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 struct percpu_rw_semaphore 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);
157 * The default hierarchy, reserved for the subsystems that are otherwise
158 * unattached - it never has more than a single cgroup, and all tasks are
159 * part of that cgroup.
161 struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
162 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
165 * The default hierarchy always exists but is hidden until mounted for the
166 * first time. This is for backward compatibility.
168 static bool cgrp_dfl_visible;
170 /* some controllers are not supported in the default hierarchy */
171 static u16 cgrp_dfl_inhibit_ss_mask;
173 /* some controllers are implicitly enabled on the default hierarchy */
174 static u16 cgrp_dfl_implicit_ss_mask;
176 /* some controllers can be threaded on the default hierarchy */
177 static u16 cgrp_dfl_threaded_ss_mask;
179 /* The list of hierarchy roots */
180 LIST_HEAD(cgroup_roots);
181 static int cgroup_root_count;
183 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
184 static DEFINE_IDR(cgroup_hierarchy_idr);
187 * Assign a monotonically increasing serial number to csses. It guarantees
188 * cgroups with bigger numbers are newer than those with smaller numbers.
189 * Also, as csses are always appended to the parent's ->children list, it
190 * guarantees that sibling csses are always sorted in the ascending serial
191 * number order on the list. Protected by cgroup_mutex.
193 static u64 css_serial_nr_next = 1;
196 * These bitmasks identify subsystems with specific features to avoid
197 * having to do iterative checks repeatedly.
199 static u16 have_fork_callback __read_mostly;
200 static u16 have_exit_callback __read_mostly;
201 static u16 have_release_callback __read_mostly;
202 static u16 have_canfork_callback __read_mostly;
204 /* cgroup namespace for init task */
205 struct cgroup_namespace init_cgroup_ns = {
206 .count = REFCOUNT_INIT(2),
207 .user_ns = &init_user_ns,
208 .ns.ops = &cgroupns_operations,
209 .ns.inum = PROC_CGROUP_INIT_INO,
210 .root_cset = &init_css_set,
213 static struct file_system_type cgroup2_fs_type;
214 static struct cftype cgroup_base_files[];
216 static int cgroup_apply_control(struct cgroup *cgrp);
217 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
218 static void css_task_iter_advance(struct css_task_iter *it);
219 static int cgroup_destroy_locked(struct cgroup *cgrp);
220 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
221 struct cgroup_subsys *ss);
222 static void css_release(struct percpu_ref *ref);
223 static void kill_css(struct cgroup_subsys_state *css);
224 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
225 struct cgroup *cgrp, struct cftype cfts[],
229 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
230 * @ssid: subsys ID of interest
232 * cgroup_subsys_enabled() can only be used with literal subsys names which
233 * is fine for individual subsystems but unsuitable for cgroup core. This
234 * is slower static_key_enabled() based test indexed by @ssid.
236 bool cgroup_ssid_enabled(int ssid)
238 if (CGROUP_SUBSYS_COUNT == 0)
241 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
245 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
246 * @cgrp: the cgroup of interest
248 * The default hierarchy is the v2 interface of cgroup and this function
249 * can be used to test whether a cgroup is on the default hierarchy for
250 * cases where a subsystem should behave differnetly depending on the
253 * The set of behaviors which change on the default hierarchy are still
254 * being determined and the mount option is prefixed with __DEVEL__.
256 * List of changed behaviors:
258 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
259 * and "name" are disallowed.
261 * - When mounting an existing superblock, mount options should match.
263 * - Remount is disallowed.
265 * - rename(2) is disallowed.
267 * - "tasks" is removed. Everything should be at process granularity. Use
268 * "cgroup.procs" instead.
270 * - "cgroup.procs" is not sorted. pids will be unique unless they got
271 * recycled inbetween reads.
273 * - "release_agent" and "notify_on_release" are removed. Replacement
274 * notification mechanism will be implemented.
276 * - "cgroup.clone_children" is removed.
278 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
279 * and its descendants contain no task; otherwise, 1. The file also
280 * generates kernfs notification which can be monitored through poll and
281 * [di]notify when the value of the file changes.
283 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
284 * take masks of ancestors with non-empty cpus/mems, instead of being
285 * moved to an ancestor.
287 * - cpuset: a task can be moved into an empty cpuset, and again it takes
288 * masks of ancestors.
290 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
293 * - blkcg: blk-throttle becomes properly hierarchical.
295 * - debug: disallowed on the default hierarchy.
297 bool cgroup_on_dfl(const struct cgroup *cgrp)
299 return cgrp->root == &cgrp_dfl_root;
302 /* IDR wrappers which synchronize using cgroup_idr_lock */
303 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
308 idr_preload(gfp_mask);
309 spin_lock_bh(&cgroup_idr_lock);
310 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
311 spin_unlock_bh(&cgroup_idr_lock);
316 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
320 spin_lock_bh(&cgroup_idr_lock);
321 ret = idr_replace(idr, ptr, id);
322 spin_unlock_bh(&cgroup_idr_lock);
326 static void cgroup_idr_remove(struct idr *idr, int id)
328 spin_lock_bh(&cgroup_idr_lock);
330 spin_unlock_bh(&cgroup_idr_lock);
333 static bool cgroup_has_tasks(struct cgroup *cgrp)
335 return cgrp->nr_populated_csets;
338 bool cgroup_is_threaded(struct cgroup *cgrp)
340 return cgrp->dom_cgrp != cgrp;
343 /* can @cgrp host both domain and threaded children? */
344 static bool cgroup_is_mixable(struct cgroup *cgrp)
347 * Root isn't under domain level resource control exempting it from
348 * the no-internal-process constraint, so it can serve as a thread
349 * root and a parent of resource domains at the same time.
351 return !cgroup_parent(cgrp);
354 /* can @cgrp become a thread root? should always be true for a thread root */
355 static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
357 /* mixables don't care */
358 if (cgroup_is_mixable(cgrp))
361 /* domain roots can't be nested under threaded */
362 if (cgroup_is_threaded(cgrp))
365 /* can only have either domain or threaded children */
366 if (cgrp->nr_populated_domain_children)
369 /* and no domain controllers can be enabled */
370 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
376 /* is @cgrp root of a threaded subtree? */
377 bool cgroup_is_thread_root(struct cgroup *cgrp)
379 /* thread root should be a domain */
380 if (cgroup_is_threaded(cgrp))
383 /* a domain w/ threaded children is a thread root */
384 if (cgrp->nr_threaded_children)
388 * A domain which has tasks and explicit threaded controllers
389 * enabled is a thread root.
391 if (cgroup_has_tasks(cgrp) &&
392 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
398 /* a domain which isn't connected to the root w/o brekage can't be used */
399 static bool cgroup_is_valid_domain(struct cgroup *cgrp)
401 /* the cgroup itself can be a thread root */
402 if (cgroup_is_threaded(cgrp))
405 /* but the ancestors can't be unless mixable */
406 while ((cgrp = cgroup_parent(cgrp))) {
407 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
409 if (cgroup_is_threaded(cgrp))
416 /* subsystems visibly enabled on a cgroup */
417 static u16 cgroup_control(struct cgroup *cgrp)
419 struct cgroup *parent = cgroup_parent(cgrp);
420 u16 root_ss_mask = cgrp->root->subsys_mask;
423 u16 ss_mask = parent->subtree_control;
425 /* threaded cgroups can only have threaded controllers */
426 if (cgroup_is_threaded(cgrp))
427 ss_mask &= cgrp_dfl_threaded_ss_mask;
431 if (cgroup_on_dfl(cgrp))
432 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
433 cgrp_dfl_implicit_ss_mask);
437 /* subsystems enabled on a cgroup */
438 static u16 cgroup_ss_mask(struct cgroup *cgrp)
440 struct cgroup *parent = cgroup_parent(cgrp);
443 u16 ss_mask = parent->subtree_ss_mask;
445 /* threaded cgroups can only have threaded controllers */
446 if (cgroup_is_threaded(cgrp))
447 ss_mask &= cgrp_dfl_threaded_ss_mask;
451 return cgrp->root->subsys_mask;
455 * cgroup_css - obtain a cgroup's css for the specified subsystem
456 * @cgrp: the cgroup of interest
457 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
459 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
460 * function must be called either under cgroup_mutex or rcu_read_lock() and
461 * the caller is responsible for pinning the returned css if it wants to
462 * keep accessing it outside the said locks. This function may return
463 * %NULL if @cgrp doesn't have @subsys_id enabled.
465 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
466 struct cgroup_subsys *ss)
469 return rcu_dereference_check(cgrp->subsys[ss->id],
470 lockdep_is_held(&cgroup_mutex));
476 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
477 * @cgrp: the cgroup of interest
478 * @ss: the subsystem of interest
480 * Find and get @cgrp's css assocaited with @ss. If the css doesn't exist
481 * or is offline, %NULL is returned.
483 static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
484 struct cgroup_subsys *ss)
486 struct cgroup_subsys_state *css;
489 css = cgroup_css(cgrp, ss);
490 if (!css || !css_tryget_online(css))
498 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
499 * @cgrp: the cgroup of interest
500 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
502 * Similar to cgroup_css() but returns the effective css, which is defined
503 * as the matching css of the nearest ancestor including self which has @ss
504 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
505 * function is guaranteed to return non-NULL css.
507 static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
508 struct cgroup_subsys *ss)
510 lockdep_assert_held(&cgroup_mutex);
516 * This function is used while updating css associations and thus
517 * can't test the csses directly. Test ss_mask.
519 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
520 cgrp = cgroup_parent(cgrp);
525 return cgroup_css(cgrp, ss);
529 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
530 * @cgrp: the cgroup of interest
531 * @ss: the subsystem of interest
533 * Find and get the effective css of @cgrp for @ss. The effective css is
534 * defined as the matching css of the nearest ancestor including self which
535 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
536 * the root css is returned, so this function always returns a valid css.
538 * The returned css is not guaranteed to be online, and therefore it is the
539 * callers responsiblity to tryget a reference for it.
541 struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
542 struct cgroup_subsys *ss)
544 struct cgroup_subsys_state *css;
547 css = cgroup_css(cgrp, ss);
551 cgrp = cgroup_parent(cgrp);
554 return init_css_set.subsys[ss->id];
558 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
559 * @cgrp: the cgroup of interest
560 * @ss: the subsystem of interest
562 * Find and get the effective css of @cgrp for @ss. The effective css is
563 * defined as the matching css of the nearest ancestor including self which
564 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
565 * the root css is returned, so this function always returns a valid css.
566 * The returned css must be put using css_put().
568 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
569 struct cgroup_subsys *ss)
571 struct cgroup_subsys_state *css;
576 css = cgroup_css(cgrp, ss);
578 if (css && css_tryget_online(css))
580 cgrp = cgroup_parent(cgrp);
583 css = init_css_set.subsys[ss->id];
590 static void cgroup_get_live(struct cgroup *cgrp)
592 WARN_ON_ONCE(cgroup_is_dead(cgrp));
593 css_get(&cgrp->self);
597 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
598 * is responsible for taking the css_set_lock.
599 * @cgrp: the cgroup in question
601 int __cgroup_task_count(const struct cgroup *cgrp)
604 struct cgrp_cset_link *link;
606 lockdep_assert_held(&css_set_lock);
608 list_for_each_entry(link, &cgrp->cset_links, cset_link)
609 count += link->cset->nr_tasks;
615 * cgroup_task_count - count the number of tasks in a cgroup.
616 * @cgrp: the cgroup in question
618 int cgroup_task_count(const struct cgroup *cgrp)
622 spin_lock_irq(&css_set_lock);
623 count = __cgroup_task_count(cgrp);
624 spin_unlock_irq(&css_set_lock);
629 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
631 struct cgroup *cgrp = of->kn->parent->priv;
632 struct cftype *cft = of_cft(of);
635 * This is open and unprotected implementation of cgroup_css().
636 * seq_css() is only called from a kernfs file operation which has
637 * an active reference on the file. Because all the subsystem
638 * files are drained before a css is disassociated with a cgroup,
639 * the matching css from the cgroup's subsys table is guaranteed to
640 * be and stay valid until the enclosing operation is complete.
643 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
647 EXPORT_SYMBOL_GPL(of_css);
650 * for_each_css - iterate all css's of a cgroup
651 * @css: the iteration cursor
652 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
653 * @cgrp: the target cgroup to iterate css's of
655 * Should be called under cgroup_[tree_]mutex.
657 #define for_each_css(css, ssid, cgrp) \
658 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
659 if (!((css) = rcu_dereference_check( \
660 (cgrp)->subsys[(ssid)], \
661 lockdep_is_held(&cgroup_mutex)))) { } \
665 * for_each_e_css - iterate all effective css's of a cgroup
666 * @css: the iteration cursor
667 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
668 * @cgrp: the target cgroup to iterate css's of
670 * Should be called under cgroup_[tree_]mutex.
672 #define for_each_e_css(css, ssid, cgrp) \
673 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
674 if (!((css) = cgroup_e_css_by_mask(cgrp, \
675 cgroup_subsys[(ssid)]))) \
680 * do_each_subsys_mask - filter for_each_subsys with a bitmask
681 * @ss: the iteration cursor
682 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
683 * @ss_mask: the bitmask
685 * The block will only run for cases where the ssid-th bit (1 << ssid) of
688 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
689 unsigned long __ss_mask = (ss_mask); \
690 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
694 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
695 (ss) = cgroup_subsys[ssid]; \
698 #define while_each_subsys_mask() \
703 /* iterate over child cgrps, lock should be held throughout iteration */
704 #define cgroup_for_each_live_child(child, cgrp) \
705 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
706 if (({ lockdep_assert_held(&cgroup_mutex); \
707 cgroup_is_dead(child); })) \
711 /* walk live descendants in preorder */
712 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
713 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
714 if (({ lockdep_assert_held(&cgroup_mutex); \
715 (dsct) = (d_css)->cgroup; \
716 cgroup_is_dead(dsct); })) \
720 /* walk live descendants in postorder */
721 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
722 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
723 if (({ lockdep_assert_held(&cgroup_mutex); \
724 (dsct) = (d_css)->cgroup; \
725 cgroup_is_dead(dsct); })) \
730 * The default css_set - used by init and its children prior to any
731 * hierarchies being mounted. It contains a pointer to the root state
732 * for each subsystem. Also used to anchor the list of css_sets. Not
733 * reference-counted, to improve performance when child cgroups
734 * haven't been created.
736 struct css_set init_css_set = {
737 .refcount = REFCOUNT_INIT(1),
738 .dom_cset = &init_css_set,
739 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
740 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
741 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
742 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
743 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
744 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
745 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
748 * The following field is re-initialized when this cset gets linked
749 * in cgroup_init(). However, let's initialize the field
750 * statically too so that the default cgroup can be accessed safely
753 .dfl_cgrp = &cgrp_dfl_root.cgrp,
756 static int css_set_count = 1; /* 1 for init_css_set */
758 static bool css_set_threaded(struct css_set *cset)
760 return cset->dom_cset != cset;
764 * css_set_populated - does a css_set contain any tasks?
765 * @cset: target css_set
767 * css_set_populated() should be the same as !!cset->nr_tasks at steady
768 * state. However, css_set_populated() can be called while a task is being
769 * added to or removed from the linked list before the nr_tasks is
770 * properly updated. Hence, we can't just look at ->nr_tasks here.
772 static bool css_set_populated(struct css_set *cset)
774 lockdep_assert_held(&css_set_lock);
776 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
780 * cgroup_update_populated - update the populated count of a cgroup
781 * @cgrp: the target cgroup
782 * @populated: inc or dec populated count
784 * One of the css_sets associated with @cgrp is either getting its first
785 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
786 * count is propagated towards root so that a given cgroup's
787 * nr_populated_children is zero iff none of its descendants contain any
790 * @cgrp's interface file "cgroup.populated" is zero if both
791 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
792 * 1 otherwise. When the sum changes from or to zero, userland is notified
793 * that the content of the interface file has changed. This can be used to
794 * detect when @cgrp and its descendants become populated or empty.
796 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
798 struct cgroup *child = NULL;
799 int adj = populated ? 1 : -1;
801 lockdep_assert_held(&css_set_lock);
804 bool was_populated = cgroup_is_populated(cgrp);
807 cgrp->nr_populated_csets += adj;
809 if (cgroup_is_threaded(child))
810 cgrp->nr_populated_threaded_children += adj;
812 cgrp->nr_populated_domain_children += adj;
815 if (was_populated == cgroup_is_populated(cgrp))
818 cgroup1_check_for_release(cgrp);
819 TRACE_CGROUP_PATH(notify_populated, cgrp,
820 cgroup_is_populated(cgrp));
821 cgroup_file_notify(&cgrp->events_file);
824 cgrp = cgroup_parent(cgrp);
829 * css_set_update_populated - update populated state of a css_set
830 * @cset: target css_set
831 * @populated: whether @cset is populated or depopulated
833 * @cset is either getting the first task or losing the last. Update the
834 * populated counters of all associated cgroups accordingly.
836 static void css_set_update_populated(struct css_set *cset, bool populated)
838 struct cgrp_cset_link *link;
840 lockdep_assert_held(&css_set_lock);
842 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
843 cgroup_update_populated(link->cgrp, populated);
847 * css_set_move_task - move a task from one css_set to another
848 * @task: task being moved
849 * @from_cset: css_set @task currently belongs to (may be NULL)
850 * @to_cset: new css_set @task is being moved to (may be NULL)
851 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
853 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
854 * css_set, @from_cset can be NULL. If @task is being disassociated
855 * instead of moved, @to_cset can be NULL.
857 * This function automatically handles populated counter updates and
858 * css_task_iter adjustments but the caller is responsible for managing
859 * @from_cset and @to_cset's reference counts.
861 static void css_set_move_task(struct task_struct *task,
862 struct css_set *from_cset, struct css_set *to_cset,
865 lockdep_assert_held(&css_set_lock);
867 if (to_cset && !css_set_populated(to_cset))
868 css_set_update_populated(to_cset, true);
871 struct css_task_iter *it, *pos;
873 WARN_ON_ONCE(list_empty(&task->cg_list));
876 * @task is leaving, advance task iterators which are
877 * pointing to it so that they can resume at the next
878 * position. Advancing an iterator might remove it from
879 * the list, use safe walk. See css_task_iter_advance*()
882 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
884 if (it->task_pos == &task->cg_list)
885 css_task_iter_advance(it);
887 list_del_init(&task->cg_list);
888 if (!css_set_populated(from_cset))
889 css_set_update_populated(from_cset, false);
891 WARN_ON_ONCE(!list_empty(&task->cg_list));
896 * We are synchronized through cgroup_threadgroup_rwsem
897 * against PF_EXITING setting such that we can't race
898 * against cgroup_exit() changing the css_set to
899 * init_css_set and dropping the old one.
901 WARN_ON_ONCE(task->flags & PF_EXITING);
903 cgroup_move_task(task, to_cset);
904 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
910 * hash table for cgroup groups. This improves the performance to find
911 * an existing css_set. This hash doesn't (currently) take into
912 * account cgroups in empty hierarchies.
914 #define CSS_SET_HASH_BITS 7
915 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
917 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
919 unsigned long key = 0UL;
920 struct cgroup_subsys *ss;
923 for_each_subsys(ss, i)
924 key += (unsigned long)css[i];
925 key = (key >> 16) ^ key;
930 void put_css_set_locked(struct css_set *cset)
932 struct cgrp_cset_link *link, *tmp_link;
933 struct cgroup_subsys *ss;
936 lockdep_assert_held(&css_set_lock);
938 if (!refcount_dec_and_test(&cset->refcount))
941 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
943 /* This css_set is dead. unlink it and release cgroup and css refs */
944 for_each_subsys(ss, ssid) {
945 list_del(&cset->e_cset_node[ssid]);
946 css_put(cset->subsys[ssid]);
948 hash_del(&cset->hlist);
951 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
952 list_del(&link->cset_link);
953 list_del(&link->cgrp_link);
954 if (cgroup_parent(link->cgrp))
955 cgroup_put(link->cgrp);
959 if (css_set_threaded(cset)) {
960 list_del(&cset->threaded_csets_node);
961 put_css_set_locked(cset->dom_cset);
964 kfree_rcu(cset, rcu_head);
968 * compare_css_sets - helper function for find_existing_css_set().
969 * @cset: candidate css_set being tested
970 * @old_cset: existing css_set for a task
971 * @new_cgrp: cgroup that's being entered by the task
972 * @template: desired set of css pointers in css_set (pre-calculated)
974 * Returns true if "cset" matches "old_cset" except for the hierarchy
975 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
977 static bool compare_css_sets(struct css_set *cset,
978 struct css_set *old_cset,
979 struct cgroup *new_cgrp,
980 struct cgroup_subsys_state *template[])
982 struct cgroup *new_dfl_cgrp;
983 struct list_head *l1, *l2;
986 * On the default hierarchy, there can be csets which are
987 * associated with the same set of cgroups but different csses.
988 * Let's first ensure that csses match.
990 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
994 /* @cset's domain should match the default cgroup's */
995 if (cgroup_on_dfl(new_cgrp))
996 new_dfl_cgrp = new_cgrp;
998 new_dfl_cgrp = old_cset->dfl_cgrp;
1000 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1004 * Compare cgroup pointers in order to distinguish between
1005 * different cgroups in hierarchies. As different cgroups may
1006 * share the same effective css, this comparison is always
1009 l1 = &cset->cgrp_links;
1010 l2 = &old_cset->cgrp_links;
1012 struct cgrp_cset_link *link1, *link2;
1013 struct cgroup *cgrp1, *cgrp2;
1017 /* See if we reached the end - both lists are equal length. */
1018 if (l1 == &cset->cgrp_links) {
1019 BUG_ON(l2 != &old_cset->cgrp_links);
1022 BUG_ON(l2 == &old_cset->cgrp_links);
1024 /* Locate the cgroups associated with these links. */
1025 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1026 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1027 cgrp1 = link1->cgrp;
1028 cgrp2 = link2->cgrp;
1029 /* Hierarchies should be linked in the same order. */
1030 BUG_ON(cgrp1->root != cgrp2->root);
1033 * If this hierarchy is the hierarchy of the cgroup
1034 * that's changing, then we need to check that this
1035 * css_set points to the new cgroup; if it's any other
1036 * hierarchy, then this css_set should point to the
1037 * same cgroup as the old css_set.
1039 if (cgrp1->root == new_cgrp->root) {
1040 if (cgrp1 != new_cgrp)
1051 * find_existing_css_set - init css array and find the matching css_set
1052 * @old_cset: the css_set that we're using before the cgroup transition
1053 * @cgrp: the cgroup that we're moving into
1054 * @template: out param for the new set of csses, should be clear on entry
1056 static struct css_set *find_existing_css_set(struct css_set *old_cset,
1057 struct cgroup *cgrp,
1058 struct cgroup_subsys_state *template[])
1060 struct cgroup_root *root = cgrp->root;
1061 struct cgroup_subsys *ss;
1062 struct css_set *cset;
1067 * Build the set of subsystem state objects that we want to see in the
1068 * new css_set. while subsystems can change globally, the entries here
1069 * won't change, so no need for locking.
1071 for_each_subsys(ss, i) {
1072 if (root->subsys_mask & (1UL << i)) {
1074 * @ss is in this hierarchy, so we want the
1075 * effective css from @cgrp.
1077 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1080 * @ss is not in this hierarchy, so we don't want
1081 * to change the css.
1083 template[i] = old_cset->subsys[i];
1087 key = css_set_hash(template);
1088 hash_for_each_possible(css_set_table, cset, hlist, key) {
1089 if (!compare_css_sets(cset, old_cset, cgrp, template))
1092 /* This css_set matches what we need */
1096 /* No existing cgroup group matched */
1100 static void free_cgrp_cset_links(struct list_head *links_to_free)
1102 struct cgrp_cset_link *link, *tmp_link;
1104 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1105 list_del(&link->cset_link);
1111 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1112 * @count: the number of links to allocate
1113 * @tmp_links: list_head the allocated links are put on
1115 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1116 * through ->cset_link. Returns 0 on success or -errno.
1118 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1120 struct cgrp_cset_link *link;
1123 INIT_LIST_HEAD(tmp_links);
1125 for (i = 0; i < count; i++) {
1126 link = kzalloc(sizeof(*link), GFP_KERNEL);
1128 free_cgrp_cset_links(tmp_links);
1131 list_add(&link->cset_link, tmp_links);
1137 * link_css_set - a helper function to link a css_set to a cgroup
1138 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1139 * @cset: the css_set to be linked
1140 * @cgrp: the destination cgroup
1142 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1143 struct cgroup *cgrp)
1145 struct cgrp_cset_link *link;
1147 BUG_ON(list_empty(tmp_links));
1149 if (cgroup_on_dfl(cgrp))
1150 cset->dfl_cgrp = cgrp;
1152 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1157 * Always add links to the tail of the lists so that the lists are
1158 * in choronological order.
1160 list_move_tail(&link->cset_link, &cgrp->cset_links);
1161 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1163 if (cgroup_parent(cgrp))
1164 cgroup_get_live(cgrp);
1168 * find_css_set - return a new css_set with one cgroup updated
1169 * @old_cset: the baseline css_set
1170 * @cgrp: the cgroup to be updated
1172 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1173 * substituted into the appropriate hierarchy.
1175 static struct css_set *find_css_set(struct css_set *old_cset,
1176 struct cgroup *cgrp)
1178 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1179 struct css_set *cset;
1180 struct list_head tmp_links;
1181 struct cgrp_cset_link *link;
1182 struct cgroup_subsys *ss;
1186 lockdep_assert_held(&cgroup_mutex);
1188 /* First see if we already have a cgroup group that matches
1189 * the desired set */
1190 spin_lock_irq(&css_set_lock);
1191 cset = find_existing_css_set(old_cset, cgrp, template);
1194 spin_unlock_irq(&css_set_lock);
1199 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1203 /* Allocate all the cgrp_cset_link objects that we'll need */
1204 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1209 refcount_set(&cset->refcount, 1);
1210 cset->dom_cset = cset;
1211 INIT_LIST_HEAD(&cset->tasks);
1212 INIT_LIST_HEAD(&cset->mg_tasks);
1213 INIT_LIST_HEAD(&cset->task_iters);
1214 INIT_LIST_HEAD(&cset->threaded_csets);
1215 INIT_HLIST_NODE(&cset->hlist);
1216 INIT_LIST_HEAD(&cset->cgrp_links);
1217 INIT_LIST_HEAD(&cset->mg_preload_node);
1218 INIT_LIST_HEAD(&cset->mg_node);
1220 /* Copy the set of subsystem state objects generated in
1221 * find_existing_css_set() */
1222 memcpy(cset->subsys, template, sizeof(cset->subsys));
1224 spin_lock_irq(&css_set_lock);
1225 /* Add reference counts and links from the new css_set. */
1226 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1227 struct cgroup *c = link->cgrp;
1229 if (c->root == cgrp->root)
1231 link_css_set(&tmp_links, cset, c);
1234 BUG_ON(!list_empty(&tmp_links));
1238 /* Add @cset to the hash table */
1239 key = css_set_hash(cset->subsys);
1240 hash_add(css_set_table, &cset->hlist, key);
1242 for_each_subsys(ss, ssid) {
1243 struct cgroup_subsys_state *css = cset->subsys[ssid];
1245 list_add_tail(&cset->e_cset_node[ssid],
1246 &css->cgroup->e_csets[ssid]);
1250 spin_unlock_irq(&css_set_lock);
1253 * If @cset should be threaded, look up the matching dom_cset and
1254 * link them up. We first fully initialize @cset then look for the
1255 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1256 * to stay empty until we return.
1258 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1259 struct css_set *dcset;
1261 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1267 spin_lock_irq(&css_set_lock);
1268 cset->dom_cset = dcset;
1269 list_add_tail(&cset->threaded_csets_node,
1270 &dcset->threaded_csets);
1271 spin_unlock_irq(&css_set_lock);
1277 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1279 struct cgroup *root_cgrp = kf_root->kn->priv;
1281 return root_cgrp->root;
1284 static int cgroup_init_root_id(struct cgroup_root *root)
1288 lockdep_assert_held(&cgroup_mutex);
1290 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1294 root->hierarchy_id = id;
1298 static void cgroup_exit_root_id(struct cgroup_root *root)
1300 lockdep_assert_held(&cgroup_mutex);
1302 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1305 void cgroup_free_root(struct cgroup_root *root)
1308 idr_destroy(&root->cgroup_idr);
1313 static void cgroup_destroy_root(struct cgroup_root *root)
1315 struct cgroup *cgrp = &root->cgrp;
1316 struct cgrp_cset_link *link, *tmp_link;
1318 trace_cgroup_destroy_root(root);
1320 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1322 BUG_ON(atomic_read(&root->nr_cgrps));
1323 BUG_ON(!list_empty(&cgrp->self.children));
1325 /* Rebind all subsystems back to the default hierarchy */
1326 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1329 * Release all the links from cset_links to this hierarchy's
1332 spin_lock_irq(&css_set_lock);
1334 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1335 list_del(&link->cset_link);
1336 list_del(&link->cgrp_link);
1340 spin_unlock_irq(&css_set_lock);
1342 if (!list_empty(&root->root_list)) {
1343 list_del(&root->root_list);
1344 cgroup_root_count--;
1347 cgroup_exit_root_id(root);
1349 mutex_unlock(&cgroup_mutex);
1351 kernfs_destroy_root(root->kf_root);
1352 cgroup_free_root(root);
1356 * look up cgroup associated with current task's cgroup namespace on the
1357 * specified hierarchy
1359 static struct cgroup *
1360 current_cgns_cgroup_from_root(struct cgroup_root *root)
1362 struct cgroup *res = NULL;
1363 struct css_set *cset;
1365 lockdep_assert_held(&css_set_lock);
1369 cset = current->nsproxy->cgroup_ns->root_cset;
1370 if (cset == &init_css_set) {
1373 struct cgrp_cset_link *link;
1375 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1376 struct cgroup *c = link->cgrp;
1378 if (c->root == root) {
1390 /* look up cgroup associated with given css_set on the specified hierarchy */
1391 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1392 struct cgroup_root *root)
1394 struct cgroup *res = NULL;
1396 lockdep_assert_held(&cgroup_mutex);
1397 lockdep_assert_held(&css_set_lock);
1399 if (cset == &init_css_set) {
1401 } else if (root == &cgrp_dfl_root) {
1402 res = cset->dfl_cgrp;
1404 struct cgrp_cset_link *link;
1406 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1407 struct cgroup *c = link->cgrp;
1409 if (c->root == root) {
1421 * Return the cgroup for "task" from the given hierarchy. Must be
1422 * called with cgroup_mutex and css_set_lock held.
1424 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1425 struct cgroup_root *root)
1428 * No need to lock the task - since we hold cgroup_mutex the
1429 * task can't change groups, so the only thing that can happen
1430 * is that it exits and its css is set back to init_css_set.
1432 return cset_cgroup_from_root(task_css_set(task), root);
1436 * A task must hold cgroup_mutex to modify cgroups.
1438 * Any task can increment and decrement the count field without lock.
1439 * So in general, code holding cgroup_mutex can't rely on the count
1440 * field not changing. However, if the count goes to zero, then only
1441 * cgroup_attach_task() can increment it again. Because a count of zero
1442 * means that no tasks are currently attached, therefore there is no
1443 * way a task attached to that cgroup can fork (the other way to
1444 * increment the count). So code holding cgroup_mutex can safely
1445 * assume that if the count is zero, it will stay zero. Similarly, if
1446 * a task holds cgroup_mutex on a cgroup with zero count, it
1447 * knows that the cgroup won't be removed, as cgroup_rmdir()
1450 * A cgroup can only be deleted if both its 'count' of using tasks
1451 * is zero, and its list of 'children' cgroups is empty. Since all
1452 * tasks in the system use _some_ cgroup, and since there is always at
1453 * least one task in the system (init, pid == 1), therefore, root cgroup
1454 * always has either children cgroups and/or using tasks. So we don't
1455 * need a special hack to ensure that root cgroup cannot be deleted.
1457 * P.S. One more locking exception. RCU is used to guard the
1458 * update of a tasks cgroup pointer by cgroup_attach_task()
1461 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1463 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1466 struct cgroup_subsys *ss = cft->ss;
1468 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1469 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1470 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1472 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1473 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1476 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1482 * cgroup_file_mode - deduce file mode of a control file
1483 * @cft: the control file in question
1485 * S_IRUGO for read, S_IWUSR for write.
1487 static umode_t cgroup_file_mode(const struct cftype *cft)
1491 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1494 if (cft->write_u64 || cft->write_s64 || cft->write) {
1495 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1505 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1506 * @subtree_control: the new subtree_control mask to consider
1507 * @this_ss_mask: available subsystems
1509 * On the default hierarchy, a subsystem may request other subsystems to be
1510 * enabled together through its ->depends_on mask. In such cases, more
1511 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1513 * This function calculates which subsystems need to be enabled if
1514 * @subtree_control is to be applied while restricted to @this_ss_mask.
1516 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1518 u16 cur_ss_mask = subtree_control;
1519 struct cgroup_subsys *ss;
1522 lockdep_assert_held(&cgroup_mutex);
1524 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1527 u16 new_ss_mask = cur_ss_mask;
1529 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1530 new_ss_mask |= ss->depends_on;
1531 } while_each_subsys_mask();
1534 * Mask out subsystems which aren't available. This can
1535 * happen only if some depended-upon subsystems were bound
1536 * to non-default hierarchies.
1538 new_ss_mask &= this_ss_mask;
1540 if (new_ss_mask == cur_ss_mask)
1542 cur_ss_mask = new_ss_mask;
1549 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1550 * @kn: the kernfs_node being serviced
1552 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1553 * the method finishes if locking succeeded. Note that once this function
1554 * returns the cgroup returned by cgroup_kn_lock_live() may become
1555 * inaccessible any time. If the caller intends to continue to access the
1556 * cgroup, it should pin it before invoking this function.
1558 void cgroup_kn_unlock(struct kernfs_node *kn)
1560 struct cgroup *cgrp;
1562 if (kernfs_type(kn) == KERNFS_DIR)
1565 cgrp = kn->parent->priv;
1567 mutex_unlock(&cgroup_mutex);
1569 kernfs_unbreak_active_protection(kn);
1574 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1575 * @kn: the kernfs_node being serviced
1576 * @drain_offline: perform offline draining on the cgroup
1578 * This helper is to be used by a cgroup kernfs method currently servicing
1579 * @kn. It breaks the active protection, performs cgroup locking and
1580 * verifies that the associated cgroup is alive. Returns the cgroup if
1581 * alive; otherwise, %NULL. A successful return should be undone by a
1582 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1583 * cgroup is drained of offlining csses before return.
1585 * Any cgroup kernfs method implementation which requires locking the
1586 * associated cgroup should use this helper. It avoids nesting cgroup
1587 * locking under kernfs active protection and allows all kernfs operations
1588 * including self-removal.
1590 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1592 struct cgroup *cgrp;
1594 if (kernfs_type(kn) == KERNFS_DIR)
1597 cgrp = kn->parent->priv;
1600 * We're gonna grab cgroup_mutex which nests outside kernfs
1601 * active_ref. cgroup liveliness check alone provides enough
1602 * protection against removal. Ensure @cgrp stays accessible and
1603 * break the active_ref protection.
1605 if (!cgroup_tryget(cgrp))
1607 kernfs_break_active_protection(kn);
1610 cgroup_lock_and_drain_offline(cgrp);
1612 mutex_lock(&cgroup_mutex);
1614 if (!cgroup_is_dead(cgrp))
1617 cgroup_kn_unlock(kn);
1621 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1623 char name[CGROUP_FILE_NAME_MAX];
1625 lockdep_assert_held(&cgroup_mutex);
1627 if (cft->file_offset) {
1628 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1629 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1631 spin_lock_irq(&cgroup_file_kn_lock);
1633 spin_unlock_irq(&cgroup_file_kn_lock);
1635 del_timer_sync(&cfile->notify_timer);
1638 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1642 * css_clear_dir - remove subsys files in a cgroup directory
1645 static void css_clear_dir(struct cgroup_subsys_state *css)
1647 struct cgroup *cgrp = css->cgroup;
1648 struct cftype *cfts;
1650 if (!(css->flags & CSS_VISIBLE))
1653 css->flags &= ~CSS_VISIBLE;
1656 if (cgroup_on_dfl(cgrp))
1657 cfts = cgroup_base_files;
1659 cfts = cgroup1_base_files;
1661 cgroup_addrm_files(css, cgrp, cfts, false);
1663 list_for_each_entry(cfts, &css->ss->cfts, node)
1664 cgroup_addrm_files(css, cgrp, cfts, false);
1669 * css_populate_dir - create subsys files in a cgroup directory
1672 * On failure, no file is added.
1674 static int css_populate_dir(struct cgroup_subsys_state *css)
1676 struct cgroup *cgrp = css->cgroup;
1677 struct cftype *cfts, *failed_cfts;
1680 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1684 if (cgroup_on_dfl(cgrp))
1685 cfts = cgroup_base_files;
1687 cfts = cgroup1_base_files;
1689 ret = cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1693 list_for_each_entry(cfts, &css->ss->cfts, node) {
1694 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1702 css->flags |= CSS_VISIBLE;
1706 list_for_each_entry(cfts, &css->ss->cfts, node) {
1707 if (cfts == failed_cfts)
1709 cgroup_addrm_files(css, cgrp, cfts, false);
1714 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1716 struct cgroup *dcgrp = &dst_root->cgrp;
1717 struct cgroup_subsys *ss;
1720 lockdep_assert_held(&cgroup_mutex);
1722 do_each_subsys_mask(ss, ssid, ss_mask) {
1724 * If @ss has non-root csses attached to it, can't move.
1725 * If @ss is an implicit controller, it is exempt from this
1726 * rule and can be stolen.
1728 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1729 !ss->implicit_on_dfl)
1732 /* can't move between two non-dummy roots either */
1733 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1735 } while_each_subsys_mask();
1737 do_each_subsys_mask(ss, ssid, ss_mask) {
1738 struct cgroup_root *src_root = ss->root;
1739 struct cgroup *scgrp = &src_root->cgrp;
1740 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1741 struct css_set *cset;
1743 WARN_ON(!css || cgroup_css(dcgrp, ss));
1745 /* disable from the source */
1746 src_root->subsys_mask &= ~(1 << ssid);
1747 WARN_ON(cgroup_apply_control(scgrp));
1748 cgroup_finalize_control(scgrp, 0);
1751 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1752 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1753 ss->root = dst_root;
1754 css->cgroup = dcgrp;
1756 spin_lock_irq(&css_set_lock);
1757 hash_for_each(css_set_table, i, cset, hlist)
1758 list_move_tail(&cset->e_cset_node[ss->id],
1759 &dcgrp->e_csets[ss->id]);
1760 spin_unlock_irq(&css_set_lock);
1762 /* default hierarchy doesn't enable controllers by default */
1763 dst_root->subsys_mask |= 1 << ssid;
1764 if (dst_root == &cgrp_dfl_root) {
1765 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1767 dcgrp->subtree_control |= 1 << ssid;
1768 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1771 ret = cgroup_apply_control(dcgrp);
1773 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1778 } while_each_subsys_mask();
1780 kernfs_activate(dcgrp->kn);
1784 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1785 struct kernfs_root *kf_root)
1789 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1790 struct cgroup *ns_cgroup;
1792 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1796 spin_lock_irq(&css_set_lock);
1797 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1798 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1799 spin_unlock_irq(&css_set_lock);
1801 if (len >= PATH_MAX)
1804 seq_escape(sf, buf, " \t\n\\");
1811 enum cgroup2_param {
1816 static const struct fs_parameter_spec cgroup2_param_specs[] = {
1817 fsparam_flag ("nsdelegate", Opt_nsdelegate),
1821 static const struct fs_parameter_description cgroup2_fs_parameters = {
1823 .specs = cgroup2_param_specs,
1826 static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1828 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1829 struct fs_parse_result result;
1832 opt = fs_parse(fc, &cgroup2_fs_parameters, param, &result);
1837 case Opt_nsdelegate:
1838 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1844 static void apply_cgroup_root_flags(unsigned int root_flags)
1846 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1847 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1848 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1850 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1854 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1856 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1857 seq_puts(seq, ",nsdelegate");
1861 static int cgroup_reconfigure(struct fs_context *fc)
1863 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1865 apply_cgroup_root_flags(ctx->flags);
1870 * To reduce the fork() overhead for systems that are not actually using
1871 * their cgroups capability, we don't maintain the lists running through
1872 * each css_set to its tasks until we see the list actually used - in other
1873 * words after the first mount.
1875 static bool use_task_css_set_links __read_mostly;
1877 static void cgroup_enable_task_cg_lists(void)
1879 struct task_struct *p, *g;
1882 * We need tasklist_lock because RCU is not safe against
1883 * while_each_thread(). Besides, a forking task that has passed
1884 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1885 * is not guaranteed to have its child immediately visible in the
1886 * tasklist if we walk through it with RCU.
1888 read_lock(&tasklist_lock);
1889 spin_lock_irq(&css_set_lock);
1891 if (use_task_css_set_links)
1894 use_task_css_set_links = true;
1896 do_each_thread(g, p) {
1897 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1898 task_css_set(p) != &init_css_set);
1901 * We should check if the process is exiting, otherwise
1902 * it will race with cgroup_exit() in that the list
1903 * entry won't be deleted though the process has exited.
1904 * Do it while holding siglock so that we don't end up
1905 * racing against cgroup_exit().
1907 * Interrupts were already disabled while acquiring
1908 * the css_set_lock, so we do not need to disable it
1909 * again when acquiring the sighand->siglock here.
1911 spin_lock(&p->sighand->siglock);
1912 if (!(p->flags & PF_EXITING)) {
1913 struct css_set *cset = task_css_set(p);
1915 if (!css_set_populated(cset))
1916 css_set_update_populated(cset, true);
1917 list_add_tail(&p->cg_list, &cset->tasks);
1921 spin_unlock(&p->sighand->siglock);
1922 } while_each_thread(g, p);
1924 spin_unlock_irq(&css_set_lock);
1925 read_unlock(&tasklist_lock);
1928 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1930 struct cgroup_subsys *ss;
1933 INIT_LIST_HEAD(&cgrp->self.sibling);
1934 INIT_LIST_HEAD(&cgrp->self.children);
1935 INIT_LIST_HEAD(&cgrp->cset_links);
1936 INIT_LIST_HEAD(&cgrp->pidlists);
1937 mutex_init(&cgrp->pidlist_mutex);
1938 cgrp->self.cgroup = cgrp;
1939 cgrp->self.flags |= CSS_ONLINE;
1940 cgrp->dom_cgrp = cgrp;
1941 cgrp->max_descendants = INT_MAX;
1942 cgrp->max_depth = INT_MAX;
1943 INIT_LIST_HEAD(&cgrp->rstat_css_list);
1944 prev_cputime_init(&cgrp->prev_cputime);
1946 for_each_subsys(ss, ssid)
1947 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1949 init_waitqueue_head(&cgrp->offline_waitq);
1950 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1953 void init_cgroup_root(struct cgroup_fs_context *ctx)
1955 struct cgroup_root *root = ctx->root;
1956 struct cgroup *cgrp = &root->cgrp;
1958 INIT_LIST_HEAD(&root->root_list);
1959 atomic_set(&root->nr_cgrps, 1);
1961 init_cgroup_housekeeping(cgrp);
1962 idr_init(&root->cgroup_idr);
1964 root->flags = ctx->flags;
1965 if (ctx->release_agent)
1966 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
1968 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
1969 if (ctx->cpuset_clone_children)
1970 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1973 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1975 LIST_HEAD(tmp_links);
1976 struct cgroup *root_cgrp = &root->cgrp;
1977 struct kernfs_syscall_ops *kf_sops;
1978 struct css_set *cset;
1981 lockdep_assert_held(&cgroup_mutex);
1983 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1986 root_cgrp->id = ret;
1987 root_cgrp->ancestor_ids[0] = ret;
1989 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1995 * We're accessing css_set_count without locking css_set_lock here,
1996 * but that's OK - it can only be increased by someone holding
1997 * cgroup_lock, and that's us. Later rebinding may disable
1998 * controllers on the default hierarchy and thus create new csets,
1999 * which can't be more than the existing ones. Allocate 2x.
2001 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2005 ret = cgroup_init_root_id(root);
2009 kf_sops = root == &cgrp_dfl_root ?
2010 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2012 root->kf_root = kernfs_create_root(kf_sops,
2013 KERNFS_ROOT_CREATE_DEACTIVATED |
2014 KERNFS_ROOT_SUPPORT_EXPORTOP,
2016 if (IS_ERR(root->kf_root)) {
2017 ret = PTR_ERR(root->kf_root);
2020 root_cgrp->kn = root->kf_root->kn;
2022 ret = css_populate_dir(&root_cgrp->self);
2026 ret = rebind_subsystems(root, ss_mask);
2030 ret = cgroup_bpf_inherit(root_cgrp);
2033 trace_cgroup_setup_root(root);
2036 * There must be no failure case after here, since rebinding takes
2037 * care of subsystems' refcounts, which are explicitly dropped in
2038 * the failure exit path.
2040 list_add(&root->root_list, &cgroup_roots);
2041 cgroup_root_count++;
2044 * Link the root cgroup in this hierarchy into all the css_set
2047 spin_lock_irq(&css_set_lock);
2048 hash_for_each(css_set_table, i, cset, hlist) {
2049 link_css_set(&tmp_links, cset, root_cgrp);
2050 if (css_set_populated(cset))
2051 cgroup_update_populated(root_cgrp, true);
2053 spin_unlock_irq(&css_set_lock);
2055 BUG_ON(!list_empty(&root_cgrp->self.children));
2056 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2058 kernfs_activate(root_cgrp->kn);
2063 kernfs_destroy_root(root->kf_root);
2064 root->kf_root = NULL;
2066 cgroup_exit_root_id(root);
2068 percpu_ref_exit(&root_cgrp->self.refcnt);
2070 free_cgrp_cset_links(&tmp_links);
2074 int cgroup_do_get_tree(struct fs_context *fc)
2076 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2079 ctx->kfc.root = ctx->root->kf_root;
2080 if (fc->fs_type == &cgroup2_fs_type)
2081 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2083 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2084 ret = kernfs_get_tree(fc);
2087 * In non-init cgroup namespace, instead of root cgroup's dentry,
2088 * we return the dentry corresponding to the cgroupns->root_cgrp.
2090 if (!ret && ctx->ns != &init_cgroup_ns) {
2091 struct dentry *nsdentry;
2092 struct super_block *sb = fc->root->d_sb;
2093 struct cgroup *cgrp;
2095 mutex_lock(&cgroup_mutex);
2096 spin_lock_irq(&css_set_lock);
2098 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2100 spin_unlock_irq(&css_set_lock);
2101 mutex_unlock(&cgroup_mutex);
2103 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2105 fc->root = nsdentry;
2106 if (IS_ERR(nsdentry)) {
2107 ret = PTR_ERR(nsdentry);
2108 deactivate_locked_super(sb);
2112 if (!ctx->kfc.new_sb_created)
2113 cgroup_put(&ctx->root->cgrp);
2119 * Destroy a cgroup filesystem context.
2121 static void cgroup_fs_context_free(struct fs_context *fc)
2123 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2126 kfree(ctx->release_agent);
2127 put_cgroup_ns(ctx->ns);
2128 kernfs_free_fs_context(fc);
2132 static int cgroup_get_tree(struct fs_context *fc)
2134 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2137 cgrp_dfl_visible = true;
2138 cgroup_get_live(&cgrp_dfl_root.cgrp);
2139 ctx->root = &cgrp_dfl_root;
2141 ret = cgroup_do_get_tree(fc);
2143 apply_cgroup_root_flags(ctx->flags);
2147 static const struct fs_context_operations cgroup_fs_context_ops = {
2148 .free = cgroup_fs_context_free,
2149 .parse_param = cgroup2_parse_param,
2150 .get_tree = cgroup_get_tree,
2151 .reconfigure = cgroup_reconfigure,
2154 static const struct fs_context_operations cgroup1_fs_context_ops = {
2155 .free = cgroup_fs_context_free,
2156 .parse_param = cgroup1_parse_param,
2157 .get_tree = cgroup1_get_tree,
2158 .reconfigure = cgroup1_reconfigure,
2162 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2163 * we select the namespace we're going to use.
2165 static int cgroup_init_fs_context(struct fs_context *fc)
2167 struct cgroup_fs_context *ctx;
2169 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2174 * The first time anyone tries to mount a cgroup, enable the list
2175 * linking each css_set to its tasks and fix up all existing tasks.
2177 if (!use_task_css_set_links)
2178 cgroup_enable_task_cg_lists();
2180 ctx->ns = current->nsproxy->cgroup_ns;
2181 get_cgroup_ns(ctx->ns);
2182 fc->fs_private = &ctx->kfc;
2183 if (fc->fs_type == &cgroup2_fs_type)
2184 fc->ops = &cgroup_fs_context_ops;
2186 fc->ops = &cgroup1_fs_context_ops;
2188 put_user_ns(fc->user_ns);
2189 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2194 static void cgroup_kill_sb(struct super_block *sb)
2196 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2197 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2200 * If @root doesn't have any children, start killing it.
2201 * This prevents new mounts by disabling percpu_ref_tryget_live().
2202 * cgroup_mount() may wait for @root's release.
2204 * And don't kill the default root.
2206 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2207 !percpu_ref_is_dying(&root->cgrp.self.refcnt))
2208 percpu_ref_kill(&root->cgrp.self.refcnt);
2209 cgroup_put(&root->cgrp);
2213 struct file_system_type cgroup_fs_type = {
2215 .init_fs_context = cgroup_init_fs_context,
2216 .parameters = &cgroup1_fs_parameters,
2217 .kill_sb = cgroup_kill_sb,
2218 .fs_flags = FS_USERNS_MOUNT,
2221 static struct file_system_type cgroup2_fs_type = {
2223 .init_fs_context = cgroup_init_fs_context,
2224 .parameters = &cgroup2_fs_parameters,
2225 .kill_sb = cgroup_kill_sb,
2226 .fs_flags = FS_USERNS_MOUNT,
2229 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2230 struct cgroup_namespace *ns)
2232 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2234 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2237 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2238 struct cgroup_namespace *ns)
2242 mutex_lock(&cgroup_mutex);
2243 spin_lock_irq(&css_set_lock);
2245 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2247 spin_unlock_irq(&css_set_lock);
2248 mutex_unlock(&cgroup_mutex);
2252 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2255 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2256 * @task: target task
2257 * @buf: the buffer to write the path into
2258 * @buflen: the length of the buffer
2260 * Determine @task's cgroup on the first (the one with the lowest non-zero
2261 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2262 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2263 * cgroup controller callbacks.
2265 * Return value is the same as kernfs_path().
2267 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2269 struct cgroup_root *root;
2270 struct cgroup *cgrp;
2271 int hierarchy_id = 1;
2274 mutex_lock(&cgroup_mutex);
2275 spin_lock_irq(&css_set_lock);
2277 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2280 cgrp = task_cgroup_from_root(task, root);
2281 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2283 /* if no hierarchy exists, everyone is in "/" */
2284 ret = strlcpy(buf, "/", buflen);
2287 spin_unlock_irq(&css_set_lock);
2288 mutex_unlock(&cgroup_mutex);
2291 EXPORT_SYMBOL_GPL(task_cgroup_path);
2294 * cgroup_migrate_add_task - add a migration target task to a migration context
2295 * @task: target task
2296 * @mgctx: target migration context
2298 * Add @task, which is a migration target, to @mgctx->tset. This function
2299 * becomes noop if @task doesn't need to be migrated. @task's css_set
2300 * should have been added as a migration source and @task->cg_list will be
2301 * moved from the css_set's tasks list to mg_tasks one.
2303 static void cgroup_migrate_add_task(struct task_struct *task,
2304 struct cgroup_mgctx *mgctx)
2306 struct css_set *cset;
2308 lockdep_assert_held(&css_set_lock);
2310 /* @task either already exited or can't exit until the end */
2311 if (task->flags & PF_EXITING)
2314 /* leave @task alone if post_fork() hasn't linked it yet */
2315 if (list_empty(&task->cg_list))
2318 cset = task_css_set(task);
2319 if (!cset->mg_src_cgrp)
2322 mgctx->tset.nr_tasks++;
2324 list_move_tail(&task->cg_list, &cset->mg_tasks);
2325 if (list_empty(&cset->mg_node))
2326 list_add_tail(&cset->mg_node,
2327 &mgctx->tset.src_csets);
2328 if (list_empty(&cset->mg_dst_cset->mg_node))
2329 list_add_tail(&cset->mg_dst_cset->mg_node,
2330 &mgctx->tset.dst_csets);
2334 * cgroup_taskset_first - reset taskset and return the first task
2335 * @tset: taskset of interest
2336 * @dst_cssp: output variable for the destination css
2338 * @tset iteration is initialized and the first task is returned.
2340 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2341 struct cgroup_subsys_state **dst_cssp)
2343 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2344 tset->cur_task = NULL;
2346 return cgroup_taskset_next(tset, dst_cssp);
2350 * cgroup_taskset_next - iterate to the next task in taskset
2351 * @tset: taskset of interest
2352 * @dst_cssp: output variable for the destination css
2354 * Return the next task in @tset. Iteration must have been initialized
2355 * with cgroup_taskset_first().
2357 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2358 struct cgroup_subsys_state **dst_cssp)
2360 struct css_set *cset = tset->cur_cset;
2361 struct task_struct *task = tset->cur_task;
2363 while (&cset->mg_node != tset->csets) {
2365 task = list_first_entry(&cset->mg_tasks,
2366 struct task_struct, cg_list);
2368 task = list_next_entry(task, cg_list);
2370 if (&task->cg_list != &cset->mg_tasks) {
2371 tset->cur_cset = cset;
2372 tset->cur_task = task;
2375 * This function may be called both before and
2376 * after cgroup_taskset_migrate(). The two cases
2377 * can be distinguished by looking at whether @cset
2378 * has its ->mg_dst_cset set.
2380 if (cset->mg_dst_cset)
2381 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2383 *dst_cssp = cset->subsys[tset->ssid];
2388 cset = list_next_entry(cset, mg_node);
2396 * cgroup_taskset_migrate - migrate a taskset
2397 * @mgctx: migration context
2399 * Migrate tasks in @mgctx as setup by migration preparation functions.
2400 * This function fails iff one of the ->can_attach callbacks fails and
2401 * guarantees that either all or none of the tasks in @mgctx are migrated.
2402 * @mgctx is consumed regardless of success.
2404 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2406 struct cgroup_taskset *tset = &mgctx->tset;
2407 struct cgroup_subsys *ss;
2408 struct task_struct *task, *tmp_task;
2409 struct css_set *cset, *tmp_cset;
2410 int ssid, failed_ssid, ret;
2412 /* check that we can legitimately attach to the cgroup */
2413 if (tset->nr_tasks) {
2414 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2415 if (ss->can_attach) {
2417 ret = ss->can_attach(tset);
2420 goto out_cancel_attach;
2423 } while_each_subsys_mask();
2427 * Now that we're guaranteed success, proceed to move all tasks to
2428 * the new cgroup. There are no failure cases after here, so this
2429 * is the commit point.
2431 spin_lock_irq(&css_set_lock);
2432 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2433 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2434 struct css_set *from_cset = task_css_set(task);
2435 struct css_set *to_cset = cset->mg_dst_cset;
2437 get_css_set(to_cset);
2438 to_cset->nr_tasks++;
2439 css_set_move_task(task, from_cset, to_cset, true);
2440 from_cset->nr_tasks--;
2442 * If the source or destination cgroup is frozen,
2443 * the task might require to change its state.
2445 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2447 put_css_set_locked(from_cset);
2451 spin_unlock_irq(&css_set_lock);
2454 * Migration is committed, all target tasks are now on dst_csets.
2455 * Nothing is sensitive to fork() after this point. Notify
2456 * controllers that migration is complete.
2458 tset->csets = &tset->dst_csets;
2460 if (tset->nr_tasks) {
2461 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2466 } while_each_subsys_mask();
2470 goto out_release_tset;
2473 if (tset->nr_tasks) {
2474 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2475 if (ssid == failed_ssid)
2477 if (ss->cancel_attach) {
2479 ss->cancel_attach(tset);
2481 } while_each_subsys_mask();
2484 spin_lock_irq(&css_set_lock);
2485 list_splice_init(&tset->dst_csets, &tset->src_csets);
2486 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2487 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2488 list_del_init(&cset->mg_node);
2490 spin_unlock_irq(&css_set_lock);
2493 * Re-initialize the cgroup_taskset structure in case it is reused
2494 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2498 tset->csets = &tset->src_csets;
2503 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2504 * @dst_cgrp: destination cgroup to test
2506 * On the default hierarchy, except for the mixable, (possible) thread root
2507 * and threaded cgroups, subtree_control must be zero for migration
2508 * destination cgroups with tasks so that child cgroups don't compete
2511 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2513 /* v1 doesn't have any restriction */
2514 if (!cgroup_on_dfl(dst_cgrp))
2517 /* verify @dst_cgrp can host resources */
2518 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2521 /* mixables don't care */
2522 if (cgroup_is_mixable(dst_cgrp))
2526 * If @dst_cgrp is already or can become a thread root or is
2527 * threaded, it doesn't matter.
2529 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2532 /* apply no-internal-process constraint */
2533 if (dst_cgrp->subtree_control)
2540 * cgroup_migrate_finish - cleanup after attach
2541 * @mgctx: migration context
2543 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2544 * those functions for details.
2546 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2548 LIST_HEAD(preloaded);
2549 struct css_set *cset, *tmp_cset;
2551 lockdep_assert_held(&cgroup_mutex);
2553 spin_lock_irq(&css_set_lock);
2555 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2556 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2558 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2559 cset->mg_src_cgrp = NULL;
2560 cset->mg_dst_cgrp = NULL;
2561 cset->mg_dst_cset = NULL;
2562 list_del_init(&cset->mg_preload_node);
2563 put_css_set_locked(cset);
2566 spin_unlock_irq(&css_set_lock);
2570 * cgroup_migrate_add_src - add a migration source css_set
2571 * @src_cset: the source css_set to add
2572 * @dst_cgrp: the destination cgroup
2573 * @mgctx: migration context
2575 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2576 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2577 * up by cgroup_migrate_finish().
2579 * This function may be called without holding cgroup_threadgroup_rwsem
2580 * even if the target is a process. Threads may be created and destroyed
2581 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2582 * into play and the preloaded css_sets are guaranteed to cover all
2585 void cgroup_migrate_add_src(struct css_set *src_cset,
2586 struct cgroup *dst_cgrp,
2587 struct cgroup_mgctx *mgctx)
2589 struct cgroup *src_cgrp;
2591 lockdep_assert_held(&cgroup_mutex);
2592 lockdep_assert_held(&css_set_lock);
2595 * If ->dead, @src_set is associated with one or more dead cgroups
2596 * and doesn't contain any migratable tasks. Ignore it early so
2597 * that the rest of migration path doesn't get confused by it.
2602 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2604 if (!list_empty(&src_cset->mg_preload_node))
2607 WARN_ON(src_cset->mg_src_cgrp);
2608 WARN_ON(src_cset->mg_dst_cgrp);
2609 WARN_ON(!list_empty(&src_cset->mg_tasks));
2610 WARN_ON(!list_empty(&src_cset->mg_node));
2612 src_cset->mg_src_cgrp = src_cgrp;
2613 src_cset->mg_dst_cgrp = dst_cgrp;
2614 get_css_set(src_cset);
2615 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2619 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2620 * @mgctx: migration context
2622 * Tasks are about to be moved and all the source css_sets have been
2623 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2624 * pins all destination css_sets, links each to its source, and append them
2625 * to @mgctx->preloaded_dst_csets.
2627 * This function must be called after cgroup_migrate_add_src() has been
2628 * called on each migration source css_set. After migration is performed
2629 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2632 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2634 struct css_set *src_cset, *tmp_cset;
2636 lockdep_assert_held(&cgroup_mutex);
2638 /* look up the dst cset for each src cset and link it to src */
2639 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2641 struct css_set *dst_cset;
2642 struct cgroup_subsys *ss;
2645 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2649 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2652 * If src cset equals dst, it's noop. Drop the src.
2653 * cgroup_migrate() will skip the cset too. Note that we
2654 * can't handle src == dst as some nodes are used by both.
2656 if (src_cset == dst_cset) {
2657 src_cset->mg_src_cgrp = NULL;
2658 src_cset->mg_dst_cgrp = NULL;
2659 list_del_init(&src_cset->mg_preload_node);
2660 put_css_set(src_cset);
2661 put_css_set(dst_cset);
2665 src_cset->mg_dst_cset = dst_cset;
2667 if (list_empty(&dst_cset->mg_preload_node))
2668 list_add_tail(&dst_cset->mg_preload_node,
2669 &mgctx->preloaded_dst_csets);
2671 put_css_set(dst_cset);
2673 for_each_subsys(ss, ssid)
2674 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2675 mgctx->ss_mask |= 1 << ssid;
2682 * cgroup_migrate - migrate a process or task to a cgroup
2683 * @leader: the leader of the process or the task to migrate
2684 * @threadgroup: whether @leader points to the whole process or a single task
2685 * @mgctx: migration context
2687 * Migrate a process or task denoted by @leader. If migrating a process,
2688 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2689 * responsible for invoking cgroup_migrate_add_src() and
2690 * cgroup_migrate_prepare_dst() on the targets before invoking this
2691 * function and following up with cgroup_migrate_finish().
2693 * As long as a controller's ->can_attach() doesn't fail, this function is
2694 * guaranteed to succeed. This means that, excluding ->can_attach()
2695 * failure, when migrating multiple targets, the success or failure can be
2696 * decided for all targets by invoking group_migrate_prepare_dst() before
2697 * actually starting migrating.
2699 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2700 struct cgroup_mgctx *mgctx)
2702 struct task_struct *task;
2705 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2706 * already PF_EXITING could be freed from underneath us unless we
2707 * take an rcu_read_lock.
2709 spin_lock_irq(&css_set_lock);
2713 cgroup_migrate_add_task(task, mgctx);
2716 } while_each_thread(leader, task);
2718 spin_unlock_irq(&css_set_lock);
2720 return cgroup_migrate_execute(mgctx);
2724 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2725 * @dst_cgrp: the cgroup to attach to
2726 * @leader: the task or the leader of the threadgroup to be attached
2727 * @threadgroup: attach the whole threadgroup?
2729 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2731 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2734 DEFINE_CGROUP_MGCTX(mgctx);
2735 struct task_struct *task;
2738 ret = cgroup_migrate_vet_dst(dst_cgrp);
2742 /* look up all src csets */
2743 spin_lock_irq(&css_set_lock);
2747 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2750 } while_each_thread(leader, task);
2752 spin_unlock_irq(&css_set_lock);
2754 /* prepare dst csets and commit */
2755 ret = cgroup_migrate_prepare_dst(&mgctx);
2757 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2759 cgroup_migrate_finish(&mgctx);
2762 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2767 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
2768 __acquires(&cgroup_threadgroup_rwsem)
2770 struct task_struct *tsk;
2773 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2774 return ERR_PTR(-EINVAL);
2776 percpu_down_write(&cgroup_threadgroup_rwsem);
2780 tsk = find_task_by_vpid(pid);
2782 tsk = ERR_PTR(-ESRCH);
2783 goto out_unlock_threadgroup;
2790 tsk = tsk->group_leader;
2793 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2794 * If userland migrates such a kthread to a non-root cgroup, it can
2795 * become trapped in a cpuset, or RT kthread may be born in a
2796 * cgroup with no rt_runtime allocated. Just say no.
2798 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2799 tsk = ERR_PTR(-EINVAL);
2800 goto out_unlock_threadgroup;
2803 get_task_struct(tsk);
2804 goto out_unlock_rcu;
2806 out_unlock_threadgroup:
2807 percpu_up_write(&cgroup_threadgroup_rwsem);
2813 void cgroup_procs_write_finish(struct task_struct *task)
2814 __releases(&cgroup_threadgroup_rwsem)
2816 struct cgroup_subsys *ss;
2819 /* release reference from cgroup_procs_write_start() */
2820 put_task_struct(task);
2822 percpu_up_write(&cgroup_threadgroup_rwsem);
2823 for_each_subsys(ss, ssid)
2824 if (ss->post_attach)
2828 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2830 struct cgroup_subsys *ss;
2831 bool printed = false;
2834 do_each_subsys_mask(ss, ssid, ss_mask) {
2837 seq_printf(seq, "%s", ss->name);
2839 } while_each_subsys_mask();
2841 seq_putc(seq, '\n');
2844 /* show controllers which are enabled from the parent */
2845 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2847 struct cgroup *cgrp = seq_css(seq)->cgroup;
2849 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2853 /* show controllers which are enabled for a given cgroup's children */
2854 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2856 struct cgroup *cgrp = seq_css(seq)->cgroup;
2858 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2863 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2864 * @cgrp: root of the subtree to update csses for
2866 * @cgrp's control masks have changed and its subtree's css associations
2867 * need to be updated accordingly. This function looks up all css_sets
2868 * which are attached to the subtree, creates the matching updated css_sets
2869 * and migrates the tasks to the new ones.
2871 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2873 DEFINE_CGROUP_MGCTX(mgctx);
2874 struct cgroup_subsys_state *d_css;
2875 struct cgroup *dsct;
2876 struct css_set *src_cset;
2879 lockdep_assert_held(&cgroup_mutex);
2881 percpu_down_write(&cgroup_threadgroup_rwsem);
2883 /* look up all csses currently attached to @cgrp's subtree */
2884 spin_lock_irq(&css_set_lock);
2885 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2886 struct cgrp_cset_link *link;
2888 list_for_each_entry(link, &dsct->cset_links, cset_link)
2889 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2891 spin_unlock_irq(&css_set_lock);
2893 /* NULL dst indicates self on default hierarchy */
2894 ret = cgroup_migrate_prepare_dst(&mgctx);
2898 spin_lock_irq(&css_set_lock);
2899 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2900 struct task_struct *task, *ntask;
2902 /* all tasks in src_csets need to be migrated */
2903 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2904 cgroup_migrate_add_task(task, &mgctx);
2906 spin_unlock_irq(&css_set_lock);
2908 ret = cgroup_migrate_execute(&mgctx);
2910 cgroup_migrate_finish(&mgctx);
2911 percpu_up_write(&cgroup_threadgroup_rwsem);
2916 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2917 * @cgrp: root of the target subtree
2919 * Because css offlining is asynchronous, userland may try to re-enable a
2920 * controller while the previous css is still around. This function grabs
2921 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2923 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2924 __acquires(&cgroup_mutex)
2926 struct cgroup *dsct;
2927 struct cgroup_subsys_state *d_css;
2928 struct cgroup_subsys *ss;
2932 mutex_lock(&cgroup_mutex);
2934 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2935 for_each_subsys(ss, ssid) {
2936 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2939 if (!css || !percpu_ref_is_dying(&css->refcnt))
2942 cgroup_get_live(dsct);
2943 prepare_to_wait(&dsct->offline_waitq, &wait,
2944 TASK_UNINTERRUPTIBLE);
2946 mutex_unlock(&cgroup_mutex);
2948 finish_wait(&dsct->offline_waitq, &wait);
2957 * cgroup_save_control - save control masks and dom_cgrp of a subtree
2958 * @cgrp: root of the target subtree
2960 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
2961 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2964 static void cgroup_save_control(struct cgroup *cgrp)
2966 struct cgroup *dsct;
2967 struct cgroup_subsys_state *d_css;
2969 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2970 dsct->old_subtree_control = dsct->subtree_control;
2971 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
2972 dsct->old_dom_cgrp = dsct->dom_cgrp;
2977 * cgroup_propagate_control - refresh control masks of a subtree
2978 * @cgrp: root of the target subtree
2980 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2981 * ->subtree_control and propagate controller availability through the
2982 * subtree so that descendants don't have unavailable controllers enabled.
2984 static void cgroup_propagate_control(struct cgroup *cgrp)
2986 struct cgroup *dsct;
2987 struct cgroup_subsys_state *d_css;
2989 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2990 dsct->subtree_control &= cgroup_control(dsct);
2991 dsct->subtree_ss_mask =
2992 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
2993 cgroup_ss_mask(dsct));
2998 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
2999 * @cgrp: root of the target subtree
3001 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3002 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3005 static void cgroup_restore_control(struct cgroup *cgrp)
3007 struct cgroup *dsct;
3008 struct cgroup_subsys_state *d_css;
3010 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3011 dsct->subtree_control = dsct->old_subtree_control;
3012 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3013 dsct->dom_cgrp = dsct->old_dom_cgrp;
3017 static bool css_visible(struct cgroup_subsys_state *css)
3019 struct cgroup_subsys *ss = css->ss;
3020 struct cgroup *cgrp = css->cgroup;
3022 if (cgroup_control(cgrp) & (1 << ss->id))
3024 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3026 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3030 * cgroup_apply_control_enable - enable or show csses according to control
3031 * @cgrp: root of the target subtree
3033 * Walk @cgrp's subtree and create new csses or make the existing ones
3034 * visible. A css is created invisible if it's being implicitly enabled
3035 * through dependency. An invisible css is made visible when the userland
3036 * explicitly enables it.
3038 * Returns 0 on success, -errno on failure. On failure, csses which have
3039 * been processed already aren't cleaned up. The caller is responsible for
3040 * cleaning up with cgroup_apply_control_disable().
3042 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3044 struct cgroup *dsct;
3045 struct cgroup_subsys_state *d_css;
3046 struct cgroup_subsys *ss;
3049 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3050 for_each_subsys(ss, ssid) {
3051 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3053 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3055 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3059 css = css_create(dsct, ss);
3061 return PTR_ERR(css);
3064 if (css_visible(css)) {
3065 ret = css_populate_dir(css);
3076 * cgroup_apply_control_disable - kill or hide csses according to control
3077 * @cgrp: root of the target subtree
3079 * Walk @cgrp's subtree and kill and hide csses so that they match
3080 * cgroup_ss_mask() and cgroup_visible_mask().
3082 * A css is hidden when the userland requests it to be disabled while other
3083 * subsystems are still depending on it. The css must not actively control
3084 * resources and be in the vanilla state if it's made visible again later.
3085 * Controllers which may be depended upon should provide ->css_reset() for
3088 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3090 struct cgroup *dsct;
3091 struct cgroup_subsys_state *d_css;
3092 struct cgroup_subsys *ss;
3095 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3096 for_each_subsys(ss, ssid) {
3097 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3099 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3105 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3107 } else if (!css_visible(css)) {
3117 * cgroup_apply_control - apply control mask updates to the subtree
3118 * @cgrp: root of the target subtree
3120 * subsystems can be enabled and disabled in a subtree using the following
3123 * 1. Call cgroup_save_control() to stash the current state.
3124 * 2. Update ->subtree_control masks in the subtree as desired.
3125 * 3. Call cgroup_apply_control() to apply the changes.
3126 * 4. Optionally perform other related operations.
3127 * 5. Call cgroup_finalize_control() to finish up.
3129 * This function implements step 3 and propagates the mask changes
3130 * throughout @cgrp's subtree, updates csses accordingly and perform
3131 * process migrations.
3133 static int cgroup_apply_control(struct cgroup *cgrp)
3137 cgroup_propagate_control(cgrp);
3139 ret = cgroup_apply_control_enable(cgrp);
3144 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3145 * making the following cgroup_update_dfl_csses() properly update
3146 * css associations of all tasks in the subtree.
3148 ret = cgroup_update_dfl_csses(cgrp);
3156 * cgroup_finalize_control - finalize control mask update
3157 * @cgrp: root of the target subtree
3158 * @ret: the result of the update
3160 * Finalize control mask update. See cgroup_apply_control() for more info.
3162 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3165 cgroup_restore_control(cgrp);
3166 cgroup_propagate_control(cgrp);
3169 cgroup_apply_control_disable(cgrp);
3172 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3174 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3176 /* if nothing is getting enabled, nothing to worry about */
3180 /* can @cgrp host any resources? */
3181 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3184 /* mixables don't care */
3185 if (cgroup_is_mixable(cgrp))
3188 if (domain_enable) {
3189 /* can't enable domain controllers inside a thread subtree */
3190 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3194 * Threaded controllers can handle internal competitions
3195 * and are always allowed inside a (prospective) thread
3198 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3203 * Controllers can't be enabled for a cgroup with tasks to avoid
3204 * child cgroups competing against tasks.
3206 if (cgroup_has_tasks(cgrp))
3212 /* change the enabled child controllers for a cgroup in the default hierarchy */
3213 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3214 char *buf, size_t nbytes,
3217 u16 enable = 0, disable = 0;
3218 struct cgroup *cgrp, *child;
3219 struct cgroup_subsys *ss;
3224 * Parse input - space separated list of subsystem names prefixed
3225 * with either + or -.
3227 buf = strstrip(buf);
3228 while ((tok = strsep(&buf, " "))) {
3231 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3232 if (!cgroup_ssid_enabled(ssid) ||
3233 strcmp(tok + 1, ss->name))
3237 enable |= 1 << ssid;
3238 disable &= ~(1 << ssid);
3239 } else if (*tok == '-') {
3240 disable |= 1 << ssid;
3241 enable &= ~(1 << ssid);
3246 } while_each_subsys_mask();
3247 if (ssid == CGROUP_SUBSYS_COUNT)
3251 cgrp = cgroup_kn_lock_live(of->kn, true);
3255 for_each_subsys(ss, ssid) {
3256 if (enable & (1 << ssid)) {
3257 if (cgrp->subtree_control & (1 << ssid)) {
3258 enable &= ~(1 << ssid);
3262 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3266 } else if (disable & (1 << ssid)) {
3267 if (!(cgrp->subtree_control & (1 << ssid))) {
3268 disable &= ~(1 << ssid);
3272 /* a child has it enabled? */
3273 cgroup_for_each_live_child(child, cgrp) {
3274 if (child->subtree_control & (1 << ssid)) {
3282 if (!enable && !disable) {
3287 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3291 /* save and update control masks and prepare csses */
3292 cgroup_save_control(cgrp);
3294 cgrp->subtree_control |= enable;
3295 cgrp->subtree_control &= ~disable;
3297 ret = cgroup_apply_control(cgrp);
3298 cgroup_finalize_control(cgrp, ret);
3302 kernfs_activate(cgrp->kn);
3304 cgroup_kn_unlock(of->kn);
3305 return ret ?: nbytes;
3309 * cgroup_enable_threaded - make @cgrp threaded
3310 * @cgrp: the target cgroup
3312 * Called when "threaded" is written to the cgroup.type interface file and
3313 * tries to make @cgrp threaded and join the parent's resource domain.
3314 * This function is never called on the root cgroup as cgroup.type doesn't
3317 static int cgroup_enable_threaded(struct cgroup *cgrp)
3319 struct cgroup *parent = cgroup_parent(cgrp);
3320 struct cgroup *dom_cgrp = parent->dom_cgrp;
3321 struct cgroup *dsct;
3322 struct cgroup_subsys_state *d_css;
3325 lockdep_assert_held(&cgroup_mutex);
3327 /* noop if already threaded */
3328 if (cgroup_is_threaded(cgrp))
3332 * If @cgroup is populated or has domain controllers enabled, it
3333 * can't be switched. While the below cgroup_can_be_thread_root()
3334 * test can catch the same conditions, that's only when @parent is
3335 * not mixable, so let's check it explicitly.
3337 if (cgroup_is_populated(cgrp) ||
3338 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3341 /* we're joining the parent's domain, ensure its validity */
3342 if (!cgroup_is_valid_domain(dom_cgrp) ||
3343 !cgroup_can_be_thread_root(dom_cgrp))
3347 * The following shouldn't cause actual migrations and should
3350 cgroup_save_control(cgrp);
3352 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3353 if (dsct == cgrp || cgroup_is_threaded(dsct))
3354 dsct->dom_cgrp = dom_cgrp;
3356 ret = cgroup_apply_control(cgrp);
3358 parent->nr_threaded_children++;
3360 cgroup_finalize_control(cgrp, ret);
3364 static int cgroup_type_show(struct seq_file *seq, void *v)
3366 struct cgroup *cgrp = seq_css(seq)->cgroup;
3368 if (cgroup_is_threaded(cgrp))
3369 seq_puts(seq, "threaded\n");
3370 else if (!cgroup_is_valid_domain(cgrp))
3371 seq_puts(seq, "domain invalid\n");
3372 else if (cgroup_is_thread_root(cgrp))
3373 seq_puts(seq, "domain threaded\n");
3375 seq_puts(seq, "domain\n");
3380 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3381 size_t nbytes, loff_t off)
3383 struct cgroup *cgrp;
3386 /* only switching to threaded mode is supported */
3387 if (strcmp(strstrip(buf), "threaded"))
3390 cgrp = cgroup_kn_lock_live(of->kn, false);
3394 /* threaded can only be enabled */
3395 ret = cgroup_enable_threaded(cgrp);
3397 cgroup_kn_unlock(of->kn);
3398 return ret ?: nbytes;
3401 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3403 struct cgroup *cgrp = seq_css(seq)->cgroup;
3404 int descendants = READ_ONCE(cgrp->max_descendants);
3406 if (descendants == INT_MAX)
3407 seq_puts(seq, "max\n");
3409 seq_printf(seq, "%d\n", descendants);
3414 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3415 char *buf, size_t nbytes, loff_t off)
3417 struct cgroup *cgrp;
3421 buf = strstrip(buf);
3422 if (!strcmp(buf, "max")) {
3423 descendants = INT_MAX;
3425 ret = kstrtoint(buf, 0, &descendants);
3430 if (descendants < 0)
3433 cgrp = cgroup_kn_lock_live(of->kn, false);
3437 cgrp->max_descendants = descendants;
3439 cgroup_kn_unlock(of->kn);
3444 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3446 struct cgroup *cgrp = seq_css(seq)->cgroup;
3447 int depth = READ_ONCE(cgrp->max_depth);
3449 if (depth == INT_MAX)
3450 seq_puts(seq, "max\n");
3452 seq_printf(seq, "%d\n", depth);
3457 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3458 char *buf, size_t nbytes, loff_t off)
3460 struct cgroup *cgrp;
3464 buf = strstrip(buf);
3465 if (!strcmp(buf, "max")) {
3468 ret = kstrtoint(buf, 0, &depth);
3476 cgrp = cgroup_kn_lock_live(of->kn, false);
3480 cgrp->max_depth = depth;
3482 cgroup_kn_unlock(of->kn);
3487 static int cgroup_events_show(struct seq_file *seq, void *v)
3489 struct cgroup *cgrp = seq_css(seq)->cgroup;
3491 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3492 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3497 static int cgroup_stat_show(struct seq_file *seq, void *v)
3499 struct cgroup *cgroup = seq_css(seq)->cgroup;
3501 seq_printf(seq, "nr_descendants %d\n",
3502 cgroup->nr_descendants);
3503 seq_printf(seq, "nr_dying_descendants %d\n",
3504 cgroup->nr_dying_descendants);
3509 static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3510 struct cgroup *cgrp, int ssid)
3512 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3513 struct cgroup_subsys_state *css;
3516 if (!ss->css_extra_stat_show)
3519 css = cgroup_tryget_css(cgrp, ss);
3523 ret = ss->css_extra_stat_show(seq, css);
3528 static int cpu_stat_show(struct seq_file *seq, void *v)
3530 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3533 cgroup_base_stat_cputime_show(seq);
3534 #ifdef CONFIG_CGROUP_SCHED
3535 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3541 static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3543 return psi_show(seq, &seq_css(seq)->cgroup->psi, PSI_IO);
3545 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3547 return psi_show(seq, &seq_css(seq)->cgroup->psi, PSI_MEM);
3549 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3551 return psi_show(seq, &seq_css(seq)->cgroup->psi, PSI_CPU);
3555 static int cgroup_freeze_show(struct seq_file *seq, void *v)
3557 struct cgroup *cgrp = seq_css(seq)->cgroup;
3559 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3564 static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3565 char *buf, size_t nbytes, loff_t off)
3567 struct cgroup *cgrp;
3571 ret = kstrtoint(strstrip(buf), 0, &freeze);
3575 if (freeze < 0 || freeze > 1)
3578 cgrp = cgroup_kn_lock_live(of->kn, false);
3582 cgroup_freeze(cgrp, freeze);
3584 cgroup_kn_unlock(of->kn);
3589 static int cgroup_file_open(struct kernfs_open_file *of)
3591 struct cftype *cft = of->kn->priv;
3594 return cft->open(of);
3598 static void cgroup_file_release(struct kernfs_open_file *of)
3600 struct cftype *cft = of->kn->priv;
3606 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3607 size_t nbytes, loff_t off)
3609 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3610 struct cgroup *cgrp = of->kn->parent->priv;
3611 struct cftype *cft = of->kn->priv;
3612 struct cgroup_subsys_state *css;
3616 * If namespaces are delegation boundaries, disallow writes to
3617 * files in an non-init namespace root from inside the namespace
3618 * except for the files explicitly marked delegatable -
3619 * cgroup.procs and cgroup.subtree_control.
3621 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3622 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3623 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3627 return cft->write(of, buf, nbytes, off);
3630 * kernfs guarantees that a file isn't deleted with operations in
3631 * flight, which means that the matching css is and stays alive and
3632 * doesn't need to be pinned. The RCU locking is not necessary
3633 * either. It's just for the convenience of using cgroup_css().
3636 css = cgroup_css(cgrp, cft->ss);
3639 if (cft->write_u64) {
3640 unsigned long long v;
3641 ret = kstrtoull(buf, 0, &v);
3643 ret = cft->write_u64(css, cft, v);
3644 } else if (cft->write_s64) {
3646 ret = kstrtoll(buf, 0, &v);
3648 ret = cft->write_s64(css, cft, v);
3653 return ret ?: nbytes;
3656 static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
3658 struct cftype *cft = of->kn->priv;
3661 return cft->poll(of, pt);
3663 return kernfs_generic_poll(of, pt);
3666 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3668 return seq_cft(seq)->seq_start(seq, ppos);
3671 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3673 return seq_cft(seq)->seq_next(seq, v, ppos);
3676 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3678 if (seq_cft(seq)->seq_stop)
3679 seq_cft(seq)->seq_stop(seq, v);
3682 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3684 struct cftype *cft = seq_cft(m);
3685 struct cgroup_subsys_state *css = seq_css(m);
3688 return cft->seq_show(m, arg);
3691 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3692 else if (cft->read_s64)
3693 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3699 static struct kernfs_ops cgroup_kf_single_ops = {
3700 .atomic_write_len = PAGE_SIZE,
3701 .open = cgroup_file_open,
3702 .release = cgroup_file_release,
3703 .write = cgroup_file_write,
3704 .poll = cgroup_file_poll,
3705 .seq_show = cgroup_seqfile_show,
3708 static struct kernfs_ops cgroup_kf_ops = {
3709 .atomic_write_len = PAGE_SIZE,
3710 .open = cgroup_file_open,
3711 .release = cgroup_file_release,
3712 .write = cgroup_file_write,
3713 .poll = cgroup_file_poll,
3714 .seq_start = cgroup_seqfile_start,
3715 .seq_next = cgroup_seqfile_next,
3716 .seq_stop = cgroup_seqfile_stop,
3717 .seq_show = cgroup_seqfile_show,
3720 /* set uid and gid of cgroup dirs and files to that of the creator */
3721 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3723 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3724 .ia_uid = current_fsuid(),
3725 .ia_gid = current_fsgid(), };
3727 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3728 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3731 return kernfs_setattr(kn, &iattr);
3734 static void cgroup_file_notify_timer(struct timer_list *timer)
3736 cgroup_file_notify(container_of(timer, struct cgroup_file,
3740 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3743 char name[CGROUP_FILE_NAME_MAX];
3744 struct kernfs_node *kn;
3745 struct lock_class_key *key = NULL;
3748 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3749 key = &cft->lockdep_key;
3751 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3752 cgroup_file_mode(cft),
3753 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
3754 0, cft->kf_ops, cft,
3759 ret = cgroup_kn_set_ugid(kn);
3765 if (cft->file_offset) {
3766 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3768 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
3770 spin_lock_irq(&cgroup_file_kn_lock);
3772 spin_unlock_irq(&cgroup_file_kn_lock);
3779 * cgroup_addrm_files - add or remove files to a cgroup directory
3780 * @css: the target css
3781 * @cgrp: the target cgroup (usually css->cgroup)
3782 * @cfts: array of cftypes to be added
3783 * @is_add: whether to add or remove
3785 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3786 * For removals, this function never fails.
3788 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3789 struct cgroup *cgrp, struct cftype cfts[],
3792 struct cftype *cft, *cft_end = NULL;
3795 lockdep_assert_held(&cgroup_mutex);
3798 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3799 /* does cft->flags tell us to skip this file on @cgrp? */
3800 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3802 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3804 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3806 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3808 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
3811 ret = cgroup_add_file(css, cgrp, cft);
3813 pr_warn("%s: failed to add %s, err=%d\n",
3814 __func__, cft->name, ret);
3820 cgroup_rm_file(cgrp, cft);
3826 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3828 struct cgroup_subsys *ss = cfts[0].ss;
3829 struct cgroup *root = &ss->root->cgrp;
3830 struct cgroup_subsys_state *css;
3833 lockdep_assert_held(&cgroup_mutex);
3835 /* add/rm files for all cgroups created before */
3836 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3837 struct cgroup *cgrp = css->cgroup;
3839 if (!(css->flags & CSS_VISIBLE))
3842 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3848 kernfs_activate(root->kn);
3852 static void cgroup_exit_cftypes(struct cftype *cfts)
3856 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3857 /* free copy for custom atomic_write_len, see init_cftypes() */
3858 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3863 /* revert flags set by cgroup core while adding @cfts */
3864 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3868 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3872 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3873 struct kernfs_ops *kf_ops;
3875 WARN_ON(cft->ss || cft->kf_ops);
3878 kf_ops = &cgroup_kf_ops;
3880 kf_ops = &cgroup_kf_single_ops;
3883 * Ugh... if @cft wants a custom max_write_len, we need to
3884 * make a copy of kf_ops to set its atomic_write_len.
3886 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3887 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3889 cgroup_exit_cftypes(cfts);
3892 kf_ops->atomic_write_len = cft->max_write_len;
3895 cft->kf_ops = kf_ops;
3902 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3904 lockdep_assert_held(&cgroup_mutex);
3906 if (!cfts || !cfts[0].ss)
3909 list_del(&cfts->node);
3910 cgroup_apply_cftypes(cfts, false);
3911 cgroup_exit_cftypes(cfts);
3916 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3917 * @cfts: zero-length name terminated array of cftypes
3919 * Unregister @cfts. Files described by @cfts are removed from all
3920 * existing cgroups and all future cgroups won't have them either. This
3921 * function can be called anytime whether @cfts' subsys is attached or not.
3923 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3926 int cgroup_rm_cftypes(struct cftype *cfts)
3930 mutex_lock(&cgroup_mutex);
3931 ret = cgroup_rm_cftypes_locked(cfts);
3932 mutex_unlock(&cgroup_mutex);
3937 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3938 * @ss: target cgroup subsystem
3939 * @cfts: zero-length name terminated array of cftypes
3941 * Register @cfts to @ss. Files described by @cfts are created for all
3942 * existing cgroups to which @ss is attached and all future cgroups will
3943 * have them too. This function can be called anytime whether @ss is
3946 * Returns 0 on successful registration, -errno on failure. Note that this
3947 * function currently returns 0 as long as @cfts registration is successful
3948 * even if some file creation attempts on existing cgroups fail.
3950 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3954 if (!cgroup_ssid_enabled(ss->id))
3957 if (!cfts || cfts[0].name[0] == '\0')
3960 ret = cgroup_init_cftypes(ss, cfts);
3964 mutex_lock(&cgroup_mutex);
3966 list_add_tail(&cfts->node, &ss->cfts);
3967 ret = cgroup_apply_cftypes(cfts, true);
3969 cgroup_rm_cftypes_locked(cfts);
3971 mutex_unlock(&cgroup_mutex);
3976 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3977 * @ss: target cgroup subsystem
3978 * @cfts: zero-length name terminated array of cftypes
3980 * Similar to cgroup_add_cftypes() but the added files are only used for
3981 * the default hierarchy.
3983 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3987 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3988 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3989 return cgroup_add_cftypes(ss, cfts);
3993 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3994 * @ss: target cgroup subsystem
3995 * @cfts: zero-length name terminated array of cftypes
3997 * Similar to cgroup_add_cftypes() but the added files are only used for
3998 * the legacy hierarchies.
4000 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4004 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4005 cft->flags |= __CFTYPE_NOT_ON_DFL;
4006 return cgroup_add_cftypes(ss, cfts);
4010 * cgroup_file_notify - generate a file modified event for a cgroup_file
4011 * @cfile: target cgroup_file
4013 * @cfile must have been obtained by setting cftype->file_offset.
4015 void cgroup_file_notify(struct cgroup_file *cfile)
4017 unsigned long flags;
4019 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4021 unsigned long last = cfile->notified_at;
4022 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4024 if (time_in_range(jiffies, last, next)) {
4025 timer_reduce(&cfile->notify_timer, next);
4027 kernfs_notify(cfile->kn);
4028 cfile->notified_at = jiffies;
4031 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4035 * css_next_child - find the next child of a given css
4036 * @pos: the current position (%NULL to initiate traversal)
4037 * @parent: css whose children to walk
4039 * This function returns the next child of @parent and should be called
4040 * under either cgroup_mutex or RCU read lock. The only requirement is
4041 * that @parent and @pos are accessible. The next sibling is guaranteed to
4042 * be returned regardless of their states.
4044 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4045 * css which finished ->css_online() is guaranteed to be visible in the
4046 * future iterations and will stay visible until the last reference is put.
4047 * A css which hasn't finished ->css_online() or already finished
4048 * ->css_offline() may show up during traversal. It's each subsystem's
4049 * responsibility to synchronize against on/offlining.
4051 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4052 struct cgroup_subsys_state *parent)
4054 struct cgroup_subsys_state *next;
4056 cgroup_assert_mutex_or_rcu_locked();
4059 * @pos could already have been unlinked from the sibling list.
4060 * Once a cgroup is removed, its ->sibling.next is no longer
4061 * updated when its next sibling changes. CSS_RELEASED is set when
4062 * @pos is taken off list, at which time its next pointer is valid,
4063 * and, as releases are serialized, the one pointed to by the next
4064 * pointer is guaranteed to not have started release yet. This
4065 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4066 * critical section, the one pointed to by its next pointer is
4067 * guaranteed to not have finished its RCU grace period even if we
4068 * have dropped rcu_read_lock() inbetween iterations.
4070 * If @pos has CSS_RELEASED set, its next pointer can't be
4071 * dereferenced; however, as each css is given a monotonically
4072 * increasing unique serial number and always appended to the
4073 * sibling list, the next one can be found by walking the parent's
4074 * children until the first css with higher serial number than
4075 * @pos's. While this path can be slower, it happens iff iteration
4076 * races against release and the race window is very small.
4079 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4080 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4081 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4083 list_for_each_entry_rcu(next, &parent->children, sibling)
4084 if (next->serial_nr > pos->serial_nr)
4089 * @next, if not pointing to the head, can be dereferenced and is
4092 if (&next->sibling != &parent->children)
4098 * css_next_descendant_pre - find the next descendant for pre-order walk
4099 * @pos: the current position (%NULL to initiate traversal)
4100 * @root: css whose descendants to walk
4102 * To be used by css_for_each_descendant_pre(). Find the next descendant
4103 * to visit for pre-order traversal of @root's descendants. @root is
4104 * included in the iteration and the first node to be visited.
4106 * While this function requires cgroup_mutex or RCU read locking, it
4107 * doesn't require the whole traversal to be contained in a single critical
4108 * section. This function will return the correct next descendant as long
4109 * as both @pos and @root are accessible and @pos is a descendant of @root.
4111 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4112 * css which finished ->css_online() is guaranteed to be visible in the
4113 * future iterations and will stay visible until the last reference is put.
4114 * A css which hasn't finished ->css_online() or already finished
4115 * ->css_offline() may show up during traversal. It's each subsystem's
4116 * responsibility to synchronize against on/offlining.
4118 struct cgroup_subsys_state *
4119 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4120 struct cgroup_subsys_state *root)
4122 struct cgroup_subsys_state *next;
4124 cgroup_assert_mutex_or_rcu_locked();
4126 /* if first iteration, visit @root */
4130 /* visit the first child if exists */
4131 next = css_next_child(NULL, pos);
4135 /* no child, visit my or the closest ancestor's next sibling */
4136 while (pos != root) {
4137 next = css_next_child(pos, pos->parent);
4147 * css_rightmost_descendant - return the rightmost descendant of a css
4148 * @pos: css of interest
4150 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4151 * is returned. This can be used during pre-order traversal to skip
4154 * While this function requires cgroup_mutex or RCU read locking, it
4155 * doesn't require the whole traversal to be contained in a single critical
4156 * section. This function will return the correct rightmost descendant as
4157 * long as @pos is accessible.
4159 struct cgroup_subsys_state *
4160 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4162 struct cgroup_subsys_state *last, *tmp;
4164 cgroup_assert_mutex_or_rcu_locked();
4168 /* ->prev isn't RCU safe, walk ->next till the end */
4170 css_for_each_child(tmp, last)
4177 static struct cgroup_subsys_state *
4178 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4180 struct cgroup_subsys_state *last;
4184 pos = css_next_child(NULL, pos);
4191 * css_next_descendant_post - find the next descendant for post-order walk
4192 * @pos: the current position (%NULL to initiate traversal)
4193 * @root: css whose descendants to walk
4195 * To be used by css_for_each_descendant_post(). Find the next descendant
4196 * to visit for post-order traversal of @root's descendants. @root is
4197 * included in the iteration and the last node to be visited.
4199 * While this function requires cgroup_mutex or RCU read locking, it
4200 * doesn't require the whole traversal to be contained in a single critical
4201 * section. This function will return the correct next descendant as long
4202 * as both @pos and @cgroup are accessible and @pos is a descendant of
4205 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4206 * css which finished ->css_online() is guaranteed to be visible in the
4207 * future iterations and will stay visible until the last reference is put.
4208 * A css which hasn't finished ->css_online() or already finished
4209 * ->css_offline() may show up during traversal. It's each subsystem's
4210 * responsibility to synchronize against on/offlining.
4212 struct cgroup_subsys_state *
4213 css_next_descendant_post(struct cgroup_subsys_state *pos,
4214 struct cgroup_subsys_state *root)
4216 struct cgroup_subsys_state *next;
4218 cgroup_assert_mutex_or_rcu_locked();
4220 /* if first iteration, visit leftmost descendant which may be @root */
4222 return css_leftmost_descendant(root);
4224 /* if we visited @root, we're done */
4228 /* if there's an unvisited sibling, visit its leftmost descendant */
4229 next = css_next_child(pos, pos->parent);
4231 return css_leftmost_descendant(next);
4233 /* no sibling left, visit parent */
4238 * css_has_online_children - does a css have online children
4239 * @css: the target css
4241 * Returns %true if @css has any online children; otherwise, %false. This
4242 * function can be called from any context but the caller is responsible
4243 * for synchronizing against on/offlining as necessary.
4245 bool css_has_online_children(struct cgroup_subsys_state *css)
4247 struct cgroup_subsys_state *child;
4251 css_for_each_child(child, css) {
4252 if (child->flags & CSS_ONLINE) {
4261 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4263 struct list_head *l;
4264 struct cgrp_cset_link *link;
4265 struct css_set *cset;
4267 lockdep_assert_held(&css_set_lock);
4269 /* find the next threaded cset */
4270 if (it->tcset_pos) {
4271 l = it->tcset_pos->next;
4273 if (l != it->tcset_head) {
4275 return container_of(l, struct css_set,
4276 threaded_csets_node);
4279 it->tcset_pos = NULL;
4282 /* find the next cset */
4285 if (l == it->cset_head) {
4286 it->cset_pos = NULL;
4291 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4293 link = list_entry(l, struct cgrp_cset_link, cset_link);
4299 /* initialize threaded css_set walking */
4300 if (it->flags & CSS_TASK_ITER_THREADED) {
4302 put_css_set_locked(it->cur_dcset);
4303 it->cur_dcset = cset;
4306 it->tcset_head = &cset->threaded_csets;
4307 it->tcset_pos = &cset->threaded_csets;
4314 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4315 * @it: the iterator to advance
4317 * Advance @it to the next css_set to walk.
4319 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4321 struct css_set *cset;
4323 lockdep_assert_held(&css_set_lock);
4325 /* Advance to the next non-empty css_set */
4327 cset = css_task_iter_next_css_set(it);
4329 it->task_pos = NULL;
4332 } while (!css_set_populated(cset));
4334 if (!list_empty(&cset->tasks))
4335 it->task_pos = cset->tasks.next;
4337 it->task_pos = cset->mg_tasks.next;
4339 it->tasks_head = &cset->tasks;
4340 it->mg_tasks_head = &cset->mg_tasks;
4343 * We don't keep css_sets locked across iteration steps and thus
4344 * need to take steps to ensure that iteration can be resumed after
4345 * the lock is re-acquired. Iteration is performed at two levels -
4346 * css_sets and tasks in them.
4348 * Once created, a css_set never leaves its cgroup lists, so a
4349 * pinned css_set is guaranteed to stay put and we can resume
4350 * iteration afterwards.
4352 * Tasks may leave @cset across iteration steps. This is resolved
4353 * by registering each iterator with the css_set currently being
4354 * walked and making css_set_move_task() advance iterators whose
4355 * next task is leaving.
4358 list_del(&it->iters_node);
4359 put_css_set_locked(it->cur_cset);
4362 it->cur_cset = cset;
4363 list_add(&it->iters_node, &cset->task_iters);
4366 static void css_task_iter_advance(struct css_task_iter *it)
4368 struct list_head *next;
4370 lockdep_assert_held(&css_set_lock);
4374 * Advance iterator to find next entry. cset->tasks is
4375 * consumed first and then ->mg_tasks. After ->mg_tasks,
4376 * we move onto the next cset.
4378 next = it->task_pos->next;
4380 if (next == it->tasks_head)
4381 next = it->mg_tasks_head->next;
4383 if (next == it->mg_tasks_head)
4384 css_task_iter_advance_css_set(it);
4386 it->task_pos = next;
4388 /* called from start, proceed to the first cset */
4389 css_task_iter_advance_css_set(it);
4392 /* if PROCS, skip over tasks which aren't group leaders */
4393 if ((it->flags & CSS_TASK_ITER_PROCS) && it->task_pos &&
4394 !thread_group_leader(list_entry(it->task_pos, struct task_struct,
4400 * css_task_iter_start - initiate task iteration
4401 * @css: the css to walk tasks of
4402 * @flags: CSS_TASK_ITER_* flags
4403 * @it: the task iterator to use
4405 * Initiate iteration through the tasks of @css. The caller can call
4406 * css_task_iter_next() to walk through the tasks until the function
4407 * returns NULL. On completion of iteration, css_task_iter_end() must be
4410 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4411 struct css_task_iter *it)
4413 /* no one should try to iterate before mounting cgroups */
4414 WARN_ON_ONCE(!use_task_css_set_links);
4416 memset(it, 0, sizeof(*it));
4418 spin_lock_irq(&css_set_lock);
4424 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4426 it->cset_pos = &css->cgroup->cset_links;
4428 it->cset_head = it->cset_pos;
4430 css_task_iter_advance(it);
4432 spin_unlock_irq(&css_set_lock);
4436 * css_task_iter_next - return the next task for the iterator
4437 * @it: the task iterator being iterated
4439 * The "next" function for task iteration. @it should have been
4440 * initialized via css_task_iter_start(). Returns NULL when the iteration
4443 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4446 put_task_struct(it->cur_task);
4447 it->cur_task = NULL;
4450 spin_lock_irq(&css_set_lock);
4453 it->cur_task = list_entry(it->task_pos, struct task_struct,
4455 get_task_struct(it->cur_task);
4456 css_task_iter_advance(it);
4459 spin_unlock_irq(&css_set_lock);
4461 return it->cur_task;
4465 * css_task_iter_end - finish task iteration
4466 * @it: the task iterator to finish
4468 * Finish task iteration started by css_task_iter_start().
4470 void css_task_iter_end(struct css_task_iter *it)
4473 spin_lock_irq(&css_set_lock);
4474 list_del(&it->iters_node);
4475 put_css_set_locked(it->cur_cset);
4476 spin_unlock_irq(&css_set_lock);
4480 put_css_set(it->cur_dcset);
4483 put_task_struct(it->cur_task);
4486 static void cgroup_procs_release(struct kernfs_open_file *of)
4489 css_task_iter_end(of->priv);
4494 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4496 struct kernfs_open_file *of = s->private;
4497 struct css_task_iter *it = of->priv;
4499 return css_task_iter_next(it);
4502 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4503 unsigned int iter_flags)
4505 struct kernfs_open_file *of = s->private;
4506 struct cgroup *cgrp = seq_css(s)->cgroup;
4507 struct css_task_iter *it = of->priv;
4510 * When a seq_file is seeked, it's always traversed sequentially
4511 * from position 0, so we can simply keep iterating on !0 *pos.
4514 if (WARN_ON_ONCE((*pos)++))
4515 return ERR_PTR(-EINVAL);
4517 it = kzalloc(sizeof(*it), GFP_KERNEL);
4519 return ERR_PTR(-ENOMEM);
4521 css_task_iter_start(&cgrp->self, iter_flags, it);
4522 } else if (!(*pos)++) {
4523 css_task_iter_end(it);
4524 css_task_iter_start(&cgrp->self, iter_flags, it);
4527 return cgroup_procs_next(s, NULL, NULL);
4530 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4532 struct cgroup *cgrp = seq_css(s)->cgroup;
4535 * All processes of a threaded subtree belong to the domain cgroup
4536 * of the subtree. Only threads can be distributed across the
4537 * subtree. Reject reads on cgroup.procs in the subtree proper.
4538 * They're always empty anyway.
4540 if (cgroup_is_threaded(cgrp))
4541 return ERR_PTR(-EOPNOTSUPP);
4543 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4544 CSS_TASK_ITER_THREADED);
4547 static int cgroup_procs_show(struct seq_file *s, void *v)
4549 seq_printf(s, "%d\n", task_pid_vnr(v));
4553 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4554 struct cgroup *dst_cgrp,
4555 struct super_block *sb)
4557 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4558 struct cgroup *com_cgrp = src_cgrp;
4559 struct inode *inode;
4562 lockdep_assert_held(&cgroup_mutex);
4564 /* find the common ancestor */
4565 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4566 com_cgrp = cgroup_parent(com_cgrp);
4568 /* %current should be authorized to migrate to the common ancestor */
4569 inode = kernfs_get_inode(sb, com_cgrp->procs_file.kn);
4573 ret = inode_permission(inode, MAY_WRITE);
4579 * If namespaces are delegation boundaries, %current must be able
4580 * to see both source and destination cgroups from its namespace.
4582 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4583 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4584 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4590 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4591 char *buf, size_t nbytes, loff_t off)
4593 struct cgroup *src_cgrp, *dst_cgrp;
4594 struct task_struct *task;
4597 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4601 task = cgroup_procs_write_start(buf, true);
4602 ret = PTR_ERR_OR_ZERO(task);
4606 /* find the source cgroup */
4607 spin_lock_irq(&css_set_lock);
4608 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4609 spin_unlock_irq(&css_set_lock);
4611 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4612 of->file->f_path.dentry->d_sb);
4616 ret = cgroup_attach_task(dst_cgrp, task, true);
4619 cgroup_procs_write_finish(task);
4621 cgroup_kn_unlock(of->kn);
4623 return ret ?: nbytes;
4626 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4628 return __cgroup_procs_start(s, pos, 0);
4631 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4632 char *buf, size_t nbytes, loff_t off)
4634 struct cgroup *src_cgrp, *dst_cgrp;
4635 struct task_struct *task;
4638 buf = strstrip(buf);
4640 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4644 task = cgroup_procs_write_start(buf, false);
4645 ret = PTR_ERR_OR_ZERO(task);
4649 /* find the source cgroup */
4650 spin_lock_irq(&css_set_lock);
4651 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4652 spin_unlock_irq(&css_set_lock);
4654 /* thread migrations follow the cgroup.procs delegation rule */
4655 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4656 of->file->f_path.dentry->d_sb);
4660 /* and must be contained in the same domain */
4662 if (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)
4665 ret = cgroup_attach_task(dst_cgrp, task, false);
4668 cgroup_procs_write_finish(task);
4670 cgroup_kn_unlock(of->kn);
4672 return ret ?: nbytes;
4675 /* cgroup core interface files for the default hierarchy */
4676 static struct cftype cgroup_base_files[] = {
4678 .name = "cgroup.type",
4679 .flags = CFTYPE_NOT_ON_ROOT,
4680 .seq_show = cgroup_type_show,
4681 .write = cgroup_type_write,
4684 .name = "cgroup.procs",
4685 .flags = CFTYPE_NS_DELEGATABLE,
4686 .file_offset = offsetof(struct cgroup, procs_file),
4687 .release = cgroup_procs_release,
4688 .seq_start = cgroup_procs_start,
4689 .seq_next = cgroup_procs_next,
4690 .seq_show = cgroup_procs_show,
4691 .write = cgroup_procs_write,
4694 .name = "cgroup.threads",
4695 .flags = CFTYPE_NS_DELEGATABLE,
4696 .release = cgroup_procs_release,
4697 .seq_start = cgroup_threads_start,
4698 .seq_next = cgroup_procs_next,
4699 .seq_show = cgroup_procs_show,
4700 .write = cgroup_threads_write,
4703 .name = "cgroup.controllers",
4704 .seq_show = cgroup_controllers_show,
4707 .name = "cgroup.subtree_control",
4708 .flags = CFTYPE_NS_DELEGATABLE,
4709 .seq_show = cgroup_subtree_control_show,
4710 .write = cgroup_subtree_control_write,
4713 .name = "cgroup.events",
4714 .flags = CFTYPE_NOT_ON_ROOT,
4715 .file_offset = offsetof(struct cgroup, events_file),
4716 .seq_show = cgroup_events_show,
4719 .name = "cgroup.max.descendants",
4720 .seq_show = cgroup_max_descendants_show,
4721 .write = cgroup_max_descendants_write,
4724 .name = "cgroup.max.depth",
4725 .seq_show = cgroup_max_depth_show,
4726 .write = cgroup_max_depth_write,
4729 .name = "cgroup.stat",
4730 .seq_show = cgroup_stat_show,
4733 .name = "cgroup.freeze",
4734 .flags = CFTYPE_NOT_ON_ROOT,
4735 .seq_show = cgroup_freeze_show,
4736 .write = cgroup_freeze_write,
4740 .flags = CFTYPE_NOT_ON_ROOT,
4741 .seq_show = cpu_stat_show,
4745 .name = "io.pressure",
4746 .flags = CFTYPE_NOT_ON_ROOT,
4747 .seq_show = cgroup_io_pressure_show,
4750 .name = "memory.pressure",
4751 .flags = CFTYPE_NOT_ON_ROOT,
4752 .seq_show = cgroup_memory_pressure_show,
4755 .name = "cpu.pressure",
4756 .flags = CFTYPE_NOT_ON_ROOT,
4757 .seq_show = cgroup_cpu_pressure_show,
4764 * css destruction is four-stage process.
4766 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4767 * Implemented in kill_css().
4769 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4770 * and thus css_tryget_online() is guaranteed to fail, the css can be
4771 * offlined by invoking offline_css(). After offlining, the base ref is
4772 * put. Implemented in css_killed_work_fn().
4774 * 3. When the percpu_ref reaches zero, the only possible remaining
4775 * accessors are inside RCU read sections. css_release() schedules the
4778 * 4. After the grace period, the css can be freed. Implemented in
4779 * css_free_work_fn().
4781 * It is actually hairier because both step 2 and 4 require process context
4782 * and thus involve punting to css->destroy_work adding two additional
4783 * steps to the already complex sequence.
4785 static void css_free_rwork_fn(struct work_struct *work)
4787 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
4788 struct cgroup_subsys_state, destroy_rwork);
4789 struct cgroup_subsys *ss = css->ss;
4790 struct cgroup *cgrp = css->cgroup;
4792 percpu_ref_exit(&css->refcnt);
4796 struct cgroup_subsys_state *parent = css->parent;
4800 cgroup_idr_remove(&ss->css_idr, id);
4806 /* cgroup free path */
4807 atomic_dec(&cgrp->root->nr_cgrps);
4808 cgroup1_pidlist_destroy_all(cgrp);
4809 cancel_work_sync(&cgrp->release_agent_work);
4811 if (cgroup_parent(cgrp)) {
4813 * We get a ref to the parent, and put the ref when
4814 * this cgroup is being freed, so it's guaranteed
4815 * that the parent won't be destroyed before its
4818 cgroup_put(cgroup_parent(cgrp));
4819 kernfs_put(cgrp->kn);
4820 psi_cgroup_free(cgrp);
4821 if (cgroup_on_dfl(cgrp))
4822 cgroup_rstat_exit(cgrp);
4826 * This is root cgroup's refcnt reaching zero,
4827 * which indicates that the root should be
4830 cgroup_destroy_root(cgrp->root);
4835 static void css_release_work_fn(struct work_struct *work)
4837 struct cgroup_subsys_state *css =
4838 container_of(work, struct cgroup_subsys_state, destroy_work);
4839 struct cgroup_subsys *ss = css->ss;
4840 struct cgroup *cgrp = css->cgroup;
4842 mutex_lock(&cgroup_mutex);
4844 css->flags |= CSS_RELEASED;
4845 list_del_rcu(&css->sibling);
4848 /* css release path */
4849 if (!list_empty(&css->rstat_css_node)) {
4850 cgroup_rstat_flush(cgrp);
4851 list_del_rcu(&css->rstat_css_node);
4854 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4855 if (ss->css_released)
4856 ss->css_released(css);
4858 struct cgroup *tcgrp;
4860 /* cgroup release path */
4861 TRACE_CGROUP_PATH(release, cgrp);
4863 if (cgroup_on_dfl(cgrp))
4864 cgroup_rstat_flush(cgrp);
4866 spin_lock_irq(&css_set_lock);
4867 for (tcgrp = cgroup_parent(cgrp); tcgrp;
4868 tcgrp = cgroup_parent(tcgrp))
4869 tcgrp->nr_dying_descendants--;
4870 spin_unlock_irq(&css_set_lock);
4872 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4876 * There are two control paths which try to determine
4877 * cgroup from dentry without going through kernfs -
4878 * cgroupstats_build() and css_tryget_online_from_dir().
4879 * Those are supported by RCU protecting clearing of
4880 * cgrp->kn->priv backpointer.
4883 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4886 cgroup_bpf_put(cgrp);
4889 mutex_unlock(&cgroup_mutex);
4891 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
4892 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
4895 static void css_release(struct percpu_ref *ref)
4897 struct cgroup_subsys_state *css =
4898 container_of(ref, struct cgroup_subsys_state, refcnt);
4900 INIT_WORK(&css->destroy_work, css_release_work_fn);
4901 queue_work(cgroup_destroy_wq, &css->destroy_work);
4904 static void init_and_link_css(struct cgroup_subsys_state *css,
4905 struct cgroup_subsys *ss, struct cgroup *cgrp)
4907 lockdep_assert_held(&cgroup_mutex);
4909 cgroup_get_live(cgrp);
4911 memset(css, 0, sizeof(*css));
4915 INIT_LIST_HEAD(&css->sibling);
4916 INIT_LIST_HEAD(&css->children);
4917 INIT_LIST_HEAD(&css->rstat_css_node);
4918 css->serial_nr = css_serial_nr_next++;
4919 atomic_set(&css->online_cnt, 0);
4921 if (cgroup_parent(cgrp)) {
4922 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4923 css_get(css->parent);
4926 if (cgroup_on_dfl(cgrp) && ss->css_rstat_flush)
4927 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
4929 BUG_ON(cgroup_css(cgrp, ss));
4932 /* invoke ->css_online() on a new CSS and mark it online if successful */
4933 static int online_css(struct cgroup_subsys_state *css)
4935 struct cgroup_subsys *ss = css->ss;
4938 lockdep_assert_held(&cgroup_mutex);
4941 ret = ss->css_online(css);
4943 css->flags |= CSS_ONLINE;
4944 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4946 atomic_inc(&css->online_cnt);
4948 atomic_inc(&css->parent->online_cnt);
4953 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4954 static void offline_css(struct cgroup_subsys_state *css)
4956 struct cgroup_subsys *ss = css->ss;
4958 lockdep_assert_held(&cgroup_mutex);
4960 if (!(css->flags & CSS_ONLINE))
4963 if (ss->css_offline)
4964 ss->css_offline(css);
4966 css->flags &= ~CSS_ONLINE;
4967 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4969 wake_up_all(&css->cgroup->offline_waitq);
4973 * css_create - create a cgroup_subsys_state
4974 * @cgrp: the cgroup new css will be associated with
4975 * @ss: the subsys of new css
4977 * Create a new css associated with @cgrp - @ss pair. On success, the new
4978 * css is online and installed in @cgrp. This function doesn't create the
4979 * interface files. Returns 0 on success, -errno on failure.
4981 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
4982 struct cgroup_subsys *ss)
4984 struct cgroup *parent = cgroup_parent(cgrp);
4985 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4986 struct cgroup_subsys_state *css;
4989 lockdep_assert_held(&cgroup_mutex);
4991 css = ss->css_alloc(parent_css);
4993 css = ERR_PTR(-ENOMEM);
4997 init_and_link_css(css, ss, cgrp);
4999 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5003 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5008 /* @css is ready to be brought online now, make it visible */
5009 list_add_tail_rcu(&css->sibling, &parent_css->children);
5010 cgroup_idr_replace(&ss->css_idr, css, css->id);
5012 err = online_css(css);
5016 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5017 cgroup_parent(parent)) {
5018 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5019 current->comm, current->pid, ss->name);
5020 if (!strcmp(ss->name, "memory"))
5021 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5022 ss->warned_broken_hierarchy = true;
5028 list_del_rcu(&css->sibling);
5030 list_del_rcu(&css->rstat_css_node);
5031 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5032 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5033 return ERR_PTR(err);
5037 * The returned cgroup is fully initialized including its control mask, but
5038 * it isn't associated with its kernfs_node and doesn't have the control
5041 static struct cgroup *cgroup_create(struct cgroup *parent)
5043 struct cgroup_root *root = parent->root;
5044 struct cgroup *cgrp, *tcgrp;
5045 int level = parent->level + 1;
5048 /* allocate the cgroup and its ID, 0 is reserved for the root */
5049 cgrp = kzalloc(struct_size(cgrp, ancestor_ids, (level + 1)),
5052 return ERR_PTR(-ENOMEM);
5054 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5058 if (cgroup_on_dfl(parent)) {
5059 ret = cgroup_rstat_init(cgrp);
5061 goto out_cancel_ref;
5065 * Temporarily set the pointer to NULL, so idr_find() won't return
5066 * a half-baked cgroup.
5068 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
5074 init_cgroup_housekeeping(cgrp);
5076 cgrp->self.parent = &parent->self;
5078 cgrp->level = level;
5080 ret = psi_cgroup_alloc(cgrp);
5084 ret = cgroup_bpf_inherit(cgrp);
5089 * New cgroup inherits effective freeze counter, and
5090 * if the parent has to be frozen, the child has too.
5092 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5093 if (cgrp->freezer.e_freeze)
5094 set_bit(CGRP_FROZEN, &cgrp->flags);
5096 spin_lock_irq(&css_set_lock);
5097 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5098 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
5100 if (tcgrp != cgrp) {
5101 tcgrp->nr_descendants++;
5104 * If the new cgroup is frozen, all ancestor cgroups
5105 * get a new frozen descendant, but their state can't
5106 * change because of this.
5108 if (cgrp->freezer.e_freeze)
5109 tcgrp->freezer.nr_frozen_descendants++;
5112 spin_unlock_irq(&css_set_lock);
5114 if (notify_on_release(parent))
5115 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5117 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5118 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5120 cgrp->self.serial_nr = css_serial_nr_next++;
5122 /* allocation complete, commit to creation */
5123 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5124 atomic_inc(&root->nr_cgrps);
5125 cgroup_get_live(parent);
5128 * @cgrp is now fully operational. If something fails after this
5129 * point, it'll be released via the normal destruction path.
5131 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5134 * On the default hierarchy, a child doesn't automatically inherit
5135 * subtree_control from the parent. Each is configured manually.
5137 if (!cgroup_on_dfl(cgrp))
5138 cgrp->subtree_control = cgroup_control(cgrp);
5140 cgroup_propagate_control(cgrp);
5145 psi_cgroup_free(cgrp);
5147 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
5149 if (cgroup_on_dfl(parent))
5150 cgroup_rstat_exit(cgrp);
5152 percpu_ref_exit(&cgrp->self.refcnt);
5155 return ERR_PTR(ret);
5158 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5160 struct cgroup *cgroup;
5164 lockdep_assert_held(&cgroup_mutex);
5166 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5167 if (cgroup->nr_descendants >= cgroup->max_descendants)
5170 if (level > cgroup->max_depth)
5181 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5183 struct cgroup *parent, *cgrp;
5184 struct kernfs_node *kn;
5187 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5188 if (strchr(name, '\n'))
5191 parent = cgroup_kn_lock_live(parent_kn, false);
5195 if (!cgroup_check_hierarchy_limits(parent)) {
5200 cgrp = cgroup_create(parent);
5202 ret = PTR_ERR(cgrp);
5206 /* create the directory */
5207 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5215 * This extra ref will be put in cgroup_free_fn() and guarantees
5216 * that @cgrp->kn is always accessible.
5220 ret = cgroup_kn_set_ugid(kn);
5224 ret = css_populate_dir(&cgrp->self);
5228 ret = cgroup_apply_control_enable(cgrp);
5232 TRACE_CGROUP_PATH(mkdir, cgrp);
5234 /* let's create and online css's */
5235 kernfs_activate(kn);
5241 cgroup_destroy_locked(cgrp);
5243 cgroup_kn_unlock(parent_kn);
5248 * This is called when the refcnt of a css is confirmed to be killed.
5249 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5250 * initate destruction and put the css ref from kill_css().
5252 static void css_killed_work_fn(struct work_struct *work)
5254 struct cgroup_subsys_state *css =
5255 container_of(work, struct cgroup_subsys_state, destroy_work);
5257 mutex_lock(&cgroup_mutex);
5262 /* @css can't go away while we're holding cgroup_mutex */
5264 } while (css && atomic_dec_and_test(&css->online_cnt));
5266 mutex_unlock(&cgroup_mutex);
5269 /* css kill confirmation processing requires process context, bounce */
5270 static void css_killed_ref_fn(struct percpu_ref *ref)
5272 struct cgroup_subsys_state *css =
5273 container_of(ref, struct cgroup_subsys_state, refcnt);
5275 if (atomic_dec_and_test(&css->online_cnt)) {
5276 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5277 queue_work(cgroup_destroy_wq, &css->destroy_work);
5282 * kill_css - destroy a css
5283 * @css: css to destroy
5285 * This function initiates destruction of @css by removing cgroup interface
5286 * files and putting its base reference. ->css_offline() will be invoked
5287 * asynchronously once css_tryget_online() is guaranteed to fail and when
5288 * the reference count reaches zero, @css will be released.
5290 static void kill_css(struct cgroup_subsys_state *css)
5292 lockdep_assert_held(&cgroup_mutex);
5294 if (css->flags & CSS_DYING)
5297 css->flags |= CSS_DYING;
5300 * This must happen before css is disassociated with its cgroup.
5301 * See seq_css() for details.
5306 * Killing would put the base ref, but we need to keep it alive
5307 * until after ->css_offline().
5312 * cgroup core guarantees that, by the time ->css_offline() is
5313 * invoked, no new css reference will be given out via
5314 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5315 * proceed to offlining css's because percpu_ref_kill() doesn't
5316 * guarantee that the ref is seen as killed on all CPUs on return.
5318 * Use percpu_ref_kill_and_confirm() to get notifications as each
5319 * css is confirmed to be seen as killed on all CPUs.
5321 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5325 * cgroup_destroy_locked - the first stage of cgroup destruction
5326 * @cgrp: cgroup to be destroyed
5328 * css's make use of percpu refcnts whose killing latency shouldn't be
5329 * exposed to userland and are RCU protected. Also, cgroup core needs to
5330 * guarantee that css_tryget_online() won't succeed by the time
5331 * ->css_offline() is invoked. To satisfy all the requirements,
5332 * destruction is implemented in the following two steps.
5334 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5335 * userland visible parts and start killing the percpu refcnts of
5336 * css's. Set up so that the next stage will be kicked off once all
5337 * the percpu refcnts are confirmed to be killed.
5339 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5340 * rest of destruction. Once all cgroup references are gone, the
5341 * cgroup is RCU-freed.
5343 * This function implements s1. After this step, @cgrp is gone as far as
5344 * the userland is concerned and a new cgroup with the same name may be
5345 * created. As cgroup doesn't care about the names internally, this
5346 * doesn't cause any problem.
5348 static int cgroup_destroy_locked(struct cgroup *cgrp)
5349 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5351 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5352 struct cgroup_subsys_state *css;
5353 struct cgrp_cset_link *link;
5356 lockdep_assert_held(&cgroup_mutex);
5359 * Only migration can raise populated from zero and we're already
5360 * holding cgroup_mutex.
5362 if (cgroup_is_populated(cgrp))
5366 * Make sure there's no live children. We can't test emptiness of
5367 * ->self.children as dead children linger on it while being
5368 * drained; otherwise, "rmdir parent/child parent" may fail.
5370 if (css_has_online_children(&cgrp->self))
5374 * Mark @cgrp and the associated csets dead. The former prevents
5375 * further task migration and child creation by disabling
5376 * cgroup_lock_live_group(). The latter makes the csets ignored by
5377 * the migration path.
5379 cgrp->self.flags &= ~CSS_ONLINE;
5381 spin_lock_irq(&css_set_lock);
5382 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5383 link->cset->dead = true;
5384 spin_unlock_irq(&css_set_lock);
5386 /* initiate massacre of all css's */
5387 for_each_css(css, ssid, cgrp)
5390 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5391 css_clear_dir(&cgrp->self);
5392 kernfs_remove(cgrp->kn);
5394 if (parent && cgroup_is_threaded(cgrp))
5395 parent->nr_threaded_children--;
5397 spin_lock_irq(&css_set_lock);
5398 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5399 tcgrp->nr_descendants--;
5400 tcgrp->nr_dying_descendants++;
5402 * If the dying cgroup is frozen, decrease frozen descendants
5403 * counters of ancestor cgroups.
5405 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5406 tcgrp->freezer.nr_frozen_descendants--;
5408 spin_unlock_irq(&css_set_lock);
5410 cgroup1_check_for_release(parent);
5412 /* put the base reference */
5413 percpu_ref_kill(&cgrp->self.refcnt);
5418 int cgroup_rmdir(struct kernfs_node *kn)
5420 struct cgroup *cgrp;
5423 cgrp = cgroup_kn_lock_live(kn, false);
5427 ret = cgroup_destroy_locked(cgrp);
5429 TRACE_CGROUP_PATH(rmdir, cgrp);
5431 cgroup_kn_unlock(kn);
5435 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5436 .show_options = cgroup_show_options,
5437 .mkdir = cgroup_mkdir,
5438 .rmdir = cgroup_rmdir,
5439 .show_path = cgroup_show_path,
5442 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5444 struct cgroup_subsys_state *css;
5446 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5448 mutex_lock(&cgroup_mutex);
5450 idr_init(&ss->css_idr);
5451 INIT_LIST_HEAD(&ss->cfts);
5453 /* Create the root cgroup state for this subsystem */
5454 ss->root = &cgrp_dfl_root;
5455 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5456 /* We don't handle early failures gracefully */
5457 BUG_ON(IS_ERR(css));
5458 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5461 * Root csses are never destroyed and we can't initialize
5462 * percpu_ref during early init. Disable refcnting.
5464 css->flags |= CSS_NO_REF;
5467 /* allocation can't be done safely during early init */
5470 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5471 BUG_ON(css->id < 0);
5474 /* Update the init_css_set to contain a subsys
5475 * pointer to this state - since the subsystem is
5476 * newly registered, all tasks and hence the
5477 * init_css_set is in the subsystem's root cgroup. */
5478 init_css_set.subsys[ss->id] = css;
5480 have_fork_callback |= (bool)ss->fork << ss->id;
5481 have_exit_callback |= (bool)ss->exit << ss->id;
5482 have_release_callback |= (bool)ss->release << ss->id;
5483 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5485 /* At system boot, before all subsystems have been
5486 * registered, no tasks have been forked, so we don't
5487 * need to invoke fork callbacks here. */
5488 BUG_ON(!list_empty(&init_task.tasks));
5490 BUG_ON(online_css(css));
5492 mutex_unlock(&cgroup_mutex);
5496 * cgroup_init_early - cgroup initialization at system boot
5498 * Initialize cgroups at system boot, and initialize any
5499 * subsystems that request early init.
5501 int __init cgroup_init_early(void)
5503 static struct cgroup_fs_context __initdata ctx;
5504 struct cgroup_subsys *ss;
5507 ctx.root = &cgrp_dfl_root;
5508 init_cgroup_root(&ctx);
5509 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5511 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5513 for_each_subsys(ss, i) {
5514 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5515 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5516 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5518 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5519 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5522 ss->name = cgroup_subsys_name[i];
5523 if (!ss->legacy_name)
5524 ss->legacy_name = cgroup_subsys_name[i];
5527 cgroup_init_subsys(ss, true);
5532 static u16 cgroup_disable_mask __initdata;
5535 * cgroup_init - cgroup initialization
5537 * Register cgroup filesystem and /proc file, and initialize
5538 * any subsystems that didn't request early init.
5540 int __init cgroup_init(void)
5542 struct cgroup_subsys *ss;
5545 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5546 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5547 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5548 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5550 cgroup_rstat_boot();
5553 * The latency of the synchronize_rcu() is too high for cgroups,
5554 * avoid it at the cost of forcing all readers into the slow path.
5556 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5558 get_user_ns(init_cgroup_ns.user_ns);
5560 mutex_lock(&cgroup_mutex);
5563 * Add init_css_set to the hash table so that dfl_root can link to
5566 hash_add(css_set_table, &init_css_set.hlist,
5567 css_set_hash(init_css_set.subsys));
5569 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5571 mutex_unlock(&cgroup_mutex);
5573 for_each_subsys(ss, ssid) {
5574 if (ss->early_init) {
5575 struct cgroup_subsys_state *css =
5576 init_css_set.subsys[ss->id];
5578 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5580 BUG_ON(css->id < 0);
5582 cgroup_init_subsys(ss, false);
5585 list_add_tail(&init_css_set.e_cset_node[ssid],
5586 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5589 * Setting dfl_root subsys_mask needs to consider the
5590 * disabled flag and cftype registration needs kmalloc,
5591 * both of which aren't available during early_init.
5593 if (cgroup_disable_mask & (1 << ssid)) {
5594 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5595 printk(KERN_INFO "Disabling %s control group subsystem\n",
5600 if (cgroup1_ssid_disabled(ssid))
5601 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5604 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5606 /* implicit controllers must be threaded too */
5607 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5609 if (ss->implicit_on_dfl)
5610 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5611 else if (!ss->dfl_cftypes)
5612 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5615 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5617 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5618 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5620 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5621 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5625 ss->bind(init_css_set.subsys[ssid]);
5627 mutex_lock(&cgroup_mutex);
5628 css_populate_dir(init_css_set.subsys[ssid]);
5629 mutex_unlock(&cgroup_mutex);
5632 /* init_css_set.subsys[] has been updated, re-hash */
5633 hash_del(&init_css_set.hlist);
5634 hash_add(css_set_table, &init_css_set.hlist,
5635 css_set_hash(init_css_set.subsys));
5637 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5638 WARN_ON(register_filesystem(&cgroup_fs_type));
5639 WARN_ON(register_filesystem(&cgroup2_fs_type));
5640 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
5645 static int __init cgroup_wq_init(void)
5648 * There isn't much point in executing destruction path in
5649 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5650 * Use 1 for @max_active.
5652 * We would prefer to do this in cgroup_init() above, but that
5653 * is called before init_workqueues(): so leave this until after.
5655 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5656 BUG_ON(!cgroup_destroy_wq);
5659 core_initcall(cgroup_wq_init);
5661 void cgroup_path_from_kernfs_id(const union kernfs_node_id *id,
5662 char *buf, size_t buflen)
5664 struct kernfs_node *kn;
5666 kn = kernfs_get_node_by_id(cgrp_dfl_root.kf_root, id);
5669 kernfs_path(kn, buf, buflen);
5674 * proc_cgroup_show()
5675 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5676 * - Used for /proc/<pid>/cgroup.
5678 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5679 struct pid *pid, struct task_struct *tsk)
5683 struct cgroup_root *root;
5686 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5690 mutex_lock(&cgroup_mutex);
5691 spin_lock_irq(&css_set_lock);
5693 for_each_root(root) {
5694 struct cgroup_subsys *ss;
5695 struct cgroup *cgrp;
5696 int ssid, count = 0;
5698 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5701 seq_printf(m, "%d:", root->hierarchy_id);
5702 if (root != &cgrp_dfl_root)
5703 for_each_subsys(ss, ssid)
5704 if (root->subsys_mask & (1 << ssid))
5705 seq_printf(m, "%s%s", count++ ? "," : "",
5707 if (strlen(root->name))
5708 seq_printf(m, "%sname=%s", count ? "," : "",
5712 cgrp = task_cgroup_from_root(tsk, root);
5715 * On traditional hierarchies, all zombie tasks show up as
5716 * belonging to the root cgroup. On the default hierarchy,
5717 * while a zombie doesn't show up in "cgroup.procs" and
5718 * thus can't be migrated, its /proc/PID/cgroup keeps
5719 * reporting the cgroup it belonged to before exiting. If
5720 * the cgroup is removed before the zombie is reaped,
5721 * " (deleted)" is appended to the cgroup path.
5723 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5724 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5725 current->nsproxy->cgroup_ns);
5726 if (retval >= PATH_MAX)
5727 retval = -ENAMETOOLONG;
5736 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5737 seq_puts(m, " (deleted)\n");
5744 spin_unlock_irq(&css_set_lock);
5745 mutex_unlock(&cgroup_mutex);
5752 * cgroup_fork - initialize cgroup related fields during copy_process()
5753 * @child: pointer to task_struct of forking parent process.
5755 * A task is associated with the init_css_set until cgroup_post_fork()
5756 * attaches it to the parent's css_set. Empty cg_list indicates that
5757 * @child isn't holding reference to its css_set.
5759 void cgroup_fork(struct task_struct *child)
5761 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5762 INIT_LIST_HEAD(&child->cg_list);
5766 * cgroup_can_fork - called on a new task before the process is exposed
5767 * @child: the task in question.
5769 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5770 * returns an error, the fork aborts with that error code. This allows for
5771 * a cgroup subsystem to conditionally allow or deny new forks.
5773 int cgroup_can_fork(struct task_struct *child)
5775 struct cgroup_subsys *ss;
5778 do_each_subsys_mask(ss, i, have_canfork_callback) {
5779 ret = ss->can_fork(child);
5782 } while_each_subsys_mask();
5787 for_each_subsys(ss, j) {
5790 if (ss->cancel_fork)
5791 ss->cancel_fork(child);
5798 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5799 * @child: the task in question
5801 * This calls the cancel_fork() callbacks if a fork failed *after*
5802 * cgroup_can_fork() succeded.
5804 void cgroup_cancel_fork(struct task_struct *child)
5806 struct cgroup_subsys *ss;
5809 for_each_subsys(ss, i)
5810 if (ss->cancel_fork)
5811 ss->cancel_fork(child);
5815 * cgroup_post_fork - called on a new task after adding it to the task list
5816 * @child: the task in question
5818 * Adds the task to the list running through its css_set if necessary and
5819 * call the subsystem fork() callbacks. Has to be after the task is
5820 * visible on the task list in case we race with the first call to
5821 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5824 void cgroup_post_fork(struct task_struct *child)
5826 struct cgroup_subsys *ss;
5830 * This may race against cgroup_enable_task_cg_lists(). As that
5831 * function sets use_task_css_set_links before grabbing
5832 * tasklist_lock and we just went through tasklist_lock to add
5833 * @child, it's guaranteed that either we see the set
5834 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5835 * @child during its iteration.
5837 * If we won the race, @child is associated with %current's
5838 * css_set. Grabbing css_set_lock guarantees both that the
5839 * association is stable, and, on completion of the parent's
5840 * migration, @child is visible in the source of migration or
5841 * already in the destination cgroup. This guarantee is necessary
5842 * when implementing operations which need to migrate all tasks of
5843 * a cgroup to another.
5845 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5846 * will remain in init_css_set. This is safe because all tasks are
5847 * in the init_css_set before cg_links is enabled and there's no
5848 * operation which transfers all tasks out of init_css_set.
5850 if (use_task_css_set_links) {
5851 struct css_set *cset;
5853 spin_lock_irq(&css_set_lock);
5854 cset = task_css_set(current);
5855 if (list_empty(&child->cg_list)) {
5858 css_set_move_task(child, NULL, cset, false);
5862 * If the cgroup has to be frozen, the new task has too.
5863 * Let's set the JOBCTL_TRAP_FREEZE jobctl bit to get
5864 * the task into the frozen state.
5866 if (unlikely(cgroup_task_freeze(child))) {
5867 spin_lock(&child->sighand->siglock);
5868 WARN_ON_ONCE(child->frozen);
5869 child->jobctl |= JOBCTL_TRAP_FREEZE;
5870 spin_unlock(&child->sighand->siglock);
5873 * Calling cgroup_update_frozen() isn't required here,
5874 * because it will be called anyway a bit later
5875 * from do_freezer_trap(). So we avoid cgroup's
5876 * transient switch from the frozen state and back.
5880 spin_unlock_irq(&css_set_lock);
5884 * Call ss->fork(). This must happen after @child is linked on
5885 * css_set; otherwise, @child might change state between ->fork()
5886 * and addition to css_set.
5888 do_each_subsys_mask(ss, i, have_fork_callback) {
5890 } while_each_subsys_mask();
5894 * cgroup_exit - detach cgroup from exiting task
5895 * @tsk: pointer to task_struct of exiting process
5897 * Description: Detach cgroup from @tsk and release it.
5899 * Note that cgroups marked notify_on_release force every task in
5900 * them to take the global cgroup_mutex mutex when exiting.
5901 * This could impact scaling on very large systems. Be reluctant to
5902 * use notify_on_release cgroups where very high task exit scaling
5903 * is required on large systems.
5905 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5906 * call cgroup_exit() while the task is still competent to handle
5907 * notify_on_release(), then leave the task attached to the root cgroup in
5908 * each hierarchy for the remainder of its exit. No need to bother with
5909 * init_css_set refcnting. init_css_set never goes away and we can't race
5910 * with migration path - PF_EXITING is visible to migration path.
5912 void cgroup_exit(struct task_struct *tsk)
5914 struct cgroup_subsys *ss;
5915 struct css_set *cset;
5919 * Unlink from @tsk from its css_set. As migration path can't race
5920 * with us, we can check css_set and cg_list without synchronization.
5922 cset = task_css_set(tsk);
5924 if (!list_empty(&tsk->cg_list)) {
5925 spin_lock_irq(&css_set_lock);
5926 css_set_move_task(tsk, cset, NULL, false);
5929 WARN_ON_ONCE(cgroup_task_frozen(tsk));
5930 if (unlikely(cgroup_task_freeze(tsk)))
5931 cgroup_update_frozen(task_dfl_cgroup(tsk));
5933 spin_unlock_irq(&css_set_lock);
5938 /* see cgroup_post_fork() for details */
5939 do_each_subsys_mask(ss, i, have_exit_callback) {
5941 } while_each_subsys_mask();
5944 void cgroup_release(struct task_struct *task)
5946 struct cgroup_subsys *ss;
5949 do_each_subsys_mask(ss, ssid, have_release_callback) {
5951 } while_each_subsys_mask();
5954 void cgroup_free(struct task_struct *task)
5956 struct css_set *cset = task_css_set(task);
5960 static int __init cgroup_disable(char *str)
5962 struct cgroup_subsys *ss;
5966 while ((token = strsep(&str, ",")) != NULL) {
5970 for_each_subsys(ss, i) {
5971 if (strcmp(token, ss->name) &&
5972 strcmp(token, ss->legacy_name))
5974 cgroup_disable_mask |= 1 << i;
5979 __setup("cgroup_disable=", cgroup_disable);
5981 void __init __weak enable_debug_cgroup(void) { }
5983 static int __init enable_cgroup_debug(char *str)
5985 cgroup_debug = true;
5986 enable_debug_cgroup();
5989 __setup("cgroup_debug", enable_cgroup_debug);
5992 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5993 * @dentry: directory dentry of interest
5994 * @ss: subsystem of interest
5996 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5997 * to get the corresponding css and return it. If such css doesn't exist
5998 * or can't be pinned, an ERR_PTR value is returned.
6000 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6001 struct cgroup_subsys *ss)
6003 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6004 struct file_system_type *s_type = dentry->d_sb->s_type;
6005 struct cgroup_subsys_state *css = NULL;
6006 struct cgroup *cgrp;
6008 /* is @dentry a cgroup dir? */
6009 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6010 !kn || kernfs_type(kn) != KERNFS_DIR)
6011 return ERR_PTR(-EBADF);
6016 * This path doesn't originate from kernfs and @kn could already
6017 * have been or be removed at any point. @kn->priv is RCU
6018 * protected for this access. See css_release_work_fn() for details.
6020 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6022 css = cgroup_css(cgrp, ss);
6024 if (!css || !css_tryget_online(css))
6025 css = ERR_PTR(-ENOENT);
6032 * css_from_id - lookup css by id
6033 * @id: the cgroup id
6034 * @ss: cgroup subsys to be looked into
6036 * Returns the css if there's valid one with @id, otherwise returns NULL.
6037 * Should be called under rcu_read_lock().
6039 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6041 WARN_ON_ONCE(!rcu_read_lock_held());
6042 return idr_find(&ss->css_idr, id);
6046 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6047 * @path: path on the default hierarchy
6049 * Find the cgroup at @path on the default hierarchy, increment its
6050 * reference count and return it. Returns pointer to the found cgroup on
6051 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6052 * if @path points to a non-directory.
6054 struct cgroup *cgroup_get_from_path(const char *path)
6056 struct kernfs_node *kn;
6057 struct cgroup *cgrp;
6059 mutex_lock(&cgroup_mutex);
6061 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6063 if (kernfs_type(kn) == KERNFS_DIR) {
6065 cgroup_get_live(cgrp);
6067 cgrp = ERR_PTR(-ENOTDIR);
6071 cgrp = ERR_PTR(-ENOENT);
6074 mutex_unlock(&cgroup_mutex);
6077 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6080 * cgroup_get_from_fd - get a cgroup pointer from a fd
6081 * @fd: fd obtained by open(cgroup2_dir)
6083 * Find the cgroup from a fd which should be obtained
6084 * by opening a cgroup directory. Returns a pointer to the
6085 * cgroup on success. ERR_PTR is returned if the cgroup
6088 struct cgroup *cgroup_get_from_fd(int fd)
6090 struct cgroup_subsys_state *css;
6091 struct cgroup *cgrp;
6096 return ERR_PTR(-EBADF);
6098 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6101 return ERR_CAST(css);
6104 if (!cgroup_on_dfl(cgrp)) {
6106 return ERR_PTR(-EBADF);
6111 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6114 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6115 * definition in cgroup-defs.h.
6117 #ifdef CONFIG_SOCK_CGROUP_DATA
6119 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6121 DEFINE_SPINLOCK(cgroup_sk_update_lock);
6122 static bool cgroup_sk_alloc_disabled __read_mostly;
6124 void cgroup_sk_alloc_disable(void)
6126 if (cgroup_sk_alloc_disabled)
6128 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6129 cgroup_sk_alloc_disabled = true;
6134 #define cgroup_sk_alloc_disabled false
6138 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6140 if (cgroup_sk_alloc_disabled)
6143 /* Socket clone path */
6146 * We might be cloning a socket which is left in an empty
6147 * cgroup and the cgroup might have already been rmdir'd.
6148 * Don't use cgroup_get_live().
6150 cgroup_get(sock_cgroup_ptr(skcd));
6157 struct css_set *cset;
6159 cset = task_css_set(current);
6160 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6161 skcd->val = (unsigned long)cset->dfl_cgrp;
6170 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6172 cgroup_put(sock_cgroup_ptr(skcd));
6175 #endif /* CONFIG_SOCK_CGROUP_DATA */
6177 #ifdef CONFIG_CGROUP_BPF
6178 int cgroup_bpf_attach(struct cgroup *cgrp, struct bpf_prog *prog,
6179 enum bpf_attach_type type, u32 flags)
6183 mutex_lock(&cgroup_mutex);
6184 ret = __cgroup_bpf_attach(cgrp, prog, type, flags);
6185 mutex_unlock(&cgroup_mutex);
6188 int cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
6189 enum bpf_attach_type type, u32 flags)
6193 mutex_lock(&cgroup_mutex);
6194 ret = __cgroup_bpf_detach(cgrp, prog, type);
6195 mutex_unlock(&cgroup_mutex);
6198 int cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
6199 union bpf_attr __user *uattr)
6203 mutex_lock(&cgroup_mutex);
6204 ret = __cgroup_bpf_query(cgrp, attr, uattr);
6205 mutex_unlock(&cgroup_mutex);
6208 #endif /* CONFIG_CGROUP_BPF */
6211 static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6212 ssize_t size, const char *prefix)
6217 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6218 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
6222 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
6224 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
6226 if (WARN_ON(ret >= size))
6233 static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
6236 struct cgroup_subsys *ss;
6240 ret = show_delegatable_files(cgroup_base_files, buf, PAGE_SIZE - ret,
6243 for_each_subsys(ss, ssid)
6244 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
6246 cgroup_subsys_name[ssid]);
6250 static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
6252 static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
6255 return snprintf(buf, PAGE_SIZE, "nsdelegate\n");
6257 static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
6259 static struct attribute *cgroup_sysfs_attrs[] = {
6260 &cgroup_delegate_attr.attr,
6261 &cgroup_features_attr.attr,
6265 static const struct attribute_group cgroup_sysfs_attr_group = {
6266 .attrs = cgroup_sysfs_attrs,
6270 static int __init cgroup_sysfs_init(void)
6272 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
6274 subsys_initcall(cgroup_sysfs_init);
6275 #endif /* CONFIG_SYSFS */