2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/percpu-rwsem.h>
49 #include <linux/string.h>
50 #include <linux/sort.h>
51 #include <linux/kmod.h>
52 #include <linux/delayacct.h>
53 #include <linux/cgroupstats.h>
54 #include <linux/hashtable.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/kthread.h>
59 #include <linux/delay.h>
60 #include <linux/atomic.h>
61 #include <linux/cpuset.h>
62 #include <linux/proc_ns.h>
63 #include <linux/nsproxy.h>
64 #include <linux/proc_ns.h>
68 * pidlists linger the following amount before being destroyed. The goal
69 * is avoiding frequent destruction in the middle of consecutive read calls
70 * Expiring in the middle is a performance problem not a correctness one.
71 * 1 sec should be enough.
73 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
75 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
79 * cgroup_mutex is the master lock. Any modification to cgroup or its
80 * hierarchy must be performed while holding it.
82 * css_set_lock protects task->cgroups pointer, the list of css_set
83 * objects, and the chain of tasks off each css_set.
85 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
86 * cgroup.h can use them for lockdep annotations.
88 #ifdef CONFIG_PROVE_RCU
89 DEFINE_MUTEX(cgroup_mutex);
90 DEFINE_SPINLOCK(css_set_lock);
91 EXPORT_SYMBOL_GPL(cgroup_mutex);
92 EXPORT_SYMBOL_GPL(css_set_lock);
94 static DEFINE_MUTEX(cgroup_mutex);
95 static DEFINE_SPINLOCK(css_set_lock);
99 * Protects cgroup_idr and css_idr so that IDs can be released without
100 * grabbing cgroup_mutex.
102 static DEFINE_SPINLOCK(cgroup_idr_lock);
105 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
106 * against file removal/re-creation across css hiding.
108 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
111 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
112 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
114 static DEFINE_SPINLOCK(release_agent_path_lock);
116 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
118 #define cgroup_assert_mutex_or_rcu_locked() \
119 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
120 !lockdep_is_held(&cgroup_mutex), \
121 "cgroup_mutex or RCU read lock required");
124 * cgroup destruction makes heavy use of work items and there can be a lot
125 * of concurrent destructions. Use a separate workqueue so that cgroup
126 * destruction work items don't end up filling up max_active of system_wq
127 * which may lead to deadlock.
129 static struct workqueue_struct *cgroup_destroy_wq;
132 * pidlist destructions need to be flushed on cgroup destruction. Use a
133 * separate workqueue as flush domain.
135 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
137 /* generate an array of cgroup subsystem pointers */
138 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
139 static struct cgroup_subsys *cgroup_subsys[] = {
140 #include <linux/cgroup_subsys.h>
144 /* array of cgroup subsystem names */
145 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
146 static const char *cgroup_subsys_name[] = {
147 #include <linux/cgroup_subsys.h>
151 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
153 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
154 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
155 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
156 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
157 #include <linux/cgroup_subsys.h>
160 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
161 static struct static_key_true *cgroup_subsys_enabled_key[] = {
162 #include <linux/cgroup_subsys.h>
166 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
167 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
168 #include <linux/cgroup_subsys.h>
173 * The default hierarchy, reserved for the subsystems that are otherwise
174 * unattached - it never has more than a single cgroup, and all tasks are
175 * part of that cgroup.
177 struct cgroup_root cgrp_dfl_root;
178 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
181 * The default hierarchy always exists but is hidden until mounted for the
182 * first time. This is for backward compatibility.
184 static bool cgrp_dfl_root_visible;
186 /* Controllers blocked by the commandline in v1 */
187 static unsigned long cgroup_no_v1_mask;
189 /* some controllers are not supported in the default hierarchy */
190 static unsigned long cgrp_dfl_root_inhibit_ss_mask;
192 /* The list of hierarchy roots */
194 static LIST_HEAD(cgroup_roots);
195 static int cgroup_root_count;
197 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
198 static DEFINE_IDR(cgroup_hierarchy_idr);
201 * Assign a monotonically increasing serial number to csses. It guarantees
202 * cgroups with bigger numbers are newer than those with smaller numbers.
203 * Also, as csses are always appended to the parent's ->children list, it
204 * guarantees that sibling csses are always sorted in the ascending serial
205 * number order on the list. Protected by cgroup_mutex.
207 static u64 css_serial_nr_next = 1;
210 * These bitmask flags indicate whether tasks in the fork and exit paths have
211 * fork/exit handlers to call. This avoids us having to do extra work in the
212 * fork/exit path to check which subsystems have fork/exit callbacks.
214 static unsigned long have_fork_callback __read_mostly;
215 static unsigned long have_exit_callback __read_mostly;
216 static unsigned long have_free_callback __read_mostly;
218 /* cgroup namespace for init task */
219 struct cgroup_namespace init_cgroup_ns = {
220 .count = { .counter = 2, },
221 .user_ns = &init_user_ns,
222 .ns.ops = &cgroupns_operations,
223 .ns.inum = PROC_CGROUP_INIT_INO,
224 .root_cset = &init_css_set,
227 /* Ditto for the can_fork callback. */
228 static unsigned long have_canfork_callback __read_mostly;
230 static struct file_system_type cgroup2_fs_type;
231 static struct cftype cgroup_dfl_base_files[];
232 static struct cftype cgroup_legacy_base_files[];
234 static int rebind_subsystems(struct cgroup_root *dst_root,
235 unsigned long ss_mask);
236 static void css_task_iter_advance(struct css_task_iter *it);
237 static int cgroup_destroy_locked(struct cgroup *cgrp);
238 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
240 static void css_release(struct percpu_ref *ref);
241 static void kill_css(struct cgroup_subsys_state *css);
242 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
243 struct cgroup *cgrp, struct cftype cfts[],
247 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
248 * @ssid: subsys ID of interest
250 * cgroup_subsys_enabled() can only be used with literal subsys names which
251 * is fine for individual subsystems but unsuitable for cgroup core. This
252 * is slower static_key_enabled() based test indexed by @ssid.
254 static bool cgroup_ssid_enabled(int ssid)
256 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
259 static bool cgroup_ssid_no_v1(int ssid)
261 return cgroup_no_v1_mask & (1 << ssid);
265 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
266 * @cgrp: the cgroup of interest
268 * The default hierarchy is the v2 interface of cgroup and this function
269 * can be used to test whether a cgroup is on the default hierarchy for
270 * cases where a subsystem should behave differnetly depending on the
273 * The set of behaviors which change on the default hierarchy are still
274 * being determined and the mount option is prefixed with __DEVEL__.
276 * List of changed behaviors:
278 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
279 * and "name" are disallowed.
281 * - When mounting an existing superblock, mount options should match.
283 * - Remount is disallowed.
285 * - rename(2) is disallowed.
287 * - "tasks" is removed. Everything should be at process granularity. Use
288 * "cgroup.procs" instead.
290 * - "cgroup.procs" is not sorted. pids will be unique unless they got
291 * recycled inbetween reads.
293 * - "release_agent" and "notify_on_release" are removed. Replacement
294 * notification mechanism will be implemented.
296 * - "cgroup.clone_children" is removed.
298 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
299 * and its descendants contain no task; otherwise, 1. The file also
300 * generates kernfs notification which can be monitored through poll and
301 * [di]notify when the value of the file changes.
303 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
304 * take masks of ancestors with non-empty cpus/mems, instead of being
305 * moved to an ancestor.
307 * - cpuset: a task can be moved into an empty cpuset, and again it takes
308 * masks of ancestors.
310 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
313 * - blkcg: blk-throttle becomes properly hierarchical.
315 * - debug: disallowed on the default hierarchy.
317 static bool cgroup_on_dfl(const struct cgroup *cgrp)
319 return cgrp->root == &cgrp_dfl_root;
322 /* IDR wrappers which synchronize using cgroup_idr_lock */
323 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
328 idr_preload(gfp_mask);
329 spin_lock_bh(&cgroup_idr_lock);
330 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
331 spin_unlock_bh(&cgroup_idr_lock);
336 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
340 spin_lock_bh(&cgroup_idr_lock);
341 ret = idr_replace(idr, ptr, id);
342 spin_unlock_bh(&cgroup_idr_lock);
346 static void cgroup_idr_remove(struct idr *idr, int id)
348 spin_lock_bh(&cgroup_idr_lock);
350 spin_unlock_bh(&cgroup_idr_lock);
353 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
355 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
358 return container_of(parent_css, struct cgroup, self);
363 * cgroup_css - obtain a cgroup's css for the specified subsystem
364 * @cgrp: the cgroup of interest
365 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
367 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
368 * function must be called either under cgroup_mutex or rcu_read_lock() and
369 * the caller is responsible for pinning the returned css if it wants to
370 * keep accessing it outside the said locks. This function may return
371 * %NULL if @cgrp doesn't have @subsys_id enabled.
373 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
374 struct cgroup_subsys *ss)
377 return rcu_dereference_check(cgrp->subsys[ss->id],
378 lockdep_is_held(&cgroup_mutex));
384 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
385 * @cgrp: the cgroup of interest
386 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
388 * Similar to cgroup_css() but returns the effective css, which is defined
389 * as the matching css of the nearest ancestor including self which has @ss
390 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
391 * function is guaranteed to return non-NULL css.
393 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
394 struct cgroup_subsys *ss)
396 lockdep_assert_held(&cgroup_mutex);
401 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
405 * This function is used while updating css associations and thus
406 * can't test the csses directly. Use ->child_subsys_mask.
408 while (cgroup_parent(cgrp) &&
409 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
410 cgrp = cgroup_parent(cgrp);
412 return cgroup_css(cgrp, ss);
416 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
417 * @cgrp: the cgroup of interest
418 * @ss: the subsystem of interest
420 * Find and get the effective css of @cgrp for @ss. The effective css is
421 * defined as the matching css of the nearest ancestor including self which
422 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
423 * the root css is returned, so this function always returns a valid css.
424 * The returned css must be put using css_put().
426 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
427 struct cgroup_subsys *ss)
429 struct cgroup_subsys_state *css;
434 css = cgroup_css(cgrp, ss);
436 if (css && css_tryget_online(css))
438 cgrp = cgroup_parent(cgrp);
441 css = init_css_set.subsys[ss->id];
448 /* convenient tests for these bits */
449 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
451 return !(cgrp->self.flags & CSS_ONLINE);
454 static void cgroup_get(struct cgroup *cgrp)
456 WARN_ON_ONCE(cgroup_is_dead(cgrp));
457 css_get(&cgrp->self);
460 static bool cgroup_tryget(struct cgroup *cgrp)
462 return css_tryget(&cgrp->self);
465 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
467 struct cgroup *cgrp = of->kn->parent->priv;
468 struct cftype *cft = of_cft(of);
471 * This is open and unprotected implementation of cgroup_css().
472 * seq_css() is only called from a kernfs file operation which has
473 * an active reference on the file. Because all the subsystem
474 * files are drained before a css is disassociated with a cgroup,
475 * the matching css from the cgroup's subsys table is guaranteed to
476 * be and stay valid until the enclosing operation is complete.
479 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
483 EXPORT_SYMBOL_GPL(of_css);
485 static int notify_on_release(const struct cgroup *cgrp)
487 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
491 * for_each_css - iterate all css's of a cgroup
492 * @css: the iteration cursor
493 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
494 * @cgrp: the target cgroup to iterate css's of
496 * Should be called under cgroup_[tree_]mutex.
498 #define for_each_css(css, ssid, cgrp) \
499 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
500 if (!((css) = rcu_dereference_check( \
501 (cgrp)->subsys[(ssid)], \
502 lockdep_is_held(&cgroup_mutex)))) { } \
506 * for_each_e_css - iterate all effective css's of a cgroup
507 * @css: the iteration cursor
508 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
509 * @cgrp: the target cgroup to iterate css's of
511 * Should be called under cgroup_[tree_]mutex.
513 #define for_each_e_css(css, ssid, cgrp) \
514 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
515 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
520 * for_each_subsys - iterate all enabled cgroup subsystems
521 * @ss: the iteration cursor
522 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
524 #define for_each_subsys(ss, ssid) \
525 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
526 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
529 * for_each_subsys_which - filter for_each_subsys with a bitmask
530 * @ss: the iteration cursor
531 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
532 * @ss_maskp: a pointer to the bitmask
534 * The block will only run for cases where the ssid-th bit (1 << ssid) of
537 #define for_each_subsys_which(ss, ssid, ss_maskp) \
538 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
541 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
542 if (((ss) = cgroup_subsys[ssid]) && false) \
546 /* iterate across the hierarchies */
547 #define for_each_root(root) \
548 list_for_each_entry((root), &cgroup_roots, root_list)
550 /* iterate over child cgrps, lock should be held throughout iteration */
551 #define cgroup_for_each_live_child(child, cgrp) \
552 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
553 if (({ lockdep_assert_held(&cgroup_mutex); \
554 cgroup_is_dead(child); })) \
558 static void cgroup_release_agent(struct work_struct *work);
559 static void check_for_release(struct cgroup *cgrp);
562 * A cgroup can be associated with multiple css_sets as different tasks may
563 * belong to different cgroups on different hierarchies. In the other
564 * direction, a css_set is naturally associated with multiple cgroups.
565 * This M:N relationship is represented by the following link structure
566 * which exists for each association and allows traversing the associations
569 struct cgrp_cset_link {
570 /* the cgroup and css_set this link associates */
572 struct css_set *cset;
574 /* list of cgrp_cset_links anchored at cgrp->cset_links */
575 struct list_head cset_link;
577 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
578 struct list_head cgrp_link;
582 * The default css_set - used by init and its children prior to any
583 * hierarchies being mounted. It contains a pointer to the root state
584 * for each subsystem. Also used to anchor the list of css_sets. Not
585 * reference-counted, to improve performance when child cgroups
586 * haven't been created.
588 struct css_set init_css_set = {
589 .refcount = ATOMIC_INIT(1),
590 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
591 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
592 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
593 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
594 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
595 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
598 static int css_set_count = 1; /* 1 for init_css_set */
601 * css_set_populated - does a css_set contain any tasks?
602 * @cset: target css_set
604 static bool css_set_populated(struct css_set *cset)
606 lockdep_assert_held(&css_set_lock);
608 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
612 * cgroup_update_populated - updated populated count of a cgroup
613 * @cgrp: the target cgroup
614 * @populated: inc or dec populated count
616 * One of the css_sets associated with @cgrp is either getting its first
617 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
618 * count is propagated towards root so that a given cgroup's populated_cnt
619 * is zero iff the cgroup and all its descendants don't contain any tasks.
621 * @cgrp's interface file "cgroup.populated" is zero if
622 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
623 * changes from or to zero, userland is notified that the content of the
624 * interface file has changed. This can be used to detect when @cgrp and
625 * its descendants become populated or empty.
627 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
629 lockdep_assert_held(&css_set_lock);
635 trigger = !cgrp->populated_cnt++;
637 trigger = !--cgrp->populated_cnt;
642 check_for_release(cgrp);
643 cgroup_file_notify(&cgrp->events_file);
645 cgrp = cgroup_parent(cgrp);
650 * css_set_update_populated - update populated state of a css_set
651 * @cset: target css_set
652 * @populated: whether @cset is populated or depopulated
654 * @cset is either getting the first task or losing the last. Update the
655 * ->populated_cnt of all associated cgroups accordingly.
657 static void css_set_update_populated(struct css_set *cset, bool populated)
659 struct cgrp_cset_link *link;
661 lockdep_assert_held(&css_set_lock);
663 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
664 cgroup_update_populated(link->cgrp, populated);
668 * css_set_move_task - move a task from one css_set to another
669 * @task: task being moved
670 * @from_cset: css_set @task currently belongs to (may be NULL)
671 * @to_cset: new css_set @task is being moved to (may be NULL)
672 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
674 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
675 * css_set, @from_cset can be NULL. If @task is being disassociated
676 * instead of moved, @to_cset can be NULL.
678 * This function automatically handles populated_cnt updates and
679 * css_task_iter adjustments but the caller is responsible for managing
680 * @from_cset and @to_cset's reference counts.
682 static void css_set_move_task(struct task_struct *task,
683 struct css_set *from_cset, struct css_set *to_cset,
686 lockdep_assert_held(&css_set_lock);
689 struct css_task_iter *it, *pos;
691 WARN_ON_ONCE(list_empty(&task->cg_list));
694 * @task is leaving, advance task iterators which are
695 * pointing to it so that they can resume at the next
696 * position. Advancing an iterator might remove it from
697 * the list, use safe walk. See css_task_iter_advance*()
700 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
702 if (it->task_pos == &task->cg_list)
703 css_task_iter_advance(it);
705 list_del_init(&task->cg_list);
706 if (!css_set_populated(from_cset))
707 css_set_update_populated(from_cset, false);
709 WARN_ON_ONCE(!list_empty(&task->cg_list));
714 * We are synchronized through cgroup_threadgroup_rwsem
715 * against PF_EXITING setting such that we can't race
716 * against cgroup_exit() changing the css_set to
717 * init_css_set and dropping the old one.
719 WARN_ON_ONCE(task->flags & PF_EXITING);
721 if (!css_set_populated(to_cset))
722 css_set_update_populated(to_cset, true);
723 rcu_assign_pointer(task->cgroups, to_cset);
724 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
730 * hash table for cgroup groups. This improves the performance to find
731 * an existing css_set. This hash doesn't (currently) take into
732 * account cgroups in empty hierarchies.
734 #define CSS_SET_HASH_BITS 7
735 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
737 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
739 unsigned long key = 0UL;
740 struct cgroup_subsys *ss;
743 for_each_subsys(ss, i)
744 key += (unsigned long)css[i];
745 key = (key >> 16) ^ key;
750 static void put_css_set_locked(struct css_set *cset)
752 struct cgrp_cset_link *link, *tmp_link;
753 struct cgroup_subsys *ss;
756 lockdep_assert_held(&css_set_lock);
758 if (!atomic_dec_and_test(&cset->refcount))
761 /* This css_set is dead. unlink it and release cgroup and css refs */
762 for_each_subsys(ss, ssid) {
763 list_del(&cset->e_cset_node[ssid]);
764 css_put(cset->subsys[ssid]);
766 hash_del(&cset->hlist);
769 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
770 list_del(&link->cset_link);
771 list_del(&link->cgrp_link);
772 if (cgroup_parent(link->cgrp))
773 cgroup_put(link->cgrp);
777 kfree_rcu(cset, rcu_head);
780 static void put_css_set(struct css_set *cset)
783 * Ensure that the refcount doesn't hit zero while any readers
784 * can see it. Similar to atomic_dec_and_lock(), but for an
787 if (atomic_add_unless(&cset->refcount, -1, 1))
790 spin_lock_bh(&css_set_lock);
791 put_css_set_locked(cset);
792 spin_unlock_bh(&css_set_lock);
796 * refcounted get/put for css_set objects
798 static inline void get_css_set(struct css_set *cset)
800 atomic_inc(&cset->refcount);
804 * compare_css_sets - helper function for find_existing_css_set().
805 * @cset: candidate css_set being tested
806 * @old_cset: existing css_set for a task
807 * @new_cgrp: cgroup that's being entered by the task
808 * @template: desired set of css pointers in css_set (pre-calculated)
810 * Returns true if "cset" matches "old_cset" except for the hierarchy
811 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
813 static bool compare_css_sets(struct css_set *cset,
814 struct css_set *old_cset,
815 struct cgroup *new_cgrp,
816 struct cgroup_subsys_state *template[])
818 struct list_head *l1, *l2;
821 * On the default hierarchy, there can be csets which are
822 * associated with the same set of cgroups but different csses.
823 * Let's first ensure that csses match.
825 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
829 * Compare cgroup pointers in order to distinguish between
830 * different cgroups in hierarchies. As different cgroups may
831 * share the same effective css, this comparison is always
834 l1 = &cset->cgrp_links;
835 l2 = &old_cset->cgrp_links;
837 struct cgrp_cset_link *link1, *link2;
838 struct cgroup *cgrp1, *cgrp2;
842 /* See if we reached the end - both lists are equal length. */
843 if (l1 == &cset->cgrp_links) {
844 BUG_ON(l2 != &old_cset->cgrp_links);
847 BUG_ON(l2 == &old_cset->cgrp_links);
849 /* Locate the cgroups associated with these links. */
850 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
851 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
854 /* Hierarchies should be linked in the same order. */
855 BUG_ON(cgrp1->root != cgrp2->root);
858 * If this hierarchy is the hierarchy of the cgroup
859 * that's changing, then we need to check that this
860 * css_set points to the new cgroup; if it's any other
861 * hierarchy, then this css_set should point to the
862 * same cgroup as the old css_set.
864 if (cgrp1->root == new_cgrp->root) {
865 if (cgrp1 != new_cgrp)
876 * find_existing_css_set - init css array and find the matching css_set
877 * @old_cset: the css_set that we're using before the cgroup transition
878 * @cgrp: the cgroup that we're moving into
879 * @template: out param for the new set of csses, should be clear on entry
881 static struct css_set *find_existing_css_set(struct css_set *old_cset,
883 struct cgroup_subsys_state *template[])
885 struct cgroup_root *root = cgrp->root;
886 struct cgroup_subsys *ss;
887 struct css_set *cset;
892 * Build the set of subsystem state objects that we want to see in the
893 * new css_set. while subsystems can change globally, the entries here
894 * won't change, so no need for locking.
896 for_each_subsys(ss, i) {
897 if (root->subsys_mask & (1UL << i)) {
899 * @ss is in this hierarchy, so we want the
900 * effective css from @cgrp.
902 template[i] = cgroup_e_css(cgrp, ss);
905 * @ss is not in this hierarchy, so we don't want
908 template[i] = old_cset->subsys[i];
912 key = css_set_hash(template);
913 hash_for_each_possible(css_set_table, cset, hlist, key) {
914 if (!compare_css_sets(cset, old_cset, cgrp, template))
917 /* This css_set matches what we need */
921 /* No existing cgroup group matched */
925 static void free_cgrp_cset_links(struct list_head *links_to_free)
927 struct cgrp_cset_link *link, *tmp_link;
929 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
930 list_del(&link->cset_link);
936 * allocate_cgrp_cset_links - allocate cgrp_cset_links
937 * @count: the number of links to allocate
938 * @tmp_links: list_head the allocated links are put on
940 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
941 * through ->cset_link. Returns 0 on success or -errno.
943 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
945 struct cgrp_cset_link *link;
948 INIT_LIST_HEAD(tmp_links);
950 for (i = 0; i < count; i++) {
951 link = kzalloc(sizeof(*link), GFP_KERNEL);
953 free_cgrp_cset_links(tmp_links);
956 list_add(&link->cset_link, tmp_links);
962 * link_css_set - a helper function to link a css_set to a cgroup
963 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
964 * @cset: the css_set to be linked
965 * @cgrp: the destination cgroup
967 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
970 struct cgrp_cset_link *link;
972 BUG_ON(list_empty(tmp_links));
974 if (cgroup_on_dfl(cgrp))
975 cset->dfl_cgrp = cgrp;
977 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
982 * Always add links to the tail of the lists so that the lists are
983 * in choronological order.
985 list_move_tail(&link->cset_link, &cgrp->cset_links);
986 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
988 if (cgroup_parent(cgrp))
993 * find_css_set - return a new css_set with one cgroup updated
994 * @old_cset: the baseline css_set
995 * @cgrp: the cgroup to be updated
997 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
998 * substituted into the appropriate hierarchy.
1000 static struct css_set *find_css_set(struct css_set *old_cset,
1001 struct cgroup *cgrp)
1003 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1004 struct css_set *cset;
1005 struct list_head tmp_links;
1006 struct cgrp_cset_link *link;
1007 struct cgroup_subsys *ss;
1011 lockdep_assert_held(&cgroup_mutex);
1013 /* First see if we already have a cgroup group that matches
1014 * the desired set */
1015 spin_lock_bh(&css_set_lock);
1016 cset = find_existing_css_set(old_cset, cgrp, template);
1019 spin_unlock_bh(&css_set_lock);
1024 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1028 /* Allocate all the cgrp_cset_link objects that we'll need */
1029 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1034 atomic_set(&cset->refcount, 1);
1035 INIT_LIST_HEAD(&cset->cgrp_links);
1036 INIT_LIST_HEAD(&cset->tasks);
1037 INIT_LIST_HEAD(&cset->mg_tasks);
1038 INIT_LIST_HEAD(&cset->mg_preload_node);
1039 INIT_LIST_HEAD(&cset->mg_node);
1040 INIT_LIST_HEAD(&cset->task_iters);
1041 INIT_HLIST_NODE(&cset->hlist);
1043 /* Copy the set of subsystem state objects generated in
1044 * find_existing_css_set() */
1045 memcpy(cset->subsys, template, sizeof(cset->subsys));
1047 spin_lock_bh(&css_set_lock);
1048 /* Add reference counts and links from the new css_set. */
1049 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1050 struct cgroup *c = link->cgrp;
1052 if (c->root == cgrp->root)
1054 link_css_set(&tmp_links, cset, c);
1057 BUG_ON(!list_empty(&tmp_links));
1061 /* Add @cset to the hash table */
1062 key = css_set_hash(cset->subsys);
1063 hash_add(css_set_table, &cset->hlist, key);
1065 for_each_subsys(ss, ssid) {
1066 struct cgroup_subsys_state *css = cset->subsys[ssid];
1068 list_add_tail(&cset->e_cset_node[ssid],
1069 &css->cgroup->e_csets[ssid]);
1073 spin_unlock_bh(&css_set_lock);
1078 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1080 struct cgroup *root_cgrp = kf_root->kn->priv;
1082 return root_cgrp->root;
1085 static int cgroup_init_root_id(struct cgroup_root *root)
1089 lockdep_assert_held(&cgroup_mutex);
1091 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1095 root->hierarchy_id = id;
1099 static void cgroup_exit_root_id(struct cgroup_root *root)
1101 lockdep_assert_held(&cgroup_mutex);
1103 if (root->hierarchy_id) {
1104 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1105 root->hierarchy_id = 0;
1109 static void cgroup_free_root(struct cgroup_root *root)
1112 /* hierarchy ID should already have been released */
1113 WARN_ON_ONCE(root->hierarchy_id);
1115 idr_destroy(&root->cgroup_idr);
1120 static void cgroup_destroy_root(struct cgroup_root *root)
1122 struct cgroup *cgrp = &root->cgrp;
1123 struct cgrp_cset_link *link, *tmp_link;
1125 mutex_lock(&cgroup_mutex);
1127 BUG_ON(atomic_read(&root->nr_cgrps));
1128 BUG_ON(!list_empty(&cgrp->self.children));
1130 /* Rebind all subsystems back to the default hierarchy */
1131 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
1134 * Release all the links from cset_links to this hierarchy's
1137 spin_lock_bh(&css_set_lock);
1139 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1140 list_del(&link->cset_link);
1141 list_del(&link->cgrp_link);
1145 spin_unlock_bh(&css_set_lock);
1147 if (!list_empty(&root->root_list)) {
1148 list_del(&root->root_list);
1149 cgroup_root_count--;
1152 cgroup_exit_root_id(root);
1154 mutex_unlock(&cgroup_mutex);
1156 kernfs_destroy_root(root->kf_root);
1157 cgroup_free_root(root);
1160 /* look up cgroup associated with given css_set on the specified hierarchy */
1161 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1162 struct cgroup_root *root)
1164 struct cgroup *res = NULL;
1166 lockdep_assert_held(&cgroup_mutex);
1167 lockdep_assert_held(&css_set_lock);
1169 if (cset == &init_css_set) {
1172 struct cgrp_cset_link *link;
1174 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1175 struct cgroup *c = link->cgrp;
1177 if (c->root == root) {
1189 * Return the cgroup for "task" from the given hierarchy. Must be
1190 * called with cgroup_mutex and css_set_lock held.
1192 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1193 struct cgroup_root *root)
1196 * No need to lock the task - since we hold cgroup_mutex the
1197 * task can't change groups, so the only thing that can happen
1198 * is that it exits and its css is set back to init_css_set.
1200 return cset_cgroup_from_root(task_css_set(task), root);
1204 * A task must hold cgroup_mutex to modify cgroups.
1206 * Any task can increment and decrement the count field without lock.
1207 * So in general, code holding cgroup_mutex can't rely on the count
1208 * field not changing. However, if the count goes to zero, then only
1209 * cgroup_attach_task() can increment it again. Because a count of zero
1210 * means that no tasks are currently attached, therefore there is no
1211 * way a task attached to that cgroup can fork (the other way to
1212 * increment the count). So code holding cgroup_mutex can safely
1213 * assume that if the count is zero, it will stay zero. Similarly, if
1214 * a task holds cgroup_mutex on a cgroup with zero count, it
1215 * knows that the cgroup won't be removed, as cgroup_rmdir()
1218 * A cgroup can only be deleted if both its 'count' of using tasks
1219 * is zero, and its list of 'children' cgroups is empty. Since all
1220 * tasks in the system use _some_ cgroup, and since there is always at
1221 * least one task in the system (init, pid == 1), therefore, root cgroup
1222 * always has either children cgroups and/or using tasks. So we don't
1223 * need a special hack to ensure that root cgroup cannot be deleted.
1225 * P.S. One more locking exception. RCU is used to guard the
1226 * update of a tasks cgroup pointer by cgroup_attach_task()
1229 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1230 static const struct file_operations proc_cgroupstats_operations;
1232 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1235 struct cgroup_subsys *ss = cft->ss;
1237 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1238 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1239 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1240 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1243 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1248 * cgroup_file_mode - deduce file mode of a control file
1249 * @cft: the control file in question
1251 * S_IRUGO for read, S_IWUSR for write.
1253 static umode_t cgroup_file_mode(const struct cftype *cft)
1257 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1260 if (cft->write_u64 || cft->write_s64 || cft->write) {
1261 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1271 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1272 * @cgrp: the target cgroup
1273 * @subtree_control: the new subtree_control mask to consider
1275 * On the default hierarchy, a subsystem may request other subsystems to be
1276 * enabled together through its ->depends_on mask. In such cases, more
1277 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1279 * This function calculates which subsystems need to be enabled if
1280 * @subtree_control is to be applied to @cgrp. The returned mask is always
1281 * a superset of @subtree_control and follows the usual hierarchy rules.
1283 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1284 unsigned long subtree_control)
1286 struct cgroup *parent = cgroup_parent(cgrp);
1287 unsigned long cur_ss_mask = subtree_control;
1288 struct cgroup_subsys *ss;
1291 lockdep_assert_held(&cgroup_mutex);
1293 if (!cgroup_on_dfl(cgrp))
1297 unsigned long new_ss_mask = cur_ss_mask;
1299 for_each_subsys_which(ss, ssid, &cur_ss_mask)
1300 new_ss_mask |= ss->depends_on;
1303 * Mask out subsystems which aren't available. This can
1304 * happen only if some depended-upon subsystems were bound
1305 * to non-default hierarchies.
1308 new_ss_mask &= parent->child_subsys_mask;
1310 new_ss_mask &= cgrp->root->subsys_mask;
1312 if (new_ss_mask == cur_ss_mask)
1314 cur_ss_mask = new_ss_mask;
1321 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1322 * @cgrp: the target cgroup
1324 * Update @cgrp->child_subsys_mask according to the current
1325 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1327 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1329 cgrp->child_subsys_mask =
1330 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1334 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1335 * @kn: the kernfs_node being serviced
1337 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1338 * the method finishes if locking succeeded. Note that once this function
1339 * returns the cgroup returned by cgroup_kn_lock_live() may become
1340 * inaccessible any time. If the caller intends to continue to access the
1341 * cgroup, it should pin it before invoking this function.
1343 static void cgroup_kn_unlock(struct kernfs_node *kn)
1345 struct cgroup *cgrp;
1347 if (kernfs_type(kn) == KERNFS_DIR)
1350 cgrp = kn->parent->priv;
1352 mutex_unlock(&cgroup_mutex);
1354 kernfs_unbreak_active_protection(kn);
1359 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1360 * @kn: the kernfs_node being serviced
1362 * This helper is to be used by a cgroup kernfs method currently servicing
1363 * @kn. It breaks the active protection, performs cgroup locking and
1364 * verifies that the associated cgroup is alive. Returns the cgroup if
1365 * alive; otherwise, %NULL. A successful return should be undone by a
1366 * matching cgroup_kn_unlock() invocation.
1368 * Any cgroup kernfs method implementation which requires locking the
1369 * associated cgroup should use this helper. It avoids nesting cgroup
1370 * locking under kernfs active protection and allows all kernfs operations
1371 * including self-removal.
1373 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1375 struct cgroup *cgrp;
1377 if (kernfs_type(kn) == KERNFS_DIR)
1380 cgrp = kn->parent->priv;
1383 * We're gonna grab cgroup_mutex which nests outside kernfs
1384 * active_ref. cgroup liveliness check alone provides enough
1385 * protection against removal. Ensure @cgrp stays accessible and
1386 * break the active_ref protection.
1388 if (!cgroup_tryget(cgrp))
1390 kernfs_break_active_protection(kn);
1392 mutex_lock(&cgroup_mutex);
1394 if (!cgroup_is_dead(cgrp))
1397 cgroup_kn_unlock(kn);
1401 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1403 char name[CGROUP_FILE_NAME_MAX];
1405 lockdep_assert_held(&cgroup_mutex);
1407 if (cft->file_offset) {
1408 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1409 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1411 spin_lock_irq(&cgroup_file_kn_lock);
1413 spin_unlock_irq(&cgroup_file_kn_lock);
1416 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1420 * css_clear_dir - remove subsys files in a cgroup directory
1422 * @cgrp_override: specify if target cgroup is different from css->cgroup
1424 static void css_clear_dir(struct cgroup_subsys_state *css,
1425 struct cgroup *cgrp_override)
1427 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1428 struct cftype *cfts;
1430 list_for_each_entry(cfts, &css->ss->cfts, node)
1431 cgroup_addrm_files(css, cgrp, cfts, false);
1435 * css_populate_dir - create subsys files in a cgroup directory
1437 * @cgrp_overried: specify if target cgroup is different from css->cgroup
1439 * On failure, no file is added.
1441 static int css_populate_dir(struct cgroup_subsys_state *css,
1442 struct cgroup *cgrp_override)
1444 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1445 struct cftype *cfts, *failed_cfts;
1449 if (cgroup_on_dfl(cgrp))
1450 cfts = cgroup_dfl_base_files;
1452 cfts = cgroup_legacy_base_files;
1454 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1457 list_for_each_entry(cfts, &css->ss->cfts, node) {
1458 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1466 list_for_each_entry(cfts, &css->ss->cfts, node) {
1467 if (cfts == failed_cfts)
1469 cgroup_addrm_files(css, cgrp, cfts, false);
1474 static int rebind_subsystems(struct cgroup_root *dst_root,
1475 unsigned long ss_mask)
1477 struct cgroup *dcgrp = &dst_root->cgrp;
1478 struct cgroup_subsys *ss;
1479 unsigned long tmp_ss_mask;
1482 lockdep_assert_held(&cgroup_mutex);
1484 for_each_subsys_which(ss, ssid, &ss_mask) {
1485 /* if @ss has non-root csses attached to it, can't move */
1486 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1489 /* can't move between two non-dummy roots either */
1490 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1494 /* skip creating root files on dfl_root for inhibited subsystems */
1495 tmp_ss_mask = ss_mask;
1496 if (dst_root == &cgrp_dfl_root)
1497 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1499 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
1500 struct cgroup *scgrp = &ss->root->cgrp;
1503 ret = css_populate_dir(cgroup_css(scgrp, ss), dcgrp);
1508 * Rebinding back to the default root is not allowed to
1509 * fail. Using both default and non-default roots should
1510 * be rare. Moving subsystems back and forth even more so.
1511 * Just warn about it and continue.
1513 if (dst_root == &cgrp_dfl_root) {
1514 if (cgrp_dfl_root_visible) {
1515 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1517 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1522 for_each_subsys_which(ss, tssid, &tmp_ss_mask) {
1525 css_clear_dir(cgroup_css(scgrp, ss), dcgrp);
1531 * Nothing can fail from this point on. Remove files for the
1532 * removed subsystems and rebind each subsystem.
1534 for_each_subsys_which(ss, ssid, &ss_mask) {
1535 struct cgroup_root *src_root = ss->root;
1536 struct cgroup *scgrp = &src_root->cgrp;
1537 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1538 struct css_set *cset;
1540 WARN_ON(!css || cgroup_css(dcgrp, ss));
1542 css_clear_dir(css, NULL);
1544 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1545 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1546 ss->root = dst_root;
1547 css->cgroup = dcgrp;
1549 spin_lock_bh(&css_set_lock);
1550 hash_for_each(css_set_table, i, cset, hlist)
1551 list_move_tail(&cset->e_cset_node[ss->id],
1552 &dcgrp->e_csets[ss->id]);
1553 spin_unlock_bh(&css_set_lock);
1555 src_root->subsys_mask &= ~(1 << ssid);
1556 scgrp->subtree_control &= ~(1 << ssid);
1557 cgroup_refresh_child_subsys_mask(scgrp);
1559 /* default hierarchy doesn't enable controllers by default */
1560 dst_root->subsys_mask |= 1 << ssid;
1561 if (dst_root == &cgrp_dfl_root) {
1562 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1564 dcgrp->subtree_control |= 1 << ssid;
1565 cgroup_refresh_child_subsys_mask(dcgrp);
1566 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1573 kernfs_activate(dcgrp->kn);
1577 static int cgroup_show_options(struct seq_file *seq,
1578 struct kernfs_root *kf_root)
1580 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1581 struct cgroup_subsys *ss;
1584 if (root != &cgrp_dfl_root)
1585 for_each_subsys(ss, ssid)
1586 if (root->subsys_mask & (1 << ssid))
1587 seq_show_option(seq, ss->legacy_name, NULL);
1588 if (root->flags & CGRP_ROOT_NOPREFIX)
1589 seq_puts(seq, ",noprefix");
1590 if (root->flags & CGRP_ROOT_XATTR)
1591 seq_puts(seq, ",xattr");
1593 spin_lock(&release_agent_path_lock);
1594 if (strlen(root->release_agent_path))
1595 seq_show_option(seq, "release_agent",
1596 root->release_agent_path);
1597 spin_unlock(&release_agent_path_lock);
1599 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1600 seq_puts(seq, ",clone_children");
1601 if (strlen(root->name))
1602 seq_show_option(seq, "name", root->name);
1606 struct cgroup_sb_opts {
1607 unsigned long subsys_mask;
1609 char *release_agent;
1610 bool cpuset_clone_children;
1612 /* User explicitly requested empty subsystem */
1616 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1618 char *token, *o = data;
1619 bool all_ss = false, one_ss = false;
1620 unsigned long mask = -1UL;
1621 struct cgroup_subsys *ss;
1625 #ifdef CONFIG_CPUSETS
1626 mask = ~(1U << cpuset_cgrp_id);
1629 memset(opts, 0, sizeof(*opts));
1631 while ((token = strsep(&o, ",")) != NULL) {
1636 if (!strcmp(token, "none")) {
1637 /* Explicitly have no subsystems */
1641 if (!strcmp(token, "all")) {
1642 /* Mutually exclusive option 'all' + subsystem name */
1648 if (!strcmp(token, "noprefix")) {
1649 opts->flags |= CGRP_ROOT_NOPREFIX;
1652 if (!strcmp(token, "clone_children")) {
1653 opts->cpuset_clone_children = true;
1656 if (!strcmp(token, "xattr")) {
1657 opts->flags |= CGRP_ROOT_XATTR;
1660 if (!strncmp(token, "release_agent=", 14)) {
1661 /* Specifying two release agents is forbidden */
1662 if (opts->release_agent)
1664 opts->release_agent =
1665 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1666 if (!opts->release_agent)
1670 if (!strncmp(token, "name=", 5)) {
1671 const char *name = token + 5;
1672 /* Can't specify an empty name */
1675 /* Must match [\w.-]+ */
1676 for (i = 0; i < strlen(name); i++) {
1680 if ((c == '.') || (c == '-') || (c == '_'))
1684 /* Specifying two names is forbidden */
1687 opts->name = kstrndup(name,
1688 MAX_CGROUP_ROOT_NAMELEN - 1,
1696 for_each_subsys(ss, i) {
1697 if (strcmp(token, ss->legacy_name))
1699 if (!cgroup_ssid_enabled(i))
1701 if (cgroup_ssid_no_v1(i))
1704 /* Mutually exclusive option 'all' + subsystem name */
1707 opts->subsys_mask |= (1 << i);
1712 if (i == CGROUP_SUBSYS_COUNT)
1717 * If the 'all' option was specified select all the subsystems,
1718 * otherwise if 'none', 'name=' and a subsystem name options were
1719 * not specified, let's default to 'all'
1721 if (all_ss || (!one_ss && !opts->none && !opts->name))
1722 for_each_subsys(ss, i)
1723 if (cgroup_ssid_enabled(i) && !cgroup_ssid_no_v1(i))
1724 opts->subsys_mask |= (1 << i);
1727 * We either have to specify by name or by subsystems. (So all
1728 * empty hierarchies must have a name).
1730 if (!opts->subsys_mask && !opts->name)
1734 * Option noprefix was introduced just for backward compatibility
1735 * with the old cpuset, so we allow noprefix only if mounting just
1736 * the cpuset subsystem.
1738 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1741 /* Can't specify "none" and some subsystems */
1742 if (opts->subsys_mask && opts->none)
1748 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1751 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1752 struct cgroup_sb_opts opts;
1753 unsigned long added_mask, removed_mask;
1755 if (root == &cgrp_dfl_root) {
1756 pr_err("remount is not allowed\n");
1760 mutex_lock(&cgroup_mutex);
1762 /* See what subsystems are wanted */
1763 ret = parse_cgroupfs_options(data, &opts);
1767 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1768 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1769 task_tgid_nr(current), current->comm);
1771 added_mask = opts.subsys_mask & ~root->subsys_mask;
1772 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1774 /* Don't allow flags or name to change at remount */
1775 if ((opts.flags ^ root->flags) ||
1776 (opts.name && strcmp(opts.name, root->name))) {
1777 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1778 opts.flags, opts.name ?: "", root->flags, root->name);
1783 /* remounting is not allowed for populated hierarchies */
1784 if (!list_empty(&root->cgrp.self.children)) {
1789 ret = rebind_subsystems(root, added_mask);
1793 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1795 if (opts.release_agent) {
1796 spin_lock(&release_agent_path_lock);
1797 strcpy(root->release_agent_path, opts.release_agent);
1798 spin_unlock(&release_agent_path_lock);
1801 kfree(opts.release_agent);
1803 mutex_unlock(&cgroup_mutex);
1808 * To reduce the fork() overhead for systems that are not actually using
1809 * their cgroups capability, we don't maintain the lists running through
1810 * each css_set to its tasks until we see the list actually used - in other
1811 * words after the first mount.
1813 static bool use_task_css_set_links __read_mostly;
1815 static void cgroup_enable_task_cg_lists(void)
1817 struct task_struct *p, *g;
1819 spin_lock_bh(&css_set_lock);
1821 if (use_task_css_set_links)
1824 use_task_css_set_links = true;
1827 * We need tasklist_lock because RCU is not safe against
1828 * while_each_thread(). Besides, a forking task that has passed
1829 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1830 * is not guaranteed to have its child immediately visible in the
1831 * tasklist if we walk through it with RCU.
1833 read_lock(&tasklist_lock);
1834 do_each_thread(g, p) {
1835 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1836 task_css_set(p) != &init_css_set);
1839 * We should check if the process is exiting, otherwise
1840 * it will race with cgroup_exit() in that the list
1841 * entry won't be deleted though the process has exited.
1842 * Do it while holding siglock so that we don't end up
1843 * racing against cgroup_exit().
1845 spin_lock_irq(&p->sighand->siglock);
1846 if (!(p->flags & PF_EXITING)) {
1847 struct css_set *cset = task_css_set(p);
1849 if (!css_set_populated(cset))
1850 css_set_update_populated(cset, true);
1851 list_add_tail(&p->cg_list, &cset->tasks);
1854 spin_unlock_irq(&p->sighand->siglock);
1855 } while_each_thread(g, p);
1856 read_unlock(&tasklist_lock);
1858 spin_unlock_bh(&css_set_lock);
1861 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1863 struct cgroup_subsys *ss;
1866 INIT_LIST_HEAD(&cgrp->self.sibling);
1867 INIT_LIST_HEAD(&cgrp->self.children);
1868 INIT_LIST_HEAD(&cgrp->cset_links);
1869 INIT_LIST_HEAD(&cgrp->pidlists);
1870 mutex_init(&cgrp->pidlist_mutex);
1871 cgrp->self.cgroup = cgrp;
1872 cgrp->self.flags |= CSS_ONLINE;
1874 for_each_subsys(ss, ssid)
1875 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1877 init_waitqueue_head(&cgrp->offline_waitq);
1878 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1881 static void init_cgroup_root(struct cgroup_root *root,
1882 struct cgroup_sb_opts *opts)
1884 struct cgroup *cgrp = &root->cgrp;
1886 INIT_LIST_HEAD(&root->root_list);
1887 atomic_set(&root->nr_cgrps, 1);
1889 init_cgroup_housekeeping(cgrp);
1890 idr_init(&root->cgroup_idr);
1892 root->flags = opts->flags;
1893 if (opts->release_agent)
1894 strcpy(root->release_agent_path, opts->release_agent);
1896 strcpy(root->name, opts->name);
1897 if (opts->cpuset_clone_children)
1898 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1901 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1903 LIST_HEAD(tmp_links);
1904 struct cgroup *root_cgrp = &root->cgrp;
1905 struct css_set *cset;
1908 lockdep_assert_held(&cgroup_mutex);
1910 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1913 root_cgrp->id = ret;
1914 root_cgrp->ancestor_ids[0] = ret;
1916 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1922 * We're accessing css_set_count without locking css_set_lock here,
1923 * but that's OK - it can only be increased by someone holding
1924 * cgroup_lock, and that's us. The worst that can happen is that we
1925 * have some link structures left over
1927 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1931 ret = cgroup_init_root_id(root);
1935 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1936 KERNFS_ROOT_CREATE_DEACTIVATED,
1938 if (IS_ERR(root->kf_root)) {
1939 ret = PTR_ERR(root->kf_root);
1942 root_cgrp->kn = root->kf_root->kn;
1944 ret = css_populate_dir(&root_cgrp->self, NULL);
1948 ret = rebind_subsystems(root, ss_mask);
1953 * There must be no failure case after here, since rebinding takes
1954 * care of subsystems' refcounts, which are explicitly dropped in
1955 * the failure exit path.
1957 list_add(&root->root_list, &cgroup_roots);
1958 cgroup_root_count++;
1961 * Link the root cgroup in this hierarchy into all the css_set
1964 spin_lock_bh(&css_set_lock);
1965 hash_for_each(css_set_table, i, cset, hlist) {
1966 link_css_set(&tmp_links, cset, root_cgrp);
1967 if (css_set_populated(cset))
1968 cgroup_update_populated(root_cgrp, true);
1970 spin_unlock_bh(&css_set_lock);
1972 BUG_ON(!list_empty(&root_cgrp->self.children));
1973 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1975 kernfs_activate(root_cgrp->kn);
1980 kernfs_destroy_root(root->kf_root);
1981 root->kf_root = NULL;
1983 cgroup_exit_root_id(root);
1985 percpu_ref_exit(&root_cgrp->self.refcnt);
1987 free_cgrp_cset_links(&tmp_links);
1991 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1992 int flags, const char *unused_dev_name,
1995 bool is_v2 = fs_type == &cgroup2_fs_type;
1996 struct super_block *pinned_sb = NULL;
1997 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
1998 struct cgroup_subsys *ss;
1999 struct cgroup_root *root;
2000 struct cgroup_sb_opts opts;
2001 struct dentry *dentry;
2008 /* Check if the caller has permission to mount. */
2009 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2011 return ERR_PTR(-EPERM);
2015 * The first time anyone tries to mount a cgroup, enable the list
2016 * linking each css_set to its tasks and fix up all existing tasks.
2018 if (!use_task_css_set_links)
2019 cgroup_enable_task_cg_lists();
2023 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data);
2025 return ERR_PTR(-EINVAL);
2027 cgrp_dfl_root_visible = true;
2028 root = &cgrp_dfl_root;
2029 cgroup_get(&root->cgrp);
2033 mutex_lock(&cgroup_mutex);
2035 /* First find the desired set of subsystems */
2036 ret = parse_cgroupfs_options(data, &opts);
2041 * Destruction of cgroup root is asynchronous, so subsystems may
2042 * still be dying after the previous unmount. Let's drain the
2043 * dying subsystems. We just need to ensure that the ones
2044 * unmounted previously finish dying and don't care about new ones
2045 * starting. Testing ref liveliness is good enough.
2047 for_each_subsys(ss, i) {
2048 if (!(opts.subsys_mask & (1 << i)) ||
2049 ss->root == &cgrp_dfl_root)
2052 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2053 mutex_unlock(&cgroup_mutex);
2055 ret = restart_syscall();
2058 cgroup_put(&ss->root->cgrp);
2061 for_each_root(root) {
2062 bool name_match = false;
2064 if (root == &cgrp_dfl_root)
2068 * If we asked for a name then it must match. Also, if
2069 * name matches but sybsys_mask doesn't, we should fail.
2070 * Remember whether name matched.
2073 if (strcmp(opts.name, root->name))
2079 * If we asked for subsystems (or explicitly for no
2080 * subsystems) then they must match.
2082 if ((opts.subsys_mask || opts.none) &&
2083 (opts.subsys_mask != root->subsys_mask)) {
2090 if (root->flags ^ opts.flags)
2091 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2094 * We want to reuse @root whose lifetime is governed by its
2095 * ->cgrp. Let's check whether @root is alive and keep it
2096 * that way. As cgroup_kill_sb() can happen anytime, we
2097 * want to block it by pinning the sb so that @root doesn't
2098 * get killed before mount is complete.
2100 * With the sb pinned, tryget_live can reliably indicate
2101 * whether @root can be reused. If it's being killed,
2102 * drain it. We can use wait_queue for the wait but this
2103 * path is super cold. Let's just sleep a bit and retry.
2105 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2106 if (IS_ERR(pinned_sb) ||
2107 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2108 mutex_unlock(&cgroup_mutex);
2109 if (!IS_ERR_OR_NULL(pinned_sb))
2110 deactivate_super(pinned_sb);
2112 ret = restart_syscall();
2121 * No such thing, create a new one. name= matching without subsys
2122 * specification is allowed for already existing hierarchies but we
2123 * can't create new one without subsys specification.
2125 if (!opts.subsys_mask && !opts.none) {
2131 * We know this subsystem has not yet been bound. Users in a non-init
2132 * user namespace may only mount hierarchies with no bound subsystems,
2133 * i.e. 'none,name=user1'
2135 if (!opts.none && !capable(CAP_SYS_ADMIN)) {
2140 root = kzalloc(sizeof(*root), GFP_KERNEL);
2146 init_cgroup_root(root, &opts);
2148 ret = cgroup_setup_root(root, opts.subsys_mask);
2150 cgroup_free_root(root);
2153 mutex_unlock(&cgroup_mutex);
2155 kfree(opts.release_agent);
2160 return ERR_PTR(ret);
2163 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2164 is_v2 ? CGROUP2_SUPER_MAGIC : CGROUP_SUPER_MAGIC,
2168 * In non-init cgroup namespace, instead of root cgroup's
2169 * dentry, we return the dentry corresponding to the
2170 * cgroupns->root_cgrp.
2172 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
2173 struct dentry *nsdentry;
2174 struct cgroup *cgrp;
2176 mutex_lock(&cgroup_mutex);
2177 spin_lock_bh(&css_set_lock);
2179 cgrp = cset_cgroup_from_root(ns->root_cset, root);
2181 spin_unlock_bh(&css_set_lock);
2182 mutex_unlock(&cgroup_mutex);
2184 nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
2189 if (IS_ERR(dentry) || !new_sb)
2190 cgroup_put(&root->cgrp);
2193 * If @pinned_sb, we're reusing an existing root and holding an
2194 * extra ref on its sb. Mount is complete. Put the extra ref.
2198 deactivate_super(pinned_sb);
2205 static void cgroup_kill_sb(struct super_block *sb)
2207 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2208 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2211 * If @root doesn't have any mounts or children, start killing it.
2212 * This prevents new mounts by disabling percpu_ref_tryget_live().
2213 * cgroup_mount() may wait for @root's release.
2215 * And don't kill the default root.
2217 if (!list_empty(&root->cgrp.self.children) ||
2218 root == &cgrp_dfl_root)
2219 cgroup_put(&root->cgrp);
2221 percpu_ref_kill(&root->cgrp.self.refcnt);
2226 static struct file_system_type cgroup_fs_type = {
2228 .mount = cgroup_mount,
2229 .kill_sb = cgroup_kill_sb,
2230 .fs_flags = FS_USERNS_MOUNT,
2233 static struct file_system_type cgroup2_fs_type = {
2235 .mount = cgroup_mount,
2236 .kill_sb = cgroup_kill_sb,
2237 .fs_flags = FS_USERNS_MOUNT,
2240 static char *cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2241 struct cgroup_namespace *ns)
2243 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2246 ret = kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2247 if (ret < 0 || ret >= buflen)
2252 char *cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2253 struct cgroup_namespace *ns)
2257 mutex_lock(&cgroup_mutex);
2258 spin_lock_bh(&css_set_lock);
2260 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2262 spin_unlock_bh(&css_set_lock);
2263 mutex_unlock(&cgroup_mutex);
2267 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2270 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2271 * @task: target task
2272 * @buf: the buffer to write the path into
2273 * @buflen: the length of the buffer
2275 * Determine @task's cgroup on the first (the one with the lowest non-zero
2276 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2277 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2278 * cgroup controller callbacks.
2280 * Return value is the same as kernfs_path().
2282 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2284 struct cgroup_root *root;
2285 struct cgroup *cgrp;
2286 int hierarchy_id = 1;
2289 mutex_lock(&cgroup_mutex);
2290 spin_lock_bh(&css_set_lock);
2292 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2295 cgrp = task_cgroup_from_root(task, root);
2296 path = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2298 /* if no hierarchy exists, everyone is in "/" */
2299 if (strlcpy(buf, "/", buflen) < buflen)
2303 spin_unlock_bh(&css_set_lock);
2304 mutex_unlock(&cgroup_mutex);
2307 EXPORT_SYMBOL_GPL(task_cgroup_path);
2309 /* used to track tasks and other necessary states during migration */
2310 struct cgroup_taskset {
2311 /* the src and dst cset list running through cset->mg_node */
2312 struct list_head src_csets;
2313 struct list_head dst_csets;
2315 /* the subsys currently being processed */
2319 * Fields for cgroup_taskset_*() iteration.
2321 * Before migration is committed, the target migration tasks are on
2322 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2323 * the csets on ->dst_csets. ->csets point to either ->src_csets
2324 * or ->dst_csets depending on whether migration is committed.
2326 * ->cur_csets and ->cur_task point to the current task position
2329 struct list_head *csets;
2330 struct css_set *cur_cset;
2331 struct task_struct *cur_task;
2334 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2335 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2336 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2337 .csets = &tset.src_csets, \
2341 * cgroup_taskset_add - try to add a migration target task to a taskset
2342 * @task: target task
2343 * @tset: target taskset
2345 * Add @task, which is a migration target, to @tset. This function becomes
2346 * noop if @task doesn't need to be migrated. @task's css_set should have
2347 * been added as a migration source and @task->cg_list will be moved from
2348 * the css_set's tasks list to mg_tasks one.
2350 static void cgroup_taskset_add(struct task_struct *task,
2351 struct cgroup_taskset *tset)
2353 struct css_set *cset;
2355 lockdep_assert_held(&css_set_lock);
2357 /* @task either already exited or can't exit until the end */
2358 if (task->flags & PF_EXITING)
2361 /* leave @task alone if post_fork() hasn't linked it yet */
2362 if (list_empty(&task->cg_list))
2365 cset = task_css_set(task);
2366 if (!cset->mg_src_cgrp)
2369 list_move_tail(&task->cg_list, &cset->mg_tasks);
2370 if (list_empty(&cset->mg_node))
2371 list_add_tail(&cset->mg_node, &tset->src_csets);
2372 if (list_empty(&cset->mg_dst_cset->mg_node))
2373 list_move_tail(&cset->mg_dst_cset->mg_node,
2378 * cgroup_taskset_first - reset taskset and return the first task
2379 * @tset: taskset of interest
2380 * @dst_cssp: output variable for the destination css
2382 * @tset iteration is initialized and the first task is returned.
2384 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2385 struct cgroup_subsys_state **dst_cssp)
2387 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2388 tset->cur_task = NULL;
2390 return cgroup_taskset_next(tset, dst_cssp);
2394 * cgroup_taskset_next - iterate to the next task in taskset
2395 * @tset: taskset of interest
2396 * @dst_cssp: output variable for the destination css
2398 * Return the next task in @tset. Iteration must have been initialized
2399 * with cgroup_taskset_first().
2401 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2402 struct cgroup_subsys_state **dst_cssp)
2404 struct css_set *cset = tset->cur_cset;
2405 struct task_struct *task = tset->cur_task;
2407 while (&cset->mg_node != tset->csets) {
2409 task = list_first_entry(&cset->mg_tasks,
2410 struct task_struct, cg_list);
2412 task = list_next_entry(task, cg_list);
2414 if (&task->cg_list != &cset->mg_tasks) {
2415 tset->cur_cset = cset;
2416 tset->cur_task = task;
2419 * This function may be called both before and
2420 * after cgroup_taskset_migrate(). The two cases
2421 * can be distinguished by looking at whether @cset
2422 * has its ->mg_dst_cset set.
2424 if (cset->mg_dst_cset)
2425 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2427 *dst_cssp = cset->subsys[tset->ssid];
2432 cset = list_next_entry(cset, mg_node);
2440 * cgroup_taskset_migrate - migrate a taskset to a cgroup
2441 * @tset: taget taskset
2442 * @dst_cgrp: destination cgroup
2444 * Migrate tasks in @tset to @dst_cgrp. This function fails iff one of the
2445 * ->can_attach callbacks fails and guarantees that either all or none of
2446 * the tasks in @tset are migrated. @tset is consumed regardless of
2449 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2450 struct cgroup *dst_cgrp)
2452 struct cgroup_subsys_state *css, *failed_css = NULL;
2453 struct task_struct *task, *tmp_task;
2454 struct css_set *cset, *tmp_cset;
2457 /* methods shouldn't be called if no task is actually migrating */
2458 if (list_empty(&tset->src_csets))
2461 /* check that we can legitimately attach to the cgroup */
2462 for_each_e_css(css, i, dst_cgrp) {
2463 if (css->ss->can_attach) {
2465 ret = css->ss->can_attach(tset);
2468 goto out_cancel_attach;
2474 * Now that we're guaranteed success, proceed to move all tasks to
2475 * the new cgroup. There are no failure cases after here, so this
2476 * is the commit point.
2478 spin_lock_bh(&css_set_lock);
2479 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2480 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2481 struct css_set *from_cset = task_css_set(task);
2482 struct css_set *to_cset = cset->mg_dst_cset;
2484 get_css_set(to_cset);
2485 css_set_move_task(task, from_cset, to_cset, true);
2486 put_css_set_locked(from_cset);
2489 spin_unlock_bh(&css_set_lock);
2492 * Migration is committed, all target tasks are now on dst_csets.
2493 * Nothing is sensitive to fork() after this point. Notify
2494 * controllers that migration is complete.
2496 tset->csets = &tset->dst_csets;
2498 for_each_e_css(css, i, dst_cgrp) {
2499 if (css->ss->attach) {
2501 css->ss->attach(tset);
2506 goto out_release_tset;
2509 for_each_e_css(css, i, dst_cgrp) {
2510 if (css == failed_css)
2512 if (css->ss->cancel_attach) {
2514 css->ss->cancel_attach(tset);
2518 spin_lock_bh(&css_set_lock);
2519 list_splice_init(&tset->dst_csets, &tset->src_csets);
2520 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2521 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2522 list_del_init(&cset->mg_node);
2524 spin_unlock_bh(&css_set_lock);
2529 * cgroup_migrate_finish - cleanup after attach
2530 * @preloaded_csets: list of preloaded css_sets
2532 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2533 * those functions for details.
2535 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2537 struct css_set *cset, *tmp_cset;
2539 lockdep_assert_held(&cgroup_mutex);
2541 spin_lock_bh(&css_set_lock);
2542 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2543 cset->mg_src_cgrp = NULL;
2544 cset->mg_dst_cset = NULL;
2545 list_del_init(&cset->mg_preload_node);
2546 put_css_set_locked(cset);
2548 spin_unlock_bh(&css_set_lock);
2552 * cgroup_migrate_add_src - add a migration source css_set
2553 * @src_cset: the source css_set to add
2554 * @dst_cgrp: the destination cgroup
2555 * @preloaded_csets: list of preloaded css_sets
2557 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2558 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2559 * up by cgroup_migrate_finish().
2561 * This function may be called without holding cgroup_threadgroup_rwsem
2562 * even if the target is a process. Threads may be created and destroyed
2563 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2564 * into play and the preloaded css_sets are guaranteed to cover all
2567 static void cgroup_migrate_add_src(struct css_set *src_cset,
2568 struct cgroup *dst_cgrp,
2569 struct list_head *preloaded_csets)
2571 struct cgroup *src_cgrp;
2573 lockdep_assert_held(&cgroup_mutex);
2574 lockdep_assert_held(&css_set_lock);
2576 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2578 if (!list_empty(&src_cset->mg_preload_node))
2581 WARN_ON(src_cset->mg_src_cgrp);
2582 WARN_ON(!list_empty(&src_cset->mg_tasks));
2583 WARN_ON(!list_empty(&src_cset->mg_node));
2585 src_cset->mg_src_cgrp = src_cgrp;
2586 get_css_set(src_cset);
2587 list_add(&src_cset->mg_preload_node, preloaded_csets);
2591 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2592 * @dst_cgrp: the destination cgroup (may be %NULL)
2593 * @preloaded_csets: list of preloaded source css_sets
2595 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2596 * have been preloaded to @preloaded_csets. This function looks up and
2597 * pins all destination css_sets, links each to its source, and append them
2598 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2599 * source css_set is assumed to be its cgroup on the default hierarchy.
2601 * This function must be called after cgroup_migrate_add_src() has been
2602 * called on each migration source css_set. After migration is performed
2603 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2606 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2607 struct list_head *preloaded_csets)
2610 struct css_set *src_cset, *tmp_cset;
2612 lockdep_assert_held(&cgroup_mutex);
2615 * Except for the root, child_subsys_mask must be zero for a cgroup
2616 * with tasks so that child cgroups don't compete against tasks.
2618 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2619 dst_cgrp->child_subsys_mask)
2622 /* look up the dst cset for each src cset and link it to src */
2623 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2624 struct css_set *dst_cset;
2626 dst_cset = find_css_set(src_cset,
2627 dst_cgrp ?: src_cset->dfl_cgrp);
2631 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2634 * If src cset equals dst, it's noop. Drop the src.
2635 * cgroup_migrate() will skip the cset too. Note that we
2636 * can't handle src == dst as some nodes are used by both.
2638 if (src_cset == dst_cset) {
2639 src_cset->mg_src_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(&dst_cset->mg_preload_node, &csets);
2651 put_css_set(dst_cset);
2654 list_splice_tail(&csets, preloaded_csets);
2657 cgroup_migrate_finish(&csets);
2662 * cgroup_migrate - migrate a process or task to a cgroup
2663 * @leader: the leader of the process or the task to migrate
2664 * @threadgroup: whether @leader points to the whole process or a single task
2665 * @cgrp: the destination cgroup
2667 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2668 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2669 * caller is also responsible for invoking cgroup_migrate_add_src() and
2670 * cgroup_migrate_prepare_dst() on the targets before invoking this
2671 * function and following up with cgroup_migrate_finish().
2673 * As long as a controller's ->can_attach() doesn't fail, this function is
2674 * guaranteed to succeed. This means that, excluding ->can_attach()
2675 * failure, when migrating multiple targets, the success or failure can be
2676 * decided for all targets by invoking group_migrate_prepare_dst() before
2677 * actually starting migrating.
2679 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2680 struct cgroup *cgrp)
2682 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
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_bh(&css_set_lock);
2694 cgroup_taskset_add(task, &tset);
2697 } while_each_thread(leader, task);
2699 spin_unlock_bh(&css_set_lock);
2701 return cgroup_taskset_migrate(&tset, cgrp);
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 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2713 struct task_struct *leader, bool threadgroup)
2715 LIST_HEAD(preloaded_csets);
2716 struct task_struct *task;
2719 /* look up all src csets */
2720 spin_lock_bh(&css_set_lock);
2724 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2728 } while_each_thread(leader, task);
2730 spin_unlock_bh(&css_set_lock);
2732 /* prepare dst csets and commit */
2733 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2735 ret = cgroup_migrate(leader, threadgroup, dst_cgrp);
2737 cgroup_migrate_finish(&preloaded_csets);
2741 static int cgroup_procs_write_permission(struct task_struct *task,
2742 struct cgroup *dst_cgrp,
2743 struct kernfs_open_file *of)
2745 const struct cred *cred = current_cred();
2746 const struct cred *tcred = get_task_cred(task);
2750 * even if we're attaching all tasks in the thread group, we only
2751 * need to check permissions on one of them.
2753 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2754 !uid_eq(cred->euid, tcred->uid) &&
2755 !uid_eq(cred->euid, tcred->suid))
2758 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2759 struct super_block *sb = of->file->f_path.dentry->d_sb;
2760 struct cgroup *cgrp;
2761 struct inode *inode;
2763 spin_lock_bh(&css_set_lock);
2764 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2765 spin_unlock_bh(&css_set_lock);
2767 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2768 cgrp = cgroup_parent(cgrp);
2771 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2773 ret = inode_permission(inode, MAY_WRITE);
2783 * Find the task_struct of the task to attach by vpid and pass it along to the
2784 * function to attach either it or all tasks in its threadgroup. Will lock
2785 * cgroup_mutex and threadgroup.
2787 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2788 size_t nbytes, loff_t off, bool threadgroup)
2790 struct task_struct *tsk;
2791 struct cgroup *cgrp;
2795 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2798 cgrp = cgroup_kn_lock_live(of->kn);
2802 percpu_down_write(&cgroup_threadgroup_rwsem);
2805 tsk = find_task_by_vpid(pid);
2808 goto out_unlock_rcu;
2815 tsk = tsk->group_leader;
2818 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2819 * trapped in a cpuset, or RT worker may be born in a cgroup
2820 * with no rt_runtime allocated. Just say no.
2822 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2824 goto out_unlock_rcu;
2827 get_task_struct(tsk);
2830 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2832 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2834 put_task_struct(tsk);
2835 goto out_unlock_threadgroup;
2839 out_unlock_threadgroup:
2840 percpu_up_write(&cgroup_threadgroup_rwsem);
2841 cgroup_kn_unlock(of->kn);
2842 cpuset_post_attach_flush();
2843 return ret ?: nbytes;
2847 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2848 * @from: attach to all cgroups of a given task
2849 * @tsk: the task to be attached
2851 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2853 struct cgroup_root *root;
2856 mutex_lock(&cgroup_mutex);
2857 for_each_root(root) {
2858 struct cgroup *from_cgrp;
2860 if (root == &cgrp_dfl_root)
2863 spin_lock_bh(&css_set_lock);
2864 from_cgrp = task_cgroup_from_root(from, root);
2865 spin_unlock_bh(&css_set_lock);
2867 retval = cgroup_attach_task(from_cgrp, tsk, false);
2871 mutex_unlock(&cgroup_mutex);
2875 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2877 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2878 char *buf, size_t nbytes, loff_t off)
2880 return __cgroup_procs_write(of, buf, nbytes, off, false);
2883 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2884 char *buf, size_t nbytes, loff_t off)
2886 return __cgroup_procs_write(of, buf, nbytes, off, true);
2889 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2890 char *buf, size_t nbytes, loff_t off)
2892 struct cgroup *cgrp;
2894 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2896 cgrp = cgroup_kn_lock_live(of->kn);
2899 spin_lock(&release_agent_path_lock);
2900 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2901 sizeof(cgrp->root->release_agent_path));
2902 spin_unlock(&release_agent_path_lock);
2903 cgroup_kn_unlock(of->kn);
2907 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2909 struct cgroup *cgrp = seq_css(seq)->cgroup;
2911 spin_lock(&release_agent_path_lock);
2912 seq_puts(seq, cgrp->root->release_agent_path);
2913 spin_unlock(&release_agent_path_lock);
2914 seq_putc(seq, '\n');
2918 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2920 seq_puts(seq, "0\n");
2924 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2926 struct cgroup_subsys *ss;
2927 bool printed = false;
2930 for_each_subsys_which(ss, ssid, &ss_mask) {
2933 seq_printf(seq, "%s", ss->name);
2937 seq_putc(seq, '\n');
2940 /* show controllers which are currently attached to the default hierarchy */
2941 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2943 struct cgroup *cgrp = seq_css(seq)->cgroup;
2945 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2946 ~cgrp_dfl_root_inhibit_ss_mask);
2950 /* show controllers which are enabled from the parent */
2951 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2953 struct cgroup *cgrp = seq_css(seq)->cgroup;
2955 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2959 /* show controllers which are enabled for a given cgroup's children */
2960 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2962 struct cgroup *cgrp = seq_css(seq)->cgroup;
2964 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2969 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2970 * @cgrp: root of the subtree to update csses for
2972 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2973 * css associations need to be updated accordingly. This function looks up
2974 * all css_sets which are attached to the subtree, creates the matching
2975 * updated css_sets and migrates the tasks to the new ones.
2977 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2979 LIST_HEAD(preloaded_csets);
2980 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2981 struct cgroup_subsys_state *css;
2982 struct css_set *src_cset;
2985 lockdep_assert_held(&cgroup_mutex);
2987 percpu_down_write(&cgroup_threadgroup_rwsem);
2989 /* look up all csses currently attached to @cgrp's subtree */
2990 spin_lock_bh(&css_set_lock);
2991 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2992 struct cgrp_cset_link *link;
2994 /* self is not affected by child_subsys_mask change */
2995 if (css->cgroup == cgrp)
2998 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2999 cgroup_migrate_add_src(link->cset, cgrp,
3002 spin_unlock_bh(&css_set_lock);
3004 /* NULL dst indicates self on default hierarchy */
3005 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
3009 spin_lock_bh(&css_set_lock);
3010 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
3011 struct task_struct *task, *ntask;
3013 /* src_csets precede dst_csets, break on the first dst_cset */
3014 if (!src_cset->mg_src_cgrp)
3017 /* all tasks in src_csets need to be migrated */
3018 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3019 cgroup_taskset_add(task, &tset);
3021 spin_unlock_bh(&css_set_lock);
3023 ret = cgroup_taskset_migrate(&tset, cgrp);
3025 cgroup_migrate_finish(&preloaded_csets);
3026 percpu_up_write(&cgroup_threadgroup_rwsem);
3030 /* change the enabled child controllers for a cgroup in the default hierarchy */
3031 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3032 char *buf, size_t nbytes,
3035 unsigned long enable = 0, disable = 0;
3036 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
3037 struct cgroup *cgrp, *child;
3038 struct cgroup_subsys *ss;
3043 * Parse input - space separated list of subsystem names prefixed
3044 * with either + or -.
3046 buf = strstrip(buf);
3047 while ((tok = strsep(&buf, " "))) {
3048 unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
3052 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
3053 if (!cgroup_ssid_enabled(ssid) ||
3054 strcmp(tok + 1, ss->name))
3058 enable |= 1 << ssid;
3059 disable &= ~(1 << ssid);
3060 } else if (*tok == '-') {
3061 disable |= 1 << ssid;
3062 enable &= ~(1 << ssid);
3068 if (ssid == CGROUP_SUBSYS_COUNT)
3072 cgrp = cgroup_kn_lock_live(of->kn);
3076 for_each_subsys(ss, ssid) {
3077 if (enable & (1 << ssid)) {
3078 if (cgrp->subtree_control & (1 << ssid)) {
3079 enable &= ~(1 << ssid);
3083 /* unavailable or not enabled on the parent? */
3084 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
3085 (cgroup_parent(cgrp) &&
3086 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
3090 } else if (disable & (1 << ssid)) {
3091 if (!(cgrp->subtree_control & (1 << ssid))) {
3092 disable &= ~(1 << ssid);
3096 /* a child has it enabled? */
3097 cgroup_for_each_live_child(child, cgrp) {
3098 if (child->subtree_control & (1 << ssid)) {
3106 if (!enable && !disable) {
3112 * Except for the root, subtree_control must be zero for a cgroup
3113 * with tasks so that child cgroups don't compete against tasks.
3115 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
3121 * Update subsys masks and calculate what needs to be done. More
3122 * subsystems than specified may need to be enabled or disabled
3123 * depending on subsystem dependencies.
3125 old_sc = cgrp->subtree_control;
3126 old_ss = cgrp->child_subsys_mask;
3127 new_sc = (old_sc | enable) & ~disable;
3128 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
3130 css_enable = ~old_ss & new_ss;
3131 css_disable = old_ss & ~new_ss;
3132 enable |= css_enable;
3133 disable |= css_disable;
3136 * Because css offlining is asynchronous, userland might try to
3137 * re-enable the same controller while the previous instance is
3138 * still around. In such cases, wait till it's gone using
3141 for_each_subsys_which(ss, ssid, &css_enable) {
3142 cgroup_for_each_live_child(child, cgrp) {
3145 if (!cgroup_css(child, ss))
3149 prepare_to_wait(&child->offline_waitq, &wait,
3150 TASK_UNINTERRUPTIBLE);
3151 cgroup_kn_unlock(of->kn);
3153 finish_wait(&child->offline_waitq, &wait);
3156 return restart_syscall();
3160 cgrp->subtree_control = new_sc;
3161 cgrp->child_subsys_mask = new_ss;
3164 * Create new csses or make the existing ones visible. A css is
3165 * created invisible if it's being implicitly enabled through
3166 * dependency. An invisible css is made visible when the userland
3167 * explicitly enables it.
3169 for_each_subsys(ss, ssid) {
3170 if (!(enable & (1 << ssid)))
3173 cgroup_for_each_live_child(child, cgrp) {
3174 if (css_enable & (1 << ssid))
3175 ret = create_css(child, ss,
3176 cgrp->subtree_control & (1 << ssid));
3178 ret = css_populate_dir(cgroup_css(child, ss),
3186 * At this point, cgroup_e_css() results reflect the new csses
3187 * making the following cgroup_update_dfl_csses() properly update
3188 * css associations of all tasks in the subtree.
3190 ret = cgroup_update_dfl_csses(cgrp);
3195 * All tasks are migrated out of disabled csses. Kill or hide
3196 * them. A css is hidden when the userland requests it to be
3197 * disabled while other subsystems are still depending on it. The
3198 * css must not actively control resources and be in the vanilla
3199 * state if it's made visible again later. Controllers which may
3200 * be depended upon should provide ->css_reset() for this purpose.
3202 for_each_subsys(ss, ssid) {
3203 if (!(disable & (1 << ssid)))
3206 cgroup_for_each_live_child(child, cgrp) {
3207 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3209 if (css_disable & (1 << ssid)) {
3212 css_clear_dir(css, NULL);
3220 * The effective csses of all the descendants (excluding @cgrp) may
3221 * have changed. Subsystems can optionally subscribe to this event
3222 * by implementing ->css_e_css_changed() which is invoked if any of
3223 * the effective csses seen from the css's cgroup may have changed.
3225 for_each_subsys(ss, ssid) {
3226 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
3227 struct cgroup_subsys_state *css;
3229 if (!ss->css_e_css_changed || !this_css)
3232 css_for_each_descendant_pre(css, this_css)
3233 if (css != this_css)
3234 ss->css_e_css_changed(css);
3237 kernfs_activate(cgrp->kn);
3240 cgroup_kn_unlock(of->kn);
3241 return ret ?: nbytes;
3244 cgrp->subtree_control = old_sc;
3245 cgrp->child_subsys_mask = old_ss;
3247 for_each_subsys(ss, ssid) {
3248 if (!(enable & (1 << ssid)))
3251 cgroup_for_each_live_child(child, cgrp) {
3252 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3257 if (css_enable & (1 << ssid))
3260 css_clear_dir(css, NULL);
3266 static int cgroup_events_show(struct seq_file *seq, void *v)
3268 seq_printf(seq, "populated %d\n",
3269 cgroup_is_populated(seq_css(seq)->cgroup));
3273 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3274 size_t nbytes, loff_t off)
3276 struct cgroup *cgrp = of->kn->parent->priv;
3277 struct cftype *cft = of->kn->priv;
3278 struct cgroup_subsys_state *css;
3282 return cft->write(of, buf, nbytes, off);
3285 * kernfs guarantees that a file isn't deleted with operations in
3286 * flight, which means that the matching css is and stays alive and
3287 * doesn't need to be pinned. The RCU locking is not necessary
3288 * either. It's just for the convenience of using cgroup_css().
3291 css = cgroup_css(cgrp, cft->ss);
3294 if (cft->write_u64) {
3295 unsigned long long v;
3296 ret = kstrtoull(buf, 0, &v);
3298 ret = cft->write_u64(css, cft, v);
3299 } else if (cft->write_s64) {
3301 ret = kstrtoll(buf, 0, &v);
3303 ret = cft->write_s64(css, cft, v);
3308 return ret ?: nbytes;
3311 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3313 return seq_cft(seq)->seq_start(seq, ppos);
3316 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3318 return seq_cft(seq)->seq_next(seq, v, ppos);
3321 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3323 seq_cft(seq)->seq_stop(seq, v);
3326 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3328 struct cftype *cft = seq_cft(m);
3329 struct cgroup_subsys_state *css = seq_css(m);
3332 return cft->seq_show(m, arg);
3335 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3336 else if (cft->read_s64)
3337 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3343 static struct kernfs_ops cgroup_kf_single_ops = {
3344 .atomic_write_len = PAGE_SIZE,
3345 .write = cgroup_file_write,
3346 .seq_show = cgroup_seqfile_show,
3349 static struct kernfs_ops cgroup_kf_ops = {
3350 .atomic_write_len = PAGE_SIZE,
3351 .write = cgroup_file_write,
3352 .seq_start = cgroup_seqfile_start,
3353 .seq_next = cgroup_seqfile_next,
3354 .seq_stop = cgroup_seqfile_stop,
3355 .seq_show = cgroup_seqfile_show,
3359 * cgroup_rename - Only allow simple rename of directories in place.
3361 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3362 const char *new_name_str)
3364 struct cgroup *cgrp = kn->priv;
3367 if (kernfs_type(kn) != KERNFS_DIR)
3369 if (kn->parent != new_parent)
3373 * This isn't a proper migration and its usefulness is very
3374 * limited. Disallow on the default hierarchy.
3376 if (cgroup_on_dfl(cgrp))
3380 * We're gonna grab cgroup_mutex which nests outside kernfs
3381 * active_ref. kernfs_rename() doesn't require active_ref
3382 * protection. Break them before grabbing cgroup_mutex.
3384 kernfs_break_active_protection(new_parent);
3385 kernfs_break_active_protection(kn);
3387 mutex_lock(&cgroup_mutex);
3389 ret = kernfs_rename(kn, new_parent, new_name_str);
3391 mutex_unlock(&cgroup_mutex);
3393 kernfs_unbreak_active_protection(kn);
3394 kernfs_unbreak_active_protection(new_parent);
3398 /* set uid and gid of cgroup dirs and files to that of the creator */
3399 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3401 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3402 .ia_uid = current_fsuid(),
3403 .ia_gid = current_fsgid(), };
3405 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3406 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3409 return kernfs_setattr(kn, &iattr);
3412 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3415 char name[CGROUP_FILE_NAME_MAX];
3416 struct kernfs_node *kn;
3417 struct lock_class_key *key = NULL;
3420 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3421 key = &cft->lockdep_key;
3423 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3424 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3429 ret = cgroup_kn_set_ugid(kn);
3435 if (cft->file_offset) {
3436 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3438 spin_lock_irq(&cgroup_file_kn_lock);
3440 spin_unlock_irq(&cgroup_file_kn_lock);
3447 * cgroup_addrm_files - add or remove files to a cgroup directory
3448 * @css: the target css
3449 * @cgrp: the target cgroup (usually css->cgroup)
3450 * @cfts: array of cftypes to be added
3451 * @is_add: whether to add or remove
3453 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3454 * For removals, this function never fails.
3456 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3457 struct cgroup *cgrp, struct cftype cfts[],
3460 struct cftype *cft, *cft_end = NULL;
3463 lockdep_assert_held(&cgroup_mutex);
3466 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3467 /* does cft->flags tell us to skip this file on @cgrp? */
3468 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3470 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3472 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3474 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3478 ret = cgroup_add_file(css, cgrp, cft);
3480 pr_warn("%s: failed to add %s, err=%d\n",
3481 __func__, cft->name, ret);
3487 cgroup_rm_file(cgrp, cft);
3493 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3496 struct cgroup_subsys *ss = cfts[0].ss;
3497 struct cgroup *root = &ss->root->cgrp;
3498 struct cgroup_subsys_state *css;
3501 lockdep_assert_held(&cgroup_mutex);
3503 /* add/rm files for all cgroups created before */
3504 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3505 struct cgroup *cgrp = css->cgroup;
3507 if (cgroup_is_dead(cgrp))
3510 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3516 kernfs_activate(root->kn);
3520 static void cgroup_exit_cftypes(struct cftype *cfts)
3524 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3525 /* free copy for custom atomic_write_len, see init_cftypes() */
3526 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3531 /* revert flags set by cgroup core while adding @cfts */
3532 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3536 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3540 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3541 struct kernfs_ops *kf_ops;
3543 WARN_ON(cft->ss || cft->kf_ops);
3546 kf_ops = &cgroup_kf_ops;
3548 kf_ops = &cgroup_kf_single_ops;
3551 * Ugh... if @cft wants a custom max_write_len, we need to
3552 * make a copy of kf_ops to set its atomic_write_len.
3554 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3555 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3557 cgroup_exit_cftypes(cfts);
3560 kf_ops->atomic_write_len = cft->max_write_len;
3563 cft->kf_ops = kf_ops;
3570 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3572 lockdep_assert_held(&cgroup_mutex);
3574 if (!cfts || !cfts[0].ss)
3577 list_del(&cfts->node);
3578 cgroup_apply_cftypes(cfts, false);
3579 cgroup_exit_cftypes(cfts);
3584 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3585 * @cfts: zero-length name terminated array of cftypes
3587 * Unregister @cfts. Files described by @cfts are removed from all
3588 * existing cgroups and all future cgroups won't have them either. This
3589 * function can be called anytime whether @cfts' subsys is attached or not.
3591 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3594 int cgroup_rm_cftypes(struct cftype *cfts)
3598 mutex_lock(&cgroup_mutex);
3599 ret = cgroup_rm_cftypes_locked(cfts);
3600 mutex_unlock(&cgroup_mutex);
3605 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3606 * @ss: target cgroup subsystem
3607 * @cfts: zero-length name terminated array of cftypes
3609 * Register @cfts to @ss. Files described by @cfts are created for all
3610 * existing cgroups to which @ss is attached and all future cgroups will
3611 * have them too. This function can be called anytime whether @ss is
3614 * Returns 0 on successful registration, -errno on failure. Note that this
3615 * function currently returns 0 as long as @cfts registration is successful
3616 * even if some file creation attempts on existing cgroups fail.
3618 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3622 if (!cgroup_ssid_enabled(ss->id))
3625 if (!cfts || cfts[0].name[0] == '\0')
3628 ret = cgroup_init_cftypes(ss, cfts);
3632 mutex_lock(&cgroup_mutex);
3634 list_add_tail(&cfts->node, &ss->cfts);
3635 ret = cgroup_apply_cftypes(cfts, true);
3637 cgroup_rm_cftypes_locked(cfts);
3639 mutex_unlock(&cgroup_mutex);
3644 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3645 * @ss: target cgroup subsystem
3646 * @cfts: zero-length name terminated array of cftypes
3648 * Similar to cgroup_add_cftypes() but the added files are only used for
3649 * the default hierarchy.
3651 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3655 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3656 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3657 return cgroup_add_cftypes(ss, cfts);
3661 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3662 * @ss: target cgroup subsystem
3663 * @cfts: zero-length name terminated array of cftypes
3665 * Similar to cgroup_add_cftypes() but the added files are only used for
3666 * the legacy hierarchies.
3668 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3672 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3673 cft->flags |= __CFTYPE_NOT_ON_DFL;
3674 return cgroup_add_cftypes(ss, cfts);
3678 * cgroup_file_notify - generate a file modified event for a cgroup_file
3679 * @cfile: target cgroup_file
3681 * @cfile must have been obtained by setting cftype->file_offset.
3683 void cgroup_file_notify(struct cgroup_file *cfile)
3685 unsigned long flags;
3687 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3689 kernfs_notify(cfile->kn);
3690 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3694 * cgroup_task_count - count the number of tasks in a cgroup.
3695 * @cgrp: the cgroup in question
3697 * Return the number of tasks in the cgroup.
3699 static int cgroup_task_count(const struct cgroup *cgrp)
3702 struct cgrp_cset_link *link;
3704 spin_lock_bh(&css_set_lock);
3705 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3706 count += atomic_read(&link->cset->refcount);
3707 spin_unlock_bh(&css_set_lock);
3712 * css_next_child - find the next child of a given css
3713 * @pos: the current position (%NULL to initiate traversal)
3714 * @parent: css whose children to walk
3716 * This function returns the next child of @parent and should be called
3717 * under either cgroup_mutex or RCU read lock. The only requirement is
3718 * that @parent and @pos are accessible. The next sibling is guaranteed to
3719 * be returned regardless of their states.
3721 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3722 * css which finished ->css_online() is guaranteed to be visible in the
3723 * future iterations and will stay visible until the last reference is put.
3724 * A css which hasn't finished ->css_online() or already finished
3725 * ->css_offline() may show up during traversal. It's each subsystem's
3726 * responsibility to synchronize against on/offlining.
3728 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3729 struct cgroup_subsys_state *parent)
3731 struct cgroup_subsys_state *next;
3733 cgroup_assert_mutex_or_rcu_locked();
3736 * @pos could already have been unlinked from the sibling list.
3737 * Once a cgroup is removed, its ->sibling.next is no longer
3738 * updated when its next sibling changes. CSS_RELEASED is set when
3739 * @pos is taken off list, at which time its next pointer is valid,
3740 * and, as releases are serialized, the one pointed to by the next
3741 * pointer is guaranteed to not have started release yet. This
3742 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3743 * critical section, the one pointed to by its next pointer is
3744 * guaranteed to not have finished its RCU grace period even if we
3745 * have dropped rcu_read_lock() inbetween iterations.
3747 * If @pos has CSS_RELEASED set, its next pointer can't be
3748 * dereferenced; however, as each css is given a monotonically
3749 * increasing unique serial number and always appended to the
3750 * sibling list, the next one can be found by walking the parent's
3751 * children until the first css with higher serial number than
3752 * @pos's. While this path can be slower, it happens iff iteration
3753 * races against release and the race window is very small.
3756 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3757 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3758 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3760 list_for_each_entry_rcu(next, &parent->children, sibling)
3761 if (next->serial_nr > pos->serial_nr)
3766 * @next, if not pointing to the head, can be dereferenced and is
3769 if (&next->sibling != &parent->children)
3775 * css_next_descendant_pre - find the next descendant for pre-order walk
3776 * @pos: the current position (%NULL to initiate traversal)
3777 * @root: css whose descendants to walk
3779 * To be used by css_for_each_descendant_pre(). Find the next descendant
3780 * to visit for pre-order traversal of @root's descendants. @root is
3781 * included in the iteration and the first node to be visited.
3783 * While this function requires cgroup_mutex or RCU read locking, it
3784 * doesn't require the whole traversal to be contained in a single critical
3785 * section. This function will return the correct next descendant as long
3786 * as both @pos and @root are accessible and @pos is a descendant of @root.
3788 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3789 * css which finished ->css_online() is guaranteed to be visible in the
3790 * future iterations and will stay visible until the last reference is put.
3791 * A css which hasn't finished ->css_online() or already finished
3792 * ->css_offline() may show up during traversal. It's each subsystem's
3793 * responsibility to synchronize against on/offlining.
3795 struct cgroup_subsys_state *
3796 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3797 struct cgroup_subsys_state *root)
3799 struct cgroup_subsys_state *next;
3801 cgroup_assert_mutex_or_rcu_locked();
3803 /* if first iteration, visit @root */
3807 /* visit the first child if exists */
3808 next = css_next_child(NULL, pos);
3812 /* no child, visit my or the closest ancestor's next sibling */
3813 while (pos != root) {
3814 next = css_next_child(pos, pos->parent);
3824 * css_rightmost_descendant - return the rightmost descendant of a css
3825 * @pos: css of interest
3827 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3828 * is returned. This can be used during pre-order traversal to skip
3831 * While this function requires cgroup_mutex or RCU read locking, it
3832 * doesn't require the whole traversal to be contained in a single critical
3833 * section. This function will return the correct rightmost descendant as
3834 * long as @pos is accessible.
3836 struct cgroup_subsys_state *
3837 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3839 struct cgroup_subsys_state *last, *tmp;
3841 cgroup_assert_mutex_or_rcu_locked();
3845 /* ->prev isn't RCU safe, walk ->next till the end */
3847 css_for_each_child(tmp, last)
3854 static struct cgroup_subsys_state *
3855 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3857 struct cgroup_subsys_state *last;
3861 pos = css_next_child(NULL, pos);
3868 * css_next_descendant_post - find the next descendant for post-order walk
3869 * @pos: the current position (%NULL to initiate traversal)
3870 * @root: css whose descendants to walk
3872 * To be used by css_for_each_descendant_post(). Find the next descendant
3873 * to visit for post-order traversal of @root's descendants. @root is
3874 * included in the iteration and the last node to be visited.
3876 * While this function requires cgroup_mutex or RCU read locking, it
3877 * doesn't require the whole traversal to be contained in a single critical
3878 * section. This function will return the correct next descendant as long
3879 * as both @pos and @cgroup are accessible and @pos is a descendant of
3882 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3883 * css which finished ->css_online() is guaranteed to be visible in the
3884 * future iterations and will stay visible until the last reference is put.
3885 * A css which hasn't finished ->css_online() or already finished
3886 * ->css_offline() may show up during traversal. It's each subsystem's
3887 * responsibility to synchronize against on/offlining.
3889 struct cgroup_subsys_state *
3890 css_next_descendant_post(struct cgroup_subsys_state *pos,
3891 struct cgroup_subsys_state *root)
3893 struct cgroup_subsys_state *next;
3895 cgroup_assert_mutex_or_rcu_locked();
3897 /* if first iteration, visit leftmost descendant which may be @root */
3899 return css_leftmost_descendant(root);
3901 /* if we visited @root, we're done */
3905 /* if there's an unvisited sibling, visit its leftmost descendant */
3906 next = css_next_child(pos, pos->parent);
3908 return css_leftmost_descendant(next);
3910 /* no sibling left, visit parent */
3915 * css_has_online_children - does a css have online children
3916 * @css: the target css
3918 * Returns %true if @css has any online children; otherwise, %false. This
3919 * function can be called from any context but the caller is responsible
3920 * for synchronizing against on/offlining as necessary.
3922 bool css_has_online_children(struct cgroup_subsys_state *css)
3924 struct cgroup_subsys_state *child;
3928 css_for_each_child(child, css) {
3929 if (child->flags & CSS_ONLINE) {
3939 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
3940 * @it: the iterator to advance
3942 * Advance @it to the next css_set to walk.
3944 static void css_task_iter_advance_css_set(struct css_task_iter *it)
3946 struct list_head *l = it->cset_pos;
3947 struct cgrp_cset_link *link;
3948 struct css_set *cset;
3950 lockdep_assert_held(&css_set_lock);
3952 /* Advance to the next non-empty css_set */
3955 if (l == it->cset_head) {
3956 it->cset_pos = NULL;
3957 it->task_pos = NULL;
3962 cset = container_of(l, struct css_set,
3963 e_cset_node[it->ss->id]);
3965 link = list_entry(l, struct cgrp_cset_link, cset_link);
3968 } while (!css_set_populated(cset));
3972 if (!list_empty(&cset->tasks))
3973 it->task_pos = cset->tasks.next;
3975 it->task_pos = cset->mg_tasks.next;
3977 it->tasks_head = &cset->tasks;
3978 it->mg_tasks_head = &cset->mg_tasks;
3981 * We don't keep css_sets locked across iteration steps and thus
3982 * need to take steps to ensure that iteration can be resumed after
3983 * the lock is re-acquired. Iteration is performed at two levels -
3984 * css_sets and tasks in them.
3986 * Once created, a css_set never leaves its cgroup lists, so a
3987 * pinned css_set is guaranteed to stay put and we can resume
3988 * iteration afterwards.
3990 * Tasks may leave @cset across iteration steps. This is resolved
3991 * by registering each iterator with the css_set currently being
3992 * walked and making css_set_move_task() advance iterators whose
3993 * next task is leaving.
3996 list_del(&it->iters_node);
3997 put_css_set_locked(it->cur_cset);
4000 it->cur_cset = cset;
4001 list_add(&it->iters_node, &cset->task_iters);
4004 static void css_task_iter_advance(struct css_task_iter *it)
4006 struct list_head *l = it->task_pos;
4008 lockdep_assert_held(&css_set_lock);
4012 * Advance iterator to find next entry. cset->tasks is consumed
4013 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4018 if (l == it->tasks_head)
4019 l = it->mg_tasks_head->next;
4021 if (l == it->mg_tasks_head)
4022 css_task_iter_advance_css_set(it);
4028 * css_task_iter_start - initiate task iteration
4029 * @css: the css to walk tasks of
4030 * @it: the task iterator to use
4032 * Initiate iteration through the tasks of @css. The caller can call
4033 * css_task_iter_next() to walk through the tasks until the function
4034 * returns NULL. On completion of iteration, css_task_iter_end() must be
4037 void css_task_iter_start(struct cgroup_subsys_state *css,
4038 struct css_task_iter *it)
4040 /* no one should try to iterate before mounting cgroups */
4041 WARN_ON_ONCE(!use_task_css_set_links);
4043 memset(it, 0, sizeof(*it));
4045 spin_lock_bh(&css_set_lock);
4050 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4052 it->cset_pos = &css->cgroup->cset_links;
4054 it->cset_head = it->cset_pos;
4056 css_task_iter_advance_css_set(it);
4058 spin_unlock_bh(&css_set_lock);
4062 * css_task_iter_next - return the next task for the iterator
4063 * @it: the task iterator being iterated
4065 * The "next" function for task iteration. @it should have been
4066 * initialized via css_task_iter_start(). Returns NULL when the iteration
4069 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4072 put_task_struct(it->cur_task);
4073 it->cur_task = NULL;
4076 spin_lock_bh(&css_set_lock);
4079 it->cur_task = list_entry(it->task_pos, struct task_struct,
4081 get_task_struct(it->cur_task);
4082 css_task_iter_advance(it);
4085 spin_unlock_bh(&css_set_lock);
4087 return it->cur_task;
4091 * css_task_iter_end - finish task iteration
4092 * @it: the task iterator to finish
4094 * Finish task iteration started by css_task_iter_start().
4096 void css_task_iter_end(struct css_task_iter *it)
4099 spin_lock_bh(&css_set_lock);
4100 list_del(&it->iters_node);
4101 put_css_set_locked(it->cur_cset);
4102 spin_unlock_bh(&css_set_lock);
4106 put_task_struct(it->cur_task);
4110 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4111 * @to: cgroup to which the tasks will be moved
4112 * @from: cgroup in which the tasks currently reside
4114 * Locking rules between cgroup_post_fork() and the migration path
4115 * guarantee that, if a task is forking while being migrated, the new child
4116 * is guaranteed to be either visible in the source cgroup after the
4117 * parent's migration is complete or put into the target cgroup. No task
4118 * can slip out of migration through forking.
4120 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4122 LIST_HEAD(preloaded_csets);
4123 struct cgrp_cset_link *link;
4124 struct css_task_iter it;
4125 struct task_struct *task;
4128 mutex_lock(&cgroup_mutex);
4130 /* all tasks in @from are being moved, all csets are source */
4131 spin_lock_bh(&css_set_lock);
4132 list_for_each_entry(link, &from->cset_links, cset_link)
4133 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4134 spin_unlock_bh(&css_set_lock);
4136 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
4141 * Migrate tasks one-by-one until @from is empty. This fails iff
4142 * ->can_attach() fails.
4145 css_task_iter_start(&from->self, &it);
4146 task = css_task_iter_next(&it);
4148 get_task_struct(task);
4149 css_task_iter_end(&it);
4152 ret = cgroup_migrate(task, false, to);
4153 put_task_struct(task);
4155 } while (task && !ret);
4157 cgroup_migrate_finish(&preloaded_csets);
4158 mutex_unlock(&cgroup_mutex);
4163 * Stuff for reading the 'tasks'/'procs' files.
4165 * Reading this file can return large amounts of data if a cgroup has
4166 * *lots* of attached tasks. So it may need several calls to read(),
4167 * but we cannot guarantee that the information we produce is correct
4168 * unless we produce it entirely atomically.
4172 /* which pidlist file are we talking about? */
4173 enum cgroup_filetype {
4179 * A pidlist is a list of pids that virtually represents the contents of one
4180 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4181 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4184 struct cgroup_pidlist {
4186 * used to find which pidlist is wanted. doesn't change as long as
4187 * this particular list stays in the list.
4189 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4192 /* how many elements the above list has */
4194 /* each of these stored in a list by its cgroup */
4195 struct list_head links;
4196 /* pointer to the cgroup we belong to, for list removal purposes */
4197 struct cgroup *owner;
4198 /* for delayed destruction */
4199 struct delayed_work destroy_dwork;
4203 * The following two functions "fix" the issue where there are more pids
4204 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4205 * TODO: replace with a kernel-wide solution to this problem
4207 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4208 static void *pidlist_allocate(int count)
4210 if (PIDLIST_TOO_LARGE(count))
4211 return vmalloc(count * sizeof(pid_t));
4213 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4216 static void pidlist_free(void *p)
4222 * Used to destroy all pidlists lingering waiting for destroy timer. None
4223 * should be left afterwards.
4225 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4227 struct cgroup_pidlist *l, *tmp_l;
4229 mutex_lock(&cgrp->pidlist_mutex);
4230 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4231 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4232 mutex_unlock(&cgrp->pidlist_mutex);
4234 flush_workqueue(cgroup_pidlist_destroy_wq);
4235 BUG_ON(!list_empty(&cgrp->pidlists));
4238 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4240 struct delayed_work *dwork = to_delayed_work(work);
4241 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4243 struct cgroup_pidlist *tofree = NULL;
4245 mutex_lock(&l->owner->pidlist_mutex);
4248 * Destroy iff we didn't get queued again. The state won't change
4249 * as destroy_dwork can only be queued while locked.
4251 if (!delayed_work_pending(dwork)) {
4252 list_del(&l->links);
4253 pidlist_free(l->list);
4254 put_pid_ns(l->key.ns);
4258 mutex_unlock(&l->owner->pidlist_mutex);
4263 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4264 * Returns the number of unique elements.
4266 static int pidlist_uniq(pid_t *list, int length)
4271 * we presume the 0th element is unique, so i starts at 1. trivial
4272 * edge cases first; no work needs to be done for either
4274 if (length == 0 || length == 1)
4276 /* src and dest walk down the list; dest counts unique elements */
4277 for (src = 1; src < length; src++) {
4278 /* find next unique element */
4279 while (list[src] == list[src-1]) {
4284 /* dest always points to where the next unique element goes */
4285 list[dest] = list[src];
4293 * The two pid files - task and cgroup.procs - guaranteed that the result
4294 * is sorted, which forced this whole pidlist fiasco. As pid order is
4295 * different per namespace, each namespace needs differently sorted list,
4296 * making it impossible to use, for example, single rbtree of member tasks
4297 * sorted by task pointer. As pidlists can be fairly large, allocating one
4298 * per open file is dangerous, so cgroup had to implement shared pool of
4299 * pidlists keyed by cgroup and namespace.
4301 * All this extra complexity was caused by the original implementation
4302 * committing to an entirely unnecessary property. In the long term, we
4303 * want to do away with it. Explicitly scramble sort order if on the
4304 * default hierarchy so that no such expectation exists in the new
4307 * Scrambling is done by swapping every two consecutive bits, which is
4308 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4310 static pid_t pid_fry(pid_t pid)
4312 unsigned a = pid & 0x55555555;
4313 unsigned b = pid & 0xAAAAAAAA;
4315 return (a << 1) | (b >> 1);
4318 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4320 if (cgroup_on_dfl(cgrp))
4321 return pid_fry(pid);
4326 static int cmppid(const void *a, const void *b)
4328 return *(pid_t *)a - *(pid_t *)b;
4331 static int fried_cmppid(const void *a, const void *b)
4333 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4336 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4337 enum cgroup_filetype type)
4339 struct cgroup_pidlist *l;
4340 /* don't need task_nsproxy() if we're looking at ourself */
4341 struct pid_namespace *ns = task_active_pid_ns(current);
4343 lockdep_assert_held(&cgrp->pidlist_mutex);
4345 list_for_each_entry(l, &cgrp->pidlists, links)
4346 if (l->key.type == type && l->key.ns == ns)
4352 * find the appropriate pidlist for our purpose (given procs vs tasks)
4353 * returns with the lock on that pidlist already held, and takes care
4354 * of the use count, or returns NULL with no locks held if we're out of
4357 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4358 enum cgroup_filetype type)
4360 struct cgroup_pidlist *l;
4362 lockdep_assert_held(&cgrp->pidlist_mutex);
4364 l = cgroup_pidlist_find(cgrp, type);
4368 /* entry not found; create a new one */
4369 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4373 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4375 /* don't need task_nsproxy() if we're looking at ourself */
4376 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4378 list_add(&l->links, &cgrp->pidlists);
4383 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4385 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4386 struct cgroup_pidlist **lp)
4390 int pid, n = 0; /* used for populating the array */
4391 struct css_task_iter it;
4392 struct task_struct *tsk;
4393 struct cgroup_pidlist *l;
4395 lockdep_assert_held(&cgrp->pidlist_mutex);
4398 * If cgroup gets more users after we read count, we won't have
4399 * enough space - tough. This race is indistinguishable to the
4400 * caller from the case that the additional cgroup users didn't
4401 * show up until sometime later on.
4403 length = cgroup_task_count(cgrp);
4404 array = pidlist_allocate(length);
4407 /* now, populate the array */
4408 css_task_iter_start(&cgrp->self, &it);
4409 while ((tsk = css_task_iter_next(&it))) {
4410 if (unlikely(n == length))
4412 /* get tgid or pid for procs or tasks file respectively */
4413 if (type == CGROUP_FILE_PROCS)
4414 pid = task_tgid_vnr(tsk);
4416 pid = task_pid_vnr(tsk);
4417 if (pid > 0) /* make sure to only use valid results */
4420 css_task_iter_end(&it);
4422 /* now sort & (if procs) strip out duplicates */
4423 if (cgroup_on_dfl(cgrp))
4424 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4426 sort(array, length, sizeof(pid_t), cmppid, NULL);
4427 if (type == CGROUP_FILE_PROCS)
4428 length = pidlist_uniq(array, length);
4430 l = cgroup_pidlist_find_create(cgrp, type);
4432 pidlist_free(array);
4436 /* store array, freeing old if necessary */
4437 pidlist_free(l->list);
4445 * cgroupstats_build - build and fill cgroupstats
4446 * @stats: cgroupstats to fill information into
4447 * @dentry: A dentry entry belonging to the cgroup for which stats have
4450 * Build and fill cgroupstats so that taskstats can export it to user
4453 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4455 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4456 struct cgroup *cgrp;
4457 struct css_task_iter it;
4458 struct task_struct *tsk;
4460 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4461 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4462 kernfs_type(kn) != KERNFS_DIR)
4465 mutex_lock(&cgroup_mutex);
4468 * We aren't being called from kernfs and there's no guarantee on
4469 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4470 * @kn->priv is RCU safe. Let's do the RCU dancing.
4473 cgrp = rcu_dereference(kn->priv);
4474 if (!cgrp || cgroup_is_dead(cgrp)) {
4476 mutex_unlock(&cgroup_mutex);
4481 css_task_iter_start(&cgrp->self, &it);
4482 while ((tsk = css_task_iter_next(&it))) {
4483 switch (tsk->state) {
4485 stats->nr_running++;
4487 case TASK_INTERRUPTIBLE:
4488 stats->nr_sleeping++;
4490 case TASK_UNINTERRUPTIBLE:
4491 stats->nr_uninterruptible++;
4494 stats->nr_stopped++;
4497 if (delayacct_is_task_waiting_on_io(tsk))
4498 stats->nr_io_wait++;
4502 css_task_iter_end(&it);
4504 mutex_unlock(&cgroup_mutex);
4510 * seq_file methods for the tasks/procs files. The seq_file position is the
4511 * next pid to display; the seq_file iterator is a pointer to the pid
4512 * in the cgroup->l->list array.
4515 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4518 * Initially we receive a position value that corresponds to
4519 * one more than the last pid shown (or 0 on the first call or
4520 * after a seek to the start). Use a binary-search to find the
4521 * next pid to display, if any
4523 struct kernfs_open_file *of = s->private;
4524 struct cgroup *cgrp = seq_css(s)->cgroup;
4525 struct cgroup_pidlist *l;
4526 enum cgroup_filetype type = seq_cft(s)->private;
4527 int index = 0, pid = *pos;
4530 mutex_lock(&cgrp->pidlist_mutex);
4533 * !NULL @of->priv indicates that this isn't the first start()
4534 * after open. If the matching pidlist is around, we can use that.
4535 * Look for it. Note that @of->priv can't be used directly. It
4536 * could already have been destroyed.
4539 of->priv = cgroup_pidlist_find(cgrp, type);
4542 * Either this is the first start() after open or the matching
4543 * pidlist has been destroyed inbetween. Create a new one.
4546 ret = pidlist_array_load(cgrp, type,
4547 (struct cgroup_pidlist **)&of->priv);
4549 return ERR_PTR(ret);
4554 int end = l->length;
4556 while (index < end) {
4557 int mid = (index + end) / 2;
4558 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4561 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4567 /* If we're off the end of the array, we're done */
4568 if (index >= l->length)
4570 /* Update the abstract position to be the actual pid that we found */
4571 iter = l->list + index;
4572 *pos = cgroup_pid_fry(cgrp, *iter);
4576 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4578 struct kernfs_open_file *of = s->private;
4579 struct cgroup_pidlist *l = of->priv;
4582 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4583 CGROUP_PIDLIST_DESTROY_DELAY);
4584 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4587 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4589 struct kernfs_open_file *of = s->private;
4590 struct cgroup_pidlist *l = of->priv;
4592 pid_t *end = l->list + l->length;
4594 * Advance to the next pid in the array. If this goes off the
4601 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4606 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4608 seq_printf(s, "%d\n", *(int *)v);
4613 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4616 return notify_on_release(css->cgroup);
4619 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4620 struct cftype *cft, u64 val)
4623 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4625 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4629 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4632 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4635 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4636 struct cftype *cft, u64 val)
4639 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4641 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4645 /* cgroup core interface files for the default hierarchy */
4646 static struct cftype cgroup_dfl_base_files[] = {
4648 .name = "cgroup.procs",
4649 .file_offset = offsetof(struct cgroup, procs_file),
4650 .seq_start = cgroup_pidlist_start,
4651 .seq_next = cgroup_pidlist_next,
4652 .seq_stop = cgroup_pidlist_stop,
4653 .seq_show = cgroup_pidlist_show,
4654 .private = CGROUP_FILE_PROCS,
4655 .write = cgroup_procs_write,
4658 .name = "cgroup.controllers",
4659 .flags = CFTYPE_ONLY_ON_ROOT,
4660 .seq_show = cgroup_root_controllers_show,
4663 .name = "cgroup.controllers",
4664 .flags = CFTYPE_NOT_ON_ROOT,
4665 .seq_show = cgroup_controllers_show,
4668 .name = "cgroup.subtree_control",
4669 .seq_show = cgroup_subtree_control_show,
4670 .write = cgroup_subtree_control_write,
4673 .name = "cgroup.events",
4674 .flags = CFTYPE_NOT_ON_ROOT,
4675 .file_offset = offsetof(struct cgroup, events_file),
4676 .seq_show = cgroup_events_show,
4681 /* cgroup core interface files for the legacy hierarchies */
4682 static struct cftype cgroup_legacy_base_files[] = {
4684 .name = "cgroup.procs",
4685 .seq_start = cgroup_pidlist_start,
4686 .seq_next = cgroup_pidlist_next,
4687 .seq_stop = cgroup_pidlist_stop,
4688 .seq_show = cgroup_pidlist_show,
4689 .private = CGROUP_FILE_PROCS,
4690 .write = cgroup_procs_write,
4693 .name = "cgroup.clone_children",
4694 .read_u64 = cgroup_clone_children_read,
4695 .write_u64 = cgroup_clone_children_write,
4698 .name = "cgroup.sane_behavior",
4699 .flags = CFTYPE_ONLY_ON_ROOT,
4700 .seq_show = cgroup_sane_behavior_show,
4704 .seq_start = cgroup_pidlist_start,
4705 .seq_next = cgroup_pidlist_next,
4706 .seq_stop = cgroup_pidlist_stop,
4707 .seq_show = cgroup_pidlist_show,
4708 .private = CGROUP_FILE_TASKS,
4709 .write = cgroup_tasks_write,
4712 .name = "notify_on_release",
4713 .read_u64 = cgroup_read_notify_on_release,
4714 .write_u64 = cgroup_write_notify_on_release,
4717 .name = "release_agent",
4718 .flags = CFTYPE_ONLY_ON_ROOT,
4719 .seq_show = cgroup_release_agent_show,
4720 .write = cgroup_release_agent_write,
4721 .max_write_len = PATH_MAX - 1,
4727 * css destruction is four-stage process.
4729 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4730 * Implemented in kill_css().
4732 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4733 * and thus css_tryget_online() is guaranteed to fail, the css can be
4734 * offlined by invoking offline_css(). After offlining, the base ref is
4735 * put. Implemented in css_killed_work_fn().
4737 * 3. When the percpu_ref reaches zero, the only possible remaining
4738 * accessors are inside RCU read sections. css_release() schedules the
4741 * 4. After the grace period, the css can be freed. Implemented in
4742 * css_free_work_fn().
4744 * It is actually hairier because both step 2 and 4 require process context
4745 * and thus involve punting to css->destroy_work adding two additional
4746 * steps to the already complex sequence.
4748 static void css_free_work_fn(struct work_struct *work)
4750 struct cgroup_subsys_state *css =
4751 container_of(work, struct cgroup_subsys_state, destroy_work);
4752 struct cgroup_subsys *ss = css->ss;
4753 struct cgroup *cgrp = css->cgroup;
4755 percpu_ref_exit(&css->refcnt);
4759 struct cgroup_subsys_state *parent = css->parent;
4763 cgroup_idr_remove(&ss->css_idr, id);
4769 /* cgroup free path */
4770 atomic_dec(&cgrp->root->nr_cgrps);
4771 cgroup_pidlist_destroy_all(cgrp);
4772 cancel_work_sync(&cgrp->release_agent_work);
4774 if (cgroup_parent(cgrp)) {
4776 * We get a ref to the parent, and put the ref when
4777 * this cgroup is being freed, so it's guaranteed
4778 * that the parent won't be destroyed before its
4781 cgroup_put(cgroup_parent(cgrp));
4782 kernfs_put(cgrp->kn);
4786 * This is root cgroup's refcnt reaching zero,
4787 * which indicates that the root should be
4790 cgroup_destroy_root(cgrp->root);
4795 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4797 struct cgroup_subsys_state *css =
4798 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4800 INIT_WORK(&css->destroy_work, css_free_work_fn);
4801 queue_work(cgroup_destroy_wq, &css->destroy_work);
4804 static void css_release_work_fn(struct work_struct *work)
4806 struct cgroup_subsys_state *css =
4807 container_of(work, struct cgroup_subsys_state, destroy_work);
4808 struct cgroup_subsys *ss = css->ss;
4809 struct cgroup *cgrp = css->cgroup;
4811 mutex_lock(&cgroup_mutex);
4813 css->flags |= CSS_RELEASED;
4814 list_del_rcu(&css->sibling);
4817 /* css release path */
4818 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4819 if (ss->css_released)
4820 ss->css_released(css);
4822 /* cgroup release path */
4823 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4827 * There are two control paths which try to determine
4828 * cgroup from dentry without going through kernfs -
4829 * cgroupstats_build() and css_tryget_online_from_dir().
4830 * Those are supported by RCU protecting clearing of
4831 * cgrp->kn->priv backpointer.
4833 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4836 mutex_unlock(&cgroup_mutex);
4838 call_rcu(&css->rcu_head, css_free_rcu_fn);
4841 static void css_release(struct percpu_ref *ref)
4843 struct cgroup_subsys_state *css =
4844 container_of(ref, struct cgroup_subsys_state, refcnt);
4846 INIT_WORK(&css->destroy_work, css_release_work_fn);
4847 queue_work(cgroup_destroy_wq, &css->destroy_work);
4850 static void init_and_link_css(struct cgroup_subsys_state *css,
4851 struct cgroup_subsys *ss, struct cgroup *cgrp)
4853 lockdep_assert_held(&cgroup_mutex);
4857 memset(css, 0, sizeof(*css));
4860 INIT_LIST_HEAD(&css->sibling);
4861 INIT_LIST_HEAD(&css->children);
4862 css->serial_nr = css_serial_nr_next++;
4863 atomic_set(&css->online_cnt, 0);
4865 if (cgroup_parent(cgrp)) {
4866 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4867 css_get(css->parent);
4870 BUG_ON(cgroup_css(cgrp, ss));
4873 /* invoke ->css_online() on a new CSS and mark it online if successful */
4874 static int online_css(struct cgroup_subsys_state *css)
4876 struct cgroup_subsys *ss = css->ss;
4879 lockdep_assert_held(&cgroup_mutex);
4882 ret = ss->css_online(css);
4884 css->flags |= CSS_ONLINE;
4885 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4887 atomic_inc(&css->online_cnt);
4889 atomic_inc(&css->parent->online_cnt);
4894 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4895 static void offline_css(struct cgroup_subsys_state *css)
4897 struct cgroup_subsys *ss = css->ss;
4899 lockdep_assert_held(&cgroup_mutex);
4901 if (!(css->flags & CSS_ONLINE))
4904 if (ss->css_offline)
4905 ss->css_offline(css);
4907 css->flags &= ~CSS_ONLINE;
4908 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4910 wake_up_all(&css->cgroup->offline_waitq);
4914 * create_css - create a cgroup_subsys_state
4915 * @cgrp: the cgroup new css will be associated with
4916 * @ss: the subsys of new css
4917 * @visible: whether to create control knobs for the new css or not
4919 * Create a new css associated with @cgrp - @ss pair. On success, the new
4920 * css is online and installed in @cgrp with all interface files created if
4921 * @visible. Returns 0 on success, -errno on failure.
4923 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4926 struct cgroup *parent = cgroup_parent(cgrp);
4927 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4928 struct cgroup_subsys_state *css;
4931 lockdep_assert_held(&cgroup_mutex);
4933 css = ss->css_alloc(parent_css);
4935 return PTR_ERR(css);
4937 init_and_link_css(css, ss, cgrp);
4939 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4943 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4945 goto err_free_percpu_ref;
4949 err = css_populate_dir(css, NULL);
4954 /* @css is ready to be brought online now, make it visible */
4955 list_add_tail_rcu(&css->sibling, &parent_css->children);
4956 cgroup_idr_replace(&ss->css_idr, css, css->id);
4958 err = online_css(css);
4962 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4963 cgroup_parent(parent)) {
4964 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4965 current->comm, current->pid, ss->name);
4966 if (!strcmp(ss->name, "memory"))
4967 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4968 ss->warned_broken_hierarchy = true;
4974 list_del_rcu(&css->sibling);
4975 css_clear_dir(css, NULL);
4977 cgroup_idr_remove(&ss->css_idr, css->id);
4978 err_free_percpu_ref:
4979 percpu_ref_exit(&css->refcnt);
4981 call_rcu(&css->rcu_head, css_free_rcu_fn);
4985 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4988 struct cgroup *parent, *cgrp, *tcgrp;
4989 struct cgroup_root *root;
4990 struct cgroup_subsys *ss;
4991 struct kernfs_node *kn;
4992 int level, ssid, ret;
4994 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4996 if (strchr(name, '\n'))
4999 parent = cgroup_kn_lock_live(parent_kn);
5002 root = parent->root;
5003 level = parent->level + 1;
5005 /* allocate the cgroup and its ID, 0 is reserved for the root */
5006 cgrp = kzalloc(sizeof(*cgrp) +
5007 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
5013 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5018 * Temporarily set the pointer to NULL, so idr_find() won't return
5019 * a half-baked cgroup.
5021 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
5024 goto out_cancel_ref;
5027 init_cgroup_housekeeping(cgrp);
5029 cgrp->self.parent = &parent->self;
5031 cgrp->level = level;
5033 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
5034 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
5036 if (notify_on_release(parent))
5037 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5039 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5040 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5042 /* create the directory */
5043 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5051 * This extra ref will be put in cgroup_free_fn() and guarantees
5052 * that @cgrp->kn is always accessible.
5056 cgrp->self.serial_nr = css_serial_nr_next++;
5058 /* allocation complete, commit to creation */
5059 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5060 atomic_inc(&root->nr_cgrps);
5064 * @cgrp is now fully operational. If something fails after this
5065 * point, it'll be released via the normal destruction path.
5067 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5069 ret = cgroup_kn_set_ugid(kn);
5073 ret = css_populate_dir(&cgrp->self, NULL);
5077 /* let's create and online css's */
5078 for_each_subsys(ss, ssid) {
5079 if (parent->child_subsys_mask & (1 << ssid)) {
5080 ret = create_css(cgrp, ss,
5081 parent->subtree_control & (1 << ssid));
5088 * On the default hierarchy, a child doesn't automatically inherit
5089 * subtree_control from the parent. Each is configured manually.
5091 if (!cgroup_on_dfl(cgrp)) {
5092 cgrp->subtree_control = parent->subtree_control;
5093 cgroup_refresh_child_subsys_mask(cgrp);
5096 kernfs_activate(kn);
5102 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
5104 percpu_ref_exit(&cgrp->self.refcnt);
5108 cgroup_kn_unlock(parent_kn);
5112 cgroup_destroy_locked(cgrp);
5117 * This is called when the refcnt of a css is confirmed to be killed.
5118 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5119 * initate destruction and put the css ref from kill_css().
5121 static void css_killed_work_fn(struct work_struct *work)
5123 struct cgroup_subsys_state *css =
5124 container_of(work, struct cgroup_subsys_state, destroy_work);
5126 mutex_lock(&cgroup_mutex);
5131 /* @css can't go away while we're holding cgroup_mutex */
5133 } while (css && atomic_dec_and_test(&css->online_cnt));
5135 mutex_unlock(&cgroup_mutex);
5138 /* css kill confirmation processing requires process context, bounce */
5139 static void css_killed_ref_fn(struct percpu_ref *ref)
5141 struct cgroup_subsys_state *css =
5142 container_of(ref, struct cgroup_subsys_state, refcnt);
5144 if (atomic_dec_and_test(&css->online_cnt)) {
5145 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5146 queue_work(cgroup_destroy_wq, &css->destroy_work);
5151 * kill_css - destroy a css
5152 * @css: css to destroy
5154 * This function initiates destruction of @css by removing cgroup interface
5155 * files and putting its base reference. ->css_offline() will be invoked
5156 * asynchronously once css_tryget_online() is guaranteed to fail and when
5157 * the reference count reaches zero, @css will be released.
5159 static void kill_css(struct cgroup_subsys_state *css)
5161 lockdep_assert_held(&cgroup_mutex);
5164 * This must happen before css is disassociated with its cgroup.
5165 * See seq_css() for details.
5167 css_clear_dir(css, NULL);
5170 * Killing would put the base ref, but we need to keep it alive
5171 * until after ->css_offline().
5176 * cgroup core guarantees that, by the time ->css_offline() is
5177 * invoked, no new css reference will be given out via
5178 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5179 * proceed to offlining css's because percpu_ref_kill() doesn't
5180 * guarantee that the ref is seen as killed on all CPUs on return.
5182 * Use percpu_ref_kill_and_confirm() to get notifications as each
5183 * css is confirmed to be seen as killed on all CPUs.
5185 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5189 * cgroup_destroy_locked - the first stage of cgroup destruction
5190 * @cgrp: cgroup to be destroyed
5192 * css's make use of percpu refcnts whose killing latency shouldn't be
5193 * exposed to userland and are RCU protected. Also, cgroup core needs to
5194 * guarantee that css_tryget_online() won't succeed by the time
5195 * ->css_offline() is invoked. To satisfy all the requirements,
5196 * destruction is implemented in the following two steps.
5198 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5199 * userland visible parts and start killing the percpu refcnts of
5200 * css's. Set up so that the next stage will be kicked off once all
5201 * the percpu refcnts are confirmed to be killed.
5203 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5204 * rest of destruction. Once all cgroup references are gone, the
5205 * cgroup is RCU-freed.
5207 * This function implements s1. After this step, @cgrp is gone as far as
5208 * the userland is concerned and a new cgroup with the same name may be
5209 * created. As cgroup doesn't care about the names internally, this
5210 * doesn't cause any problem.
5212 static int cgroup_destroy_locked(struct cgroup *cgrp)
5213 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5215 struct cgroup_subsys_state *css;
5218 lockdep_assert_held(&cgroup_mutex);
5221 * Only migration can raise populated from zero and we're already
5222 * holding cgroup_mutex.
5224 if (cgroup_is_populated(cgrp))
5228 * Make sure there's no live children. We can't test emptiness of
5229 * ->self.children as dead children linger on it while being
5230 * drained; otherwise, "rmdir parent/child parent" may fail.
5232 if (css_has_online_children(&cgrp->self))
5236 * Mark @cgrp dead. This prevents further task migration and child
5237 * creation by disabling cgroup_lock_live_group().
5239 cgrp->self.flags &= ~CSS_ONLINE;
5241 /* initiate massacre of all css's */
5242 for_each_css(css, ssid, cgrp)
5246 * Remove @cgrp directory along with the base files. @cgrp has an
5247 * extra ref on its kn.
5249 kernfs_remove(cgrp->kn);
5251 check_for_release(cgroup_parent(cgrp));
5253 /* put the base reference */
5254 percpu_ref_kill(&cgrp->self.refcnt);
5259 static int cgroup_rmdir(struct kernfs_node *kn)
5261 struct cgroup *cgrp;
5264 cgrp = cgroup_kn_lock_live(kn);
5268 ret = cgroup_destroy_locked(cgrp);
5270 cgroup_kn_unlock(kn);
5274 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5275 .remount_fs = cgroup_remount,
5276 .show_options = cgroup_show_options,
5277 .mkdir = cgroup_mkdir,
5278 .rmdir = cgroup_rmdir,
5279 .rename = cgroup_rename,
5282 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5284 struct cgroup_subsys_state *css;
5286 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5288 mutex_lock(&cgroup_mutex);
5290 idr_init(&ss->css_idr);
5291 INIT_LIST_HEAD(&ss->cfts);
5293 /* Create the root cgroup state for this subsystem */
5294 ss->root = &cgrp_dfl_root;
5295 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5296 /* We don't handle early failures gracefully */
5297 BUG_ON(IS_ERR(css));
5298 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5301 * Root csses are never destroyed and we can't initialize
5302 * percpu_ref during early init. Disable refcnting.
5304 css->flags |= CSS_NO_REF;
5307 /* allocation can't be done safely during early init */
5310 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5311 BUG_ON(css->id < 0);
5314 /* Update the init_css_set to contain a subsys
5315 * pointer to this state - since the subsystem is
5316 * newly registered, all tasks and hence the
5317 * init_css_set is in the subsystem's root cgroup. */
5318 init_css_set.subsys[ss->id] = css;
5320 have_fork_callback |= (bool)ss->fork << ss->id;
5321 have_exit_callback |= (bool)ss->exit << ss->id;
5322 have_free_callback |= (bool)ss->free << ss->id;
5323 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5325 /* At system boot, before all subsystems have been
5326 * registered, no tasks have been forked, so we don't
5327 * need to invoke fork callbacks here. */
5328 BUG_ON(!list_empty(&init_task.tasks));
5330 BUG_ON(online_css(css));
5332 mutex_unlock(&cgroup_mutex);
5336 * cgroup_init_early - cgroup initialization at system boot
5338 * Initialize cgroups at system boot, and initialize any
5339 * subsystems that request early init.
5341 int __init cgroup_init_early(void)
5343 static struct cgroup_sb_opts __initdata opts;
5344 struct cgroup_subsys *ss;
5347 init_cgroup_root(&cgrp_dfl_root, &opts);
5348 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5350 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5352 for_each_subsys(ss, i) {
5353 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5354 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
5355 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5357 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5358 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5361 ss->name = cgroup_subsys_name[i];
5362 if (!ss->legacy_name)
5363 ss->legacy_name = cgroup_subsys_name[i];
5366 cgroup_init_subsys(ss, true);
5371 static unsigned long cgroup_disable_mask __initdata;
5374 * cgroup_init - cgroup initialization
5376 * Register cgroup filesystem and /proc file, and initialize
5377 * any subsystems that didn't request early init.
5379 int __init cgroup_init(void)
5381 struct cgroup_subsys *ss;
5385 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5386 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5387 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5389 get_user_ns(init_cgroup_ns.user_ns);
5391 mutex_lock(&cgroup_mutex);
5393 /* Add init_css_set to the hash table */
5394 key = css_set_hash(init_css_set.subsys);
5395 hash_add(css_set_table, &init_css_set.hlist, key);
5397 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5399 mutex_unlock(&cgroup_mutex);
5401 for_each_subsys(ss, ssid) {
5402 if (ss->early_init) {
5403 struct cgroup_subsys_state *css =
5404 init_css_set.subsys[ss->id];
5406 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5408 BUG_ON(css->id < 0);
5410 cgroup_init_subsys(ss, false);
5413 list_add_tail(&init_css_set.e_cset_node[ssid],
5414 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5417 * Setting dfl_root subsys_mask needs to consider the
5418 * disabled flag and cftype registration needs kmalloc,
5419 * both of which aren't available during early_init.
5421 if (cgroup_disable_mask & (1 << ssid)) {
5422 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5423 printk(KERN_INFO "Disabling %s control group subsystem\n",
5428 if (cgroup_ssid_no_v1(ssid))
5429 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5432 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5434 if (!ss->dfl_cftypes)
5435 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5437 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5438 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5440 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5441 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5445 ss->bind(init_css_set.subsys[ssid]);
5448 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5449 WARN_ON(register_filesystem(&cgroup_fs_type));
5450 WARN_ON(register_filesystem(&cgroup2_fs_type));
5451 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5456 static int __init cgroup_wq_init(void)
5459 * There isn't much point in executing destruction path in
5460 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5461 * Use 1 for @max_active.
5463 * We would prefer to do this in cgroup_init() above, but that
5464 * is called before init_workqueues(): so leave this until after.
5466 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5467 BUG_ON(!cgroup_destroy_wq);
5470 * Used to destroy pidlists and separate to serve as flush domain.
5471 * Cap @max_active to 1 too.
5473 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5475 BUG_ON(!cgroup_pidlist_destroy_wq);
5479 core_initcall(cgroup_wq_init);
5482 * proc_cgroup_show()
5483 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5484 * - Used for /proc/<pid>/cgroup.
5486 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5487 struct pid *pid, struct task_struct *tsk)
5491 struct cgroup_root *root;
5494 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5498 mutex_lock(&cgroup_mutex);
5499 spin_lock_bh(&css_set_lock);
5501 for_each_root(root) {
5502 struct cgroup_subsys *ss;
5503 struct cgroup *cgrp;
5504 int ssid, count = 0;
5506 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5509 seq_printf(m, "%d:", root->hierarchy_id);
5510 if (root != &cgrp_dfl_root)
5511 for_each_subsys(ss, ssid)
5512 if (root->subsys_mask & (1 << ssid))
5513 seq_printf(m, "%s%s", count++ ? "," : "",
5515 if (strlen(root->name))
5516 seq_printf(m, "%sname=%s", count ? "," : "",
5520 cgrp = task_cgroup_from_root(tsk, root);
5523 * On traditional hierarchies, all zombie tasks show up as
5524 * belonging to the root cgroup. On the default hierarchy,
5525 * while a zombie doesn't show up in "cgroup.procs" and
5526 * thus can't be migrated, its /proc/PID/cgroup keeps
5527 * reporting the cgroup it belonged to before exiting. If
5528 * the cgroup is removed before the zombie is reaped,
5529 * " (deleted)" is appended to the cgroup path.
5531 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5532 path = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5533 current->nsproxy->cgroup_ns);
5535 retval = -ENAMETOOLONG;
5544 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5545 seq_puts(m, " (deleted)\n");
5552 spin_unlock_bh(&css_set_lock);
5553 mutex_unlock(&cgroup_mutex);
5559 /* Display information about each subsystem and each hierarchy */
5560 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5562 struct cgroup_subsys *ss;
5565 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5567 * ideally we don't want subsystems moving around while we do this.
5568 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5569 * subsys/hierarchy state.
5571 mutex_lock(&cgroup_mutex);
5573 for_each_subsys(ss, i)
5574 seq_printf(m, "%s\t%d\t%d\t%d\n",
5575 ss->legacy_name, ss->root->hierarchy_id,
5576 atomic_read(&ss->root->nr_cgrps),
5577 cgroup_ssid_enabled(i));
5579 mutex_unlock(&cgroup_mutex);
5583 static int cgroupstats_open(struct inode *inode, struct file *file)
5585 return single_open(file, proc_cgroupstats_show, NULL);
5588 static const struct file_operations proc_cgroupstats_operations = {
5589 .open = cgroupstats_open,
5591 .llseek = seq_lseek,
5592 .release = single_release,
5596 * cgroup_fork - initialize cgroup related fields during copy_process()
5597 * @child: pointer to task_struct of forking parent process.
5599 * A task is associated with the init_css_set until cgroup_post_fork()
5600 * attaches it to the parent's css_set. Empty cg_list indicates that
5601 * @child isn't holding reference to its css_set.
5603 void cgroup_fork(struct task_struct *child)
5605 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5606 INIT_LIST_HEAD(&child->cg_list);
5610 * cgroup_can_fork - called on a new task before the process is exposed
5611 * @child: the task in question.
5613 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5614 * returns an error, the fork aborts with that error code. This allows for
5615 * a cgroup subsystem to conditionally allow or deny new forks.
5617 int cgroup_can_fork(struct task_struct *child)
5619 struct cgroup_subsys *ss;
5622 for_each_subsys_which(ss, i, &have_canfork_callback) {
5623 ret = ss->can_fork(child);
5631 for_each_subsys(ss, j) {
5634 if (ss->cancel_fork)
5635 ss->cancel_fork(child);
5642 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5643 * @child: the task in question
5645 * This calls the cancel_fork() callbacks if a fork failed *after*
5646 * cgroup_can_fork() succeded.
5648 void cgroup_cancel_fork(struct task_struct *child)
5650 struct cgroup_subsys *ss;
5653 for_each_subsys(ss, i)
5654 if (ss->cancel_fork)
5655 ss->cancel_fork(child);
5659 * cgroup_post_fork - called on a new task after adding it to the task list
5660 * @child: the task in question
5662 * Adds the task to the list running through its css_set if necessary and
5663 * call the subsystem fork() callbacks. Has to be after the task is
5664 * visible on the task list in case we race with the first call to
5665 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5668 void cgroup_post_fork(struct task_struct *child)
5670 struct cgroup_subsys *ss;
5674 * This may race against cgroup_enable_task_cg_lists(). As that
5675 * function sets use_task_css_set_links before grabbing
5676 * tasklist_lock and we just went through tasklist_lock to add
5677 * @child, it's guaranteed that either we see the set
5678 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5679 * @child during its iteration.
5681 * If we won the race, @child is associated with %current's
5682 * css_set. Grabbing css_set_lock guarantees both that the
5683 * association is stable, and, on completion of the parent's
5684 * migration, @child is visible in the source of migration or
5685 * already in the destination cgroup. This guarantee is necessary
5686 * when implementing operations which need to migrate all tasks of
5687 * a cgroup to another.
5689 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5690 * will remain in init_css_set. This is safe because all tasks are
5691 * in the init_css_set before cg_links is enabled and there's no
5692 * operation which transfers all tasks out of init_css_set.
5694 if (use_task_css_set_links) {
5695 struct css_set *cset;
5697 spin_lock_bh(&css_set_lock);
5698 cset = task_css_set(current);
5699 if (list_empty(&child->cg_list)) {
5701 css_set_move_task(child, NULL, cset, false);
5703 spin_unlock_bh(&css_set_lock);
5707 * Call ss->fork(). This must happen after @child is linked on
5708 * css_set; otherwise, @child might change state between ->fork()
5709 * and addition to css_set.
5711 for_each_subsys_which(ss, i, &have_fork_callback)
5716 * cgroup_exit - detach cgroup from exiting task
5717 * @tsk: pointer to task_struct of exiting process
5719 * Description: Detach cgroup from @tsk and release it.
5721 * Note that cgroups marked notify_on_release force every task in
5722 * them to take the global cgroup_mutex mutex when exiting.
5723 * This could impact scaling on very large systems. Be reluctant to
5724 * use notify_on_release cgroups where very high task exit scaling
5725 * is required on large systems.
5727 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5728 * call cgroup_exit() while the task is still competent to handle
5729 * notify_on_release(), then leave the task attached to the root cgroup in
5730 * each hierarchy for the remainder of its exit. No need to bother with
5731 * init_css_set refcnting. init_css_set never goes away and we can't race
5732 * with migration path - PF_EXITING is visible to migration path.
5734 void cgroup_exit(struct task_struct *tsk)
5736 struct cgroup_subsys *ss;
5737 struct css_set *cset;
5741 * Unlink from @tsk from its css_set. As migration path can't race
5742 * with us, we can check css_set and cg_list without synchronization.
5744 cset = task_css_set(tsk);
5746 if (!list_empty(&tsk->cg_list)) {
5747 spin_lock_bh(&css_set_lock);
5748 css_set_move_task(tsk, cset, NULL, false);
5749 spin_unlock_bh(&css_set_lock);
5754 /* see cgroup_post_fork() for details */
5755 for_each_subsys_which(ss, i, &have_exit_callback)
5759 void cgroup_free(struct task_struct *task)
5761 struct css_set *cset = task_css_set(task);
5762 struct cgroup_subsys *ss;
5765 for_each_subsys_which(ss, ssid, &have_free_callback)
5771 static void check_for_release(struct cgroup *cgrp)
5773 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
5774 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5775 schedule_work(&cgrp->release_agent_work);
5779 * Notify userspace when a cgroup is released, by running the
5780 * configured release agent with the name of the cgroup (path
5781 * relative to the root of cgroup file system) as the argument.
5783 * Most likely, this user command will try to rmdir this cgroup.
5785 * This races with the possibility that some other task will be
5786 * attached to this cgroup before it is removed, or that some other
5787 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5788 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5789 * unused, and this cgroup will be reprieved from its death sentence,
5790 * to continue to serve a useful existence. Next time it's released,
5791 * we will get notified again, if it still has 'notify_on_release' set.
5793 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5794 * means only wait until the task is successfully execve()'d. The
5795 * separate release agent task is forked by call_usermodehelper(),
5796 * then control in this thread returns here, without waiting for the
5797 * release agent task. We don't bother to wait because the caller of
5798 * this routine has no use for the exit status of the release agent
5799 * task, so no sense holding our caller up for that.
5801 static void cgroup_release_agent(struct work_struct *work)
5803 struct cgroup *cgrp =
5804 container_of(work, struct cgroup, release_agent_work);
5805 char *pathbuf = NULL, *agentbuf = NULL, *path;
5806 char *argv[3], *envp[3];
5808 mutex_lock(&cgroup_mutex);
5810 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5811 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5812 if (!pathbuf || !agentbuf)
5815 spin_lock_bh(&css_set_lock);
5816 path = cgroup_path_ns_locked(cgrp, pathbuf, PATH_MAX, &init_cgroup_ns);
5817 spin_unlock_bh(&css_set_lock);
5825 /* minimal command environment */
5827 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5830 mutex_unlock(&cgroup_mutex);
5831 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5834 mutex_unlock(&cgroup_mutex);
5840 static int __init cgroup_disable(char *str)
5842 struct cgroup_subsys *ss;
5846 while ((token = strsep(&str, ",")) != NULL) {
5850 for_each_subsys(ss, i) {
5851 if (strcmp(token, ss->name) &&
5852 strcmp(token, ss->legacy_name))
5854 cgroup_disable_mask |= 1 << i;
5859 __setup("cgroup_disable=", cgroup_disable);
5861 static int __init cgroup_no_v1(char *str)
5863 struct cgroup_subsys *ss;
5867 while ((token = strsep(&str, ",")) != NULL) {
5871 if (!strcmp(token, "all")) {
5872 cgroup_no_v1_mask = ~0UL;
5876 for_each_subsys(ss, i) {
5877 if (strcmp(token, ss->name) &&
5878 strcmp(token, ss->legacy_name))
5881 cgroup_no_v1_mask |= 1 << i;
5886 __setup("cgroup_no_v1=", cgroup_no_v1);
5889 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5890 * @dentry: directory dentry of interest
5891 * @ss: subsystem of interest
5893 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5894 * to get the corresponding css and return it. If such css doesn't exist
5895 * or can't be pinned, an ERR_PTR value is returned.
5897 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5898 struct cgroup_subsys *ss)
5900 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5901 struct cgroup_subsys_state *css = NULL;
5902 struct cgroup *cgrp;
5904 /* is @dentry a cgroup dir? */
5905 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5906 kernfs_type(kn) != KERNFS_DIR)
5907 return ERR_PTR(-EBADF);
5912 * This path doesn't originate from kernfs and @kn could already
5913 * have been or be removed at any point. @kn->priv is RCU
5914 * protected for this access. See css_release_work_fn() for details.
5916 cgrp = rcu_dereference(kn->priv);
5918 css = cgroup_css(cgrp, ss);
5920 if (!css || !css_tryget_online(css))
5921 css = ERR_PTR(-ENOENT);
5928 * css_from_id - lookup css by id
5929 * @id: the cgroup id
5930 * @ss: cgroup subsys to be looked into
5932 * Returns the css if there's valid one with @id, otherwise returns NULL.
5933 * Should be called under rcu_read_lock().
5935 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5937 WARN_ON_ONCE(!rcu_read_lock_held());
5938 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5942 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5943 * @path: path on the default hierarchy
5945 * Find the cgroup at @path on the default hierarchy, increment its
5946 * reference count and return it. Returns pointer to the found cgroup on
5947 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5948 * if @path points to a non-directory.
5950 struct cgroup *cgroup_get_from_path(const char *path)
5952 struct kernfs_node *kn;
5953 struct cgroup *cgrp;
5955 mutex_lock(&cgroup_mutex);
5957 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
5959 if (kernfs_type(kn) == KERNFS_DIR) {
5963 cgrp = ERR_PTR(-ENOTDIR);
5967 cgrp = ERR_PTR(-ENOENT);
5970 mutex_unlock(&cgroup_mutex);
5973 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
5976 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
5977 * definition in cgroup-defs.h.
5979 #ifdef CONFIG_SOCK_CGROUP_DATA
5981 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5983 DEFINE_SPINLOCK(cgroup_sk_update_lock);
5984 static bool cgroup_sk_alloc_disabled __read_mostly;
5986 void cgroup_sk_alloc_disable(void)
5988 if (cgroup_sk_alloc_disabled)
5990 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5991 cgroup_sk_alloc_disabled = true;
5996 #define cgroup_sk_alloc_disabled false
6000 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6002 if (cgroup_sk_alloc_disabled)
6008 struct css_set *cset;
6010 cset = task_css_set(current);
6011 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6012 skcd->val = (unsigned long)cset->dfl_cgrp;
6021 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6023 cgroup_put(sock_cgroup_ptr(skcd));
6026 #endif /* CONFIG_SOCK_CGROUP_DATA */
6028 /* cgroup namespaces */
6030 static struct cgroup_namespace *alloc_cgroup_ns(void)
6032 struct cgroup_namespace *new_ns;
6035 new_ns = kzalloc(sizeof(struct cgroup_namespace), GFP_KERNEL);
6037 return ERR_PTR(-ENOMEM);
6038 ret = ns_alloc_inum(&new_ns->ns);
6041 return ERR_PTR(ret);
6043 atomic_set(&new_ns->count, 1);
6044 new_ns->ns.ops = &cgroupns_operations;
6048 void free_cgroup_ns(struct cgroup_namespace *ns)
6050 put_css_set(ns->root_cset);
6051 put_user_ns(ns->user_ns);
6052 ns_free_inum(&ns->ns);
6055 EXPORT_SYMBOL(free_cgroup_ns);
6057 struct cgroup_namespace *copy_cgroup_ns(unsigned long flags,
6058 struct user_namespace *user_ns,
6059 struct cgroup_namespace *old_ns)
6061 struct cgroup_namespace *new_ns;
6062 struct css_set *cset;
6066 if (!(flags & CLONE_NEWCGROUP)) {
6067 get_cgroup_ns(old_ns);
6071 /* Allow only sysadmin to create cgroup namespace. */
6072 if (!ns_capable(user_ns, CAP_SYS_ADMIN))
6073 return ERR_PTR(-EPERM);
6075 mutex_lock(&cgroup_mutex);
6076 spin_lock_bh(&css_set_lock);
6078 cset = task_css_set(current);
6081 spin_unlock_bh(&css_set_lock);
6082 mutex_unlock(&cgroup_mutex);
6084 new_ns = alloc_cgroup_ns();
6085 if (IS_ERR(new_ns)) {
6090 new_ns->user_ns = get_user_ns(user_ns);
6091 new_ns->root_cset = cset;
6096 static inline struct cgroup_namespace *to_cg_ns(struct ns_common *ns)
6098 return container_of(ns, struct cgroup_namespace, ns);
6101 static int cgroupns_install(struct nsproxy *nsproxy, struct ns_common *ns)
6103 struct cgroup_namespace *cgroup_ns = to_cg_ns(ns);
6105 if (!ns_capable(current_user_ns(), CAP_SYS_ADMIN) ||
6106 !ns_capable(cgroup_ns->user_ns, CAP_SYS_ADMIN))
6109 /* Don't need to do anything if we are attaching to our own cgroupns. */
6110 if (cgroup_ns == nsproxy->cgroup_ns)
6113 get_cgroup_ns(cgroup_ns);
6114 put_cgroup_ns(nsproxy->cgroup_ns);
6115 nsproxy->cgroup_ns = cgroup_ns;
6120 static struct ns_common *cgroupns_get(struct task_struct *task)
6122 struct cgroup_namespace *ns = NULL;
6123 struct nsproxy *nsproxy;
6126 nsproxy = task->nsproxy;
6128 ns = nsproxy->cgroup_ns;
6133 return ns ? &ns->ns : NULL;
6136 static void cgroupns_put(struct ns_common *ns)
6138 put_cgroup_ns(to_cg_ns(ns));
6141 const struct proc_ns_operations cgroupns_operations = {
6143 .type = CLONE_NEWCGROUP,
6144 .get = cgroupns_get,
6145 .put = cgroupns_put,
6146 .install = cgroupns_install,
6149 static __init int cgroup_namespaces_init(void)
6153 subsys_initcall(cgroup_namespaces_init);
6155 #ifdef CONFIG_CGROUP_DEBUG
6156 static struct cgroup_subsys_state *
6157 debug_css_alloc(struct cgroup_subsys_state *parent_css)
6159 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
6162 return ERR_PTR(-ENOMEM);
6167 static void debug_css_free(struct cgroup_subsys_state *css)
6172 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
6175 return cgroup_task_count(css->cgroup);
6178 static u64 current_css_set_read(struct cgroup_subsys_state *css,
6181 return (u64)(unsigned long)current->cgroups;
6184 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
6190 count = atomic_read(&task_css_set(current)->refcount);
6195 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
6197 struct cgrp_cset_link *link;
6198 struct css_set *cset;
6201 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
6205 spin_lock_bh(&css_set_lock);
6207 cset = rcu_dereference(current->cgroups);
6208 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
6209 struct cgroup *c = link->cgrp;
6211 cgroup_name(c, name_buf, NAME_MAX + 1);
6212 seq_printf(seq, "Root %d group %s\n",
6213 c->root->hierarchy_id, name_buf);
6216 spin_unlock_bh(&css_set_lock);
6221 #define MAX_TASKS_SHOWN_PER_CSS 25
6222 static int cgroup_css_links_read(struct seq_file *seq, void *v)
6224 struct cgroup_subsys_state *css = seq_css(seq);
6225 struct cgrp_cset_link *link;
6227 spin_lock_bh(&css_set_lock);
6228 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
6229 struct css_set *cset = link->cset;
6230 struct task_struct *task;
6233 seq_printf(seq, "css_set %p\n", cset);
6235 list_for_each_entry(task, &cset->tasks, cg_list) {
6236 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6238 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6241 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
6242 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6244 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6248 seq_puts(seq, " ...\n");
6250 spin_unlock_bh(&css_set_lock);
6254 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
6256 return (!cgroup_is_populated(css->cgroup) &&
6257 !css_has_online_children(&css->cgroup->self));
6260 static struct cftype debug_files[] = {
6262 .name = "taskcount",
6263 .read_u64 = debug_taskcount_read,
6267 .name = "current_css_set",
6268 .read_u64 = current_css_set_read,
6272 .name = "current_css_set_refcount",
6273 .read_u64 = current_css_set_refcount_read,
6277 .name = "current_css_set_cg_links",
6278 .seq_show = current_css_set_cg_links_read,
6282 .name = "cgroup_css_links",
6283 .seq_show = cgroup_css_links_read,
6287 .name = "releasable",
6288 .read_u64 = releasable_read,
6294 struct cgroup_subsys debug_cgrp_subsys = {
6295 .css_alloc = debug_css_alloc,
6296 .css_free = debug_css_free,
6297 .legacy_cftypes = debug_files,
6299 #endif /* CONFIG_CGROUP_DEBUG */