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 "cgroup-internal.h"
33 #include <linux/cred.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/magic.h>
38 #include <linux/mutex.h>
39 #include <linux/mount.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/sched/task.h>
45 #include <linux/slab.h>
46 #include <linux/spinlock.h>
47 #include <linux/percpu-rwsem.h>
48 #include <linux/string.h>
49 #include <linux/hashtable.h>
50 #include <linux/idr.h>
51 #include <linux/kthread.h>
52 #include <linux/atomic.h>
53 #include <linux/cpuset.h>
54 #include <linux/proc_ns.h>
55 #include <linux/nsproxy.h>
56 #include <linux/file.h>
57 #include <linux/fs_parser.h>
58 #include <linux/sched/cputime.h>
59 #include <linux/psi.h>
62 #define CREATE_TRACE_POINTS
63 #include <trace/events/cgroup.h>
65 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
67 /* let's not notify more than 100 times per second */
68 #define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
71 * cgroup_mutex is the master lock. Any modification to cgroup or its
72 * hierarchy must be performed while holding it.
74 * css_set_lock protects task->cgroups pointer, the list of css_set
75 * objects, and the chain of tasks off each css_set.
77 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
78 * cgroup.h can use them for lockdep annotations.
80 DEFINE_MUTEX(cgroup_mutex);
81 DEFINE_SPINLOCK(css_set_lock);
83 #ifdef CONFIG_PROVE_RCU
84 EXPORT_SYMBOL_GPL(cgroup_mutex);
85 EXPORT_SYMBOL_GPL(css_set_lock);
88 DEFINE_SPINLOCK(trace_cgroup_path_lock);
89 char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
90 bool cgroup_debug __read_mostly;
93 * Protects cgroup_idr and css_idr so that IDs can be released without
94 * grabbing cgroup_mutex.
96 static DEFINE_SPINLOCK(cgroup_idr_lock);
99 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
100 * against file removal/re-creation across css hiding.
102 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
104 DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
106 #define cgroup_assert_mutex_or_rcu_locked() \
107 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
108 !lockdep_is_held(&cgroup_mutex), \
109 "cgroup_mutex or RCU read lock required");
112 * cgroup destruction makes heavy use of work items and there can be a lot
113 * of concurrent destructions. Use a separate workqueue so that cgroup
114 * destruction work items don't end up filling up max_active of system_wq
115 * which may lead to deadlock.
117 static struct workqueue_struct *cgroup_destroy_wq;
119 /* generate an array of cgroup subsystem pointers */
120 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
121 struct cgroup_subsys *cgroup_subsys[] = {
122 #include <linux/cgroup_subsys.h>
126 /* array of cgroup subsystem names */
127 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
128 static const char *cgroup_subsys_name[] = {
129 #include <linux/cgroup_subsys.h>
133 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
135 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
136 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
137 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
138 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
139 #include <linux/cgroup_subsys.h>
142 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
143 static struct static_key_true *cgroup_subsys_enabled_key[] = {
144 #include <linux/cgroup_subsys.h>
148 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
149 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
150 #include <linux/cgroup_subsys.h>
154 static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu);
156 /* the default hierarchy */
157 struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
158 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
161 * The default hierarchy always exists but is hidden until mounted for the
162 * first time. This is for backward compatibility.
164 static bool cgrp_dfl_visible;
166 /* some controllers are not supported in the default hierarchy */
167 static u16 cgrp_dfl_inhibit_ss_mask;
169 /* some controllers are implicitly enabled on the default hierarchy */
170 static u16 cgrp_dfl_implicit_ss_mask;
172 /* some controllers can be threaded on the default hierarchy */
173 static u16 cgrp_dfl_threaded_ss_mask;
175 /* The list of hierarchy roots */
176 LIST_HEAD(cgroup_roots);
177 static int cgroup_root_count;
179 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
180 static DEFINE_IDR(cgroup_hierarchy_idr);
183 * Assign a monotonically increasing serial number to csses. It guarantees
184 * cgroups with bigger numbers are newer than those with smaller numbers.
185 * Also, as csses are always appended to the parent's ->children list, it
186 * guarantees that sibling csses are always sorted in the ascending serial
187 * number order on the list. Protected by cgroup_mutex.
189 static u64 css_serial_nr_next = 1;
192 * These bitmasks identify subsystems with specific features to avoid
193 * having to do iterative checks repeatedly.
195 static u16 have_fork_callback __read_mostly;
196 static u16 have_exit_callback __read_mostly;
197 static u16 have_release_callback __read_mostly;
198 static u16 have_canfork_callback __read_mostly;
200 /* cgroup namespace for init task */
201 struct cgroup_namespace init_cgroup_ns = {
202 .ns.count = REFCOUNT_INIT(2),
203 .user_ns = &init_user_ns,
204 .ns.ops = &cgroupns_operations,
205 .ns.inum = PROC_CGROUP_INIT_INO,
206 .root_cset = &init_css_set,
209 static struct file_system_type cgroup2_fs_type;
210 static struct cftype cgroup_base_files[];
212 static int cgroup_apply_control(struct cgroup *cgrp);
213 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
214 static void css_task_iter_skip(struct css_task_iter *it,
215 struct task_struct *task);
216 static int cgroup_destroy_locked(struct cgroup *cgrp);
217 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
218 struct cgroup_subsys *ss);
219 static void css_release(struct percpu_ref *ref);
220 static void kill_css(struct cgroup_subsys_state *css);
221 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
222 struct cgroup *cgrp, struct cftype cfts[],
226 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
227 * @ssid: subsys ID of interest
229 * cgroup_subsys_enabled() can only be used with literal subsys names which
230 * is fine for individual subsystems but unsuitable for cgroup core. This
231 * is slower static_key_enabled() based test indexed by @ssid.
233 bool cgroup_ssid_enabled(int ssid)
235 if (CGROUP_SUBSYS_COUNT == 0)
238 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
242 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
243 * @cgrp: the cgroup of interest
245 * The default hierarchy is the v2 interface of cgroup and this function
246 * can be used to test whether a cgroup is on the default hierarchy for
247 * cases where a subsystem should behave differently depending on the
250 * List of changed behaviors:
252 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
253 * and "name" are disallowed.
255 * - When mounting an existing superblock, mount options should match.
257 * - Remount is disallowed.
259 * - rename(2) is disallowed.
261 * - "tasks" is removed. Everything should be at process granularity. Use
262 * "cgroup.procs" instead.
264 * - "cgroup.procs" is not sorted. pids will be unique unless they got
265 * recycled in-between reads.
267 * - "release_agent" and "notify_on_release" are removed. Replacement
268 * notification mechanism will be implemented.
270 * - "cgroup.clone_children" is removed.
272 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
273 * and its descendants contain no task; otherwise, 1. The file also
274 * generates kernfs notification which can be monitored through poll and
275 * [di]notify when the value of the file changes.
277 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
278 * take masks of ancestors with non-empty cpus/mems, instead of being
279 * moved to an ancestor.
281 * - cpuset: a task can be moved into an empty cpuset, and again it takes
282 * masks of ancestors.
284 * - blkcg: blk-throttle becomes properly hierarchical.
286 * - debug: disallowed on the default hierarchy.
288 bool cgroup_on_dfl(const struct cgroup *cgrp)
290 return cgrp->root == &cgrp_dfl_root;
293 /* IDR wrappers which synchronize using cgroup_idr_lock */
294 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
299 idr_preload(gfp_mask);
300 spin_lock_bh(&cgroup_idr_lock);
301 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
302 spin_unlock_bh(&cgroup_idr_lock);
307 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
311 spin_lock_bh(&cgroup_idr_lock);
312 ret = idr_replace(idr, ptr, id);
313 spin_unlock_bh(&cgroup_idr_lock);
317 static void cgroup_idr_remove(struct idr *idr, int id)
319 spin_lock_bh(&cgroup_idr_lock);
321 spin_unlock_bh(&cgroup_idr_lock);
324 static bool cgroup_has_tasks(struct cgroup *cgrp)
326 return cgrp->nr_populated_csets;
329 bool cgroup_is_threaded(struct cgroup *cgrp)
331 return cgrp->dom_cgrp != cgrp;
334 /* can @cgrp host both domain and threaded children? */
335 static bool cgroup_is_mixable(struct cgroup *cgrp)
338 * Root isn't under domain level resource control exempting it from
339 * the no-internal-process constraint, so it can serve as a thread
340 * root and a parent of resource domains at the same time.
342 return !cgroup_parent(cgrp);
345 /* can @cgrp become a thread root? Should always be true for a thread root */
346 static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
348 /* mixables don't care */
349 if (cgroup_is_mixable(cgrp))
352 /* domain roots can't be nested under threaded */
353 if (cgroup_is_threaded(cgrp))
356 /* can only have either domain or threaded children */
357 if (cgrp->nr_populated_domain_children)
360 /* and no domain controllers can be enabled */
361 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
367 /* is @cgrp root of a threaded subtree? */
368 bool cgroup_is_thread_root(struct cgroup *cgrp)
370 /* thread root should be a domain */
371 if (cgroup_is_threaded(cgrp))
374 /* a domain w/ threaded children is a thread root */
375 if (cgrp->nr_threaded_children)
379 * A domain which has tasks and explicit threaded controllers
380 * enabled is a thread root.
382 if (cgroup_has_tasks(cgrp) &&
383 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
389 /* a domain which isn't connected to the root w/o brekage can't be used */
390 static bool cgroup_is_valid_domain(struct cgroup *cgrp)
392 /* the cgroup itself can be a thread root */
393 if (cgroup_is_threaded(cgrp))
396 /* but the ancestors can't be unless mixable */
397 while ((cgrp = cgroup_parent(cgrp))) {
398 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
400 if (cgroup_is_threaded(cgrp))
407 /* subsystems visibly enabled on a cgroup */
408 static u16 cgroup_control(struct cgroup *cgrp)
410 struct cgroup *parent = cgroup_parent(cgrp);
411 u16 root_ss_mask = cgrp->root->subsys_mask;
414 u16 ss_mask = parent->subtree_control;
416 /* threaded cgroups can only have threaded controllers */
417 if (cgroup_is_threaded(cgrp))
418 ss_mask &= cgrp_dfl_threaded_ss_mask;
422 if (cgroup_on_dfl(cgrp))
423 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
424 cgrp_dfl_implicit_ss_mask);
428 /* subsystems enabled on a cgroup */
429 static u16 cgroup_ss_mask(struct cgroup *cgrp)
431 struct cgroup *parent = cgroup_parent(cgrp);
434 u16 ss_mask = parent->subtree_ss_mask;
436 /* threaded cgroups can only have threaded controllers */
437 if (cgroup_is_threaded(cgrp))
438 ss_mask &= cgrp_dfl_threaded_ss_mask;
442 return cgrp->root->subsys_mask;
446 * cgroup_css - obtain a cgroup's css for the specified subsystem
447 * @cgrp: the cgroup of interest
448 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
450 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
451 * function must be called either under cgroup_mutex or rcu_read_lock() and
452 * the caller is responsible for pinning the returned css if it wants to
453 * keep accessing it outside the said locks. This function may return
454 * %NULL if @cgrp doesn't have @subsys_id enabled.
456 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
457 struct cgroup_subsys *ss)
460 return rcu_dereference_check(cgrp->subsys[ss->id],
461 lockdep_is_held(&cgroup_mutex));
467 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
468 * @cgrp: the cgroup of interest
469 * @ss: the subsystem of interest
471 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
472 * or is offline, %NULL is returned.
474 static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
475 struct cgroup_subsys *ss)
477 struct cgroup_subsys_state *css;
480 css = cgroup_css(cgrp, ss);
481 if (css && !css_tryget_online(css))
489 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
490 * @cgrp: the cgroup of interest
491 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
493 * Similar to cgroup_css() but returns the effective css, which is defined
494 * as the matching css of the nearest ancestor including self which has @ss
495 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
496 * function is guaranteed to return non-NULL css.
498 static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
499 struct cgroup_subsys *ss)
501 lockdep_assert_held(&cgroup_mutex);
507 * This function is used while updating css associations and thus
508 * can't test the csses directly. Test ss_mask.
510 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
511 cgrp = cgroup_parent(cgrp);
516 return cgroup_css(cgrp, ss);
520 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
521 * @cgrp: the cgroup of interest
522 * @ss: the subsystem of interest
524 * Find and get the effective css of @cgrp for @ss. The effective css is
525 * defined as the matching css of the nearest ancestor including self which
526 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
527 * the root css is returned, so this function always returns a valid css.
529 * The returned css is not guaranteed to be online, and therefore it is the
530 * callers responsibility to try get a reference for it.
532 struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
533 struct cgroup_subsys *ss)
535 struct cgroup_subsys_state *css;
538 css = cgroup_css(cgrp, ss);
542 cgrp = cgroup_parent(cgrp);
545 return init_css_set.subsys[ss->id];
549 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
550 * @cgrp: the cgroup of interest
551 * @ss: the subsystem of interest
553 * Find and get the effective css of @cgrp for @ss. The effective css is
554 * defined as the matching css of the nearest ancestor including self which
555 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
556 * the root css is returned, so this function always returns a valid css.
557 * The returned css must be put using css_put().
559 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
560 struct cgroup_subsys *ss)
562 struct cgroup_subsys_state *css;
567 css = cgroup_css(cgrp, ss);
569 if (css && css_tryget_online(css))
571 cgrp = cgroup_parent(cgrp);
574 css = init_css_set.subsys[ss->id];
581 static void cgroup_get_live(struct cgroup *cgrp)
583 WARN_ON_ONCE(cgroup_is_dead(cgrp));
584 css_get(&cgrp->self);
588 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
589 * is responsible for taking the css_set_lock.
590 * @cgrp: the cgroup in question
592 int __cgroup_task_count(const struct cgroup *cgrp)
595 struct cgrp_cset_link *link;
597 lockdep_assert_held(&css_set_lock);
599 list_for_each_entry(link, &cgrp->cset_links, cset_link)
600 count += link->cset->nr_tasks;
606 * cgroup_task_count - count the number of tasks in a cgroup.
607 * @cgrp: the cgroup in question
609 int cgroup_task_count(const struct cgroup *cgrp)
613 spin_lock_irq(&css_set_lock);
614 count = __cgroup_task_count(cgrp);
615 spin_unlock_irq(&css_set_lock);
620 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
622 struct cgroup *cgrp = of->kn->parent->priv;
623 struct cftype *cft = of_cft(of);
626 * This is open and unprotected implementation of cgroup_css().
627 * seq_css() is only called from a kernfs file operation which has
628 * an active reference on the file. Because all the subsystem
629 * files are drained before a css is disassociated with a cgroup,
630 * the matching css from the cgroup's subsys table is guaranteed to
631 * be and stay valid until the enclosing operation is complete.
634 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
638 EXPORT_SYMBOL_GPL(of_css);
641 * for_each_css - iterate all css's of a cgroup
642 * @css: the iteration cursor
643 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
644 * @cgrp: the target cgroup to iterate css's of
646 * Should be called under cgroup_[tree_]mutex.
648 #define for_each_css(css, ssid, cgrp) \
649 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
650 if (!((css) = rcu_dereference_check( \
651 (cgrp)->subsys[(ssid)], \
652 lockdep_is_held(&cgroup_mutex)))) { } \
656 * for_each_e_css - iterate all effective css's of a cgroup
657 * @css: the iteration cursor
658 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
659 * @cgrp: the target cgroup to iterate css's of
661 * Should be called under cgroup_[tree_]mutex.
663 #define for_each_e_css(css, ssid, cgrp) \
664 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
665 if (!((css) = cgroup_e_css_by_mask(cgrp, \
666 cgroup_subsys[(ssid)]))) \
671 * do_each_subsys_mask - filter for_each_subsys with a bitmask
672 * @ss: the iteration cursor
673 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
674 * @ss_mask: the bitmask
676 * The block will only run for cases where the ssid-th bit (1 << ssid) of
679 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
680 unsigned long __ss_mask = (ss_mask); \
681 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
685 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
686 (ss) = cgroup_subsys[ssid]; \
689 #define while_each_subsys_mask() \
694 /* iterate over child cgrps, lock should be held throughout iteration */
695 #define cgroup_for_each_live_child(child, cgrp) \
696 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
697 if (({ lockdep_assert_held(&cgroup_mutex); \
698 cgroup_is_dead(child); })) \
702 /* walk live descendants in pre order */
703 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
704 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
705 if (({ lockdep_assert_held(&cgroup_mutex); \
706 (dsct) = (d_css)->cgroup; \
707 cgroup_is_dead(dsct); })) \
711 /* walk live descendants in postorder */
712 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
713 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
714 if (({ lockdep_assert_held(&cgroup_mutex); \
715 (dsct) = (d_css)->cgroup; \
716 cgroup_is_dead(dsct); })) \
721 * The default css_set - used by init and its children prior to any
722 * hierarchies being mounted. It contains a pointer to the root state
723 * for each subsystem. Also used to anchor the list of css_sets. Not
724 * reference-counted, to improve performance when child cgroups
725 * haven't been created.
727 struct css_set init_css_set = {
728 .refcount = REFCOUNT_INIT(1),
729 .dom_cset = &init_css_set,
730 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
731 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
732 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
733 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
734 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
735 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
736 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
737 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
740 * The following field is re-initialized when this cset gets linked
741 * in cgroup_init(). However, let's initialize the field
742 * statically too so that the default cgroup can be accessed safely
745 .dfl_cgrp = &cgrp_dfl_root.cgrp,
748 static int css_set_count = 1; /* 1 for init_css_set */
750 static bool css_set_threaded(struct css_set *cset)
752 return cset->dom_cset != cset;
756 * css_set_populated - does a css_set contain any tasks?
757 * @cset: target css_set
759 * css_set_populated() should be the same as !!cset->nr_tasks at steady
760 * state. However, css_set_populated() can be called while a task is being
761 * added to or removed from the linked list before the nr_tasks is
762 * properly updated. Hence, we can't just look at ->nr_tasks here.
764 static bool css_set_populated(struct css_set *cset)
766 lockdep_assert_held(&css_set_lock);
768 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
772 * cgroup_update_populated - update the populated count of a cgroup
773 * @cgrp: the target cgroup
774 * @populated: inc or dec populated count
776 * One of the css_sets associated with @cgrp is either getting its first
777 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
778 * count is propagated towards root so that a given cgroup's
779 * nr_populated_children is zero iff none of its descendants contain any
782 * @cgrp's interface file "cgroup.populated" is zero if both
783 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
784 * 1 otherwise. When the sum changes from or to zero, userland is notified
785 * that the content of the interface file has changed. This can be used to
786 * detect when @cgrp and its descendants become populated or empty.
788 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
790 struct cgroup *child = NULL;
791 int adj = populated ? 1 : -1;
793 lockdep_assert_held(&css_set_lock);
796 bool was_populated = cgroup_is_populated(cgrp);
799 cgrp->nr_populated_csets += adj;
801 if (cgroup_is_threaded(child))
802 cgrp->nr_populated_threaded_children += adj;
804 cgrp->nr_populated_domain_children += adj;
807 if (was_populated == cgroup_is_populated(cgrp))
810 cgroup1_check_for_release(cgrp);
811 TRACE_CGROUP_PATH(notify_populated, cgrp,
812 cgroup_is_populated(cgrp));
813 cgroup_file_notify(&cgrp->events_file);
816 cgrp = cgroup_parent(cgrp);
821 * css_set_update_populated - update populated state of a css_set
822 * @cset: target css_set
823 * @populated: whether @cset is populated or depopulated
825 * @cset is either getting the first task or losing the last. Update the
826 * populated counters of all associated cgroups accordingly.
828 static void css_set_update_populated(struct css_set *cset, bool populated)
830 struct cgrp_cset_link *link;
832 lockdep_assert_held(&css_set_lock);
834 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
835 cgroup_update_populated(link->cgrp, populated);
839 * @task is leaving, advance task iterators which are pointing to it so
840 * that they can resume at the next position. Advancing an iterator might
841 * remove it from the list, use safe walk. See css_task_iter_skip() for
844 static void css_set_skip_task_iters(struct css_set *cset,
845 struct task_struct *task)
847 struct css_task_iter *it, *pos;
849 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
850 css_task_iter_skip(it, task);
854 * css_set_move_task - move a task from one css_set to another
855 * @task: task being moved
856 * @from_cset: css_set @task currently belongs to (may be NULL)
857 * @to_cset: new css_set @task is being moved to (may be NULL)
858 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
860 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
861 * css_set, @from_cset can be NULL. If @task is being disassociated
862 * instead of moved, @to_cset can be NULL.
864 * This function automatically handles populated counter updates and
865 * css_task_iter adjustments but the caller is responsible for managing
866 * @from_cset and @to_cset's reference counts.
868 static void css_set_move_task(struct task_struct *task,
869 struct css_set *from_cset, struct css_set *to_cset,
872 lockdep_assert_held(&css_set_lock);
874 if (to_cset && !css_set_populated(to_cset))
875 css_set_update_populated(to_cset, true);
878 WARN_ON_ONCE(list_empty(&task->cg_list));
880 css_set_skip_task_iters(from_cset, task);
881 list_del_init(&task->cg_list);
882 if (!css_set_populated(from_cset))
883 css_set_update_populated(from_cset, false);
885 WARN_ON_ONCE(!list_empty(&task->cg_list));
890 * We are synchronized through cgroup_threadgroup_rwsem
891 * against PF_EXITING setting such that we can't race
892 * against cgroup_exit()/cgroup_free() dropping the css_set.
894 WARN_ON_ONCE(task->flags & PF_EXITING);
896 cgroup_move_task(task, to_cset);
897 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
903 * hash table for cgroup groups. This improves the performance to find
904 * an existing css_set. This hash doesn't (currently) take into
905 * account cgroups in empty hierarchies.
907 #define CSS_SET_HASH_BITS 7
908 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
910 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
912 unsigned long key = 0UL;
913 struct cgroup_subsys *ss;
916 for_each_subsys(ss, i)
917 key += (unsigned long)css[i];
918 key = (key >> 16) ^ key;
923 void put_css_set_locked(struct css_set *cset)
925 struct cgrp_cset_link *link, *tmp_link;
926 struct cgroup_subsys *ss;
929 lockdep_assert_held(&css_set_lock);
931 if (!refcount_dec_and_test(&cset->refcount))
934 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
936 /* This css_set is dead. Unlink it and release cgroup and css refs */
937 for_each_subsys(ss, ssid) {
938 list_del(&cset->e_cset_node[ssid]);
939 css_put(cset->subsys[ssid]);
941 hash_del(&cset->hlist);
944 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
945 list_del(&link->cset_link);
946 list_del(&link->cgrp_link);
947 if (cgroup_parent(link->cgrp))
948 cgroup_put(link->cgrp);
952 if (css_set_threaded(cset)) {
953 list_del(&cset->threaded_csets_node);
954 put_css_set_locked(cset->dom_cset);
957 kfree_rcu(cset, rcu_head);
961 * compare_css_sets - helper function for find_existing_css_set().
962 * @cset: candidate css_set being tested
963 * @old_cset: existing css_set for a task
964 * @new_cgrp: cgroup that's being entered by the task
965 * @template: desired set of css pointers in css_set (pre-calculated)
967 * Returns true if "cset" matches "old_cset" except for the hierarchy
968 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
970 static bool compare_css_sets(struct css_set *cset,
971 struct css_set *old_cset,
972 struct cgroup *new_cgrp,
973 struct cgroup_subsys_state *template[])
975 struct cgroup *new_dfl_cgrp;
976 struct list_head *l1, *l2;
979 * On the default hierarchy, there can be csets which are
980 * associated with the same set of cgroups but different csses.
981 * Let's first ensure that csses match.
983 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
987 /* @cset's domain should match the default cgroup's */
988 if (cgroup_on_dfl(new_cgrp))
989 new_dfl_cgrp = new_cgrp;
991 new_dfl_cgrp = old_cset->dfl_cgrp;
993 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
997 * Compare cgroup pointers in order to distinguish between
998 * different cgroups in hierarchies. As different cgroups may
999 * share the same effective css, this comparison is always
1002 l1 = &cset->cgrp_links;
1003 l2 = &old_cset->cgrp_links;
1005 struct cgrp_cset_link *link1, *link2;
1006 struct cgroup *cgrp1, *cgrp2;
1010 /* See if we reached the end - both lists are equal length. */
1011 if (l1 == &cset->cgrp_links) {
1012 BUG_ON(l2 != &old_cset->cgrp_links);
1015 BUG_ON(l2 == &old_cset->cgrp_links);
1017 /* Locate the cgroups associated with these links. */
1018 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1019 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1020 cgrp1 = link1->cgrp;
1021 cgrp2 = link2->cgrp;
1022 /* Hierarchies should be linked in the same order. */
1023 BUG_ON(cgrp1->root != cgrp2->root);
1026 * If this hierarchy is the hierarchy of the cgroup
1027 * that's changing, then we need to check that this
1028 * css_set points to the new cgroup; if it's any other
1029 * hierarchy, then this css_set should point to the
1030 * same cgroup as the old css_set.
1032 if (cgrp1->root == new_cgrp->root) {
1033 if (cgrp1 != new_cgrp)
1044 * find_existing_css_set - init css array and find the matching css_set
1045 * @old_cset: the css_set that we're using before the cgroup transition
1046 * @cgrp: the cgroup that we're moving into
1047 * @template: out param for the new set of csses, should be clear on entry
1049 static struct css_set *find_existing_css_set(struct css_set *old_cset,
1050 struct cgroup *cgrp,
1051 struct cgroup_subsys_state *template[])
1053 struct cgroup_root *root = cgrp->root;
1054 struct cgroup_subsys *ss;
1055 struct css_set *cset;
1060 * Build the set of subsystem state objects that we want to see in the
1061 * new css_set. While subsystems can change globally, the entries here
1062 * won't change, so no need for locking.
1064 for_each_subsys(ss, i) {
1065 if (root->subsys_mask & (1UL << i)) {
1067 * @ss is in this hierarchy, so we want the
1068 * effective css from @cgrp.
1070 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1073 * @ss is not in this hierarchy, so we don't want
1074 * to change the css.
1076 template[i] = old_cset->subsys[i];
1080 key = css_set_hash(template);
1081 hash_for_each_possible(css_set_table, cset, hlist, key) {
1082 if (!compare_css_sets(cset, old_cset, cgrp, template))
1085 /* This css_set matches what we need */
1089 /* No existing cgroup group matched */
1093 static void free_cgrp_cset_links(struct list_head *links_to_free)
1095 struct cgrp_cset_link *link, *tmp_link;
1097 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1098 list_del(&link->cset_link);
1104 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1105 * @count: the number of links to allocate
1106 * @tmp_links: list_head the allocated links are put on
1108 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1109 * through ->cset_link. Returns 0 on success or -errno.
1111 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1113 struct cgrp_cset_link *link;
1116 INIT_LIST_HEAD(tmp_links);
1118 for (i = 0; i < count; i++) {
1119 link = kzalloc(sizeof(*link), GFP_KERNEL);
1121 free_cgrp_cset_links(tmp_links);
1124 list_add(&link->cset_link, tmp_links);
1130 * link_css_set - a helper function to link a css_set to a cgroup
1131 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1132 * @cset: the css_set to be linked
1133 * @cgrp: the destination cgroup
1135 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1136 struct cgroup *cgrp)
1138 struct cgrp_cset_link *link;
1140 BUG_ON(list_empty(tmp_links));
1142 if (cgroup_on_dfl(cgrp))
1143 cset->dfl_cgrp = cgrp;
1145 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1150 * Always add links to the tail of the lists so that the lists are
1151 * in chronological order.
1153 list_move_tail(&link->cset_link, &cgrp->cset_links);
1154 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1156 if (cgroup_parent(cgrp))
1157 cgroup_get_live(cgrp);
1161 * find_css_set - return a new css_set with one cgroup updated
1162 * @old_cset: the baseline css_set
1163 * @cgrp: the cgroup to be updated
1165 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1166 * substituted into the appropriate hierarchy.
1168 static struct css_set *find_css_set(struct css_set *old_cset,
1169 struct cgroup *cgrp)
1171 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1172 struct css_set *cset;
1173 struct list_head tmp_links;
1174 struct cgrp_cset_link *link;
1175 struct cgroup_subsys *ss;
1179 lockdep_assert_held(&cgroup_mutex);
1181 /* First see if we already have a cgroup group that matches
1182 * the desired set */
1183 spin_lock_irq(&css_set_lock);
1184 cset = find_existing_css_set(old_cset, cgrp, template);
1187 spin_unlock_irq(&css_set_lock);
1192 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1196 /* Allocate all the cgrp_cset_link objects that we'll need */
1197 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1202 refcount_set(&cset->refcount, 1);
1203 cset->dom_cset = cset;
1204 INIT_LIST_HEAD(&cset->tasks);
1205 INIT_LIST_HEAD(&cset->mg_tasks);
1206 INIT_LIST_HEAD(&cset->dying_tasks);
1207 INIT_LIST_HEAD(&cset->task_iters);
1208 INIT_LIST_HEAD(&cset->threaded_csets);
1209 INIT_HLIST_NODE(&cset->hlist);
1210 INIT_LIST_HEAD(&cset->cgrp_links);
1211 INIT_LIST_HEAD(&cset->mg_preload_node);
1212 INIT_LIST_HEAD(&cset->mg_node);
1214 /* Copy the set of subsystem state objects generated in
1215 * find_existing_css_set() */
1216 memcpy(cset->subsys, template, sizeof(cset->subsys));
1218 spin_lock_irq(&css_set_lock);
1219 /* Add reference counts and links from the new css_set. */
1220 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1221 struct cgroup *c = link->cgrp;
1223 if (c->root == cgrp->root)
1225 link_css_set(&tmp_links, cset, c);
1228 BUG_ON(!list_empty(&tmp_links));
1232 /* Add @cset to the hash table */
1233 key = css_set_hash(cset->subsys);
1234 hash_add(css_set_table, &cset->hlist, key);
1236 for_each_subsys(ss, ssid) {
1237 struct cgroup_subsys_state *css = cset->subsys[ssid];
1239 list_add_tail(&cset->e_cset_node[ssid],
1240 &css->cgroup->e_csets[ssid]);
1244 spin_unlock_irq(&css_set_lock);
1247 * If @cset should be threaded, look up the matching dom_cset and
1248 * link them up. We first fully initialize @cset then look for the
1249 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1250 * to stay empty until we return.
1252 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1253 struct css_set *dcset;
1255 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1261 spin_lock_irq(&css_set_lock);
1262 cset->dom_cset = dcset;
1263 list_add_tail(&cset->threaded_csets_node,
1264 &dcset->threaded_csets);
1265 spin_unlock_irq(&css_set_lock);
1271 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1273 struct cgroup *root_cgrp = kf_root->kn->priv;
1275 return root_cgrp->root;
1278 static int cgroup_init_root_id(struct cgroup_root *root)
1282 lockdep_assert_held(&cgroup_mutex);
1284 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1288 root->hierarchy_id = id;
1292 static void cgroup_exit_root_id(struct cgroup_root *root)
1294 lockdep_assert_held(&cgroup_mutex);
1296 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1299 void cgroup_free_root(struct cgroup_root *root)
1304 static void cgroup_destroy_root(struct cgroup_root *root)
1306 struct cgroup *cgrp = &root->cgrp;
1307 struct cgrp_cset_link *link, *tmp_link;
1309 trace_cgroup_destroy_root(root);
1311 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1313 BUG_ON(atomic_read(&root->nr_cgrps));
1314 BUG_ON(!list_empty(&cgrp->self.children));
1316 /* Rebind all subsystems back to the default hierarchy */
1317 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1320 * Release all the links from cset_links to this hierarchy's
1323 spin_lock_irq(&css_set_lock);
1325 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1326 list_del(&link->cset_link);
1327 list_del(&link->cgrp_link);
1331 spin_unlock_irq(&css_set_lock);
1333 if (!list_empty(&root->root_list)) {
1334 list_del(&root->root_list);
1335 cgroup_root_count--;
1338 cgroup_exit_root_id(root);
1340 mutex_unlock(&cgroup_mutex);
1342 cgroup_rstat_exit(cgrp);
1343 kernfs_destroy_root(root->kf_root);
1344 cgroup_free_root(root);
1348 * look up cgroup associated with current task's cgroup namespace on the
1349 * specified hierarchy
1351 static struct cgroup *
1352 current_cgns_cgroup_from_root(struct cgroup_root *root)
1354 struct cgroup *res = NULL;
1355 struct css_set *cset;
1357 lockdep_assert_held(&css_set_lock);
1361 cset = current->nsproxy->cgroup_ns->root_cset;
1362 if (cset == &init_css_set) {
1364 } else if (root == &cgrp_dfl_root) {
1365 res = cset->dfl_cgrp;
1367 struct cgrp_cset_link *link;
1369 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1370 struct cgroup *c = link->cgrp;
1372 if (c->root == root) {
1384 /* look up cgroup associated with given css_set on the specified hierarchy */
1385 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1386 struct cgroup_root *root)
1388 struct cgroup *res = NULL;
1390 lockdep_assert_held(&cgroup_mutex);
1391 lockdep_assert_held(&css_set_lock);
1393 if (cset == &init_css_set) {
1395 } else if (root == &cgrp_dfl_root) {
1396 res = cset->dfl_cgrp;
1398 struct cgrp_cset_link *link;
1400 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1401 struct cgroup *c = link->cgrp;
1403 if (c->root == root) {
1415 * Return the cgroup for "task" from the given hierarchy. Must be
1416 * called with cgroup_mutex and css_set_lock held.
1418 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1419 struct cgroup_root *root)
1422 * No need to lock the task - since we hold css_set_lock the
1423 * task can't change groups.
1425 return cset_cgroup_from_root(task_css_set(task), root);
1429 * A task must hold cgroup_mutex to modify cgroups.
1431 * Any task can increment and decrement the count field without lock.
1432 * So in general, code holding cgroup_mutex can't rely on the count
1433 * field not changing. However, if the count goes to zero, then only
1434 * cgroup_attach_task() can increment it again. Because a count of zero
1435 * means that no tasks are currently attached, therefore there is no
1436 * way a task attached to that cgroup can fork (the other way to
1437 * increment the count). So code holding cgroup_mutex can safely
1438 * assume that if the count is zero, it will stay zero. Similarly, if
1439 * a task holds cgroup_mutex on a cgroup with zero count, it
1440 * knows that the cgroup won't be removed, as cgroup_rmdir()
1443 * A cgroup can only be deleted if both its 'count' of using tasks
1444 * is zero, and its list of 'children' cgroups is empty. Since all
1445 * tasks in the system use _some_ cgroup, and since there is always at
1446 * least one task in the system (init, pid == 1), therefore, root cgroup
1447 * always has either children cgroups and/or using tasks. So we don't
1448 * need a special hack to ensure that root cgroup cannot be deleted.
1450 * P.S. One more locking exception. RCU is used to guard the
1451 * update of a tasks cgroup pointer by cgroup_attach_task()
1454 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1456 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1459 struct cgroup_subsys *ss = cft->ss;
1461 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1462 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1463 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1465 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1466 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1469 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1475 * cgroup_file_mode - deduce file mode of a control file
1476 * @cft: the control file in question
1478 * S_IRUGO for read, S_IWUSR for write.
1480 static umode_t cgroup_file_mode(const struct cftype *cft)
1484 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1487 if (cft->write_u64 || cft->write_s64 || cft->write) {
1488 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1498 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1499 * @subtree_control: the new subtree_control mask to consider
1500 * @this_ss_mask: available subsystems
1502 * On the default hierarchy, a subsystem may request other subsystems to be
1503 * enabled together through its ->depends_on mask. In such cases, more
1504 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1506 * This function calculates which subsystems need to be enabled if
1507 * @subtree_control is to be applied while restricted to @this_ss_mask.
1509 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1511 u16 cur_ss_mask = subtree_control;
1512 struct cgroup_subsys *ss;
1515 lockdep_assert_held(&cgroup_mutex);
1517 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1520 u16 new_ss_mask = cur_ss_mask;
1522 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1523 new_ss_mask |= ss->depends_on;
1524 } while_each_subsys_mask();
1527 * Mask out subsystems which aren't available. This can
1528 * happen only if some depended-upon subsystems were bound
1529 * to non-default hierarchies.
1531 new_ss_mask &= this_ss_mask;
1533 if (new_ss_mask == cur_ss_mask)
1535 cur_ss_mask = new_ss_mask;
1542 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1543 * @kn: the kernfs_node being serviced
1545 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1546 * the method finishes if locking succeeded. Note that once this function
1547 * returns the cgroup returned by cgroup_kn_lock_live() may become
1548 * inaccessible any time. If the caller intends to continue to access the
1549 * cgroup, it should pin it before invoking this function.
1551 void cgroup_kn_unlock(struct kernfs_node *kn)
1553 struct cgroup *cgrp;
1555 if (kernfs_type(kn) == KERNFS_DIR)
1558 cgrp = kn->parent->priv;
1560 mutex_unlock(&cgroup_mutex);
1562 kernfs_unbreak_active_protection(kn);
1567 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1568 * @kn: the kernfs_node being serviced
1569 * @drain_offline: perform offline draining on the cgroup
1571 * This helper is to be used by a cgroup kernfs method currently servicing
1572 * @kn. It breaks the active protection, performs cgroup locking and
1573 * verifies that the associated cgroup is alive. Returns the cgroup if
1574 * alive; otherwise, %NULL. A successful return should be undone by a
1575 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1576 * cgroup is drained of offlining csses before return.
1578 * Any cgroup kernfs method implementation which requires locking the
1579 * associated cgroup should use this helper. It avoids nesting cgroup
1580 * locking under kernfs active protection and allows all kernfs operations
1581 * including self-removal.
1583 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1585 struct cgroup *cgrp;
1587 if (kernfs_type(kn) == KERNFS_DIR)
1590 cgrp = kn->parent->priv;
1593 * We're gonna grab cgroup_mutex which nests outside kernfs
1594 * active_ref. cgroup liveliness check alone provides enough
1595 * protection against removal. Ensure @cgrp stays accessible and
1596 * break the active_ref protection.
1598 if (!cgroup_tryget(cgrp))
1600 kernfs_break_active_protection(kn);
1603 cgroup_lock_and_drain_offline(cgrp);
1605 mutex_lock(&cgroup_mutex);
1607 if (!cgroup_is_dead(cgrp))
1610 cgroup_kn_unlock(kn);
1614 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1616 char name[CGROUP_FILE_NAME_MAX];
1618 lockdep_assert_held(&cgroup_mutex);
1620 if (cft->file_offset) {
1621 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1622 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1624 spin_lock_irq(&cgroup_file_kn_lock);
1626 spin_unlock_irq(&cgroup_file_kn_lock);
1628 del_timer_sync(&cfile->notify_timer);
1631 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1635 * css_clear_dir - remove subsys files in a cgroup directory
1638 static void css_clear_dir(struct cgroup_subsys_state *css)
1640 struct cgroup *cgrp = css->cgroup;
1641 struct cftype *cfts;
1643 if (!(css->flags & CSS_VISIBLE))
1646 css->flags &= ~CSS_VISIBLE;
1649 if (cgroup_on_dfl(cgrp))
1650 cfts = cgroup_base_files;
1652 cfts = cgroup1_base_files;
1654 cgroup_addrm_files(css, cgrp, cfts, false);
1656 list_for_each_entry(cfts, &css->ss->cfts, node)
1657 cgroup_addrm_files(css, cgrp, cfts, false);
1662 * css_populate_dir - create subsys files in a cgroup directory
1665 * On failure, no file is added.
1667 static int css_populate_dir(struct cgroup_subsys_state *css)
1669 struct cgroup *cgrp = css->cgroup;
1670 struct cftype *cfts, *failed_cfts;
1673 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1677 if (cgroup_on_dfl(cgrp))
1678 cfts = cgroup_base_files;
1680 cfts = cgroup1_base_files;
1682 ret = cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1686 list_for_each_entry(cfts, &css->ss->cfts, node) {
1687 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1695 css->flags |= CSS_VISIBLE;
1699 list_for_each_entry(cfts, &css->ss->cfts, node) {
1700 if (cfts == failed_cfts)
1702 cgroup_addrm_files(css, cgrp, cfts, false);
1707 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1709 struct cgroup *dcgrp = &dst_root->cgrp;
1710 struct cgroup_subsys *ss;
1713 lockdep_assert_held(&cgroup_mutex);
1715 do_each_subsys_mask(ss, ssid, ss_mask) {
1717 * If @ss has non-root csses attached to it, can't move.
1718 * If @ss is an implicit controller, it is exempt from this
1719 * rule and can be stolen.
1721 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1722 !ss->implicit_on_dfl)
1725 /* can't move between two non-dummy roots either */
1726 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1728 } while_each_subsys_mask();
1730 do_each_subsys_mask(ss, ssid, ss_mask) {
1731 struct cgroup_root *src_root = ss->root;
1732 struct cgroup *scgrp = &src_root->cgrp;
1733 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1734 struct css_set *cset;
1736 WARN_ON(!css || cgroup_css(dcgrp, ss));
1738 /* disable from the source */
1739 src_root->subsys_mask &= ~(1 << ssid);
1740 WARN_ON(cgroup_apply_control(scgrp));
1741 cgroup_finalize_control(scgrp, 0);
1744 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1745 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1746 ss->root = dst_root;
1747 css->cgroup = dcgrp;
1749 spin_lock_irq(&css_set_lock);
1750 hash_for_each(css_set_table, i, cset, hlist)
1751 list_move_tail(&cset->e_cset_node[ss->id],
1752 &dcgrp->e_csets[ss->id]);
1753 spin_unlock_irq(&css_set_lock);
1755 if (ss->css_rstat_flush) {
1756 list_del_rcu(&css->rstat_css_node);
1757 list_add_rcu(&css->rstat_css_node,
1758 &dcgrp->rstat_css_list);
1761 /* default hierarchy doesn't enable controllers by default */
1762 dst_root->subsys_mask |= 1 << ssid;
1763 if (dst_root == &cgrp_dfl_root) {
1764 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1766 dcgrp->subtree_control |= 1 << ssid;
1767 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1770 ret = cgroup_apply_control(dcgrp);
1772 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1777 } while_each_subsys_mask();
1779 kernfs_activate(dcgrp->kn);
1783 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1784 struct kernfs_root *kf_root)
1788 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1789 struct cgroup *ns_cgroup;
1791 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1795 spin_lock_irq(&css_set_lock);
1796 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1797 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1798 spin_unlock_irq(&css_set_lock);
1800 if (len >= PATH_MAX)
1803 seq_escape(sf, buf, " \t\n\\");
1810 enum cgroup2_param {
1812 Opt_memory_localevents,
1813 Opt_memory_recursiveprot,
1817 static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1818 fsparam_flag("nsdelegate", Opt_nsdelegate),
1819 fsparam_flag("memory_localevents", Opt_memory_localevents),
1820 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
1824 static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1826 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1827 struct fs_parse_result result;
1830 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1835 case Opt_nsdelegate:
1836 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1838 case Opt_memory_localevents:
1839 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1841 case Opt_memory_recursiveprot:
1842 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1848 static void apply_cgroup_root_flags(unsigned int root_flags)
1850 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1851 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1852 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1854 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1856 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1857 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1859 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1861 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1862 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1864 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1868 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1870 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1871 seq_puts(seq, ",nsdelegate");
1872 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1873 seq_puts(seq, ",memory_localevents");
1874 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1875 seq_puts(seq, ",memory_recursiveprot");
1879 static int cgroup_reconfigure(struct fs_context *fc)
1881 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1883 apply_cgroup_root_flags(ctx->flags);
1887 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1889 struct cgroup_subsys *ss;
1892 INIT_LIST_HEAD(&cgrp->self.sibling);
1893 INIT_LIST_HEAD(&cgrp->self.children);
1894 INIT_LIST_HEAD(&cgrp->cset_links);
1895 INIT_LIST_HEAD(&cgrp->pidlists);
1896 mutex_init(&cgrp->pidlist_mutex);
1897 cgrp->self.cgroup = cgrp;
1898 cgrp->self.flags |= CSS_ONLINE;
1899 cgrp->dom_cgrp = cgrp;
1900 cgrp->max_descendants = INT_MAX;
1901 cgrp->max_depth = INT_MAX;
1902 INIT_LIST_HEAD(&cgrp->rstat_css_list);
1903 prev_cputime_init(&cgrp->prev_cputime);
1905 for_each_subsys(ss, ssid)
1906 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1908 init_waitqueue_head(&cgrp->offline_waitq);
1909 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1912 void init_cgroup_root(struct cgroup_fs_context *ctx)
1914 struct cgroup_root *root = ctx->root;
1915 struct cgroup *cgrp = &root->cgrp;
1917 INIT_LIST_HEAD(&root->root_list);
1918 atomic_set(&root->nr_cgrps, 1);
1920 init_cgroup_housekeeping(cgrp);
1922 root->flags = ctx->flags;
1923 if (ctx->release_agent)
1924 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
1926 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
1927 if (ctx->cpuset_clone_children)
1928 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1931 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1933 LIST_HEAD(tmp_links);
1934 struct cgroup *root_cgrp = &root->cgrp;
1935 struct kernfs_syscall_ops *kf_sops;
1936 struct css_set *cset;
1939 lockdep_assert_held(&cgroup_mutex);
1941 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1947 * We're accessing css_set_count without locking css_set_lock here,
1948 * but that's OK - it can only be increased by someone holding
1949 * cgroup_lock, and that's us. Later rebinding may disable
1950 * controllers on the default hierarchy and thus create new csets,
1951 * which can't be more than the existing ones. Allocate 2x.
1953 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1957 ret = cgroup_init_root_id(root);
1961 kf_sops = root == &cgrp_dfl_root ?
1962 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1964 root->kf_root = kernfs_create_root(kf_sops,
1965 KERNFS_ROOT_CREATE_DEACTIVATED |
1966 KERNFS_ROOT_SUPPORT_EXPORTOP |
1967 KERNFS_ROOT_SUPPORT_USER_XATTR,
1969 if (IS_ERR(root->kf_root)) {
1970 ret = PTR_ERR(root->kf_root);
1973 root_cgrp->kn = root->kf_root->kn;
1974 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
1975 root_cgrp->ancestor_ids[0] = cgroup_id(root_cgrp);
1977 ret = css_populate_dir(&root_cgrp->self);
1981 ret = cgroup_rstat_init(root_cgrp);
1985 ret = rebind_subsystems(root, ss_mask);
1989 ret = cgroup_bpf_inherit(root_cgrp);
1992 trace_cgroup_setup_root(root);
1995 * There must be no failure case after here, since rebinding takes
1996 * care of subsystems' refcounts, which are explicitly dropped in
1997 * the failure exit path.
1999 list_add(&root->root_list, &cgroup_roots);
2000 cgroup_root_count++;
2003 * Link the root cgroup in this hierarchy into all the css_set
2006 spin_lock_irq(&css_set_lock);
2007 hash_for_each(css_set_table, i, cset, hlist) {
2008 link_css_set(&tmp_links, cset, root_cgrp);
2009 if (css_set_populated(cset))
2010 cgroup_update_populated(root_cgrp, true);
2012 spin_unlock_irq(&css_set_lock);
2014 BUG_ON(!list_empty(&root_cgrp->self.children));
2015 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2021 cgroup_rstat_exit(root_cgrp);
2023 kernfs_destroy_root(root->kf_root);
2024 root->kf_root = NULL;
2026 cgroup_exit_root_id(root);
2028 percpu_ref_exit(&root_cgrp->self.refcnt);
2030 free_cgrp_cset_links(&tmp_links);
2034 int cgroup_do_get_tree(struct fs_context *fc)
2036 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2039 ctx->kfc.root = ctx->root->kf_root;
2040 if (fc->fs_type == &cgroup2_fs_type)
2041 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2043 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2044 ret = kernfs_get_tree(fc);
2047 * In non-init cgroup namespace, instead of root cgroup's dentry,
2048 * we return the dentry corresponding to the cgroupns->root_cgrp.
2050 if (!ret && ctx->ns != &init_cgroup_ns) {
2051 struct dentry *nsdentry;
2052 struct super_block *sb = fc->root->d_sb;
2053 struct cgroup *cgrp;
2055 mutex_lock(&cgroup_mutex);
2056 spin_lock_irq(&css_set_lock);
2058 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2060 spin_unlock_irq(&css_set_lock);
2061 mutex_unlock(&cgroup_mutex);
2063 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2065 if (IS_ERR(nsdentry)) {
2066 deactivate_locked_super(sb);
2067 ret = PTR_ERR(nsdentry);
2070 fc->root = nsdentry;
2073 if (!ctx->kfc.new_sb_created)
2074 cgroup_put(&ctx->root->cgrp);
2080 * Destroy a cgroup filesystem context.
2082 static void cgroup_fs_context_free(struct fs_context *fc)
2084 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2087 kfree(ctx->release_agent);
2088 put_cgroup_ns(ctx->ns);
2089 kernfs_free_fs_context(fc);
2093 static int cgroup_get_tree(struct fs_context *fc)
2095 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2098 cgrp_dfl_visible = true;
2099 cgroup_get_live(&cgrp_dfl_root.cgrp);
2100 ctx->root = &cgrp_dfl_root;
2102 ret = cgroup_do_get_tree(fc);
2104 apply_cgroup_root_flags(ctx->flags);
2108 static const struct fs_context_operations cgroup_fs_context_ops = {
2109 .free = cgroup_fs_context_free,
2110 .parse_param = cgroup2_parse_param,
2111 .get_tree = cgroup_get_tree,
2112 .reconfigure = cgroup_reconfigure,
2115 static const struct fs_context_operations cgroup1_fs_context_ops = {
2116 .free = cgroup_fs_context_free,
2117 .parse_param = cgroup1_parse_param,
2118 .get_tree = cgroup1_get_tree,
2119 .reconfigure = cgroup1_reconfigure,
2123 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2124 * we select the namespace we're going to use.
2126 static int cgroup_init_fs_context(struct fs_context *fc)
2128 struct cgroup_fs_context *ctx;
2130 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2134 ctx->ns = current->nsproxy->cgroup_ns;
2135 get_cgroup_ns(ctx->ns);
2136 fc->fs_private = &ctx->kfc;
2137 if (fc->fs_type == &cgroup2_fs_type)
2138 fc->ops = &cgroup_fs_context_ops;
2140 fc->ops = &cgroup1_fs_context_ops;
2141 put_user_ns(fc->user_ns);
2142 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2147 static void cgroup_kill_sb(struct super_block *sb)
2149 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2150 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2153 * If @root doesn't have any children, start killing it.
2154 * This prevents new mounts by disabling percpu_ref_tryget_live().
2155 * cgroup_mount() may wait for @root's release.
2157 * And don't kill the default root.
2159 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2160 !percpu_ref_is_dying(&root->cgrp.self.refcnt))
2161 percpu_ref_kill(&root->cgrp.self.refcnt);
2162 cgroup_put(&root->cgrp);
2166 struct file_system_type cgroup_fs_type = {
2168 .init_fs_context = cgroup_init_fs_context,
2169 .parameters = cgroup1_fs_parameters,
2170 .kill_sb = cgroup_kill_sb,
2171 .fs_flags = FS_USERNS_MOUNT,
2174 static struct file_system_type cgroup2_fs_type = {
2176 .init_fs_context = cgroup_init_fs_context,
2177 .parameters = cgroup2_fs_parameters,
2178 .kill_sb = cgroup_kill_sb,
2179 .fs_flags = FS_USERNS_MOUNT,
2182 #ifdef CONFIG_CPUSETS
2183 static const struct fs_context_operations cpuset_fs_context_ops = {
2184 .get_tree = cgroup1_get_tree,
2185 .free = cgroup_fs_context_free,
2189 * This is ugly, but preserves the userspace API for existing cpuset
2190 * users. If someone tries to mount the "cpuset" filesystem, we
2191 * silently switch it to mount "cgroup" instead
2193 static int cpuset_init_fs_context(struct fs_context *fc)
2195 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2196 struct cgroup_fs_context *ctx;
2199 err = cgroup_init_fs_context(fc);
2205 fc->ops = &cpuset_fs_context_ops;
2207 ctx = cgroup_fc2context(fc);
2208 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2209 ctx->flags |= CGRP_ROOT_NOPREFIX;
2210 ctx->release_agent = agent;
2212 get_filesystem(&cgroup_fs_type);
2213 put_filesystem(fc->fs_type);
2214 fc->fs_type = &cgroup_fs_type;
2219 static struct file_system_type cpuset_fs_type = {
2221 .init_fs_context = cpuset_init_fs_context,
2222 .fs_flags = FS_USERNS_MOUNT,
2226 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2227 struct cgroup_namespace *ns)
2229 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2231 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2234 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2235 struct cgroup_namespace *ns)
2239 mutex_lock(&cgroup_mutex);
2240 spin_lock_irq(&css_set_lock);
2242 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2244 spin_unlock_irq(&css_set_lock);
2245 mutex_unlock(&cgroup_mutex);
2249 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2252 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2253 * @task: target task
2254 * @buf: the buffer to write the path into
2255 * @buflen: the length of the buffer
2257 * Determine @task's cgroup on the first (the one with the lowest non-zero
2258 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2259 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2260 * cgroup controller callbacks.
2262 * Return value is the same as kernfs_path().
2264 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2266 struct cgroup_root *root;
2267 struct cgroup *cgrp;
2268 int hierarchy_id = 1;
2271 mutex_lock(&cgroup_mutex);
2272 spin_lock_irq(&css_set_lock);
2274 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2277 cgrp = task_cgroup_from_root(task, root);
2278 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2280 /* if no hierarchy exists, everyone is in "/" */
2281 ret = strlcpy(buf, "/", buflen);
2284 spin_unlock_irq(&css_set_lock);
2285 mutex_unlock(&cgroup_mutex);
2288 EXPORT_SYMBOL_GPL(task_cgroup_path);
2291 * cgroup_migrate_add_task - add a migration target task to a migration context
2292 * @task: target task
2293 * @mgctx: target migration context
2295 * Add @task, which is a migration target, to @mgctx->tset. This function
2296 * becomes noop if @task doesn't need to be migrated. @task's css_set
2297 * should have been added as a migration source and @task->cg_list will be
2298 * moved from the css_set's tasks list to mg_tasks one.
2300 static void cgroup_migrate_add_task(struct task_struct *task,
2301 struct cgroup_mgctx *mgctx)
2303 struct css_set *cset;
2305 lockdep_assert_held(&css_set_lock);
2307 /* @task either already exited or can't exit until the end */
2308 if (task->flags & PF_EXITING)
2311 /* cgroup_threadgroup_rwsem protects racing against forks */
2312 WARN_ON_ONCE(list_empty(&task->cg_list));
2314 cset = task_css_set(task);
2315 if (!cset->mg_src_cgrp)
2318 mgctx->tset.nr_tasks++;
2320 list_move_tail(&task->cg_list, &cset->mg_tasks);
2321 if (list_empty(&cset->mg_node))
2322 list_add_tail(&cset->mg_node,
2323 &mgctx->tset.src_csets);
2324 if (list_empty(&cset->mg_dst_cset->mg_node))
2325 list_add_tail(&cset->mg_dst_cset->mg_node,
2326 &mgctx->tset.dst_csets);
2330 * cgroup_taskset_first - reset taskset and return the first task
2331 * @tset: taskset of interest
2332 * @dst_cssp: output variable for the destination css
2334 * @tset iteration is initialized and the first task is returned.
2336 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2337 struct cgroup_subsys_state **dst_cssp)
2339 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2340 tset->cur_task = NULL;
2342 return cgroup_taskset_next(tset, dst_cssp);
2346 * cgroup_taskset_next - iterate to the next task in taskset
2347 * @tset: taskset of interest
2348 * @dst_cssp: output variable for the destination css
2350 * Return the next task in @tset. Iteration must have been initialized
2351 * with cgroup_taskset_first().
2353 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2354 struct cgroup_subsys_state **dst_cssp)
2356 struct css_set *cset = tset->cur_cset;
2357 struct task_struct *task = tset->cur_task;
2359 while (&cset->mg_node != tset->csets) {
2361 task = list_first_entry(&cset->mg_tasks,
2362 struct task_struct, cg_list);
2364 task = list_next_entry(task, cg_list);
2366 if (&task->cg_list != &cset->mg_tasks) {
2367 tset->cur_cset = cset;
2368 tset->cur_task = task;
2371 * This function may be called both before and
2372 * after cgroup_taskset_migrate(). The two cases
2373 * can be distinguished by looking at whether @cset
2374 * has its ->mg_dst_cset set.
2376 if (cset->mg_dst_cset)
2377 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2379 *dst_cssp = cset->subsys[tset->ssid];
2384 cset = list_next_entry(cset, mg_node);
2392 * cgroup_taskset_migrate - migrate a taskset
2393 * @mgctx: migration context
2395 * Migrate tasks in @mgctx as setup by migration preparation functions.
2396 * This function fails iff one of the ->can_attach callbacks fails and
2397 * guarantees that either all or none of the tasks in @mgctx are migrated.
2398 * @mgctx is consumed regardless of success.
2400 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2402 struct cgroup_taskset *tset = &mgctx->tset;
2403 struct cgroup_subsys *ss;
2404 struct task_struct *task, *tmp_task;
2405 struct css_set *cset, *tmp_cset;
2406 int ssid, failed_ssid, ret;
2408 /* check that we can legitimately attach to the cgroup */
2409 if (tset->nr_tasks) {
2410 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2411 if (ss->can_attach) {
2413 ret = ss->can_attach(tset);
2416 goto out_cancel_attach;
2419 } while_each_subsys_mask();
2423 * Now that we're guaranteed success, proceed to move all tasks to
2424 * the new cgroup. There are no failure cases after here, so this
2425 * is the commit point.
2427 spin_lock_irq(&css_set_lock);
2428 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2429 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2430 struct css_set *from_cset = task_css_set(task);
2431 struct css_set *to_cset = cset->mg_dst_cset;
2433 get_css_set(to_cset);
2434 to_cset->nr_tasks++;
2435 css_set_move_task(task, from_cset, to_cset, true);
2436 from_cset->nr_tasks--;
2438 * If the source or destination cgroup is frozen,
2439 * the task might require to change its state.
2441 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2443 put_css_set_locked(from_cset);
2447 spin_unlock_irq(&css_set_lock);
2450 * Migration is committed, all target tasks are now on dst_csets.
2451 * Nothing is sensitive to fork() after this point. Notify
2452 * controllers that migration is complete.
2454 tset->csets = &tset->dst_csets;
2456 if (tset->nr_tasks) {
2457 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2462 } while_each_subsys_mask();
2466 goto out_release_tset;
2469 if (tset->nr_tasks) {
2470 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2471 if (ssid == failed_ssid)
2473 if (ss->cancel_attach) {
2475 ss->cancel_attach(tset);
2477 } while_each_subsys_mask();
2480 spin_lock_irq(&css_set_lock);
2481 list_splice_init(&tset->dst_csets, &tset->src_csets);
2482 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2483 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2484 list_del_init(&cset->mg_node);
2486 spin_unlock_irq(&css_set_lock);
2489 * Re-initialize the cgroup_taskset structure in case it is reused
2490 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2494 tset->csets = &tset->src_csets;
2499 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2500 * @dst_cgrp: destination cgroup to test
2502 * On the default hierarchy, except for the mixable, (possible) thread root
2503 * and threaded cgroups, subtree_control must be zero for migration
2504 * destination cgroups with tasks so that child cgroups don't compete
2507 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2509 /* v1 doesn't have any restriction */
2510 if (!cgroup_on_dfl(dst_cgrp))
2513 /* verify @dst_cgrp can host resources */
2514 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2517 /* mixables don't care */
2518 if (cgroup_is_mixable(dst_cgrp))
2522 * If @dst_cgrp is already or can become a thread root or is
2523 * threaded, it doesn't matter.
2525 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2528 /* apply no-internal-process constraint */
2529 if (dst_cgrp->subtree_control)
2536 * cgroup_migrate_finish - cleanup after attach
2537 * @mgctx: migration context
2539 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2540 * those functions for details.
2542 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2544 LIST_HEAD(preloaded);
2545 struct css_set *cset, *tmp_cset;
2547 lockdep_assert_held(&cgroup_mutex);
2549 spin_lock_irq(&css_set_lock);
2551 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2552 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2554 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2555 cset->mg_src_cgrp = NULL;
2556 cset->mg_dst_cgrp = NULL;
2557 cset->mg_dst_cset = NULL;
2558 list_del_init(&cset->mg_preload_node);
2559 put_css_set_locked(cset);
2562 spin_unlock_irq(&css_set_lock);
2566 * cgroup_migrate_add_src - add a migration source css_set
2567 * @src_cset: the source css_set to add
2568 * @dst_cgrp: the destination cgroup
2569 * @mgctx: migration context
2571 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2572 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2573 * up by cgroup_migrate_finish().
2575 * This function may be called without holding cgroup_threadgroup_rwsem
2576 * even if the target is a process. Threads may be created and destroyed
2577 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2578 * into play and the preloaded css_sets are guaranteed to cover all
2581 void cgroup_migrate_add_src(struct css_set *src_cset,
2582 struct cgroup *dst_cgrp,
2583 struct cgroup_mgctx *mgctx)
2585 struct cgroup *src_cgrp;
2587 lockdep_assert_held(&cgroup_mutex);
2588 lockdep_assert_held(&css_set_lock);
2591 * If ->dead, @src_set is associated with one or more dead cgroups
2592 * and doesn't contain any migratable tasks. Ignore it early so
2593 * that the rest of migration path doesn't get confused by it.
2598 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2600 if (!list_empty(&src_cset->mg_preload_node))
2603 WARN_ON(src_cset->mg_src_cgrp);
2604 WARN_ON(src_cset->mg_dst_cgrp);
2605 WARN_ON(!list_empty(&src_cset->mg_tasks));
2606 WARN_ON(!list_empty(&src_cset->mg_node));
2608 src_cset->mg_src_cgrp = src_cgrp;
2609 src_cset->mg_dst_cgrp = dst_cgrp;
2610 get_css_set(src_cset);
2611 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2615 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2616 * @mgctx: migration context
2618 * Tasks are about to be moved and all the source css_sets have been
2619 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2620 * pins all destination css_sets, links each to its source, and append them
2621 * to @mgctx->preloaded_dst_csets.
2623 * This function must be called after cgroup_migrate_add_src() has been
2624 * called on each migration source css_set. After migration is performed
2625 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2628 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2630 struct css_set *src_cset, *tmp_cset;
2632 lockdep_assert_held(&cgroup_mutex);
2634 /* look up the dst cset for each src cset and link it to src */
2635 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2637 struct css_set *dst_cset;
2638 struct cgroup_subsys *ss;
2641 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2645 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2648 * If src cset equals dst, it's noop. Drop the src.
2649 * cgroup_migrate() will skip the cset too. Note that we
2650 * can't handle src == dst as some nodes are used by both.
2652 if (src_cset == dst_cset) {
2653 src_cset->mg_src_cgrp = NULL;
2654 src_cset->mg_dst_cgrp = NULL;
2655 list_del_init(&src_cset->mg_preload_node);
2656 put_css_set(src_cset);
2657 put_css_set(dst_cset);
2661 src_cset->mg_dst_cset = dst_cset;
2663 if (list_empty(&dst_cset->mg_preload_node))
2664 list_add_tail(&dst_cset->mg_preload_node,
2665 &mgctx->preloaded_dst_csets);
2667 put_css_set(dst_cset);
2669 for_each_subsys(ss, ssid)
2670 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2671 mgctx->ss_mask |= 1 << ssid;
2678 * cgroup_migrate - migrate a process or task to a cgroup
2679 * @leader: the leader of the process or the task to migrate
2680 * @threadgroup: whether @leader points to the whole process or a single task
2681 * @mgctx: migration context
2683 * Migrate a process or task denoted by @leader. If migrating a process,
2684 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2685 * responsible for invoking cgroup_migrate_add_src() and
2686 * cgroup_migrate_prepare_dst() on the targets before invoking this
2687 * function and following up with cgroup_migrate_finish().
2689 * As long as a controller's ->can_attach() doesn't fail, this function is
2690 * guaranteed to succeed. This means that, excluding ->can_attach()
2691 * failure, when migrating multiple targets, the success or failure can be
2692 * decided for all targets by invoking group_migrate_prepare_dst() before
2693 * actually starting migrating.
2695 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2696 struct cgroup_mgctx *mgctx)
2698 struct task_struct *task;
2701 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2702 * already PF_EXITING could be freed from underneath us unless we
2703 * take an rcu_read_lock.
2705 spin_lock_irq(&css_set_lock);
2709 cgroup_migrate_add_task(task, mgctx);
2712 } while_each_thread(leader, task);
2714 spin_unlock_irq(&css_set_lock);
2716 return cgroup_migrate_execute(mgctx);
2720 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2721 * @dst_cgrp: the cgroup to attach to
2722 * @leader: the task or the leader of the threadgroup to be attached
2723 * @threadgroup: attach the whole threadgroup?
2725 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2727 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2730 DEFINE_CGROUP_MGCTX(mgctx);
2731 struct task_struct *task;
2734 /* look up all src csets */
2735 spin_lock_irq(&css_set_lock);
2739 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2742 } while_each_thread(leader, task);
2744 spin_unlock_irq(&css_set_lock);
2746 /* prepare dst csets and commit */
2747 ret = cgroup_migrate_prepare_dst(&mgctx);
2749 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2751 cgroup_migrate_finish(&mgctx);
2754 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2759 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2761 __acquires(&cgroup_threadgroup_rwsem)
2763 struct task_struct *tsk;
2766 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2767 return ERR_PTR(-EINVAL);
2770 * If we migrate a single thread, we don't care about threadgroup
2771 * stability. If the thread is `current`, it won't exit(2) under our
2772 * hands or change PID through exec(2). We exclude
2773 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2774 * callers by cgroup_mutex.
2775 * Therefore, we can skip the global lock.
2777 lockdep_assert_held(&cgroup_mutex);
2778 if (pid || threadgroup) {
2779 percpu_down_write(&cgroup_threadgroup_rwsem);
2787 tsk = find_task_by_vpid(pid);
2789 tsk = ERR_PTR(-ESRCH);
2790 goto out_unlock_threadgroup;
2797 tsk = tsk->group_leader;
2800 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2801 * If userland migrates such a kthread to a non-root cgroup, it can
2802 * become trapped in a cpuset, or RT kthread may be born in a
2803 * cgroup with no rt_runtime allocated. Just say no.
2805 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2806 tsk = ERR_PTR(-EINVAL);
2807 goto out_unlock_threadgroup;
2810 get_task_struct(tsk);
2811 goto out_unlock_rcu;
2813 out_unlock_threadgroup:
2815 percpu_up_write(&cgroup_threadgroup_rwsem);
2823 void cgroup_procs_write_finish(struct task_struct *task, bool locked)
2824 __releases(&cgroup_threadgroup_rwsem)
2826 struct cgroup_subsys *ss;
2829 /* release reference from cgroup_procs_write_start() */
2830 put_task_struct(task);
2833 percpu_up_write(&cgroup_threadgroup_rwsem);
2834 for_each_subsys(ss, ssid)
2835 if (ss->post_attach)
2839 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2841 struct cgroup_subsys *ss;
2842 bool printed = false;
2845 do_each_subsys_mask(ss, ssid, ss_mask) {
2848 seq_puts(seq, ss->name);
2850 } while_each_subsys_mask();
2852 seq_putc(seq, '\n');
2855 /* show controllers which are enabled from the parent */
2856 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2858 struct cgroup *cgrp = seq_css(seq)->cgroup;
2860 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2864 /* show controllers which are enabled for a given cgroup's children */
2865 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2867 struct cgroup *cgrp = seq_css(seq)->cgroup;
2869 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2874 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2875 * @cgrp: root of the subtree to update csses for
2877 * @cgrp's control masks have changed and its subtree's css associations
2878 * need to be updated accordingly. This function looks up all css_sets
2879 * which are attached to the subtree, creates the matching updated css_sets
2880 * and migrates the tasks to the new ones.
2882 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2884 DEFINE_CGROUP_MGCTX(mgctx);
2885 struct cgroup_subsys_state *d_css;
2886 struct cgroup *dsct;
2887 struct css_set *src_cset;
2890 lockdep_assert_held(&cgroup_mutex);
2892 percpu_down_write(&cgroup_threadgroup_rwsem);
2894 /* look up all csses currently attached to @cgrp's subtree */
2895 spin_lock_irq(&css_set_lock);
2896 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2897 struct cgrp_cset_link *link;
2899 list_for_each_entry(link, &dsct->cset_links, cset_link)
2900 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2902 spin_unlock_irq(&css_set_lock);
2904 /* NULL dst indicates self on default hierarchy */
2905 ret = cgroup_migrate_prepare_dst(&mgctx);
2909 spin_lock_irq(&css_set_lock);
2910 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2911 struct task_struct *task, *ntask;
2913 /* all tasks in src_csets need to be migrated */
2914 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2915 cgroup_migrate_add_task(task, &mgctx);
2917 spin_unlock_irq(&css_set_lock);
2919 ret = cgroup_migrate_execute(&mgctx);
2921 cgroup_migrate_finish(&mgctx);
2922 percpu_up_write(&cgroup_threadgroup_rwsem);
2927 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2928 * @cgrp: root of the target subtree
2930 * Because css offlining is asynchronous, userland may try to re-enable a
2931 * controller while the previous css is still around. This function grabs
2932 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2934 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2935 __acquires(&cgroup_mutex)
2937 struct cgroup *dsct;
2938 struct cgroup_subsys_state *d_css;
2939 struct cgroup_subsys *ss;
2943 mutex_lock(&cgroup_mutex);
2945 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2946 for_each_subsys(ss, ssid) {
2947 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2950 if (!css || !percpu_ref_is_dying(&css->refcnt))
2953 cgroup_get_live(dsct);
2954 prepare_to_wait(&dsct->offline_waitq, &wait,
2955 TASK_UNINTERRUPTIBLE);
2957 mutex_unlock(&cgroup_mutex);
2959 finish_wait(&dsct->offline_waitq, &wait);
2968 * cgroup_save_control - save control masks and dom_cgrp of a subtree
2969 * @cgrp: root of the target subtree
2971 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
2972 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2975 static void cgroup_save_control(struct cgroup *cgrp)
2977 struct cgroup *dsct;
2978 struct cgroup_subsys_state *d_css;
2980 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2981 dsct->old_subtree_control = dsct->subtree_control;
2982 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
2983 dsct->old_dom_cgrp = dsct->dom_cgrp;
2988 * cgroup_propagate_control - refresh control masks of a subtree
2989 * @cgrp: root of the target subtree
2991 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2992 * ->subtree_control and propagate controller availability through the
2993 * subtree so that descendants don't have unavailable controllers enabled.
2995 static void cgroup_propagate_control(struct cgroup *cgrp)
2997 struct cgroup *dsct;
2998 struct cgroup_subsys_state *d_css;
3000 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3001 dsct->subtree_control &= cgroup_control(dsct);
3002 dsct->subtree_ss_mask =
3003 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3004 cgroup_ss_mask(dsct));
3009 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3010 * @cgrp: root of the target subtree
3012 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3013 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3016 static void cgroup_restore_control(struct cgroup *cgrp)
3018 struct cgroup *dsct;
3019 struct cgroup_subsys_state *d_css;
3021 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3022 dsct->subtree_control = dsct->old_subtree_control;
3023 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3024 dsct->dom_cgrp = dsct->old_dom_cgrp;
3028 static bool css_visible(struct cgroup_subsys_state *css)
3030 struct cgroup_subsys *ss = css->ss;
3031 struct cgroup *cgrp = css->cgroup;
3033 if (cgroup_control(cgrp) & (1 << ss->id))
3035 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3037 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3041 * cgroup_apply_control_enable - enable or show csses according to control
3042 * @cgrp: root of the target subtree
3044 * Walk @cgrp's subtree and create new csses or make the existing ones
3045 * visible. A css is created invisible if it's being implicitly enabled
3046 * through dependency. An invisible css is made visible when the userland
3047 * explicitly enables it.
3049 * Returns 0 on success, -errno on failure. On failure, csses which have
3050 * been processed already aren't cleaned up. The caller is responsible for
3051 * cleaning up with cgroup_apply_control_disable().
3053 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3055 struct cgroup *dsct;
3056 struct cgroup_subsys_state *d_css;
3057 struct cgroup_subsys *ss;
3060 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3061 for_each_subsys(ss, ssid) {
3062 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3064 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3068 css = css_create(dsct, ss);
3070 return PTR_ERR(css);
3073 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3075 if (css_visible(css)) {
3076 ret = css_populate_dir(css);
3087 * cgroup_apply_control_disable - kill or hide csses according to control
3088 * @cgrp: root of the target subtree
3090 * Walk @cgrp's subtree and kill and hide csses so that they match
3091 * cgroup_ss_mask() and cgroup_visible_mask().
3093 * A css is hidden when the userland requests it to be disabled while other
3094 * subsystems are still depending on it. The css must not actively control
3095 * resources and be in the vanilla state if it's made visible again later.
3096 * Controllers which may be depended upon should provide ->css_reset() for
3099 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3101 struct cgroup *dsct;
3102 struct cgroup_subsys_state *d_css;
3103 struct cgroup_subsys *ss;
3106 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3107 for_each_subsys(ss, ssid) {
3108 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3113 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3116 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3118 } else if (!css_visible(css)) {
3128 * cgroup_apply_control - apply control mask updates to the subtree
3129 * @cgrp: root of the target subtree
3131 * subsystems can be enabled and disabled in a subtree using the following
3134 * 1. Call cgroup_save_control() to stash the current state.
3135 * 2. Update ->subtree_control masks in the subtree as desired.
3136 * 3. Call cgroup_apply_control() to apply the changes.
3137 * 4. Optionally perform other related operations.
3138 * 5. Call cgroup_finalize_control() to finish up.
3140 * This function implements step 3 and propagates the mask changes
3141 * throughout @cgrp's subtree, updates csses accordingly and perform
3142 * process migrations.
3144 static int cgroup_apply_control(struct cgroup *cgrp)
3148 cgroup_propagate_control(cgrp);
3150 ret = cgroup_apply_control_enable(cgrp);
3155 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3156 * making the following cgroup_update_dfl_csses() properly update
3157 * css associations of all tasks in the subtree.
3159 ret = cgroup_update_dfl_csses(cgrp);
3167 * cgroup_finalize_control - finalize control mask update
3168 * @cgrp: root of the target subtree
3169 * @ret: the result of the update
3171 * Finalize control mask update. See cgroup_apply_control() for more info.
3173 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3176 cgroup_restore_control(cgrp);
3177 cgroup_propagate_control(cgrp);
3180 cgroup_apply_control_disable(cgrp);
3183 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3185 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3187 /* if nothing is getting enabled, nothing to worry about */
3191 /* can @cgrp host any resources? */
3192 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3195 /* mixables don't care */
3196 if (cgroup_is_mixable(cgrp))
3199 if (domain_enable) {
3200 /* can't enable domain controllers inside a thread subtree */
3201 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3205 * Threaded controllers can handle internal competitions
3206 * and are always allowed inside a (prospective) thread
3209 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3214 * Controllers can't be enabled for a cgroup with tasks to avoid
3215 * child cgroups competing against tasks.
3217 if (cgroup_has_tasks(cgrp))
3223 /* change the enabled child controllers for a cgroup in the default hierarchy */
3224 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3225 char *buf, size_t nbytes,
3228 u16 enable = 0, disable = 0;
3229 struct cgroup *cgrp, *child;
3230 struct cgroup_subsys *ss;
3235 * Parse input - space separated list of subsystem names prefixed
3236 * with either + or -.
3238 buf = strstrip(buf);
3239 while ((tok = strsep(&buf, " "))) {
3242 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3243 if (!cgroup_ssid_enabled(ssid) ||
3244 strcmp(tok + 1, ss->name))
3248 enable |= 1 << ssid;
3249 disable &= ~(1 << ssid);
3250 } else if (*tok == '-') {
3251 disable |= 1 << ssid;
3252 enable &= ~(1 << ssid);
3257 } while_each_subsys_mask();
3258 if (ssid == CGROUP_SUBSYS_COUNT)
3262 cgrp = cgroup_kn_lock_live(of->kn, true);
3266 for_each_subsys(ss, ssid) {
3267 if (enable & (1 << ssid)) {
3268 if (cgrp->subtree_control & (1 << ssid)) {
3269 enable &= ~(1 << ssid);
3273 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3277 } else if (disable & (1 << ssid)) {
3278 if (!(cgrp->subtree_control & (1 << ssid))) {
3279 disable &= ~(1 << ssid);
3283 /* a child has it enabled? */
3284 cgroup_for_each_live_child(child, cgrp) {
3285 if (child->subtree_control & (1 << ssid)) {
3293 if (!enable && !disable) {
3298 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3302 /* save and update control masks and prepare csses */
3303 cgroup_save_control(cgrp);
3305 cgrp->subtree_control |= enable;
3306 cgrp->subtree_control &= ~disable;
3308 ret = cgroup_apply_control(cgrp);
3309 cgroup_finalize_control(cgrp, ret);
3313 kernfs_activate(cgrp->kn);
3315 cgroup_kn_unlock(of->kn);
3316 return ret ?: nbytes;
3320 * cgroup_enable_threaded - make @cgrp threaded
3321 * @cgrp: the target cgroup
3323 * Called when "threaded" is written to the cgroup.type interface file and
3324 * tries to make @cgrp threaded and join the parent's resource domain.
3325 * This function is never called on the root cgroup as cgroup.type doesn't
3328 static int cgroup_enable_threaded(struct cgroup *cgrp)
3330 struct cgroup *parent = cgroup_parent(cgrp);
3331 struct cgroup *dom_cgrp = parent->dom_cgrp;
3332 struct cgroup *dsct;
3333 struct cgroup_subsys_state *d_css;
3336 lockdep_assert_held(&cgroup_mutex);
3338 /* noop if already threaded */
3339 if (cgroup_is_threaded(cgrp))
3343 * If @cgroup is populated or has domain controllers enabled, it
3344 * can't be switched. While the below cgroup_can_be_thread_root()
3345 * test can catch the same conditions, that's only when @parent is
3346 * not mixable, so let's check it explicitly.
3348 if (cgroup_is_populated(cgrp) ||
3349 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3352 /* we're joining the parent's domain, ensure its validity */
3353 if (!cgroup_is_valid_domain(dom_cgrp) ||
3354 !cgroup_can_be_thread_root(dom_cgrp))
3358 * The following shouldn't cause actual migrations and should
3361 cgroup_save_control(cgrp);
3363 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3364 if (dsct == cgrp || cgroup_is_threaded(dsct))
3365 dsct->dom_cgrp = dom_cgrp;
3367 ret = cgroup_apply_control(cgrp);
3369 parent->nr_threaded_children++;
3371 cgroup_finalize_control(cgrp, ret);
3375 static int cgroup_type_show(struct seq_file *seq, void *v)
3377 struct cgroup *cgrp = seq_css(seq)->cgroup;
3379 if (cgroup_is_threaded(cgrp))
3380 seq_puts(seq, "threaded\n");
3381 else if (!cgroup_is_valid_domain(cgrp))
3382 seq_puts(seq, "domain invalid\n");
3383 else if (cgroup_is_thread_root(cgrp))
3384 seq_puts(seq, "domain threaded\n");
3386 seq_puts(seq, "domain\n");
3391 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3392 size_t nbytes, loff_t off)
3394 struct cgroup *cgrp;
3397 /* only switching to threaded mode is supported */
3398 if (strcmp(strstrip(buf), "threaded"))
3401 /* drain dying csses before we re-apply (threaded) subtree control */
3402 cgrp = cgroup_kn_lock_live(of->kn, true);
3406 /* threaded can only be enabled */
3407 ret = cgroup_enable_threaded(cgrp);
3409 cgroup_kn_unlock(of->kn);
3410 return ret ?: nbytes;
3413 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3415 struct cgroup *cgrp = seq_css(seq)->cgroup;
3416 int descendants = READ_ONCE(cgrp->max_descendants);
3418 if (descendants == INT_MAX)
3419 seq_puts(seq, "max\n");
3421 seq_printf(seq, "%d\n", descendants);
3426 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3427 char *buf, size_t nbytes, loff_t off)
3429 struct cgroup *cgrp;
3433 buf = strstrip(buf);
3434 if (!strcmp(buf, "max")) {
3435 descendants = INT_MAX;
3437 ret = kstrtoint(buf, 0, &descendants);
3442 if (descendants < 0)
3445 cgrp = cgroup_kn_lock_live(of->kn, false);
3449 cgrp->max_descendants = descendants;
3451 cgroup_kn_unlock(of->kn);
3456 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3458 struct cgroup *cgrp = seq_css(seq)->cgroup;
3459 int depth = READ_ONCE(cgrp->max_depth);
3461 if (depth == INT_MAX)
3462 seq_puts(seq, "max\n");
3464 seq_printf(seq, "%d\n", depth);
3469 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3470 char *buf, size_t nbytes, loff_t off)
3472 struct cgroup *cgrp;
3476 buf = strstrip(buf);
3477 if (!strcmp(buf, "max")) {
3480 ret = kstrtoint(buf, 0, &depth);
3488 cgrp = cgroup_kn_lock_live(of->kn, false);
3492 cgrp->max_depth = depth;
3494 cgroup_kn_unlock(of->kn);
3499 static int cgroup_events_show(struct seq_file *seq, void *v)
3501 struct cgroup *cgrp = seq_css(seq)->cgroup;
3503 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3504 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3509 static int cgroup_stat_show(struct seq_file *seq, void *v)
3511 struct cgroup *cgroup = seq_css(seq)->cgroup;
3513 seq_printf(seq, "nr_descendants %d\n",
3514 cgroup->nr_descendants);
3515 seq_printf(seq, "nr_dying_descendants %d\n",
3516 cgroup->nr_dying_descendants);
3521 static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3522 struct cgroup *cgrp, int ssid)
3524 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3525 struct cgroup_subsys_state *css;
3528 if (!ss->css_extra_stat_show)
3531 css = cgroup_tryget_css(cgrp, ss);
3535 ret = ss->css_extra_stat_show(seq, css);
3540 static int cpu_stat_show(struct seq_file *seq, void *v)
3542 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3545 cgroup_base_stat_cputime_show(seq);
3546 #ifdef CONFIG_CGROUP_SCHED
3547 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3553 static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3555 struct cgroup *cgrp = seq_css(seq)->cgroup;
3556 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3558 return psi_show(seq, psi, PSI_IO);
3560 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3562 struct cgroup *cgrp = seq_css(seq)->cgroup;
3563 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3565 return psi_show(seq, psi, PSI_MEM);
3567 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3569 struct cgroup *cgrp = seq_css(seq)->cgroup;
3570 struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3572 return psi_show(seq, psi, PSI_CPU);
3575 static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf,
3576 size_t nbytes, enum psi_res res)
3578 struct psi_trigger *new;
3579 struct cgroup *cgrp;
3580 struct psi_group *psi;
3582 cgrp = cgroup_kn_lock_live(of->kn, false);
3587 cgroup_kn_unlock(of->kn);
3589 psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3590 new = psi_trigger_create(psi, buf, nbytes, res);
3593 return PTR_ERR(new);
3596 psi_trigger_replace(&of->priv, new);
3603 static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3604 char *buf, size_t nbytes,
3607 return cgroup_pressure_write(of, buf, nbytes, PSI_IO);
3610 static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3611 char *buf, size_t nbytes,
3614 return cgroup_pressure_write(of, buf, nbytes, PSI_MEM);
3617 static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3618 char *buf, size_t nbytes,
3621 return cgroup_pressure_write(of, buf, nbytes, PSI_CPU);
3624 static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3627 return psi_trigger_poll(&of->priv, of->file, pt);
3630 static void cgroup_pressure_release(struct kernfs_open_file *of)
3632 psi_trigger_replace(&of->priv, NULL);
3634 #endif /* CONFIG_PSI */
3636 static int cgroup_freeze_show(struct seq_file *seq, void *v)
3638 struct cgroup *cgrp = seq_css(seq)->cgroup;
3640 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3645 static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3646 char *buf, size_t nbytes, loff_t off)
3648 struct cgroup *cgrp;
3652 ret = kstrtoint(strstrip(buf), 0, &freeze);
3656 if (freeze < 0 || freeze > 1)
3659 cgrp = cgroup_kn_lock_live(of->kn, false);
3663 cgroup_freeze(cgrp, freeze);
3665 cgroup_kn_unlock(of->kn);
3670 static int cgroup_file_open(struct kernfs_open_file *of)
3672 struct cftype *cft = of_cft(of);
3675 return cft->open(of);
3679 static void cgroup_file_release(struct kernfs_open_file *of)
3681 struct cftype *cft = of_cft(of);
3687 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3688 size_t nbytes, loff_t off)
3690 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3691 struct cgroup *cgrp = of->kn->parent->priv;
3692 struct cftype *cft = of_cft(of);
3693 struct cgroup_subsys_state *css;
3700 * If namespaces are delegation boundaries, disallow writes to
3701 * files in an non-init namespace root from inside the namespace
3702 * except for the files explicitly marked delegatable -
3703 * cgroup.procs and cgroup.subtree_control.
3705 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3706 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3707 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3711 return cft->write(of, buf, nbytes, off);
3714 * kernfs guarantees that a file isn't deleted with operations in
3715 * flight, which means that the matching css is and stays alive and
3716 * doesn't need to be pinned. The RCU locking is not necessary
3717 * either. It's just for the convenience of using cgroup_css().
3720 css = cgroup_css(cgrp, cft->ss);
3723 if (cft->write_u64) {
3724 unsigned long long v;
3725 ret = kstrtoull(buf, 0, &v);
3727 ret = cft->write_u64(css, cft, v);
3728 } else if (cft->write_s64) {
3730 ret = kstrtoll(buf, 0, &v);
3732 ret = cft->write_s64(css, cft, v);
3737 return ret ?: nbytes;
3740 static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
3742 struct cftype *cft = of_cft(of);
3745 return cft->poll(of, pt);
3747 return kernfs_generic_poll(of, pt);
3750 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3752 return seq_cft(seq)->seq_start(seq, ppos);
3755 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3757 return seq_cft(seq)->seq_next(seq, v, ppos);
3760 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3762 if (seq_cft(seq)->seq_stop)
3763 seq_cft(seq)->seq_stop(seq, v);
3766 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3768 struct cftype *cft = seq_cft(m);
3769 struct cgroup_subsys_state *css = seq_css(m);
3772 return cft->seq_show(m, arg);
3775 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3776 else if (cft->read_s64)
3777 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3783 static struct kernfs_ops cgroup_kf_single_ops = {
3784 .atomic_write_len = PAGE_SIZE,
3785 .open = cgroup_file_open,
3786 .release = cgroup_file_release,
3787 .write = cgroup_file_write,
3788 .poll = cgroup_file_poll,
3789 .seq_show = cgroup_seqfile_show,
3792 static struct kernfs_ops cgroup_kf_ops = {
3793 .atomic_write_len = PAGE_SIZE,
3794 .open = cgroup_file_open,
3795 .release = cgroup_file_release,
3796 .write = cgroup_file_write,
3797 .poll = cgroup_file_poll,
3798 .seq_start = cgroup_seqfile_start,
3799 .seq_next = cgroup_seqfile_next,
3800 .seq_stop = cgroup_seqfile_stop,
3801 .seq_show = cgroup_seqfile_show,
3804 /* set uid and gid of cgroup dirs and files to that of the creator */
3805 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3807 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3808 .ia_uid = current_fsuid(),
3809 .ia_gid = current_fsgid(), };
3811 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3812 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3815 return kernfs_setattr(kn, &iattr);
3818 static void cgroup_file_notify_timer(struct timer_list *timer)
3820 cgroup_file_notify(container_of(timer, struct cgroup_file,
3824 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3827 char name[CGROUP_FILE_NAME_MAX];
3828 struct kernfs_node *kn;
3829 struct lock_class_key *key = NULL;
3832 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3833 key = &cft->lockdep_key;
3835 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3836 cgroup_file_mode(cft),
3837 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
3838 0, cft->kf_ops, cft,
3843 ret = cgroup_kn_set_ugid(kn);
3849 if (cft->file_offset) {
3850 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3852 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
3854 spin_lock_irq(&cgroup_file_kn_lock);
3856 spin_unlock_irq(&cgroup_file_kn_lock);
3863 * cgroup_addrm_files - add or remove files to a cgroup directory
3864 * @css: the target css
3865 * @cgrp: the target cgroup (usually css->cgroup)
3866 * @cfts: array of cftypes to be added
3867 * @is_add: whether to add or remove
3869 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3870 * For removals, this function never fails.
3872 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3873 struct cgroup *cgrp, struct cftype cfts[],
3876 struct cftype *cft, *cft_end = NULL;
3879 lockdep_assert_held(&cgroup_mutex);
3882 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3883 /* does cft->flags tell us to skip this file on @cgrp? */
3884 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3886 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3888 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3890 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3892 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
3895 ret = cgroup_add_file(css, cgrp, cft);
3897 pr_warn("%s: failed to add %s, err=%d\n",
3898 __func__, cft->name, ret);
3904 cgroup_rm_file(cgrp, cft);
3910 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3912 struct cgroup_subsys *ss = cfts[0].ss;
3913 struct cgroup *root = &ss->root->cgrp;
3914 struct cgroup_subsys_state *css;
3917 lockdep_assert_held(&cgroup_mutex);
3919 /* add/rm files for all cgroups created before */
3920 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3921 struct cgroup *cgrp = css->cgroup;
3923 if (!(css->flags & CSS_VISIBLE))
3926 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3932 kernfs_activate(root->kn);
3936 static void cgroup_exit_cftypes(struct cftype *cfts)
3940 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3941 /* free copy for custom atomic_write_len, see init_cftypes() */
3942 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3947 /* revert flags set by cgroup core while adding @cfts */
3948 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3952 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3956 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3957 struct kernfs_ops *kf_ops;
3959 WARN_ON(cft->ss || cft->kf_ops);
3962 kf_ops = &cgroup_kf_ops;
3964 kf_ops = &cgroup_kf_single_ops;
3967 * Ugh... if @cft wants a custom max_write_len, we need to
3968 * make a copy of kf_ops to set its atomic_write_len.
3970 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3971 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3973 cgroup_exit_cftypes(cfts);
3976 kf_ops->atomic_write_len = cft->max_write_len;
3979 cft->kf_ops = kf_ops;
3986 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3988 lockdep_assert_held(&cgroup_mutex);
3990 if (!cfts || !cfts[0].ss)
3993 list_del(&cfts->node);
3994 cgroup_apply_cftypes(cfts, false);
3995 cgroup_exit_cftypes(cfts);
4000 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4001 * @cfts: zero-length name terminated array of cftypes
4003 * Unregister @cfts. Files described by @cfts are removed from all
4004 * existing cgroups and all future cgroups won't have them either. This
4005 * function can be called anytime whether @cfts' subsys is attached or not.
4007 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4010 int cgroup_rm_cftypes(struct cftype *cfts)
4014 mutex_lock(&cgroup_mutex);
4015 ret = cgroup_rm_cftypes_locked(cfts);
4016 mutex_unlock(&cgroup_mutex);
4021 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4022 * @ss: target cgroup subsystem
4023 * @cfts: zero-length name terminated array of cftypes
4025 * Register @cfts to @ss. Files described by @cfts are created for all
4026 * existing cgroups to which @ss is attached and all future cgroups will
4027 * have them too. This function can be called anytime whether @ss is
4030 * Returns 0 on successful registration, -errno on failure. Note that this
4031 * function currently returns 0 as long as @cfts registration is successful
4032 * even if some file creation attempts on existing cgroups fail.
4034 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4038 if (!cgroup_ssid_enabled(ss->id))
4041 if (!cfts || cfts[0].name[0] == '\0')
4044 ret = cgroup_init_cftypes(ss, cfts);
4048 mutex_lock(&cgroup_mutex);
4050 list_add_tail(&cfts->node, &ss->cfts);
4051 ret = cgroup_apply_cftypes(cfts, true);
4053 cgroup_rm_cftypes_locked(cfts);
4055 mutex_unlock(&cgroup_mutex);
4060 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4061 * @ss: target cgroup subsystem
4062 * @cfts: zero-length name terminated array of cftypes
4064 * Similar to cgroup_add_cftypes() but the added files are only used for
4065 * the default hierarchy.
4067 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4071 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4072 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4073 return cgroup_add_cftypes(ss, cfts);
4077 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4078 * @ss: target cgroup subsystem
4079 * @cfts: zero-length name terminated array of cftypes
4081 * Similar to cgroup_add_cftypes() but the added files are only used for
4082 * the legacy hierarchies.
4084 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4088 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4089 cft->flags |= __CFTYPE_NOT_ON_DFL;
4090 return cgroup_add_cftypes(ss, cfts);
4094 * cgroup_file_notify - generate a file modified event for a cgroup_file
4095 * @cfile: target cgroup_file
4097 * @cfile must have been obtained by setting cftype->file_offset.
4099 void cgroup_file_notify(struct cgroup_file *cfile)
4101 unsigned long flags;
4103 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4105 unsigned long last = cfile->notified_at;
4106 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4108 if (time_in_range(jiffies, last, next)) {
4109 timer_reduce(&cfile->notify_timer, next);
4111 kernfs_notify(cfile->kn);
4112 cfile->notified_at = jiffies;
4115 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4119 * css_next_child - find the next child of a given css
4120 * @pos: the current position (%NULL to initiate traversal)
4121 * @parent: css whose children to walk
4123 * This function returns the next child of @parent and should be called
4124 * under either cgroup_mutex or RCU read lock. The only requirement is
4125 * that @parent and @pos are accessible. The next sibling is guaranteed to
4126 * be returned regardless of their states.
4128 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4129 * css which finished ->css_online() is guaranteed to be visible in the
4130 * future iterations and will stay visible until the last reference is put.
4131 * A css which hasn't finished ->css_online() or already finished
4132 * ->css_offline() may show up during traversal. It's each subsystem's
4133 * responsibility to synchronize against on/offlining.
4135 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4136 struct cgroup_subsys_state *parent)
4138 struct cgroup_subsys_state *next;
4140 cgroup_assert_mutex_or_rcu_locked();
4143 * @pos could already have been unlinked from the sibling list.
4144 * Once a cgroup is removed, its ->sibling.next is no longer
4145 * updated when its next sibling changes. CSS_RELEASED is set when
4146 * @pos is taken off list, at which time its next pointer is valid,
4147 * and, as releases are serialized, the one pointed to by the next
4148 * pointer is guaranteed to not have started release yet. This
4149 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4150 * critical section, the one pointed to by its next pointer is
4151 * guaranteed to not have finished its RCU grace period even if we
4152 * have dropped rcu_read_lock() in-between iterations.
4154 * If @pos has CSS_RELEASED set, its next pointer can't be
4155 * dereferenced; however, as each css is given a monotonically
4156 * increasing unique serial number and always appended to the
4157 * sibling list, the next one can be found by walking the parent's
4158 * children until the first css with higher serial number than
4159 * @pos's. While this path can be slower, it happens iff iteration
4160 * races against release and the race window is very small.
4163 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4164 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4165 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4167 list_for_each_entry_rcu(next, &parent->children, sibling,
4168 lockdep_is_held(&cgroup_mutex))
4169 if (next->serial_nr > pos->serial_nr)
4174 * @next, if not pointing to the head, can be dereferenced and is
4177 if (&next->sibling != &parent->children)
4183 * css_next_descendant_pre - find the next descendant for pre-order walk
4184 * @pos: the current position (%NULL to initiate traversal)
4185 * @root: css whose descendants to walk
4187 * To be used by css_for_each_descendant_pre(). Find the next descendant
4188 * to visit for pre-order traversal of @root's descendants. @root is
4189 * included in the iteration and the first node to be visited.
4191 * While this function requires cgroup_mutex or RCU read locking, it
4192 * doesn't require the whole traversal to be contained in a single critical
4193 * section. This function will return the correct next descendant as long
4194 * as both @pos and @root are accessible and @pos is a descendant of @root.
4196 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4197 * css which finished ->css_online() is guaranteed to be visible in the
4198 * future iterations and will stay visible until the last reference is put.
4199 * A css which hasn't finished ->css_online() or already finished
4200 * ->css_offline() may show up during traversal. It's each subsystem's
4201 * responsibility to synchronize against on/offlining.
4203 struct cgroup_subsys_state *
4204 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4205 struct cgroup_subsys_state *root)
4207 struct cgroup_subsys_state *next;
4209 cgroup_assert_mutex_or_rcu_locked();
4211 /* if first iteration, visit @root */
4215 /* visit the first child if exists */
4216 next = css_next_child(NULL, pos);
4220 /* no child, visit my or the closest ancestor's next sibling */
4221 while (pos != root) {
4222 next = css_next_child(pos, pos->parent);
4230 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4233 * css_rightmost_descendant - return the rightmost descendant of a css
4234 * @pos: css of interest
4236 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4237 * is returned. This can be used during pre-order traversal to skip
4240 * While this function requires cgroup_mutex or RCU read locking, it
4241 * doesn't require the whole traversal to be contained in a single critical
4242 * section. This function will return the correct rightmost descendant as
4243 * long as @pos is accessible.
4245 struct cgroup_subsys_state *
4246 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4248 struct cgroup_subsys_state *last, *tmp;
4250 cgroup_assert_mutex_or_rcu_locked();
4254 /* ->prev isn't RCU safe, walk ->next till the end */
4256 css_for_each_child(tmp, last)
4263 static struct cgroup_subsys_state *
4264 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4266 struct cgroup_subsys_state *last;
4270 pos = css_next_child(NULL, pos);
4277 * css_next_descendant_post - find the next descendant for post-order walk
4278 * @pos: the current position (%NULL to initiate traversal)
4279 * @root: css whose descendants to walk
4281 * To be used by css_for_each_descendant_post(). Find the next descendant
4282 * to visit for post-order traversal of @root's descendants. @root is
4283 * included in the iteration and the last node to be visited.
4285 * While this function requires cgroup_mutex or RCU read locking, it
4286 * doesn't require the whole traversal to be contained in a single critical
4287 * section. This function will return the correct next descendant as long
4288 * as both @pos and @cgroup are accessible and @pos is a descendant of
4291 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4292 * css which finished ->css_online() is guaranteed to be visible in the
4293 * future iterations and will stay visible until the last reference is put.
4294 * A css which hasn't finished ->css_online() or already finished
4295 * ->css_offline() may show up during traversal. It's each subsystem's
4296 * responsibility to synchronize against on/offlining.
4298 struct cgroup_subsys_state *
4299 css_next_descendant_post(struct cgroup_subsys_state *pos,
4300 struct cgroup_subsys_state *root)
4302 struct cgroup_subsys_state *next;
4304 cgroup_assert_mutex_or_rcu_locked();
4306 /* if first iteration, visit leftmost descendant which may be @root */
4308 return css_leftmost_descendant(root);
4310 /* if we visited @root, we're done */
4314 /* if there's an unvisited sibling, visit its leftmost descendant */
4315 next = css_next_child(pos, pos->parent);
4317 return css_leftmost_descendant(next);
4319 /* no sibling left, visit parent */
4324 * css_has_online_children - does a css have online children
4325 * @css: the target css
4327 * Returns %true if @css has any online children; otherwise, %false. This
4328 * function can be called from any context but the caller is responsible
4329 * for synchronizing against on/offlining as necessary.
4331 bool css_has_online_children(struct cgroup_subsys_state *css)
4333 struct cgroup_subsys_state *child;
4337 css_for_each_child(child, css) {
4338 if (child->flags & CSS_ONLINE) {
4347 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4349 struct list_head *l;
4350 struct cgrp_cset_link *link;
4351 struct css_set *cset;
4353 lockdep_assert_held(&css_set_lock);
4355 /* find the next threaded cset */
4356 if (it->tcset_pos) {
4357 l = it->tcset_pos->next;
4359 if (l != it->tcset_head) {
4361 return container_of(l, struct css_set,
4362 threaded_csets_node);
4365 it->tcset_pos = NULL;
4368 /* find the next cset */
4371 if (l == it->cset_head) {
4372 it->cset_pos = NULL;
4377 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4379 link = list_entry(l, struct cgrp_cset_link, cset_link);
4385 /* initialize threaded css_set walking */
4386 if (it->flags & CSS_TASK_ITER_THREADED) {
4388 put_css_set_locked(it->cur_dcset);
4389 it->cur_dcset = cset;
4392 it->tcset_head = &cset->threaded_csets;
4393 it->tcset_pos = &cset->threaded_csets;
4400 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4401 * @it: the iterator to advance
4403 * Advance @it to the next css_set to walk.
4405 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4407 struct css_set *cset;
4409 lockdep_assert_held(&css_set_lock);
4411 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4412 while ((cset = css_task_iter_next_css_set(it))) {
4413 if (!list_empty(&cset->tasks)) {
4414 it->cur_tasks_head = &cset->tasks;
4416 } else if (!list_empty(&cset->mg_tasks)) {
4417 it->cur_tasks_head = &cset->mg_tasks;
4419 } else if (!list_empty(&cset->dying_tasks)) {
4420 it->cur_tasks_head = &cset->dying_tasks;
4425 it->task_pos = NULL;
4428 it->task_pos = it->cur_tasks_head->next;
4431 * We don't keep css_sets locked across iteration steps and thus
4432 * need to take steps to ensure that iteration can be resumed after
4433 * the lock is re-acquired. Iteration is performed at two levels -
4434 * css_sets and tasks in them.
4436 * Once created, a css_set never leaves its cgroup lists, so a
4437 * pinned css_set is guaranteed to stay put and we can resume
4438 * iteration afterwards.
4440 * Tasks may leave @cset across iteration steps. This is resolved
4441 * by registering each iterator with the css_set currently being
4442 * walked and making css_set_move_task() advance iterators whose
4443 * next task is leaving.
4446 list_del(&it->iters_node);
4447 put_css_set_locked(it->cur_cset);
4450 it->cur_cset = cset;
4451 list_add(&it->iters_node, &cset->task_iters);
4454 static void css_task_iter_skip(struct css_task_iter *it,
4455 struct task_struct *task)
4457 lockdep_assert_held(&css_set_lock);
4459 if (it->task_pos == &task->cg_list) {
4460 it->task_pos = it->task_pos->next;
4461 it->flags |= CSS_TASK_ITER_SKIPPED;
4465 static void css_task_iter_advance(struct css_task_iter *it)
4467 struct task_struct *task;
4469 lockdep_assert_held(&css_set_lock);
4473 * Advance iterator to find next entry. We go through cset
4474 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4477 if (it->flags & CSS_TASK_ITER_SKIPPED)
4478 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4480 it->task_pos = it->task_pos->next;
4482 if (it->task_pos == &it->cur_cset->tasks) {
4483 it->cur_tasks_head = &it->cur_cset->mg_tasks;
4484 it->task_pos = it->cur_tasks_head->next;
4486 if (it->task_pos == &it->cur_cset->mg_tasks) {
4487 it->cur_tasks_head = &it->cur_cset->dying_tasks;
4488 it->task_pos = it->cur_tasks_head->next;
4490 if (it->task_pos == &it->cur_cset->dying_tasks)
4491 css_task_iter_advance_css_set(it);
4493 /* called from start, proceed to the first cset */
4494 css_task_iter_advance_css_set(it);
4500 task = list_entry(it->task_pos, struct task_struct, cg_list);
4502 if (it->flags & CSS_TASK_ITER_PROCS) {
4503 /* if PROCS, skip over tasks which aren't group leaders */
4504 if (!thread_group_leader(task))
4507 /* and dying leaders w/o live member threads */
4508 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4509 !atomic_read(&task->signal->live))
4512 /* skip all dying ones */
4513 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4519 * css_task_iter_start - initiate task iteration
4520 * @css: the css to walk tasks of
4521 * @flags: CSS_TASK_ITER_* flags
4522 * @it: the task iterator to use
4524 * Initiate iteration through the tasks of @css. The caller can call
4525 * css_task_iter_next() to walk through the tasks until the function
4526 * returns NULL. On completion of iteration, css_task_iter_end() must be
4529 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4530 struct css_task_iter *it)
4532 memset(it, 0, sizeof(*it));
4534 spin_lock_irq(&css_set_lock);
4540 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4542 it->cset_pos = &css->cgroup->cset_links;
4544 it->cset_head = it->cset_pos;
4546 css_task_iter_advance(it);
4548 spin_unlock_irq(&css_set_lock);
4552 * css_task_iter_next - return the next task for the iterator
4553 * @it: the task iterator being iterated
4555 * The "next" function for task iteration. @it should have been
4556 * initialized via css_task_iter_start(). Returns NULL when the iteration
4559 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4562 put_task_struct(it->cur_task);
4563 it->cur_task = NULL;
4566 spin_lock_irq(&css_set_lock);
4568 /* @it may be half-advanced by skips, finish advancing */
4569 if (it->flags & CSS_TASK_ITER_SKIPPED)
4570 css_task_iter_advance(it);
4573 it->cur_task = list_entry(it->task_pos, struct task_struct,
4575 get_task_struct(it->cur_task);
4576 css_task_iter_advance(it);
4579 spin_unlock_irq(&css_set_lock);
4581 return it->cur_task;
4585 * css_task_iter_end - finish task iteration
4586 * @it: the task iterator to finish
4588 * Finish task iteration started by css_task_iter_start().
4590 void css_task_iter_end(struct css_task_iter *it)
4593 spin_lock_irq(&css_set_lock);
4594 list_del(&it->iters_node);
4595 put_css_set_locked(it->cur_cset);
4596 spin_unlock_irq(&css_set_lock);
4600 put_css_set(it->cur_dcset);
4603 put_task_struct(it->cur_task);
4606 static void cgroup_procs_release(struct kernfs_open_file *of)
4609 css_task_iter_end(of->priv);
4614 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4616 struct kernfs_open_file *of = s->private;
4617 struct css_task_iter *it = of->priv;
4622 return css_task_iter_next(it);
4625 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4626 unsigned int iter_flags)
4628 struct kernfs_open_file *of = s->private;
4629 struct cgroup *cgrp = seq_css(s)->cgroup;
4630 struct css_task_iter *it = of->priv;
4633 * When a seq_file is seeked, it's always traversed sequentially
4634 * from position 0, so we can simply keep iterating on !0 *pos.
4637 if (WARN_ON_ONCE((*pos)))
4638 return ERR_PTR(-EINVAL);
4640 it = kzalloc(sizeof(*it), GFP_KERNEL);
4642 return ERR_PTR(-ENOMEM);
4644 css_task_iter_start(&cgrp->self, iter_flags, it);
4645 } else if (!(*pos)) {
4646 css_task_iter_end(it);
4647 css_task_iter_start(&cgrp->self, iter_flags, it);
4649 return it->cur_task;
4651 return cgroup_procs_next(s, NULL, NULL);
4654 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4656 struct cgroup *cgrp = seq_css(s)->cgroup;
4659 * All processes of a threaded subtree belong to the domain cgroup
4660 * of the subtree. Only threads can be distributed across the
4661 * subtree. Reject reads on cgroup.procs in the subtree proper.
4662 * They're always empty anyway.
4664 if (cgroup_is_threaded(cgrp))
4665 return ERR_PTR(-EOPNOTSUPP);
4667 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4668 CSS_TASK_ITER_THREADED);
4671 static int cgroup_procs_show(struct seq_file *s, void *v)
4673 seq_printf(s, "%d\n", task_pid_vnr(v));
4677 static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
4680 struct inode *inode;
4682 lockdep_assert_held(&cgroup_mutex);
4684 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
4688 ret = inode_permission(&init_user_ns, inode, MAY_WRITE);
4693 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4694 struct cgroup *dst_cgrp,
4695 struct super_block *sb)
4697 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4698 struct cgroup *com_cgrp = src_cgrp;
4701 lockdep_assert_held(&cgroup_mutex);
4703 /* find the common ancestor */
4704 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4705 com_cgrp = cgroup_parent(com_cgrp);
4707 /* %current should be authorized to migrate to the common ancestor */
4708 ret = cgroup_may_write(com_cgrp, sb);
4713 * If namespaces are delegation boundaries, %current must be able
4714 * to see both source and destination cgroups from its namespace.
4716 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4717 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4718 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4724 static int cgroup_attach_permissions(struct cgroup *src_cgrp,
4725 struct cgroup *dst_cgrp,
4726 struct super_block *sb, bool threadgroup)
4730 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb);
4734 ret = cgroup_migrate_vet_dst(dst_cgrp);
4738 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
4744 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
4747 struct cgroup *src_cgrp, *dst_cgrp;
4748 struct task_struct *task;
4752 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4756 task = cgroup_procs_write_start(buf, threadgroup, &locked);
4757 ret = PTR_ERR_OR_ZERO(task);
4761 /* find the source cgroup */
4762 spin_lock_irq(&css_set_lock);
4763 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4764 spin_unlock_irq(&css_set_lock);
4766 /* process and thread migrations follow same delegation rule */
4767 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
4768 of->file->f_path.dentry->d_sb, threadgroup);
4772 ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
4775 cgroup_procs_write_finish(task, locked);
4777 cgroup_kn_unlock(of->kn);
4782 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4783 char *buf, size_t nbytes, loff_t off)
4785 return __cgroup_procs_write(of, buf, true) ?: nbytes;
4788 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4790 return __cgroup_procs_start(s, pos, 0);
4793 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4794 char *buf, size_t nbytes, loff_t off)
4796 return __cgroup_procs_write(of, buf, false) ?: nbytes;
4799 /* cgroup core interface files for the default hierarchy */
4800 static struct cftype cgroup_base_files[] = {
4802 .name = "cgroup.type",
4803 .flags = CFTYPE_NOT_ON_ROOT,
4804 .seq_show = cgroup_type_show,
4805 .write = cgroup_type_write,
4808 .name = "cgroup.procs",
4809 .flags = CFTYPE_NS_DELEGATABLE,
4810 .file_offset = offsetof(struct cgroup, procs_file),
4811 .release = cgroup_procs_release,
4812 .seq_start = cgroup_procs_start,
4813 .seq_next = cgroup_procs_next,
4814 .seq_show = cgroup_procs_show,
4815 .write = cgroup_procs_write,
4818 .name = "cgroup.threads",
4819 .flags = CFTYPE_NS_DELEGATABLE,
4820 .release = cgroup_procs_release,
4821 .seq_start = cgroup_threads_start,
4822 .seq_next = cgroup_procs_next,
4823 .seq_show = cgroup_procs_show,
4824 .write = cgroup_threads_write,
4827 .name = "cgroup.controllers",
4828 .seq_show = cgroup_controllers_show,
4831 .name = "cgroup.subtree_control",
4832 .flags = CFTYPE_NS_DELEGATABLE,
4833 .seq_show = cgroup_subtree_control_show,
4834 .write = cgroup_subtree_control_write,
4837 .name = "cgroup.events",
4838 .flags = CFTYPE_NOT_ON_ROOT,
4839 .file_offset = offsetof(struct cgroup, events_file),
4840 .seq_show = cgroup_events_show,
4843 .name = "cgroup.max.descendants",
4844 .seq_show = cgroup_max_descendants_show,
4845 .write = cgroup_max_descendants_write,
4848 .name = "cgroup.max.depth",
4849 .seq_show = cgroup_max_depth_show,
4850 .write = cgroup_max_depth_write,
4853 .name = "cgroup.stat",
4854 .seq_show = cgroup_stat_show,
4857 .name = "cgroup.freeze",
4858 .flags = CFTYPE_NOT_ON_ROOT,
4859 .seq_show = cgroup_freeze_show,
4860 .write = cgroup_freeze_write,
4864 .seq_show = cpu_stat_show,
4868 .name = "io.pressure",
4869 .seq_show = cgroup_io_pressure_show,
4870 .write = cgroup_io_pressure_write,
4871 .poll = cgroup_pressure_poll,
4872 .release = cgroup_pressure_release,
4875 .name = "memory.pressure",
4876 .seq_show = cgroup_memory_pressure_show,
4877 .write = cgroup_memory_pressure_write,
4878 .poll = cgroup_pressure_poll,
4879 .release = cgroup_pressure_release,
4882 .name = "cpu.pressure",
4883 .seq_show = cgroup_cpu_pressure_show,
4884 .write = cgroup_cpu_pressure_write,
4885 .poll = cgroup_pressure_poll,
4886 .release = cgroup_pressure_release,
4888 #endif /* CONFIG_PSI */
4893 * css destruction is four-stage process.
4895 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4896 * Implemented in kill_css().
4898 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4899 * and thus css_tryget_online() is guaranteed to fail, the css can be
4900 * offlined by invoking offline_css(). After offlining, the base ref is
4901 * put. Implemented in css_killed_work_fn().
4903 * 3. When the percpu_ref reaches zero, the only possible remaining
4904 * accessors are inside RCU read sections. css_release() schedules the
4907 * 4. After the grace period, the css can be freed. Implemented in
4908 * css_free_work_fn().
4910 * It is actually hairier because both step 2 and 4 require process context
4911 * and thus involve punting to css->destroy_work adding two additional
4912 * steps to the already complex sequence.
4914 static void css_free_rwork_fn(struct work_struct *work)
4916 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
4917 struct cgroup_subsys_state, destroy_rwork);
4918 struct cgroup_subsys *ss = css->ss;
4919 struct cgroup *cgrp = css->cgroup;
4921 percpu_ref_exit(&css->refcnt);
4925 struct cgroup_subsys_state *parent = css->parent;
4929 cgroup_idr_remove(&ss->css_idr, id);
4935 /* cgroup free path */
4936 atomic_dec(&cgrp->root->nr_cgrps);
4937 cgroup1_pidlist_destroy_all(cgrp);
4938 cancel_work_sync(&cgrp->release_agent_work);
4940 if (cgroup_parent(cgrp)) {
4942 * We get a ref to the parent, and put the ref when
4943 * this cgroup is being freed, so it's guaranteed
4944 * that the parent won't be destroyed before its
4947 cgroup_put(cgroup_parent(cgrp));
4948 kernfs_put(cgrp->kn);
4949 psi_cgroup_free(cgrp);
4950 cgroup_rstat_exit(cgrp);
4954 * This is root cgroup's refcnt reaching zero,
4955 * which indicates that the root should be
4958 cgroup_destroy_root(cgrp->root);
4963 static void css_release_work_fn(struct work_struct *work)
4965 struct cgroup_subsys_state *css =
4966 container_of(work, struct cgroup_subsys_state, destroy_work);
4967 struct cgroup_subsys *ss = css->ss;
4968 struct cgroup *cgrp = css->cgroup;
4970 mutex_lock(&cgroup_mutex);
4972 css->flags |= CSS_RELEASED;
4973 list_del_rcu(&css->sibling);
4976 /* css release path */
4977 if (!list_empty(&css->rstat_css_node)) {
4978 cgroup_rstat_flush(cgrp);
4979 list_del_rcu(&css->rstat_css_node);
4982 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4983 if (ss->css_released)
4984 ss->css_released(css);
4986 struct cgroup *tcgrp;
4988 /* cgroup release path */
4989 TRACE_CGROUP_PATH(release, cgrp);
4991 cgroup_rstat_flush(cgrp);
4993 spin_lock_irq(&css_set_lock);
4994 for (tcgrp = cgroup_parent(cgrp); tcgrp;
4995 tcgrp = cgroup_parent(tcgrp))
4996 tcgrp->nr_dying_descendants--;
4997 spin_unlock_irq(&css_set_lock);
5000 * There are two control paths which try to determine
5001 * cgroup from dentry without going through kernfs -
5002 * cgroupstats_build() and css_tryget_online_from_dir().
5003 * Those are supported by RCU protecting clearing of
5004 * cgrp->kn->priv backpointer.
5007 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5011 mutex_unlock(&cgroup_mutex);
5013 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5014 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5017 static void css_release(struct percpu_ref *ref)
5019 struct cgroup_subsys_state *css =
5020 container_of(ref, struct cgroup_subsys_state, refcnt);
5022 INIT_WORK(&css->destroy_work, css_release_work_fn);
5023 queue_work(cgroup_destroy_wq, &css->destroy_work);
5026 static void init_and_link_css(struct cgroup_subsys_state *css,
5027 struct cgroup_subsys *ss, struct cgroup *cgrp)
5029 lockdep_assert_held(&cgroup_mutex);
5031 cgroup_get_live(cgrp);
5033 memset(css, 0, sizeof(*css));
5037 INIT_LIST_HEAD(&css->sibling);
5038 INIT_LIST_HEAD(&css->children);
5039 INIT_LIST_HEAD(&css->rstat_css_node);
5040 css->serial_nr = css_serial_nr_next++;
5041 atomic_set(&css->online_cnt, 0);
5043 if (cgroup_parent(cgrp)) {
5044 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5045 css_get(css->parent);
5048 if (ss->css_rstat_flush)
5049 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5051 BUG_ON(cgroup_css(cgrp, ss));
5054 /* invoke ->css_online() on a new CSS and mark it online if successful */
5055 static int online_css(struct cgroup_subsys_state *css)
5057 struct cgroup_subsys *ss = css->ss;
5060 lockdep_assert_held(&cgroup_mutex);
5063 ret = ss->css_online(css);
5065 css->flags |= CSS_ONLINE;
5066 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5068 atomic_inc(&css->online_cnt);
5070 atomic_inc(&css->parent->online_cnt);
5075 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5076 static void offline_css(struct cgroup_subsys_state *css)
5078 struct cgroup_subsys *ss = css->ss;
5080 lockdep_assert_held(&cgroup_mutex);
5082 if (!(css->flags & CSS_ONLINE))
5085 if (ss->css_offline)
5086 ss->css_offline(css);
5088 css->flags &= ~CSS_ONLINE;
5089 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5091 wake_up_all(&css->cgroup->offline_waitq);
5095 * css_create - create a cgroup_subsys_state
5096 * @cgrp: the cgroup new css will be associated with
5097 * @ss: the subsys of new css
5099 * Create a new css associated with @cgrp - @ss pair. On success, the new
5100 * css is online and installed in @cgrp. This function doesn't create the
5101 * interface files. Returns 0 on success, -errno on failure.
5103 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5104 struct cgroup_subsys *ss)
5106 struct cgroup *parent = cgroup_parent(cgrp);
5107 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5108 struct cgroup_subsys_state *css;
5111 lockdep_assert_held(&cgroup_mutex);
5113 css = ss->css_alloc(parent_css);
5115 css = ERR_PTR(-ENOMEM);
5119 init_and_link_css(css, ss, cgrp);
5121 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5125 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5130 /* @css is ready to be brought online now, make it visible */
5131 list_add_tail_rcu(&css->sibling, &parent_css->children);
5132 cgroup_idr_replace(&ss->css_idr, css, css->id);
5134 err = online_css(css);
5141 list_del_rcu(&css->sibling);
5143 list_del_rcu(&css->rstat_css_node);
5144 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5145 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5146 return ERR_PTR(err);
5150 * The returned cgroup is fully initialized including its control mask, but
5151 * it isn't associated with its kernfs_node and doesn't have the control
5154 static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5157 struct cgroup_root *root = parent->root;
5158 struct cgroup *cgrp, *tcgrp;
5159 struct kernfs_node *kn;
5160 int level = parent->level + 1;
5163 /* allocate the cgroup and its ID, 0 is reserved for the root */
5164 cgrp = kzalloc(struct_size(cgrp, ancestor_ids, (level + 1)),
5167 return ERR_PTR(-ENOMEM);
5169 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5173 ret = cgroup_rstat_init(cgrp);
5175 goto out_cancel_ref;
5177 /* create the directory */
5178 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5185 init_cgroup_housekeeping(cgrp);
5187 cgrp->self.parent = &parent->self;
5189 cgrp->level = level;
5191 ret = psi_cgroup_alloc(cgrp);
5193 goto out_kernfs_remove;
5195 ret = cgroup_bpf_inherit(cgrp);
5200 * New cgroup inherits effective freeze counter, and
5201 * if the parent has to be frozen, the child has too.
5203 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5204 if (cgrp->freezer.e_freeze) {
5206 * Set the CGRP_FREEZE flag, so when a process will be
5207 * attached to the child cgroup, it will become frozen.
5208 * At this point the new cgroup is unpopulated, so we can
5209 * consider it frozen immediately.
5211 set_bit(CGRP_FREEZE, &cgrp->flags);
5212 set_bit(CGRP_FROZEN, &cgrp->flags);
5215 spin_lock_irq(&css_set_lock);
5216 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5217 cgrp->ancestor_ids[tcgrp->level] = cgroup_id(tcgrp);
5219 if (tcgrp != cgrp) {
5220 tcgrp->nr_descendants++;
5223 * If the new cgroup is frozen, all ancestor cgroups
5224 * get a new frozen descendant, but their state can't
5225 * change because of this.
5227 if (cgrp->freezer.e_freeze)
5228 tcgrp->freezer.nr_frozen_descendants++;
5231 spin_unlock_irq(&css_set_lock);
5233 if (notify_on_release(parent))
5234 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5236 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5237 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5239 cgrp->self.serial_nr = css_serial_nr_next++;
5241 /* allocation complete, commit to creation */
5242 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5243 atomic_inc(&root->nr_cgrps);
5244 cgroup_get_live(parent);
5247 * On the default hierarchy, a child doesn't automatically inherit
5248 * subtree_control from the parent. Each is configured manually.
5250 if (!cgroup_on_dfl(cgrp))
5251 cgrp->subtree_control = cgroup_control(cgrp);
5253 cgroup_propagate_control(cgrp);
5258 psi_cgroup_free(cgrp);
5260 kernfs_remove(cgrp->kn);
5262 cgroup_rstat_exit(cgrp);
5264 percpu_ref_exit(&cgrp->self.refcnt);
5267 return ERR_PTR(ret);
5270 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5272 struct cgroup *cgroup;
5276 lockdep_assert_held(&cgroup_mutex);
5278 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5279 if (cgroup->nr_descendants >= cgroup->max_descendants)
5282 if (level > cgroup->max_depth)
5293 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5295 struct cgroup *parent, *cgrp;
5298 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5299 if (strchr(name, '\n'))
5302 parent = cgroup_kn_lock_live(parent_kn, false);
5306 if (!cgroup_check_hierarchy_limits(parent)) {
5311 cgrp = cgroup_create(parent, name, mode);
5313 ret = PTR_ERR(cgrp);
5318 * This extra ref will be put in cgroup_free_fn() and guarantees
5319 * that @cgrp->kn is always accessible.
5321 kernfs_get(cgrp->kn);
5323 ret = cgroup_kn_set_ugid(cgrp->kn);
5327 ret = css_populate_dir(&cgrp->self);
5331 ret = cgroup_apply_control_enable(cgrp);
5335 TRACE_CGROUP_PATH(mkdir, cgrp);
5337 /* let's create and online css's */
5338 kernfs_activate(cgrp->kn);
5344 cgroup_destroy_locked(cgrp);
5346 cgroup_kn_unlock(parent_kn);
5351 * This is called when the refcnt of a css is confirmed to be killed.
5352 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5353 * initiate destruction and put the css ref from kill_css().
5355 static void css_killed_work_fn(struct work_struct *work)
5357 struct cgroup_subsys_state *css =
5358 container_of(work, struct cgroup_subsys_state, destroy_work);
5360 mutex_lock(&cgroup_mutex);
5365 /* @css can't go away while we're holding cgroup_mutex */
5367 } while (css && atomic_dec_and_test(&css->online_cnt));
5369 mutex_unlock(&cgroup_mutex);
5372 /* css kill confirmation processing requires process context, bounce */
5373 static void css_killed_ref_fn(struct percpu_ref *ref)
5375 struct cgroup_subsys_state *css =
5376 container_of(ref, struct cgroup_subsys_state, refcnt);
5378 if (atomic_dec_and_test(&css->online_cnt)) {
5379 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5380 queue_work(cgroup_destroy_wq, &css->destroy_work);
5385 * kill_css - destroy a css
5386 * @css: css to destroy
5388 * This function initiates destruction of @css by removing cgroup interface
5389 * files and putting its base reference. ->css_offline() will be invoked
5390 * asynchronously once css_tryget_online() is guaranteed to fail and when
5391 * the reference count reaches zero, @css will be released.
5393 static void kill_css(struct cgroup_subsys_state *css)
5395 lockdep_assert_held(&cgroup_mutex);
5397 if (css->flags & CSS_DYING)
5400 css->flags |= CSS_DYING;
5403 * This must happen before css is disassociated with its cgroup.
5404 * See seq_css() for details.
5409 * Killing would put the base ref, but we need to keep it alive
5410 * until after ->css_offline().
5415 * cgroup core guarantees that, by the time ->css_offline() is
5416 * invoked, no new css reference will be given out via
5417 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5418 * proceed to offlining css's because percpu_ref_kill() doesn't
5419 * guarantee that the ref is seen as killed on all CPUs on return.
5421 * Use percpu_ref_kill_and_confirm() to get notifications as each
5422 * css is confirmed to be seen as killed on all CPUs.
5424 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5428 * cgroup_destroy_locked - the first stage of cgroup destruction
5429 * @cgrp: cgroup to be destroyed
5431 * css's make use of percpu refcnts whose killing latency shouldn't be
5432 * exposed to userland and are RCU protected. Also, cgroup core needs to
5433 * guarantee that css_tryget_online() won't succeed by the time
5434 * ->css_offline() is invoked. To satisfy all the requirements,
5435 * destruction is implemented in the following two steps.
5437 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5438 * userland visible parts and start killing the percpu refcnts of
5439 * css's. Set up so that the next stage will be kicked off once all
5440 * the percpu refcnts are confirmed to be killed.
5442 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5443 * rest of destruction. Once all cgroup references are gone, the
5444 * cgroup is RCU-freed.
5446 * This function implements s1. After this step, @cgrp is gone as far as
5447 * the userland is concerned and a new cgroup with the same name may be
5448 * created. As cgroup doesn't care about the names internally, this
5449 * doesn't cause any problem.
5451 static int cgroup_destroy_locked(struct cgroup *cgrp)
5452 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5454 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5455 struct cgroup_subsys_state *css;
5456 struct cgrp_cset_link *link;
5459 lockdep_assert_held(&cgroup_mutex);
5462 * Only migration can raise populated from zero and we're already
5463 * holding cgroup_mutex.
5465 if (cgroup_is_populated(cgrp))
5469 * Make sure there's no live children. We can't test emptiness of
5470 * ->self.children as dead children linger on it while being
5471 * drained; otherwise, "rmdir parent/child parent" may fail.
5473 if (css_has_online_children(&cgrp->self))
5477 * Mark @cgrp and the associated csets dead. The former prevents
5478 * further task migration and child creation by disabling
5479 * cgroup_lock_live_group(). The latter makes the csets ignored by
5480 * the migration path.
5482 cgrp->self.flags &= ~CSS_ONLINE;
5484 spin_lock_irq(&css_set_lock);
5485 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5486 link->cset->dead = true;
5487 spin_unlock_irq(&css_set_lock);
5489 /* initiate massacre of all css's */
5490 for_each_css(css, ssid, cgrp)
5493 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5494 css_clear_dir(&cgrp->self);
5495 kernfs_remove(cgrp->kn);
5497 if (parent && cgroup_is_threaded(cgrp))
5498 parent->nr_threaded_children--;
5500 spin_lock_irq(&css_set_lock);
5501 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5502 tcgrp->nr_descendants--;
5503 tcgrp->nr_dying_descendants++;
5505 * If the dying cgroup is frozen, decrease frozen descendants
5506 * counters of ancestor cgroups.
5508 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5509 tcgrp->freezer.nr_frozen_descendants--;
5511 spin_unlock_irq(&css_set_lock);
5513 cgroup1_check_for_release(parent);
5515 cgroup_bpf_offline(cgrp);
5517 /* put the base reference */
5518 percpu_ref_kill(&cgrp->self.refcnt);
5523 int cgroup_rmdir(struct kernfs_node *kn)
5525 struct cgroup *cgrp;
5528 cgrp = cgroup_kn_lock_live(kn, false);
5532 ret = cgroup_destroy_locked(cgrp);
5534 TRACE_CGROUP_PATH(rmdir, cgrp);
5536 cgroup_kn_unlock(kn);
5540 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5541 .show_options = cgroup_show_options,
5542 .mkdir = cgroup_mkdir,
5543 .rmdir = cgroup_rmdir,
5544 .show_path = cgroup_show_path,
5547 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5549 struct cgroup_subsys_state *css;
5551 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5553 mutex_lock(&cgroup_mutex);
5555 idr_init(&ss->css_idr);
5556 INIT_LIST_HEAD(&ss->cfts);
5558 /* Create the root cgroup state for this subsystem */
5559 ss->root = &cgrp_dfl_root;
5560 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5561 /* We don't handle early failures gracefully */
5562 BUG_ON(IS_ERR(css));
5563 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5566 * Root csses are never destroyed and we can't initialize
5567 * percpu_ref during early init. Disable refcnting.
5569 css->flags |= CSS_NO_REF;
5572 /* allocation can't be done safely during early init */
5575 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5576 BUG_ON(css->id < 0);
5579 /* Update the init_css_set to contain a subsys
5580 * pointer to this state - since the subsystem is
5581 * newly registered, all tasks and hence the
5582 * init_css_set is in the subsystem's root cgroup. */
5583 init_css_set.subsys[ss->id] = css;
5585 have_fork_callback |= (bool)ss->fork << ss->id;
5586 have_exit_callback |= (bool)ss->exit << ss->id;
5587 have_release_callback |= (bool)ss->release << ss->id;
5588 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5590 /* At system boot, before all subsystems have been
5591 * registered, no tasks have been forked, so we don't
5592 * need to invoke fork callbacks here. */
5593 BUG_ON(!list_empty(&init_task.tasks));
5595 BUG_ON(online_css(css));
5597 mutex_unlock(&cgroup_mutex);
5601 * cgroup_init_early - cgroup initialization at system boot
5603 * Initialize cgroups at system boot, and initialize any
5604 * subsystems that request early init.
5606 int __init cgroup_init_early(void)
5608 static struct cgroup_fs_context __initdata ctx;
5609 struct cgroup_subsys *ss;
5612 ctx.root = &cgrp_dfl_root;
5613 init_cgroup_root(&ctx);
5614 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5616 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5618 for_each_subsys(ss, i) {
5619 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5620 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5621 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5623 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5624 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5627 ss->name = cgroup_subsys_name[i];
5628 if (!ss->legacy_name)
5629 ss->legacy_name = cgroup_subsys_name[i];
5632 cgroup_init_subsys(ss, true);
5638 * cgroup_init - cgroup initialization
5640 * Register cgroup filesystem and /proc file, and initialize
5641 * any subsystems that didn't request early init.
5643 int __init cgroup_init(void)
5645 struct cgroup_subsys *ss;
5648 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5649 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5650 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5652 cgroup_rstat_boot();
5655 * The latency of the synchronize_rcu() is too high for cgroups,
5656 * avoid it at the cost of forcing all readers into the slow path.
5658 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5660 get_user_ns(init_cgroup_ns.user_ns);
5662 mutex_lock(&cgroup_mutex);
5665 * Add init_css_set to the hash table so that dfl_root can link to
5668 hash_add(css_set_table, &init_css_set.hlist,
5669 css_set_hash(init_css_set.subsys));
5671 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5673 mutex_unlock(&cgroup_mutex);
5675 for_each_subsys(ss, ssid) {
5676 if (ss->early_init) {
5677 struct cgroup_subsys_state *css =
5678 init_css_set.subsys[ss->id];
5680 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5682 BUG_ON(css->id < 0);
5684 cgroup_init_subsys(ss, false);
5687 list_add_tail(&init_css_set.e_cset_node[ssid],
5688 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5691 * Setting dfl_root subsys_mask needs to consider the
5692 * disabled flag and cftype registration needs kmalloc,
5693 * both of which aren't available during early_init.
5695 if (!cgroup_ssid_enabled(ssid))
5698 if (cgroup1_ssid_disabled(ssid))
5699 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5702 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5704 /* implicit controllers must be threaded too */
5705 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5707 if (ss->implicit_on_dfl)
5708 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5709 else if (!ss->dfl_cftypes)
5710 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5713 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5715 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5716 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5718 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5719 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5723 ss->bind(init_css_set.subsys[ssid]);
5725 mutex_lock(&cgroup_mutex);
5726 css_populate_dir(init_css_set.subsys[ssid]);
5727 mutex_unlock(&cgroup_mutex);
5730 /* init_css_set.subsys[] has been updated, re-hash */
5731 hash_del(&init_css_set.hlist);
5732 hash_add(css_set_table, &init_css_set.hlist,
5733 css_set_hash(init_css_set.subsys));
5735 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5736 WARN_ON(register_filesystem(&cgroup_fs_type));
5737 WARN_ON(register_filesystem(&cgroup2_fs_type));
5738 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
5739 #ifdef CONFIG_CPUSETS
5740 WARN_ON(register_filesystem(&cpuset_fs_type));
5746 static int __init cgroup_wq_init(void)
5749 * There isn't much point in executing destruction path in
5750 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5751 * Use 1 for @max_active.
5753 * We would prefer to do this in cgroup_init() above, but that
5754 * is called before init_workqueues(): so leave this until after.
5756 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5757 BUG_ON(!cgroup_destroy_wq);
5760 core_initcall(cgroup_wq_init);
5762 void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
5764 struct kernfs_node *kn;
5766 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
5769 kernfs_path(kn, buf, buflen);
5774 * proc_cgroup_show()
5775 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5776 * - Used for /proc/<pid>/cgroup.
5778 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5779 struct pid *pid, struct task_struct *tsk)
5783 struct cgroup_root *root;
5786 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5790 mutex_lock(&cgroup_mutex);
5791 spin_lock_irq(&css_set_lock);
5793 for_each_root(root) {
5794 struct cgroup_subsys *ss;
5795 struct cgroup *cgrp;
5796 int ssid, count = 0;
5798 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5801 seq_printf(m, "%d:", root->hierarchy_id);
5802 if (root != &cgrp_dfl_root)
5803 for_each_subsys(ss, ssid)
5804 if (root->subsys_mask & (1 << ssid))
5805 seq_printf(m, "%s%s", count++ ? "," : "",
5807 if (strlen(root->name))
5808 seq_printf(m, "%sname=%s", count ? "," : "",
5812 cgrp = task_cgroup_from_root(tsk, root);
5815 * On traditional hierarchies, all zombie tasks show up as
5816 * belonging to the root cgroup. On the default hierarchy,
5817 * while a zombie doesn't show up in "cgroup.procs" and
5818 * thus can't be migrated, its /proc/PID/cgroup keeps
5819 * reporting the cgroup it belonged to before exiting. If
5820 * the cgroup is removed before the zombie is reaped,
5821 * " (deleted)" is appended to the cgroup path.
5823 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5824 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5825 current->nsproxy->cgroup_ns);
5826 if (retval >= PATH_MAX)
5827 retval = -ENAMETOOLONG;
5836 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5837 seq_puts(m, " (deleted)\n");
5844 spin_unlock_irq(&css_set_lock);
5845 mutex_unlock(&cgroup_mutex);
5852 * cgroup_fork - initialize cgroup related fields during copy_process()
5853 * @child: pointer to task_struct of forking parent process.
5855 * A task is associated with the init_css_set until cgroup_post_fork()
5856 * attaches it to the target css_set.
5858 void cgroup_fork(struct task_struct *child)
5860 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5861 INIT_LIST_HEAD(&child->cg_list);
5864 static struct cgroup *cgroup_get_from_file(struct file *f)
5866 struct cgroup_subsys_state *css;
5867 struct cgroup *cgrp;
5869 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
5871 return ERR_CAST(css);
5874 if (!cgroup_on_dfl(cgrp)) {
5876 return ERR_PTR(-EBADF);
5883 * cgroup_css_set_fork - find or create a css_set for a child process
5884 * @kargs: the arguments passed to create the child process
5886 * This functions finds or creates a new css_set which the child
5887 * process will be attached to in cgroup_post_fork(). By default,
5888 * the child process will be given the same css_set as its parent.
5890 * If CLONE_INTO_CGROUP is specified this function will try to find an
5891 * existing css_set which includes the requested cgroup and if not create
5892 * a new css_set that the child will be attached to later. If this function
5893 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
5894 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
5895 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
5896 * to the target cgroup.
5898 static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
5899 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
5902 struct cgroup *dst_cgrp = NULL;
5903 struct css_set *cset;
5904 struct super_block *sb;
5907 if (kargs->flags & CLONE_INTO_CGROUP)
5908 mutex_lock(&cgroup_mutex);
5910 cgroup_threadgroup_change_begin(current);
5912 spin_lock_irq(&css_set_lock);
5913 cset = task_css_set(current);
5915 spin_unlock_irq(&css_set_lock);
5917 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
5922 f = fget_raw(kargs->cgroup);
5927 sb = f->f_path.dentry->d_sb;
5929 dst_cgrp = cgroup_get_from_file(f);
5930 if (IS_ERR(dst_cgrp)) {
5931 ret = PTR_ERR(dst_cgrp);
5936 if (cgroup_is_dead(dst_cgrp)) {
5942 * Verify that we the target cgroup is writable for us. This is
5943 * usually done by the vfs layer but since we're not going through
5944 * the vfs layer here we need to do it "manually".
5946 ret = cgroup_may_write(dst_cgrp, sb);
5950 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
5951 !(kargs->flags & CLONE_THREAD));
5955 kargs->cset = find_css_set(cset, dst_cgrp);
5963 kargs->cgrp = dst_cgrp;
5967 cgroup_threadgroup_change_end(current);
5968 mutex_unlock(&cgroup_mutex);
5972 cgroup_put(dst_cgrp);
5975 put_css_set(kargs->cset);
5980 * cgroup_css_set_put_fork - drop references we took during fork
5981 * @kargs: the arguments passed to create the child process
5983 * Drop references to the prepared css_set and target cgroup if
5984 * CLONE_INTO_CGROUP was requested.
5986 static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
5987 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
5989 cgroup_threadgroup_change_end(current);
5991 if (kargs->flags & CLONE_INTO_CGROUP) {
5992 struct cgroup *cgrp = kargs->cgrp;
5993 struct css_set *cset = kargs->cset;
5995 mutex_unlock(&cgroup_mutex);
6010 * cgroup_can_fork - called on a new task before the process is exposed
6011 * @child: the child process
6013 * This prepares a new css_set for the child process which the child will
6014 * be attached to in cgroup_post_fork().
6015 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6016 * callback returns an error, the fork aborts with that error code. This
6017 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6019 int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6021 struct cgroup_subsys *ss;
6024 ret = cgroup_css_set_fork(kargs);
6028 do_each_subsys_mask(ss, i, have_canfork_callback) {
6029 ret = ss->can_fork(child, kargs->cset);
6032 } while_each_subsys_mask();
6037 for_each_subsys(ss, j) {
6040 if (ss->cancel_fork)
6041 ss->cancel_fork(child, kargs->cset);
6044 cgroup_css_set_put_fork(kargs);
6050 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6051 * @child: the child process
6052 * @kargs: the arguments passed to create the child process
6054 * This calls the cancel_fork() callbacks if a fork failed *after*
6055 * cgroup_can_fork() succeeded and cleans up references we took to
6056 * prepare a new css_set for the child process in cgroup_can_fork().
6058 void cgroup_cancel_fork(struct task_struct *child,
6059 struct kernel_clone_args *kargs)
6061 struct cgroup_subsys *ss;
6064 for_each_subsys(ss, i)
6065 if (ss->cancel_fork)
6066 ss->cancel_fork(child, kargs->cset);
6068 cgroup_css_set_put_fork(kargs);
6072 * cgroup_post_fork - finalize cgroup setup for the child process
6073 * @child: the child process
6075 * Attach the child process to its css_set calling the subsystem fork()
6078 void cgroup_post_fork(struct task_struct *child,
6079 struct kernel_clone_args *kargs)
6080 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6082 struct cgroup_subsys *ss;
6083 struct css_set *cset;
6089 spin_lock_irq(&css_set_lock);
6091 /* init tasks are special, only link regular threads */
6092 if (likely(child->pid)) {
6093 WARN_ON_ONCE(!list_empty(&child->cg_list));
6095 css_set_move_task(child, NULL, cset, false);
6102 * If the cgroup has to be frozen, the new task has too. Let's set
6103 * the JOBCTL_TRAP_FREEZE jobctl bit to get the task into the
6106 if (unlikely(cgroup_task_freeze(child))) {
6107 spin_lock(&child->sighand->siglock);
6108 WARN_ON_ONCE(child->frozen);
6109 child->jobctl |= JOBCTL_TRAP_FREEZE;
6110 spin_unlock(&child->sighand->siglock);
6113 * Calling cgroup_update_frozen() isn't required here,
6114 * because it will be called anyway a bit later from
6115 * do_freezer_trap(). So we avoid cgroup's transient switch
6116 * from the frozen state and back.
6120 spin_unlock_irq(&css_set_lock);
6123 * Call ss->fork(). This must happen after @child is linked on
6124 * css_set; otherwise, @child might change state between ->fork()
6125 * and addition to css_set.
6127 do_each_subsys_mask(ss, i, have_fork_callback) {
6129 } while_each_subsys_mask();
6131 /* Make the new cset the root_cset of the new cgroup namespace. */
6132 if (kargs->flags & CLONE_NEWCGROUP) {
6133 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6136 child->nsproxy->cgroup_ns->root_cset = cset;
6140 cgroup_css_set_put_fork(kargs);
6144 * cgroup_exit - detach cgroup from exiting task
6145 * @tsk: pointer to task_struct of exiting process
6147 * Description: Detach cgroup from @tsk.
6150 void cgroup_exit(struct task_struct *tsk)
6152 struct cgroup_subsys *ss;
6153 struct css_set *cset;
6156 spin_lock_irq(&css_set_lock);
6158 WARN_ON_ONCE(list_empty(&tsk->cg_list));
6159 cset = task_css_set(tsk);
6160 css_set_move_task(tsk, cset, NULL, false);
6161 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6164 WARN_ON_ONCE(cgroup_task_frozen(tsk));
6165 if (unlikely(cgroup_task_freeze(tsk)))
6166 cgroup_update_frozen(task_dfl_cgroup(tsk));
6168 spin_unlock_irq(&css_set_lock);
6170 /* see cgroup_post_fork() for details */
6171 do_each_subsys_mask(ss, i, have_exit_callback) {
6173 } while_each_subsys_mask();
6176 void cgroup_release(struct task_struct *task)
6178 struct cgroup_subsys *ss;
6181 do_each_subsys_mask(ss, ssid, have_release_callback) {
6183 } while_each_subsys_mask();
6185 spin_lock_irq(&css_set_lock);
6186 css_set_skip_task_iters(task_css_set(task), task);
6187 list_del_init(&task->cg_list);
6188 spin_unlock_irq(&css_set_lock);
6191 void cgroup_free(struct task_struct *task)
6193 struct css_set *cset = task_css_set(task);
6197 static int __init cgroup_disable(char *str)
6199 struct cgroup_subsys *ss;
6203 while ((token = strsep(&str, ",")) != NULL) {
6207 for_each_subsys(ss, i) {
6208 if (strcmp(token, ss->name) &&
6209 strcmp(token, ss->legacy_name))
6212 static_branch_disable(cgroup_subsys_enabled_key[i]);
6213 pr_info("Disabling %s control group subsystem\n",
6219 __setup("cgroup_disable=", cgroup_disable);
6221 void __init __weak enable_debug_cgroup(void) { }
6223 static int __init enable_cgroup_debug(char *str)
6225 cgroup_debug = true;
6226 enable_debug_cgroup();
6229 __setup("cgroup_debug", enable_cgroup_debug);
6232 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6233 * @dentry: directory dentry of interest
6234 * @ss: subsystem of interest
6236 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6237 * to get the corresponding css and return it. If such css doesn't exist
6238 * or can't be pinned, an ERR_PTR value is returned.
6240 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6241 struct cgroup_subsys *ss)
6243 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6244 struct file_system_type *s_type = dentry->d_sb->s_type;
6245 struct cgroup_subsys_state *css = NULL;
6246 struct cgroup *cgrp;
6248 /* is @dentry a cgroup dir? */
6249 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6250 !kn || kernfs_type(kn) != KERNFS_DIR)
6251 return ERR_PTR(-EBADF);
6256 * This path doesn't originate from kernfs and @kn could already
6257 * have been or be removed at any point. @kn->priv is RCU
6258 * protected for this access. See css_release_work_fn() for details.
6260 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6262 css = cgroup_css(cgrp, ss);
6264 if (!css || !css_tryget_online(css))
6265 css = ERR_PTR(-ENOENT);
6272 * css_from_id - lookup css by id
6273 * @id: the cgroup id
6274 * @ss: cgroup subsys to be looked into
6276 * Returns the css if there's valid one with @id, otherwise returns NULL.
6277 * Should be called under rcu_read_lock().
6279 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6281 WARN_ON_ONCE(!rcu_read_lock_held());
6282 return idr_find(&ss->css_idr, id);
6286 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6287 * @path: path on the default hierarchy
6289 * Find the cgroup at @path on the default hierarchy, increment its
6290 * reference count and return it. Returns pointer to the found cgroup on
6291 * success, ERR_PTR(-ENOENT) if @path doesn't exist and ERR_PTR(-ENOTDIR)
6292 * if @path points to a non-directory.
6294 struct cgroup *cgroup_get_from_path(const char *path)
6296 struct kernfs_node *kn;
6297 struct cgroup *cgrp;
6299 mutex_lock(&cgroup_mutex);
6301 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6303 if (kernfs_type(kn) == KERNFS_DIR) {
6305 cgroup_get_live(cgrp);
6307 cgrp = ERR_PTR(-ENOTDIR);
6311 cgrp = ERR_PTR(-ENOENT);
6314 mutex_unlock(&cgroup_mutex);
6317 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6320 * cgroup_get_from_fd - get a cgroup pointer from a fd
6321 * @fd: fd obtained by open(cgroup2_dir)
6323 * Find the cgroup from a fd which should be obtained
6324 * by opening a cgroup directory. Returns a pointer to the
6325 * cgroup on success. ERR_PTR is returned if the cgroup
6328 struct cgroup *cgroup_get_from_fd(int fd)
6330 struct cgroup *cgrp;
6335 return ERR_PTR(-EBADF);
6337 cgrp = cgroup_get_from_file(f);
6341 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6343 static u64 power_of_ten(int power)
6352 * cgroup_parse_float - parse a floating number
6353 * @input: input string
6354 * @dec_shift: number of decimal digits to shift
6357 * Parse a decimal floating point number in @input and store the result in
6358 * @v with decimal point right shifted @dec_shift times. For example, if
6359 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6360 * Returns 0 on success, -errno otherwise.
6362 * There's nothing cgroup specific about this function except that it's
6363 * currently the only user.
6365 int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6367 s64 whole, frac = 0;
6368 int fstart = 0, fend = 0, flen;
6370 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6375 flen = fend > fstart ? fend - fstart : 0;
6376 if (flen < dec_shift)
6377 frac *= power_of_ten(dec_shift - flen);
6379 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6381 *v = whole * power_of_ten(dec_shift) + frac;
6386 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6387 * definition in cgroup-defs.h.
6389 #ifdef CONFIG_SOCK_CGROUP_DATA
6391 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6393 DEFINE_SPINLOCK(cgroup_sk_update_lock);
6394 static bool cgroup_sk_alloc_disabled __read_mostly;
6396 void cgroup_sk_alloc_disable(void)
6398 if (cgroup_sk_alloc_disabled)
6400 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6401 cgroup_sk_alloc_disabled = true;
6406 #define cgroup_sk_alloc_disabled false
6410 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6412 if (cgroup_sk_alloc_disabled) {
6413 skcd->no_refcnt = 1;
6417 /* Don't associate the sock with unrelated interrupted task's cgroup. */
6424 struct css_set *cset;
6426 cset = task_css_set(current);
6427 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6428 skcd->val = (unsigned long)cset->dfl_cgrp;
6429 cgroup_bpf_get(cset->dfl_cgrp);
6438 void cgroup_sk_clone(struct sock_cgroup_data *skcd)
6441 if (skcd->no_refcnt)
6444 * We might be cloning a socket which is left in an empty
6445 * cgroup and the cgroup might have already been rmdir'd.
6446 * Don't use cgroup_get_live().
6448 cgroup_get(sock_cgroup_ptr(skcd));
6449 cgroup_bpf_get(sock_cgroup_ptr(skcd));
6453 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6455 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6457 if (skcd->no_refcnt)
6459 cgroup_bpf_put(cgrp);
6463 #endif /* CONFIG_SOCK_CGROUP_DATA */
6465 #ifdef CONFIG_CGROUP_BPF
6466 int cgroup_bpf_attach(struct cgroup *cgrp,
6467 struct bpf_prog *prog, struct bpf_prog *replace_prog,
6468 struct bpf_cgroup_link *link,
6469 enum bpf_attach_type type,
6474 mutex_lock(&cgroup_mutex);
6475 ret = __cgroup_bpf_attach(cgrp, prog, replace_prog, link, type, flags);
6476 mutex_unlock(&cgroup_mutex);
6480 int cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
6481 enum bpf_attach_type type)
6485 mutex_lock(&cgroup_mutex);
6486 ret = __cgroup_bpf_detach(cgrp, prog, NULL, type);
6487 mutex_unlock(&cgroup_mutex);
6491 int cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
6492 union bpf_attr __user *uattr)
6496 mutex_lock(&cgroup_mutex);
6497 ret = __cgroup_bpf_query(cgrp, attr, uattr);
6498 mutex_unlock(&cgroup_mutex);
6501 #endif /* CONFIG_CGROUP_BPF */
6504 static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6505 ssize_t size, const char *prefix)
6510 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6511 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
6515 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
6517 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
6519 if (WARN_ON(ret >= size))
6526 static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
6529 struct cgroup_subsys *ss;
6533 ret = show_delegatable_files(cgroup_base_files, buf, PAGE_SIZE - ret,
6536 for_each_subsys(ss, ssid)
6537 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
6539 cgroup_subsys_name[ssid]);
6543 static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
6545 static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
6548 return snprintf(buf, PAGE_SIZE,
6550 "memory_localevents\n"
6551 "memory_recursiveprot\n");
6553 static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
6555 static struct attribute *cgroup_sysfs_attrs[] = {
6556 &cgroup_delegate_attr.attr,
6557 &cgroup_features_attr.attr,
6561 static const struct attribute_group cgroup_sysfs_attr_group = {
6562 .attrs = cgroup_sysfs_attrs,
6566 static int __init cgroup_sysfs_init(void)
6568 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
6570 subsys_initcall(cgroup_sysfs_init);
6572 #endif /* CONFIG_SYSFS */