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_fs_parameters[] = {
1820 fsparam_flag("nsdelegate", Opt_nsdelegate),
1821 fsparam_flag("memory_localevents", Opt_memory_localevents),
1825 static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1827 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1828 struct fs_parse_result result;
1831 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1836 case Opt_nsdelegate:
1837 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1839 case Opt_memory_localevents:
1840 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1846 static void apply_cgroup_root_flags(unsigned int root_flags)
1848 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1849 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1850 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1852 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1854 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1855 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1857 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1861 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1863 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1864 seq_puts(seq, ",nsdelegate");
1865 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1866 seq_puts(seq, ",memory_localevents");
1870 static int cgroup_reconfigure(struct fs_context *fc)
1872 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1874 apply_cgroup_root_flags(ctx->flags);
1878 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1880 struct cgroup_subsys *ss;
1883 INIT_LIST_HEAD(&cgrp->self.sibling);
1884 INIT_LIST_HEAD(&cgrp->self.children);
1885 INIT_LIST_HEAD(&cgrp->cset_links);
1886 INIT_LIST_HEAD(&cgrp->pidlists);
1887 mutex_init(&cgrp->pidlist_mutex);
1888 cgrp->self.cgroup = cgrp;
1889 cgrp->self.flags |= CSS_ONLINE;
1890 cgrp->dom_cgrp = cgrp;
1891 cgrp->max_descendants = INT_MAX;
1892 cgrp->max_depth = INT_MAX;
1893 INIT_LIST_HEAD(&cgrp->rstat_css_list);
1894 prev_cputime_init(&cgrp->prev_cputime);
1896 for_each_subsys(ss, ssid)
1897 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1899 init_waitqueue_head(&cgrp->offline_waitq);
1900 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1903 void init_cgroup_root(struct cgroup_fs_context *ctx)
1905 struct cgroup_root *root = ctx->root;
1906 struct cgroup *cgrp = &root->cgrp;
1908 INIT_LIST_HEAD(&root->root_list);
1909 atomic_set(&root->nr_cgrps, 1);
1911 init_cgroup_housekeeping(cgrp);
1913 root->flags = ctx->flags;
1914 if (ctx->release_agent)
1915 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
1917 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
1918 if (ctx->cpuset_clone_children)
1919 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1922 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1924 LIST_HEAD(tmp_links);
1925 struct cgroup *root_cgrp = &root->cgrp;
1926 struct kernfs_syscall_ops *kf_sops;
1927 struct css_set *cset;
1930 lockdep_assert_held(&cgroup_mutex);
1932 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1938 * We're accessing css_set_count without locking css_set_lock here,
1939 * but that's OK - it can only be increased by someone holding
1940 * cgroup_lock, and that's us. Later rebinding may disable
1941 * controllers on the default hierarchy and thus create new csets,
1942 * which can't be more than the existing ones. Allocate 2x.
1944 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1948 ret = cgroup_init_root_id(root);
1952 kf_sops = root == &cgrp_dfl_root ?
1953 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1955 root->kf_root = kernfs_create_root(kf_sops,
1956 KERNFS_ROOT_CREATE_DEACTIVATED |
1957 KERNFS_ROOT_SUPPORT_EXPORTOP,
1959 if (IS_ERR(root->kf_root)) {
1960 ret = PTR_ERR(root->kf_root);
1963 root_cgrp->kn = root->kf_root->kn;
1964 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
1965 root_cgrp->ancestor_ids[0] = cgroup_id(root_cgrp);
1967 ret = css_populate_dir(&root_cgrp->self);
1971 ret = rebind_subsystems(root, ss_mask);
1975 ret = cgroup_bpf_inherit(root_cgrp);
1978 trace_cgroup_setup_root(root);
1981 * There must be no failure case after here, since rebinding takes
1982 * care of subsystems' refcounts, which are explicitly dropped in
1983 * the failure exit path.
1985 list_add(&root->root_list, &cgroup_roots);
1986 cgroup_root_count++;
1989 * Link the root cgroup in this hierarchy into all the css_set
1992 spin_lock_irq(&css_set_lock);
1993 hash_for_each(css_set_table, i, cset, hlist) {
1994 link_css_set(&tmp_links, cset, root_cgrp);
1995 if (css_set_populated(cset))
1996 cgroup_update_populated(root_cgrp, true);
1998 spin_unlock_irq(&css_set_lock);
2000 BUG_ON(!list_empty(&root_cgrp->self.children));
2001 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2003 kernfs_activate(root_cgrp->kn);
2008 kernfs_destroy_root(root->kf_root);
2009 root->kf_root = NULL;
2011 cgroup_exit_root_id(root);
2013 percpu_ref_exit(&root_cgrp->self.refcnt);
2015 free_cgrp_cset_links(&tmp_links);
2019 int cgroup_do_get_tree(struct fs_context *fc)
2021 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2024 ctx->kfc.root = ctx->root->kf_root;
2025 if (fc->fs_type == &cgroup2_fs_type)
2026 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2028 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2029 ret = kernfs_get_tree(fc);
2032 * In non-init cgroup namespace, instead of root cgroup's dentry,
2033 * we return the dentry corresponding to the cgroupns->root_cgrp.
2035 if (!ret && ctx->ns != &init_cgroup_ns) {
2036 struct dentry *nsdentry;
2037 struct super_block *sb = fc->root->d_sb;
2038 struct cgroup *cgrp;
2040 mutex_lock(&cgroup_mutex);
2041 spin_lock_irq(&css_set_lock);
2043 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2045 spin_unlock_irq(&css_set_lock);
2046 mutex_unlock(&cgroup_mutex);
2048 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2050 if (IS_ERR(nsdentry)) {
2051 deactivate_locked_super(sb);
2052 ret = PTR_ERR(nsdentry);
2055 fc->root = nsdentry;
2058 if (!ctx->kfc.new_sb_created)
2059 cgroup_put(&ctx->root->cgrp);
2065 * Destroy a cgroup filesystem context.
2067 static void cgroup_fs_context_free(struct fs_context *fc)
2069 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2072 kfree(ctx->release_agent);
2073 put_cgroup_ns(ctx->ns);
2074 kernfs_free_fs_context(fc);
2078 static int cgroup_get_tree(struct fs_context *fc)
2080 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2083 cgrp_dfl_visible = true;
2084 cgroup_get_live(&cgrp_dfl_root.cgrp);
2085 ctx->root = &cgrp_dfl_root;
2087 ret = cgroup_do_get_tree(fc);
2089 apply_cgroup_root_flags(ctx->flags);
2093 static const struct fs_context_operations cgroup_fs_context_ops = {
2094 .free = cgroup_fs_context_free,
2095 .parse_param = cgroup2_parse_param,
2096 .get_tree = cgroup_get_tree,
2097 .reconfigure = cgroup_reconfigure,
2100 static const struct fs_context_operations cgroup1_fs_context_ops = {
2101 .free = cgroup_fs_context_free,
2102 .parse_param = cgroup1_parse_param,
2103 .get_tree = cgroup1_get_tree,
2104 .reconfigure = cgroup1_reconfigure,
2108 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2109 * we select the namespace we're going to use.
2111 static int cgroup_init_fs_context(struct fs_context *fc)
2113 struct cgroup_fs_context *ctx;
2115 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2119 ctx->ns = current->nsproxy->cgroup_ns;
2120 get_cgroup_ns(ctx->ns);
2121 fc->fs_private = &ctx->kfc;
2122 if (fc->fs_type == &cgroup2_fs_type)
2123 fc->ops = &cgroup_fs_context_ops;
2125 fc->ops = &cgroup1_fs_context_ops;
2126 put_user_ns(fc->user_ns);
2127 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2132 static void cgroup_kill_sb(struct super_block *sb)
2134 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2135 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2138 * If @root doesn't have any children, start killing it.
2139 * This prevents new mounts by disabling percpu_ref_tryget_live().
2140 * cgroup_mount() may wait for @root's release.
2142 * And don't kill the default root.
2144 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2145 !percpu_ref_is_dying(&root->cgrp.self.refcnt))
2146 percpu_ref_kill(&root->cgrp.self.refcnt);
2147 cgroup_put(&root->cgrp);
2151 struct file_system_type cgroup_fs_type = {
2153 .init_fs_context = cgroup_init_fs_context,
2154 .parameters = cgroup1_fs_parameters,
2155 .kill_sb = cgroup_kill_sb,
2156 .fs_flags = FS_USERNS_MOUNT,
2159 static struct file_system_type cgroup2_fs_type = {
2161 .init_fs_context = cgroup_init_fs_context,
2162 .parameters = cgroup2_fs_parameters,
2163 .kill_sb = cgroup_kill_sb,
2164 .fs_flags = FS_USERNS_MOUNT,
2167 #ifdef CONFIG_CPUSETS
2168 static const struct fs_context_operations cpuset_fs_context_ops = {
2169 .get_tree = cgroup1_get_tree,
2170 .free = cgroup_fs_context_free,
2174 * This is ugly, but preserves the userspace API for existing cpuset
2175 * users. If someone tries to mount the "cpuset" filesystem, we
2176 * silently switch it to mount "cgroup" instead
2178 static int cpuset_init_fs_context(struct fs_context *fc)
2180 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2181 struct cgroup_fs_context *ctx;
2184 err = cgroup_init_fs_context(fc);
2190 fc->ops = &cpuset_fs_context_ops;
2192 ctx = cgroup_fc2context(fc);
2193 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2194 ctx->flags |= CGRP_ROOT_NOPREFIX;
2195 ctx->release_agent = agent;
2197 get_filesystem(&cgroup_fs_type);
2198 put_filesystem(fc->fs_type);
2199 fc->fs_type = &cgroup_fs_type;
2204 static struct file_system_type cpuset_fs_type = {
2206 .init_fs_context = cpuset_init_fs_context,
2207 .fs_flags = FS_USERNS_MOUNT,
2211 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2212 struct cgroup_namespace *ns)
2214 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2216 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2219 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2220 struct cgroup_namespace *ns)
2224 mutex_lock(&cgroup_mutex);
2225 spin_lock_irq(&css_set_lock);
2227 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2229 spin_unlock_irq(&css_set_lock);
2230 mutex_unlock(&cgroup_mutex);
2234 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2237 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2238 * @task: target task
2239 * @buf: the buffer to write the path into
2240 * @buflen: the length of the buffer
2242 * Determine @task's cgroup on the first (the one with the lowest non-zero
2243 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2244 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2245 * cgroup controller callbacks.
2247 * Return value is the same as kernfs_path().
2249 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2251 struct cgroup_root *root;
2252 struct cgroup *cgrp;
2253 int hierarchy_id = 1;
2256 mutex_lock(&cgroup_mutex);
2257 spin_lock_irq(&css_set_lock);
2259 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2262 cgrp = task_cgroup_from_root(task, root);
2263 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2265 /* if no hierarchy exists, everyone is in "/" */
2266 ret = strlcpy(buf, "/", buflen);
2269 spin_unlock_irq(&css_set_lock);
2270 mutex_unlock(&cgroup_mutex);
2273 EXPORT_SYMBOL_GPL(task_cgroup_path);
2276 * cgroup_migrate_add_task - add a migration target task to a migration context
2277 * @task: target task
2278 * @mgctx: target migration context
2280 * Add @task, which is a migration target, to @mgctx->tset. This function
2281 * becomes noop if @task doesn't need to be migrated. @task's css_set
2282 * should have been added as a migration source and @task->cg_list will be
2283 * moved from the css_set's tasks list to mg_tasks one.
2285 static void cgroup_migrate_add_task(struct task_struct *task,
2286 struct cgroup_mgctx *mgctx)
2288 struct css_set *cset;
2290 lockdep_assert_held(&css_set_lock);
2292 /* @task either already exited or can't exit until the end */
2293 if (task->flags & PF_EXITING)
2296 /* cgroup_threadgroup_rwsem protects racing against forks */
2297 WARN_ON_ONCE(list_empty(&task->cg_list));
2299 cset = task_css_set(task);
2300 if (!cset->mg_src_cgrp)
2303 mgctx->tset.nr_tasks++;
2305 list_move_tail(&task->cg_list, &cset->mg_tasks);
2306 if (list_empty(&cset->mg_node))
2307 list_add_tail(&cset->mg_node,
2308 &mgctx->tset.src_csets);
2309 if (list_empty(&cset->mg_dst_cset->mg_node))
2310 list_add_tail(&cset->mg_dst_cset->mg_node,
2311 &mgctx->tset.dst_csets);
2315 * cgroup_taskset_first - reset taskset and return the first task
2316 * @tset: taskset of interest
2317 * @dst_cssp: output variable for the destination css
2319 * @tset iteration is initialized and the first task is returned.
2321 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2322 struct cgroup_subsys_state **dst_cssp)
2324 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2325 tset->cur_task = NULL;
2327 return cgroup_taskset_next(tset, dst_cssp);
2331 * cgroup_taskset_next - iterate to the next task in taskset
2332 * @tset: taskset of interest
2333 * @dst_cssp: output variable for the destination css
2335 * Return the next task in @tset. Iteration must have been initialized
2336 * with cgroup_taskset_first().
2338 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2339 struct cgroup_subsys_state **dst_cssp)
2341 struct css_set *cset = tset->cur_cset;
2342 struct task_struct *task = tset->cur_task;
2344 while (&cset->mg_node != tset->csets) {
2346 task = list_first_entry(&cset->mg_tasks,
2347 struct task_struct, cg_list);
2349 task = list_next_entry(task, cg_list);
2351 if (&task->cg_list != &cset->mg_tasks) {
2352 tset->cur_cset = cset;
2353 tset->cur_task = task;
2356 * This function may be called both before and
2357 * after cgroup_taskset_migrate(). The two cases
2358 * can be distinguished by looking at whether @cset
2359 * has its ->mg_dst_cset set.
2361 if (cset->mg_dst_cset)
2362 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2364 *dst_cssp = cset->subsys[tset->ssid];
2369 cset = list_next_entry(cset, mg_node);
2377 * cgroup_taskset_migrate - migrate a taskset
2378 * @mgctx: migration context
2380 * Migrate tasks in @mgctx as setup by migration preparation functions.
2381 * This function fails iff one of the ->can_attach callbacks fails and
2382 * guarantees that either all or none of the tasks in @mgctx are migrated.
2383 * @mgctx is consumed regardless of success.
2385 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2387 struct cgroup_taskset *tset = &mgctx->tset;
2388 struct cgroup_subsys *ss;
2389 struct task_struct *task, *tmp_task;
2390 struct css_set *cset, *tmp_cset;
2391 int ssid, failed_ssid, ret;
2393 /* check that we can legitimately attach to the cgroup */
2394 if (tset->nr_tasks) {
2395 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2396 if (ss->can_attach) {
2398 ret = ss->can_attach(tset);
2401 goto out_cancel_attach;
2404 } while_each_subsys_mask();
2408 * Now that we're guaranteed success, proceed to move all tasks to
2409 * the new cgroup. There are no failure cases after here, so this
2410 * is the commit point.
2412 spin_lock_irq(&css_set_lock);
2413 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2414 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2415 struct css_set *from_cset = task_css_set(task);
2416 struct css_set *to_cset = cset->mg_dst_cset;
2418 get_css_set(to_cset);
2419 to_cset->nr_tasks++;
2420 css_set_move_task(task, from_cset, to_cset, true);
2421 from_cset->nr_tasks--;
2423 * If the source or destination cgroup is frozen,
2424 * the task might require to change its state.
2426 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2428 put_css_set_locked(from_cset);
2432 spin_unlock_irq(&css_set_lock);
2435 * Migration is committed, all target tasks are now on dst_csets.
2436 * Nothing is sensitive to fork() after this point. Notify
2437 * controllers that migration is complete.
2439 tset->csets = &tset->dst_csets;
2441 if (tset->nr_tasks) {
2442 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2447 } while_each_subsys_mask();
2451 goto out_release_tset;
2454 if (tset->nr_tasks) {
2455 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2456 if (ssid == failed_ssid)
2458 if (ss->cancel_attach) {
2460 ss->cancel_attach(tset);
2462 } while_each_subsys_mask();
2465 spin_lock_irq(&css_set_lock);
2466 list_splice_init(&tset->dst_csets, &tset->src_csets);
2467 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2468 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2469 list_del_init(&cset->mg_node);
2471 spin_unlock_irq(&css_set_lock);
2474 * Re-initialize the cgroup_taskset structure in case it is reused
2475 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2479 tset->csets = &tset->src_csets;
2484 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2485 * @dst_cgrp: destination cgroup to test
2487 * On the default hierarchy, except for the mixable, (possible) thread root
2488 * and threaded cgroups, subtree_control must be zero for migration
2489 * destination cgroups with tasks so that child cgroups don't compete
2492 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2494 /* v1 doesn't have any restriction */
2495 if (!cgroup_on_dfl(dst_cgrp))
2498 /* verify @dst_cgrp can host resources */
2499 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2502 /* mixables don't care */
2503 if (cgroup_is_mixable(dst_cgrp))
2507 * If @dst_cgrp is already or can become a thread root or is
2508 * threaded, it doesn't matter.
2510 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2513 /* apply no-internal-process constraint */
2514 if (dst_cgrp->subtree_control)
2521 * cgroup_migrate_finish - cleanup after attach
2522 * @mgctx: migration context
2524 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2525 * those functions for details.
2527 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2529 LIST_HEAD(preloaded);
2530 struct css_set *cset, *tmp_cset;
2532 lockdep_assert_held(&cgroup_mutex);
2534 spin_lock_irq(&css_set_lock);
2536 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2537 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2539 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2540 cset->mg_src_cgrp = NULL;
2541 cset->mg_dst_cgrp = NULL;
2542 cset->mg_dst_cset = NULL;
2543 list_del_init(&cset->mg_preload_node);
2544 put_css_set_locked(cset);
2547 spin_unlock_irq(&css_set_lock);
2551 * cgroup_migrate_add_src - add a migration source css_set
2552 * @src_cset: the source css_set to add
2553 * @dst_cgrp: the destination cgroup
2554 * @mgctx: migration context
2556 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2557 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2558 * up by cgroup_migrate_finish().
2560 * This function may be called without holding cgroup_threadgroup_rwsem
2561 * even if the target is a process. Threads may be created and destroyed
2562 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2563 * into play and the preloaded css_sets are guaranteed to cover all
2566 void cgroup_migrate_add_src(struct css_set *src_cset,
2567 struct cgroup *dst_cgrp,
2568 struct cgroup_mgctx *mgctx)
2570 struct cgroup *src_cgrp;
2572 lockdep_assert_held(&cgroup_mutex);
2573 lockdep_assert_held(&css_set_lock);
2576 * If ->dead, @src_set is associated with one or more dead cgroups
2577 * and doesn't contain any migratable tasks. Ignore it early so
2578 * that the rest of migration path doesn't get confused by it.
2583 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2585 if (!list_empty(&src_cset->mg_preload_node))
2588 WARN_ON(src_cset->mg_src_cgrp);
2589 WARN_ON(src_cset->mg_dst_cgrp);
2590 WARN_ON(!list_empty(&src_cset->mg_tasks));
2591 WARN_ON(!list_empty(&src_cset->mg_node));
2593 src_cset->mg_src_cgrp = src_cgrp;
2594 src_cset->mg_dst_cgrp = dst_cgrp;
2595 get_css_set(src_cset);
2596 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2600 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2601 * @mgctx: migration context
2603 * Tasks are about to be moved and all the source css_sets have been
2604 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2605 * pins all destination css_sets, links each to its source, and append them
2606 * to @mgctx->preloaded_dst_csets.
2608 * This function must be called after cgroup_migrate_add_src() has been
2609 * called on each migration source css_set. After migration is performed
2610 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2613 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2615 struct css_set *src_cset, *tmp_cset;
2617 lockdep_assert_held(&cgroup_mutex);
2619 /* look up the dst cset for each src cset and link it to src */
2620 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2622 struct css_set *dst_cset;
2623 struct cgroup_subsys *ss;
2626 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2630 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2633 * If src cset equals dst, it's noop. Drop the src.
2634 * cgroup_migrate() will skip the cset too. Note that we
2635 * can't handle src == dst as some nodes are used by both.
2637 if (src_cset == dst_cset) {
2638 src_cset->mg_src_cgrp = NULL;
2639 src_cset->mg_dst_cgrp = NULL;
2640 list_del_init(&src_cset->mg_preload_node);
2641 put_css_set(src_cset);
2642 put_css_set(dst_cset);
2646 src_cset->mg_dst_cset = dst_cset;
2648 if (list_empty(&dst_cset->mg_preload_node))
2649 list_add_tail(&dst_cset->mg_preload_node,
2650 &mgctx->preloaded_dst_csets);
2652 put_css_set(dst_cset);
2654 for_each_subsys(ss, ssid)
2655 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2656 mgctx->ss_mask |= 1 << ssid;
2663 * cgroup_migrate - migrate a process or task to a cgroup
2664 * @leader: the leader of the process or the task to migrate
2665 * @threadgroup: whether @leader points to the whole process or a single task
2666 * @mgctx: migration context
2668 * Migrate a process or task denoted by @leader. If migrating a process,
2669 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2670 * responsible for invoking cgroup_migrate_add_src() and
2671 * cgroup_migrate_prepare_dst() on the targets before invoking this
2672 * function and following up with cgroup_migrate_finish().
2674 * As long as a controller's ->can_attach() doesn't fail, this function is
2675 * guaranteed to succeed. This means that, excluding ->can_attach()
2676 * failure, when migrating multiple targets, the success or failure can be
2677 * decided for all targets by invoking group_migrate_prepare_dst() before
2678 * actually starting migrating.
2680 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2681 struct cgroup_mgctx *mgctx)
2683 struct task_struct *task;
2686 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2687 * already PF_EXITING could be freed from underneath us unless we
2688 * take an rcu_read_lock.
2690 spin_lock_irq(&css_set_lock);
2694 cgroup_migrate_add_task(task, mgctx);
2697 } while_each_thread(leader, task);
2699 spin_unlock_irq(&css_set_lock);
2701 return cgroup_migrate_execute(mgctx);
2705 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2706 * @dst_cgrp: the cgroup to attach to
2707 * @leader: the task or the leader of the threadgroup to be attached
2708 * @threadgroup: attach the whole threadgroup?
2710 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2712 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2715 DEFINE_CGROUP_MGCTX(mgctx);
2716 struct task_struct *task;
2719 ret = cgroup_migrate_vet_dst(dst_cgrp);
2723 /* look up all src csets */
2724 spin_lock_irq(&css_set_lock);
2728 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2731 } while_each_thread(leader, task);
2733 spin_unlock_irq(&css_set_lock);
2735 /* prepare dst csets and commit */
2736 ret = cgroup_migrate_prepare_dst(&mgctx);
2738 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2740 cgroup_migrate_finish(&mgctx);
2743 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2748 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2750 __acquires(&cgroup_threadgroup_rwsem)
2752 struct task_struct *tsk;
2755 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2756 return ERR_PTR(-EINVAL);
2759 * If we migrate a single thread, we don't care about threadgroup
2760 * stability. If the thread is `current`, it won't exit(2) under our
2761 * hands or change PID through exec(2). We exclude
2762 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2763 * callers by cgroup_mutex.
2764 * Therefore, we can skip the global lock.
2766 lockdep_assert_held(&cgroup_mutex);
2767 if (pid || threadgroup) {
2768 percpu_down_write(&cgroup_threadgroup_rwsem);
2776 tsk = find_task_by_vpid(pid);
2778 tsk = ERR_PTR(-ESRCH);
2779 goto out_unlock_threadgroup;
2786 tsk = tsk->group_leader;
2789 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2790 * If userland migrates such a kthread to a non-root cgroup, it can
2791 * become trapped in a cpuset, or RT kthread may be born in a
2792 * cgroup with no rt_runtime allocated. Just say no.
2794 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2795 tsk = ERR_PTR(-EINVAL);
2796 goto out_unlock_threadgroup;
2799 get_task_struct(tsk);
2800 goto out_unlock_rcu;
2802 out_unlock_threadgroup:
2804 percpu_up_write(&cgroup_threadgroup_rwsem);
2812 void cgroup_procs_write_finish(struct task_struct *task, bool locked)
2813 __releases(&cgroup_threadgroup_rwsem)
2815 struct cgroup_subsys *ss;
2818 /* release reference from cgroup_procs_write_start() */
2819 put_task_struct(task);
2822 percpu_up_write(&cgroup_threadgroup_rwsem);
2823 for_each_subsys(ss, ssid)
2824 if (ss->post_attach)
2828 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2830 struct cgroup_subsys *ss;
2831 bool printed = false;
2834 do_each_subsys_mask(ss, ssid, ss_mask) {
2837 seq_puts(seq, ss->name);
2839 } while_each_subsys_mask();
2841 seq_putc(seq, '\n');
2844 /* show controllers which are enabled from the parent */
2845 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2847 struct cgroup *cgrp = seq_css(seq)->cgroup;
2849 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2853 /* show controllers which are enabled for a given cgroup's children */
2854 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2856 struct cgroup *cgrp = seq_css(seq)->cgroup;
2858 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2863 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2864 * @cgrp: root of the subtree to update csses for
2866 * @cgrp's control masks have changed and its subtree's css associations
2867 * need to be updated accordingly. This function looks up all css_sets
2868 * which are attached to the subtree, creates the matching updated css_sets
2869 * and migrates the tasks to the new ones.
2871 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2873 DEFINE_CGROUP_MGCTX(mgctx);
2874 struct cgroup_subsys_state *d_css;
2875 struct cgroup *dsct;
2876 struct css_set *src_cset;
2879 lockdep_assert_held(&cgroup_mutex);
2881 percpu_down_write(&cgroup_threadgroup_rwsem);
2883 /* look up all csses currently attached to @cgrp's subtree */
2884 spin_lock_irq(&css_set_lock);
2885 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2886 struct cgrp_cset_link *link;
2888 list_for_each_entry(link, &dsct->cset_links, cset_link)
2889 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2891 spin_unlock_irq(&css_set_lock);
2893 /* NULL dst indicates self on default hierarchy */
2894 ret = cgroup_migrate_prepare_dst(&mgctx);
2898 spin_lock_irq(&css_set_lock);
2899 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2900 struct task_struct *task, *ntask;
2902 /* all tasks in src_csets need to be migrated */
2903 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2904 cgroup_migrate_add_task(task, &mgctx);
2906 spin_unlock_irq(&css_set_lock);
2908 ret = cgroup_migrate_execute(&mgctx);
2910 cgroup_migrate_finish(&mgctx);
2911 percpu_up_write(&cgroup_threadgroup_rwsem);
2916 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2917 * @cgrp: root of the target subtree
2919 * Because css offlining is asynchronous, userland may try to re-enable a
2920 * controller while the previous css is still around. This function grabs
2921 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2923 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2924 __acquires(&cgroup_mutex)
2926 struct cgroup *dsct;
2927 struct cgroup_subsys_state *d_css;
2928 struct cgroup_subsys *ss;
2932 mutex_lock(&cgroup_mutex);
2934 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2935 for_each_subsys(ss, ssid) {
2936 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2939 if (!css || !percpu_ref_is_dying(&css->refcnt))
2942 cgroup_get_live(dsct);
2943 prepare_to_wait(&dsct->offline_waitq, &wait,
2944 TASK_UNINTERRUPTIBLE);
2946 mutex_unlock(&cgroup_mutex);
2948 finish_wait(&dsct->offline_waitq, &wait);
2957 * cgroup_save_control - save control masks and dom_cgrp of a subtree
2958 * @cgrp: root of the target subtree
2960 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
2961 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2964 static void cgroup_save_control(struct cgroup *cgrp)
2966 struct cgroup *dsct;
2967 struct cgroup_subsys_state *d_css;
2969 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2970 dsct->old_subtree_control = dsct->subtree_control;
2971 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
2972 dsct->old_dom_cgrp = dsct->dom_cgrp;
2977 * cgroup_propagate_control - refresh control masks of a subtree
2978 * @cgrp: root of the target subtree
2980 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2981 * ->subtree_control and propagate controller availability through the
2982 * subtree so that descendants don't have unavailable controllers enabled.
2984 static void cgroup_propagate_control(struct cgroup *cgrp)
2986 struct cgroup *dsct;
2987 struct cgroup_subsys_state *d_css;
2989 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2990 dsct->subtree_control &= cgroup_control(dsct);
2991 dsct->subtree_ss_mask =
2992 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
2993 cgroup_ss_mask(dsct));
2998 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
2999 * @cgrp: root of the target subtree
3001 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3002 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3005 static void cgroup_restore_control(struct cgroup *cgrp)
3007 struct cgroup *dsct;
3008 struct cgroup_subsys_state *d_css;
3010 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3011 dsct->subtree_control = dsct->old_subtree_control;
3012 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3013 dsct->dom_cgrp = dsct->old_dom_cgrp;
3017 static bool css_visible(struct cgroup_subsys_state *css)
3019 struct cgroup_subsys *ss = css->ss;
3020 struct cgroup *cgrp = css->cgroup;
3022 if (cgroup_control(cgrp) & (1 << ss->id))
3024 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3026 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3030 * cgroup_apply_control_enable - enable or show csses according to control
3031 * @cgrp: root of the target subtree
3033 * Walk @cgrp's subtree and create new csses or make the existing ones
3034 * visible. A css is created invisible if it's being implicitly enabled
3035 * through dependency. An invisible css is made visible when the userland
3036 * explicitly enables it.
3038 * Returns 0 on success, -errno on failure. On failure, csses which have
3039 * been processed already aren't cleaned up. The caller is responsible for
3040 * cleaning up with cgroup_apply_control_disable().
3042 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3044 struct cgroup *dsct;
3045 struct cgroup_subsys_state *d_css;
3046 struct cgroup_subsys *ss;
3049 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3050 for_each_subsys(ss, ssid) {
3051 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3053 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3057 css = css_create(dsct, ss);
3059 return PTR_ERR(css);
3062 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3064 if (css_visible(css)) {
3065 ret = css_populate_dir(css);
3076 * cgroup_apply_control_disable - kill or hide csses according to control
3077 * @cgrp: root of the target subtree
3079 * Walk @cgrp's subtree and kill and hide csses so that they match
3080 * cgroup_ss_mask() and cgroup_visible_mask().
3082 * A css is hidden when the userland requests it to be disabled while other
3083 * subsystems are still depending on it. The css must not actively control
3084 * resources and be in the vanilla state if it's made visible again later.
3085 * Controllers which may be depended upon should provide ->css_reset() for
3088 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3090 struct cgroup *dsct;
3091 struct cgroup_subsys_state *d_css;
3092 struct cgroup_subsys *ss;
3095 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3096 for_each_subsys(ss, ssid) {
3097 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3102 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3105 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3107 } else if (!css_visible(css)) {
3117 * cgroup_apply_control - apply control mask updates to the subtree
3118 * @cgrp: root of the target subtree
3120 * subsystems can be enabled and disabled in a subtree using the following
3123 * 1. Call cgroup_save_control() to stash the current state.
3124 * 2. Update ->subtree_control masks in the subtree as desired.
3125 * 3. Call cgroup_apply_control() to apply the changes.
3126 * 4. Optionally perform other related operations.
3127 * 5. Call cgroup_finalize_control() to finish up.
3129 * This function implements step 3 and propagates the mask changes
3130 * throughout @cgrp's subtree, updates csses accordingly and perform
3131 * process migrations.
3133 static int cgroup_apply_control(struct cgroup *cgrp)
3137 cgroup_propagate_control(cgrp);
3139 ret = cgroup_apply_control_enable(cgrp);
3144 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3145 * making the following cgroup_update_dfl_csses() properly update
3146 * css associations of all tasks in the subtree.
3148 ret = cgroup_update_dfl_csses(cgrp);
3156 * cgroup_finalize_control - finalize control mask update
3157 * @cgrp: root of the target subtree
3158 * @ret: the result of the update
3160 * Finalize control mask update. See cgroup_apply_control() for more info.
3162 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3165 cgroup_restore_control(cgrp);
3166 cgroup_propagate_control(cgrp);
3169 cgroup_apply_control_disable(cgrp);
3172 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3174 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3176 /* if nothing is getting enabled, nothing to worry about */
3180 /* can @cgrp host any resources? */
3181 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3184 /* mixables don't care */
3185 if (cgroup_is_mixable(cgrp))
3188 if (domain_enable) {
3189 /* can't enable domain controllers inside a thread subtree */
3190 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3194 * Threaded controllers can handle internal competitions
3195 * and are always allowed inside a (prospective) thread
3198 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3203 * Controllers can't be enabled for a cgroup with tasks to avoid
3204 * child cgroups competing against tasks.
3206 if (cgroup_has_tasks(cgrp))
3212 /* change the enabled child controllers for a cgroup in the default hierarchy */
3213 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3214 char *buf, size_t nbytes,
3217 u16 enable = 0, disable = 0;
3218 struct cgroup *cgrp, *child;
3219 struct cgroup_subsys *ss;
3224 * Parse input - space separated list of subsystem names prefixed
3225 * with either + or -.
3227 buf = strstrip(buf);
3228 while ((tok = strsep(&buf, " "))) {
3231 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3232 if (!cgroup_ssid_enabled(ssid) ||
3233 strcmp(tok + 1, ss->name))
3237 enable |= 1 << ssid;
3238 disable &= ~(1 << ssid);
3239 } else if (*tok == '-') {
3240 disable |= 1 << ssid;
3241 enable &= ~(1 << ssid);
3246 } while_each_subsys_mask();
3247 if (ssid == CGROUP_SUBSYS_COUNT)
3251 cgrp = cgroup_kn_lock_live(of->kn, true);
3255 for_each_subsys(ss, ssid) {
3256 if (enable & (1 << ssid)) {
3257 if (cgrp->subtree_control & (1 << ssid)) {
3258 enable &= ~(1 << ssid);
3262 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3266 } else if (disable & (1 << ssid)) {
3267 if (!(cgrp->subtree_control & (1 << ssid))) {
3268 disable &= ~(1 << ssid);
3272 /* a child has it enabled? */
3273 cgroup_for_each_live_child(child, cgrp) {
3274 if (child->subtree_control & (1 << ssid)) {
3282 if (!enable && !disable) {
3287 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3291 /* save and update control masks and prepare csses */
3292 cgroup_save_control(cgrp);
3294 cgrp->subtree_control |= enable;
3295 cgrp->subtree_control &= ~disable;
3297 ret = cgroup_apply_control(cgrp);
3298 cgroup_finalize_control(cgrp, ret);
3302 kernfs_activate(cgrp->kn);
3304 cgroup_kn_unlock(of->kn);
3305 return ret ?: nbytes;
3309 * cgroup_enable_threaded - make @cgrp threaded
3310 * @cgrp: the target cgroup
3312 * Called when "threaded" is written to the cgroup.type interface file and
3313 * tries to make @cgrp threaded and join the parent's resource domain.
3314 * This function is never called on the root cgroup as cgroup.type doesn't
3317 static int cgroup_enable_threaded(struct cgroup *cgrp)
3319 struct cgroup *parent = cgroup_parent(cgrp);
3320 struct cgroup *dom_cgrp = parent->dom_cgrp;
3321 struct cgroup *dsct;
3322 struct cgroup_subsys_state *d_css;
3325 lockdep_assert_held(&cgroup_mutex);
3327 /* noop if already threaded */
3328 if (cgroup_is_threaded(cgrp))
3332 * If @cgroup is populated or has domain controllers enabled, it
3333 * can't be switched. While the below cgroup_can_be_thread_root()
3334 * test can catch the same conditions, that's only when @parent is
3335 * not mixable, so let's check it explicitly.
3337 if (cgroup_is_populated(cgrp) ||
3338 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3341 /* we're joining the parent's domain, ensure its validity */
3342 if (!cgroup_is_valid_domain(dom_cgrp) ||
3343 !cgroup_can_be_thread_root(dom_cgrp))
3347 * The following shouldn't cause actual migrations and should
3350 cgroup_save_control(cgrp);
3352 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3353 if (dsct == cgrp || cgroup_is_threaded(dsct))
3354 dsct->dom_cgrp = dom_cgrp;
3356 ret = cgroup_apply_control(cgrp);
3358 parent->nr_threaded_children++;
3360 cgroup_finalize_control(cgrp, ret);
3364 static int cgroup_type_show(struct seq_file *seq, void *v)
3366 struct cgroup *cgrp = seq_css(seq)->cgroup;
3368 if (cgroup_is_threaded(cgrp))
3369 seq_puts(seq, "threaded\n");
3370 else if (!cgroup_is_valid_domain(cgrp))
3371 seq_puts(seq, "domain invalid\n");
3372 else if (cgroup_is_thread_root(cgrp))
3373 seq_puts(seq, "domain threaded\n");
3375 seq_puts(seq, "domain\n");
3380 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3381 size_t nbytes, loff_t off)
3383 struct cgroup *cgrp;
3386 /* only switching to threaded mode is supported */
3387 if (strcmp(strstrip(buf), "threaded"))
3390 /* drain dying csses before we re-apply (threaded) subtree control */
3391 cgrp = cgroup_kn_lock_live(of->kn, true);
3395 /* threaded can only be enabled */
3396 ret = cgroup_enable_threaded(cgrp);
3398 cgroup_kn_unlock(of->kn);
3399 return ret ?: nbytes;
3402 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3404 struct cgroup *cgrp = seq_css(seq)->cgroup;
3405 int descendants = READ_ONCE(cgrp->max_descendants);
3407 if (descendants == INT_MAX)
3408 seq_puts(seq, "max\n");
3410 seq_printf(seq, "%d\n", descendants);
3415 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3416 char *buf, size_t nbytes, loff_t off)
3418 struct cgroup *cgrp;
3422 buf = strstrip(buf);
3423 if (!strcmp(buf, "max")) {
3424 descendants = INT_MAX;
3426 ret = kstrtoint(buf, 0, &descendants);
3431 if (descendants < 0)
3434 cgrp = cgroup_kn_lock_live(of->kn, false);
3438 cgrp->max_descendants = descendants;
3440 cgroup_kn_unlock(of->kn);
3445 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3447 struct cgroup *cgrp = seq_css(seq)->cgroup;
3448 int depth = READ_ONCE(cgrp->max_depth);
3450 if (depth == INT_MAX)
3451 seq_puts(seq, "max\n");
3453 seq_printf(seq, "%d\n", depth);
3458 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3459 char *buf, size_t nbytes, loff_t off)
3461 struct cgroup *cgrp;
3465 buf = strstrip(buf);
3466 if (!strcmp(buf, "max")) {
3469 ret = kstrtoint(buf, 0, &depth);
3477 cgrp = cgroup_kn_lock_live(of->kn, false);
3481 cgrp->max_depth = depth;
3483 cgroup_kn_unlock(of->kn);
3488 static int cgroup_events_show(struct seq_file *seq, void *v)
3490 struct cgroup *cgrp = seq_css(seq)->cgroup;
3492 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3493 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3498 static int cgroup_stat_show(struct seq_file *seq, void *v)
3500 struct cgroup *cgroup = seq_css(seq)->cgroup;
3502 seq_printf(seq, "nr_descendants %d\n",
3503 cgroup->nr_descendants);
3504 seq_printf(seq, "nr_dying_descendants %d\n",
3505 cgroup->nr_dying_descendants);
3510 static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3511 struct cgroup *cgrp, int ssid)
3513 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3514 struct cgroup_subsys_state *css;
3517 if (!ss->css_extra_stat_show)
3520 css = cgroup_tryget_css(cgrp, ss);
3524 ret = ss->css_extra_stat_show(seq, css);
3529 static int cpu_stat_show(struct seq_file *seq, void *v)
3531 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3534 cgroup_base_stat_cputime_show(seq);
3535 #ifdef CONFIG_CGROUP_SCHED
3536 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3542 static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3544 struct cgroup *cgrp = seq_css(seq)->cgroup;
3545 struct psi_group *psi = cgroup_id(cgrp) == 1 ? &psi_system : &cgrp->psi;
3547 return psi_show(seq, psi, PSI_IO);
3549 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3551 struct cgroup *cgrp = seq_css(seq)->cgroup;
3552 struct psi_group *psi = cgroup_id(cgrp) == 1 ? &psi_system : &cgrp->psi;
3554 return psi_show(seq, psi, PSI_MEM);
3556 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3558 struct cgroup *cgrp = seq_css(seq)->cgroup;
3559 struct psi_group *psi = cgroup_id(cgrp) == 1 ? &psi_system : &cgrp->psi;
3561 return psi_show(seq, psi, PSI_CPU);
3564 static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf,
3565 size_t nbytes, enum psi_res res)
3567 struct psi_trigger *new;
3568 struct cgroup *cgrp;
3570 cgrp = cgroup_kn_lock_live(of->kn, false);
3575 cgroup_kn_unlock(of->kn);
3577 new = psi_trigger_create(&cgrp->psi, buf, nbytes, res);
3580 return PTR_ERR(new);
3583 psi_trigger_replace(&of->priv, new);
3590 static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3591 char *buf, size_t nbytes,
3594 return cgroup_pressure_write(of, buf, nbytes, PSI_IO);
3597 static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3598 char *buf, size_t nbytes,
3601 return cgroup_pressure_write(of, buf, nbytes, PSI_MEM);
3604 static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3605 char *buf, size_t nbytes,
3608 return cgroup_pressure_write(of, buf, nbytes, PSI_CPU);
3611 static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3614 return psi_trigger_poll(&of->priv, of->file, pt);
3617 static void cgroup_pressure_release(struct kernfs_open_file *of)
3619 psi_trigger_replace(&of->priv, NULL);
3621 #endif /* CONFIG_PSI */
3623 static int cgroup_freeze_show(struct seq_file *seq, void *v)
3625 struct cgroup *cgrp = seq_css(seq)->cgroup;
3627 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3632 static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3633 char *buf, size_t nbytes, loff_t off)
3635 struct cgroup *cgrp;
3639 ret = kstrtoint(strstrip(buf), 0, &freeze);
3643 if (freeze < 0 || freeze > 1)
3646 cgrp = cgroup_kn_lock_live(of->kn, false);
3650 cgroup_freeze(cgrp, freeze);
3652 cgroup_kn_unlock(of->kn);
3657 static int cgroup_file_open(struct kernfs_open_file *of)
3659 struct cftype *cft = of->kn->priv;
3662 return cft->open(of);
3666 static void cgroup_file_release(struct kernfs_open_file *of)
3668 struct cftype *cft = of->kn->priv;
3674 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3675 size_t nbytes, loff_t off)
3677 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3678 struct cgroup *cgrp = of->kn->parent->priv;
3679 struct cftype *cft = of->kn->priv;
3680 struct cgroup_subsys_state *css;
3684 * If namespaces are delegation boundaries, disallow writes to
3685 * files in an non-init namespace root from inside the namespace
3686 * except for the files explicitly marked delegatable -
3687 * cgroup.procs and cgroup.subtree_control.
3689 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3690 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3691 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3695 return cft->write(of, buf, nbytes, off);
3698 * kernfs guarantees that a file isn't deleted with operations in
3699 * flight, which means that the matching css is and stays alive and
3700 * doesn't need to be pinned. The RCU locking is not necessary
3701 * either. It's just for the convenience of using cgroup_css().
3704 css = cgroup_css(cgrp, cft->ss);
3707 if (cft->write_u64) {
3708 unsigned long long v;
3709 ret = kstrtoull(buf, 0, &v);
3711 ret = cft->write_u64(css, cft, v);
3712 } else if (cft->write_s64) {
3714 ret = kstrtoll(buf, 0, &v);
3716 ret = cft->write_s64(css, cft, v);
3721 return ret ?: nbytes;
3724 static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
3726 struct cftype *cft = of->kn->priv;
3729 return cft->poll(of, pt);
3731 return kernfs_generic_poll(of, pt);
3734 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3736 return seq_cft(seq)->seq_start(seq, ppos);
3739 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3741 return seq_cft(seq)->seq_next(seq, v, ppos);
3744 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3746 if (seq_cft(seq)->seq_stop)
3747 seq_cft(seq)->seq_stop(seq, v);
3750 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3752 struct cftype *cft = seq_cft(m);
3753 struct cgroup_subsys_state *css = seq_css(m);
3756 return cft->seq_show(m, arg);
3759 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3760 else if (cft->read_s64)
3761 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3767 static struct kernfs_ops cgroup_kf_single_ops = {
3768 .atomic_write_len = PAGE_SIZE,
3769 .open = cgroup_file_open,
3770 .release = cgroup_file_release,
3771 .write = cgroup_file_write,
3772 .poll = cgroup_file_poll,
3773 .seq_show = cgroup_seqfile_show,
3776 static struct kernfs_ops cgroup_kf_ops = {
3777 .atomic_write_len = PAGE_SIZE,
3778 .open = cgroup_file_open,
3779 .release = cgroup_file_release,
3780 .write = cgroup_file_write,
3781 .poll = cgroup_file_poll,
3782 .seq_start = cgroup_seqfile_start,
3783 .seq_next = cgroup_seqfile_next,
3784 .seq_stop = cgroup_seqfile_stop,
3785 .seq_show = cgroup_seqfile_show,
3788 /* set uid and gid of cgroup dirs and files to that of the creator */
3789 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3791 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3792 .ia_uid = current_fsuid(),
3793 .ia_gid = current_fsgid(), };
3795 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3796 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3799 return kernfs_setattr(kn, &iattr);
3802 static void cgroup_file_notify_timer(struct timer_list *timer)
3804 cgroup_file_notify(container_of(timer, struct cgroup_file,
3808 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3811 char name[CGROUP_FILE_NAME_MAX];
3812 struct kernfs_node *kn;
3813 struct lock_class_key *key = NULL;
3816 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3817 key = &cft->lockdep_key;
3819 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3820 cgroup_file_mode(cft),
3821 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
3822 0, cft->kf_ops, cft,
3827 ret = cgroup_kn_set_ugid(kn);
3833 if (cft->file_offset) {
3834 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3836 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
3838 spin_lock_irq(&cgroup_file_kn_lock);
3840 spin_unlock_irq(&cgroup_file_kn_lock);
3847 * cgroup_addrm_files - add or remove files to a cgroup directory
3848 * @css: the target css
3849 * @cgrp: the target cgroup (usually css->cgroup)
3850 * @cfts: array of cftypes to be added
3851 * @is_add: whether to add or remove
3853 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3854 * For removals, this function never fails.
3856 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3857 struct cgroup *cgrp, struct cftype cfts[],
3860 struct cftype *cft, *cft_end = NULL;
3863 lockdep_assert_held(&cgroup_mutex);
3866 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3867 /* does cft->flags tell us to skip this file on @cgrp? */
3868 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3870 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3872 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3874 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3876 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
3879 ret = cgroup_add_file(css, cgrp, cft);
3881 pr_warn("%s: failed to add %s, err=%d\n",
3882 __func__, cft->name, ret);
3888 cgroup_rm_file(cgrp, cft);
3894 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3896 struct cgroup_subsys *ss = cfts[0].ss;
3897 struct cgroup *root = &ss->root->cgrp;
3898 struct cgroup_subsys_state *css;
3901 lockdep_assert_held(&cgroup_mutex);
3903 /* add/rm files for all cgroups created before */
3904 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3905 struct cgroup *cgrp = css->cgroup;
3907 if (!(css->flags & CSS_VISIBLE))
3910 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3916 kernfs_activate(root->kn);
3920 static void cgroup_exit_cftypes(struct cftype *cfts)
3924 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3925 /* free copy for custom atomic_write_len, see init_cftypes() */
3926 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3931 /* revert flags set by cgroup core while adding @cfts */
3932 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3936 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3940 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3941 struct kernfs_ops *kf_ops;
3943 WARN_ON(cft->ss || cft->kf_ops);
3946 kf_ops = &cgroup_kf_ops;
3948 kf_ops = &cgroup_kf_single_ops;
3951 * Ugh... if @cft wants a custom max_write_len, we need to
3952 * make a copy of kf_ops to set its atomic_write_len.
3954 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3955 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3957 cgroup_exit_cftypes(cfts);
3960 kf_ops->atomic_write_len = cft->max_write_len;
3963 cft->kf_ops = kf_ops;
3970 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3972 lockdep_assert_held(&cgroup_mutex);
3974 if (!cfts || !cfts[0].ss)
3977 list_del(&cfts->node);
3978 cgroup_apply_cftypes(cfts, false);
3979 cgroup_exit_cftypes(cfts);
3984 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3985 * @cfts: zero-length name terminated array of cftypes
3987 * Unregister @cfts. Files described by @cfts are removed from all
3988 * existing cgroups and all future cgroups won't have them either. This
3989 * function can be called anytime whether @cfts' subsys is attached or not.
3991 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3994 int cgroup_rm_cftypes(struct cftype *cfts)
3998 mutex_lock(&cgroup_mutex);
3999 ret = cgroup_rm_cftypes_locked(cfts);
4000 mutex_unlock(&cgroup_mutex);
4005 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4006 * @ss: target cgroup subsystem
4007 * @cfts: zero-length name terminated array of cftypes
4009 * Register @cfts to @ss. Files described by @cfts are created for all
4010 * existing cgroups to which @ss is attached and all future cgroups will
4011 * have them too. This function can be called anytime whether @ss is
4014 * Returns 0 on successful registration, -errno on failure. Note that this
4015 * function currently returns 0 as long as @cfts registration is successful
4016 * even if some file creation attempts on existing cgroups fail.
4018 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4022 if (!cgroup_ssid_enabled(ss->id))
4025 if (!cfts || cfts[0].name[0] == '\0')
4028 ret = cgroup_init_cftypes(ss, cfts);
4032 mutex_lock(&cgroup_mutex);
4034 list_add_tail(&cfts->node, &ss->cfts);
4035 ret = cgroup_apply_cftypes(cfts, true);
4037 cgroup_rm_cftypes_locked(cfts);
4039 mutex_unlock(&cgroup_mutex);
4044 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4045 * @ss: target cgroup subsystem
4046 * @cfts: zero-length name terminated array of cftypes
4048 * Similar to cgroup_add_cftypes() but the added files are only used for
4049 * the default hierarchy.
4051 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4055 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4056 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4057 return cgroup_add_cftypes(ss, cfts);
4061 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4062 * @ss: target cgroup subsystem
4063 * @cfts: zero-length name terminated array of cftypes
4065 * Similar to cgroup_add_cftypes() but the added files are only used for
4066 * the legacy hierarchies.
4068 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4072 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4073 cft->flags |= __CFTYPE_NOT_ON_DFL;
4074 return cgroup_add_cftypes(ss, cfts);
4078 * cgroup_file_notify - generate a file modified event for a cgroup_file
4079 * @cfile: target cgroup_file
4081 * @cfile must have been obtained by setting cftype->file_offset.
4083 void cgroup_file_notify(struct cgroup_file *cfile)
4085 unsigned long flags;
4087 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4089 unsigned long last = cfile->notified_at;
4090 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4092 if (time_in_range(jiffies, last, next)) {
4093 timer_reduce(&cfile->notify_timer, next);
4095 kernfs_notify(cfile->kn);
4096 cfile->notified_at = jiffies;
4099 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4103 * css_next_child - find the next child of a given css
4104 * @pos: the current position (%NULL to initiate traversal)
4105 * @parent: css whose children to walk
4107 * This function returns the next child of @parent and should be called
4108 * under either cgroup_mutex or RCU read lock. The only requirement is
4109 * that @parent and @pos are accessible. The next sibling is guaranteed to
4110 * be returned regardless of their states.
4112 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4113 * css which finished ->css_online() is guaranteed to be visible in the
4114 * future iterations and will stay visible until the last reference is put.
4115 * A css which hasn't finished ->css_online() or already finished
4116 * ->css_offline() may show up during traversal. It's each subsystem's
4117 * responsibility to synchronize against on/offlining.
4119 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4120 struct cgroup_subsys_state *parent)
4122 struct cgroup_subsys_state *next;
4124 cgroup_assert_mutex_or_rcu_locked();
4127 * @pos could already have been unlinked from the sibling list.
4128 * Once a cgroup is removed, its ->sibling.next is no longer
4129 * updated when its next sibling changes. CSS_RELEASED is set when
4130 * @pos is taken off list, at which time its next pointer is valid,
4131 * and, as releases are serialized, the one pointed to by the next
4132 * pointer is guaranteed to not have started release yet. This
4133 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4134 * critical section, the one pointed to by its next pointer is
4135 * guaranteed to not have finished its RCU grace period even if we
4136 * have dropped rcu_read_lock() inbetween iterations.
4138 * If @pos has CSS_RELEASED set, its next pointer can't be
4139 * dereferenced; however, as each css is given a monotonically
4140 * increasing unique serial number and always appended to the
4141 * sibling list, the next one can be found by walking the parent's
4142 * children until the first css with higher serial number than
4143 * @pos's. While this path can be slower, it happens iff iteration
4144 * races against release and the race window is very small.
4147 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4148 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4149 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4151 list_for_each_entry_rcu(next, &parent->children, sibling)
4152 if (next->serial_nr > pos->serial_nr)
4157 * @next, if not pointing to the head, can be dereferenced and is
4160 if (&next->sibling != &parent->children)
4166 * css_next_descendant_pre - find the next descendant for pre-order walk
4167 * @pos: the current position (%NULL to initiate traversal)
4168 * @root: css whose descendants to walk
4170 * To be used by css_for_each_descendant_pre(). Find the next descendant
4171 * to visit for pre-order traversal of @root's descendants. @root is
4172 * included in the iteration and the first node to be visited.
4174 * While this function requires cgroup_mutex or RCU read locking, it
4175 * doesn't require the whole traversal to be contained in a single critical
4176 * section. This function will return the correct next descendant as long
4177 * as both @pos and @root are accessible and @pos is a descendant of @root.
4179 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4180 * css which finished ->css_online() is guaranteed to be visible in the
4181 * future iterations and will stay visible until the last reference is put.
4182 * A css which hasn't finished ->css_online() or already finished
4183 * ->css_offline() may show up during traversal. It's each subsystem's
4184 * responsibility to synchronize against on/offlining.
4186 struct cgroup_subsys_state *
4187 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4188 struct cgroup_subsys_state *root)
4190 struct cgroup_subsys_state *next;
4192 cgroup_assert_mutex_or_rcu_locked();
4194 /* if first iteration, visit @root */
4198 /* visit the first child if exists */
4199 next = css_next_child(NULL, pos);
4203 /* no child, visit my or the closest ancestor's next sibling */
4204 while (pos != root) {
4205 next = css_next_child(pos, pos->parent);
4213 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4216 * css_rightmost_descendant - return the rightmost descendant of a css
4217 * @pos: css of interest
4219 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4220 * is returned. This can be used during pre-order traversal to skip
4223 * While this function requires cgroup_mutex or RCU read locking, it
4224 * doesn't require the whole traversal to be contained in a single critical
4225 * section. This function will return the correct rightmost descendant as
4226 * long as @pos is accessible.
4228 struct cgroup_subsys_state *
4229 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4231 struct cgroup_subsys_state *last, *tmp;
4233 cgroup_assert_mutex_or_rcu_locked();
4237 /* ->prev isn't RCU safe, walk ->next till the end */
4239 css_for_each_child(tmp, last)
4246 static struct cgroup_subsys_state *
4247 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4249 struct cgroup_subsys_state *last;
4253 pos = css_next_child(NULL, pos);
4260 * css_next_descendant_post - find the next descendant for post-order walk
4261 * @pos: the current position (%NULL to initiate traversal)
4262 * @root: css whose descendants to walk
4264 * To be used by css_for_each_descendant_post(). Find the next descendant
4265 * to visit for post-order traversal of @root's descendants. @root is
4266 * included in the iteration and the last node to be visited.
4268 * While this function requires cgroup_mutex or RCU read locking, it
4269 * doesn't require the whole traversal to be contained in a single critical
4270 * section. This function will return the correct next descendant as long
4271 * as both @pos and @cgroup are accessible and @pos is a descendant of
4274 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4275 * css which finished ->css_online() is guaranteed to be visible in the
4276 * future iterations and will stay visible until the last reference is put.
4277 * A css which hasn't finished ->css_online() or already finished
4278 * ->css_offline() may show up during traversal. It's each subsystem's
4279 * responsibility to synchronize against on/offlining.
4281 struct cgroup_subsys_state *
4282 css_next_descendant_post(struct cgroup_subsys_state *pos,
4283 struct cgroup_subsys_state *root)
4285 struct cgroup_subsys_state *next;
4287 cgroup_assert_mutex_or_rcu_locked();
4289 /* if first iteration, visit leftmost descendant which may be @root */
4291 return css_leftmost_descendant(root);
4293 /* if we visited @root, we're done */
4297 /* if there's an unvisited sibling, visit its leftmost descendant */
4298 next = css_next_child(pos, pos->parent);
4300 return css_leftmost_descendant(next);
4302 /* no sibling left, visit parent */
4307 * css_has_online_children - does a css have online children
4308 * @css: the target css
4310 * Returns %true if @css has any online children; otherwise, %false. This
4311 * function can be called from any context but the caller is responsible
4312 * for synchronizing against on/offlining as necessary.
4314 bool css_has_online_children(struct cgroup_subsys_state *css)
4316 struct cgroup_subsys_state *child;
4320 css_for_each_child(child, css) {
4321 if (child->flags & CSS_ONLINE) {
4330 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4332 struct list_head *l;
4333 struct cgrp_cset_link *link;
4334 struct css_set *cset;
4336 lockdep_assert_held(&css_set_lock);
4338 /* find the next threaded cset */
4339 if (it->tcset_pos) {
4340 l = it->tcset_pos->next;
4342 if (l != it->tcset_head) {
4344 return container_of(l, struct css_set,
4345 threaded_csets_node);
4348 it->tcset_pos = NULL;
4351 /* find the next cset */
4354 if (l == it->cset_head) {
4355 it->cset_pos = NULL;
4360 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4362 link = list_entry(l, struct cgrp_cset_link, cset_link);
4368 /* initialize threaded css_set walking */
4369 if (it->flags & CSS_TASK_ITER_THREADED) {
4371 put_css_set_locked(it->cur_dcset);
4372 it->cur_dcset = cset;
4375 it->tcset_head = &cset->threaded_csets;
4376 it->tcset_pos = &cset->threaded_csets;
4383 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4384 * @it: the iterator to advance
4386 * Advance @it to the next css_set to walk.
4388 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4390 struct css_set *cset;
4392 lockdep_assert_held(&css_set_lock);
4394 /* Advance to the next non-empty css_set */
4396 cset = css_task_iter_next_css_set(it);
4398 it->task_pos = NULL;
4401 } while (!css_set_populated(cset) && list_empty(&cset->dying_tasks));
4403 if (!list_empty(&cset->tasks))
4404 it->task_pos = cset->tasks.next;
4405 else if (!list_empty(&cset->mg_tasks))
4406 it->task_pos = cset->mg_tasks.next;
4408 it->task_pos = cset->dying_tasks.next;
4410 it->tasks_head = &cset->tasks;
4411 it->mg_tasks_head = &cset->mg_tasks;
4412 it->dying_tasks_head = &cset->dying_tasks;
4415 * We don't keep css_sets locked across iteration steps and thus
4416 * need to take steps to ensure that iteration can be resumed after
4417 * the lock is re-acquired. Iteration is performed at two levels -
4418 * css_sets and tasks in them.
4420 * Once created, a css_set never leaves its cgroup lists, so a
4421 * pinned css_set is guaranteed to stay put and we can resume
4422 * iteration afterwards.
4424 * Tasks may leave @cset across iteration steps. This is resolved
4425 * by registering each iterator with the css_set currently being
4426 * walked and making css_set_move_task() advance iterators whose
4427 * next task is leaving.
4430 list_del(&it->iters_node);
4431 put_css_set_locked(it->cur_cset);
4434 it->cur_cset = cset;
4435 list_add(&it->iters_node, &cset->task_iters);
4438 static void css_task_iter_skip(struct css_task_iter *it,
4439 struct task_struct *task)
4441 lockdep_assert_held(&css_set_lock);
4443 if (it->task_pos == &task->cg_list) {
4444 it->task_pos = it->task_pos->next;
4445 it->flags |= CSS_TASK_ITER_SKIPPED;
4449 static void css_task_iter_advance(struct css_task_iter *it)
4451 struct task_struct *task;
4453 lockdep_assert_held(&css_set_lock);
4457 * Advance iterator to find next entry. cset->tasks is
4458 * consumed first and then ->mg_tasks. After ->mg_tasks,
4459 * we move onto the next cset.
4461 if (it->flags & CSS_TASK_ITER_SKIPPED)
4462 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4464 it->task_pos = it->task_pos->next;
4466 if (it->task_pos == it->tasks_head)
4467 it->task_pos = it->mg_tasks_head->next;
4468 if (it->task_pos == it->mg_tasks_head)
4469 it->task_pos = it->dying_tasks_head->next;
4470 if (it->task_pos == it->dying_tasks_head)
4471 css_task_iter_advance_css_set(it);
4473 /* called from start, proceed to the first cset */
4474 css_task_iter_advance_css_set(it);
4480 task = list_entry(it->task_pos, struct task_struct, cg_list);
4482 if (it->flags & CSS_TASK_ITER_PROCS) {
4483 /* if PROCS, skip over tasks which aren't group leaders */
4484 if (!thread_group_leader(task))
4487 /* and dying leaders w/o live member threads */
4488 if (!atomic_read(&task->signal->live))
4491 /* skip all dying ones */
4492 if (task->flags & PF_EXITING)
4498 * css_task_iter_start - initiate task iteration
4499 * @css: the css to walk tasks of
4500 * @flags: CSS_TASK_ITER_* flags
4501 * @it: the task iterator to use
4503 * Initiate iteration through the tasks of @css. The caller can call
4504 * css_task_iter_next() to walk through the tasks until the function
4505 * returns NULL. On completion of iteration, css_task_iter_end() must be
4508 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4509 struct css_task_iter *it)
4511 memset(it, 0, sizeof(*it));
4513 spin_lock_irq(&css_set_lock);
4519 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4521 it->cset_pos = &css->cgroup->cset_links;
4523 it->cset_head = it->cset_pos;
4525 css_task_iter_advance(it);
4527 spin_unlock_irq(&css_set_lock);
4531 * css_task_iter_next - return the next task for the iterator
4532 * @it: the task iterator being iterated
4534 * The "next" function for task iteration. @it should have been
4535 * initialized via css_task_iter_start(). Returns NULL when the iteration
4538 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4541 put_task_struct(it->cur_task);
4542 it->cur_task = NULL;
4545 spin_lock_irq(&css_set_lock);
4547 /* @it may be half-advanced by skips, finish advancing */
4548 if (it->flags & CSS_TASK_ITER_SKIPPED)
4549 css_task_iter_advance(it);
4552 it->cur_task = list_entry(it->task_pos, struct task_struct,
4554 get_task_struct(it->cur_task);
4555 css_task_iter_advance(it);
4558 spin_unlock_irq(&css_set_lock);
4560 return it->cur_task;
4564 * css_task_iter_end - finish task iteration
4565 * @it: the task iterator to finish
4567 * Finish task iteration started by css_task_iter_start().
4569 void css_task_iter_end(struct css_task_iter *it)
4572 spin_lock_irq(&css_set_lock);
4573 list_del(&it->iters_node);
4574 put_css_set_locked(it->cur_cset);
4575 spin_unlock_irq(&css_set_lock);
4579 put_css_set(it->cur_dcset);
4582 put_task_struct(it->cur_task);
4585 static void cgroup_procs_release(struct kernfs_open_file *of)
4588 css_task_iter_end(of->priv);
4593 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4595 struct kernfs_open_file *of = s->private;
4596 struct css_task_iter *it = of->priv;
4598 return css_task_iter_next(it);
4601 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4602 unsigned int iter_flags)
4604 struct kernfs_open_file *of = s->private;
4605 struct cgroup *cgrp = seq_css(s)->cgroup;
4606 struct css_task_iter *it = of->priv;
4609 * When a seq_file is seeked, it's always traversed sequentially
4610 * from position 0, so we can simply keep iterating on !0 *pos.
4613 if (WARN_ON_ONCE((*pos)++))
4614 return ERR_PTR(-EINVAL);
4616 it = kzalloc(sizeof(*it), GFP_KERNEL);
4618 return ERR_PTR(-ENOMEM);
4620 css_task_iter_start(&cgrp->self, iter_flags, it);
4621 } else if (!(*pos)++) {
4622 css_task_iter_end(it);
4623 css_task_iter_start(&cgrp->self, iter_flags, it);
4626 return cgroup_procs_next(s, NULL, NULL);
4629 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4631 struct cgroup *cgrp = seq_css(s)->cgroup;
4634 * All processes of a threaded subtree belong to the domain cgroup
4635 * of the subtree. Only threads can be distributed across the
4636 * subtree. Reject reads on cgroup.procs in the subtree proper.
4637 * They're always empty anyway.
4639 if (cgroup_is_threaded(cgrp))
4640 return ERR_PTR(-EOPNOTSUPP);
4642 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4643 CSS_TASK_ITER_THREADED);
4646 static int cgroup_procs_show(struct seq_file *s, void *v)
4648 seq_printf(s, "%d\n", task_pid_vnr(v));
4652 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4653 struct cgroup *dst_cgrp,
4654 struct super_block *sb)
4656 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4657 struct cgroup *com_cgrp = src_cgrp;
4658 struct inode *inode;
4661 lockdep_assert_held(&cgroup_mutex);
4663 /* find the common ancestor */
4664 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4665 com_cgrp = cgroup_parent(com_cgrp);
4667 /* %current should be authorized to migrate to the common ancestor */
4668 inode = kernfs_get_inode(sb, com_cgrp->procs_file.kn);
4672 ret = inode_permission(inode, MAY_WRITE);
4678 * If namespaces are delegation boundaries, %current must be able
4679 * to see both source and destination cgroups from its namespace.
4681 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4682 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4683 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4689 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4690 char *buf, size_t nbytes, loff_t off)
4692 struct cgroup *src_cgrp, *dst_cgrp;
4693 struct task_struct *task;
4697 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4701 task = cgroup_procs_write_start(buf, true, &locked);
4702 ret = PTR_ERR_OR_ZERO(task);
4706 /* find the source cgroup */
4707 spin_lock_irq(&css_set_lock);
4708 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4709 spin_unlock_irq(&css_set_lock);
4711 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4712 of->file->f_path.dentry->d_sb);
4716 ret = cgroup_attach_task(dst_cgrp, task, true);
4719 cgroup_procs_write_finish(task, locked);
4721 cgroup_kn_unlock(of->kn);
4723 return ret ?: nbytes;
4726 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4728 return __cgroup_procs_start(s, pos, 0);
4731 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4732 char *buf, size_t nbytes, loff_t off)
4734 struct cgroup *src_cgrp, *dst_cgrp;
4735 struct task_struct *task;
4739 buf = strstrip(buf);
4741 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4745 task = cgroup_procs_write_start(buf, false, &locked);
4746 ret = PTR_ERR_OR_ZERO(task);
4750 /* find the source cgroup */
4751 spin_lock_irq(&css_set_lock);
4752 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4753 spin_unlock_irq(&css_set_lock);
4755 /* thread migrations follow the cgroup.procs delegation rule */
4756 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4757 of->file->f_path.dentry->d_sb);
4761 /* and must be contained in the same domain */
4763 if (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)
4766 ret = cgroup_attach_task(dst_cgrp, task, false);
4769 cgroup_procs_write_finish(task, locked);
4771 cgroup_kn_unlock(of->kn);
4773 return ret ?: nbytes;
4776 /* cgroup core interface files for the default hierarchy */
4777 static struct cftype cgroup_base_files[] = {
4779 .name = "cgroup.type",
4780 .flags = CFTYPE_NOT_ON_ROOT,
4781 .seq_show = cgroup_type_show,
4782 .write = cgroup_type_write,
4785 .name = "cgroup.procs",
4786 .flags = CFTYPE_NS_DELEGATABLE,
4787 .file_offset = offsetof(struct cgroup, procs_file),
4788 .release = cgroup_procs_release,
4789 .seq_start = cgroup_procs_start,
4790 .seq_next = cgroup_procs_next,
4791 .seq_show = cgroup_procs_show,
4792 .write = cgroup_procs_write,
4795 .name = "cgroup.threads",
4796 .flags = CFTYPE_NS_DELEGATABLE,
4797 .release = cgroup_procs_release,
4798 .seq_start = cgroup_threads_start,
4799 .seq_next = cgroup_procs_next,
4800 .seq_show = cgroup_procs_show,
4801 .write = cgroup_threads_write,
4804 .name = "cgroup.controllers",
4805 .seq_show = cgroup_controllers_show,
4808 .name = "cgroup.subtree_control",
4809 .flags = CFTYPE_NS_DELEGATABLE,
4810 .seq_show = cgroup_subtree_control_show,
4811 .write = cgroup_subtree_control_write,
4814 .name = "cgroup.events",
4815 .flags = CFTYPE_NOT_ON_ROOT,
4816 .file_offset = offsetof(struct cgroup, events_file),
4817 .seq_show = cgroup_events_show,
4820 .name = "cgroup.max.descendants",
4821 .seq_show = cgroup_max_descendants_show,
4822 .write = cgroup_max_descendants_write,
4825 .name = "cgroup.max.depth",
4826 .seq_show = cgroup_max_depth_show,
4827 .write = cgroup_max_depth_write,
4830 .name = "cgroup.stat",
4831 .seq_show = cgroup_stat_show,
4834 .name = "cgroup.freeze",
4835 .flags = CFTYPE_NOT_ON_ROOT,
4836 .seq_show = cgroup_freeze_show,
4837 .write = cgroup_freeze_write,
4841 .flags = CFTYPE_NOT_ON_ROOT,
4842 .seq_show = cpu_stat_show,
4846 .name = "io.pressure",
4847 .seq_show = cgroup_io_pressure_show,
4848 .write = cgroup_io_pressure_write,
4849 .poll = cgroup_pressure_poll,
4850 .release = cgroup_pressure_release,
4853 .name = "memory.pressure",
4854 .seq_show = cgroup_memory_pressure_show,
4855 .write = cgroup_memory_pressure_write,
4856 .poll = cgroup_pressure_poll,
4857 .release = cgroup_pressure_release,
4860 .name = "cpu.pressure",
4861 .seq_show = cgroup_cpu_pressure_show,
4862 .write = cgroup_cpu_pressure_write,
4863 .poll = cgroup_pressure_poll,
4864 .release = cgroup_pressure_release,
4866 #endif /* CONFIG_PSI */
4871 * css destruction is four-stage process.
4873 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4874 * Implemented in kill_css().
4876 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4877 * and thus css_tryget_online() is guaranteed to fail, the css can be
4878 * offlined by invoking offline_css(). After offlining, the base ref is
4879 * put. Implemented in css_killed_work_fn().
4881 * 3. When the percpu_ref reaches zero, the only possible remaining
4882 * accessors are inside RCU read sections. css_release() schedules the
4885 * 4. After the grace period, the css can be freed. Implemented in
4886 * css_free_work_fn().
4888 * It is actually hairier because both step 2 and 4 require process context
4889 * and thus involve punting to css->destroy_work adding two additional
4890 * steps to the already complex sequence.
4892 static void css_free_rwork_fn(struct work_struct *work)
4894 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
4895 struct cgroup_subsys_state, destroy_rwork);
4896 struct cgroup_subsys *ss = css->ss;
4897 struct cgroup *cgrp = css->cgroup;
4899 percpu_ref_exit(&css->refcnt);
4903 struct cgroup_subsys_state *parent = css->parent;
4907 cgroup_idr_remove(&ss->css_idr, id);
4913 /* cgroup free path */
4914 atomic_dec(&cgrp->root->nr_cgrps);
4915 cgroup1_pidlist_destroy_all(cgrp);
4916 cancel_work_sync(&cgrp->release_agent_work);
4918 if (cgroup_parent(cgrp)) {
4920 * We get a ref to the parent, and put the ref when
4921 * this cgroup is being freed, so it's guaranteed
4922 * that the parent won't be destroyed before its
4925 cgroup_put(cgroup_parent(cgrp));
4926 kernfs_put(cgrp->kn);
4927 psi_cgroup_free(cgrp);
4928 if (cgroup_on_dfl(cgrp))
4929 cgroup_rstat_exit(cgrp);
4933 * This is root cgroup's refcnt reaching zero,
4934 * which indicates that the root should be
4937 cgroup_destroy_root(cgrp->root);
4942 static void css_release_work_fn(struct work_struct *work)
4944 struct cgroup_subsys_state *css =
4945 container_of(work, struct cgroup_subsys_state, destroy_work);
4946 struct cgroup_subsys *ss = css->ss;
4947 struct cgroup *cgrp = css->cgroup;
4949 mutex_lock(&cgroup_mutex);
4951 css->flags |= CSS_RELEASED;
4952 list_del_rcu(&css->sibling);
4955 /* css release path */
4956 if (!list_empty(&css->rstat_css_node)) {
4957 cgroup_rstat_flush(cgrp);
4958 list_del_rcu(&css->rstat_css_node);
4961 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4962 if (ss->css_released)
4963 ss->css_released(css);
4965 struct cgroup *tcgrp;
4967 /* cgroup release path */
4968 TRACE_CGROUP_PATH(release, cgrp);
4970 if (cgroup_on_dfl(cgrp))
4971 cgroup_rstat_flush(cgrp);
4973 spin_lock_irq(&css_set_lock);
4974 for (tcgrp = cgroup_parent(cgrp); tcgrp;
4975 tcgrp = cgroup_parent(tcgrp))
4976 tcgrp->nr_dying_descendants--;
4977 spin_unlock_irq(&css_set_lock);
4980 * There are two control paths which try to determine
4981 * cgroup from dentry without going through kernfs -
4982 * cgroupstats_build() and css_tryget_online_from_dir().
4983 * Those are supported by RCU protecting clearing of
4984 * cgrp->kn->priv backpointer.
4987 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4991 mutex_unlock(&cgroup_mutex);
4993 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
4994 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
4997 static void css_release(struct percpu_ref *ref)
4999 struct cgroup_subsys_state *css =
5000 container_of(ref, struct cgroup_subsys_state, refcnt);
5002 INIT_WORK(&css->destroy_work, css_release_work_fn);
5003 queue_work(cgroup_destroy_wq, &css->destroy_work);
5006 static void init_and_link_css(struct cgroup_subsys_state *css,
5007 struct cgroup_subsys *ss, struct cgroup *cgrp)
5009 lockdep_assert_held(&cgroup_mutex);
5011 cgroup_get_live(cgrp);
5013 memset(css, 0, sizeof(*css));
5017 INIT_LIST_HEAD(&css->sibling);
5018 INIT_LIST_HEAD(&css->children);
5019 INIT_LIST_HEAD(&css->rstat_css_node);
5020 css->serial_nr = css_serial_nr_next++;
5021 atomic_set(&css->online_cnt, 0);
5023 if (cgroup_parent(cgrp)) {
5024 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5025 css_get(css->parent);
5028 if (cgroup_on_dfl(cgrp) && ss->css_rstat_flush)
5029 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5031 BUG_ON(cgroup_css(cgrp, ss));
5034 /* invoke ->css_online() on a new CSS and mark it online if successful */
5035 static int online_css(struct cgroup_subsys_state *css)
5037 struct cgroup_subsys *ss = css->ss;
5040 lockdep_assert_held(&cgroup_mutex);
5043 ret = ss->css_online(css);
5045 css->flags |= CSS_ONLINE;
5046 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5048 atomic_inc(&css->online_cnt);
5050 atomic_inc(&css->parent->online_cnt);
5055 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5056 static void offline_css(struct cgroup_subsys_state *css)
5058 struct cgroup_subsys *ss = css->ss;
5060 lockdep_assert_held(&cgroup_mutex);
5062 if (!(css->flags & CSS_ONLINE))
5065 if (ss->css_offline)
5066 ss->css_offline(css);
5068 css->flags &= ~CSS_ONLINE;
5069 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5071 wake_up_all(&css->cgroup->offline_waitq);
5075 * css_create - create a cgroup_subsys_state
5076 * @cgrp: the cgroup new css will be associated with
5077 * @ss: the subsys of new css
5079 * Create a new css associated with @cgrp - @ss pair. On success, the new
5080 * css is online and installed in @cgrp. This function doesn't create the
5081 * interface files. Returns 0 on success, -errno on failure.
5083 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5084 struct cgroup_subsys *ss)
5086 struct cgroup *parent = cgroup_parent(cgrp);
5087 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5088 struct cgroup_subsys_state *css;
5091 lockdep_assert_held(&cgroup_mutex);
5093 css = ss->css_alloc(parent_css);
5095 css = ERR_PTR(-ENOMEM);
5099 init_and_link_css(css, ss, cgrp);
5101 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5105 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5110 /* @css is ready to be brought online now, make it visible */
5111 list_add_tail_rcu(&css->sibling, &parent_css->children);
5112 cgroup_idr_replace(&ss->css_idr, css, css->id);
5114 err = online_css(css);
5118 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5119 cgroup_parent(parent)) {
5120 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5121 current->comm, current->pid, ss->name);
5122 if (!strcmp(ss->name, "memory"))
5123 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5124 ss->warned_broken_hierarchy = true;
5130 list_del_rcu(&css->sibling);
5132 list_del_rcu(&css->rstat_css_node);
5133 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5134 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5135 return ERR_PTR(err);
5139 * The returned cgroup is fully initialized including its control mask, but
5140 * it isn't associated with its kernfs_node and doesn't have the control
5143 static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5146 struct cgroup_root *root = parent->root;
5147 struct cgroup *cgrp, *tcgrp;
5148 struct kernfs_node *kn;
5149 int level = parent->level + 1;
5152 /* allocate the cgroup and its ID, 0 is reserved for the root */
5153 cgrp = kzalloc(struct_size(cgrp, ancestor_ids, (level + 1)),
5156 return ERR_PTR(-ENOMEM);
5158 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5162 if (cgroup_on_dfl(parent)) {
5163 ret = cgroup_rstat_init(cgrp);
5165 goto out_cancel_ref;
5168 /* create the directory */
5169 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5176 init_cgroup_housekeeping(cgrp);
5178 cgrp->self.parent = &parent->self;
5180 cgrp->level = level;
5182 ret = psi_cgroup_alloc(cgrp);
5184 goto out_kernfs_remove;
5186 ret = cgroup_bpf_inherit(cgrp);
5191 * New cgroup inherits effective freeze counter, and
5192 * if the parent has to be frozen, the child has too.
5194 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5195 if (cgrp->freezer.e_freeze) {
5197 * Set the CGRP_FREEZE flag, so when a process will be
5198 * attached to the child cgroup, it will become frozen.
5199 * At this point the new cgroup is unpopulated, so we can
5200 * consider it frozen immediately.
5202 set_bit(CGRP_FREEZE, &cgrp->flags);
5203 set_bit(CGRP_FROZEN, &cgrp->flags);
5206 spin_lock_irq(&css_set_lock);
5207 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5208 cgrp->ancestor_ids[tcgrp->level] = cgroup_id(tcgrp);
5210 if (tcgrp != cgrp) {
5211 tcgrp->nr_descendants++;
5214 * If the new cgroup is frozen, all ancestor cgroups
5215 * get a new frozen descendant, but their state can't
5216 * change because of this.
5218 if (cgrp->freezer.e_freeze)
5219 tcgrp->freezer.nr_frozen_descendants++;
5222 spin_unlock_irq(&css_set_lock);
5224 if (notify_on_release(parent))
5225 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5227 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5228 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5230 cgrp->self.serial_nr = css_serial_nr_next++;
5232 /* allocation complete, commit to creation */
5233 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5234 atomic_inc(&root->nr_cgrps);
5235 cgroup_get_live(parent);
5238 * On the default hierarchy, a child doesn't automatically inherit
5239 * subtree_control from the parent. Each is configured manually.
5241 if (!cgroup_on_dfl(cgrp))
5242 cgrp->subtree_control = cgroup_control(cgrp);
5244 cgroup_propagate_control(cgrp);
5249 psi_cgroup_free(cgrp);
5251 kernfs_remove(cgrp->kn);
5253 if (cgroup_on_dfl(parent))
5254 cgroup_rstat_exit(cgrp);
5256 percpu_ref_exit(&cgrp->self.refcnt);
5259 return ERR_PTR(ret);
5262 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5264 struct cgroup *cgroup;
5268 lockdep_assert_held(&cgroup_mutex);
5270 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5271 if (cgroup->nr_descendants >= cgroup->max_descendants)
5274 if (level > cgroup->max_depth)
5285 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5287 struct cgroup *parent, *cgrp;
5290 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5291 if (strchr(name, '\n'))
5294 parent = cgroup_kn_lock_live(parent_kn, false);
5298 if (!cgroup_check_hierarchy_limits(parent)) {
5303 cgrp = cgroup_create(parent, name, mode);
5305 ret = PTR_ERR(cgrp);
5310 * This extra ref will be put in cgroup_free_fn() and guarantees
5311 * that @cgrp->kn is always accessible.
5313 kernfs_get(cgrp->kn);
5315 ret = cgroup_kn_set_ugid(cgrp->kn);
5319 ret = css_populate_dir(&cgrp->self);
5323 ret = cgroup_apply_control_enable(cgrp);
5327 TRACE_CGROUP_PATH(mkdir, cgrp);
5329 /* let's create and online css's */
5330 kernfs_activate(cgrp->kn);
5336 cgroup_destroy_locked(cgrp);
5338 cgroup_kn_unlock(parent_kn);
5343 * This is called when the refcnt of a css is confirmed to be killed.
5344 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5345 * initate destruction and put the css ref from kill_css().
5347 static void css_killed_work_fn(struct work_struct *work)
5349 struct cgroup_subsys_state *css =
5350 container_of(work, struct cgroup_subsys_state, destroy_work);
5352 mutex_lock(&cgroup_mutex);
5357 /* @css can't go away while we're holding cgroup_mutex */
5359 } while (css && atomic_dec_and_test(&css->online_cnt));
5361 mutex_unlock(&cgroup_mutex);
5364 /* css kill confirmation processing requires process context, bounce */
5365 static void css_killed_ref_fn(struct percpu_ref *ref)
5367 struct cgroup_subsys_state *css =
5368 container_of(ref, struct cgroup_subsys_state, refcnt);
5370 if (atomic_dec_and_test(&css->online_cnt)) {
5371 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5372 queue_work(cgroup_destroy_wq, &css->destroy_work);
5377 * kill_css - destroy a css
5378 * @css: css to destroy
5380 * This function initiates destruction of @css by removing cgroup interface
5381 * files and putting its base reference. ->css_offline() will be invoked
5382 * asynchronously once css_tryget_online() is guaranteed to fail and when
5383 * the reference count reaches zero, @css will be released.
5385 static void kill_css(struct cgroup_subsys_state *css)
5387 lockdep_assert_held(&cgroup_mutex);
5389 if (css->flags & CSS_DYING)
5392 css->flags |= CSS_DYING;
5395 * This must happen before css is disassociated with its cgroup.
5396 * See seq_css() for details.
5401 * Killing would put the base ref, but we need to keep it alive
5402 * until after ->css_offline().
5407 * cgroup core guarantees that, by the time ->css_offline() is
5408 * invoked, no new css reference will be given out via
5409 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5410 * proceed to offlining css's because percpu_ref_kill() doesn't
5411 * guarantee that the ref is seen as killed on all CPUs on return.
5413 * Use percpu_ref_kill_and_confirm() to get notifications as each
5414 * css is confirmed to be seen as killed on all CPUs.
5416 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5420 * cgroup_destroy_locked - the first stage of cgroup destruction
5421 * @cgrp: cgroup to be destroyed
5423 * css's make use of percpu refcnts whose killing latency shouldn't be
5424 * exposed to userland and are RCU protected. Also, cgroup core needs to
5425 * guarantee that css_tryget_online() won't succeed by the time
5426 * ->css_offline() is invoked. To satisfy all the requirements,
5427 * destruction is implemented in the following two steps.
5429 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5430 * userland visible parts and start killing the percpu refcnts of
5431 * css's. Set up so that the next stage will be kicked off once all
5432 * the percpu refcnts are confirmed to be killed.
5434 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5435 * rest of destruction. Once all cgroup references are gone, the
5436 * cgroup is RCU-freed.
5438 * This function implements s1. After this step, @cgrp is gone as far as
5439 * the userland is concerned and a new cgroup with the same name may be
5440 * created. As cgroup doesn't care about the names internally, this
5441 * doesn't cause any problem.
5443 static int cgroup_destroy_locked(struct cgroup *cgrp)
5444 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5446 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5447 struct cgroup_subsys_state *css;
5448 struct cgrp_cset_link *link;
5451 lockdep_assert_held(&cgroup_mutex);
5454 * Only migration can raise populated from zero and we're already
5455 * holding cgroup_mutex.
5457 if (cgroup_is_populated(cgrp))
5461 * Make sure there's no live children. We can't test emptiness of
5462 * ->self.children as dead children linger on it while being
5463 * drained; otherwise, "rmdir parent/child parent" may fail.
5465 if (css_has_online_children(&cgrp->self))
5469 * Mark @cgrp and the associated csets dead. The former prevents
5470 * further task migration and child creation by disabling
5471 * cgroup_lock_live_group(). The latter makes the csets ignored by
5472 * the migration path.
5474 cgrp->self.flags &= ~CSS_ONLINE;
5476 spin_lock_irq(&css_set_lock);
5477 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5478 link->cset->dead = true;
5479 spin_unlock_irq(&css_set_lock);
5481 /* initiate massacre of all css's */
5482 for_each_css(css, ssid, cgrp)
5485 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5486 css_clear_dir(&cgrp->self);
5487 kernfs_remove(cgrp->kn);
5489 if (parent && cgroup_is_threaded(cgrp))
5490 parent->nr_threaded_children--;
5492 spin_lock_irq(&css_set_lock);
5493 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5494 tcgrp->nr_descendants--;
5495 tcgrp->nr_dying_descendants++;
5497 * If the dying cgroup is frozen, decrease frozen descendants
5498 * counters of ancestor cgroups.
5500 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5501 tcgrp->freezer.nr_frozen_descendants--;
5503 spin_unlock_irq(&css_set_lock);
5505 cgroup1_check_for_release(parent);
5507 cgroup_bpf_offline(cgrp);
5509 /* put the base reference */
5510 percpu_ref_kill(&cgrp->self.refcnt);
5515 int cgroup_rmdir(struct kernfs_node *kn)
5517 struct cgroup *cgrp;
5520 cgrp = cgroup_kn_lock_live(kn, false);
5524 ret = cgroup_destroy_locked(cgrp);
5526 TRACE_CGROUP_PATH(rmdir, cgrp);
5528 cgroup_kn_unlock(kn);
5532 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5533 .show_options = cgroup_show_options,
5534 .mkdir = cgroup_mkdir,
5535 .rmdir = cgroup_rmdir,
5536 .show_path = cgroup_show_path,
5539 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5541 struct cgroup_subsys_state *css;
5543 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5545 mutex_lock(&cgroup_mutex);
5547 idr_init(&ss->css_idr);
5548 INIT_LIST_HEAD(&ss->cfts);
5550 /* Create the root cgroup state for this subsystem */
5551 ss->root = &cgrp_dfl_root;
5552 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5553 /* We don't handle early failures gracefully */
5554 BUG_ON(IS_ERR(css));
5555 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5558 * Root csses are never destroyed and we can't initialize
5559 * percpu_ref during early init. Disable refcnting.
5561 css->flags |= CSS_NO_REF;
5564 /* allocation can't be done safely during early init */
5567 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5568 BUG_ON(css->id < 0);
5571 /* Update the init_css_set to contain a subsys
5572 * pointer to this state - since the subsystem is
5573 * newly registered, all tasks and hence the
5574 * init_css_set is in the subsystem's root cgroup. */
5575 init_css_set.subsys[ss->id] = css;
5577 have_fork_callback |= (bool)ss->fork << ss->id;
5578 have_exit_callback |= (bool)ss->exit << ss->id;
5579 have_release_callback |= (bool)ss->release << ss->id;
5580 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5582 /* At system boot, before all subsystems have been
5583 * registered, no tasks have been forked, so we don't
5584 * need to invoke fork callbacks here. */
5585 BUG_ON(!list_empty(&init_task.tasks));
5587 BUG_ON(online_css(css));
5589 mutex_unlock(&cgroup_mutex);
5593 * cgroup_init_early - cgroup initialization at system boot
5595 * Initialize cgroups at system boot, and initialize any
5596 * subsystems that request early init.
5598 int __init cgroup_init_early(void)
5600 static struct cgroup_fs_context __initdata ctx;
5601 struct cgroup_subsys *ss;
5604 ctx.root = &cgrp_dfl_root;
5605 init_cgroup_root(&ctx);
5606 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5608 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5610 for_each_subsys(ss, i) {
5611 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5612 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5613 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5615 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5616 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5619 ss->name = cgroup_subsys_name[i];
5620 if (!ss->legacy_name)
5621 ss->legacy_name = cgroup_subsys_name[i];
5624 cgroup_init_subsys(ss, true);
5629 static u16 cgroup_disable_mask __initdata;
5632 * cgroup_init - cgroup initialization
5634 * Register cgroup filesystem and /proc file, and initialize
5635 * any subsystems that didn't request early init.
5637 int __init cgroup_init(void)
5639 struct cgroup_subsys *ss;
5642 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5643 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5644 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5646 cgroup_rstat_boot();
5649 * The latency of the synchronize_rcu() is too high for cgroups,
5650 * avoid it at the cost of forcing all readers into the slow path.
5652 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5654 get_user_ns(init_cgroup_ns.user_ns);
5656 mutex_lock(&cgroup_mutex);
5659 * Add init_css_set to the hash table so that dfl_root can link to
5662 hash_add(css_set_table, &init_css_set.hlist,
5663 css_set_hash(init_css_set.subsys));
5665 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5667 mutex_unlock(&cgroup_mutex);
5669 for_each_subsys(ss, ssid) {
5670 if (ss->early_init) {
5671 struct cgroup_subsys_state *css =
5672 init_css_set.subsys[ss->id];
5674 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5676 BUG_ON(css->id < 0);
5678 cgroup_init_subsys(ss, false);
5681 list_add_tail(&init_css_set.e_cset_node[ssid],
5682 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5685 * Setting dfl_root subsys_mask needs to consider the
5686 * disabled flag and cftype registration needs kmalloc,
5687 * both of which aren't available during early_init.
5689 if (cgroup_disable_mask & (1 << ssid)) {
5690 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5691 printk(KERN_INFO "Disabling %s control group subsystem\n",
5696 if (cgroup1_ssid_disabled(ssid))
5697 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5700 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5702 /* implicit controllers must be threaded too */
5703 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5705 if (ss->implicit_on_dfl)
5706 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5707 else if (!ss->dfl_cftypes)
5708 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5711 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5713 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5714 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5716 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5717 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5721 ss->bind(init_css_set.subsys[ssid]);
5723 mutex_lock(&cgroup_mutex);
5724 css_populate_dir(init_css_set.subsys[ssid]);
5725 mutex_unlock(&cgroup_mutex);
5728 /* init_css_set.subsys[] has been updated, re-hash */
5729 hash_del(&init_css_set.hlist);
5730 hash_add(css_set_table, &init_css_set.hlist,
5731 css_set_hash(init_css_set.subsys));
5733 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5734 WARN_ON(register_filesystem(&cgroup_fs_type));
5735 WARN_ON(register_filesystem(&cgroup2_fs_type));
5736 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
5737 #ifdef CONFIG_CPUSETS
5738 WARN_ON(register_filesystem(&cpuset_fs_type));
5744 static int __init cgroup_wq_init(void)
5747 * There isn't much point in executing destruction path in
5748 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5749 * Use 1 for @max_active.
5751 * We would prefer to do this in cgroup_init() above, but that
5752 * is called before init_workqueues(): so leave this until after.
5754 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5755 BUG_ON(!cgroup_destroy_wq);
5758 core_initcall(cgroup_wq_init);
5760 void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
5762 struct kernfs_node *kn;
5764 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
5767 kernfs_path(kn, buf, buflen);
5772 * proc_cgroup_show()
5773 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5774 * - Used for /proc/<pid>/cgroup.
5776 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5777 struct pid *pid, struct task_struct *tsk)
5781 struct cgroup_root *root;
5784 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5788 mutex_lock(&cgroup_mutex);
5789 spin_lock_irq(&css_set_lock);
5791 for_each_root(root) {
5792 struct cgroup_subsys *ss;
5793 struct cgroup *cgrp;
5794 int ssid, count = 0;
5796 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5799 seq_printf(m, "%d:", root->hierarchy_id);
5800 if (root != &cgrp_dfl_root)
5801 for_each_subsys(ss, ssid)
5802 if (root->subsys_mask & (1 << ssid))
5803 seq_printf(m, "%s%s", count++ ? "," : "",
5805 if (strlen(root->name))
5806 seq_printf(m, "%sname=%s", count ? "," : "",
5810 cgrp = task_cgroup_from_root(tsk, root);
5813 * On traditional hierarchies, all zombie tasks show up as
5814 * belonging to the root cgroup. On the default hierarchy,
5815 * while a zombie doesn't show up in "cgroup.procs" and
5816 * thus can't be migrated, its /proc/PID/cgroup keeps
5817 * reporting the cgroup it belonged to before exiting. If
5818 * the cgroup is removed before the zombie is reaped,
5819 * " (deleted)" is appended to the cgroup path.
5821 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5822 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5823 current->nsproxy->cgroup_ns);
5824 if (retval >= PATH_MAX)
5825 retval = -ENAMETOOLONG;
5834 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5835 seq_puts(m, " (deleted)\n");
5842 spin_unlock_irq(&css_set_lock);
5843 mutex_unlock(&cgroup_mutex);
5850 * cgroup_fork - initialize cgroup related fields during copy_process()
5851 * @child: pointer to task_struct of forking parent process.
5853 * A task is associated with the init_css_set until cgroup_post_fork()
5854 * attaches it to the parent's css_set. Empty cg_list indicates that
5855 * @child isn't holding reference to its css_set.
5857 void cgroup_fork(struct task_struct *child)
5859 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5860 INIT_LIST_HEAD(&child->cg_list);
5864 * cgroup_can_fork - called on a new task before the process is exposed
5865 * @child: the task in question.
5867 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5868 * returns an error, the fork aborts with that error code. This allows for
5869 * a cgroup subsystem to conditionally allow or deny new forks.
5871 int cgroup_can_fork(struct task_struct *child)
5873 struct cgroup_subsys *ss;
5876 do_each_subsys_mask(ss, i, have_canfork_callback) {
5877 ret = ss->can_fork(child);
5880 } while_each_subsys_mask();
5885 for_each_subsys(ss, j) {
5888 if (ss->cancel_fork)
5889 ss->cancel_fork(child);
5896 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5897 * @child: the task in question
5899 * This calls the cancel_fork() callbacks if a fork failed *after*
5900 * cgroup_can_fork() succeded.
5902 void cgroup_cancel_fork(struct task_struct *child)
5904 struct cgroup_subsys *ss;
5907 for_each_subsys(ss, i)
5908 if (ss->cancel_fork)
5909 ss->cancel_fork(child);
5913 * cgroup_post_fork - called on a new task after adding it to the task list
5914 * @child: the task in question
5916 * Adds the task to the list running through its css_set if necessary and
5917 * call the subsystem fork() callbacks. Has to be after the task is
5918 * visible on the task list in case we race with the first call to
5919 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5922 void cgroup_post_fork(struct task_struct *child)
5924 struct cgroup_subsys *ss;
5925 struct css_set *cset;
5928 spin_lock_irq(&css_set_lock);
5930 /* init tasks are special, only link regular threads */
5931 if (likely(child->pid)) {
5932 WARN_ON_ONCE(!list_empty(&child->cg_list));
5933 cset = task_css_set(current); /* current is @child's parent */
5936 css_set_move_task(child, NULL, cset, false);
5940 * If the cgroup has to be frozen, the new task has too. Let's set
5941 * the JOBCTL_TRAP_FREEZE jobctl bit to get the task into the
5944 if (unlikely(cgroup_task_freeze(child))) {
5945 spin_lock(&child->sighand->siglock);
5946 WARN_ON_ONCE(child->frozen);
5947 child->jobctl |= JOBCTL_TRAP_FREEZE;
5948 spin_unlock(&child->sighand->siglock);
5951 * Calling cgroup_update_frozen() isn't required here,
5952 * because it will be called anyway a bit later from
5953 * do_freezer_trap(). So we avoid cgroup's transient switch
5954 * from the frozen state and back.
5958 spin_unlock_irq(&css_set_lock);
5961 * Call ss->fork(). This must happen after @child is linked on
5962 * css_set; otherwise, @child might change state between ->fork()
5963 * and addition to css_set.
5965 do_each_subsys_mask(ss, i, have_fork_callback) {
5967 } while_each_subsys_mask();
5971 * cgroup_exit - detach cgroup from exiting task
5972 * @tsk: pointer to task_struct of exiting process
5974 * Description: Detach cgroup from @tsk.
5977 void cgroup_exit(struct task_struct *tsk)
5979 struct cgroup_subsys *ss;
5980 struct css_set *cset;
5983 spin_lock_irq(&css_set_lock);
5985 WARN_ON_ONCE(list_empty(&tsk->cg_list));
5986 cset = task_css_set(tsk);
5987 css_set_move_task(tsk, cset, NULL, false);
5988 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
5991 WARN_ON_ONCE(cgroup_task_frozen(tsk));
5992 if (unlikely(cgroup_task_freeze(tsk)))
5993 cgroup_update_frozen(task_dfl_cgroup(tsk));
5995 spin_unlock_irq(&css_set_lock);
5997 /* see cgroup_post_fork() for details */
5998 do_each_subsys_mask(ss, i, have_exit_callback) {
6000 } while_each_subsys_mask();
6003 void cgroup_release(struct task_struct *task)
6005 struct cgroup_subsys *ss;
6008 do_each_subsys_mask(ss, ssid, have_release_callback) {
6010 } while_each_subsys_mask();
6012 spin_lock_irq(&css_set_lock);
6013 css_set_skip_task_iters(task_css_set(task), task);
6014 list_del_init(&task->cg_list);
6015 spin_unlock_irq(&css_set_lock);
6018 void cgroup_free(struct task_struct *task)
6020 struct css_set *cset = task_css_set(task);
6024 static int __init cgroup_disable(char *str)
6026 struct cgroup_subsys *ss;
6030 while ((token = strsep(&str, ",")) != NULL) {
6034 for_each_subsys(ss, i) {
6035 if (strcmp(token, ss->name) &&
6036 strcmp(token, ss->legacy_name))
6038 cgroup_disable_mask |= 1 << i;
6043 __setup("cgroup_disable=", cgroup_disable);
6045 void __init __weak enable_debug_cgroup(void) { }
6047 static int __init enable_cgroup_debug(char *str)
6049 cgroup_debug = true;
6050 enable_debug_cgroup();
6053 __setup("cgroup_debug", enable_cgroup_debug);
6056 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6057 * @dentry: directory dentry of interest
6058 * @ss: subsystem of interest
6060 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6061 * to get the corresponding css and return it. If such css doesn't exist
6062 * or can't be pinned, an ERR_PTR value is returned.
6064 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6065 struct cgroup_subsys *ss)
6067 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6068 struct file_system_type *s_type = dentry->d_sb->s_type;
6069 struct cgroup_subsys_state *css = NULL;
6070 struct cgroup *cgrp;
6072 /* is @dentry a cgroup dir? */
6073 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6074 !kn || kernfs_type(kn) != KERNFS_DIR)
6075 return ERR_PTR(-EBADF);
6080 * This path doesn't originate from kernfs and @kn could already
6081 * have been or be removed at any point. @kn->priv is RCU
6082 * protected for this access. See css_release_work_fn() for details.
6084 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6086 css = cgroup_css(cgrp, ss);
6088 if (!css || !css_tryget_online(css))
6089 css = ERR_PTR(-ENOENT);
6096 * css_from_id - lookup css by id
6097 * @id: the cgroup id
6098 * @ss: cgroup subsys to be looked into
6100 * Returns the css if there's valid one with @id, otherwise returns NULL.
6101 * Should be called under rcu_read_lock().
6103 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6105 WARN_ON_ONCE(!rcu_read_lock_held());
6106 return idr_find(&ss->css_idr, id);
6110 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6111 * @path: path on the default hierarchy
6113 * Find the cgroup at @path on the default hierarchy, increment its
6114 * reference count and return it. Returns pointer to the found cgroup on
6115 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6116 * if @path points to a non-directory.
6118 struct cgroup *cgroup_get_from_path(const char *path)
6120 struct kernfs_node *kn;
6121 struct cgroup *cgrp;
6123 mutex_lock(&cgroup_mutex);
6125 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6127 if (kernfs_type(kn) == KERNFS_DIR) {
6129 cgroup_get_live(cgrp);
6131 cgrp = ERR_PTR(-ENOTDIR);
6135 cgrp = ERR_PTR(-ENOENT);
6138 mutex_unlock(&cgroup_mutex);
6141 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6144 * cgroup_get_from_fd - get a cgroup pointer from a fd
6145 * @fd: fd obtained by open(cgroup2_dir)
6147 * Find the cgroup from a fd which should be obtained
6148 * by opening a cgroup directory. Returns a pointer to the
6149 * cgroup on success. ERR_PTR is returned if the cgroup
6152 struct cgroup *cgroup_get_from_fd(int fd)
6154 struct cgroup_subsys_state *css;
6155 struct cgroup *cgrp;
6160 return ERR_PTR(-EBADF);
6162 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6165 return ERR_CAST(css);
6168 if (!cgroup_on_dfl(cgrp)) {
6170 return ERR_PTR(-EBADF);
6175 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6177 static u64 power_of_ten(int power)
6186 * cgroup_parse_float - parse a floating number
6187 * @input: input string
6188 * @dec_shift: number of decimal digits to shift
6191 * Parse a decimal floating point number in @input and store the result in
6192 * @v with decimal point right shifted @dec_shift times. For example, if
6193 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6194 * Returns 0 on success, -errno otherwise.
6196 * There's nothing cgroup specific about this function except that it's
6197 * currently the only user.
6199 int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6201 s64 whole, frac = 0;
6202 int fstart = 0, fend = 0, flen;
6204 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6209 flen = fend > fstart ? fend - fstart : 0;
6210 if (flen < dec_shift)
6211 frac *= power_of_ten(dec_shift - flen);
6213 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6215 *v = whole * power_of_ten(dec_shift) + frac;
6220 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6221 * definition in cgroup-defs.h.
6223 #ifdef CONFIG_SOCK_CGROUP_DATA
6225 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6227 DEFINE_SPINLOCK(cgroup_sk_update_lock);
6228 static bool cgroup_sk_alloc_disabled __read_mostly;
6230 void cgroup_sk_alloc_disable(void)
6232 if (cgroup_sk_alloc_disabled)
6234 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6235 cgroup_sk_alloc_disabled = true;
6240 #define cgroup_sk_alloc_disabled false
6244 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6246 if (cgroup_sk_alloc_disabled)
6249 /* Socket clone path */
6252 * We might be cloning a socket which is left in an empty
6253 * cgroup and the cgroup might have already been rmdir'd.
6254 * Don't use cgroup_get_live().
6256 cgroup_get(sock_cgroup_ptr(skcd));
6257 cgroup_bpf_get(sock_cgroup_ptr(skcd));
6264 struct css_set *cset;
6266 cset = task_css_set(current);
6267 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6268 skcd->val = (unsigned long)cset->dfl_cgrp;
6269 cgroup_bpf_get(cset->dfl_cgrp);
6278 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6280 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6282 cgroup_bpf_put(cgrp);
6286 #endif /* CONFIG_SOCK_CGROUP_DATA */
6288 #ifdef CONFIG_CGROUP_BPF
6289 int cgroup_bpf_attach(struct cgroup *cgrp, struct bpf_prog *prog,
6290 struct bpf_prog *replace_prog, enum bpf_attach_type type,
6295 mutex_lock(&cgroup_mutex);
6296 ret = __cgroup_bpf_attach(cgrp, prog, replace_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 */