659 if (cft->file_offset) {
3660 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3662 spin_lock_irq(&cgroup_file_kn_lock);
3664 spin_unlock_irq(&cgroup_file_kn_lock);
3671 * cgroup_addrm_files - add or remove files to a cgroup directory
3672 * @css: the target css
3673 * @cgrp: the target cgroup (usually css->cgroup)
3674 * @cfts: array of cftypes to be added
3675 * @is_add: whether to add or remove
3677 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3678 * For removals, this function never fails.
3680 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3681 struct cgroup *cgrp, struct cftype cfts[],
3684 struct cftype *cft, *cft_end = NULL;
3687 lockdep_assert_held(&cgroup_mutex);
3690 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3691 /* does cft->flags tell us to skip this file on @cgrp? */
3692 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3694 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3696 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3698 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3702 ret = cgroup_add_file(css, cgrp, cft);
3704 pr_warn("%s: failed to add %s, err=%d\n",
3705 __func__, cft->name, ret);
3711 cgroup_rm_file(cgrp, cft);
3717 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3720 struct cgroup_subsys *ss = cfts[0].ss;
3721 struct cgroup *root = &ss->root->cgrp;
3722 struct cgroup_subsys_state *css;
3725 lockdep_assert_held(&cgroup_mutex);
3727 /* add/rm files for all cgroups created before */
3728 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3729 struct cgroup *cgrp = css->cgroup;
3731 if (!(css->flags & CSS_VISIBLE))
3734 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3740 kernfs_activate(root->kn);
3744 static void cgroup_exit_cftypes(struct cftype *cfts)
3748 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3749 /* free copy for custom atomic_write_len, see init_cftypes() */
3750 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3755 /* revert flags set by cgroup core while adding @cfts */
3756 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3760 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3764 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3765 struct kernfs_ops *kf_ops;
3767 WARN_ON(cft->ss || cft->kf_ops);
3770 kf_ops = &cgroup_kf_ops;
3772 kf_ops = &cgroup_kf_single_ops;
3775 * Ugh... if @cft wants a custom max_write_len, we need to
3776 * make a copy of kf_ops to set its atomic_write_len.
3778 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3779 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3781 cgroup_exit_cftypes(cfts);
3784 kf_ops->atomic_write_len = cft->max_write_len;
3787 cft->kf_ops = kf_ops;
3794 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3796 lockdep_assert_held(&cgroup_mutex);
3798 if (!cfts || !cfts[0].ss)
3801 list_del(&cfts->node);
3802 cgroup_apply_cftypes(cfts, false);
3803 cgroup_exit_cftypes(cfts);
3808 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3809 * @cfts: zero-length name terminated array of cftypes
3811 * Unregister @cfts. Files described by @cfts are removed from all
3812 * existing cgroups and all future cgroups won't have them either. This
3813 * function can be called anytime whether @cfts' subsys is attached or not.
3815 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3818 int cgroup_rm_cftypes(struct cftype *cfts)
3822 mutex_lock(&cgroup_mutex);
3823 ret = cgroup_rm_cftypes_locked(cfts);
3824 mutex_unlock(&cgroup_mutex);
3829 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3830 * @ss: target cgroup subsystem
3831 * @cfts: zero-length name terminated array of cftypes
3833 * Register @cfts to @ss. Files described by @cfts are created for all
3834 * existing cgroups to which @ss is attached and all future cgroups will
3835 * have them too. This function can be called anytime whether @ss is
3838 * Returns 0 on successful registration, -errno on failure. Note that this
3839 * function currently returns 0 as long as @cfts registration is successful
3840 * even if some file creation attempts on existing cgroups fail.
3842 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3846 if (!cgroup_ssid_enabled(ss->id))
3849 if (!cfts || cfts[0].name[0] == '\0')
3852 ret = cgroup_init_cftypes(ss, cfts);
3856 mutex_lock(&cgroup_mutex);
3858 list_add_tail(&cfts->node, &ss->cfts);
3859 ret = cgroup_apply_cftypes(cfts, true);
3861 cgroup_rm_cftypes_locked(cfts);
3863 mutex_unlock(&cgroup_mutex);
3868 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3869 * @ss: target cgroup subsystem
3870 * @cfts: zero-length name terminated array of cftypes
3872 * Similar to cgroup_add_cftypes() but the added files are only used for
3873 * the default hierarchy.
3875 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3879 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3880 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3881 return cgroup_add_cftypes(ss, cfts);
3885 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3886 * @ss: target cgroup subsystem
3887 * @cfts: zero-length name terminated array of cftypes
3889 * Similar to cgroup_add_cftypes() but the added files are only used for
3890 * the legacy hierarchies.
3892 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3896 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3897 cft->flags |= __CFTYPE_NOT_ON_DFL;
3898 return cgroup_add_cftypes(ss, cfts);
3902 * cgroup_file_notify - generate a file modified event for a cgroup_file
3903 * @cfile: target cgroup_file
3905 * @cfile must have been obtained by setting cftype->file_offset.
3907 void cgroup_file_notify(struct cgroup_file *cfile)
3909 unsigned long flags;
3911 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3913 kernfs_notify(cfile->kn);
3914 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3918 * cgroup_task_count - count the number of tasks in a cgroup.
3919 * @cgrp: the cgroup in question
3921 * Return the number of tasks in the cgroup. The returned number can be
3922 * higher than the actual number of tasks due to css_set references from
3923 * namespace roots and temporary usages.
3925 static int cgroup_task_count(const struct cgroup *cgrp)
3928 struct cgrp_cset_link *link;
3930 spin_lock_irq(&css_set_lock);
3931 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3932 count += atomic_read(&link->cset->refcount);
3933 spin_unlock_irq(&css_set_lock);
3938 * css_next_child - find the next child of a given css
3939 * @pos: the current position (%NULL to initiate traversal)
3940 * @parent: css whose children to walk
3942 * This function returns the next child of @parent and should be called
3943 * under either cgroup_mutex or RCU read lock. The only requirement is
3944 * that @parent and @pos are accessible. The next sibling is guaranteed to
3945 * be returned regardless of their states.
3947 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3948 * css which finished ->css_online() is guaranteed to be visible in the
3949 * future iterations and will stay visible until the last reference is put.
3950 * A css which hasn't finished ->css_online() or already finished
3951 * ->css_offline() may show up during traversal. It's each subsystem's
3952 * responsibility to synchronize against on/offlining.
3954 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3955 struct cgroup_subsys_state *parent)
3957 struct cgroup_subsys_state *next;
3959 cgroup_assert_mutex_or_rcu_locked();
3962 * @pos could already have been unlinked from the sibling list.
3963 * Once a cgroup is removed, its ->sibling.next is no longer
3964 * updated when its next sibling changes. CSS_RELEASED is set when
3965 * @pos is taken off list, at which time its next pointer is valid,
3966 * and, as releases are serialized, the one pointed to by the next
3967 * pointer is guaranteed to not have started release yet. This
3968 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3969 * critical section, the one pointed to by its next pointer is
3970 * guaranteed to not have finished its RCU grace period even if we
3971 * have dropped rcu_read_lock() inbetween iterations.
3973 * If @pos has CSS_RELEASED set, its next pointer can't be
3974 * dereferenced; however, as each css is given a monotonically
3975 * increasing unique serial number and always appended to the
3976 * sibling list, the next one can be found by walking the parent's
3977 * children until the first css with higher serial number than
3978 * @pos's. While this path can be slower, it happens iff iteration
3979 * races against release and the race window is very small.
3982 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3983 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3984 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3986 list_for_each_entry_rcu(next, &parent->children, sibling)
3987 if (next->serial_nr > pos->serial_nr)
3992 * @next, if not pointing to the head, can be dereferenced and is
3995 if (&next->sibling != &parent->children)
4001 * css_next_descendant_pre - find the next descendant for pre-order walk
4002 * @pos: the current position (%NULL to initiate traversal)
4003 * @root: css whose descendants to walk
4005 * To be used by css_for_each_descendant_pre(). Find the next descendant
4006 * to visit for pre-order traversal of @root's descendants. @root is
4007 * included in the iteration and the first node to be visited.
4009 * While this function requires cgroup_mutex or RCU read locking, it
4010 * doesn't require the whole traversal to be contained in a single critical
4011 * section. This function will return the correct next descendant as long
4012 * as both @pos and @root are accessible and @pos is a descendant of @root.
4014 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4015 * css which finished ->css_online() is guaranteed to be visible in the
4016 * future iterations and will stay visible until the last reference is put.
4017 * A css which hasn't finished ->css_online() or already finished
4018 * ->css_offline() may show up during traversal. It's each subsystem's
4019 * responsibility to synchronize against on/offlining.
4021 struct cgroup_subsys_state *
4022 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4023 struct cgroup_subsys_state *root)
4025 struct cgroup_subsys_state *next;
4027 cgroup_assert_mutex_or_rcu_locked();
4029 /* if first iteration, visit @root */
4033 /* visit the first child if exists */
4034 next = css_next_child(NULL, pos);
4038 /* no child, visit my or the closest ancestor's next sibling */
4039 while (pos != root) {
4040 next = css_next_child(pos, pos->parent);
4050 * css_rightmost_descendant - return the rightmost descendant of a css
4051 * @pos: css of interest
4053 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4054 * is returned. This can be used during pre-order traversal to skip
4057 * While this function requires cgroup_mutex or RCU read locking, it
4058 * doesn't require the whole traversal to be contained in a single critical
4059 * section. This function will return the correct rightmost descendant as
4060 * long as @pos is accessible.
4062 struct cgroup_subsys_state *
4063 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4065 struct cgroup_subsys_state *last, *tmp;
4067 cgroup_assert_mutex_or_rcu_locked();
4071 /* ->prev isn't RCU safe, walk ->next till the end */
4073 css_for_each_child(tmp, last)
4080 static struct cgroup_subsys_state *
4081 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4083 struct cgroup_subsys_state *last;
4087 pos = css_next_child(NULL, pos);
4094 * css_next_descendant_post - find the next descendant for post-order walk
4095 * @pos: the current position (%NULL to initiate traversal)
4096 * @root: css whose descendants to walk
4098 * To be used by css_for_each_descendant_post(). Find the next descendant
4099 * to visit for post-order traversal of @root's descendants. @root is
4100 * included in the iteration and the last node to be visited.
4102 * While this function requires cgroup_mutex or RCU read locking, it
4103 * doesn't require the whole traversal to be contained in a single critical
4104 * section. This function will return the correct next descendant as long
4105 * as both @pos and @cgroup are accessible and @pos is a descendant of
4108 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4109 * css which finished ->css_online() is guaranteed to be visible in the
4110 * future iterations and will stay visible until the last reference is put.
4111 * A css which hasn't finished ->css_online() or already finished
4112 * ->css_offline() may show up during traversal. It's each subsystem's
4113 * responsibility to synchronize against on/offlining.
4115 struct cgroup_subsys_state *
4116 css_next_descendant_post(struct cgroup_subsys_state *pos,
4117 struct cgroup_subsys_state *root)
4119 struct cgroup_subsys_state *next;
4121 cgroup_assert_mutex_or_rcu_locked();
4123 /* if first iteration, visit leftmost descendant which may be @root */
4125 return css_leftmost_descendant(root);
4127 /* if we visited @root, we're done */
4131 /* if there's an unvisited sibling, visit its leftmost descendant */
4132 next = css_next_child(pos, pos->parent);
4134 return css_leftmost_descendant(next);
4136 /* no sibling left, visit parent */
4141 * css_has_online_children - does a css have online children
4142 * @css: the target css
4144 * Returns %true if @css has any online children; otherwise, %false. This
4145 * function can be called from any context but the caller is responsible
4146 * for synchronizing against on/offlining as necessary.
4148 bool css_has_online_children(struct cgroup_subsys_state *css)
4150 struct cgroup_subsys_state *child;
4154 css_for_each_child(child, css) {
4155 if (child->flags & CSS_ONLINE) {
4165 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4166 * @it: the iterator to advance
4168 * Advance @it to the next css_set to walk.
4170 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4172 struct list_head *l = it->cset_pos;
4173 struct cgrp_cset_link *link;
4174 struct css_set *cset;
4176 lockdep_assert_held(&css_set_lock);
4178 /* Advance to the next non-empty css_set */
4181 if (l == it->cset_head) {
4182 it->cset_pos = NULL;
4183 it->task_pos = NULL;
4188 cset = container_of(l, struct css_set,
4189 e_cset_node[it->ss->id]);
4191 link = list_entry(l, struct cgrp_cset_link, cset_link);
4194 } while (!css_set_populated(cset));
4198 if (!list_empty(&cset->tasks))
4199 it->task_pos = cset->tasks.next;
4201 it->task_pos = cset->mg_tasks.next;
4203 it->tasks_head = &cset->tasks;
4204 it->mg_tasks_head = &cset->mg_tasks;
4207 * We don't keep css_sets locked across iteration steps and thus
4208 * need to take steps to ensure that iteration can be resumed after
4209 * the lock is re-acquired. Iteration is performed at two levels -
4210 * css_sets and tasks in them.
4212 * Once created, a css_set never leaves its cgroup lists, so a
4213 * pinned css_set is guaranteed to stay put and we can resume
4214 * iteration afterwards.
4216 * Tasks may leave @cset across iteration steps. This is resolved
4217 * by registering each iterator with the css_set currently being
4218 * walked and making css_set_move_task() advance iterators whose
4219 * next task is leaving.
4222 list_del(&it->iters_node);
4223 put_css_set_locked(it->cur_cset);
4226 it->cur_cset = cset;
4227 list_add(&it->iters_node, &cset->task_iters);
4230 static void css_task_iter_advance(struct css_task_iter *it)
4232 struct list_head *l = it->task_pos;
4234 lockdep_assert_held(&css_set_lock);
4238 * Advance iterator to find next entry. cset->tasks is consumed
4239 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4244 if (l == it->tasks_head)
4245 l = it->mg_tasks_head->next;
4247 if (l == it->mg_tasks_head)
4248 css_task_iter_advance_css_set(it);
4254 * css_task_iter_start - initiate task iteration
4255 * @css: the css to walk tasks of
4256 * @it: the task iterator to use
4258 * Initiate iteration through the tasks of @css. The caller can call
4259 * css_task_iter_next() to walk through the tasks until the function
4260 * returns NULL. On completion of iteration, css_task_iter_end() must be
4263 void css_task_iter_start(struct cgroup_subsys_state *css,
4264 struct css_task_iter *it)
4266 /* no one should try to iterate before mounting cgroups */
4267 WARN_ON_ONCE(!use_task_css_set_links);
4269 memset(it, 0, sizeof(*it));
4271 spin_lock_irq(&css_set_lock);
4276 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4278 it->cset_pos = &css->cgroup->cset_links;
4280 it->cset_head = it->cset_pos;
4282 css_task_iter_advance_css_set(it);
4284 spin_unlock_irq(&css_set_lock);
4288 * css_task_iter_next - return the next task for the iterator
4289 * @it: the task iterator being iterated
4291 * The "next" function for task iteration. @it should have been
4292 * initialized via css_task_iter_start(). Returns NULL when the iteration
4295 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4298 put_task_struct(it->cur_task);
4299 it->cur_task = NULL;
4302 spin_lock_irq(&css_set_lock);
4305 it->cur_task = list_entry(it->task_pos, struct task_struct,
4307 get_task_struct(it->cur_task);
4308 css_task_iter_advance(it);
4311 spin_unlock_irq(&css_set_lock);
4313 return it->cur_task;
4317 * css_task_iter_end - finish task iteration
4318 * @it: the task iterator to finish
4320 * Finish task iteration started by css_task_iter_start().
4322 void css_task_iter_end(struct css_task_iter *it)
4325 spin_lock_irq(&css_set_lock);
4326 list_del(&it->iters_node);
4327 put_css_set_locked(it->cur_cset);
4328 spin_unlock_irq(&css_set_lock);
4332 put_task_struct(it->cur_task);
4336 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4337 * @to: cgroup to which the tasks will be moved
4338 * @from: cgroup in which the tasks currently reside
4340 * Locking rules between cgroup_post_fork() and the migration path
4341 * guarantee that, if a task is forking while being migrated, the new child
4342 * is guaranteed to be either visible in the source cgroup after the
4343 * parent's migration is complete or put into the target cgroup. No task
4344 * can slip out of migration through forking.
4346 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4348 LIST_HEAD(preloaded_csets);
4349 struct cgrp_cset_link *link;
4350 struct css_task_iter it;
4351 struct task_struct *task;
4354 if (!cgroup_may_migrate_to(to))
4357 mutex_lock(&cgroup_mutex);
4359 percpu_down_write(&cgroup_threadgroup_rwsem);
4361 /* all tasks in @from are being moved, all csets are source */
4362 spin_lock_irq(&css_set_lock);
4363 list_for_each_entry(link, &from->cset_links, cset_link)
4364 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4365 spin_unlock_irq(&css_set_lock);
4367 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
4372 * Migrate tasks one-by-one until @from is empty. This fails iff
4373 * ->can_attach() fails.
4376 css_task_iter_start(&from->self, &it);
4377 task = css_task_iter_next(&it);
4379 get_task_struct(task);
4380 css_task_iter_end(&it);
4383 ret = cgroup_migrate(task, false, to->root);
4384 put_task_struct(task);
4386 } while (task && !ret);
4388 cgroup_migrate_finish(&preloaded_csets);
4389 percpu_up_write(&cgroup_threadgroup_rwsem);
4390 mutex_unlock(&cgroup_mutex);
4395 * Stuff for reading the 'tasks'/'procs' files.
4397 * Reading this file can return large amounts of data if a cgroup has
4398 * *lots* of attached tasks. So it may need several calls to read(),
4399 * but we cannot guarantee that the information we produce is correct
4400 * unless we produce it entirely atomically.
4404 /* which pidlist file are we talking about? */
4405 enum cgroup_filetype {
4411 * A pidlist is a list of pids that virtually represents the contents of one
4412 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4413 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4416 struct cgroup_pidlist {
4418 * used to find which pidlist is wanted. doesn't change as long as
4419 * this particular list stays in the list.
4421 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4424 /* how many elements the above list has */
4426 /* each of these stored in a list by its cgroup */
4427 struct list_head links;
4428 /* pointer to the cgroup we belong to, for list removal purposes */
4429 struct cgroup *owner;
4430 /* for delayed destruction */
4431 struct delayed_work destroy_dwork;
4435 * The following two functions "fix" the issue where there are more pids
4436 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4437 * TODO: replace with a kernel-wide solution to this problem
4439 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4440 static void *pidlist_allocate(int count)
4442 if (PIDLIST_TOO_LARGE(count))
4443 return vmalloc(count * sizeof(pid_t));
4445 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4448 static void pidlist_free(void *p)
4454 * Used to destroy all pidlists lingering waiting for destroy timer. None
4455 * should be left afterwards.
4457 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4459 struct cgroup_pidlist *l, *tmp_l;
4461 mutex_lock(&cgrp->pidlist_mutex);
4462 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4463 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4464 mutex_unlock(&cgrp->pidlist_mutex);
4466 flush_workqueue(cgroup_pidlist_destroy_wq);
4467 BUG_ON(!list_empty(&cgrp->pidlists));
4470 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4472 struct delayed_work *dwork = to_delayed_work(work);
4473 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4475 struct cgroup_pidlist *tofree = NULL;
4477 mutex_lock(&l->owner->pidlist_mutex);
4480 * Destroy iff we didn't get queued again. The state won't change
4481 * as destroy_dwork can only be queued while locked.
4483 if (!delayed_work_pending(dwork)) {
4484 list_del(&l->links);
4485 pidlist_free(l->list);
4486 put_pid_ns(l->key.ns);
4490 mutex_unlock(&l->owner->pidlist_mutex);
4495 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4496 * Returns the number of unique elements.
4498 static int pidlist_uniq(pid_t *list, int length)
4503 * we presume the 0th element is unique, so i starts at 1. trivial
4504 * edge cases first; no work needs to be done for either
4506 if (length == 0 || length == 1)
4508 /* src and dest walk down the list; dest counts unique elements */
4509 for (src = 1; src < length; src++) {
4510 /* find next unique element */
4511 while (list[src] == list[src-1]) {
4516 /* dest always points to where the next unique element goes */
4517 list[dest] = list[src];
4525 * The two pid files - task and cgroup.procs - guaranteed that the result
4526 * is sorted, which forced this whole pidlist fiasco. As pid order is
4527 * different per namespace, each namespace needs differently sorted list,
4528 * making it impossible to use, for example, single rbtree of member tasks
4529 * sorted by task pointer. As pidlists can be fairly large, allocating one
4530 * per open file is dangerous, so cgroup had to implement shared pool of
4531 * pidlists keyed by cgroup and namespace.
4533 * All this extra complexity was caused by the original implementation
4534 * committing to an entirely unnecessary property. In the long term, we
4535 * want to do away with it. Explicitly scramble sort order if on the
4536 * default hierarchy so that no such expectation exists in the new
4539 * Scrambling is done by swapping every two consecutive bits, which is
4540 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4542 static pid_t pid_fry(pid_t pid)
4544 unsigned a = pid & 0x55555555;
4545 unsigned b = pid & 0xAAAAAAAA;
4547 return (a << 1) | (b >> 1);
4550 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4552 if (cgroup_on_dfl(cgrp))
4553 return pid_fry(pid);
4558 static int cmppid(const void *a, const void *b)
4560 return *(pid_t *)a - *(pid_t *)b;
4563 static int fried_cmppid(const void *a, const void *b)
4565 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4568 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4569 enum cgroup_filetype type)
4571 struct cgroup_pidlist *l;
4572 /* don't need task_nsproxy() if we're looking at ourself */
4573 struct pid_namespace *ns = task_active_pid_ns(current);
4575 lockdep_assert_held(&cgrp->pidlist_mutex);
4577 list_for_each_entry(l, &cgrp->pidlists, links)
4578 if (l->key.type == type && l->key.ns == ns)
4584 * find the appropriate pidlist for our purpose (given procs vs tasks)
4585 * returns with the lock on that pidlist already held, and takes care
4586 * of the use count, or returns NULL with no locks held if we're out of
4589 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4590 enum cgroup_filetype type)
4592 struct cgroup_pidlist *l;
4594 lockdep_assert_held(&cgrp->pidlist_mutex);
4596 l = cgroup_pidlist_find(cgrp, type);
4600 /* entry not found; create a new one */
4601 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4605 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4607 /* don't need task_nsproxy() if we're looking at ourself */
4608 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4610 list_add(&l->links, &cgrp->pidlists);
4615 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4617 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4618 struct cgroup_pidlist **lp)
4622 int pid, n = 0; /* used for populating the array */
4623 struct css_task_iter it;
4624 struct task_struct *tsk;
4625 struct cgroup_pidlist *l;
4627 lockdep_assert_held(&cgrp->pidlist_mutex);
4630 * If cgroup gets more users after we read count, we won't have
4631 * enough space - tough. This race is indistinguishable to the
4632 * caller from the case that the additional cgroup users didn't
4633 * show up until sometime later on.
4635 length = cgroup_task_count(cgrp);
4636 array = pidlist_allocate(length);
4639 /* now, populate the array */
4640 css_task_iter_start(&cgrp->self, &it);
4641 while ((tsk = css_task_iter_next(&it))) {
4642 if (unlikely(n == length))
4644 /* get tgid or pid for procs or tasks file respectively */
4645 if (type == CGROUP_FILE_PROCS)
4646 pid = task_tgid_vnr(tsk);
4648 pid = task_pid_vnr(tsk);
4649 if (pid > 0) /* make sure to only use valid results */
4652 css_task_iter_end(&it);
4654 /* now sort & (if procs) strip out duplicates */
4655 if (cgroup_on_dfl(cgrp))
4656 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4658 sort(array, length, sizeof(pid_t), cmppid, NULL);
4659 if (type == CGROUP_FILE_PROCS)
4660 length = pidlist_uniq(array, length);
4662 l = cgroup_pidlist_find_create(cgrp, type);
4664 pidlist_free(array);
4668 /* store array, freeing old if necessary */
4669 pidlist_free(l->list);
4677 * cgroupstats_build - build and fill cgroupstats
4678 * @stats: cgroupstats to fill information into
4679 * @dentry: A dentry entry belonging to the cgroup for which stats have
4682 * Build and fill cgroupstats so that taskstats can export it to user
4685 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4687 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4688 struct cgroup *cgrp;
4689 struct css_task_iter it;
4690 struct task_struct *tsk;
4692 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4693 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4694 kernfs_type(kn) != KERNFS_DIR)
4697 mutex_lock(&cgroup_mutex);
4700 * We aren't being called from kernfs and there's no guarantee on
4701 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4702 * @kn->priv is RCU safe. Let's do the RCU dancing.
4705 cgrp = rcu_dereference(kn->priv);
4706 if (!cgrp || cgroup_is_dead(cgrp)) {
4708 mutex_unlock(&cgroup_mutex);
4713 css_task_iter_start(&cgrp->self, &it);
4714 while ((tsk = css_task_iter_next(&it))) {
4715 switch (tsk->state) {
4717 stats->nr_running++;
4719 case TASK_INTERRUPTIBLE:
4720 stats->nr_sleeping++;
4722 case TASK_UNINTERRUPTIBLE:
4723 stats->nr_uninterruptible++;
4726 stats->nr_stopped++;
4729 if (delayacct_is_task_waiting_on_io(tsk))
4730 stats->nr_io_wait++;
4734 css_task_iter_end(&it);
4736 mutex_unlock(&cgroup_mutex);
4742 * seq_file methods for the tasks/procs files. The seq_file position is the
4743 * next pid to display; the seq_file iterator is a pointer to the pid
4744 * in the cgroup->l->list array.
4747 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4750 * Initially we receive a position value that corresponds to
4751 * one more than the last pid shown (or 0 on the first call or
4752 * after a seek to the start). Use a binary-search to find the
4753 * next pid to display, if any
4755 struct kernfs_open_file *of = s->private;
4756 struct cgroup *cgrp = seq_css(s)->cgroup;
4757 struct cgroup_pidlist *l;
4758 enum cgroup_filetype type = seq_cft(s)->private;
4759 int index = 0, pid = *pos;
4762 mutex_lock(&cgrp->pidlist_mutex);
4765 * !NULL @of->priv indicates that this isn't the first start()
4766 * after open. If the matching pidlist is around, we can use that.
4767 * Look for it. Note that @of->priv can't be used directly. It
4768 * could already have been destroyed.
4771 of->priv = cgroup_pidlist_find(cgrp, type);
4774 * Either this is the first start() after open or the matching
4775 * pidlist has been destroyed inbetween. Create a new one.
4778 ret = pidlist_array_load(cgrp, type,
4779 (struct cgroup_pidlist **)&of->priv);
4781 return ERR_PTR(ret);
4786 int end = l->length;
4788 while (index < end) {
4789 int mid = (index + end) / 2;
4790 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4793 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4799 /* If we're off the end of the array, we're done */
4800 if (index >= l->length)
4802 /* Update the abstract position to be the actual pid that we found */
4803 iter = l->list + index;
4804 *pos = cgroup_pid_fry(cgrp, *iter);
4808 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4810 struct kernfs_open_file *of = s->private;
4811 struct cgroup_pidlist *l = of->priv;
4814 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4815 CGROUP_PIDLIST_DESTROY_DELAY);
4816 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4819 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4821 struct kernfs_open_file *of = s->private;
4822 struct cgroup_pidlist *l = of->priv;
4824 pid_t *end = l->list + l->length;
4826 * Advance to the next pid in the array. If this goes off the
4833 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4838 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4840 seq_printf(s, "%d\n", *(int *)v);
4845 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4848 return notify_on_release(css->cgroup);
4851 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4852 struct cftype *cft, u64 val)
4855 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4857 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4861 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4864 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4867 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4868 struct cftype *cft, u64 val)
4871 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4873 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4877 /* cgroup core interface files for the default hierarchy */
4878 static struct cftype cgroup_dfl_base_files[] = {
4880 .name = "cgroup.procs",
4881 .file_offset = offsetof(struct cgroup, procs_file),
4882 .seq_start = cgroup_pidlist_start,
4883 .seq_next = cgroup_pidlist_next,
4884 .seq_stop = cgroup_pidlist_stop,
4885 .seq_show = cgroup_pidlist_show,
4886 .private = CGROUP_FILE_PROCS,
4887 .write = cgroup_procs_write,
4890 .name = "cgroup.controllers",
4891 .seq_show = cgroup_controllers_show,
4894 .name = "cgroup.subtree_control",
4895 .seq_show = cgroup_subtree_control_show,
4896 .write = cgroup_subtree_control_write,
4899 .name = "cgroup.events",
4900 .flags = CFTYPE_NOT_ON_ROOT,
4901 .file_offset = offsetof(struct cgroup, events_file),
4902 .seq_show = cgroup_events_show,
4907 /* cgroup core interface files for the legacy hierarchies */
4908 static struct cftype cgroup_legacy_base_files[] = {
4910 .name = "cgroup.procs",
4911 .seq_start = cgroup_pidlist_start,
4912 .seq_next = cgroup_pidlist_next,
4913 .seq_stop = cgroup_pidlist_stop,
4914 .seq_show = cgroup_pidlist_show,
4915 .private = CGROUP_FILE_PROCS,
4916 .write = cgroup_procs_write,
4919 .name = "cgroup.clone_children",
4920 .read_u64 = cgroup_clone_children_read,
4921 .write_u64 = cgroup_clone_children_write,
4924 .name = "cgroup.sane_behavior",
4925 .flags = CFTYPE_ONLY_ON_ROOT,
4926 .seq_show = cgroup_sane_behavior_show,
4930 .seq_start = cgroup_pidlist_start,
4931 .seq_next = cgroup_pidlist_next,
4932 .seq_stop = cgroup_pidlist_stop,
4933 .seq_show = cgroup_pidlist_show,
4934 .private = CGROUP_FILE_TASKS,
4935 .write = cgroup_tasks_write,
4938 .name = "notify_on_release",
4939 .read_u64 = cgroup_read_notify_on_release,
4940 .write_u64 = cgroup_write_notify_on_release,
4943 .name = "release_agent",
4944 .flags = CFTYPE_ONLY_ON_ROOT,
4945 .seq_show = cgroup_release_agent_show,
4946 .write = cgroup_release_agent_write,
4947 .max_write_len = PATH_MAX - 1,
4953 * css destruction is four-stage process.
4955 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4956 * Implemented in kill_css().
4958 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4959 * and thus css_tryget_online() is guaranteed to fail, the css can be
4960 * offlined by invoking offline_css(). After offlining, the base ref is
4961 * put. Implemented in css_killed_work_fn().
4963 * 3. When the percpu_ref reaches zero, the only possible remaining
4964 * accessors are inside RCU read sections. css_release() schedules the
4967 * 4. After the grace period, the css can be freed. Implemented in
4968 * css_free_work_fn().
4970 * It is actually hairier because both step 2 and 4 require process context
4971 * and thus involve punting to css->destroy_work adding two additional
4972 * steps to the already complex sequence.
4974 static void css_free_work_fn(struct work_struct *work)
4976 struct cgroup_subsys_state *css =
4977 container_of(work, struct cgroup_subsys_state, destroy_work);
4978 struct cgroup_subsys *ss = css->ss;
4979 struct cgroup *cgrp = css->cgroup;
4981 percpu_ref_exit(&css->refcnt);
4985 struct cgroup_subsys_state *parent = css->parent;
4989 cgroup_idr_remove(&ss->css_idr, id);
4995 /* cgroup free path */
4996 atomic_dec(&cgrp->root->nr_cgrps);
4997 cgroup_pidlist_destroy_all(cgrp);
4998 cancel_work_sync(&cgrp->release_agent_work);
5000 if (cgroup_parent(cgrp)) {
5002 * We get a ref to the parent, and put the ref when
5003 * this cgroup is being freed, so it's guaranteed
5004 * that the parent won't be destroyed before its
5007 cgroup_put(cgroup_parent(cgrp));
5008 kernfs_put(cgrp->kn);
5012 * This is root cgroup's refcnt reaching zero,
5013 * which indicates that the root should be
5016 cgroup_destroy_root(cgrp->root);
5021 static void css_free_rcu_fn(struct rcu_head *rcu_head)
5023 struct cgroup_subsys_state *css =
5024 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
5026 INIT_WORK(&css->destroy_work, css_free_work_fn);
5027 queue_work(cgroup_destroy_wq, &css->destroy_work);
5030 static void css_release_work_fn(struct work_struct *work)
5032 struct cgroup_subsys_state *css =
5033 container_of(work, struct cgroup_subsys_state, destroy_work);
5034 struct cgroup_subsys *ss = css->ss;
5035 struct cgroup *cgrp = css->cgroup;
5037 mutex_lock(&cgroup_mutex);
5039 css->flags |= CSS_RELEASED;
5040 list_del_rcu(&css->sibling);
5043 /* css release path */
5044 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5045 if (ss->css_released)
5046 ss->css_released(css);
5048 /* cgroup release path */
5049 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
5053 * There are two control paths which try to determine
5054 * cgroup from dentry without going through kernfs -
5055 * cgroupstats_build() and css_tryget_online_from_dir().
5056 * Those are supported by RCU protecting clearing of
5057 * cgrp->kn->priv backpointer.
5060 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5064 mutex_unlock(&cgroup_mutex);
5066 call_rcu(&css->rcu_head, css_free_rcu_fn);
5069 static void css_release(struct percpu_ref *ref)
5071 struct cgroup_subsys_state *css =
5072 container_of(ref, struct cgroup_subsys_state, refcnt);
5074 INIT_WORK(&css->destroy_work, css_release_work_fn);
5075 queue_work(cgroup_destroy_wq, &css->destroy_work);
5078 static void init_and_link_css(struct cgroup_subsys_state *css,
5079 struct cgroup_subsys *ss, struct cgroup *cgrp)
5081 lockdep_assert_held(&cgroup_mutex);
5085 memset(css, 0, sizeof(*css));
5089 INIT_LIST_HEAD(&css->sibling);
5090 INIT_LIST_HEAD(&css->children);
5091 css->serial_nr = css_serial_nr_next++;
5092 atomic_set(&css->online_cnt, 0);
5094 if (cgroup_parent(cgrp)) {
5095 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5096 css_get(css->parent);
5099 BUG_ON(cgroup_css(cgrp, ss));
5102 /* invoke ->css_online() on a new CSS and mark it online if successful */
5103 static int online_css(struct cgroup_subsys_state *css)
5105 struct cgroup_subsys *ss = css->ss;
5108 lockdep_assert_held(&cgroup_mutex);
5111 ret = ss->css_online(css);
5113 css->flags |= CSS_ONLINE;
5114 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5116 atomic_inc(&css->online_cnt);
5118 atomic_inc(&css->parent->online_cnt);
5123 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5124 static void offline_css(struct cgroup_subsys_state *css)
5126 struct cgroup_subsys *ss = css->ss;
5128 lockdep_assert_held(&cgroup_mutex);
5130 if (!(css->flags & CSS_ONLINE))
5136 if (ss->css_offline)
5137 ss->css_offline(css);
5139 css->flags &= ~CSS_ONLINE;
5140 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5142 wake_up_all(&css->cgroup->offline_waitq);
5146 * css_create - create a cgroup_subsys_state
5147 * @cgrp: the cgroup new css will be associated with
5148 * @ss: the subsys of new css
5150 * Create a new css associated with @cgrp - @ss pair. On success, the new
5151 * css is online and installed in @cgrp. This function doesn't create the
5152 * interface files. Returns 0 on success, -errno on failure.
5154 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5155 struct cgroup_subsys *ss)
5157 struct cgroup *parent = cgroup_parent(cgrp);
5158 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5159 struct cgroup_subsys_state *css;
5162 lockdep_assert_held(&cgroup_mutex);
5164 css = ss->css_alloc(parent_css);
5166 css = ERR_PTR(-ENOMEM);
5170 init_and_link_css(css, ss, cgrp);
5172 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5176 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5181 /* @css is ready to be brought online now, make it visible */
5182 list_add_tail_rcu(&css->sibling, &parent_css->children);
5183 cgroup_idr_replace(&ss->css_idr, css, css->id);
5185 err = online_css(css);
5189 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5190 cgroup_parent(parent)) {
5191 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5192 current->comm, current->pid, ss->name);
5193 if (!strcmp(ss->name, "memory"))
5194 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5195 ss->warned_broken_hierarchy = true;
5201 list_del_rcu(&css->sibling);
5203 call_rcu(&css->rcu_head, css_free_rcu_fn);
5204 return ERR_PTR(err);
5207 static struct cgroup *cgroup_create(struct cgroup *parent)
5209 struct cgroup_root *root = parent->root;
5210 struct cgroup *cgrp, *tcgrp;
5211 int level = parent->level + 1;
5214 /* allocate the cgroup and its ID, 0 is reserved for the root */
5215 cgrp = kzalloc(sizeof(*cgrp) +
5216 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
5218 return ERR_PTR(-ENOMEM);
5220 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5225 * Temporarily set the pointer to NULL, so idr_find() won't return
5226 * a half-baked cgroup.
5228 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
5231 goto out_cancel_ref;
5234 init_cgroup_housekeeping(cgrp);
5236 cgrp->self.parent = &parent->self;
5238 cgrp->level = level;
5240 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
5241 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
5243 if (notify_on_release(parent))
5244 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5246 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5247 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5249 cgrp->self.serial_nr = css_serial_nr_next++;
5251 /* allocation complete, commit to creation */
5252 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5253 atomic_inc(&root->nr_cgrps);
5257 * @cgrp is now fully operational. If something fails after this
5258 * point, it'll be released via the normal destruction path.
5260 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5263 * On the default hierarchy, a child doesn't automatically inherit
5264 * subtree_control from the parent. Each is configured manually.
5266 if (!cgroup_on_dfl(cgrp))
5267 cgrp->subtree_control = cgroup_control(cgrp);
5269 cgroup_propagate_control(cgrp);
5271 /* @cgrp doesn't have dir yet so the following will only create csses */
5272 ret = cgroup_apply_control_enable(cgrp);
5279 percpu_ref_exit(&cgrp->self.refcnt);
5282 return ERR_PTR(ret);
5284 cgroup_destroy_locked(cgrp);
5285 return ERR_PTR(ret);
5288 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
5291 struct cgroup *parent, *cgrp;
5292 struct kernfs_node *kn;
5295 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5296 if (strchr(name, '\n'))
5299 parent = cgroup_kn_lock_live(parent_kn, false);
5303 cgrp = cgroup_create(parent);
5305 ret = PTR_ERR(cgrp);
5309 /* create the directory */
5310 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5318 * This extra ref will be put in cgroup_free_fn() and guarantees
5319 * that @cgrp->kn is always accessible.
5323 ret = cgroup_kn_set_ugid(kn);
5327 ret = css_populate_dir(&cgrp->self);
5331 ret = cgroup_apply_control_enable(cgrp);
5335 /* let's create and online css's */
5336 kernfs_activate(kn);
5342 cgroup_destroy_locked(cgrp);
5344 cgroup_kn_unlock(parent_kn);
5349 * This is called when the refcnt of a css is confirmed to be killed.
5350 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5351 * initate destruction and put the css ref from kill_css().
5353 static void css_killed_work_fn(struct work_struct *work)
5355 struct cgroup_subsys_state *css =
5356 container_of(work, struct cgroup_subsys_state, destroy_work);
5358 mutex_lock(&cgroup_mutex);
5363 /* @css can't go away while we're holding cgroup_mutex */
5365 } while (css && atomic_dec_and_test(&css->online_cnt));
5367 mutex_unlock(&cgroup_mutex);
5370 /* css kill confirmation processing requires process context, bounce */
5371 static void css_killed_ref_fn(struct percpu_ref *ref)
5373 struct cgroup_subsys_state *css =
5374 container_of(ref, struct cgroup_subsys_state, refcnt);
5376 if (atomic_dec_and_test(&css->online_cnt)) {
5377 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5378 queue_work(cgroup_destroy_wq, &css->destroy_work);
5383 * kill_css - destroy a css
5384 * @css: css to destroy
5386 * This function initiates destruction of @css by removing cgroup interface
5387 * files and putting its base reference. ->css_offline() will be invoked
5388 * asynchronously once css_tryget_online() is guaranteed to fail and when
5389 * the reference count reaches zero, @css will be released.
5391 static void kill_css(struct cgroup_subsys_state *css)
5393 lockdep_assert_held(&cgroup_mutex);
5396 * This must happen before css is disassociated with its cgroup.
5397 * See seq_css() for details.
5402 * Killing would put the base ref, but we need to keep it alive
5403 * until after ->css_offline().
5408 * cgroup core guarantees that, by the time ->css_offline() is
5409 * invoked, no new css reference will be given out via
5410 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5411 * proceed to offlining css's because percpu_ref_kill() doesn't
5412 * guarantee that the ref is seen as killed on all CPUs on return.
5414 * Use percpu_ref_kill_and_confirm() to get notifications as each
5415 * css is confirmed to be seen as killed on all CPUs.
5417 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5421 * cgroup_destroy_locked - the first stage of cgroup destruction
5422 * @cgrp: cgroup to be destroyed
5424 * css's make use of percpu refcnts whose killing latency shouldn't be
5425 * exposed to userland and are RCU protected. Also, cgroup core needs to
5426 * guarantee that css_tryget_online() won't succeed by the time
5427 * ->css_offline() is invoked. To satisfy all the requirements,
5428 * destruction is implemented in the following two steps.
5430 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5431 * userland visible parts and start killing the percpu refcnts of
5432 * css's. Set up so that the next stage will be kicked off once all
5433 * the percpu refcnts are confirmed to be killed.
5435 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5436 * rest of destruction. Once all cgroup references are gone, the
5437 * cgroup is RCU-freed.
5439 * This function implements s1. After this step, @cgrp is gone as far as
5440 * the userland is concerned and a new cgroup with the same name may be
5441 * created. As cgroup doesn't care about the names internally, this
5442 * doesn't cause any problem.
5444 static int cgroup_destroy_locked(struct cgroup *cgrp)
5445 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5447 struct cgroup_subsys_state *css;
5448 struct cgrp_cset_link *link;
5451 lockdep_assert_held(&cgroup_mutex);
5454 * Only migration can raise populated from zero and we're already
5455 * holding cgroup_mutex.
5457 if (cgroup_is_populated(cgrp))
5461 * Make sure there's no live children. We can't test emptiness of
5462 * ->self.children as dead children linger on it while being
5463 * drained; otherwise, "rmdir parent/child parent" may fail.
5465 if (css_has_online_children(&cgrp->self))
5469 * Mark @cgrp and the associated csets dead. The former prevents
5470 * further task migration and child creation by disabling
5471 * cgroup_lock_live_group(). The latter makes the csets ignored by
5472 * the migration path.
5474 cgrp->self.flags &= ~CSS_ONLINE;
5476 spin_lock_irq(&css_set_lock);
5477 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5478 link->cset->dead = true;
5479 spin_unlock_irq(&css_set_lock);
5481 /* initiate massacre of all css's */
5482 for_each_css(css, ssid, cgrp)
5486 * Remove @cgrp directory along with the base files. @cgrp has an
5487 * extra ref on its kn.
5489 kernfs_remove(cgrp->kn);
5491 check_for_release(cgroup_parent(cgrp));
5493 /* put the base reference */
5494 percpu_ref_kill(&cgrp->self.refcnt);
5499 static int cgroup_rmdir(struct kernfs_node *kn)
5501 struct cgroup *cgrp;
5504 cgrp = cgroup_kn_lock_live(kn, false);
5508 ret = cgroup_destroy_locked(cgrp);
5510 cgroup_kn_unlock(kn);
5514 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5515 .remount_fs = cgroup_remount,
5516 .show_options = cgroup_show_options,
5517 .mkdir = cgroup_mkdir,
5518 .rmdir = cgroup_rmdir,
5519 .rename = cgroup_rename,
5520 .show_path = cgroup_show_path,
5523 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5525 struct cgroup_subsys_state *css;
5527 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5529 mutex_lock(&cgroup_mutex);
5531 idr_init(&ss->css_idr);
5532 INIT_LIST_HEAD(&ss->cfts);
5534 /* Create the root cgroup state for this subsystem */
5535 ss->root = &cgrp_dfl_root;
5536 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5537 /* We don't handle early failures gracefully */
5538 BUG_ON(IS_ERR(css));
5539 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5542 * Root csses are never destroyed and we can't initialize
5543 * percpu_ref during early init. Disable refcnting.
5545 css->flags |= CSS_NO_REF;
5548 /* allocation can't be done safely during early init */
5551 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5552 BUG_ON(css->id < 0);
5555 /* Update the init_css_set to contain a subsys
5556 * pointer to this state - since the subsystem is
5557 * newly registered, all tasks and hence the
5558 * init_css_set is in the subsystem's root cgroup. */
5559 init_css_set.subsys[ss->id] = css;
5561 have_fork_callback |= (bool)ss->fork << ss->id;
5562 have_exit_callback |= (bool)ss->exit << ss->id;
5563 have_free_callback |= (bool)ss->free << ss->id;
5564 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5566 /* At system boot, before all subsystems have been
5567 * registered, no tasks have been forked, so we don't
5568 * need to invoke fork callbacks here. */
5569 BUG_ON(!list_empty(&init_task.tasks));
5571 BUG_ON(online_css(css));
5573 mutex_unlock(&cgroup_mutex);
5577 * cgroup_init_early - cgroup initialization at system boot
5579 * Initialize cgroups at system boot, and initialize any
5580 * subsystems that request early init.
5582 int __init cgroup_init_early(void)
5584 static struct cgroup_sb_opts __initdata opts;
5585 struct cgroup_subsys *ss;
5588 init_cgroup_root(&cgrp_dfl_root, &opts);
5589 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5591 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5593 for_each_subsys(ss, i) {
5594 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5595 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5596 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5598 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5599 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5602 ss->name = cgroup_subsys_name[i];
5603 if (!ss->legacy_name)
5604 ss->legacy_name = cgroup_subsys_name[i];
5607 cgroup_init_subsys(ss, true);
5612 static u16 cgroup_disable_mask __initdata;
5615 * cgroup_init - cgroup initialization
5617 * Register cgroup filesystem and /proc file, and initialize
5618 * any subsystems that didn't request early init.
5620 int __init cgroup_init(void)
5622 struct cgroup_subsys *ss;
5625 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5626 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5627 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5628 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5631 * The latency of the synchronize_sched() is too high for cgroups,
5632 * avoid it at the cost of forcing all readers into the slow path.
5634 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5636 get_user_ns(init_cgroup_ns.user_ns);
5638 mutex_lock(&cgroup_mutex);
5641 * Add init_css_set to the hash table so that dfl_root can link to
5644 hash_add(css_set_table, &init_css_set.hlist,
5645 css_set_hash(init_css_set.subsys));
5647 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5649 mutex_unlock(&cgroup_mutex);
5651 for_each_subsys(ss, ssid) {
5652 if (ss->early_init) {
5653 struct cgroup_subsys_state *css =
5654 init_css_set.subsys[ss->id];
5656 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5658 BUG_ON(css->id < 0);
5660 cgroup_init_subsys(ss, false);
5663 list_add_tail(&init_css_set.e_cset_node[ssid],
5664 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5667 * Setting dfl_root subsys_mask needs to consider the
5668 * disabled flag and cftype registration needs kmalloc,
5669 * both of which aren't available during early_init.
5671 if (cgroup_disable_mask & (1 << ssid)) {
5672 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5673 printk(KERN_INFO "Disabling %s control group subsystem\n",
5678 if (cgroup_ssid_no_v1(ssid))
5679 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5682 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5684 if (ss->implicit_on_dfl)
5685 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5686 else if (!ss->dfl_cftypes)
5687 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5689 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5690 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5692 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5693 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5697 ss->bind(init_css_set.subsys[ssid]);
5700 /* init_css_set.subsys[] has been updated, re-hash */
5701 hash_del(&init_css_set.hlist);
5702 hash_add(css_set_table, &init_css_set.hlist,
5703 css_set_hash(init_css_set.subsys));
5705 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5706 WARN_ON(register_filesystem(&cgroup_fs_type));
5707 WARN_ON(register_filesystem(&cgroup2_fs_type));
5708 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5713 static int __init cgroup_wq_init(void)
5716 * There isn't much point in executing destruction path in
5717 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5718 * Use 1 for @max_active.
5720 * We would prefer to do this in cgroup_init() above, but that
5721 * is called before init_workqueues(): so leave this until after.
5723 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5724 BUG_ON(!cgroup_destroy_wq);
5727 * Used to destroy pidlists and separate to serve as flush domain.
5728 * Cap @max_active to 1 too.
5730 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5732 BUG_ON(!cgroup_pidlist_destroy_wq);
5736 core_initcall(cgroup_wq_init);
5739 * proc_cgroup_show()
5740 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5741 * - Used for /proc/<pid>/cgroup.
5743 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5744 struct pid *pid, struct task_struct *tsk)
5748 struct cgroup_root *root;
5751 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5755 mutex_lock(&cgroup_mutex);
5756 spin_lock_irq(&css_set_lock);
5758 for_each_root(root) {
5759 struct cgroup_subsys *ss;
5760 struct cgroup *cgrp;
5761 int ssid, count = 0;
5763 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5766 seq_printf(m, "%d:", root->hierarchy_id);
5767 if (root != &cgrp_dfl_root)
5768 for_each_subsys(ss, ssid)
5769 if (root->subsys_mask & (1 << ssid))
5770 seq_printf(m, "%s%s", count++ ? "," : "",
5772 if (strlen(root->name))
5773 seq_printf(m, "%sname=%s", count ? "," : "",
5777 cgrp = task_cgroup_from_root(tsk, root);
5780 * On traditional hierarchies, all zombie tasks show up as
5781 * belonging to the root cgroup. On the default hierarchy,
5782 * while a zombie doesn't show up in "cgroup.procs" and
5783 * thus can't be migrated, its /proc/PID/cgroup keeps
5784 * reporting the cgroup it belonged to before exiting. If
5785 * the cgroup is removed before the zombie is reaped,
5786 * " (deleted)" is appended to the cgroup path.
5788 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5789 path = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5790 current->nsproxy->cgroup_ns);
5792 retval = -ENAMETOOLONG;
5801 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5802 seq_puts(m, " (deleted)\n");
5809 spin_unlock_irq(&css_set_lock);
5810 mutex_unlock(&cgroup_mutex);
5816 /* Display information about each subsystem and each hierarchy */
5817 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5819 struct cgroup_subsys *ss;
5822 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5824 * ideally we don't want subsystems moving around while we do this.
5825 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5826 * subsys/hierarchy state.
5828 mutex_lock(&cgroup_mutex);
5830 for_each_subsys(ss, i)
5831 seq_printf(m, "%s\t%d\t%d\t%d\n",
5832 ss->legacy_name, ss->root->hierarchy_id,
5833 atomic_read(&ss->root->nr_cgrps),
5834 cgroup_ssid_enabled(i));
5836 mutex_unlock(&cgroup_mutex);
5840 static int cgroupstats_open(struct inode *inode, struct file *file)
5842 return single_open(file, proc_cgroupstats_show, NULL);
5845 static const struct file_operations proc_cgroupstats_operations = {
5846 .open = cgroupstats_open,
5848 .llseek = seq_lseek,
5849 .release = single_release,
5853 * cgroup_fork - initialize cgroup related fields during copy_process()
5854 * @child: pointer to task_struct of forking parent process.
5856 * A task is associated with the init_css_set until cgroup_post_fork()
5857 * attaches it to the parent's css_set. Empty cg_list indicates that
5858 * @child isn't holding reference to its css_set.
5860 void cgroup_fork(struct task_struct *child)
5862 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5863 INIT_LIST_HEAD(&child->cg_list);
5867 * cgroup_can_fork - called on a new task before the process is exposed
5868 * @child: the task in question.
5870 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5871 * returns an error, the fork aborts with that error code. This allows for
5872 * a cgroup subsystem to conditionally allow or deny new forks.
5874 int cgroup_can_fork(struct task_struct *child)
5876 struct cgroup_subsys *ss;
5879 do_each_subsys_mask(ss, i, have_canfork_callback) {
5880 ret = ss->can_fork(child);
5883 } while_each_subsys_mask();
5888 for_each_subsys(ss, j) {
5891 if (ss->cancel_fork)
5892 ss->cancel_fork(child);
5899 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5900 * @child: the task in question
5902 * This calls the cancel_fork() callbacks if a fork failed *after*
5903 * cgroup_can_fork() succeded.
5905 void cgroup_cancel_fork(struct task_struct *child)
5907 struct cgroup_subsys *ss;
5910 for_each_subsys(ss, i)
5911 if (ss->cancel_fork)
5912 ss->cancel_fork(child);
5916 * cgroup_post_fork - called on a new task after adding it to the task list
5917 * @child: the task in question
5919 * Adds the task to the list running through its css_set if necessary and
5920 * call the subsystem fork() callbacks. Has to be after the task is
5921 * visible on the task list in case we race with the first call to
5922 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5925 void cgroup_post_fork(struct task_struct *child)
5927 struct cgroup_subsys *ss;
5931 * This may race against cgroup_enable_task_cg_lists(). As that
5932 * function sets use_task_css_set_links before grabbing
5933 * tasklist_lock and we just went through tasklist_lock to add
5934 * @child, it's guaranteed that either we see the set
5935 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5936 * @child during its iteration.
5938 * If we won the race, @child is associated with %current's
5939 * css_set. Grabbing css_set_lock guarantees both that the
5940 * association is stable, and, on completion of the parent's
5941 * migration, @child is visible in the source of migration or
5942 * already in the destination cgroup. This guarantee is necessary
5943 * when implementing operations which need to migrate all tasks of
5944 * a cgroup to another.
5946 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5947 * will remain in init_css_set. This is safe because all tasks are
5948 * in the init_css_set before cg_links is enabled and there's no
5949 * operation which transfers all tasks out of init_css_set.
5951 if (use_task_css_set_links) {
5952 struct css_set *cset;
5954 spin_lock_irq(&css_set_lock);
5955 cset = task_css_set(current);
5956 if (list_empty(&child->cg_list)) {
5958 css_set_move_task(child, NULL, cset, false);
5960 spin_unlock_irq(&css_set_lock);
5964 * Call ss->fork(). This must happen after @child is linked on
5965 * css_set; otherwise, @child might change state between ->fork()
5966 * and addition to css_set.
5968 do_each_subsys_mask(ss, i, have_fork_callback) {
5970 } while_each_subsys_mask();
5974 * cgroup_exit - detach cgroup from exiting task
5975 * @tsk: pointer to task_struct of exiting process
5977 * Description: Detach cgroup from @tsk and release it.
5979 * Note that cgroups marked notify_on_release force every task in
5980 * them to take the global cgroup_mutex mutex when exiting.
5981 * This could impact scaling on very large systems. Be reluctant to
5982 * use notify_on_release cgroups where very high task exit scaling
5983 * is required on large systems.
5985 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5986 * call cgroup_exit() while the task is still competent to handle
5987 * notify_on_release(), then leave the task attached to the root cgroup in
5988 * each hierarchy for the remainder of its exit. No need to bother with
5989 * init_css_set refcnting. init_css_set never goes away and we can't race
5990 * with migration path - PF_EXITING is visible to migration path.
5992 void cgroup_exit(struct task_struct *tsk)
5994 struct cgroup_subsys *ss;
5995 struct css_set *cset;
5999 * Unlink from @tsk from its css_set. As migration path can't race
6000 * with us, we can check css_set and cg_list without synchronization.
6002 cset = task_css_set(tsk);
6004 if (!list_empty(&tsk->cg_list)) {
6005 spin_lock_irq(&css_set_lock);
6006 css_set_move_task(tsk, cset, NULL, false);
6007 spin_unlock_irq(&css_set_lock);
6012 /* see cgroup_post_fork() for details */
6013 do_each_subsys_mask(ss, i, have_exit_callback) {
6015 } while_each_subsys_mask();
6018 void cgroup_free(struct task_struct *task)
6020 struct css_set *cset = task_css_set(task);
6021 struct cgroup_subsys *ss;
6024 do_each_subsys_mask(ss, ssid, have_free_callback) {
6026 } while_each_subsys_mask();
6031 static void check_for_release(struct cgroup *cgrp)
6033 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
6034 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
6035 schedule_work(&cgrp->release_agent_work);
6039 * Notify userspace when a cgroup is released, by running the
6040 * configured release agent with the name of the cgroup (path
6041 * relative to the root of cgroup file system) as the argument.
6043 * Most likely, this user command will try to rmdir this cgroup.
6045 * This races with the possibility that some other task will be
6046 * attached to this cgroup before it is removed, or that some other
6047 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
6048 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
6049 * unused, and this cgroup will be reprieved from its death sentence,
6050 * to continue to serve a useful existence. Next time it's released,
6051 * we will get notified again, if it still has 'notify_on_release' set.
6053 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
6054 * means only wait until the task is successfully execve()'d. The
6055 * separate release agent task is forked by call_usermodehelper(),
6056 * then control in this thread returns here, without waiting for the
6057 * release agent task. We don't bother to wait because the caller of
6058 * this routine has no use for the exit status of the release agent
6059 * task, so no sense holding our caller up for that.
6061 static void cgroup_release_agent(struct work_struct *work)
6063 struct cgroup *cgrp =
6064 container_of(work, struct cgroup, release_agent_work);
6065 char *pathbuf = NULL, *agentbuf = NULL, *path;
6066 char *argv[3], *envp[3];
6068 mutex_lock(&cgroup_mutex);
6070 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
6071 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
6072 if (!pathbuf || !agentbuf)
6075 spin_lock_irq(&css_set_lock);
6076 path = cgroup_path_ns_locked(cgrp, pathbuf, PATH_MAX, &init_cgroup_ns);
6077 spin_unlock_irq(&css_set_lock);
6085 /* minimal command environment */
6087 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
6090 mutex_unlock(&cgroup_mutex);
6091 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
6094 mutex_unlock(&cgroup_mutex);
6100 static int __init cgroup_disable(char *str)
6102 struct cgroup_subsys *ss;
6106 while ((token = strsep(&str, ",")) != NULL) {
6110 for_each_subsys(ss, i) {
6111 if (strcmp(token, ss->name) &&
6112 strcmp(token, ss->legacy_name))
6114 cgroup_disable_mask |= 1 << i;
6119 __setup("cgroup_disable=", cgroup_disable);
6121 static int __init cgroup_no_v1(char *str)
6123 struct cgroup_subsys *ss;
6127 while ((token = strsep(&str, ",")) != NULL) {
6131 if (!strcmp(token, "all")) {
6132 cgroup_no_v1_mask = U16_MAX;
6136 for_each_subsys(ss, i) {
6137 if (strcmp(token, ss->name) &&
6138 strcmp(token, ss->legacy_name))
6141 cgroup_no_v1_mask |= 1 << i;
6146 __setup("cgroup_no_v1=", cgroup_no_v1);
6149 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6150 * @dentry: directory dentry of interest
6151 * @ss: subsystem of interest
6153 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6154 * to get the corresponding css and return it. If such css doesn't exist
6155 * or can't be pinned, an ERR_PTR value is returned.
6157 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6158 struct cgroup_subsys *ss)
6160 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6161 struct file_system_type *s_type = dentry->d_sb->s_type;
6162 struct cgroup_subsys_state *css = NULL;
6163 struct cgroup *cgrp;
6165 /* is @dentry a cgroup dir? */
6166 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6167 !kn || kernfs_type(kn) != KERNFS_DIR)
6168 return ERR_PTR(-EBADF);
6173 * This path doesn't originate from kernfs and @kn could already
6174 * have been or be removed at any point. @kn->priv is RCU
6175 * protected for this access. See css_release_work_fn() for details.
6177 cgrp = rcu_dereference(kn->priv);
6179 css = cgroup_css(cgrp, ss);
6181 if (!css || !css_tryget_online(css))
6182 css = ERR_PTR(-ENOENT);
6189 * css_from_id - lookup css by id
6190 * @id: the cgroup id
6191 * @ss: cgroup subsys to be looked into
6193 * Returns the css if there's valid one with @id, otherwise returns NULL.
6194 * Should be called under rcu_read_lock().
6196 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6198 WARN_ON_ONCE(!rcu_read_lock_held());
6199 return idr_find(&ss->css_idr, id);
6203 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6204 * @path: path on the default hierarchy
6206 * Find the cgroup at @path on the default hierarchy, increment its
6207 * reference count and return it. Returns pointer to the found cgroup on
6208 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6209 * if @path points to a non-directory.
6211 struct cgroup *cgroup_get_from_path(const char *path)
6213 struct kernfs_node *kn;
6214 struct cgroup *cgrp;
6216 mutex_lock(&cgroup_mutex);
6218 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6220 if (kernfs_type(kn) == KERNFS_DIR) {
6224 cgrp = ERR_PTR(-ENOTDIR);
6228 cgrp = ERR_PTR(-ENOENT);
6231 mutex_unlock(&cgroup_mutex);
6234 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6237 * cgroup_get_from_fd - get a cgroup pointer from a fd
6238 * @fd: fd obtained by open(cgroup2_dir)
6240 * Find the cgroup from a fd which should be obtained
6241 * by opening a cgroup directory. Returns a pointer to the
6242 * cgroup on success. ERR_PTR is returned if the cgroup
6245 struct cgroup *cgroup_get_from_fd(int fd)
6247 struct cgroup_subsys_state *css;
6248 struct cgroup *cgrp;
6253 return ERR_PTR(-EBADF);
6255 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6258 return ERR_CAST(css);
6261 if (!cgroup_on_dfl(cgrp)) {
6263 return ERR_PTR(-EBADF);
6268 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6271 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6272 * definition in cgroup-defs.h.
6274 #ifdef CONFIG_SOCK_CGROUP_DATA
6276 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6278 DEFINE_SPINLOCK(cgroup_sk_update_lock);
6279 static bool cgroup_sk_alloc_disabled __read_mostly;
6281 void cgroup_sk_alloc_disable(void)
6283 if (cgroup_sk_alloc_disabled)
6285 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6286 cgroup_sk_alloc_disabled = true;
6291 #define cgroup_sk_alloc_disabled false
6295 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6297 if (cgroup_sk_alloc_disabled)
6300 /* Socket clone path */
6302 cgroup_get(sock_cgroup_ptr(skcd));
6309 struct css_set *cset;
6311 cset = task_css_set(current);
6312 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6313 skcd->val = (unsigned long)cset->dfl_cgrp;
6322 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6324 cgroup_put(sock_cgroup_ptr(skcd));
6327 #endif /* CONFIG_SOCK_CGROUP_DATA */
6329 /* cgroup namespaces */
6331 static struct ucounts *inc_cgroup_namespaces(struct user_namespace *ns)
6333 return inc_ucount(ns, current_euid(), UCOUNT_CGROUP_NAMESPACES);
6336 static void dec_cgroup_namespaces(struct ucounts *ucounts)
6338 dec_ucount(ucounts, UCOUNT_CGROUP_NAMESPACES);
6341 static struct cgroup_namespace *alloc_cgroup_ns(void)
6343 struct cgroup_namespace *new_ns;
6346 new_ns = kzalloc(sizeof(struct cgroup_namespace), GFP_KERNEL);
6348 return ERR_PTR(-ENOMEM);
6349 ret = ns_alloc_inum(&new_ns->ns);
6352 return ERR_PTR(ret);
6354 atomic_set(&new_ns->count, 1);
6355 new_ns->ns.ops = &cgroupns_operations;
6359 void free_cgroup_ns(struct cgroup_namespace *ns)
6361 put_css_set(ns->root_cset);
6362 dec_cgroup_namespaces(ns->ucounts);
6363 put_user_ns(ns->user_ns);
6364 ns_free_inum(&ns->ns);
6367 EXPORT_SYMBOL(free_cgroup_ns);
6369 struct cgroup_namespace *copy_cgroup_ns(unsigned long flags,
6370 struct user_namespace *user_ns,
6371 struct cgroup_namespace *old_ns)
6373 struct cgroup_namespace *new_ns;
6374 struct ucounts *ucounts;
6375 struct css_set *cset;
6379 if (!(flags & CLONE_NEWCGROUP)) {
6380 get_cgroup_ns(old_ns);
6384 /* Allow only sysadmin to create cgroup namespace. */
6385 if (!ns_capable(user_ns, CAP_SYS_ADMIN))
6386 return ERR_PTR(-EPERM);
6388 ucounts = inc_cgroup_namespaces(user_ns);
6390 return ERR_PTR(-ENOSPC);
6392 /* It is not safe to take cgroup_mutex here */
6393 spin_lock_irq(&css_set_lock);
6394 cset = task_css_set(current);
6396 spin_unlock_irq(&css_set_lock);
6398 new_ns = alloc_cgroup_ns();
6399 if (IS_ERR(new_ns)) {
6401 dec_cgroup_namespaces(ucounts);
6405 new_ns->user_ns = get_user_ns(user_ns);
6406 new_ns->ucounts = ucounts;
6407 new_ns->root_cset = cset;
6412 static inline struct cgroup_namespace *to_cg_ns(struct ns_common *ns)
6414 return container_of(ns, struct cgroup_namespace, ns);
6417 static int cgroupns_install(struct nsproxy *nsproxy, struct ns_common *ns)
6419 struct cgroup_namespace *cgroup_ns = to_cg_ns(ns);
6421 if (!ns_capable(current_user_ns(), CAP_SYS_ADMIN) ||
6422 !ns_capable(cgroup_ns->user_ns, CAP_SYS_ADMIN))
6425 /* Don't need to do anything if we are attaching to our own cgroupns. */
6426 if (cgroup_ns == nsproxy->cgroup_ns)
6429 get_cgroup_ns(cgroup_ns);
6430 put_cgroup_ns(nsproxy->cgroup_ns);
6431 nsproxy->cgroup_ns = cgroup_ns;
6436 static struct ns_common *cgroupns_get(struct task_struct *task)
6438 struct cgroup_namespace *ns = NULL;
6439 struct nsproxy *nsproxy;
6442 nsproxy = task->nsproxy;
6444 ns = nsproxy->cgroup_ns;
6449 return ns ? &ns->ns : NULL;
6452 static void cgroupns_put(struct ns_common *ns)
6454 put_cgroup_ns(to_cg_ns(ns));
6457 static struct user_namespace *cgroupns_owner(struct ns_common *ns)
6459 return to_cg_ns(ns)->user_ns;
6462 const struct proc_ns_operations cgroupns_operations = {
6464 .type = CLONE_NEWCGROUP,
6465 .get = cgroupns_get,
6466 .put = cgroupns_put,
6467 .install = cgroupns_install,
6468 .owner = cgroupns_owner,
6471 static __init int cgroup_namespaces_init(void)
6475 subsys_initcall(cgroup_namespaces_init);
6477 #ifdef CONFIG_CGROUP_DEBUG
6478 static struct cgroup_subsys_state *
6479 debug_css_alloc(struct cgroup_subsys_state *parent_css)
6481 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
6484 return ERR_PTR(-ENOMEM);
6489 static void debug_css_free(struct cgroup_subsys_state *css)
6494 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
6497 return cgroup_task_count(css->cgroup);
6500 static u64 current_css_set_read(struct cgroup_subsys_state *css,
6503 return (u64)(unsigned long)current->cgroups;
6506 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
6512 count = atomic_read(&task_css_set(current)->refcount);
6517 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
6519 struct cgrp_cset_link *link;
6520 struct css_set *cset;
6523 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
6527 spin_lock_irq(&css_set_lock);
6529 cset = rcu_dereference(current->cgroups);
6530 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
6531 struct cgroup *c = link->cgrp;
6533 cgroup_name(c, name_buf, NAME_MAX + 1);
6534 seq_printf(seq, "Root %d group %s\n",
6535 c->root->hierarchy_id, name_buf);
6538 spin_unlock_irq(&css_set_lock);
6543 #define MAX_TASKS_SHOWN_PER_CSS 25
6544 static int cgroup_css_links_read(struct seq_file *seq, void *v)
6546 struct cgroup_subsys_state *css = seq_css(seq);
6547 struct cgrp_cset_link *link;
6549 spin_lock_irq(&css_set_lock);
6550 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
6551 struct css_set *cset = link->cset;
6552 struct task_struct *task;
6555 seq_printf(seq, "css_set %p\n", cset);
6557 list_for_each_entry(task, &cset->tasks, cg_list) {
6558 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6560 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6563 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
6564 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6566 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6570 seq_puts(seq, " ...\n");
6572 spin_unlock_irq(&css_set_lock);
6576 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
6578 return (!cgroup_is_populated(css->cgroup) &&
6579 !css_has_online_children(&css->cgroup->self));
6582 static struct cftype debug_files[] = {
6584 .name = "taskcount",
6585 .read_u64 = debug_taskcount_read,
6589 .name = "current_css_set",
6590 .read_u64 = current_css_set_read,
6594 .name = "current_css_set_refcount",
6595 .read_u64 = current_css_set_refcount_read,
6599 .name = "current_css_set_cg_links",
6600 .seq_show = current_css_set_cg_links_read,
6604 .name = "cgroup_css_links",
6605 .seq_show = cgroup_css_links_read,
6609 .name = "releasable",
6610 .read_u64 = releasable_read,
6616 struct cgroup_subsys debug_cgrp_subsys = {
6617 .css_alloc = debug_css_alloc,
6618 .css_free = debug_css_free,
6619 .legacy_cftypes = debug_files,
6621 #endif /* CONFIG_CGROUP_DEBUG */