2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include "cgroup-internal.h"
33 #include <linux/cred.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/magic.h>
38 #include <linux/mutex.h>
39 #include <linux/mount.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/sched/task.h>
45 #include <linux/slab.h>
46 #include <linux/spinlock.h>
47 #include <linux/percpu-rwsem.h>
48 #include <linux/string.h>
49 #include <linux/hashtable.h>
50 #include <linux/idr.h>
51 #include <linux/kthread.h>
52 #include <linux/atomic.h>
53 #include <linux/cpuset.h>
54 #include <linux/proc_ns.h>
55 #include <linux/nsproxy.h>
56 #include <linux/file.h>
57 #include <linux/fs_parser.h>
58 #include <linux/sched/cputime.h>
59 #include <linux/psi.h>
62 #define CREATE_TRACE_POINTS
63 #include <trace/events/cgroup.h>
65 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
67 /* let's not notify more than 100 times per second */
68 #define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
71 * cgroup_mutex is the master lock. Any modification to cgroup or its
72 * hierarchy must be performed while holding it.
74 * css_set_lock protects task->cgroups pointer, the list of css_set
75 * objects, and the chain of tasks off each css_set.
77 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
78 * cgroup.h can use them for lockdep annotations.
80 DEFINE_MUTEX(cgroup_mutex);
81 DEFINE_SPINLOCK(css_set_lock);
83 #ifdef CONFIG_PROVE_RCU
84 EXPORT_SYMBOL_GPL(cgroup_mutex);
85 EXPORT_SYMBOL_GPL(css_set_lock);
88 DEFINE_SPINLOCK(trace_cgroup_path_lock);
89 char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
90 bool cgroup_debug __read_mostly;
93 * Protects cgroup_idr and css_idr so that IDs can be released without
94 * grabbing cgroup_mutex.
96 static DEFINE_SPINLOCK(cgroup_idr_lock);
99 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
100 * against file removal/re-creation across css hiding.
102 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
104 DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
106 #define cgroup_assert_mutex_or_rcu_locked() \
107 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
108 !lockdep_is_held(&cgroup_mutex), \
109 "cgroup_mutex or RCU read lock required");
112 * cgroup destruction makes heavy use of work items and there can be a lot
113 * of concurrent destructions. Use a separate workqueue so that cgroup
114 * destruction work items don't end up filling up max_active of system_wq
115 * which may lead to deadlock.
117 static struct workqueue_struct *cgroup_destroy_wq;
119 /* generate an array of cgroup subsystem pointers */
120 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
121 struct cgroup_subsys *cgroup_subsys[] = {
122 #include <linux/cgroup_subsys.h>
126 /* array of cgroup subsystem names */
127 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
128 static const char *cgroup_subsys_name[] = {
129 #include <linux/cgroup_subsys.h>
133 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
135 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
136 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
137 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
138 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
139 #include <linux/cgroup_subsys.h>
142 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
143 static struct static_key_true *cgroup_subsys_enabled_key[] = {
144 #include <linux/cgroup_subsys.h>
148 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
149 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
150 #include <linux/cgroup_subsys.h>
154 static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu);
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_skip(struct css_task_iter *it,
219 struct task_struct *task);
220 static int cgroup_destroy_locked(struct cgroup *cgrp);
221 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
222 struct cgroup_subsys *ss);
223 static void css_release(struct percpu_ref *ref);
224 static void kill_css(struct cgroup_subsys_state *css);
225 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
226 struct cgroup *cgrp, struct cftype cfts[],
230 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
231 * @ssid: subsys ID of interest
233 * cgroup_subsys_enabled() can only be used with literal subsys names which
234 * is fine for individual subsystems but unsuitable for cgroup core. This
235 * is slower static_key_enabled() based test indexed by @ssid.
237 bool cgroup_ssid_enabled(int ssid)
239 if (CGROUP_SUBSYS_COUNT == 0)
242 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
246 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
247 * @cgrp: the cgroup of interest
249 * The default hierarchy is the v2 interface of cgroup and this function
250 * can be used to test whether a cgroup is on the default hierarchy for
251 * cases where a subsystem should behave differnetly depending on the
254 * The set of behaviors which change on the default hierarchy are still
255 * being determined and the mount option is prefixed with __DEVEL__.
257 * List of changed behaviors:
259 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
260 * and "name" are disallowed.
262 * - When mounting an existing superblock, mount options should match.
264 * - Remount is disallowed.
266 * - rename(2) is disallowed.
268 * - "tasks" is removed. Everything should be at process granularity. Use
269 * "cgroup.procs" instead.
271 * - "cgroup.procs" is not sorted. pids will be unique unless they got
272 * recycled inbetween reads.
274 * - "release_agent" and "notify_on_release" are removed. Replacement
275 * notification mechanism will be implemented.
277 * - "cgroup.clone_children" is removed.
279 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
280 * and its descendants contain no task; otherwise, 1. The file also
281 * generates kernfs notification which can be monitored through poll and
282 * [di]notify when the value of the file changes.
284 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
285 * take masks of ancestors with non-empty cpus/mems, instead of being
286 * moved to an ancestor.
288 * - cpuset: a task can be moved into an empty cpuset, and again it takes
289 * masks of ancestors.
291 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
294 * - blkcg: blk-throttle becomes properly hierarchical.
296 * - debug: disallowed on the default hierarchy.
298 bool cgroup_on_dfl(const struct cgroup *cgrp)
300 return cgrp->root == &cgrp_dfl_root;
303 /* IDR wrappers which synchronize using cgroup_idr_lock */
304 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
309 idr_preload(gfp_mask);
310 spin_lock_bh(&cgroup_idr_lock);
311 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
312 spin_unlock_bh(&cgroup_idr_lock);
317 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
321 spin_lock_bh(&cgroup_idr_lock);
322 ret = idr_replace(idr, ptr, id);
323 spin_unlock_bh(&cgroup_idr_lock);
327 static void cgroup_idr_remove(struct idr *idr, int id)
329 spin_lock_bh(&cgroup_idr_lock);
331 spin_unlock_bh(&cgroup_idr_lock);
334 static bool cgroup_has_tasks(struct cgroup *cgrp)
336 return cgrp->nr_populated_csets;
339 bool cgroup_is_threaded(struct cgroup *cgrp)
341 return cgrp->dom_cgrp != cgrp;
344 /* can @cgrp host both domain and threaded children? */
345 static bool cgroup_is_mixable(struct cgroup *cgrp)
348 * Root isn't under domain level resource control exempting it from
349 * the no-internal-process constraint, so it can serve as a thread
350 * root and a parent of resource domains at the same time.
352 return !cgroup_parent(cgrp);
355 /* can @cgrp become a thread root? should always be true for a thread root */
356 static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
358 /* mixables don't care */
359 if (cgroup_is_mixable(cgrp))
362 /* domain roots can't be nested under threaded */
363 if (cgroup_is_threaded(cgrp))
366 /* can only have either domain or threaded children */
367 if (cgrp->nr_populated_domain_children)
370 /* and no domain controllers can be enabled */
371 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
377 /* is @cgrp root of a threaded subtree? */
378 bool cgroup_is_thread_root(struct cgroup *cgrp)
380 /* thread root should be a domain */
381 if (cgroup_is_threaded(cgrp))
384 /* a domain w/ threaded children is a thread root */
385 if (cgrp->nr_threaded_children)
389 * A domain which has tasks and explicit threaded controllers
390 * enabled is a thread root.
392 if (cgroup_has_tasks(cgrp) &&
393 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
399 /* a domain which isn't connected to the root w/o brekage can't be used */
400 static bool cgroup_is_valid_domain(struct cgroup *cgrp)
402 /* the cgroup itself can be a thread root */
403 if (cgroup_is_threaded(cgrp))
406 /* but the ancestors can't be unless mixable */
407 while ((cgrp = cgroup_parent(cgrp))) {
408 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
410 if (cgroup_is_threaded(cgrp))
417 /* subsystems visibly enabled on a cgroup */
418 static u16 cgroup_control(struct cgroup *cgrp)
420 struct cgroup *parent = cgroup_parent(cgrp);
421 u16 root_ss_mask = cgrp->root->subsys_mask;
424 u16 ss_mask = parent->subtree_control;
426 /* threaded cgroups can only have threaded controllers */
427 if (cgroup_is_threaded(cgrp))
428 ss_mask &= cgrp_dfl_threaded_ss_mask;
432 if (cgroup_on_dfl(cgrp))
433 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
434 cgrp_dfl_implicit_ss_mask);
438 /* subsystems enabled on a cgroup */
439 static u16 cgroup_ss_mask(struct cgroup *cgrp)
441 struct cgroup *parent = cgroup_parent(cgrp);
444 u16 ss_mask = parent->subtree_ss_mask;
446 /* threaded cgroups can only have threaded controllers */
447 if (cgroup_is_threaded(cgrp))
448 ss_mask &= cgrp_dfl_threaded_ss_mask;
452 return cgrp->root->subsys_mask;
456 * cgroup_css - obtain a cgroup's css for the specified subsystem
457 * @cgrp: the cgroup of interest
458 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
460 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
461 * function must be called either under cgroup_mutex or rcu_read_lock() and
462 * the caller is responsible for pinning the returned css if it wants to
463 * keep accessing it outside the said locks. This function may return
464 * %NULL if @cgrp doesn't have @subsys_id enabled.
466 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
467 struct cgroup_subsys *ss)
470 return rcu_dereference_check(cgrp->subsys[ss->id],
471 lockdep_is_held(&cgroup_mutex));
477 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
478 * @cgrp: the cgroup of interest
479 * @ss: the subsystem of interest
481 * Find and get @cgrp's css assocaited with @ss. If the css doesn't exist
482 * or is offline, %NULL is returned.
484 static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
485 struct cgroup_subsys *ss)
487 struct cgroup_subsys_state *css;
490 css = cgroup_css(cgrp, ss);
491 if (css && !css_tryget_online(css))
499 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
500 * @cgrp: the cgroup of interest
501 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
503 * Similar to cgroup_css() but returns the effective css, which is defined
504 * as the matching css of the nearest ancestor including self which has @ss
505 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
506 * function is guaranteed to return non-NULL css.
508 static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
509 struct cgroup_subsys *ss)
511 lockdep_assert_held(&cgroup_mutex);
517 * This function is used while updating css associations and thus
518 * can't test the csses directly. Test ss_mask.
520 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
521 cgrp = cgroup_parent(cgrp);
526 return cgroup_css(cgrp, ss);
530 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
531 * @cgrp: the cgroup of interest
532 * @ss: the subsystem of interest
534 * Find and get the effective css of @cgrp for @ss. The effective css is
535 * defined as the matching css of the nearest ancestor including self which
536 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
537 * the root css is returned, so this function always returns a valid css.
539 * The returned css is not guaranteed to be online, and therefore it is the
540 * callers responsiblity to tryget a reference for it.
542 struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
543 struct cgroup_subsys *ss)
545 struct cgroup_subsys_state *css;
548 css = cgroup_css(cgrp, ss);
552 cgrp = cgroup_parent(cgrp);
555 return init_css_set.subsys[ss->id];
559 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
560 * @cgrp: the cgroup of interest
561 * @ss: the subsystem of interest
563 * Find and get the effective css of @cgrp for @ss. The effective css is
564 * defined as the matching css of the nearest ancestor including self which
565 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
566 * the root css is returned, so this function always returns a valid css.
567 * The returned css must be put using css_put().
569 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
570 struct cgroup_subsys *ss)
572 struct cgroup_subsys_state *css;
577 css = cgroup_css(cgrp, ss);
579 if (css && css_tryget_online(css))
581 cgrp = cgroup_parent(cgrp);
584 css = init_css_set.subsys[ss->id];
591 static void cgroup_get_live(struct cgroup *cgrp)
593 WARN_ON_ONCE(cgroup_is_dead(cgrp));
594 css_get(&cgrp->self);
598 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
599 * is responsible for taking the css_set_lock.
600 * @cgrp: the cgroup in question
602 int __cgroup_task_count(const struct cgroup *cgrp)
605 struct cgrp_cset_link *link;
607 lockdep_assert_held(&css_set_lock);
609 list_for_each_entry(link, &cgrp->cset_links, cset_link)
610 count += link->cset->nr_tasks;
616 * cgroup_task_count - count the number of tasks in a cgroup.
617 * @cgrp: the cgroup in question
619 int cgroup_task_count(const struct cgroup *cgrp)
623 spin_lock_irq(&css_set_lock);
624 count = __cgroup_task_count(cgrp);
625 spin_unlock_irq(&css_set_lock);
630 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
632 struct cgroup *cgrp = of->kn->parent->priv;
633 struct cftype *cft = of_cft(of);
636 * This is open and unprotected implementation of cgroup_css().
637 * seq_css() is only called from a kernfs file operation which has
638 * an active reference on the file. Because all the subsystem
639 * files are drained before a css is disassociated with a cgroup,
640 * the matching css from the cgroup's subsys table is guaranteed to
641 * be and stay valid until the enclosing operation is complete.
644 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
648 EXPORT_SYMBOL_GPL(of_css);
651 * for_each_css - iterate all css's of a cgroup
652 * @css: the iteration cursor
653 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
654 * @cgrp: the target cgroup to iterate css's of
656 * Should be called under cgroup_[tree_]mutex.
658 #define for_each_css(css, ssid, cgrp) \
659 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
660 if (!((css) = rcu_dereference_check( \
661 (cgrp)->subsys[(ssid)], \
662 lockdep_is_held(&cgroup_mutex)))) { } \
666 * for_each_e_css - iterate all effective css's of a cgroup
667 * @css: the iteration cursor
668 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
669 * @cgrp: the target cgroup to iterate css's of
671 * Should be called under cgroup_[tree_]mutex.
673 #define for_each_e_css(css, ssid, cgrp) \
674 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
675 if (!((css) = cgroup_e_css_by_mask(cgrp, \
676 cgroup_subsys[(ssid)]))) \
681 * do_each_subsys_mask - filter for_each_subsys with a bitmask
682 * @ss: the iteration cursor
683 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
684 * @ss_mask: the bitmask
686 * The block will only run for cases where the ssid-th bit (1 << ssid) of
689 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
690 unsigned long __ss_mask = (ss_mask); \
691 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
695 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
696 (ss) = cgroup_subsys[ssid]; \
699 #define while_each_subsys_mask() \
704 /* iterate over child cgrps, lock should be held throughout iteration */
705 #define cgroup_for_each_live_child(child, cgrp) \
706 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
707 if (({ lockdep_assert_held(&cgroup_mutex); \
708 cgroup_is_dead(child); })) \
712 /* walk live descendants in preorder */
713 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
714 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
715 if (({ lockdep_assert_held(&cgroup_mutex); \
716 (dsct) = (d_css)->cgroup; \
717 cgroup_is_dead(dsct); })) \
721 /* walk live descendants in postorder */
722 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
723 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
724 if (({ lockdep_assert_held(&cgroup_mutex); \
725 (dsct) = (d_css)->cgroup; \
726 cgroup_is_dead(dsct); })) \
731 * The default css_set - used by init and its children prior to any
732 * hierarchies being mounted. It contains a pointer to the root state
733 * for each subsystem. Also used to anchor the list of css_sets. Not
734 * reference-counted, to improve performance when child cgroups
735 * haven't been created.
737 struct css_set init_css_set = {
738 .refcount = REFCOUNT_INIT(1),
739 .dom_cset = &init_css_set,
740 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
741 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
742 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
743 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
744 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
745 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
746 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
747 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
750 * The following field is re-initialized when this cset gets linked
751 * in cgroup_init(). However, let's initialize the field
752 * statically too so that the default cgroup can be accessed safely
755 .dfl_cgrp = &cgrp_dfl_root.cgrp,
758 static int css_set_count = 1; /* 1 for init_css_set */
760 static bool css_set_threaded(struct css_set *cset)
762 return cset->dom_cset != cset;
766 * css_set_populated - does a css_set contain any tasks?
767 * @cset: target css_set
769 * css_set_populated() should be the same as !!cset->nr_tasks at steady
770 * state. However, css_set_populated() can be called while a task is being
771 * added to or removed from the linked list before the nr_tasks is
772 * properly updated. Hence, we can't just look at ->nr_tasks here.
774 static bool css_set_populated(struct css_set *cset)
776 lockdep_assert_held(&css_set_lock);
778 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
782 * cgroup_update_populated - update the populated count of a cgroup
783 * @cgrp: the target cgroup
784 * @populated: inc or dec populated count
786 * One of the css_sets associated with @cgrp is either getting its first
787 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
788 * count is propagated towards root so that a given cgroup's
789 * nr_populated_children is zero iff none of its descendants contain any
792 * @cgrp's interface file "cgroup.populated" is zero if both
793 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
794 * 1 otherwise. When the sum changes from or to zero, userland is notified
795 * that the content of the interface file has changed. This can be used to
796 * detect when @cgrp and its descendants become populated or empty.
798 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
800 struct cgroup *child = NULL;
801 int adj = populated ? 1 : -1;
803 lockdep_assert_held(&css_set_lock);
806 bool was_populated = cgroup_is_populated(cgrp);
809 cgrp->nr_populated_csets += adj;
811 if (cgroup_is_threaded(child))
812 cgrp->nr_populated_threaded_children += adj;
814 cgrp->nr_populated_domain_children += adj;
817 if (was_populated == cgroup_is_populated(cgrp))
820 cgroup1_check_for_release(cgrp);
821 TRACE_CGROUP_PATH(notify_populated, cgrp,
822 cgroup_is_populated(cgrp));
823 cgroup_file_notify(&cgrp->events_file);
826 cgrp = cgroup_parent(cgrp);
831 * css_set_update_populated - update populated state of a css_set
832 * @cset: target css_set
833 * @populated: whether @cset is populated or depopulated
835 * @cset is either getting the first task or losing the last. Update the
836 * populated counters of all associated cgroups accordingly.
838 static void css_set_update_populated(struct css_set *cset, bool populated)
840 struct cgrp_cset_link *link;
842 lockdep_assert_held(&css_set_lock);
844 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
845 cgroup_update_populated(link->cgrp, populated);
849 * @task is leaving, advance task iterators which are pointing to it so
850 * that they can resume at the next position. Advancing an iterator might
851 * remove it from the list, use safe walk. See css_task_iter_skip() for
854 static void css_set_skip_task_iters(struct css_set *cset,
855 struct task_struct *task)
857 struct css_task_iter *it, *pos;
859 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
860 css_task_iter_skip(it, task);
864 * css_set_move_task - move a task from one css_set to another
865 * @task: task being moved
866 * @from_cset: css_set @task currently belongs to (may be NULL)
867 * @to_cset: new css_set @task is being moved to (may be NULL)
868 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
870 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
871 * css_set, @from_cset can be NULL. If @task is being disassociated
872 * instead of moved, @to_cset can be NULL.
874 * This function automatically handles populated counter updates and
875 * css_task_iter adjustments but the caller is responsible for managing
876 * @from_cset and @to_cset's reference counts.
878 static void css_set_move_task(struct task_struct *task,
879 struct css_set *from_cset, struct css_set *to_cset,
882 lockdep_assert_held(&css_set_lock);
884 if (to_cset && !css_set_populated(to_cset))
885 css_set_update_populated(to_cset, true);
888 WARN_ON_ONCE(list_empty(&task->cg_list));
890 css_set_skip_task_iters(from_cset, task);
891 list_del_init(&task->cg_list);
892 if (!css_set_populated(from_cset))
893 css_set_update_populated(from_cset, false);
895 WARN_ON_ONCE(!list_empty(&task->cg_list));
900 * We are synchronized through cgroup_threadgroup_rwsem
901 * against PF_EXITING setting such that we can't race
902 * against cgroup_exit()/cgroup_free() dropping the css_set.
904 WARN_ON_ONCE(task->flags & PF_EXITING);
906 cgroup_move_task(task, to_cset);
907 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
913 * hash table for cgroup groups. This improves the performance to find
914 * an existing css_set. This hash doesn't (currently) take into
915 * account cgroups in empty hierarchies.
917 #define CSS_SET_HASH_BITS 7
918 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
920 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
922 unsigned long key = 0UL;
923 struct cgroup_subsys *ss;
926 for_each_subsys(ss, i)
927 key += (unsigned long)css[i];
928 key = (key >> 16) ^ key;
933 void put_css_set_locked(struct css_set *cset)
935 struct cgrp_cset_link *link, *tmp_link;
936 struct cgroup_subsys *ss;
939 lockdep_assert_held(&css_set_lock);
941 if (!refcount_dec_and_test(&cset->refcount))
944 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
946 /* This css_set is dead. unlink it and release cgroup and css refs */
947 for_each_subsys(ss, ssid) {
948 list_del(&cset->e_cset_node[ssid]);
949 css_put(cset->subsys[ssid]);
951 hash_del(&cset->hlist);
954 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
955 list_del(&link->cset_link);
956 list_del(&link->cgrp_link);
957 if (cgroup_parent(link->cgrp))
958 cgroup_put(link->cgrp);
962 if (css_set_threaded(cset)) {
963 list_del(&cset->threaded_csets_node);
964 put_css_set_locked(cset->dom_cset);
967 kfree_rcu(cset, rcu_head);
971 * compare_css_sets - helper function for find_existing_css_set().
972 * @cset: candidate css_set being tested
973 * @old_cset: existing css_set for a task
974 * @new_cgrp: cgroup that's being entered by the task
975 * @template: desired set of css pointers in css_set (pre-calculated)
977 * Returns true if "cset" matches "old_cset" except for the hierarchy
978 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
980 static bool compare_css_sets(struct css_set *cset,
981 struct css_set *old_cset,
982 struct cgroup *new_cgrp,
983 struct cgroup_subsys_state *template[])
985 struct cgroup *new_dfl_cgrp;
986 struct list_head *l1, *l2;
989 * On the default hierarchy, there can be csets which are
990 * associated with the same set of cgroups but different csses.
991 * Let's first ensure that csses match.
993 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
997 /* @cset's domain should match the default cgroup's */
998 if (cgroup_on_dfl(new_cgrp))
999 new_dfl_cgrp = new_cgrp;
1001 new_dfl_cgrp = old_cset->dfl_cgrp;
1003 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1007 * Compare cgroup pointers in order to distinguish between
1008 * different cgroups in hierarchies. As different cgroups may
1009 * share the same effective css, this comparison is always
1012 l1 = &cset->cgrp_links;
1013 l2 = &old_cset->cgrp_links;
1015 struct cgrp_cset_link *link1, *link2;
1016 struct cgroup *cgrp1, *cgrp2;
1020 /* See if we reached the end - both lists are equal length. */
1021 if (l1 == &cset->cgrp_links) {
1022 BUG_ON(l2 != &old_cset->cgrp_links);
1025 BUG_ON(l2 == &old_cset->cgrp_links);
1027 /* Locate the cgroups associated with these links. */
1028 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1029 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1030 cgrp1 = link1->cgrp;
1031 cgrp2 = link2->cgrp;
1032 /* Hierarchies should be linked in the same order. */
1033 BUG_ON(cgrp1->root != cgrp2->root);
1036 * If this hierarchy is the hierarchy of the cgroup
1037 * that's changing, then we need to check that this
1038 * css_set points to the new cgroup; if it's any other
1039 * hierarchy, then this css_set should point to the
1040 * same cgroup as the old css_set.
1042 if (cgrp1->root == new_cgrp->root) {
1043 if (cgrp1 != new_cgrp)
1054 * find_existing_css_set - init css array and find the matching css_set
1055 * @old_cset: the css_set that we're using before the cgroup transition
1056 * @cgrp: the cgroup that we're moving into
1057 * @template: out param for the new set of csses, should be clear on entry
1059 static struct css_set *find_existing_css_set(struct css_set *old_cset,
1060 struct cgroup *cgrp,
1061 struct cgroup_subsys_state *template[])
1063 struct cgroup_root *root = cgrp->root;
1064 struct cgroup_subsys *ss;
1065 struct css_set *cset;
1070 * Build the set of subsystem state objects that we want to see in the
1071 * new css_set. while subsystems can change globally, the entries here
1072 * won't change, so no need for locking.
1074 for_each_subsys(ss, i) {
1075 if (root->subsys_mask & (1UL << i)) {
1077 * @ss is in this hierarchy, so we want the
1078 * effective css from @cgrp.
1080 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1083 * @ss is not in this hierarchy, so we don't want
1084 * to change the css.
1086 template[i] = old_cset->subsys[i];
1090 key = css_set_hash(template);
1091 hash_for_each_possible(css_set_table, cset, hlist, key) {
1092 if (!compare_css_sets(cset, old_cset, cgrp, template))
1095 /* This css_set matches what we need */
1099 /* No existing cgroup group matched */
1103 static void free_cgrp_cset_links(struct list_head *links_to_free)
1105 struct cgrp_cset_link *link, *tmp_link;
1107 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1108 list_del(&link->cset_link);
1114 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1115 * @count: the number of links to allocate
1116 * @tmp_links: list_head the allocated links are put on
1118 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1119 * through ->cset_link. Returns 0 on success or -errno.
1121 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1123 struct cgrp_cset_link *link;
1126 INIT_LIST_HEAD(tmp_links);
1128 for (i = 0; i < count; i++) {
1129 link = kzalloc(sizeof(*link), GFP_KERNEL);
1131 free_cgrp_cset_links(tmp_links);
1134 list_add(&link->cset_link, tmp_links);
1140 * link_css_set - a helper function to link a css_set to a cgroup
1141 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1142 * @cset: the css_set to be linked
1143 * @cgrp: the destination cgroup
1145 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1146 struct cgroup *cgrp)
1148 struct cgrp_cset_link *link;
1150 BUG_ON(list_empty(tmp_links));
1152 if (cgroup_on_dfl(cgrp))
1153 cset->dfl_cgrp = cgrp;
1155 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1160 * Always add links to the tail of the lists so that the lists are
1161 * in choronological order.
1163 list_move_tail(&link->cset_link, &cgrp->cset_links);
1164 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1166 if (cgroup_parent(cgrp))
1167 cgroup_get_live(cgrp);
1171 * find_css_set - return a new css_set with one cgroup updated
1172 * @old_cset: the baseline css_set
1173 * @cgrp: the cgroup to be updated
1175 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1176 * substituted into the appropriate hierarchy.
1178 static struct css_set *find_css_set(struct css_set *old_cset,
1179 struct cgroup *cgrp)
1181 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1182 struct css_set *cset;
1183 struct list_head tmp_links;
1184 struct cgrp_cset_link *link;
1185 struct cgroup_subsys *ss;
1189 lockdep_assert_held(&cgroup_mutex);
1191 /* First see if we already have a cgroup group that matches
1192 * the desired set */
1193 spin_lock_irq(&css_set_lock);
1194 cset = find_existing_css_set(old_cset, cgrp, template);
1197 spin_unlock_irq(&css_set_lock);
1202 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1206 /* Allocate all the cgrp_cset_link objects that we'll need */
1207 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1212 refcount_set(&cset->refcount, 1);
1213 cset->dom_cset = cset;
1214 INIT_LIST_HEAD(&cset->tasks);
1215 INIT_LIST_HEAD(&cset->mg_tasks);
1216 INIT_LIST_HEAD(&cset->dying_tasks);
1217 INIT_LIST_HEAD(&cset->task_iters);
1218 INIT_LIST_HEAD(&cset->threaded_csets);
1219 INIT_HLIST_NODE(&cset->hlist);
1220 INIT_LIST_HEAD(&cset->cgrp_links);
1221 INIT_LIST_HEAD(&cset->mg_preload_node);
1222 INIT_LIST_HEAD(&cset->mg_node);
1224 /* Copy the set of subsystem state objects generated in
1225 * find_existing_css_set() */
1226 memcpy(cset->subsys, template, sizeof(cset->subsys));
1228 spin_lock_irq(&css_set_lock);
1229 /* Add reference counts and links from the new css_set. */
1230 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1231 struct cgroup *c = link->cgrp;
1233 if (c->root == cgrp->root)
1235 link_css_set(&tmp_links, cset, c);
1238 BUG_ON(!list_empty(&tmp_links));
1242 /* Add @cset to the hash table */
1243 key = css_set_hash(cset->subsys);
1244 hash_add(css_set_table, &cset->hlist, key);
1246 for_each_subsys(ss, ssid) {
1247 struct cgroup_subsys_state *css = cset->subsys[ssid];
1249 list_add_tail(&cset->e_cset_node[ssid],
1250 &css->cgroup->e_csets[ssid]);
1254 spin_unlock_irq(&css_set_lock);
1257 * If @cset should be threaded, look up the matching dom_cset and
1258 * link them up. We first fully initialize @cset then look for the
1259 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1260 * to stay empty until we return.
1262 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1263 struct css_set *dcset;
1265 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1271 spin_lock_irq(&css_set_lock);
1272 cset->dom_cset = dcset;
1273 list_add_tail(&cset->threaded_csets_node,
1274 &dcset->threaded_csets);
1275 spin_unlock_irq(&css_set_lock);
1281 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1283 struct cgroup *root_cgrp = kf_root->kn->priv;
1285 return root_cgrp->root;
1288 static int cgroup_init_root_id(struct cgroup_root *root)
1292 lockdep_assert_held(&cgroup_mutex);
1294 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1298 root->hierarchy_id = id;
1302 static void cgroup_exit_root_id(struct cgroup_root *root)
1304 lockdep_assert_held(&cgroup_mutex);
1306 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1309 void cgroup_free_root(struct cgroup_root *root)
1314 static void cgroup_destroy_root(struct cgroup_root *root)
1316 struct cgroup *cgrp = &root->cgrp;
1317 struct cgrp_cset_link *link, *tmp_link;
1319 trace_cgroup_destroy_root(root);
1321 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1323 BUG_ON(atomic_read(&root->nr_cgrps));
1324 BUG_ON(!list_empty(&cgrp->self.children));
1326 /* Rebind all subsystems back to the default hierarchy */
1327 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1330 * Release all the links from cset_links to this hierarchy's
1333 spin_lock_irq(&css_set_lock);
1335 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1336 list_del(&link->cset_link);
1337 list_del(&link->cgrp_link);
1341 spin_unlock_irq(&css_set_lock);
1343 if (!list_empty(&root->root_list)) {
1344 list_del(&root->root_list);
1345 cgroup_root_count--;
1348 cgroup_exit_root_id(root);
1350 mutex_unlock(&cgroup_mutex);
1352 kernfs_destroy_root(root->kf_root);
1353 cgroup_free_root(root);
1357 * look up cgroup associated with current task's cgroup namespace on the
1358 * specified hierarchy
1360 static struct cgroup *
1361 current_cgns_cgroup_from_root(struct cgroup_root *root)
1363 struct cgroup *res = NULL;
1364 struct css_set *cset;
1366 lockdep_assert_held(&css_set_lock);
1370 cset = current->nsproxy->cgroup_ns->root_cset;
1371 if (cset == &init_css_set) {
1373 } else if (root == &cgrp_dfl_root) {
1374 res = cset->dfl_cgrp;
1376 struct cgrp_cset_link *link;
1378 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1379 struct cgroup *c = link->cgrp;
1381 if (c->root == root) {
1393 /* look up cgroup associated with given css_set on the specified hierarchy */
1394 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1395 struct cgroup_root *root)
1397 struct cgroup *res = NULL;
1399 lockdep_assert_held(&cgroup_mutex);
1400 lockdep_assert_held(&css_set_lock);
1402 if (cset == &init_css_set) {
1404 } else if (root == &cgrp_dfl_root) {
1405 res = cset->dfl_cgrp;
1407 struct cgrp_cset_link *link;
1409 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1410 struct cgroup *c = link->cgrp;
1412 if (c->root == root) {
1424 * Return the cgroup for "task" from the given hierarchy. Must be
1425 * called with cgroup_mutex and css_set_lock held.
1427 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1428 struct cgroup_root *root)
1431 * No need to lock the task - since we hold css_set_lock the
1432 * task can't change groups.
1434 return cset_cgroup_from_root(task_css_set(task), root);
1438 * A task must hold cgroup_mutex to modify cgroups.
1440 * Any task can increment and decrement the count field without lock.
1441 * So in general, code holding cgroup_mutex can't rely on the count
1442 * field not changing. However, if the count goes to zero, then only
1443 * cgroup_attach_task() can increment it again. Because a count of zero
1444 * means that no tasks are currently attached, therefore there is no
1445 * way a task attached to that cgroup can fork (the other way to
1446 * increment the count). So code holding cgroup_mutex can safely
1447 * assume that if the count is zero, it will stay zero. Similarly, if
1448 * a task holds cgroup_mutex on a cgroup with zero count, it
1449 * knows that the cgroup won't be removed, as cgroup_rmdir()
1452 * A cgroup can only be deleted if both its 'count' of using tasks
1453 * is zero, and its list of 'children' cgroups is empty. Since all
1454 * tasks in the system use _some_ cgroup, and since there is always at
1455 * least one task in the system (init, pid == 1), therefore, root cgroup
1456 * always has either children cgroups and/or using tasks. So we don't
1457 * need a special hack to ensure that root cgroup cannot be deleted.
1459 * P.S. One more locking exception. RCU is used to guard the
1460 * update of a tasks cgroup pointer by cgroup_attach_task()
1463 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1465 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1468 struct cgroup_subsys *ss = cft->ss;
1470 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1471 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1472 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1474 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1475 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1478 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1484 * cgroup_file_mode - deduce file mode of a control file
1485 * @cft: the control file in question
1487 * S_IRUGO for read, S_IWUSR for write.
1489 static umode_t cgroup_file_mode(const struct cftype *cft)
1493 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1496 if (cft->write_u64 || cft->write_s64 || cft->write) {
1497 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1507 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1508 * @subtree_control: the new subtree_control mask to consider
1509 * @this_ss_mask: available subsystems
1511 * On the default hierarchy, a subsystem may request other subsystems to be
1512 * enabled together through its ->depends_on mask. In such cases, more
1513 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1515 * This function calculates which subsystems need to be enabled if
1516 * @subtree_control is to be applied while restricted to @this_ss_mask.
1518 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1520 u16 cur_ss_mask = subtree_control;
1521 struct cgroup_subsys *ss;
1524 lockdep_assert_held(&cgroup_mutex);
1526 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1529 u16 new_ss_mask = cur_ss_mask;
1531 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1532 new_ss_mask |= ss->depends_on;
1533 } while_each_subsys_mask();
1536 * Mask out subsystems which aren't available. This can
1537 * happen only if some depended-upon subsystems were bound
1538 * to non-default hierarchies.
1540 new_ss_mask &= this_ss_mask;
1542 if (new_ss_mask == cur_ss_mask)
1544 cur_ss_mask = new_ss_mask;
1551 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1552 * @kn: the kernfs_node being serviced
1554 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1555 * the method finishes if locking succeeded. Note that once this function
1556 * returns the cgroup returned by cgroup_kn_lock_live() may become
1557 * inaccessible any time. If the caller intends to continue to access the
1558 * cgroup, it should pin it before invoking this function.
1560 void cgroup_kn_unlock(struct kernfs_node *kn)
1562 struct cgroup *cgrp;
1564 if (kernfs_type(kn) == KERNFS_DIR)
1567 cgrp = kn->parent->priv;
1569 mutex_unlock(&cgroup_mutex);
1571 kernfs_unbreak_active_protection(kn);
1576 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1577 * @kn: the kernfs_node being serviced
1578 * @drain_offline: perform offline draining on the cgroup
1580 * This helper is to be used by a cgroup kernfs method currently servicing
1581 * @kn. It breaks the active protection, performs cgroup locking and
1582 * verifies that the associated cgroup is alive. Returns the cgroup if
1583 * alive; otherwise, %NULL. A successful return should be undone by a
1584 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1585 * cgroup is drained of offlining csses before return.
1587 * Any cgroup kernfs method implementation which requires locking the
1588 * associated cgroup should use this helper. It avoids nesting cgroup
1589 * locking under kernfs active protection and allows all kernfs operations
1590 * including self-removal.
1592 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1594 struct cgroup *cgrp;
1596 if (kernfs_type(kn) == KERNFS_DIR)
1599 cgrp = kn->parent->priv;
1602 * We're gonna grab cgroup_mutex which nests outside kernfs
1603 * active_ref. cgroup liveliness check alone provides enough
1604 * protection against removal. Ensure @cgrp stays accessible and
1605 * break the active_ref protection.
1607 if (!cgroup_tryget(cgrp))
1609 kernfs_break_active_protection(kn);
1612 cgroup_lock_and_drain_offline(cgrp);
1614 mutex_lock(&cgroup_mutex);
1616 if (!cgroup_is_dead(cgrp))
1619 cgroup_kn_unlock(kn);
1623 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1625 char name[CGROUP_FILE_NAME_MAX];
1627 lockdep_assert_held(&cgroup_mutex);
1629 if (cft->file_offset) {
1630 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1631 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1633 spin_lock_irq(&cgroup_file_kn_lock);
1635 spin_unlock_irq(&cgroup_file_kn_lock);
1637 del_timer_sync(&cfile->notify_timer);
1640 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1644 * css_clear_dir - remove subsys files in a cgroup directory
1647 static void css_clear_dir(struct cgroup_subsys_state *css)
1649 struct cgroup *cgrp = css->cgroup;
1650 struct cftype *cfts;
1652 if (!(css->flags & CSS_VISIBLE))
1655 css->flags &= ~CSS_VISIBLE;
1658 if (cgroup_on_dfl(cgrp))
1659 cfts = cgroup_base_files;
1661 cfts = cgroup1_base_files;
1663 cgroup_addrm_files(css, cgrp, cfts, false);
1665 list_for_each_entry(cfts, &css->ss->cfts, node)
1666 cgroup_addrm_files(css, cgrp, cfts, false);
1671 * css_populate_dir - create subsys files in a cgroup directory
1674 * On failure, no file is added.
1676 static int css_populate_dir(struct cgroup_subsys_state *css)
1678 struct cgroup *cgrp = css->cgroup;
1679 struct cftype *cfts, *failed_cfts;
1682 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1686 if (cgroup_on_dfl(cgrp))
1687 cfts = cgroup_base_files;
1689 cfts = cgroup1_base_files;
1691 ret = cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1695 list_for_each_entry(cfts, &css->ss->cfts, node) {
1696 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1704 css->flags |= CSS_VISIBLE;
1708 list_for_each_entry(cfts, &css->ss->cfts, node) {
1709 if (cfts == failed_cfts)
1711 cgroup_addrm_files(css, cgrp, cfts, false);
1716 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1718 struct cgroup *dcgrp = &dst_root->cgrp;
1719 struct cgroup_subsys *ss;
1722 lockdep_assert_held(&cgroup_mutex);
1724 do_each_subsys_mask(ss, ssid, ss_mask) {
1726 * If @ss has non-root csses attached to it, can't move.
1727 * If @ss is an implicit controller, it is exempt from this
1728 * rule and can be stolen.
1730 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1731 !ss->implicit_on_dfl)
1734 /* can't move between two non-dummy roots either */
1735 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1737 } while_each_subsys_mask();
1739 do_each_subsys_mask(ss, ssid, ss_mask) {
1740 struct cgroup_root *src_root = ss->root;
1741 struct cgroup *scgrp = &src_root->cgrp;
1742 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1743 struct css_set *cset;
1745 WARN_ON(!css || cgroup_css(dcgrp, ss));
1747 /* disable from the source */
1748 src_root->subsys_mask &= ~(1 << ssid);
1749 WARN_ON(cgroup_apply_control(scgrp));
1750 cgroup_finalize_control(scgrp, 0);
1753 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1754 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1755 ss->root = dst_root;
1756 css->cgroup = dcgrp;
1758 spin_lock_irq(&css_set_lock);
1759 hash_for_each(css_set_table, i, cset, hlist)
1760 list_move_tail(&cset->e_cset_node[ss->id],
1761 &dcgrp->e_csets[ss->id]);
1762 spin_unlock_irq(&css_set_lock);
1764 /* default hierarchy doesn't enable controllers by default */
1765 dst_root->subsys_mask |= 1 << ssid;
1766 if (dst_root == &cgrp_dfl_root) {
1767 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1769 dcgrp->subtree_control |= 1 << ssid;
1770 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1773 ret = cgroup_apply_control(dcgrp);
1775 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1780 } while_each_subsys_mask();
1782 kernfs_activate(dcgrp->kn);
1786 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1787 struct kernfs_root *kf_root)
1791 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1792 struct cgroup *ns_cgroup;
1794 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1798 spin_lock_irq(&css_set_lock);
1799 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1800 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1801 spin_unlock_irq(&css_set_lock);
1803 if (len >= PATH_MAX)
1806 seq_escape(sf, buf, " \t\n\\");
1813 enum cgroup2_param {
1815 Opt_memory_localevents,
1819 static const struct fs_parameter_spec cgroup2_param_specs[] = {
1820 fsparam_flag("nsdelegate", Opt_nsdelegate),
1821 fsparam_flag("memory_localevents", Opt_memory_localevents),
1825 static const struct fs_parameter_description cgroup2_fs_parameters = {
1827 .specs = cgroup2_param_specs,
1830 static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1832 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1833 struct fs_parse_result result;
1836 opt = fs_parse(fc, &cgroup2_fs_parameters, param, &result);
1841 case Opt_nsdelegate:
1842 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1844 case Opt_memory_localevents:
1845 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1851 static void apply_cgroup_root_flags(unsigned int root_flags)
1853 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1854 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1855 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1857 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1859 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1860 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1862 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1866 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1868 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1869 seq_puts(seq, ",nsdelegate");
1870 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1871 seq_puts(seq, ",memory_localevents");
1875 static int cgroup_reconfigure(struct fs_context *fc)
1877 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1879 apply_cgroup_root_flags(ctx->flags);
1883 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1885 struct cgroup_subsys *ss;
1888 INIT_LIST_HEAD(&cgrp->self.sibling);
1889 INIT_LIST_HEAD(&cgrp->self.children);
1890 INIT_LIST_HEAD(&cgrp->cset_links);
1891 INIT_LIST_HEAD(&cgrp->pidlists);
1892 mutex_init(&cgrp->pidlist_mutex);
1893 cgrp->self.cgroup = cgrp;
1894 cgrp->self.flags |= CSS_ONLINE;
1895 cgrp->dom_cgrp = cgrp;
1896 cgrp->max_descendants = INT_MAX;
1897 cgrp->max_depth = INT_MAX;
1898 INIT_LIST_HEAD(&cgrp->rstat_css_list);
1899 prev_cputime_init(&cgrp->prev_cputime);
1901 for_each_subsys(ss, ssid)
1902 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1904 init_waitqueue_head(&cgrp->offline_waitq);
1905 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1908 void init_cgroup_root(struct cgroup_fs_context *ctx)
1910 struct cgroup_root *root = ctx->root;
1911 struct cgroup *cgrp = &root->cgrp;
1913 INIT_LIST_HEAD(&root->root_list);
1914 atomic_set(&root->nr_cgrps, 1);
1916 init_cgroup_housekeeping(cgrp);
1918 root->flags = ctx->flags;
1919 if (ctx->release_agent)
1920 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
1922 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
1923 if (ctx->cpuset_clone_children)
1924 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1927 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1929 LIST_HEAD(tmp_links);
1930 struct cgroup *root_cgrp = &root->cgrp;
1931 struct kernfs_syscall_ops *kf_sops;
1932 struct css_set *cset;
1935 lockdep_assert_held(&cgroup_mutex);
1937 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1943 * We're accessing css_set_count without locking css_set_lock here,
1944 * but that's OK - it can only be increased by someone holding
1945 * cgroup_lock, and that's us. Later rebinding may disable
1946 * controllers on the default hierarchy and thus create new csets,
1947 * which can't be more than the existing ones. Allocate 2x.
1949 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1953 ret = cgroup_init_root_id(root);
1957 kf_sops = root == &cgrp_dfl_root ?
1958 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1960 root->kf_root = kernfs_create_root(kf_sops,
1961 KERNFS_ROOT_CREATE_DEACTIVATED |
1962 KERNFS_ROOT_SUPPORT_EXPORTOP,
1964 if (IS_ERR(root->kf_root)) {
1965 ret = PTR_ERR(root->kf_root);
1968 root_cgrp->kn = root->kf_root->kn;
1969 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
1970 root_cgrp->ancestor_ids[0] = cgroup_id(root_cgrp);
1972 ret = css_populate_dir(&root_cgrp->self);
1976 ret = rebind_subsystems(root, ss_mask);
1980 ret = cgroup_bpf_inherit(root_cgrp);
1983 trace_cgroup_setup_root(root);
1986 * There must be no failure case after here, since rebinding takes
1987 * care of subsystems' refcounts, which are explicitly dropped in
1988 * the failure exit path.
1990 list_add(&root->root_list, &cgroup_roots);
1991 cgroup_root_count++;
1994 * Link the root cgroup in this hierarchy into all the css_set
1997 spin_lock_irq(&css_set_lock);
1998 hash_for_each(css_set_table, i, cset, hlist) {
1999 link_css_set(&tmp_links, cset, root_cgrp);
2000 if (css_set_populated(cset))
2001 cgroup_update_populated(root_cgrp, true);
2003 spin_unlock_irq(&css_set_lock);
2005 BUG_ON(!list_empty(&root_cgrp->self.children));
2006 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2008 kernfs_activate(root_cgrp->kn);
2013 kernfs_destroy_root(root->kf_root);
2014 root->kf_root = NULL;
2016 cgroup_exit_root_id(root);
2018 percpu_ref_exit(&root_cgrp->self.refcnt);
2020 free_cgrp_cset_links(&tmp_links);
2024 int cgroup_do_get_tree(struct fs_context *fc)
2026 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2029 ctx->kfc.root = ctx->root->kf_root;
2030 if (fc->fs_type == &cgroup2_fs_type)
2031 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2033 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2034 ret = kernfs_get_tree(fc);
2037 * In non-init cgroup namespace, instead of root cgroup's dentry,
2038 * we return the dentry corresponding to the cgroupns->root_cgrp.
2040 if (!ret && ctx->ns != &init_cgroup_ns) {
2041 struct dentry *nsdentry;
2042 struct super_block *sb = fc->root->d_sb;
2043 struct cgroup *cgrp;
2045 mutex_lock(&cgroup_mutex);
2046 spin_lock_irq(&css_set_lock);
2048 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2050 spin_unlock_irq(&css_set_lock);
2051 mutex_unlock(&cgroup_mutex);
2053 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2055 if (IS_ERR(nsdentry)) {
2056 deactivate_locked_super(sb);
2057 ret = PTR_ERR(nsdentry);
2060 fc->root = nsdentry;
2063 if (!ctx->kfc.new_sb_created)
2064 cgroup_put(&ctx->root->cgrp);
2070 * Destroy a cgroup filesystem context.
2072 static void cgroup_fs_context_free(struct fs_context *fc)
2074 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2077 kfree(ctx->release_agent);
2078 put_cgroup_ns(ctx->ns);
2079 kernfs_free_fs_context(fc);
2083 static int cgroup_get_tree(struct fs_context *fc)
2085 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2088 cgrp_dfl_visible = true;
2089 cgroup_get_live(&cgrp_dfl_root.cgrp);
2090 ctx->root = &cgrp_dfl_root;
2092 ret = cgroup_do_get_tree(fc);
2094 apply_cgroup_root_flags(ctx->flags);
2098 static const struct fs_context_operations cgroup_fs_context_ops = {
2099 .free = cgroup_fs_context_free,
2100 .parse_param = cgroup2_parse_param,
2101 .get_tree = cgroup_get_tree,
2102 .reconfigure = cgroup_reconfigure,
2105 static const struct fs_context_operations cgroup1_fs_context_ops = {
2106 .free = cgroup_fs_context_free,
2107 .parse_param = cgroup1_parse_param,
2108 .get_tree = cgroup1_get_tree,
2109 .reconfigure = cgroup1_reconfigure,
2113 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2114 * we select the namespace we're going to use.
2116 static int cgroup_init_fs_context(struct fs_context *fc)
2118 struct cgroup_fs_context *ctx;
2120 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2124 ctx->ns = current->nsproxy->cgroup_ns;
2125 get_cgroup_ns(ctx->ns);
2126 fc->fs_private = &ctx->kfc;
2127 if (fc->fs_type == &cgroup2_fs_type)
2128 fc->ops = &cgroup_fs_context_ops;
2130 fc->ops = &cgroup1_fs_context_ops;
2131 put_user_ns(fc->user_ns);
2132 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2137 static void cgroup_kill_sb(struct super_block *sb)
2139 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2140 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2143 * If @root doesn't have any children, start killing it.
2144 * This prevents new mounts by disabling percpu_ref_tryget_live().
2145 * cgroup_mount() may wait for @root's release.
2147 * And don't kill the default root.
2149 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2150 !percpu_ref_is_dying(&root->cgrp.self.refcnt))
2151 percpu_ref_kill(&root->cgrp.self.refcnt);
2152 cgroup_put(&root->cgrp);
2156 struct file_system_type cgroup_fs_type = {
2158 .init_fs_context = cgroup_init_fs_context,
2159 .parameters = &cgroup1_fs_parameters,
2160 .kill_sb = cgroup_kill_sb,
2161 .fs_flags = FS_USERNS_MOUNT,
2164 static struct file_system_type cgroup2_fs_type = {
2166 .init_fs_context = cgroup_init_fs_context,
2167 .parameters = &cgroup2_fs_parameters,
2168 .kill_sb = cgroup_kill_sb,
2169 .fs_flags = FS_USERNS_MOUNT,
2172 #ifdef CONFIG_CPUSETS
2173 static const struct fs_context_operations cpuset_fs_context_ops = {
2174 .get_tree = cgroup1_get_tree,
2175 .free = cgroup_fs_context_free,
2179 * This is ugly, but preserves the userspace API for existing cpuset
2180 * users. If someone tries to mount the "cpuset" filesystem, we
2181 * silently switch it to mount "cgroup" instead
2183 static int cpuset_init_fs_context(struct fs_context *fc)
2185 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2186 struct cgroup_fs_context *ctx;
2189 err = cgroup_init_fs_context(fc);
2195 fc->ops = &cpuset_fs_context_ops;
2197 ctx = cgroup_fc2context(fc);
2198 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2199 ctx->flags |= CGRP_ROOT_NOPREFIX;
2200 ctx->release_agent = agent;
2202 get_filesystem(&cgroup_fs_type);
2203 put_filesystem(fc->fs_type);
2204 fc->fs_type = &cgroup_fs_type;
2209 static struct file_system_type cpuset_fs_type = {
2211 .init_fs_context = cpuset_init_fs_context,
2212 .fs_flags = FS_USERNS_MOUNT,
2216 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2217 struct cgroup_namespace *ns)
2219 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2221 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2224 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2225 struct cgroup_namespace *ns)
2229 mutex_lock(&cgroup_mutex);
2230 spin_lock_irq(&css_set_lock);
2232 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2234 spin_unlock_irq(&css_set_lock);
2235 mutex_unlock(&cgroup_mutex);
2239 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2242 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2243 * @task: target task
2244 * @buf: the buffer to write the path into
2245 * @buflen: the length of the buffer
2247 * Determine @task's cgroup on the first (the one with the lowest non-zero
2248 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2249 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2250 * cgroup controller callbacks.
2252 * Return value is the same as kernfs_path().
2254 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2256 struct cgroup_root *root;
2257 struct cgroup *cgrp;
2258 int hierarchy_id = 1;
2261 mutex_lock(&cgroup_mutex);
2262 spin_lock_irq(&css_set_lock);
2264 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2267 cgrp = task_cgroup_from_root(task, root);
2268 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2270 /* if no hierarchy exists, everyone is in "/" */
2271 ret = strlcpy(buf, "/", buflen);
2274 spin_unlock_irq(&css_set_lock);
2275 mutex_unlock(&cgroup_mutex);
2278 EXPORT_SYMBOL_GPL(task_cgroup_path);
2281 * cgroup_migrate_add_task - add a migration target task to a migration context
2282 * @task: target task
2283 * @mgctx: target migration context
2285 * Add @task, which is a migration target, to @mgctx->tset. This function
2286 * becomes noop if @task doesn't need to be migrated. @task's css_set
2287 * should have been added as a migration source and @task->cg_list will be
2288 * moved from the css_set's tasks list to mg_tasks one.
2290 static void cgroup_migrate_add_task(struct task_struct *task,
2291 struct cgroup_mgctx *mgctx)
2293 struct css_set *cset;
2295 lockdep_assert_held(&css_set_lock);
2297 /* @task either already exited or can't exit until the end */
2298 if (task->flags & PF_EXITING)
2301 /* cgroup_threadgroup_rwsem protects racing against forks */
2302 WARN_ON_ONCE(list_empty(&task->cg_list));
2304 cset = task_css_set(task);
2305 if (!cset->mg_src_cgrp)
2308 mgctx->tset.nr_tasks++;
2310 list_move_tail(&task->cg_list, &cset->mg_tasks);
2311 if (list_empty(&cset->mg_node))
2312 list_add_tail(&cset->mg_node,
2313 &mgctx->tset.src_csets);
2314 if (list_empty(&cset->mg_dst_cset->mg_node))
2315 list_add_tail(&cset->mg_dst_cset->mg_node,
2316 &mgctx->tset.dst_csets);
2320 * cgroup_taskset_first - reset taskset and return the first task
2321 * @tset: taskset of interest
2322 * @dst_cssp: output variable for the destination css
2324 * @tset iteration is initialized and the first task is returned.
2326 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2327 struct cgroup_subsys_state **dst_cssp)
2329 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2330 tset->cur_task = NULL;
2332 return cgroup_taskset_next(tset, dst_cssp);
2336 * cgroup_taskset_next - iterate to the next task in taskset
2337 * @tset: taskset of interest
2338 * @dst_cssp: output variable for the destination css
2340 * Return the next task in @tset. Iteration must have been initialized
2341 * with cgroup_taskset_first().
2343 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2344 struct cgroup_subsys_state **dst_cssp)
2346 struct css_set *cset = tset->cur_cset;
2347 struct task_struct *task = tset->cur_task;
2349 while (&cset->mg_node != tset->csets) {
2351 task = list_first_entry(&cset->mg_tasks,
2352 struct task_struct, cg_list);
2354 task = list_next_entry(task, cg_list);
2356 if (&task->cg_list != &cset->mg_tasks) {
2357 tset->cur_cset = cset;
2358 tset->cur_task = task;
2361 * This function may be called both before and
2362 * after cgroup_taskset_migrate(). The two cases
2363 * can be distinguished by looking at whether @cset
2364 * has its ->mg_dst_cset set.
2366 if (cset->mg_dst_cset)
2367 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2369 *dst_cssp = cset->subsys[tset->ssid];
2374 cset = list_next_entry(cset, mg_node);
2382 * cgroup_taskset_migrate - migrate a taskset
2383 * @mgctx: migration context
2385 * Migrate tasks in @mgctx as setup by migration preparation functions.
2386 * This function fails iff one of the ->can_attach callbacks fails and
2387 * guarantees that either all or none of the tasks in @mgctx are migrated.
2388 * @mgctx is consumed regardless of success.
2390 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2392 struct cgroup_taskset *tset = &mgctx->tset;
2393 struct cgroup_subsys *ss;
2394 struct task_struct *task, *tmp_task;
2395 struct css_set *cset, *tmp_cset;
2396 int ssid, failed_ssid, ret;
2398 /* check that we can legitimately attach to the cgroup */
2399 if (tset->nr_tasks) {
2400 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2401 if (ss->can_attach) {
2403 ret = ss->can_attach(tset);
2406 goto out_cancel_attach;
2409 } while_each_subsys_mask();
2413 * Now that we're guaranteed success, proceed to move all tasks to
2414 * the new cgroup. There are no failure cases after here, so this
2415 * is the commit point.
2417 spin_lock_irq(&css_set_lock);
2418 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2419 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2420 struct css_set *from_cset = task_css_set(task);
2421 struct css_set *to_cset = cset->mg_dst_cset;
2423 get_css_set(to_cset);
2424 to_cset->nr_tasks++;
2425 css_set_move_task(task, from_cset, to_cset, true);
2426 from_cset->nr_tasks--;
2428 * If the source or destination cgroup is frozen,
2429 * the task might require to change its state.
2431 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2433 put_css_set_locked(from_cset);
2437 spin_unlock_irq(&css_set_lock);
2440 * Migration is committed, all target tasks are now on dst_csets.
2441 * Nothing is sensitive to fork() after this point. Notify
2442 * controllers that migration is complete.
2444 tset->csets = &tset->dst_csets;
2446 if (tset->nr_tasks) {
2447 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2452 } while_each_subsys_mask();
2456 goto out_release_tset;
2459 if (tset->nr_tasks) {
2460 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2461 if (ssid == failed_ssid)
2463 if (ss->cancel_attach) {
2465 ss->cancel_attach(tset);
2467 } while_each_subsys_mask();
2470 spin_lock_irq(&css_set_lock);
2471 list_splice_init(&tset->dst_csets, &tset->src_csets);
2472 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2473 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2474 list_del_init(&cset->mg_node);
2476 spin_unlock_irq(&css_set_lock);
2479 * Re-initialize the cgroup_taskset structure in case it is reused
2480 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2484 tset->csets = &tset->src_csets;
2489 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2490 * @dst_cgrp: destination cgroup to test
2492 * On the default hierarchy, except for the mixable, (possible) thread root
2493 * and threaded cgroups, subtree_control must be zero for migration
2494 * destination cgroups with tasks so that child cgroups don't compete
2497 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2499 /* v1 doesn't have any restriction */
2500 if (!cgroup_on_dfl(dst_cgrp))
2503 /* verify @dst_cgrp can host resources */
2504 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2507 /* mixables don't care */
2508 if (cgroup_is_mixable(dst_cgrp))
2512 * If @dst_cgrp is already or can become a thread root or is
2513 * threaded, it doesn't matter.
2515 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2518 /* apply no-internal-process constraint */
2519 if (dst_cgrp->subtree_control)
2526 * cgroup_migrate_finish - cleanup after attach
2527 * @mgctx: migration context
2529 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2530 * those functions for details.
2532 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2534 LIST_HEAD(preloaded);
2535 struct css_set *cset, *tmp_cset;
2537 lockdep_assert_held(&cgroup_mutex);
2539 spin_lock_irq(&css_set_lock);
2541 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2542 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2544 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2545 cset->mg_src_cgrp = NULL;
2546 cset->mg_dst_cgrp = NULL;
2547 cset->mg_dst_cset = NULL;
2548 list_del_init(&cset->mg_preload_node);
2549 put_css_set_locked(cset);
2552 spin_unlock_irq(&css_set_lock);
2556 * cgroup_migrate_add_src - add a migration source css_set
2557 * @src_cset: the source css_set to add
2558 * @dst_cgrp: the destination cgroup
2559 * @mgctx: migration context
2561 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2562 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2563 * up by cgroup_migrate_finish().
2565 * This function may be called without holding cgroup_threadgroup_rwsem
2566 * even if the target is a process. Threads may be created and destroyed
2567 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2568 * into play and the preloaded css_sets are guaranteed to cover all
2571 void cgroup_migrate_add_src(struct css_set *src_cset,
2572 struct cgroup *dst_cgrp,
2573 struct cgroup_mgctx *mgctx)
2575 struct cgroup *src_cgrp;
2577 lockdep_assert_held(&cgroup_mutex);
2578 lockdep_assert_held(&css_set_lock);
2581 * If ->dead, @src_set is associated with one or more dead cgroups
2582 * and doesn't contain any migratable tasks. Ignore it early so
2583 * that the rest of migration path doesn't get confused by it.
2588 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2590 if (!list_empty(&src_cset->mg_preload_node))
2593 WARN_ON(src_cset->mg_src_cgrp);
2594 WARN_ON(src_cset->mg_dst_cgrp);
2595 WARN_ON(!list_empty(&src_cset->mg_tasks));
2596 WARN_ON(!list_empty(&src_cset->mg_node));
2598 src_cset->mg_src_cgrp = src_cgrp;
2599 src_cset->mg_dst_cgrp = dst_cgrp;
2600 get_css_set(src_cset);
2601 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2605 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2606 * @mgctx: migration context
2608 * Tasks are about to be moved and all the source css_sets have been
2609 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2610 * pins all destination css_sets, links each to its source, and append them
2611 * to @mgctx->preloaded_dst_csets.
2613 * This function must be called after cgroup_migrate_add_src() has been
2614 * called on each migration source css_set. After migration is performed
2615 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2618 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2620 struct css_set *src_cset, *tmp_cset;
2622 lockdep_assert_held(&cgroup_mutex);
2624 /* look up the dst cset for each src cset and link it to src */
2625 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2627 struct css_set *dst_cset;
2628 struct cgroup_subsys *ss;
2631 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2635 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2638 * If src cset equals dst, it's noop. Drop the src.
2639 * cgroup_migrate() will skip the cset too. Note that we
2640 * can't handle src == dst as some nodes are used by both.
2642 if (src_cset == dst_cset) {
2643 src_cset->mg_src_cgrp = NULL;
2644 src_cset->mg_dst_cgrp = NULL;
2645 list_del_init(&src_cset->mg_preload_node);
2646 put_css_set(src_cset);
2647 put_css_set(dst_cset);
2651 src_cset->mg_dst_cset = dst_cset;
2653 if (list_empty(&dst_cset->mg_preload_node))
2654 list_add_tail(&dst_cset->mg_preload_node,
2655 &mgctx->preloaded_dst_csets);
2657 put_css_set(dst_cset);
2659 for_each_subsys(ss, ssid)
2660 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2661 mgctx->ss_mask |= 1 << ssid;
2668 * cgroup_migrate - migrate a process or task to a cgroup
2669 * @leader: the leader of the process or the task to migrate
2670 * @threadgroup: whether @leader points to the whole process or a single task
2671 * @mgctx: migration context
2673 * Migrate a process or task denoted by @leader. If migrating a process,
2674 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2675 * responsible for invoking cgroup_migrate_add_src() and
2676 * cgroup_migrate_prepare_dst() on the targets before invoking this
2677 * function and following up with cgroup_migrate_finish().
2679 * As long as a controller's ->can_attach() doesn't fail, this function is
2680 * guaranteed to succeed. This means that, excluding ->can_attach()
2681 * failure, when migrating multiple targets, the success or failure can be
2682 * decided for all targets by invoking group_migrate_prepare_dst() before
2683 * actually starting migrating.
2685 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2686 struct cgroup_mgctx *mgctx)
2688 struct task_struct *task;
2691 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2692 * already PF_EXITING could be freed from underneath us unless we
2693 * take an rcu_read_lock.
2695 spin_lock_irq(&css_set_lock);
2699 cgroup_migrate_add_task(task, mgctx);
2702 } while_each_thread(leader, task);
2704 spin_unlock_irq(&css_set_lock);
2706 return cgroup_migrate_execute(mgctx);
2710 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2711 * @dst_cgrp: the cgroup to attach to
2712 * @leader: the task or the leader of the threadgroup to be attached
2713 * @threadgroup: attach the whole threadgroup?
2715 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2717 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2720 DEFINE_CGROUP_MGCTX(mgctx);
2721 struct task_struct *task;
2724 ret = cgroup_migrate_vet_dst(dst_cgrp);
2728 /* look up all src csets */
2729 spin_lock_irq(&css_set_lock);
2733 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2736 } while_each_thread(leader, task);
2738 spin_unlock_irq(&css_set_lock);
2740 /* prepare dst csets and commit */
2741 ret = cgroup_migrate_prepare_dst(&mgctx);
2743 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2745 cgroup_migrate_finish(&mgctx);
2748 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2753 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2755 __acquires(&cgroup_threadgroup_rwsem)
2757 struct task_struct *tsk;
2760 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2761 return ERR_PTR(-EINVAL);
2764 * If we migrate a single thread, we don't care about threadgroup
2765 * stability. If the thread is `current`, it won't exit(2) under our
2766 * hands or change PID through exec(2). We exclude
2767 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2768 * callers by cgroup_mutex.
2769 * Therefore, we can skip the global lock.
2771 lockdep_assert_held(&cgroup_mutex);
2772 if (pid || threadgroup) {
2773 percpu_down_write(&cgroup_threadgroup_rwsem);
2781 tsk = find_task_by_vpid(pid);
2783 tsk = ERR_PTR(-ESRCH);
2784 goto out_unlock_threadgroup;
2791 tsk = tsk->group_leader;
2794 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2795 * If userland migrates such a kthread to a non-root cgroup, it can
2796 * become trapped in a cpuset, or RT kthread may be born in a
2797 * cgroup with no rt_runtime allocated. Just say no.
2799 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2800 tsk = ERR_PTR(-EINVAL);
2801 goto out_unlock_threadgroup;
2804 get_task_struct(tsk);
2805 goto out_unlock_rcu;
2807 out_unlock_threadgroup:
2809 percpu_up_write(&cgroup_threadgroup_rwsem);
2817 void cgroup_procs_write_finish(struct task_struct *task, bool locked)
2818 __releases(&cgroup_threadgroup_rwsem)
2820 struct cgroup_subsys *ss;
2823 /* release reference from cgroup_procs_write_start() */
2824 put_task_struct(task);
2827 percpu_up_write(&cgroup_threadgroup_rwsem);
2828 for_each_subsys(ss, ssid)
2829 if (ss->post_attach)
2833 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2835 struct cgroup_subsys *ss;
2836 bool printed = false;
2839 do_each_subsys_mask(ss, ssid, ss_mask) {
2842 seq_puts(seq, ss->name);
2844 } while_each_subsys_mask();
2846 seq_putc(seq, '\n');
2849 /* show controllers which are enabled from the parent */
2850 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2852 struct cgroup *cgrp = seq_css(seq)->cgroup;
2854 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2858 /* show controllers which are enabled for a given cgroup's children */
2859 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2861 struct cgroup *cgrp = seq_css(seq)->cgroup;
2863 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2868 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2869 * @cgrp: root of the subtree to update csses for
2871 * @cgrp's control masks have changed and its subtree's css associations
2872 * need to be updated accordingly. This function looks up all css_sets
2873 * which are attached to the subtree, creates the matching updated css_sets
2874 * and migrates the tasks to the new ones.
2876 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2878 DEFINE_CGROUP_MGCTX(mgctx);
2879 struct cgroup_subsys_state *d_css;
2880 struct cgroup *dsct;
2881 struct css_set *src_cset;
2884 lockdep_assert_held(&cgroup_mutex);
2886 percpu_down_write(&cgroup_threadgroup_rwsem);
2888 /* look up all csses currently attached to @cgrp's subtree */
2889 spin_lock_irq(&css_set_lock);
2890 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2891 struct cgrp_cset_link *link;
2893 list_for_each_entry(link, &dsct->cset_links, cset_link)
2894 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2896 spin_unlock_irq(&css_set_lock);
2898 /* NULL dst indicates self on default hierarchy */
2899 ret = cgroup_migrate_prepare_dst(&mgctx);
2903 spin_lock_irq(&css_set_lock);
2904 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2905 struct task_struct *task, *ntask;
2907 /* all tasks in src_csets need to be migrated */
2908 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2909 cgroup_migrate_add_task(task, &mgctx);
2911 spin_unlock_irq(&css_set_lock);
2913 ret = cgroup_migrate_execute(&mgctx);
2915 cgroup_migrate_finish(&mgctx);
2916 percpu_up_write(&cgroup_threadgroup_rwsem);
2921 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2922 * @cgrp: root of the target subtree
2924 * Because css offlining is asynchronous, userland may try to re-enable a
2925 * controller while the previous css is still around. This function grabs
2926 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2928 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2929 __acquires(&cgroup_mutex)
2931 struct cgroup *dsct;
2932 struct cgroup_subsys_state *d_css;
2933 struct cgroup_subsys *ss;
2937 mutex_lock(&cgroup_mutex);
2939 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2940 for_each_subsys(ss, ssid) {
2941 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2944 if (!css || !percpu_ref_is_dying(&css->refcnt))
2947 cgroup_get_live(dsct);
2948 prepare_to_wait(&dsct->offline_waitq, &wait,
2949 TASK_UNINTERRUPTIBLE);
2951 mutex_unlock(&cgroup_mutex);
2953 finish_wait(&dsct->offline_waitq, &wait);
2962 * cgroup_save_control - save control masks and dom_cgrp of a subtree
2963 * @cgrp: root of the target subtree
2965 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
2966 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2969 static void cgroup_save_control(struct cgroup *cgrp)
2971 struct cgroup *dsct;
2972 struct cgroup_subsys_state *d_css;
2974 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2975 dsct->old_subtree_control = dsct->subtree_control;
2976 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
2977 dsct->old_dom_cgrp = dsct->dom_cgrp;
2982 * cgroup_propagate_control - refresh control masks of a subtree
2983 * @cgrp: root of the target subtree
2985 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2986 * ->subtree_control and propagate controller availability through the
2987 * subtree so that descendants don't have unavailable controllers enabled.
2989 static void cgroup_propagate_control(struct cgroup *cgrp)
2991 struct cgroup *dsct;
2992 struct cgroup_subsys_state *d_css;
2994 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2995 dsct->subtree_control &= cgroup_control(dsct);
2996 dsct->subtree_ss_mask =
2997 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
2998 cgroup_ss_mask(dsct));
3003 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3004 * @cgrp: root of the target subtree
3006 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3007 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3010 static void cgroup_restore_control(struct cgroup *cgrp)
3012 struct cgroup *dsct;
3013 struct cgroup_subsys_state *d_css;
3015 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3016 dsct->subtree_control = dsct->old_subtree_control;
3017 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3018 dsct->dom_cgrp = dsct->old_dom_cgrp;
3022 static bool css_visible(struct cgroup_subsys_state *css)
3024 struct cgroup_subsys *ss = css->ss;
3025 struct cgroup *cgrp = css->cgroup;
3027 if (cgroup_control(cgrp) & (1 << ss->id))
3029 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3031 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3035 * cgroup_apply_control_enable - enable or show csses according to control
3036 * @cgrp: root of the target subtree
3038 * Walk @cgrp's subtree and create new csses or make the existing ones
3039 * visible. A css is created invisible if it's being implicitly enabled
3040 * through dependency. An invisible css is made visible when the userland
3041 * explicitly enables it.
3043 * Returns 0 on success, -errno on failure. On failure, csses which have
3044 * been processed already aren't cleaned up. The caller is responsible for
3045 * cleaning up with cgroup_apply_control_disable().
3047 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3049 struct cgroup *dsct;
3050 struct cgroup_subsys_state *d_css;
3051 struct cgroup_subsys *ss;
3054 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3055 for_each_subsys(ss, ssid) {
3056 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3058 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3060 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3064 css = css_create(dsct, ss);
3066 return PTR_ERR(css);
3069 if (css_visible(css)) {
3070 ret = css_populate_dir(css);
3081 * cgroup_apply_control_disable - kill or hide csses according to control
3082 * @cgrp: root of the target subtree
3084 * Walk @cgrp's subtree and kill and hide csses so that they match
3085 * cgroup_ss_mask() and cgroup_visible_mask().
3087 * A css is hidden when the userland requests it to be disabled while other
3088 * subsystems are still depending on it. The css must not actively control
3089 * resources and be in the vanilla state if it's made visible again later.
3090 * Controllers which may be depended upon should provide ->css_reset() for
3093 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3095 struct cgroup *dsct;
3096 struct cgroup_subsys_state *d_css;
3097 struct cgroup_subsys *ss;
3100 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3101 for_each_subsys(ss, ssid) {
3102 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3104 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3110 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3112 } else if (!css_visible(css)) {
3122 * cgroup_apply_control - apply control mask updates to the subtree
3123 * @cgrp: root of the target subtree
3125 * subsystems can be enabled and disabled in a subtree using the following
3128 * 1. Call cgroup_save_control() to stash the current state.
3129 * 2. Update ->subtree_control masks in the subtree as desired.
3130 * 3. Call cgroup_apply_control() to apply the changes.
3131 * 4. Optionally perform other related operations.
3132 * 5. Call cgroup_finalize_control() to finish up.
3134 * This function implements step 3 and propagates the mask changes
3135 * throughout @cgrp's subtree, updates csses accordingly and perform
3136 * process migrations.
3138 static int cgroup_apply_control(struct cgroup *cgrp)
3142 cgroup_propagate_control(cgrp);
3144 ret = cgroup_apply_control_enable(cgrp);
3149 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3150 * making the following cgroup_update_dfl_csses() properly update
3151 * css associations of all tasks in the subtree.
3153 ret = cgroup_update_dfl_csses(cgrp);
3161 * cgroup_finalize_control - finalize control mask update
3162 * @cgrp: root of the target subtree
3163 * @ret: the result of the update
3165 * Finalize control mask update. See cgroup_apply_control() for more info.
3167 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3170 cgroup_restore_control(cgrp);
3171 cgroup_propagate_control(cgrp);
3174 cgroup_apply_control_disable(cgrp);
3177 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3179 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3181 /* if nothing is getting enabled, nothing to worry about */
3185 /* can @cgrp host any resources? */
3186 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3189 /* mixables don't care */
3190 if (cgroup_is_mixable(cgrp))
3193 if (domain_enable) {
3194 /* can't enable domain controllers inside a thread subtree */
3195 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3199 * Threaded controllers can handle internal competitions
3200 * and are always allowed inside a (prospective) thread
3203 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3208 * Controllers can't be enabled for a cgroup with tasks to avoid
3209 * child cgroups competing against tasks.
3211 if (cgroup_has_tasks(cgrp))
3217 /* change the enabled child controllers for a cgroup in the default hierarchy */
3218 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3219 char *buf, size_t nbytes,
3222 u16 enable = 0, disable = 0;
3223 struct cgroup *cgrp, *child;
3224 struct cgroup_subsys *ss;
3229 * Parse input - space separated list of subsystem names prefixed
3230 * with either + or -.
3232 buf = strstrip(buf);
3233 while ((tok = strsep(&buf, " "))) {
3236 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3237 if (!cgroup_ssid_enabled(ssid) ||
3238 strcmp(tok + 1, ss->name))
3242 enable |= 1 << ssid;
3243 disable &= ~(1 << ssid);
3244 } else if (*tok == '-') {
3245 disable |= 1 << ssid;
3246 enable &= ~(1 << ssid);
3251 } while_each_subsys_mask();
3252 if (ssid == CGROUP_SUBSYS_COUNT)
3256 cgrp = cgroup_kn_lock_live(of->kn, true);
3260 for_each_subsys(ss, ssid) {
3261 if (enable & (1 << ssid)) {
3262 if (cgrp->subtree_control & (1 << ssid)) {
3263 enable &= ~(1 << ssid);
3267 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3271 } else if (disable & (1 << ssid)) {
3272 if (!(cgrp->subtree_control & (1 << ssid))) {
3273 disable &= ~(1 << ssid);
3277 /* a child has it enabled? */
3278 cgroup_for_each_live_child(child, cgrp) {
3279 if (child->subtree_control & (1 << ssid)) {
3287 if (!enable && !disable) {
3292 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3296 /* save and update control masks and prepare csses */
3297 cgroup_save_control(cgrp);
3299 cgrp->subtree_control |= enable;
3300 cgrp->subtree_control &= ~disable;
3302 ret = cgroup_apply_control(cgrp);
3303 cgroup_finalize_control(cgrp, ret);
3307 kernfs_activate(cgrp->kn);
3309 cgroup_kn_unlock(of->kn);
3310 return ret ?: nbytes;
3314 * cgroup_enable_threaded - make @cgrp threaded
3315 * @cgrp: the target cgroup
3317 * Called when "threaded" is written to the cgroup.type interface file and
3318 * tries to make @cgrp threaded and join the parent's resource domain.
3319 * This function is never called on the root cgroup as cgroup.type doesn't
3322 static int cgroup_enable_threaded(struct cgroup *cgrp)
3324 struct cgroup *parent = cgroup_parent(cgrp);
3325 struct cgroup *dom_cgrp = parent->dom_cgrp;
3326 struct cgroup *dsct;
3327 struct cgroup_subsys_state *d_css;
3330 lockdep_assert_held(&cgroup_mutex);
3332 /* noop if already threaded */
3333 if (cgroup_is_threaded(cgrp))
3337 * If @cgroup is populated or has domain controllers enabled, it
3338 * can't be switched. While the below cgroup_can_be_thread_root()
3339 * test can catch the same conditions, that's only when @parent is
3340 * not mixable, so let's check it explicitly.
3342 if (cgroup_is_populated(cgrp) ||
3343 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3346 /* we're joining the parent's domain, ensure its validity */
3347 if (!cgroup_is_valid_domain(dom_cgrp) ||
3348 !cgroup_can_be_thread_root(dom_cgrp))
3352 * The following shouldn't cause actual migrations and should
3355 cgroup_save_control(cgrp);
3357 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3358 if (dsct == cgrp || cgroup_is_threaded(dsct))
3359 dsct->dom_cgrp = dom_cgrp;
3361 ret = cgroup_apply_control(cgrp);
3363 parent->nr_threaded_children++;
3365 cgroup_finalize_control(cgrp, ret);
3369 static int cgroup_type_show(struct seq_file *seq, void *v)
3371 struct cgroup *cgrp = seq_css(seq)->cgroup;
3373 if (cgroup_is_threaded(cgrp))
3374 seq_puts(seq, "threaded\n");
3375 else if (!cgroup_is_valid_domain(cgrp))
3376 seq_puts(seq, "domain invalid\n");
3377 else if (cgroup_is_thread_root(cgrp))
3378 seq_puts(seq, "domain threaded\n");
3380 seq_puts(seq, "domain\n");
3385 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3386 size_t nbytes, loff_t off)
3388 struct cgroup *cgrp;
3391 /* only switching to threaded mode is supported */
3392 if (strcmp(strstrip(buf), "threaded"))
3395 cgrp = cgroup_kn_lock_live(of->kn, false);
3399 /* threaded can only be enabled */
3400 ret = cgroup_enable_threaded(cgrp);
3402 cgroup_kn_unlock(of->kn);
3403 return ret ?: nbytes;
3406 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3408 struct cgroup *cgrp = seq_css(seq)->cgroup;
3409 int descendants = READ_ONCE(cgrp->max_descendants);
3411 if (descendants == INT_MAX)
3412 seq_puts(seq, "max\n");
3414 seq_printf(seq, "%d\n", descendants);
3419 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3420 char *buf, size_t nbytes, loff_t off)
3422 struct cgroup *cgrp;
3426 buf = strstrip(buf);
3427 if (!strcmp(buf, "max")) {
3428 descendants = INT_MAX;
3430 ret = kstrtoint(buf, 0, &descendants);
3435 if (descendants < 0)
3438 cgrp = cgroup_kn_lock_live(of->kn, false);
3442 cgrp->max_descendants = descendants;
3444 cgroup_kn_unlock(of->kn);
3449 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3451 struct cgroup *cgrp = seq_css(seq)->cgroup;
3452 int depth = READ_ONCE(cgrp->max_depth);
3454 if (depth == INT_MAX)
3455 seq_puts(seq, "max\n");
3457 seq_printf(seq, "%d\n", depth);
3462 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3463 char *buf, size_t nbytes, loff_t off)
3465 struct cgroup *cgrp;
3469 buf = strstrip(buf);
3470 if (!strcmp(buf, "max")) {
3473 ret = kstrtoint(buf, 0, &depth);
3481 cgrp = cgroup_kn_lock_live(of->kn, false);
3485 cgrp->max_depth = depth;
3487 cgroup_kn_unlock(of->kn);
3492 static int cgroup_events_show(struct seq_file *seq, void *v)
3494 struct cgroup *cgrp = seq_css(seq)->cgroup;
3496 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3497 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3502 static int cgroup_stat_show(struct seq_file *seq, void *v)
3504 struct cgroup *cgroup = seq_css(seq)->cgroup;
3506 seq_printf(seq, "nr_descendants %d\n",
3507 cgroup->nr_descendants);
3508 seq_printf(seq, "nr_dying_descendants %d\n",
3509 cgroup->nr_dying_descendants);
3514 static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3515 struct cgroup *cgrp, int ssid)
3517 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3518 struct cgroup_subsys_state *css;
3521 if (!ss->css_extra_stat_show)
3524 css = cgroup_tryget_css(cgrp, ss);
3528 ret = ss->css_extra_stat_show(seq, css);
3533 static int cpu_stat_show(struct seq_file *seq, void *v)
3535 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3538 cgroup_base_stat_cputime_show(seq);
3539 #ifdef CONFIG_CGROUP_SCHED
3540 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3546 static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3548 struct cgroup *cgrp = seq_css(seq)->cgroup;
3549 struct psi_group *psi = cgroup_id(cgrp) == 1 ? &psi_system : &cgrp->psi;
3551 return psi_show(seq, psi, PSI_IO);
3553 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3555 struct cgroup *cgrp = seq_css(seq)->cgroup;
3556 struct psi_group *psi = cgroup_id(cgrp) == 1 ? &psi_system : &cgrp->psi;
3558 return psi_show(seq, psi, PSI_MEM);
3560 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3562 struct cgroup *cgrp = seq_css(seq)->cgroup;
3563 struct psi_group *psi = cgroup_id(cgrp) == 1 ? &psi_system : &cgrp->psi;
3565 return psi_show(seq, psi, PSI_CPU);
3568 static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf,
3569 size_t nbytes, enum psi_res res)
3571 struct psi_trigger *new;
3572 struct cgroup *cgrp;
3574 cgrp = cgroup_kn_lock_live(of->kn, false);
3579 cgroup_kn_unlock(of->kn);
3581 new = psi_trigger_create(&cgrp->psi, buf, nbytes, res);
3584 return PTR_ERR(new);
3587 psi_trigger_replace(&of->priv, new);
3594 static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3595 char *buf, size_t nbytes,
3598 return cgroup_pressure_write(of, buf, nbytes, PSI_IO);
3601 static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3602 char *buf, size_t nbytes,
3605 return cgroup_pressure_write(of, buf, nbytes, PSI_MEM);
3608 static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3609 char *buf, size_t nbytes,
3612 return cgroup_pressure_write(of, buf, nbytes, PSI_CPU);
3615 static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3618 return psi_trigger_poll(&of->priv, of->file, pt);
3621 static void cgroup_pressure_release(struct kernfs_open_file *of)
3623 psi_trigger_replace(&of->priv, NULL);
3625 #endif /* CONFIG_PSI */
3627 static int cgroup_freeze_show(struct seq_file *seq, void *v)
3629 struct cgroup *cgrp = seq_css(seq)->cgroup;
3631 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3636 static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3637 char *buf, size_t nbytes, loff_t off)
3639 struct cgroup *cgrp;
3643 ret = kstrtoint(strstrip(buf), 0, &freeze);
3647 if (freeze < 0 || freeze > 1)
3650 cgrp = cgroup_kn_lock_live(of->kn, false);
3654 cgroup_freeze(cgrp, freeze);
3656 cgroup_kn_unlock(of->kn);
3661 static int cgroup_file_open(struct kernfs_open_file *of)
3663 struct cftype *cft = of->kn->priv;
3666 return cft->open(of);
3670 static void cgroup_file_release(struct kernfs_open_file *of)
3672 struct cftype *cft = of->kn->priv;
3678 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3679 size_t nbytes, loff_t off)
3681 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3682 struct cgroup *cgrp = of->kn->parent->priv;
3683 struct cftype *cft = of->kn->priv;
3684 struct cgroup_subsys_state *css;
3688 * If namespaces are delegation boundaries, disallow writes to
3689 * files in an non-init namespace root from inside the namespace
3690 * except for the files explicitly marked delegatable -
3691 * cgroup.procs and cgroup.subtree_control.
3693 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3694 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3695 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3699 return cft->write(of, buf, nbytes, off);
3702 * kernfs guarantees that a file isn't deleted with operations in
3703 * flight, which means that the matching css is and stays alive and
3704 * doesn't need to be pinned. The RCU locking is not necessary
3705 * either. It's just for the convenience of using cgroup_css().
3708 css = cgroup_css(cgrp, cft->ss);
3711 if (cft->write_u64) {
3712 unsigned long long v;
3713 ret = kstrtoull(buf, 0, &v);
3715 ret = cft->write_u64(css, cft, v);
3716 } else if (cft->write_s64) {
3718 ret = kstrtoll(buf, 0, &v);
3720 ret = cft->write_s64(css, cft, v);
3725 return ret ?: nbytes;
3728 static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
3730 struct cftype *cft = of->kn->priv;
3733 return cft->poll(of, pt);
3735 return kernfs_generic_poll(of, pt);
3738 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3740 return seq_cft(seq)->seq_start(seq, ppos);
3743 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3745 return seq_cft(seq)->seq_next(seq, v, ppos);
3748 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3750 if (seq_cft(seq)->seq_stop)
3751 seq_cft(seq)->seq_stop(seq, v);
3754 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3756 struct cftype *cft = seq_cft(m);
3757 struct cgroup_subsys_state *css = seq_css(m);
3760 return cft->seq_show(m, arg);
3763 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3764 else if (cft->read_s64)
3765 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3771 static struct kernfs_ops cgroup_kf_single_ops = {
3772 .atomic_write_len = PAGE_SIZE,
3773 .open = cgroup_file_open,
3774 .release = cgroup_file_release,
3775 .write = cgroup_file_write,
3776 .poll = cgroup_file_poll,
3777 .seq_show = cgroup_seqfile_show,
3780 static struct kernfs_ops cgroup_kf_ops = {
3781 .atomic_write_len = PAGE_SIZE,
3782 .open = cgroup_file_open,
3783 .release = cgroup_file_release,
3784 .write = cgroup_file_write,
3785 .poll = cgroup_file_poll,
3786 .seq_start = cgroup_seqfile_start,
3787 .seq_next = cgroup_seqfile_next,
3788 .seq_stop = cgroup_seqfile_stop,
3789 .seq_show = cgroup_seqfile_show,
3792 /* set uid and gid of cgroup dirs and files to that of the creator */
3793 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3795 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3796 .ia_uid = current_fsuid(),
3797 .ia_gid = current_fsgid(), };
3799 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3800 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3803 return kernfs_setattr(kn, &iattr);
3806 static void cgroup_file_notify_timer(struct timer_list *timer)
3808 cgroup_file_notify(container_of(timer, struct cgroup_file,
3812 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3815 char name[CGROUP_FILE_NAME_MAX];
3816 struct kernfs_node *kn;
3817 struct lock_class_key *key = NULL;
3820 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3821 key = &cft->lockdep_key;
3823 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3824 cgroup_file_mode(cft),
3825 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
3826 0, cft->kf_ops, cft,
3831 ret = cgroup_kn_set_ugid(kn);
3837 if (cft->file_offset) {
3838 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3840 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
3842 spin_lock_irq(&cgroup_file_kn_lock);
3844 spin_unlock_irq(&cgroup_file_kn_lock);
3851 * cgroup_addrm_files - add or remove files to a cgroup directory
3852 * @css: the target css
3853 * @cgrp: the target cgroup (usually css->cgroup)
3854 * @cfts: array of cftypes to be added
3855 * @is_add: whether to add or remove
3857 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3858 * For removals, this function never fails.
3860 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3861 struct cgroup *cgrp, struct cftype cfts[],
3864 struct cftype *cft, *cft_end = NULL;
3867 lockdep_assert_held(&cgroup_mutex);
3870 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3871 /* does cft->flags tell us to skip this file on @cgrp? */
3872 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3874 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3876 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3878 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3880 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
3883 ret = cgroup_add_file(css, cgrp, cft);
3885 pr_warn("%s: failed to add %s, err=%d\n",
3886 __func__, cft->name, ret);
3892 cgroup_rm_file(cgrp, cft);
3898 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3900 struct cgroup_subsys *ss = cfts[0].ss;
3901 struct cgroup *root = &ss->root->cgrp;
3902 struct cgroup_subsys_state *css;
3905 lockdep_assert_held(&cgroup_mutex);
3907 /* add/rm files for all cgroups created before */
3908 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3909 struct cgroup *cgrp = css->cgroup;
3911 if (!(css->flags & CSS_VISIBLE))
3914 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3920 kernfs_activate(root->kn);
3924 static void cgroup_exit_cftypes(struct cftype *cfts)
3928 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3929 /* free copy for custom atomic_write_len, see init_cftypes() */
3930 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3935 /* revert flags set by cgroup core while adding @cfts */
3936 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3940 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3944 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3945 struct kernfs_ops *kf_ops;
3947 WARN_ON(cft->ss || cft->kf_ops);
3950 kf_ops = &cgroup_kf_ops;
3952 kf_ops = &cgroup_kf_single_ops;
3955 * Ugh... if @cft wants a custom max_write_len, we need to
3956 * make a copy of kf_ops to set its atomic_write_len.
3958 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3959 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3961 cgroup_exit_cftypes(cfts);
3964 kf_ops->atomic_write_len = cft->max_write_len;
3967 cft->kf_ops = kf_ops;
3974 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3976 lockdep_assert_held(&cgroup_mutex);
3978 if (!cfts || !cfts[0].ss)
3981 list_del(&cfts->node);
3982 cgroup_apply_cftypes(cfts, false);
3983 cgroup_exit_cftypes(cfts);
3988 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3989 * @cfts: zero-length name terminated array of cftypes
3991 * Unregister @cfts. Files described by @cfts are removed from all
3992 * existing cgroups and all future cgroups won't have them either. This
3993 * function can be called anytime whether @cfts' subsys is attached or not.
3995 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3998 int cgroup_rm_cftypes(struct cftype *cfts)
4002 mutex_lock(&cgroup_mutex);
4003 ret = cgroup_rm_cftypes_locked(cfts);
4004 mutex_unlock(&cgroup_mutex);
4009 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4010 * @ss: target cgroup subsystem
4011 * @cfts: zero-length name terminated array of cftypes
4013 * Register @cfts to @ss. Files described by @cfts are created for all
4014 * existing cgroups to which @ss is attached and all future cgroups will
4015 * have them too. This function can be called anytime whether @ss is
4018 * Returns 0 on successful registration, -errno on failure. Note that this
4019 * function currently returns 0 as long as @cfts registration is successful
4020 * even if some file creation attempts on existing cgroups fail.
4022 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4026 if (!cgroup_ssid_enabled(ss->id))
4029 if (!cfts || cfts[0].name[0] == '\0')
4032 ret = cgroup_init_cftypes(ss, cfts);
4036 mutex_lock(&cgroup_mutex);
4038 list_add_tail(&cfts->node, &ss->cfts);
4039 ret = cgroup_apply_cftypes(cfts, true);
4041 cgroup_rm_cftypes_locked(cfts);
4043 mutex_unlock(&cgroup_mutex);
4048 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4049 * @ss: target cgroup subsystem
4050 * @cfts: zero-length name terminated array of cftypes
4052 * Similar to cgroup_add_cftypes() but the added files are only used for
4053 * the default hierarchy.
4055 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4059 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4060 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4061 return cgroup_add_cftypes(ss, cfts);
4065 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4066 * @ss: target cgroup subsystem
4067 * @cfts: zero-length name terminated array of cftypes
4069 * Similar to cgroup_add_cftypes() but the added files are only used for
4070 * the legacy hierarchies.
4072 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4076 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4077 cft->flags |= __CFTYPE_NOT_ON_DFL;
4078 return cgroup_add_cftypes(ss, cfts);
4082 * cgroup_file_notify - generate a file modified event for a cgroup_file
4083 * @cfile: target cgroup_file
4085 * @cfile must have been obtained by setting cftype->file_offset.
4087 void cgroup_file_notify(struct cgroup_file *cfile)
4089 unsigned long flags;
4091 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4093 unsigned long last = cfile->notified_at;
4094 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4096 if (time_in_range(jiffies, last, next)) {
4097 timer_reduce(&cfile->notify_timer, next);
4099 kernfs_notify(cfile->kn);
4100 cfile->notified_at = jiffies;
4103 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4107 * css_next_child - find the next child of a given css
4108 * @pos: the current position (%NULL to initiate traversal)
4109 * @parent: css whose children to walk
4111 * This function returns the next child of @parent and should be called
4112 * under either cgroup_mutex or RCU read lock. The only requirement is
4113 * that @parent and @pos are accessible. The next sibling is guaranteed to
4114 * be returned regardless of their states.
4116 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4117 * css which finished ->css_online() is guaranteed to be visible in the
4118 * future iterations and will stay visible until the last reference is put.
4119 * A css which hasn't finished ->css_online() or already finished
4120 * ->css_offline() may show up during traversal. It's each subsystem's
4121 * responsibility to synchronize against on/offlining.
4123 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4124 struct cgroup_subsys_state *parent)
4126 struct cgroup_subsys_state *next;
4128 cgroup_assert_mutex_or_rcu_locked();
4131 * @pos could already have been unlinked from the sibling list.
4132 * Once a cgroup is removed, its ->sibling.next is no longer
4133 * updated when its next sibling changes. CSS_RELEASED is set when
4134 * @pos is taken off list, at which time its next pointer is valid,
4135 * and, as releases are serialized, the one pointed to by the next
4136 * pointer is guaranteed to not have started release yet. This
4137 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4138 * critical section, the one pointed to by its next pointer is
4139 * guaranteed to not have finished its RCU grace period even if we
4140 * have dropped rcu_read_lock() inbetween iterations.
4142 * If @pos has CSS_RELEASED set, its next pointer can't be
4143 * dereferenced; however, as each css is given a monotonically
4144 * increasing unique serial number and always appended to the
4145 * sibling list, the next one can be found by walking the parent's
4146 * children until the first css with higher serial number than
4147 * @pos's. While this path can be slower, it happens iff iteration
4148 * races against release and the race window is very small.
4151 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4152 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4153 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4155 list_for_each_entry_rcu(next, &parent->children, sibling)
4156 if (next->serial_nr > pos->serial_nr)
4161 * @next, if not pointing to the head, can be dereferenced and is
4164 if (&next->sibling != &parent->children)
4170 * css_next_descendant_pre - find the next descendant for pre-order walk
4171 * @pos: the current position (%NULL to initiate traversal)
4172 * @root: css whose descendants to walk
4174 * To be used by css_for_each_descendant_pre(). Find the next descendant
4175 * to visit for pre-order traversal of @root's descendants. @root is
4176 * included in the iteration and the first node to be visited.
4178 * While this function requires cgroup_mutex or RCU read locking, it
4179 * doesn't require the whole traversal to be contained in a single critical
4180 * section. This function will return the correct next descendant as long
4181 * as both @pos and @root are accessible and @pos is a descendant of @root.
4183 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4184 * css which finished ->css_online() is guaranteed to be visible in the
4185 * future iterations and will stay visible until the last reference is put.
4186 * A css which hasn't finished ->css_online() or already finished
4187 * ->css_offline() may show up during traversal. It's each subsystem's
4188 * responsibility to synchronize against on/offlining.
4190 struct cgroup_subsys_state *
4191 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4192 struct cgroup_subsys_state *root)
4194 struct cgroup_subsys_state *next;
4196 cgroup_assert_mutex_or_rcu_locked();
4198 /* if first iteration, visit @root */
4202 /* visit the first child if exists */
4203 next = css_next_child(NULL, pos);
4207 /* no child, visit my or the closest ancestor's next sibling */
4208 while (pos != root) {
4209 next = css_next_child(pos, pos->parent);
4217 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4220 * css_rightmost_descendant - return the rightmost descendant of a css
4221 * @pos: css of interest
4223 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4224 * is returned. This can be used during pre-order traversal to skip
4227 * While this function requires cgroup_mutex or RCU read locking, it
4228 * doesn't require the whole traversal to be contained in a single critical
4229 * section. This function will return the correct rightmost descendant as
4230 * long as @pos is accessible.
4232 struct cgroup_subsys_state *
4233 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4235 struct cgroup_subsys_state *last, *tmp;
4237 cgroup_assert_mutex_or_rcu_locked();
4241 /* ->prev isn't RCU safe, walk ->next till the end */
4243 css_for_each_child(tmp, last)
4250 static struct cgroup_subsys_state *
4251 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4253 struct cgroup_subsys_state *last;
4257 pos = css_next_child(NULL, pos);
4264 * css_next_descendant_post - find the next descendant for post-order walk
4265 * @pos: the current position (%NULL to initiate traversal)
4266 * @root: css whose descendants to walk
4268 * To be used by css_for_each_descendant_post(). Find the next descendant
4269 * to visit for post-order traversal of @root's descendants. @root is
4270 * included in the iteration and the last node to be visited.
4272 * While this function requires cgroup_mutex or RCU read locking, it
4273 * doesn't require the whole traversal to be contained in a single critical
4274 * section. This function will return the correct next descendant as long
4275 * as both @pos and @cgroup are accessible and @pos is a descendant of
4278 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4279 * css which finished ->css_online() is guaranteed to be visible in the
4280 * future iterations and will stay visible until the last reference is put.
4281 * A css which hasn't finished ->css_online() or already finished
4282 * ->css_offline() may show up during traversal. It's each subsystem's
4283 * responsibility to synchronize against on/offlining.
4285 struct cgroup_subsys_state *
4286 css_next_descendant_post(struct cgroup_subsys_state *pos,
4287 struct cgroup_subsys_state *root)
4289 struct cgroup_subsys_state *next;
4291 cgroup_assert_mutex_or_rcu_locked();
4293 /* if first iteration, visit leftmost descendant which may be @root */
4295 return css_leftmost_descendant(root);
4297 /* if we visited @root, we're done */
4301 /* if there's an unvisited sibling, visit its leftmost descendant */
4302 next = css_next_child(pos, pos->parent);
4304 return css_leftmost_descendant(next);
4306 /* no sibling left, visit parent */
4311 * css_has_online_children - does a css have online children
4312 * @css: the target css
4314 * Returns %true if @css has any online children; otherwise, %false. This
4315 * function can be called from any context but the caller is responsible
4316 * for synchronizing against on/offlining as necessary.
4318 bool css_has_online_children(struct cgroup_subsys_state *css)
4320 struct cgroup_subsys_state *child;
4324 css_for_each_child(child, css) {
4325 if (child->flags & CSS_ONLINE) {
4334 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4336 struct list_head *l;
4337 struct cgrp_cset_link *link;
4338 struct css_set *cset;
4340 lockdep_assert_held(&css_set_lock);
4342 /* find the next threaded cset */
4343 if (it->tcset_pos) {
4344 l = it->tcset_pos->next;
4346 if (l != it->tcset_head) {
4348 return container_of(l, struct css_set,
4349 threaded_csets_node);
4352 it->tcset_pos = NULL;
4355 /* find the next cset */
4358 if (l == it->cset_head) {
4359 it->cset_pos = NULL;
4364 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4366 link = list_entry(l, struct cgrp_cset_link, cset_link);
4372 /* initialize threaded css_set walking */
4373 if (it->flags & CSS_TASK_ITER_THREADED) {
4375 put_css_set_locked(it->cur_dcset);
4376 it->cur_dcset = cset;
4379 it->tcset_head = &cset->threaded_csets;
4380 it->tcset_pos = &cset->threaded_csets;
4387 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4388 * @it: the iterator to advance
4390 * Advance @it to the next css_set to walk.
4392 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4394 struct css_set *cset;
4396 lockdep_assert_held(&css_set_lock);
4398 /* Advance to the next non-empty css_set */
4400 cset = css_task_iter_next_css_set(it);
4402 it->task_pos = NULL;
4405 } while (!css_set_populated(cset) && list_empty(&cset->dying_tasks));
4407 if (!list_empty(&cset->tasks))
4408 it->task_pos = cset->tasks.next;
4409 else if (!list_empty(&cset->mg_tasks))
4410 it->task_pos = cset->mg_tasks.next;
4412 it->task_pos = cset->dying_tasks.next;
4414 it->tasks_head = &cset->tasks;
4415 it->mg_tasks_head = &cset->mg_tasks;
4416 it->dying_tasks_head = &cset->dying_tasks;
4419 * We don't keep css_sets locked across iteration steps and thus
4420 * need to take steps to ensure that iteration can be resumed after
4421 * the lock is re-acquired. Iteration is performed at two levels -
4422 * css_sets and tasks in them.
4424 * Once created, a css_set never leaves its cgroup lists, so a
4425 * pinned css_set is guaranteed to stay put and we can resume
4426 * iteration afterwards.
4428 * Tasks may leave @cset across iteration steps. This is resolved
4429 * by registering each iterator with the css_set currently being
4430 * walked and making css_set_move_task() advance iterators whose
4431 * next task is leaving.
4434 list_del(&it->iters_node);
4435 put_css_set_locked(it->cur_cset);
4438 it->cur_cset = cset;
4439 list_add(&it->iters_node, &cset->task_iters);
4442 static void css_task_iter_skip(struct css_task_iter *it,
4443 struct task_struct *task)
4445 lockdep_assert_held(&css_set_lock);
4447 if (it->task_pos == &task->cg_list) {
4448 it->task_pos = it->task_pos->next;
4449 it->flags |= CSS_TASK_ITER_SKIPPED;
4453 static void css_task_iter_advance(struct css_task_iter *it)
4455 struct task_struct *task;
4457 lockdep_assert_held(&css_set_lock);
4461 * Advance iterator to find next entry. cset->tasks is
4462 * consumed first and then ->mg_tasks. After ->mg_tasks,
4463 * we move onto the next cset.
4465 if (it->flags & CSS_TASK_ITER_SKIPPED)
4466 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4468 it->task_pos = it->task_pos->next;
4470 if (it->task_pos == it->tasks_head)
4471 it->task_pos = it->mg_tasks_head->next;
4472 if (it->task_pos == it->mg_tasks_head)
4473 it->task_pos = it->dying_tasks_head->next;
4474 if (it->task_pos == it->dying_tasks_head)
4475 css_task_iter_advance_css_set(it);
4477 /* called from start, proceed to the first cset */
4478 css_task_iter_advance_css_set(it);
4484 task = list_entry(it->task_pos, struct task_struct, cg_list);
4486 if (it->flags & CSS_TASK_ITER_PROCS) {
4487 /* if PROCS, skip over tasks which aren't group leaders */
4488 if (!thread_group_leader(task))
4491 /* and dying leaders w/o live member threads */
4492 if (!atomic_read(&task->signal->live))
4495 /* skip all dying ones */
4496 if (task->flags & PF_EXITING)
4502 * css_task_iter_start - initiate task iteration
4503 * @css: the css to walk tasks of
4504 * @flags: CSS_TASK_ITER_* flags
4505 * @it: the task iterator to use
4507 * Initiate iteration through the tasks of @css. The caller can call
4508 * css_task_iter_next() to walk through the tasks until the function
4509 * returns NULL. On completion of iteration, css_task_iter_end() must be
4512 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4513 struct css_task_iter *it)
4515 memset(it, 0, sizeof(*it));
4517 spin_lock_irq(&css_set_lock);
4523 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4525 it->cset_pos = &css->cgroup->cset_links;
4527 it->cset_head = it->cset_pos;
4529 css_task_iter_advance(it);
4531 spin_unlock_irq(&css_set_lock);
4535 * css_task_iter_next - return the next task for the iterator
4536 * @it: the task iterator being iterated
4538 * The "next" function for task iteration. @it should have been
4539 * initialized via css_task_iter_start(). Returns NULL when the iteration
4542 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4545 put_task_struct(it->cur_task);
4546 it->cur_task = NULL;
4549 spin_lock_irq(&css_set_lock);
4551 /* @it may be half-advanced by skips, finish advancing */
4552 if (it->flags & CSS_TASK_ITER_SKIPPED)
4553 css_task_iter_advance(it);
4556 it->cur_task = list_entry(it->task_pos, struct task_struct,
4558 get_task_struct(it->cur_task);
4559 css_task_iter_advance(it);
4562 spin_unlock_irq(&css_set_lock);
4564 return it->cur_task;
4568 * css_task_iter_end - finish task iteration
4569 * @it: the task iterator to finish
4571 * Finish task iteration started by css_task_iter_start().
4573 void css_task_iter_end(struct css_task_iter *it)
4576 spin_lock_irq(&css_set_lock);
4577 list_del(&it->iters_node);
4578 put_css_set_locked(it->cur_cset);
4579 spin_unlock_irq(&css_set_lock);
4583 put_css_set(it->cur_dcset);
4586 put_task_struct(it->cur_task);
4589 static void cgroup_procs_release(struct kernfs_open_file *of)
4592 css_task_iter_end(of->priv);
4597 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4599 struct kernfs_open_file *of = s->private;
4600 struct css_task_iter *it = of->priv;
4602 return css_task_iter_next(it);
4605 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4606 unsigned int iter_flags)
4608 struct kernfs_open_file *of = s->private;
4609 struct cgroup *cgrp = seq_css(s)->cgroup;
4610 struct css_task_iter *it = of->priv;
4613 * When a seq_file is seeked, it's always traversed sequentially
4614 * from position 0, so we can simply keep iterating on !0 *pos.
4617 if (WARN_ON_ONCE((*pos)++))
4618 return ERR_PTR(-EINVAL);
4620 it = kzalloc(sizeof(*it), GFP_KERNEL);
4622 return ERR_PTR(-ENOMEM);
4624 css_task_iter_start(&cgrp->self, iter_flags, it);
4625 } else if (!(*pos)++) {
4626 css_task_iter_end(it);
4627 css_task_iter_start(&cgrp->self, iter_flags, it);
4630 return cgroup_procs_next(s, NULL, NULL);
4633 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4635 struct cgroup *cgrp = seq_css(s)->cgroup;
4638 * All processes of a threaded subtree belong to the domain cgroup
4639 * of the subtree. Only threads can be distributed across the
4640 * subtree. Reject reads on cgroup.procs in the subtree proper.
4641 * They're always empty anyway.
4643 if (cgroup_is_threaded(cgrp))
4644 return ERR_PTR(-EOPNOTSUPP);
4646 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4647 CSS_TASK_ITER_THREADED);
4650 static int cgroup_procs_show(struct seq_file *s, void *v)
4652 seq_printf(s, "%d\n", task_pid_vnr(v));
4656 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4657 struct cgroup *dst_cgrp,
4658 struct super_block *sb)
4660 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4661 struct cgroup *com_cgrp = src_cgrp;
4662 struct inode *inode;
4665 lockdep_assert_held(&cgroup_mutex);
4667 /* find the common ancestor */
4668 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4669 com_cgrp = cgroup_parent(com_cgrp);
4671 /* %current should be authorized to migrate to the common ancestor */
4672 inode = kernfs_get_inode(sb, com_cgrp->procs_file.kn);
4676 ret = inode_permission(inode, MAY_WRITE);
4682 * If namespaces are delegation boundaries, %current must be able
4683 * to see both source and destination cgroups from its namespace.
4685 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4686 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4687 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4693 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4694 char *buf, size_t nbytes, loff_t off)
4696 struct cgroup *src_cgrp, *dst_cgrp;
4697 struct task_struct *task;
4701 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4705 task = cgroup_procs_write_start(buf, true, &locked);
4706 ret = PTR_ERR_OR_ZERO(task);
4710 /* find the source cgroup */
4711 spin_lock_irq(&css_set_lock);
4712 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4713 spin_unlock_irq(&css_set_lock);
4715 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4716 of->file->f_path.dentry->d_sb);
4720 ret = cgroup_attach_task(dst_cgrp, task, true);
4723 cgroup_procs_write_finish(task, locked);
4725 cgroup_kn_unlock(of->kn);
4727 return ret ?: nbytes;
4730 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4732 return __cgroup_procs_start(s, pos, 0);
4735 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4736 char *buf, size_t nbytes, loff_t off)
4738 struct cgroup *src_cgrp, *dst_cgrp;
4739 struct task_struct *task;
4743 buf = strstrip(buf);
4745 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4749 task = cgroup_procs_write_start(buf, false, &locked);
4750 ret = PTR_ERR_OR_ZERO(task);
4754 /* find the source cgroup */
4755 spin_lock_irq(&css_set_lock);
4756 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4757 spin_unlock_irq(&css_set_lock);
4759 /* thread migrations follow the cgroup.procs delegation rule */
4760 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4761 of->file->f_path.dentry->d_sb);
4765 /* and must be contained in the same domain */
4767 if (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)
4770 ret = cgroup_attach_task(dst_cgrp, task, false);
4773 cgroup_procs_write_finish(task, locked);
4775 cgroup_kn_unlock(of->kn);
4777 return ret ?: nbytes;
4780 /* cgroup core interface files for the default hierarchy */
4781 static struct cftype cgroup_base_files[] = {
4783 .name = "cgroup.type",
4784 .flags = CFTYPE_NOT_ON_ROOT,
4785 .seq_show = cgroup_type_show,
4786 .write = cgroup_type_write,
4789 .name = "cgroup.procs",
4790 .flags = CFTYPE_NS_DELEGATABLE,
4791 .file_offset = offsetof(struct cgroup, procs_file),
4792 .release = cgroup_procs_release,
4793 .seq_start = cgroup_procs_start,
4794 .seq_next = cgroup_procs_next,
4795 .seq_show = cgroup_procs_show,
4796 .write = cgroup_procs_write,
4799 .name = "cgroup.threads",
4800 .flags = CFTYPE_NS_DELEGATABLE,
4801 .release = cgroup_procs_release,
4802 .seq_start = cgroup_threads_start,
4803 .seq_next = cgroup_procs_next,
4804 .seq_show = cgroup_procs_show,
4805 .write = cgroup_threads_write,
4808 .name = "cgroup.controllers",
4809 .seq_show = cgroup_controllers_show,
4812 .name = "cgroup.subtree_control",
4813 .flags = CFTYPE_NS_DELEGATABLE,
4814 .seq_show = cgroup_subtree_control_show,
4815 .write = cgroup_subtree_control_write,
4818 .name = "cgroup.events",
4819 .flags = CFTYPE_NOT_ON_ROOT,
4820 .file_offset = offsetof(struct cgroup, events_file),
4821 .seq_show = cgroup_events_show,
4824 .name = "cgroup.max.descendants",
4825 .seq_show = cgroup_max_descendants_show,
4826 .write = cgroup_max_descendants_write,
4829 .name = "cgroup.max.depth",
4830 .seq_show = cgroup_max_depth_show,
4831 .write = cgroup_max_depth_write,
4834 .name = "cgroup.stat",
4835 .seq_show = cgroup_stat_show,
4838 .name = "cgroup.freeze",
4839 .flags = CFTYPE_NOT_ON_ROOT,
4840 .seq_show = cgroup_freeze_show,
4841 .write = cgroup_freeze_write,
4845 .flags = CFTYPE_NOT_ON_ROOT,
4846 .seq_show = cpu_stat_show,
4850 .name = "io.pressure",
4851 .seq_show = cgroup_io_pressure_show,
4852 .write = cgroup_io_pressure_write,
4853 .poll = cgroup_pressure_poll,
4854 .release = cgroup_pressure_release,
4857 .name = "memory.pressure",
4858 .seq_show = cgroup_memory_pressure_show,
4859 .write = cgroup_memory_pressure_write,
4860 .poll = cgroup_pressure_poll,
4861 .release = cgroup_pressure_release,
4864 .name = "cpu.pressure",
4865 .seq_show = cgroup_cpu_pressure_show,
4866 .write = cgroup_cpu_pressure_write,
4867 .poll = cgroup_pressure_poll,
4868 .release = cgroup_pressure_release,
4870 #endif /* CONFIG_PSI */
4875 * css destruction is four-stage process.
4877 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4878 * Implemented in kill_css().
4880 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4881 * and thus css_tryget_online() is guaranteed to fail, the css can be
4882 * offlined by invoking offline_css(). After offlining, the base ref is
4883 * put. Implemented in css_killed_work_fn().
4885 * 3. When the percpu_ref reaches zero, the only possible remaining
4886 * accessors are inside RCU read sections. css_release() schedules the
4889 * 4. After the grace period, the css can be freed. Implemented in
4890 * css_free_work_fn().
4892 * It is actually hairier because both step 2 and 4 require process context
4893 * and thus involve punting to css->destroy_work adding two additional
4894 * steps to the already complex sequence.
4896 static void css_free_rwork_fn(struct work_struct *work)
4898 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
4899 struct cgroup_subsys_state, destroy_rwork);
4900 struct cgroup_subsys *ss = css->ss;
4901 struct cgroup *cgrp = css->cgroup;
4903 percpu_ref_exit(&css->refcnt);
4907 struct cgroup_subsys_state *parent = css->parent;
4911 cgroup_idr_remove(&ss->css_idr, id);
4917 /* cgroup free path */
4918 atomic_dec(&cgrp->root->nr_cgrps);
4919 cgroup1_pidlist_destroy_all(cgrp);
4920 cancel_work_sync(&cgrp->release_agent_work);
4922 if (cgroup_parent(cgrp)) {
4924 * We get a ref to the parent, and put the ref when
4925 * this cgroup is being freed, so it's guaranteed
4926 * that the parent won't be destroyed before its
4929 cgroup_put(cgroup_parent(cgrp));
4930 kernfs_put(cgrp->kn);
4931 psi_cgroup_free(cgrp);
4932 if (cgroup_on_dfl(cgrp))
4933 cgroup_rstat_exit(cgrp);
4937 * This is root cgroup's refcnt reaching zero,
4938 * which indicates that the root should be
4941 cgroup_destroy_root(cgrp->root);
4946 static void css_release_work_fn(struct work_struct *work)
4948 struct cgroup_subsys_state *css =
4949 container_of(work, struct cgroup_subsys_state, destroy_work);
4950 struct cgroup_subsys *ss = css->ss;
4951 struct cgroup *cgrp = css->cgroup;
4953 mutex_lock(&cgroup_mutex);
4955 css->flags |= CSS_RELEASED;
4956 list_del_rcu(&css->sibling);
4959 /* css release path */
4960 if (!list_empty(&css->rstat_css_node)) {
4961 cgroup_rstat_flush(cgrp);
4962 list_del_rcu(&css->rstat_css_node);
4965 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4966 if (ss->css_released)
4967 ss->css_released(css);
4969 struct cgroup *tcgrp;
4971 /* cgroup release path */
4972 TRACE_CGROUP_PATH(release, cgrp);
4974 if (cgroup_on_dfl(cgrp))
4975 cgroup_rstat_flush(cgrp);
4977 spin_lock_irq(&css_set_lock);
4978 for (tcgrp = cgroup_parent(cgrp); tcgrp;
4979 tcgrp = cgroup_parent(tcgrp))
4980 tcgrp->nr_dying_descendants--;
4981 spin_unlock_irq(&css_set_lock);
4984 * There are two control paths which try to determine
4985 * cgroup from dentry without going through kernfs -
4986 * cgroupstats_build() and css_tryget_online_from_dir().
4987 * Those are supported by RCU protecting clearing of
4988 * cgrp->kn->priv backpointer.
4991 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4995 mutex_unlock(&cgroup_mutex);
4997 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
4998 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5001 static void css_release(struct percpu_ref *ref)
5003 struct cgroup_subsys_state *css =
5004 container_of(ref, struct cgroup_subsys_state, refcnt);
5006 INIT_WORK(&css->destroy_work, css_release_work_fn);
5007 queue_work(cgroup_destroy_wq, &css->destroy_work);
5010 static void init_and_link_css(struct cgroup_subsys_state *css,
5011 struct cgroup_subsys *ss, struct cgroup *cgrp)
5013 lockdep_assert_held(&cgroup_mutex);
5015 cgroup_get_live(cgrp);
5017 memset(css, 0, sizeof(*css));
5021 INIT_LIST_HEAD(&css->sibling);
5022 INIT_LIST_HEAD(&css->children);
5023 INIT_LIST_HEAD(&css->rstat_css_node);
5024 css->serial_nr = css_serial_nr_next++;
5025 atomic_set(&css->online_cnt, 0);
5027 if (cgroup_parent(cgrp)) {
5028 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5029 css_get(css->parent);
5032 if (cgroup_on_dfl(cgrp) && ss->css_rstat_flush)
5033 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5035 BUG_ON(cgroup_css(cgrp, ss));
5038 /* invoke ->css_online() on a new CSS and mark it online if successful */
5039 static int online_css(struct cgroup_subsys_state *css)
5041 struct cgroup_subsys *ss = css->ss;
5044 lockdep_assert_held(&cgroup_mutex);
5047 ret = ss->css_online(css);
5049 css->flags |= CSS_ONLINE;
5050 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5052 atomic_inc(&css->online_cnt);
5054 atomic_inc(&css->parent->online_cnt);
5059 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5060 static void offline_css(struct cgroup_subsys_state *css)
5062 struct cgroup_subsys *ss = css->ss;
5064 lockdep_assert_held(&cgroup_mutex);
5066 if (!(css->flags & CSS_ONLINE))
5069 if (ss->css_offline)
5070 ss->css_offline(css);
5072 css->flags &= ~CSS_ONLINE;
5073 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5075 wake_up_all(&css->cgroup->offline_waitq);
5079 * css_create - create a cgroup_subsys_state
5080 * @cgrp: the cgroup new css will be associated with
5081 * @ss: the subsys of new css
5083 * Create a new css associated with @cgrp - @ss pair. On success, the new
5084 * css is online and installed in @cgrp. This function doesn't create the
5085 * interface files. Returns 0 on success, -errno on failure.
5087 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5088 struct cgroup_subsys *ss)
5090 struct cgroup *parent = cgroup_parent(cgrp);
5091 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5092 struct cgroup_subsys_state *css;
5095 lockdep_assert_held(&cgroup_mutex);
5097 css = ss->css_alloc(parent_css);
5099 css = ERR_PTR(-ENOMEM);
5103 init_and_link_css(css, ss, cgrp);
5105 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5109 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5114 /* @css is ready to be brought online now, make it visible */
5115 list_add_tail_rcu(&css->sibling, &parent_css->children);
5116 cgroup_idr_replace(&ss->css_idr, css, css->id);
5118 err = online_css(css);
5122 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5123 cgroup_parent(parent)) {
5124 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5125 current->comm, current->pid, ss->name);
5126 if (!strcmp(ss->name, "memory"))
5127 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5128 ss->warned_broken_hierarchy = true;
5134 list_del_rcu(&css->sibling);
5136 list_del_rcu(&css->rstat_css_node);
5137 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5138 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5139 return ERR_PTR(err);
5143 * The returned cgroup is fully initialized including its control mask, but
5144 * it isn't associated with its kernfs_node and doesn't have the control
5147 static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5150 struct cgroup_root *root = parent->root;
5151 struct cgroup *cgrp, *tcgrp;
5152 struct kernfs_node *kn;
5153 int level = parent->level + 1;
5156 /* allocate the cgroup and its ID, 0 is reserved for the root */
5157 cgrp = kzalloc(struct_size(cgrp, ancestor_ids, (level + 1)),
5160 return ERR_PTR(-ENOMEM);
5162 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5166 if (cgroup_on_dfl(parent)) {
5167 ret = cgroup_rstat_init(cgrp);
5169 goto out_cancel_ref;
5172 /* create the directory */
5173 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5180 init_cgroup_housekeeping(cgrp);
5182 cgrp->self.parent = &parent->self;
5184 cgrp->level = level;
5186 ret = psi_cgroup_alloc(cgrp);
5188 goto out_kernfs_remove;
5190 ret = cgroup_bpf_inherit(cgrp);
5195 * New cgroup inherits effective freeze counter, and
5196 * if the parent has to be frozen, the child has too.
5198 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5199 if (cgrp->freezer.e_freeze) {
5201 * Set the CGRP_FREEZE flag, so when a process will be
5202 * attached to the child cgroup, it will become frozen.
5203 * At this point the new cgroup is unpopulated, so we can
5204 * consider it frozen immediately.
5206 set_bit(CGRP_FREEZE, &cgrp->flags);
5207 set_bit(CGRP_FROZEN, &cgrp->flags);
5210 spin_lock_irq(&css_set_lock);
5211 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5212 cgrp->ancestor_ids[tcgrp->level] = cgroup_id(tcgrp);
5214 if (tcgrp != cgrp) {
5215 tcgrp->nr_descendants++;
5218 * If the new cgroup is frozen, all ancestor cgroups
5219 * get a new frozen descendant, but their state can't
5220 * change because of this.
5222 if (cgrp->freezer.e_freeze)
5223 tcgrp->freezer.nr_frozen_descendants++;
5226 spin_unlock_irq(&css_set_lock);
5228 if (notify_on_release(parent))
5229 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5231 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5232 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5234 cgrp->self.serial_nr = css_serial_nr_next++;
5236 /* allocation complete, commit to creation */
5237 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5238 atomic_inc(&root->nr_cgrps);
5239 cgroup_get_live(parent);
5242 * On the default hierarchy, a child doesn't automatically inherit
5243 * subtree_control from the parent. Each is configured manually.
5245 if (!cgroup_on_dfl(cgrp))
5246 cgrp->subtree_control = cgroup_control(cgrp);
5248 cgroup_propagate_control(cgrp);
5253 psi_cgroup_free(cgrp);
5255 kernfs_remove(cgrp->kn);
5257 if (cgroup_on_dfl(parent))
5258 cgroup_rstat_exit(cgrp);
5260 percpu_ref_exit(&cgrp->self.refcnt);
5263 return ERR_PTR(ret);
5266 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5268 struct cgroup *cgroup;
5272 lockdep_assert_held(&cgroup_mutex);
5274 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5275 if (cgroup->nr_descendants >= cgroup->max_descendants)
5278 if (level > cgroup->max_depth)
5289 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5291 struct cgroup *parent, *cgrp;
5294 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5295 if (strchr(name, '\n'))
5298 parent = cgroup_kn_lock_live(parent_kn, false);
5302 if (!cgroup_check_hierarchy_limits(parent)) {
5307 cgrp = cgroup_create(parent, name, mode);
5309 ret = PTR_ERR(cgrp);
5314 * This extra ref will be put in cgroup_free_fn() and guarantees
5315 * that @cgrp->kn is always accessible.
5317 kernfs_get(cgrp->kn);
5319 ret = cgroup_kn_set_ugid(cgrp->kn);
5323 ret = css_populate_dir(&cgrp->self);
5327 ret = cgroup_apply_control_enable(cgrp);
5331 TRACE_CGROUP_PATH(mkdir, cgrp);
5333 /* let's create and online css's */
5334 kernfs_activate(cgrp->kn);
5340 cgroup_destroy_locked(cgrp);
5342 cgroup_kn_unlock(parent_kn);
5347 * This is called when the refcnt of a css is confirmed to be killed.
5348 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5349 * initate destruction and put the css ref from kill_css().
5351 static void css_killed_work_fn(struct work_struct *work)
5353 struct cgroup_subsys_state *css =
5354 container_of(work, struct cgroup_subsys_state, destroy_work);
5356 mutex_lock(&cgroup_mutex);
5361 /* @css can't go away while we're holding cgroup_mutex */
5363 } while (css && atomic_dec_and_test(&css->online_cnt));
5365 mutex_unlock(&cgroup_mutex);
5368 /* css kill confirmation processing requires process context, bounce */
5369 static void css_killed_ref_fn(struct percpu_ref *ref)
5371 struct cgroup_subsys_state *css =
5372 container_of(ref, struct cgroup_subsys_state, refcnt);
5374 if (atomic_dec_and_test(&css->online_cnt)) {
5375 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5376 queue_work(cgroup_destroy_wq, &css->destroy_work);
5381 * kill_css - destroy a css
5382 * @css: css to destroy
5384 * This function initiates destruction of @css by removing cgroup interface
5385 * files and putting its base reference. ->css_offline() will be invoked
5386 * asynchronously once css_tryget_online() is guaranteed to fail and when
5387 * the reference count reaches zero, @css will be released.
5389 static void kill_css(struct cgroup_subsys_state *css)
5391 lockdep_assert_held(&cgroup_mutex);
5393 if (css->flags & CSS_DYING)
5396 css->flags |= CSS_DYING;
5399 * This must happen before css is disassociated with its cgroup.
5400 * See seq_css() for details.
5405 * Killing would put the base ref, but we need to keep it alive
5406 * until after ->css_offline().
5411 * cgroup core guarantees that, by the time ->css_offline() is
5412 * invoked, no new css reference will be given out via
5413 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5414 * proceed to offlining css's because percpu_ref_kill() doesn't
5415 * guarantee that the ref is seen as killed on all CPUs on return.
5417 * Use percpu_ref_kill_and_confirm() to get notifications as each
5418 * css is confirmed to be seen as killed on all CPUs.
5420 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5424 * cgroup_destroy_locked - the first stage of cgroup destruction
5425 * @cgrp: cgroup to be destroyed
5427 * css's make use of percpu refcnts whose killing latency shouldn't be
5428 * exposed to userland and are RCU protected. Also, cgroup core needs to
5429 * guarantee that css_tryget_online() won't succeed by the time
5430 * ->css_offline() is invoked. To satisfy all the requirements,
5431 * destruction is implemented in the following two steps.
5433 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5434 * userland visible parts and start killing the percpu refcnts of
5435 * css's. Set up so that the next stage will be kicked off once all
5436 * the percpu refcnts are confirmed to be killed.
5438 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5439 * rest of destruction. Once all cgroup references are gone, the
5440 * cgroup is RCU-freed.
5442 * This function implements s1. After this step, @cgrp is gone as far as
5443 * the userland is concerned and a new cgroup with the same name may be
5444 * created. As cgroup doesn't care about the names internally, this
5445 * doesn't cause any problem.
5447 static int cgroup_destroy_locked(struct cgroup *cgrp)
5448 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5450 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5451 struct cgroup_subsys_state *css;
5452 struct cgrp_cset_link *link;
5455 lockdep_assert_held(&cgroup_mutex);
5458 * Only migration can raise populated from zero and we're already
5459 * holding cgroup_mutex.
5461 if (cgroup_is_populated(cgrp))
5465 * Make sure there's no live children. We can't test emptiness of
5466 * ->self.children as dead children linger on it while being
5467 * drained; otherwise, "rmdir parent/child parent" may fail.
5469 if (css_has_online_children(&cgrp->self))
5473 * Mark @cgrp and the associated csets dead. The former prevents
5474 * further task migration and child creation by disabling
5475 * cgroup_lock_live_group(). The latter makes the csets ignored by
5476 * the migration path.
5478 cgrp->self.flags &= ~CSS_ONLINE;
5480 spin_lock_irq(&css_set_lock);
5481 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5482 link->cset->dead = true;
5483 spin_unlock_irq(&css_set_lock);
5485 /* initiate massacre of all css's */
5486 for_each_css(css, ssid, cgrp)
5489 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5490 css_clear_dir(&cgrp->self);
5491 kernfs_remove(cgrp->kn);
5493 if (parent && cgroup_is_threaded(cgrp))
5494 parent->nr_threaded_children--;
5496 spin_lock_irq(&css_set_lock);
5497 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5498 tcgrp->nr_descendants--;
5499 tcgrp->nr_dying_descendants++;
5501 * If the dying cgroup is frozen, decrease frozen descendants
5502 * counters of ancestor cgroups.
5504 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5505 tcgrp->freezer.nr_frozen_descendants--;
5507 spin_unlock_irq(&css_set_lock);
5509 cgroup1_check_for_release(parent);
5511 cgroup_bpf_offline(cgrp);
5513 /* put the base reference */
5514 percpu_ref_kill(&cgrp->self.refcnt);
5519 int cgroup_rmdir(struct kernfs_node *kn)
5521 struct cgroup *cgrp;
5524 cgrp = cgroup_kn_lock_live(kn, false);
5528 ret = cgroup_destroy_locked(cgrp);
5530 TRACE_CGROUP_PATH(rmdir, cgrp);
5532 cgroup_kn_unlock(kn);
5536 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5537 .show_options = cgroup_show_options,
5538 .mkdir = cgroup_mkdir,
5539 .rmdir = cgroup_rmdir,
5540 .show_path = cgroup_show_path,
5543 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5545 struct cgroup_subsys_state *css;
5547 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5549 mutex_lock(&cgroup_mutex);
5551 idr_init(&ss->css_idr);
5552 INIT_LIST_HEAD(&ss->cfts);
5554 /* Create the root cgroup state for this subsystem */
5555 ss->root = &cgrp_dfl_root;
5556 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5557 /* We don't handle early failures gracefully */
5558 BUG_ON(IS_ERR(css));
5559 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5562 * Root csses are never destroyed and we can't initialize
5563 * percpu_ref during early init. Disable refcnting.
5565 css->flags |= CSS_NO_REF;
5568 /* allocation can't be done safely during early init */
5571 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5572 BUG_ON(css->id < 0);
5575 /* Update the init_css_set to contain a subsys
5576 * pointer to this state - since the subsystem is
5577 * newly registered, all tasks and hence the
5578 * init_css_set is in the subsystem's root cgroup. */
5579 init_css_set.subsys[ss->id] = css;
5581 have_fork_callback |= (bool)ss->fork << ss->id;
5582 have_exit_callback |= (bool)ss->exit << ss->id;
5583 have_release_callback |= (bool)ss->release << ss->id;
5584 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5586 /* At system boot, before all subsystems have been
5587 * registered, no tasks have been forked, so we don't
5588 * need to invoke fork callbacks here. */
5589 BUG_ON(!list_empty(&init_task.tasks));
5591 BUG_ON(online_css(css));
5593 mutex_unlock(&cgroup_mutex);
5597 * cgroup_init_early - cgroup initialization at system boot
5599 * Initialize cgroups at system boot, and initialize any
5600 * subsystems that request early init.
5602 int __init cgroup_init_early(void)
5604 static struct cgroup_fs_context __initdata ctx;
5605 struct cgroup_subsys *ss;
5608 ctx.root = &cgrp_dfl_root;
5609 init_cgroup_root(&ctx);
5610 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5612 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5614 for_each_subsys(ss, i) {
5615 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5616 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5617 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5619 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5620 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5623 ss->name = cgroup_subsys_name[i];
5624 if (!ss->legacy_name)
5625 ss->legacy_name = cgroup_subsys_name[i];
5628 cgroup_init_subsys(ss, true);
5633 static u16 cgroup_disable_mask __initdata;
5636 * cgroup_init - cgroup initialization
5638 * Register cgroup filesystem and /proc file, and initialize
5639 * any subsystems that didn't request early init.
5641 int __init cgroup_init(void)
5643 struct cgroup_subsys *ss;
5646 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5647 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5648 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5650 cgroup_rstat_boot();
5653 * The latency of the synchronize_rcu() is too high for cgroups,
5654 * avoid it at the cost of forcing all readers into the slow path.
5656 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5658 get_user_ns(init_cgroup_ns.user_ns);
5660 mutex_lock(&cgroup_mutex);
5663 * Add init_css_set to the hash table so that dfl_root can link to
5666 hash_add(css_set_table, &init_css_set.hlist,
5667 css_set_hash(init_css_set.subsys));
5669 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5671 mutex_unlock(&cgroup_mutex);
5673 for_each_subsys(ss, ssid) {
5674 if (ss->early_init) {
5675 struct cgroup_subsys_state *css =
5676 init_css_set.subsys[ss->id];
5678 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5680 BUG_ON(css->id < 0);
5682 cgroup_init_subsys(ss, false);
5685 list_add_tail(&init_css_set.e_cset_node[ssid],
5686 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5689 * Setting dfl_root subsys_mask needs to consider the
5690 * disabled flag and cftype registration needs kmalloc,
5691 * both of which aren't available during early_init.
5693 if (cgroup_disable_mask & (1 << ssid)) {
5694 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5695 printk(KERN_INFO "Disabling %s control group subsystem\n",
5700 if (cgroup1_ssid_disabled(ssid))
5701 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5704 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5706 /* implicit controllers must be threaded too */
5707 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5709 if (ss->implicit_on_dfl)
5710 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5711 else if (!ss->dfl_cftypes)
5712 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5715 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5717 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5718 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5720 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5721 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5725 ss->bind(init_css_set.subsys[ssid]);
5727 mutex_lock(&cgroup_mutex);
5728 css_populate_dir(init_css_set.subsys[ssid]);
5729 mutex_unlock(&cgroup_mutex);
5732 /* init_css_set.subsys[] has been updated, re-hash */
5733 hash_del(&init_css_set.hlist);
5734 hash_add(css_set_table, &init_css_set.hlist,
5735 css_set_hash(init_css_set.subsys));
5737 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5738 WARN_ON(register_filesystem(&cgroup_fs_type));
5739 WARN_ON(register_filesystem(&cgroup2_fs_type));
5740 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
5741 #ifdef CONFIG_CPUSETS
5742 WARN_ON(register_filesystem(&cpuset_fs_type));
5748 static int __init cgroup_wq_init(void)
5751 * There isn't much point in executing destruction path in
5752 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5753 * Use 1 for @max_active.
5755 * We would prefer to do this in cgroup_init() above, but that
5756 * is called before init_workqueues(): so leave this until after.
5758 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5759 BUG_ON(!cgroup_destroy_wq);
5762 core_initcall(cgroup_wq_init);
5764 void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
5766 struct kernfs_node *kn;
5768 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
5771 kernfs_path(kn, buf, buflen);
5776 * proc_cgroup_show()
5777 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5778 * - Used for /proc/<pid>/cgroup.
5780 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5781 struct pid *pid, struct task_struct *tsk)
5785 struct cgroup_root *root;
5788 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5792 mutex_lock(&cgroup_mutex);
5793 spin_lock_irq(&css_set_lock);
5795 for_each_root(root) {
5796 struct cgroup_subsys *ss;
5797 struct cgroup *cgrp;
5798 int ssid, count = 0;
5800 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5803 seq_printf(m, "%d:", root->hierarchy_id);
5804 if (root != &cgrp_dfl_root)
5805 for_each_subsys(ss, ssid)
5806 if (root->subsys_mask & (1 << ssid))
5807 seq_printf(m, "%s%s", count++ ? "," : "",
5809 if (strlen(root->name))
5810 seq_printf(m, "%sname=%s", count ? "," : "",
5814 cgrp = task_cgroup_from_root(tsk, root);
5817 * On traditional hierarchies, all zombie tasks show up as
5818 * belonging to the root cgroup. On the default hierarchy,
5819 * while a zombie doesn't show up in "cgroup.procs" and
5820 * thus can't be migrated, its /proc/PID/cgroup keeps
5821 * reporting the cgroup it belonged to before exiting. If
5822 * the cgroup is removed before the zombie is reaped,
5823 * " (deleted)" is appended to the cgroup path.
5825 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5826 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5827 current->nsproxy->cgroup_ns);
5828 if (retval >= PATH_MAX)
5829 retval = -ENAMETOOLONG;
5838 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5839 seq_puts(m, " (deleted)\n");
5846 spin_unlock_irq(&css_set_lock);
5847 mutex_unlock(&cgroup_mutex);
5854 * cgroup_fork - initialize cgroup related fields during copy_process()
5855 * @child: pointer to task_struct of forking parent process.
5857 * A task is associated with the init_css_set until cgroup_post_fork()
5858 * attaches it to the parent's css_set. Empty cg_list indicates that
5859 * @child isn't holding reference to its css_set.
5861 void cgroup_fork(struct task_struct *child)
5863 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5864 INIT_LIST_HEAD(&child->cg_list);
5868 * cgroup_can_fork - called on a new task before the process is exposed
5869 * @child: the task in question.
5871 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5872 * returns an error, the fork aborts with that error code. This allows for
5873 * a cgroup subsystem to conditionally allow or deny new forks.
5875 int cgroup_can_fork(struct task_struct *child)
5877 struct cgroup_subsys *ss;
5880 do_each_subsys_mask(ss, i, have_canfork_callback) {
5881 ret = ss->can_fork(child);
5884 } while_each_subsys_mask();
5889 for_each_subsys(ss, j) {
5892 if (ss->cancel_fork)
5893 ss->cancel_fork(child);
5900 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5901 * @child: the task in question
5903 * This calls the cancel_fork() callbacks if a fork failed *after*
5904 * cgroup_can_fork() succeded.
5906 void cgroup_cancel_fork(struct task_struct *child)
5908 struct cgroup_subsys *ss;
5911 for_each_subsys(ss, i)
5912 if (ss->cancel_fork)
5913 ss->cancel_fork(child);
5917 * cgroup_post_fork - called on a new task after adding it to the task list
5918 * @child: the task in question
5920 * Adds the task to the list running through its css_set if necessary and
5921 * call the subsystem fork() callbacks. Has to be after the task is
5922 * visible on the task list in case we race with the first call to
5923 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5926 void cgroup_post_fork(struct task_struct *child)
5928 struct cgroup_subsys *ss;
5929 struct css_set *cset;
5932 spin_lock_irq(&css_set_lock);
5934 WARN_ON_ONCE(!list_empty(&child->cg_list));
5935 cset = task_css_set(current); /* current is @child's parent */
5938 css_set_move_task(child, NULL, cset, false);
5941 * If the cgroup has to be frozen, the new task has too. Let's set
5942 * the JOBCTL_TRAP_FREEZE jobctl bit to get the task into the
5945 if (unlikely(cgroup_task_freeze(child))) {
5946 spin_lock(&child->sighand->siglock);
5947 WARN_ON_ONCE(child->frozen);
5948 child->jobctl |= JOBCTL_TRAP_FREEZE;
5949 spin_unlock(&child->sighand->siglock);
5952 * Calling cgroup_update_frozen() isn't required here,
5953 * because it will be called anyway a bit later from
5954 * do_freezer_trap(). So we avoid cgroup's transient switch
5955 * from the frozen state and back.
5959 spin_unlock_irq(&css_set_lock);
5962 * Call ss->fork(). This must happen after @child is linked on
5963 * css_set; otherwise, @child might change state between ->fork()
5964 * and addition to css_set.
5966 do_each_subsys_mask(ss, i, have_fork_callback) {
5968 } while_each_subsys_mask();
5972 * cgroup_exit - detach cgroup from exiting task
5973 * @tsk: pointer to task_struct of exiting process
5975 * Description: Detach cgroup from @tsk.
5978 void cgroup_exit(struct task_struct *tsk)
5980 struct cgroup_subsys *ss;
5981 struct css_set *cset;
5984 spin_lock_irq(&css_set_lock);
5986 WARN_ON_ONCE(list_empty(&tsk->cg_list));
5987 cset = task_css_set(tsk);
5988 css_set_move_task(tsk, cset, NULL, false);
5989 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
5992 WARN_ON_ONCE(cgroup_task_frozen(tsk));
5993 if (unlikely(cgroup_task_freeze(tsk)))
5994 cgroup_update_frozen(task_dfl_cgroup(tsk));
5996 spin_unlock_irq(&css_set_lock);
5998 /* see cgroup_post_fork() for details */
5999 do_each_subsys_mask(ss, i, have_exit_callback) {
6001 } while_each_subsys_mask();
6004 void cgroup_release(struct task_struct *task)
6006 struct cgroup_subsys *ss;
6009 do_each_subsys_mask(ss, ssid, have_release_callback) {
6011 } while_each_subsys_mask();
6013 spin_lock_irq(&css_set_lock);
6014 css_set_skip_task_iters(task_css_set(task), task);
6015 list_del_init(&task->cg_list);
6016 spin_unlock_irq(&css_set_lock);
6019 void cgroup_free(struct task_struct *task)
6021 struct css_set *cset = task_css_set(task);
6025 static int __init cgroup_disable(char *str)
6027 struct cgroup_subsys *ss;
6031 while ((token = strsep(&str, ",")) != NULL) {
6035 for_each_subsys(ss, i) {
6036 if (strcmp(token, ss->name) &&
6037 strcmp(token, ss->legacy_name))
6039 cgroup_disable_mask |= 1 << i;
6044 __setup("cgroup_disable=", cgroup_disable);
6046 void __init __weak enable_debug_cgroup(void) { }
6048 static int __init enable_cgroup_debug(char *str)
6050 cgroup_debug = true;
6051 enable_debug_cgroup();
6054 __setup("cgroup_debug", enable_cgroup_debug);
6057 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6058 * @dentry: directory dentry of interest
6059 * @ss: subsystem of interest
6061 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6062 * to get the corresponding css and return it. If such css doesn't exist
6063 * or can't be pinned, an ERR_PTR value is returned.
6065 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6066 struct cgroup_subsys *ss)
6068 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6069 struct file_system_type *s_type = dentry->d_sb->s_type;
6070 struct cgroup_subsys_state *css = NULL;
6071 struct cgroup *cgrp;
6073 /* is @dentry a cgroup dir? */
6074 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6075 !kn || kernfs_type(kn) != KERNFS_DIR)
6076 return ERR_PTR(-EBADF);
6081 * This path doesn't originate from kernfs and @kn could already
6082 * have been or be removed at any point. @kn->priv is RCU
6083 * protected for this access. See css_release_work_fn() for details.
6085 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6087 css = cgroup_css(cgrp, ss);
6089 if (!css || !css_tryget_online(css))
6090 css = ERR_PTR(-ENOENT);
6097 * css_from_id - lookup css by id
6098 * @id: the cgroup id
6099 * @ss: cgroup subsys to be looked into
6101 * Returns the css if there's valid one with @id, otherwise returns NULL.
6102 * Should be called under rcu_read_lock().
6104 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6106 WARN_ON_ONCE(!rcu_read_lock_held());
6107 return idr_find(&ss->css_idr, id);
6111 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6112 * @path: path on the default hierarchy
6114 * Find the cgroup at @path on the default hierarchy, increment its
6115 * reference count and return it. Returns pointer to the found cgroup on
6116 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6117 * if @path points to a non-directory.
6119 struct cgroup *cgroup_get_from_path(const char *path)
6121 struct kernfs_node *kn;
6122 struct cgroup *cgrp;
6124 mutex_lock(&cgroup_mutex);
6126 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6128 if (kernfs_type(kn) == KERNFS_DIR) {
6130 cgroup_get_live(cgrp);
6132 cgrp = ERR_PTR(-ENOTDIR);
6136 cgrp = ERR_PTR(-ENOENT);
6139 mutex_unlock(&cgroup_mutex);
6142 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6145 * cgroup_get_from_fd - get a cgroup pointer from a fd
6146 * @fd: fd obtained by open(cgroup2_dir)
6148 * Find the cgroup from a fd which should be obtained
6149 * by opening a cgroup directory. Returns a pointer to the
6150 * cgroup on success. ERR_PTR is returned if the cgroup
6153 struct cgroup *cgroup_get_from_fd(int fd)
6155 struct cgroup_subsys_state *css;
6156 struct cgroup *cgrp;
6161 return ERR_PTR(-EBADF);
6163 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6166 return ERR_CAST(css);
6169 if (!cgroup_on_dfl(cgrp)) {
6171 return ERR_PTR(-EBADF);
6176 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6178 static u64 power_of_ten(int power)
6187 * cgroup_parse_float - parse a floating number
6188 * @input: input string
6189 * @dec_shift: number of decimal digits to shift
6192 * Parse a decimal floating point number in @input and store the result in
6193 * @v with decimal point right shifted @dec_shift times. For example, if
6194 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6195 * Returns 0 on success, -errno otherwise.
6197 * There's nothing cgroup specific about this function except that it's
6198 * currently the only user.
6200 int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6202 s64 whole, frac = 0;
6203 int fstart = 0, fend = 0, flen;
6205 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6210 flen = fend > fstart ? fend - fstart : 0;
6211 if (flen < dec_shift)
6212 frac *= power_of_ten(dec_shift - flen);
6214 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6216 *v = whole * power_of_ten(dec_shift) + frac;
6221 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6222 * definition in cgroup-defs.h.
6224 #ifdef CONFIG_SOCK_CGROUP_DATA
6226 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6228 DEFINE_SPINLOCK(cgroup_sk_update_lock);
6229 static bool cgroup_sk_alloc_disabled __read_mostly;
6231 void cgroup_sk_alloc_disable(void)
6233 if (cgroup_sk_alloc_disabled)
6235 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6236 cgroup_sk_alloc_disabled = true;
6241 #define cgroup_sk_alloc_disabled false
6245 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6247 if (cgroup_sk_alloc_disabled)
6250 /* Socket clone path */
6253 * We might be cloning a socket which is left in an empty
6254 * cgroup and the cgroup might have already been rmdir'd.
6255 * Don't use cgroup_get_live().
6257 cgroup_get(sock_cgroup_ptr(skcd));
6258 cgroup_bpf_get(sock_cgroup_ptr(skcd));
6265 struct css_set *cset;
6267 cset = task_css_set(current);
6268 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6269 skcd->val = (unsigned long)cset->dfl_cgrp;
6270 cgroup_bpf_get(cset->dfl_cgrp);
6279 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6281 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6283 cgroup_bpf_put(cgrp);
6287 #endif /* CONFIG_SOCK_CGROUP_DATA */
6289 #ifdef CONFIG_CGROUP_BPF
6290 int cgroup_bpf_attach(struct cgroup *cgrp, struct bpf_prog *prog,
6291 enum bpf_attach_type type, u32 flags)
6295 mutex_lock(&cgroup_mutex);
6296 ret = __cgroup_bpf_attach(cgrp, prog, type, flags);
6297 mutex_unlock(&cgroup_mutex);
6300 int cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
6301 enum bpf_attach_type type, u32 flags)
6305 mutex_lock(&cgroup_mutex);
6306 ret = __cgroup_bpf_detach(cgrp, prog, type);
6307 mutex_unlock(&cgroup_mutex);
6310 int cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
6311 union bpf_attr __user *uattr)
6315 mutex_lock(&cgroup_mutex);
6316 ret = __cgroup_bpf_query(cgrp, attr, uattr);
6317 mutex_unlock(&cgroup_mutex);
6320 #endif /* CONFIG_CGROUP_BPF */
6323 static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6324 ssize_t size, const char *prefix)
6329 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6330 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
6334 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
6336 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
6338 if (WARN_ON(ret >= size))
6345 static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
6348 struct cgroup_subsys *ss;
6352 ret = show_delegatable_files(cgroup_base_files, buf, PAGE_SIZE - ret,
6355 for_each_subsys(ss, ssid)
6356 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
6358 cgroup_subsys_name[ssid]);
6362 static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
6364 static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
6367 return snprintf(buf, PAGE_SIZE, "nsdelegate\nmemory_localevents\n");
6369 static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
6371 static struct attribute *cgroup_sysfs_attrs[] = {
6372 &cgroup_delegate_attr.attr,
6373 &cgroup_features_attr.attr,
6377 static const struct attribute_group cgroup_sysfs_attr_group = {
6378 .attrs = cgroup_sysfs_attrs,
6382 static int __init cgroup_sysfs_init(void)
6384 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
6386 subsys_initcall(cgroup_sysfs_init);
6388 #endif /* CONFIG_SYSFS */