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 #include <linux/cgroup.h>
30 #include <linux/cred.h>
31 #include <linux/ctype.h>
32 #include <linux/errno.h>
33 #include <linux/init_task.h>
34 #include <linux/kernel.h>
35 #include <linux/list.h>
36 #include <linux/magic.h>
38 #include <linux/mutex.h>
39 #include <linux/mount.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/slab.h>
45 #include <linux/spinlock.h>
46 #include <linux/rwsem.h>
47 #include <linux/string.h>
48 #include <linux/sort.h>
49 #include <linux/kmod.h>
50 #include <linux/delayacct.h>
51 #include <linux/cgroupstats.h>
52 #include <linux/hashtable.h>
53 #include <linux/pid_namespace.h>
54 #include <linux/idr.h>
55 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
56 #include <linux/kthread.h>
57 #include <linux/delay.h>
59 #include <linux/atomic.h>
62 * pidlists linger the following amount before being destroyed. The goal
63 * is avoiding frequent destruction in the middle of consecutive read calls
64 * Expiring in the middle is a performance problem not a correctness one.
65 * 1 sec should be enough.
67 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
69 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
73 * cgroup_tree_mutex nests above cgroup_mutex and protects cftypes, file
74 * creation/removal and hierarchy changing operations including cgroup
75 * creation, removal, css association and controller rebinding. This outer
76 * lock is needed mainly to resolve the circular dependency between kernfs
77 * active ref and cgroup_mutex. cgroup_tree_mutex nests above both.
79 static DEFINE_MUTEX(cgroup_tree_mutex);
82 * cgroup_mutex is the master lock. Any modification to cgroup or its
83 * hierarchy must be performed while holding it.
85 * css_set_rwsem protects task->cgroups pointer, the list of css_set
86 * objects, and the chain of tasks off each css_set.
88 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
89 * cgroup.h can use them for lockdep annotations.
91 #ifdef CONFIG_PROVE_RCU
92 DEFINE_MUTEX(cgroup_mutex);
93 DECLARE_RWSEM(css_set_rwsem);
94 EXPORT_SYMBOL_GPL(cgroup_mutex);
95 EXPORT_SYMBOL_GPL(css_set_rwsem);
97 static DEFINE_MUTEX(cgroup_mutex);
98 static DECLARE_RWSEM(css_set_rwsem);
102 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
103 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
105 static DEFINE_SPINLOCK(release_agent_path_lock);
107 #define cgroup_assert_mutexes_or_rcu_locked() \
108 rcu_lockdep_assert(rcu_read_lock_held() || \
109 lockdep_is_held(&cgroup_tree_mutex) || \
110 lockdep_is_held(&cgroup_mutex), \
111 "cgroup_[tree_]mutex or RCU read lock required");
114 * cgroup destruction makes heavy use of work items and there can be a lot
115 * of concurrent destructions. Use a separate workqueue so that cgroup
116 * destruction work items don't end up filling up max_active of system_wq
117 * which may lead to deadlock.
119 static struct workqueue_struct *cgroup_destroy_wq;
122 * pidlist destructions need to be flushed on cgroup destruction. Use a
123 * separate workqueue as flush domain.
125 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
127 /* generate an array of cgroup subsystem pointers */
128 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
129 static struct cgroup_subsys *cgroup_subsys[] = {
130 #include <linux/cgroup_subsys.h>
134 /* array of cgroup subsystem names */
135 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
136 static const char *cgroup_subsys_name[] = {
137 #include <linux/cgroup_subsys.h>
142 * The default hierarchy, reserved for the subsystems that are otherwise
143 * unattached - it never has more than a single cgroup, and all tasks are
144 * part of that cgroup.
146 struct cgroup_root cgrp_dfl_root;
149 * The default hierarchy always exists but is hidden until mounted for the
150 * first time. This is for backward compatibility.
152 static bool cgrp_dfl_root_visible;
154 /* The list of hierarchy roots */
156 static LIST_HEAD(cgroup_roots);
157 static int cgroup_root_count;
159 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
160 static DEFINE_IDR(cgroup_hierarchy_idr);
163 * Assign a monotonically increasing serial number to cgroups. It
164 * guarantees cgroups with bigger numbers are newer than those with smaller
165 * numbers. Also, as cgroups are always appended to the parent's
166 * ->children list, it guarantees that sibling cgroups are always sorted in
167 * the ascending serial number order on the list. Protected by
170 static u64 cgroup_serial_nr_next = 1;
172 /* This flag indicates whether tasks in the fork and exit paths should
173 * check for fork/exit handlers to call. This avoids us having to do
174 * extra work in the fork/exit path if none of the subsystems need to
177 static int need_forkexit_callback __read_mostly;
179 static struct cftype cgroup_base_files[];
181 static void cgroup_put(struct cgroup *cgrp);
182 static int rebind_subsystems(struct cgroup_root *dst_root,
183 unsigned long ss_mask);
184 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
185 static int cgroup_destroy_locked(struct cgroup *cgrp);
186 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
188 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
191 * cgroup_css - obtain a cgroup's css for the specified subsystem
192 * @cgrp: the cgroup of interest
193 * @ss: the subsystem of interest (%NULL returns the dummy_css)
195 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
196 * function must be called either under cgroup_mutex or rcu_read_lock() and
197 * the caller is responsible for pinning the returned css if it wants to
198 * keep accessing it outside the said locks. This function may return
199 * %NULL if @cgrp doesn't have @subsys_id enabled.
201 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
202 struct cgroup_subsys *ss)
205 return rcu_dereference_check(cgrp->subsys[ss->id],
206 lockdep_is_held(&cgroup_tree_mutex) ||
207 lockdep_is_held(&cgroup_mutex));
209 return &cgrp->dummy_css;
212 /* convenient tests for these bits */
213 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
215 return test_bit(CGRP_DEAD, &cgrp->flags);
218 struct cgroup_subsys_state *seq_css(struct seq_file *seq)
220 struct kernfs_open_file *of = seq->private;
221 struct cgroup *cgrp = of->kn->parent->priv;
222 struct cftype *cft = seq_cft(seq);
225 * This is open and unprotected implementation of cgroup_css().
226 * seq_css() is only called from a kernfs file operation which has
227 * an active reference on the file. Because all the subsystem
228 * files are drained before a css is disassociated with a cgroup,
229 * the matching css from the cgroup's subsys table is guaranteed to
230 * be and stay valid until the enclosing operation is complete.
233 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
235 return &cgrp->dummy_css;
237 EXPORT_SYMBOL_GPL(seq_css);
240 * cgroup_is_descendant - test ancestry
241 * @cgrp: the cgroup to be tested
242 * @ancestor: possible ancestor of @cgrp
244 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
245 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
246 * and @ancestor are accessible.
248 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
251 if (cgrp == ancestor)
258 static int cgroup_is_releasable(const struct cgroup *cgrp)
261 (1 << CGRP_RELEASABLE) |
262 (1 << CGRP_NOTIFY_ON_RELEASE);
263 return (cgrp->flags & bits) == bits;
266 static int notify_on_release(const struct cgroup *cgrp)
268 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
272 * for_each_css - iterate all css's of a cgroup
273 * @css: the iteration cursor
274 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
275 * @cgrp: the target cgroup to iterate css's of
277 * Should be called under cgroup_mutex.
279 #define for_each_css(css, ssid, cgrp) \
280 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
281 if (!((css) = rcu_dereference_check( \
282 (cgrp)->subsys[(ssid)], \
283 lockdep_is_held(&cgroup_tree_mutex) || \
284 lockdep_is_held(&cgroup_mutex)))) { } \
288 * for_each_subsys - iterate all enabled cgroup subsystems
289 * @ss: the iteration cursor
290 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
292 #define for_each_subsys(ss, ssid) \
293 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
294 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
296 /* iterate across the hierarchies */
297 #define for_each_root(root) \
298 list_for_each_entry((root), &cgroup_roots, root_list)
301 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
302 * @cgrp: the cgroup to be checked for liveness
304 * On success, returns true; the mutex should be later unlocked. On
305 * failure returns false with no lock held.
307 static bool cgroup_lock_live_group(struct cgroup *cgrp)
309 mutex_lock(&cgroup_mutex);
310 if (cgroup_is_dead(cgrp)) {
311 mutex_unlock(&cgroup_mutex);
317 /* the list of cgroups eligible for automatic release. Protected by
318 * release_list_lock */
319 static LIST_HEAD(release_list);
320 static DEFINE_RAW_SPINLOCK(release_list_lock);
321 static void cgroup_release_agent(struct work_struct *work);
322 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
323 static void check_for_release(struct cgroup *cgrp);
326 * A cgroup can be associated with multiple css_sets as different tasks may
327 * belong to different cgroups on different hierarchies. In the other
328 * direction, a css_set is naturally associated with multiple cgroups.
329 * This M:N relationship is represented by the following link structure
330 * which exists for each association and allows traversing the associations
333 struct cgrp_cset_link {
334 /* the cgroup and css_set this link associates */
336 struct css_set *cset;
338 /* list of cgrp_cset_links anchored at cgrp->cset_links */
339 struct list_head cset_link;
341 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
342 struct list_head cgrp_link;
346 * The default css_set - used by init and its children prior to any
347 * hierarchies being mounted. It contains a pointer to the root state
348 * for each subsystem. Also used to anchor the list of css_sets. Not
349 * reference-counted, to improve performance when child cgroups
350 * haven't been created.
352 struct css_set init_css_set = {
353 .refcount = ATOMIC_INIT(1),
354 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
355 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
356 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
357 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
358 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
361 static int css_set_count = 1; /* 1 for init_css_set */
364 * hash table for cgroup groups. This improves the performance to find
365 * an existing css_set. This hash doesn't (currently) take into
366 * account cgroups in empty hierarchies.
368 #define CSS_SET_HASH_BITS 7
369 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
371 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
373 unsigned long key = 0UL;
374 struct cgroup_subsys *ss;
377 for_each_subsys(ss, i)
378 key += (unsigned long)css[i];
379 key = (key >> 16) ^ key;
384 static void put_css_set_locked(struct css_set *cset, bool taskexit)
386 struct cgrp_cset_link *link, *tmp_link;
388 lockdep_assert_held(&css_set_rwsem);
390 if (!atomic_dec_and_test(&cset->refcount))
393 /* This css_set is dead. unlink it and release cgroup refcounts */
394 hash_del(&cset->hlist);
397 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
398 struct cgroup *cgrp = link->cgrp;
400 list_del(&link->cset_link);
401 list_del(&link->cgrp_link);
403 /* @cgrp can't go away while we're holding css_set_rwsem */
404 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
406 set_bit(CGRP_RELEASABLE, &cgrp->flags);
407 check_for_release(cgrp);
413 kfree_rcu(cset, rcu_head);
416 static void put_css_set(struct css_set *cset, bool taskexit)
419 * Ensure that the refcount doesn't hit zero while any readers
420 * can see it. Similar to atomic_dec_and_lock(), but for an
423 if (atomic_add_unless(&cset->refcount, -1, 1))
426 down_write(&css_set_rwsem);
427 put_css_set_locked(cset, taskexit);
428 up_write(&css_set_rwsem);
432 * refcounted get/put for css_set objects
434 static inline void get_css_set(struct css_set *cset)
436 atomic_inc(&cset->refcount);
440 * compare_css_sets - helper function for find_existing_css_set().
441 * @cset: candidate css_set being tested
442 * @old_cset: existing css_set for a task
443 * @new_cgrp: cgroup that's being entered by the task
444 * @template: desired set of css pointers in css_set (pre-calculated)
446 * Returns true if "cset" matches "old_cset" except for the hierarchy
447 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
449 static bool compare_css_sets(struct css_set *cset,
450 struct css_set *old_cset,
451 struct cgroup *new_cgrp,
452 struct cgroup_subsys_state *template[])
454 struct list_head *l1, *l2;
456 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
457 /* Not all subsystems matched */
462 * Compare cgroup pointers in order to distinguish between
463 * different cgroups in heirarchies with no subsystems. We
464 * could get by with just this check alone (and skip the
465 * memcmp above) but on most setups the memcmp check will
466 * avoid the need for this more expensive check on almost all
470 l1 = &cset->cgrp_links;
471 l2 = &old_cset->cgrp_links;
473 struct cgrp_cset_link *link1, *link2;
474 struct cgroup *cgrp1, *cgrp2;
478 /* See if we reached the end - both lists are equal length. */
479 if (l1 == &cset->cgrp_links) {
480 BUG_ON(l2 != &old_cset->cgrp_links);
483 BUG_ON(l2 == &old_cset->cgrp_links);
485 /* Locate the cgroups associated with these links. */
486 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
487 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
490 /* Hierarchies should be linked in the same order. */
491 BUG_ON(cgrp1->root != cgrp2->root);
494 * If this hierarchy is the hierarchy of the cgroup
495 * that's changing, then we need to check that this
496 * css_set points to the new cgroup; if it's any other
497 * hierarchy, then this css_set should point to the
498 * same cgroup as the old css_set.
500 if (cgrp1->root == new_cgrp->root) {
501 if (cgrp1 != new_cgrp)
512 * find_existing_css_set - init css array and find the matching css_set
513 * @old_cset: the css_set that we're using before the cgroup transition
514 * @cgrp: the cgroup that we're moving into
515 * @template: out param for the new set of csses, should be clear on entry
517 static struct css_set *find_existing_css_set(struct css_set *old_cset,
519 struct cgroup_subsys_state *template[])
521 struct cgroup_root *root = cgrp->root;
522 struct cgroup_subsys *ss;
523 struct css_set *cset;
528 * Build the set of subsystem state objects that we want to see in the
529 * new css_set. while subsystems can change globally, the entries here
530 * won't change, so no need for locking.
532 for_each_subsys(ss, i) {
533 if (root->cgrp.subsys_mask & (1UL << i)) {
534 /* Subsystem is in this hierarchy. So we want
535 * the subsystem state from the new
537 template[i] = cgroup_css(cgrp, ss);
539 /* Subsystem is not in this hierarchy, so we
540 * don't want to change the subsystem state */
541 template[i] = old_cset->subsys[i];
545 key = css_set_hash(template);
546 hash_for_each_possible(css_set_table, cset, hlist, key) {
547 if (!compare_css_sets(cset, old_cset, cgrp, template))
550 /* This css_set matches what we need */
554 /* No existing cgroup group matched */
558 static void free_cgrp_cset_links(struct list_head *links_to_free)
560 struct cgrp_cset_link *link, *tmp_link;
562 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
563 list_del(&link->cset_link);
569 * allocate_cgrp_cset_links - allocate cgrp_cset_links
570 * @count: the number of links to allocate
571 * @tmp_links: list_head the allocated links are put on
573 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
574 * through ->cset_link. Returns 0 on success or -errno.
576 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
578 struct cgrp_cset_link *link;
581 INIT_LIST_HEAD(tmp_links);
583 for (i = 0; i < count; i++) {
584 link = kzalloc(sizeof(*link), GFP_KERNEL);
586 free_cgrp_cset_links(tmp_links);
589 list_add(&link->cset_link, tmp_links);
595 * link_css_set - a helper function to link a css_set to a cgroup
596 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
597 * @cset: the css_set to be linked
598 * @cgrp: the destination cgroup
600 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
603 struct cgrp_cset_link *link;
605 BUG_ON(list_empty(tmp_links));
606 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
609 list_move(&link->cset_link, &cgrp->cset_links);
611 * Always add links to the tail of the list so that the list
612 * is sorted by order of hierarchy creation
614 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
618 * find_css_set - return a new css_set with one cgroup updated
619 * @old_cset: the baseline css_set
620 * @cgrp: the cgroup to be updated
622 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
623 * substituted into the appropriate hierarchy.
625 static struct css_set *find_css_set(struct css_set *old_cset,
628 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
629 struct css_set *cset;
630 struct list_head tmp_links;
631 struct cgrp_cset_link *link;
634 lockdep_assert_held(&cgroup_mutex);
636 /* First see if we already have a cgroup group that matches
638 down_read(&css_set_rwsem);
639 cset = find_existing_css_set(old_cset, cgrp, template);
642 up_read(&css_set_rwsem);
647 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
651 /* Allocate all the cgrp_cset_link objects that we'll need */
652 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
657 atomic_set(&cset->refcount, 1);
658 INIT_LIST_HEAD(&cset->cgrp_links);
659 INIT_LIST_HEAD(&cset->tasks);
660 INIT_LIST_HEAD(&cset->mg_tasks);
661 INIT_LIST_HEAD(&cset->mg_preload_node);
662 INIT_LIST_HEAD(&cset->mg_node);
663 INIT_HLIST_NODE(&cset->hlist);
665 /* Copy the set of subsystem state objects generated in
666 * find_existing_css_set() */
667 memcpy(cset->subsys, template, sizeof(cset->subsys));
669 down_write(&css_set_rwsem);
670 /* Add reference counts and links from the new css_set. */
671 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
672 struct cgroup *c = link->cgrp;
674 if (c->root == cgrp->root)
676 link_css_set(&tmp_links, cset, c);
679 BUG_ON(!list_empty(&tmp_links));
683 /* Add this cgroup group to the hash table */
684 key = css_set_hash(cset->subsys);
685 hash_add(css_set_table, &cset->hlist, key);
687 up_write(&css_set_rwsem);
692 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
694 struct cgroup *root_cgrp = kf_root->kn->priv;
696 return root_cgrp->root;
699 static int cgroup_init_root_id(struct cgroup_root *root)
703 lockdep_assert_held(&cgroup_mutex);
705 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
709 root->hierarchy_id = id;
713 static void cgroup_exit_root_id(struct cgroup_root *root)
715 lockdep_assert_held(&cgroup_mutex);
717 if (root->hierarchy_id) {
718 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
719 root->hierarchy_id = 0;
723 static void cgroup_free_root(struct cgroup_root *root)
726 /* hierarhcy ID shoulid already have been released */
727 WARN_ON_ONCE(root->hierarchy_id);
729 idr_destroy(&root->cgroup_idr);
734 static void cgroup_destroy_root(struct cgroup_root *root)
736 struct cgroup *cgrp = &root->cgrp;
737 struct cgrp_cset_link *link, *tmp_link;
739 mutex_lock(&cgroup_tree_mutex);
740 mutex_lock(&cgroup_mutex);
742 BUG_ON(atomic_read(&root->nr_cgrps));
743 BUG_ON(!list_empty(&cgrp->children));
745 /* Rebind all subsystems back to the default hierarchy */
746 rebind_subsystems(&cgrp_dfl_root, cgrp->subsys_mask);
749 * Release all the links from cset_links to this hierarchy's
752 down_write(&css_set_rwsem);
754 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
755 list_del(&link->cset_link);
756 list_del(&link->cgrp_link);
759 up_write(&css_set_rwsem);
761 if (!list_empty(&root->root_list)) {
762 list_del(&root->root_list);
766 cgroup_exit_root_id(root);
768 mutex_unlock(&cgroup_mutex);
769 mutex_unlock(&cgroup_tree_mutex);
771 kernfs_destroy_root(root->kf_root);
772 cgroup_free_root(root);
775 /* look up cgroup associated with given css_set on the specified hierarchy */
776 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
777 struct cgroup_root *root)
779 struct cgroup *res = NULL;
781 lockdep_assert_held(&cgroup_mutex);
782 lockdep_assert_held(&css_set_rwsem);
784 if (cset == &init_css_set) {
787 struct cgrp_cset_link *link;
789 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
790 struct cgroup *c = link->cgrp;
792 if (c->root == root) {
804 * Return the cgroup for "task" from the given hierarchy. Must be
805 * called with cgroup_mutex and css_set_rwsem held.
807 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
808 struct cgroup_root *root)
811 * No need to lock the task - since we hold cgroup_mutex the
812 * task can't change groups, so the only thing that can happen
813 * is that it exits and its css is set back to init_css_set.
815 return cset_cgroup_from_root(task_css_set(task), root);
819 * A task must hold cgroup_mutex to modify cgroups.
821 * Any task can increment and decrement the count field without lock.
822 * So in general, code holding cgroup_mutex can't rely on the count
823 * field not changing. However, if the count goes to zero, then only
824 * cgroup_attach_task() can increment it again. Because a count of zero
825 * means that no tasks are currently attached, therefore there is no
826 * way a task attached to that cgroup can fork (the other way to
827 * increment the count). So code holding cgroup_mutex can safely
828 * assume that if the count is zero, it will stay zero. Similarly, if
829 * a task holds cgroup_mutex on a cgroup with zero count, it
830 * knows that the cgroup won't be removed, as cgroup_rmdir()
833 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
834 * (usually) take cgroup_mutex. These are the two most performance
835 * critical pieces of code here. The exception occurs on cgroup_exit(),
836 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
837 * is taken, and if the cgroup count is zero, a usermode call made
838 * to the release agent with the name of the cgroup (path relative to
839 * the root of cgroup file system) as the argument.
841 * A cgroup can only be deleted if both its 'count' of using tasks
842 * is zero, and its list of 'children' cgroups is empty. Since all
843 * tasks in the system use _some_ cgroup, and since there is always at
844 * least one task in the system (init, pid == 1), therefore, root cgroup
845 * always has either children cgroups and/or using tasks. So we don't
846 * need a special hack to ensure that root cgroup cannot be deleted.
848 * P.S. One more locking exception. RCU is used to guard the
849 * update of a tasks cgroup pointer by cgroup_attach_task()
852 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
853 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
854 static const struct file_operations proc_cgroupstats_operations;
856 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
859 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
860 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
861 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
862 cft->ss->name, cft->name);
864 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
869 * cgroup_file_mode - deduce file mode of a control file
870 * @cft: the control file in question
872 * returns cft->mode if ->mode is not 0
873 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
874 * returns S_IRUGO if it has only a read handler
875 * returns S_IWUSR if it has only a write hander
877 static umode_t cgroup_file_mode(const struct cftype *cft)
884 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
887 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
894 static void cgroup_free_fn(struct work_struct *work)
896 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
898 atomic_dec(&cgrp->root->nr_cgrps);
899 cgroup_pidlist_destroy_all(cgrp);
903 * We get a ref to the parent, and put the ref when this
904 * cgroup is being freed, so it's guaranteed that the
905 * parent won't be destroyed before its children.
907 cgroup_put(cgrp->parent);
908 kernfs_put(cgrp->kn);
912 * This is root cgroup's refcnt reaching zero, which
913 * indicates that the root should be released.
915 cgroup_destroy_root(cgrp->root);
919 static void cgroup_free_rcu(struct rcu_head *head)
921 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
923 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
924 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
927 static void cgroup_get(struct cgroup *cgrp)
929 WARN_ON_ONCE(cgroup_is_dead(cgrp));
930 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
931 atomic_inc(&cgrp->refcnt);
934 static void cgroup_put(struct cgroup *cgrp)
936 if (!atomic_dec_and_test(&cgrp->refcnt))
938 if (WARN_ON_ONCE(cgrp->parent && !cgroup_is_dead(cgrp)))
942 * XXX: cgrp->id is only used to look up css's. As cgroup and
943 * css's lifetimes will be decoupled, it should be made
944 * per-subsystem and moved to css->id so that lookups are
945 * successful until the target css is released.
947 mutex_lock(&cgroup_mutex);
948 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
949 mutex_unlock(&cgroup_mutex);
952 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
955 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
957 char name[CGROUP_FILE_NAME_MAX];
959 lockdep_assert_held(&cgroup_tree_mutex);
960 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
964 * cgroup_clear_dir - remove subsys files in a cgroup directory
965 * @cgrp: target cgroup
966 * @subsys_mask: mask of the subsystem ids whose files should be removed
968 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
970 struct cgroup_subsys *ss;
973 for_each_subsys(ss, i) {
976 if (!test_bit(i, &subsys_mask))
978 list_for_each_entry(cfts, &ss->cfts, node)
979 cgroup_addrm_files(cgrp, cfts, false);
983 static int rebind_subsystems(struct cgroup_root *dst_root,
984 unsigned long ss_mask)
986 struct cgroup_subsys *ss;
989 lockdep_assert_held(&cgroup_tree_mutex);
990 lockdep_assert_held(&cgroup_mutex);
992 for_each_subsys(ss, ssid) {
993 if (!(ss_mask & (1 << ssid)))
996 /* if @ss is on the dummy_root, we can always move it */
997 if (ss->root == &cgrp_dfl_root)
1000 /* if @ss has non-root cgroups attached to it, can't move */
1001 if (!list_empty(&ss->root->cgrp.children))
1004 /* can't move between two non-dummy roots either */
1005 if (dst_root != &cgrp_dfl_root)
1009 ret = cgroup_populate_dir(&dst_root->cgrp, ss_mask);
1011 if (dst_root != &cgrp_dfl_root)
1015 * Rebinding back to the default root is not allowed to
1016 * fail. Using both default and non-default roots should
1017 * be rare. Moving subsystems back and forth even more so.
1018 * Just warn about it and continue.
1020 if (cgrp_dfl_root_visible) {
1021 pr_warning("cgroup: failed to create files (%d) while rebinding 0x%lx to default root\n",
1023 pr_warning("cgroup: you may retry by moving them to a different hierarchy and unbinding\n");
1028 * Nothing can fail from this point on. Remove files for the
1029 * removed subsystems and rebind each subsystem.
1031 mutex_unlock(&cgroup_mutex);
1032 for_each_subsys(ss, ssid)
1033 if (ss_mask & (1 << ssid))
1034 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1035 mutex_lock(&cgroup_mutex);
1037 for_each_subsys(ss, ssid) {
1038 struct cgroup_root *src_root;
1039 struct cgroup_subsys_state *css;
1041 if (!(ss_mask & (1 << ssid)))
1044 src_root = ss->root;
1045 css = cgroup_css(&src_root->cgrp, ss);
1047 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1049 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1050 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1051 ss->root = dst_root;
1052 css->cgroup = &dst_root->cgrp;
1054 src_root->cgrp.subsys_mask &= ~(1 << ssid);
1055 dst_root->cgrp.subsys_mask |= 1 << ssid;
1061 kernfs_activate(dst_root->cgrp.kn);
1065 static int cgroup_show_options(struct seq_file *seq,
1066 struct kernfs_root *kf_root)
1068 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1069 struct cgroup_subsys *ss;
1072 for_each_subsys(ss, ssid)
1073 if (root->cgrp.subsys_mask & (1 << ssid))
1074 seq_printf(seq, ",%s", ss->name);
1075 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1076 seq_puts(seq, ",sane_behavior");
1077 if (root->flags & CGRP_ROOT_NOPREFIX)
1078 seq_puts(seq, ",noprefix");
1079 if (root->flags & CGRP_ROOT_XATTR)
1080 seq_puts(seq, ",xattr");
1082 spin_lock(&release_agent_path_lock);
1083 if (strlen(root->release_agent_path))
1084 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1085 spin_unlock(&release_agent_path_lock);
1087 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1088 seq_puts(seq, ",clone_children");
1089 if (strlen(root->name))
1090 seq_printf(seq, ",name=%s", root->name);
1094 struct cgroup_sb_opts {
1095 unsigned long subsys_mask;
1096 unsigned long flags;
1097 char *release_agent;
1098 bool cpuset_clone_children;
1100 /* User explicitly requested empty subsystem */
1105 * Convert a hierarchy specifier into a bitmask of subsystems and
1106 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1107 * array. This function takes refcounts on subsystems to be used, unless it
1108 * returns error, in which case no refcounts are taken.
1110 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1112 char *token, *o = data;
1113 bool all_ss = false, one_ss = false;
1114 unsigned long mask = (unsigned long)-1;
1115 struct cgroup_subsys *ss;
1118 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1120 #ifdef CONFIG_CPUSETS
1121 mask = ~(1UL << cpuset_cgrp_id);
1124 memset(opts, 0, sizeof(*opts));
1126 while ((token = strsep(&o, ",")) != NULL) {
1129 if (!strcmp(token, "none")) {
1130 /* Explicitly have no subsystems */
1134 if (!strcmp(token, "all")) {
1135 /* Mutually exclusive option 'all' + subsystem name */
1141 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1142 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1145 if (!strcmp(token, "noprefix")) {
1146 opts->flags |= CGRP_ROOT_NOPREFIX;
1149 if (!strcmp(token, "clone_children")) {
1150 opts->cpuset_clone_children = true;
1153 if (!strcmp(token, "xattr")) {
1154 opts->flags |= CGRP_ROOT_XATTR;
1157 if (!strncmp(token, "release_agent=", 14)) {
1158 /* Specifying two release agents is forbidden */
1159 if (opts->release_agent)
1161 opts->release_agent =
1162 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1163 if (!opts->release_agent)
1167 if (!strncmp(token, "name=", 5)) {
1168 const char *name = token + 5;
1169 /* Can't specify an empty name */
1172 /* Must match [\w.-]+ */
1173 for (i = 0; i < strlen(name); i++) {
1177 if ((c == '.') || (c == '-') || (c == '_'))
1181 /* Specifying two names is forbidden */
1184 opts->name = kstrndup(name,
1185 MAX_CGROUP_ROOT_NAMELEN - 1,
1193 for_each_subsys(ss, i) {
1194 if (strcmp(token, ss->name))
1199 /* Mutually exclusive option 'all' + subsystem name */
1202 set_bit(i, &opts->subsys_mask);
1207 if (i == CGROUP_SUBSYS_COUNT)
1211 /* Consistency checks */
1213 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1214 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1216 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1217 opts->cpuset_clone_children || opts->release_agent ||
1219 pr_err("cgroup: sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1224 * If the 'all' option was specified select all the
1225 * subsystems, otherwise if 'none', 'name=' and a subsystem
1226 * name options were not specified, let's default to 'all'
1228 if (all_ss || (!one_ss && !opts->none && !opts->name))
1229 for_each_subsys(ss, i)
1231 set_bit(i, &opts->subsys_mask);
1234 * We either have to specify by name or by subsystems. (So
1235 * all empty hierarchies must have a name).
1237 if (!opts->subsys_mask && !opts->name)
1242 * Option noprefix was introduced just for backward compatibility
1243 * with the old cpuset, so we allow noprefix only if mounting just
1244 * the cpuset subsystem.
1246 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1250 /* Can't specify "none" and some subsystems */
1251 if (opts->subsys_mask && opts->none)
1257 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1260 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1261 struct cgroup_sb_opts opts;
1262 unsigned long added_mask, removed_mask;
1264 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1265 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1269 mutex_lock(&cgroup_tree_mutex);
1270 mutex_lock(&cgroup_mutex);
1272 /* See what subsystems are wanted */
1273 ret = parse_cgroupfs_options(data, &opts);
1277 if (opts.subsys_mask != root->cgrp.subsys_mask || opts.release_agent)
1278 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1279 task_tgid_nr(current), current->comm);
1281 added_mask = opts.subsys_mask & ~root->cgrp.subsys_mask;
1282 removed_mask = root->cgrp.subsys_mask & ~opts.subsys_mask;
1284 /* Don't allow flags or name to change at remount */
1285 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1286 (opts.name && strcmp(opts.name, root->name))) {
1287 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1288 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1289 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1294 /* remounting is not allowed for populated hierarchies */
1295 if (!list_empty(&root->cgrp.children)) {
1300 ret = rebind_subsystems(root, added_mask);
1304 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1306 if (opts.release_agent) {
1307 spin_lock(&release_agent_path_lock);
1308 strcpy(root->release_agent_path, opts.release_agent);
1309 spin_unlock(&release_agent_path_lock);
1312 kfree(opts.release_agent);
1314 mutex_unlock(&cgroup_mutex);
1315 mutex_unlock(&cgroup_tree_mutex);
1320 * To reduce the fork() overhead for systems that are not actually using
1321 * their cgroups capability, we don't maintain the lists running through
1322 * each css_set to its tasks until we see the list actually used - in other
1323 * words after the first mount.
1325 static bool use_task_css_set_links __read_mostly;
1327 static void cgroup_enable_task_cg_lists(void)
1329 struct task_struct *p, *g;
1331 down_write(&css_set_rwsem);
1333 if (use_task_css_set_links)
1336 use_task_css_set_links = true;
1339 * We need tasklist_lock because RCU is not safe against
1340 * while_each_thread(). Besides, a forking task that has passed
1341 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1342 * is not guaranteed to have its child immediately visible in the
1343 * tasklist if we walk through it with RCU.
1345 read_lock(&tasklist_lock);
1346 do_each_thread(g, p) {
1347 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1348 task_css_set(p) != &init_css_set);
1351 * We should check if the process is exiting, otherwise
1352 * it will race with cgroup_exit() in that the list
1353 * entry won't be deleted though the process has exited.
1354 * Do it while holding siglock so that we don't end up
1355 * racing against cgroup_exit().
1357 spin_lock_irq(&p->sighand->siglock);
1358 if (!(p->flags & PF_EXITING)) {
1359 struct css_set *cset = task_css_set(p);
1361 list_add(&p->cg_list, &cset->tasks);
1364 spin_unlock_irq(&p->sighand->siglock);
1365 } while_each_thread(g, p);
1366 read_unlock(&tasklist_lock);
1368 up_write(&css_set_rwsem);
1371 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1373 atomic_set(&cgrp->refcnt, 1);
1374 INIT_LIST_HEAD(&cgrp->sibling);
1375 INIT_LIST_HEAD(&cgrp->children);
1376 INIT_LIST_HEAD(&cgrp->cset_links);
1377 INIT_LIST_HEAD(&cgrp->release_list);
1378 INIT_LIST_HEAD(&cgrp->pidlists);
1379 mutex_init(&cgrp->pidlist_mutex);
1380 cgrp->dummy_css.cgroup = cgrp;
1383 static void init_cgroup_root(struct cgroup_root *root,
1384 struct cgroup_sb_opts *opts)
1386 struct cgroup *cgrp = &root->cgrp;
1388 INIT_LIST_HEAD(&root->root_list);
1389 atomic_set(&root->nr_cgrps, 1);
1391 init_cgroup_housekeeping(cgrp);
1392 idr_init(&root->cgroup_idr);
1394 root->flags = opts->flags;
1395 if (opts->release_agent)
1396 strcpy(root->release_agent_path, opts->release_agent);
1398 strcpy(root->name, opts->name);
1399 if (opts->cpuset_clone_children)
1400 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1403 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1405 LIST_HEAD(tmp_links);
1406 struct cgroup *root_cgrp = &root->cgrp;
1407 struct css_set *cset;
1410 lockdep_assert_held(&cgroup_tree_mutex);
1411 lockdep_assert_held(&cgroup_mutex);
1413 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1416 root_cgrp->id = ret;
1419 * We're accessing css_set_count without locking css_set_rwsem here,
1420 * but that's OK - it can only be increased by someone holding
1421 * cgroup_lock, and that's us. The worst that can happen is that we
1422 * have some link structures left over
1424 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1428 ret = cgroup_init_root_id(root);
1432 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1433 KERNFS_ROOT_CREATE_DEACTIVATED,
1435 if (IS_ERR(root->kf_root)) {
1436 ret = PTR_ERR(root->kf_root);
1439 root_cgrp->kn = root->kf_root->kn;
1441 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1445 ret = rebind_subsystems(root, ss_mask);
1450 * There must be no failure case after here, since rebinding takes
1451 * care of subsystems' refcounts, which are explicitly dropped in
1452 * the failure exit path.
1454 list_add(&root->root_list, &cgroup_roots);
1455 cgroup_root_count++;
1458 * Link the root cgroup in this hierarchy into all the css_set
1461 down_write(&css_set_rwsem);
1462 hash_for_each(css_set_table, i, cset, hlist)
1463 link_css_set(&tmp_links, cset, root_cgrp);
1464 up_write(&css_set_rwsem);
1466 BUG_ON(!list_empty(&root_cgrp->children));
1467 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1469 kernfs_activate(root_cgrp->kn);
1474 kernfs_destroy_root(root->kf_root);
1475 root->kf_root = NULL;
1477 cgroup_exit_root_id(root);
1479 free_cgrp_cset_links(&tmp_links);
1483 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1484 int flags, const char *unused_dev_name,
1487 struct cgroup_root *root;
1488 struct cgroup_sb_opts opts;
1489 struct dentry *dentry;
1494 * The first time anyone tries to mount a cgroup, enable the list
1495 * linking each css_set to its tasks and fix up all existing tasks.
1497 if (!use_task_css_set_links)
1498 cgroup_enable_task_cg_lists();
1500 mutex_lock(&cgroup_tree_mutex);
1501 mutex_lock(&cgroup_mutex);
1503 /* First find the desired set of subsystems */
1504 ret = parse_cgroupfs_options(data, &opts);
1508 /* look for a matching existing root */
1509 if (!opts.subsys_mask && !opts.none && !opts.name) {
1510 cgrp_dfl_root_visible = true;
1511 root = &cgrp_dfl_root;
1512 cgroup_get(&root->cgrp);
1517 for_each_root(root) {
1518 bool name_match = false;
1520 if (root == &cgrp_dfl_root)
1524 * If we asked for a name then it must match. Also, if
1525 * name matches but sybsys_mask doesn't, we should fail.
1526 * Remember whether name matched.
1529 if (strcmp(opts.name, root->name))
1535 * If we asked for subsystems (or explicitly for no
1536 * subsystems) then they must match.
1538 if ((opts.subsys_mask || opts.none) &&
1539 (opts.subsys_mask != root->cgrp.subsys_mask)) {
1546 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1547 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1548 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1552 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1557 * A root's lifetime is governed by its root cgroup. Zero
1558 * ref indicate that the root is being destroyed. Wait for
1559 * destruction to complete so that the subsystems are free.
1560 * We can use wait_queue for the wait but this path is
1561 * super cold. Let's just sleep for a bit and retry.
1563 if (!atomic_inc_not_zero(&root->cgrp.refcnt)) {
1564 mutex_unlock(&cgroup_mutex);
1565 mutex_unlock(&cgroup_tree_mutex);
1567 mutex_lock(&cgroup_tree_mutex);
1568 mutex_lock(&cgroup_mutex);
1577 * No such thing, create a new one. name= matching without subsys
1578 * specification is allowed for already existing hierarchies but we
1579 * can't create new one without subsys specification.
1581 if (!opts.subsys_mask && !opts.none) {
1586 root = kzalloc(sizeof(*root), GFP_KERNEL);
1592 init_cgroup_root(root, &opts);
1594 ret = cgroup_setup_root(root, opts.subsys_mask);
1596 cgroup_free_root(root);
1599 mutex_unlock(&cgroup_mutex);
1600 mutex_unlock(&cgroup_tree_mutex);
1602 kfree(opts.release_agent);
1606 return ERR_PTR(ret);
1608 dentry = kernfs_mount(fs_type, flags, root->kf_root,
1609 CGROUP_SUPER_MAGIC, &new_sb);
1610 if (IS_ERR(dentry) || !new_sb)
1611 cgroup_put(&root->cgrp);
1615 static void cgroup_kill_sb(struct super_block *sb)
1617 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1618 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1620 cgroup_put(&root->cgrp);
1624 static struct file_system_type cgroup_fs_type = {
1626 .mount = cgroup_mount,
1627 .kill_sb = cgroup_kill_sb,
1630 static struct kobject *cgroup_kobj;
1633 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1634 * @task: target task
1635 * @buf: the buffer to write the path into
1636 * @buflen: the length of the buffer
1638 * Determine @task's cgroup on the first (the one with the lowest non-zero
1639 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1640 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1641 * cgroup controller callbacks.
1643 * Return value is the same as kernfs_path().
1645 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1647 struct cgroup_root *root;
1648 struct cgroup *cgrp;
1649 int hierarchy_id = 1;
1652 mutex_lock(&cgroup_mutex);
1653 down_read(&css_set_rwsem);
1655 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1658 cgrp = task_cgroup_from_root(task, root);
1659 path = cgroup_path(cgrp, buf, buflen);
1661 /* if no hierarchy exists, everyone is in "/" */
1662 if (strlcpy(buf, "/", buflen) < buflen)
1666 up_read(&css_set_rwsem);
1667 mutex_unlock(&cgroup_mutex);
1670 EXPORT_SYMBOL_GPL(task_cgroup_path);
1672 /* used to track tasks and other necessary states during migration */
1673 struct cgroup_taskset {
1674 /* the src and dst cset list running through cset->mg_node */
1675 struct list_head src_csets;
1676 struct list_head dst_csets;
1679 * Fields for cgroup_taskset_*() iteration.
1681 * Before migration is committed, the target migration tasks are on
1682 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1683 * the csets on ->dst_csets. ->csets point to either ->src_csets
1684 * or ->dst_csets depending on whether migration is committed.
1686 * ->cur_csets and ->cur_task point to the current task position
1689 struct list_head *csets;
1690 struct css_set *cur_cset;
1691 struct task_struct *cur_task;
1695 * cgroup_taskset_first - reset taskset and return the first task
1696 * @tset: taskset of interest
1698 * @tset iteration is initialized and the first task is returned.
1700 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1702 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
1703 tset->cur_task = NULL;
1705 return cgroup_taskset_next(tset);
1709 * cgroup_taskset_next - iterate to the next task in taskset
1710 * @tset: taskset of interest
1712 * Return the next task in @tset. Iteration must have been initialized
1713 * with cgroup_taskset_first().
1715 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1717 struct css_set *cset = tset->cur_cset;
1718 struct task_struct *task = tset->cur_task;
1720 while (&cset->mg_node != tset->csets) {
1722 task = list_first_entry(&cset->mg_tasks,
1723 struct task_struct, cg_list);
1725 task = list_next_entry(task, cg_list);
1727 if (&task->cg_list != &cset->mg_tasks) {
1728 tset->cur_cset = cset;
1729 tset->cur_task = task;
1733 cset = list_next_entry(cset, mg_node);
1741 * cgroup_task_migrate - move a task from one cgroup to another.
1742 * @old_cgrp; the cgroup @tsk is being migrated from
1743 * @tsk: the task being migrated
1744 * @new_cset: the new css_set @tsk is being attached to
1746 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1748 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1749 struct task_struct *tsk,
1750 struct css_set *new_cset)
1752 struct css_set *old_cset;
1754 lockdep_assert_held(&cgroup_mutex);
1755 lockdep_assert_held(&css_set_rwsem);
1758 * We are synchronized through threadgroup_lock() against PF_EXITING
1759 * setting such that we can't race against cgroup_exit() changing the
1760 * css_set to init_css_set and dropping the old one.
1762 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1763 old_cset = task_css_set(tsk);
1765 get_css_set(new_cset);
1766 rcu_assign_pointer(tsk->cgroups, new_cset);
1769 * Use move_tail so that cgroup_taskset_first() still returns the
1770 * leader after migration. This works because cgroup_migrate()
1771 * ensures that the dst_cset of the leader is the first on the
1772 * tset's dst_csets list.
1774 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
1777 * We just gained a reference on old_cset by taking it from the
1778 * task. As trading it for new_cset is protected by cgroup_mutex,
1779 * we're safe to drop it here; it will be freed under RCU.
1781 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1782 put_css_set_locked(old_cset, false);
1786 * cgroup_migrate_finish - cleanup after attach
1787 * @preloaded_csets: list of preloaded css_sets
1789 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
1790 * those functions for details.
1792 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
1794 struct css_set *cset, *tmp_cset;
1796 lockdep_assert_held(&cgroup_mutex);
1798 down_write(&css_set_rwsem);
1799 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
1800 cset->mg_src_cgrp = NULL;
1801 cset->mg_dst_cset = NULL;
1802 list_del_init(&cset->mg_preload_node);
1803 put_css_set_locked(cset, false);
1805 up_write(&css_set_rwsem);
1809 * cgroup_migrate_add_src - add a migration source css_set
1810 * @src_cset: the source css_set to add
1811 * @dst_cgrp: the destination cgroup
1812 * @preloaded_csets: list of preloaded css_sets
1814 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
1815 * @src_cset and add it to @preloaded_csets, which should later be cleaned
1816 * up by cgroup_migrate_finish().
1818 * This function may be called without holding threadgroup_lock even if the
1819 * target is a process. Threads may be created and destroyed but as long
1820 * as cgroup_mutex is not dropped, no new css_set can be put into play and
1821 * the preloaded css_sets are guaranteed to cover all migrations.
1823 static void cgroup_migrate_add_src(struct css_set *src_cset,
1824 struct cgroup *dst_cgrp,
1825 struct list_head *preloaded_csets)
1827 struct cgroup *src_cgrp;
1829 lockdep_assert_held(&cgroup_mutex);
1830 lockdep_assert_held(&css_set_rwsem);
1832 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
1834 /* nothing to do if this cset already belongs to the cgroup */
1835 if (src_cgrp == dst_cgrp)
1838 if (!list_empty(&src_cset->mg_preload_node))
1841 WARN_ON(src_cset->mg_src_cgrp);
1842 WARN_ON(!list_empty(&src_cset->mg_tasks));
1843 WARN_ON(!list_empty(&src_cset->mg_node));
1845 src_cset->mg_src_cgrp = src_cgrp;
1846 get_css_set(src_cset);
1847 list_add(&src_cset->mg_preload_node, preloaded_csets);
1851 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
1852 * @dst_cgrp: the destination cgroup
1853 * @preloaded_csets: list of preloaded source css_sets
1855 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
1856 * have been preloaded to @preloaded_csets. This function looks up and
1857 * pins all destination css_sets, links each to its source, and put them on
1860 * This function must be called after cgroup_migrate_add_src() has been
1861 * called on each migration source css_set. After migration is performed
1862 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
1865 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
1866 struct list_head *preloaded_csets)
1869 struct css_set *src_cset;
1871 lockdep_assert_held(&cgroup_mutex);
1873 /* look up the dst cset for each src cset and link it to src */
1874 list_for_each_entry(src_cset, preloaded_csets, mg_preload_node) {
1875 struct css_set *dst_cset;
1877 dst_cset = find_css_set(src_cset, dst_cgrp);
1881 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
1882 src_cset->mg_dst_cset = dst_cset;
1884 if (list_empty(&dst_cset->mg_preload_node))
1885 list_add(&dst_cset->mg_preload_node, &csets);
1887 put_css_set(dst_cset, false);
1890 list_splice(&csets, preloaded_csets);
1893 cgroup_migrate_finish(&csets);
1898 * cgroup_migrate - migrate a process or task to a cgroup
1899 * @cgrp: the destination cgroup
1900 * @leader: the leader of the process or the task to migrate
1901 * @threadgroup: whether @leader points to the whole process or a single task
1903 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
1904 * process, the caller must be holding threadgroup_lock of @leader. The
1905 * caller is also responsible for invoking cgroup_migrate_add_src() and
1906 * cgroup_migrate_prepare_dst() on the targets before invoking this
1907 * function and following up with cgroup_migrate_finish().
1909 * As long as a controller's ->can_attach() doesn't fail, this function is
1910 * guaranteed to succeed. This means that, excluding ->can_attach()
1911 * failure, when migrating multiple targets, the success or failure can be
1912 * decided for all targets by invoking group_migrate_prepare_dst() before
1913 * actually starting migrating.
1915 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
1918 struct cgroup_taskset tset = {
1919 .src_csets = LIST_HEAD_INIT(tset.src_csets),
1920 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
1921 .csets = &tset.src_csets,
1923 struct cgroup_subsys_state *css, *failed_css = NULL;
1924 struct css_set *cset, *tmp_cset;
1925 struct task_struct *task, *tmp_task;
1929 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1930 * already PF_EXITING could be freed from underneath us unless we
1931 * take an rcu_read_lock.
1933 down_write(&css_set_rwsem);
1937 /* @task either already exited or can't exit until the end */
1938 if (task->flags & PF_EXITING)
1941 /* leave @task alone if post_fork() hasn't linked it yet */
1942 if (list_empty(&task->cg_list))
1945 cset = task_css_set(task);
1946 if (!cset->mg_src_cgrp)
1950 * cgroup_taskset_first() must always return the leader.
1951 * Take care to avoid disturbing the ordering.
1953 list_move_tail(&task->cg_list, &cset->mg_tasks);
1954 if (list_empty(&cset->mg_node))
1955 list_add_tail(&cset->mg_node, &tset.src_csets);
1956 if (list_empty(&cset->mg_dst_cset->mg_node))
1957 list_move_tail(&cset->mg_dst_cset->mg_node,
1962 } while_each_thread(leader, task);
1964 up_write(&css_set_rwsem);
1966 /* methods shouldn't be called if no task is actually migrating */
1967 if (list_empty(&tset.src_csets))
1970 /* check that we can legitimately attach to the cgroup */
1971 for_each_css(css, i, cgrp) {
1972 if (css->ss->can_attach) {
1973 ret = css->ss->can_attach(css, &tset);
1976 goto out_cancel_attach;
1982 * Now that we're guaranteed success, proceed to move all tasks to
1983 * the new cgroup. There are no failure cases after here, so this
1984 * is the commit point.
1986 down_write(&css_set_rwsem);
1987 list_for_each_entry(cset, &tset.src_csets, mg_node) {
1988 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
1989 cgroup_task_migrate(cset->mg_src_cgrp, task,
1992 up_write(&css_set_rwsem);
1995 * Migration is committed, all target tasks are now on dst_csets.
1996 * Nothing is sensitive to fork() after this point. Notify
1997 * controllers that migration is complete.
1999 tset.csets = &tset.dst_csets;
2001 for_each_css(css, i, cgrp)
2002 if (css->ss->attach)
2003 css->ss->attach(css, &tset);
2006 goto out_release_tset;
2009 for_each_css(css, i, cgrp) {
2010 if (css == failed_css)
2012 if (css->ss->cancel_attach)
2013 css->ss->cancel_attach(css, &tset);
2016 down_write(&css_set_rwsem);
2017 list_splice_init(&tset.dst_csets, &tset.src_csets);
2018 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2019 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2020 list_del_init(&cset->mg_node);
2022 up_write(&css_set_rwsem);
2027 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2028 * @dst_cgrp: the cgroup to attach to
2029 * @leader: the task or the leader of the threadgroup to be attached
2030 * @threadgroup: attach the whole threadgroup?
2032 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2034 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2035 struct task_struct *leader, bool threadgroup)
2037 LIST_HEAD(preloaded_csets);
2038 struct task_struct *task;
2041 /* look up all src csets */
2042 down_read(&css_set_rwsem);
2046 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2050 } while_each_thread(leader, task);
2052 up_read(&css_set_rwsem);
2054 /* prepare dst csets and commit */
2055 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2057 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2059 cgroup_migrate_finish(&preloaded_csets);
2064 * Find the task_struct of the task to attach by vpid and pass it along to the
2065 * function to attach either it or all tasks in its threadgroup. Will lock
2066 * cgroup_mutex and threadgroup.
2068 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2070 struct task_struct *tsk;
2071 const struct cred *cred = current_cred(), *tcred;
2074 if (!cgroup_lock_live_group(cgrp))
2080 tsk = find_task_by_vpid(pid);
2084 goto out_unlock_cgroup;
2087 * even if we're attaching all tasks in the thread group, we
2088 * only need to check permissions on one of them.
2090 tcred = __task_cred(tsk);
2091 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2092 !uid_eq(cred->euid, tcred->uid) &&
2093 !uid_eq(cred->euid, tcred->suid)) {
2096 goto out_unlock_cgroup;
2102 tsk = tsk->group_leader;
2105 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2106 * trapped in a cpuset, or RT worker may be born in a cgroup
2107 * with no rt_runtime allocated. Just say no.
2109 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2112 goto out_unlock_cgroup;
2115 get_task_struct(tsk);
2118 threadgroup_lock(tsk);
2120 if (!thread_group_leader(tsk)) {
2122 * a race with de_thread from another thread's exec()
2123 * may strip us of our leadership, if this happens,
2124 * there is no choice but to throw this task away and
2125 * try again; this is
2126 * "double-double-toil-and-trouble-check locking".
2128 threadgroup_unlock(tsk);
2129 put_task_struct(tsk);
2130 goto retry_find_task;
2134 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2136 threadgroup_unlock(tsk);
2138 put_task_struct(tsk);
2140 mutex_unlock(&cgroup_mutex);
2145 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2146 * @from: attach to all cgroups of a given task
2147 * @tsk: the task to be attached
2149 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2151 struct cgroup_root *root;
2154 mutex_lock(&cgroup_mutex);
2155 for_each_root(root) {
2156 struct cgroup *from_cgrp;
2158 if (root == &cgrp_dfl_root)
2161 down_read(&css_set_rwsem);
2162 from_cgrp = task_cgroup_from_root(from, root);
2163 up_read(&css_set_rwsem);
2165 retval = cgroup_attach_task(from_cgrp, tsk, false);
2169 mutex_unlock(&cgroup_mutex);
2173 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2175 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2176 struct cftype *cft, u64 pid)
2178 return attach_task_by_pid(css->cgroup, pid, false);
2181 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2182 struct cftype *cft, u64 tgid)
2184 return attach_task_by_pid(css->cgroup, tgid, true);
2187 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2188 struct cftype *cft, char *buffer)
2190 struct cgroup_root *root = css->cgroup->root;
2192 BUILD_BUG_ON(sizeof(root->release_agent_path) < PATH_MAX);
2193 if (!cgroup_lock_live_group(css->cgroup))
2195 spin_lock(&release_agent_path_lock);
2196 strlcpy(root->release_agent_path, buffer,
2197 sizeof(root->release_agent_path));
2198 spin_unlock(&release_agent_path_lock);
2199 mutex_unlock(&cgroup_mutex);
2203 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2205 struct cgroup *cgrp = seq_css(seq)->cgroup;
2207 if (!cgroup_lock_live_group(cgrp))
2209 seq_puts(seq, cgrp->root->release_agent_path);
2210 seq_putc(seq, '\n');
2211 mutex_unlock(&cgroup_mutex);
2215 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2217 struct cgroup *cgrp = seq_css(seq)->cgroup;
2219 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2223 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2224 size_t nbytes, loff_t off)
2226 struct cgroup *cgrp = of->kn->parent->priv;
2227 struct cftype *cft = of->kn->priv;
2228 struct cgroup_subsys_state *css;
2232 * kernfs guarantees that a file isn't deleted with operations in
2233 * flight, which means that the matching css is and stays alive and
2234 * doesn't need to be pinned. The RCU locking is not necessary
2235 * either. It's just for the convenience of using cgroup_css().
2238 css = cgroup_css(cgrp, cft->ss);
2241 if (cft->write_string) {
2242 ret = cft->write_string(css, cft, strstrip(buf));
2243 } else if (cft->write_u64) {
2244 unsigned long long v;
2245 ret = kstrtoull(buf, 0, &v);
2247 ret = cft->write_u64(css, cft, v);
2248 } else if (cft->write_s64) {
2250 ret = kstrtoll(buf, 0, &v);
2252 ret = cft->write_s64(css, cft, v);
2253 } else if (cft->trigger) {
2254 ret = cft->trigger(css, (unsigned int)cft->private);
2259 return ret ?: nbytes;
2262 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2264 return seq_cft(seq)->seq_start(seq, ppos);
2267 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2269 return seq_cft(seq)->seq_next(seq, v, ppos);
2272 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2274 seq_cft(seq)->seq_stop(seq, v);
2277 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2279 struct cftype *cft = seq_cft(m);
2280 struct cgroup_subsys_state *css = seq_css(m);
2283 return cft->seq_show(m, arg);
2286 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2287 else if (cft->read_s64)
2288 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2294 static struct kernfs_ops cgroup_kf_single_ops = {
2295 .atomic_write_len = PAGE_SIZE,
2296 .write = cgroup_file_write,
2297 .seq_show = cgroup_seqfile_show,
2300 static struct kernfs_ops cgroup_kf_ops = {
2301 .atomic_write_len = PAGE_SIZE,
2302 .write = cgroup_file_write,
2303 .seq_start = cgroup_seqfile_start,
2304 .seq_next = cgroup_seqfile_next,
2305 .seq_stop = cgroup_seqfile_stop,
2306 .seq_show = cgroup_seqfile_show,
2310 * cgroup_rename - Only allow simple rename of directories in place.
2312 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2313 const char *new_name_str)
2315 struct cgroup *cgrp = kn->priv;
2318 if (kernfs_type(kn) != KERNFS_DIR)
2320 if (kn->parent != new_parent)
2324 * This isn't a proper migration and its usefulness is very
2325 * limited. Disallow if sane_behavior.
2327 if (cgroup_sane_behavior(cgrp))
2331 * We're gonna grab cgroup_tree_mutex which nests outside kernfs
2332 * active_ref. kernfs_rename() doesn't require active_ref
2333 * protection. Break them before grabbing cgroup_tree_mutex.
2335 kernfs_break_active_protection(new_parent);
2336 kernfs_break_active_protection(kn);
2338 mutex_lock(&cgroup_tree_mutex);
2339 mutex_lock(&cgroup_mutex);
2341 ret = kernfs_rename(kn, new_parent, new_name_str);
2343 mutex_unlock(&cgroup_mutex);
2344 mutex_unlock(&cgroup_tree_mutex);
2346 kernfs_unbreak_active_protection(kn);
2347 kernfs_unbreak_active_protection(new_parent);
2351 /* set uid and gid of cgroup dirs and files to that of the creator */
2352 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
2354 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
2355 .ia_uid = current_fsuid(),
2356 .ia_gid = current_fsgid(), };
2358 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
2359 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
2362 return kernfs_setattr(kn, &iattr);
2365 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2367 char name[CGROUP_FILE_NAME_MAX];
2368 struct kernfs_node *kn;
2369 struct lock_class_key *key = NULL;
2372 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2373 key = &cft->lockdep_key;
2375 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2376 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2381 ret = cgroup_kn_set_ugid(kn);
2388 * cgroup_addrm_files - add or remove files to a cgroup directory
2389 * @cgrp: the target cgroup
2390 * @cfts: array of cftypes to be added
2391 * @is_add: whether to add or remove
2393 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2394 * For removals, this function never fails. If addition fails, this
2395 * function doesn't remove files already added. The caller is responsible
2398 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2404 lockdep_assert_held(&cgroup_tree_mutex);
2406 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2407 /* does cft->flags tell us to skip this file on @cgrp? */
2408 if ((cft->flags & CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
2410 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2412 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2414 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2418 ret = cgroup_add_file(cgrp, cft);
2420 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2425 cgroup_rm_file(cgrp, cft);
2431 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2434 struct cgroup_subsys *ss = cfts[0].ss;
2435 struct cgroup *root = &ss->root->cgrp;
2436 struct cgroup_subsys_state *css;
2439 lockdep_assert_held(&cgroup_tree_mutex);
2441 /* don't bother if @ss isn't attached */
2442 if (ss->root == &cgrp_dfl_root)
2445 /* add/rm files for all cgroups created before */
2446 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2447 struct cgroup *cgrp = css->cgroup;
2449 if (cgroup_is_dead(cgrp))
2452 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2458 kernfs_activate(root->kn);
2462 static void cgroup_exit_cftypes(struct cftype *cfts)
2466 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2467 /* free copy for custom atomic_write_len, see init_cftypes() */
2468 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2475 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2479 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2480 struct kernfs_ops *kf_ops;
2482 WARN_ON(cft->ss || cft->kf_ops);
2485 kf_ops = &cgroup_kf_ops;
2487 kf_ops = &cgroup_kf_single_ops;
2490 * Ugh... if @cft wants a custom max_write_len, we need to
2491 * make a copy of kf_ops to set its atomic_write_len.
2493 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
2494 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
2496 cgroup_exit_cftypes(cfts);
2499 kf_ops->atomic_write_len = cft->max_write_len;
2502 cft->kf_ops = kf_ops;
2509 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
2511 lockdep_assert_held(&cgroup_tree_mutex);
2513 if (!cfts || !cfts[0].ss)
2516 list_del(&cfts->node);
2517 cgroup_apply_cftypes(cfts, false);
2518 cgroup_exit_cftypes(cfts);
2523 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2524 * @cfts: zero-length name terminated array of cftypes
2526 * Unregister @cfts. Files described by @cfts are removed from all
2527 * existing cgroups and all future cgroups won't have them either. This
2528 * function can be called anytime whether @cfts' subsys is attached or not.
2530 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2533 int cgroup_rm_cftypes(struct cftype *cfts)
2537 mutex_lock(&cgroup_tree_mutex);
2538 ret = cgroup_rm_cftypes_locked(cfts);
2539 mutex_unlock(&cgroup_tree_mutex);
2544 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2545 * @ss: target cgroup subsystem
2546 * @cfts: zero-length name terminated array of cftypes
2548 * Register @cfts to @ss. Files described by @cfts are created for all
2549 * existing cgroups to which @ss is attached and all future cgroups will
2550 * have them too. This function can be called anytime whether @ss is
2553 * Returns 0 on successful registration, -errno on failure. Note that this
2554 * function currently returns 0 as long as @cfts registration is successful
2555 * even if some file creation attempts on existing cgroups fail.
2557 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2561 if (!cfts || cfts[0].name[0] == '\0')
2564 ret = cgroup_init_cftypes(ss, cfts);
2568 mutex_lock(&cgroup_tree_mutex);
2570 list_add_tail(&cfts->node, &ss->cfts);
2571 ret = cgroup_apply_cftypes(cfts, true);
2573 cgroup_rm_cftypes_locked(cfts);
2575 mutex_unlock(&cgroup_tree_mutex);
2580 * cgroup_task_count - count the number of tasks in a cgroup.
2581 * @cgrp: the cgroup in question
2583 * Return the number of tasks in the cgroup.
2585 static int cgroup_task_count(const struct cgroup *cgrp)
2588 struct cgrp_cset_link *link;
2590 down_read(&css_set_rwsem);
2591 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2592 count += atomic_read(&link->cset->refcount);
2593 up_read(&css_set_rwsem);
2598 * css_next_child - find the next child of a given css
2599 * @pos_css: the current position (%NULL to initiate traversal)
2600 * @parent_css: css whose children to walk
2602 * This function returns the next child of @parent_css and should be called
2603 * under either cgroup_mutex or RCU read lock. The only requirement is
2604 * that @parent_css and @pos_css are accessible. The next sibling is
2605 * guaranteed to be returned regardless of their states.
2607 struct cgroup_subsys_state *
2608 css_next_child(struct cgroup_subsys_state *pos_css,
2609 struct cgroup_subsys_state *parent_css)
2611 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2612 struct cgroup *cgrp = parent_css->cgroup;
2613 struct cgroup *next;
2615 cgroup_assert_mutexes_or_rcu_locked();
2618 * @pos could already have been removed. Once a cgroup is removed,
2619 * its ->sibling.next is no longer updated when its next sibling
2620 * changes. As CGRP_DEAD assertion is serialized and happens
2621 * before the cgroup is taken off the ->sibling list, if we see it
2622 * unasserted, it's guaranteed that the next sibling hasn't
2623 * finished its grace period even if it's already removed, and thus
2624 * safe to dereference from this RCU critical section. If
2625 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2626 * to be visible as %true here.
2628 * If @pos is dead, its next pointer can't be dereferenced;
2629 * however, as each cgroup is given a monotonically increasing
2630 * unique serial number and always appended to the sibling list,
2631 * the next one can be found by walking the parent's children until
2632 * we see a cgroup with higher serial number than @pos's. While
2633 * this path can be slower, it's taken only when either the current
2634 * cgroup is removed or iteration and removal race.
2637 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2638 } else if (likely(!cgroup_is_dead(pos))) {
2639 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2641 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2642 if (next->serial_nr > pos->serial_nr)
2646 if (&next->sibling == &cgrp->children)
2649 return cgroup_css(next, parent_css->ss);
2653 * css_next_descendant_pre - find the next descendant for pre-order walk
2654 * @pos: the current position (%NULL to initiate traversal)
2655 * @root: css whose descendants to walk
2657 * To be used by css_for_each_descendant_pre(). Find the next descendant
2658 * to visit for pre-order traversal of @root's descendants. @root is
2659 * included in the iteration and the first node to be visited.
2661 * While this function requires cgroup_mutex or RCU read locking, it
2662 * doesn't require the whole traversal to be contained in a single critical
2663 * section. This function will return the correct next descendant as long
2664 * as both @pos and @root are accessible and @pos is a descendant of @root.
2666 struct cgroup_subsys_state *
2667 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2668 struct cgroup_subsys_state *root)
2670 struct cgroup_subsys_state *next;
2672 cgroup_assert_mutexes_or_rcu_locked();
2674 /* if first iteration, visit @root */
2678 /* visit the first child if exists */
2679 next = css_next_child(NULL, pos);
2683 /* no child, visit my or the closest ancestor's next sibling */
2684 while (pos != root) {
2685 next = css_next_child(pos, css_parent(pos));
2688 pos = css_parent(pos);
2695 * css_rightmost_descendant - return the rightmost descendant of a css
2696 * @pos: css of interest
2698 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2699 * is returned. This can be used during pre-order traversal to skip
2702 * While this function requires cgroup_mutex or RCU read locking, it
2703 * doesn't require the whole traversal to be contained in a single critical
2704 * section. This function will return the correct rightmost descendant as
2705 * long as @pos is accessible.
2707 struct cgroup_subsys_state *
2708 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2710 struct cgroup_subsys_state *last, *tmp;
2712 cgroup_assert_mutexes_or_rcu_locked();
2716 /* ->prev isn't RCU safe, walk ->next till the end */
2718 css_for_each_child(tmp, last)
2725 static struct cgroup_subsys_state *
2726 css_leftmost_descendant(struct cgroup_subsys_state *pos)
2728 struct cgroup_subsys_state *last;
2732 pos = css_next_child(NULL, pos);
2739 * css_next_descendant_post - find the next descendant for post-order walk
2740 * @pos: the current position (%NULL to initiate traversal)
2741 * @root: css whose descendants to walk
2743 * To be used by css_for_each_descendant_post(). Find the next descendant
2744 * to visit for post-order traversal of @root's descendants. @root is
2745 * included in the iteration and the last node to be visited.
2747 * While this function requires cgroup_mutex or RCU read locking, it
2748 * doesn't require the whole traversal to be contained in a single critical
2749 * section. This function will return the correct next descendant as long
2750 * as both @pos and @cgroup are accessible and @pos is a descendant of
2753 struct cgroup_subsys_state *
2754 css_next_descendant_post(struct cgroup_subsys_state *pos,
2755 struct cgroup_subsys_state *root)
2757 struct cgroup_subsys_state *next;
2759 cgroup_assert_mutexes_or_rcu_locked();
2761 /* if first iteration, visit leftmost descendant which may be @root */
2763 return css_leftmost_descendant(root);
2765 /* if we visited @root, we're done */
2769 /* if there's an unvisited sibling, visit its leftmost descendant */
2770 next = css_next_child(pos, css_parent(pos));
2772 return css_leftmost_descendant(next);
2774 /* no sibling left, visit parent */
2775 return css_parent(pos);
2779 * css_advance_task_iter - advance a task itererator to the next css_set
2780 * @it: the iterator to advance
2782 * Advance @it to the next css_set to walk.
2784 static void css_advance_task_iter(struct css_task_iter *it)
2786 struct list_head *l = it->cset_link;
2787 struct cgrp_cset_link *link;
2788 struct css_set *cset;
2790 /* Advance to the next non-empty css_set */
2793 if (l == &it->origin_css->cgroup->cset_links) {
2794 it->cset_link = NULL;
2797 link = list_entry(l, struct cgrp_cset_link, cset_link);
2799 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
2803 if (!list_empty(&cset->tasks))
2804 it->task = cset->tasks.next;
2806 it->task = cset->mg_tasks.next;
2810 * css_task_iter_start - initiate task iteration
2811 * @css: the css to walk tasks of
2812 * @it: the task iterator to use
2814 * Initiate iteration through the tasks of @css. The caller can call
2815 * css_task_iter_next() to walk through the tasks until the function
2816 * returns NULL. On completion of iteration, css_task_iter_end() must be
2819 * Note that this function acquires a lock which is released when the
2820 * iteration finishes. The caller can't sleep while iteration is in
2823 void css_task_iter_start(struct cgroup_subsys_state *css,
2824 struct css_task_iter *it)
2825 __acquires(css_set_rwsem)
2827 /* no one should try to iterate before mounting cgroups */
2828 WARN_ON_ONCE(!use_task_css_set_links);
2830 down_read(&css_set_rwsem);
2832 it->origin_css = css;
2833 it->cset_link = &css->cgroup->cset_links;
2835 css_advance_task_iter(it);
2839 * css_task_iter_next - return the next task for the iterator
2840 * @it: the task iterator being iterated
2842 * The "next" function for task iteration. @it should have been
2843 * initialized via css_task_iter_start(). Returns NULL when the iteration
2846 struct task_struct *css_task_iter_next(struct css_task_iter *it)
2848 struct task_struct *res;
2849 struct list_head *l = it->task;
2850 struct cgrp_cset_link *link = list_entry(it->cset_link,
2851 struct cgrp_cset_link, cset_link);
2853 /* If the iterator cg is NULL, we have no tasks */
2856 res = list_entry(l, struct task_struct, cg_list);
2859 * Advance iterator to find next entry. cset->tasks is consumed
2860 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
2865 if (l == &link->cset->tasks)
2866 l = link->cset->mg_tasks.next;
2868 if (l == &link->cset->mg_tasks)
2869 css_advance_task_iter(it);
2877 * css_task_iter_end - finish task iteration
2878 * @it: the task iterator to finish
2880 * Finish task iteration started by css_task_iter_start().
2882 void css_task_iter_end(struct css_task_iter *it)
2883 __releases(css_set_rwsem)
2885 up_read(&css_set_rwsem);
2889 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
2890 * @to: cgroup to which the tasks will be moved
2891 * @from: cgroup in which the tasks currently reside
2893 * Locking rules between cgroup_post_fork() and the migration path
2894 * guarantee that, if a task is forking while being migrated, the new child
2895 * is guaranteed to be either visible in the source cgroup after the
2896 * parent's migration is complete or put into the target cgroup. No task
2897 * can slip out of migration through forking.
2899 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
2901 LIST_HEAD(preloaded_csets);
2902 struct cgrp_cset_link *link;
2903 struct css_task_iter it;
2904 struct task_struct *task;
2907 mutex_lock(&cgroup_mutex);
2909 /* all tasks in @from are being moved, all csets are source */
2910 down_read(&css_set_rwsem);
2911 list_for_each_entry(link, &from->cset_links, cset_link)
2912 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
2913 up_read(&css_set_rwsem);
2915 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
2920 * Migrate tasks one-by-one until @form is empty. This fails iff
2921 * ->can_attach() fails.
2924 css_task_iter_start(&from->dummy_css, &it);
2925 task = css_task_iter_next(&it);
2927 get_task_struct(task);
2928 css_task_iter_end(&it);
2931 ret = cgroup_migrate(to, task, false);
2932 put_task_struct(task);
2934 } while (task && !ret);
2936 cgroup_migrate_finish(&preloaded_csets);
2937 mutex_unlock(&cgroup_mutex);
2942 * Stuff for reading the 'tasks'/'procs' files.
2944 * Reading this file can return large amounts of data if a cgroup has
2945 * *lots* of attached tasks. So it may need several calls to read(),
2946 * but we cannot guarantee that the information we produce is correct
2947 * unless we produce it entirely atomically.
2951 /* which pidlist file are we talking about? */
2952 enum cgroup_filetype {
2958 * A pidlist is a list of pids that virtually represents the contents of one
2959 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
2960 * a pair (one each for procs, tasks) for each pid namespace that's relevant
2963 struct cgroup_pidlist {
2965 * used to find which pidlist is wanted. doesn't change as long as
2966 * this particular list stays in the list.
2968 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
2971 /* how many elements the above list has */
2973 /* each of these stored in a list by its cgroup */
2974 struct list_head links;
2975 /* pointer to the cgroup we belong to, for list removal purposes */
2976 struct cgroup *owner;
2977 /* for delayed destruction */
2978 struct delayed_work destroy_dwork;
2982 * The following two functions "fix" the issue where there are more pids
2983 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2984 * TODO: replace with a kernel-wide solution to this problem
2986 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2987 static void *pidlist_allocate(int count)
2989 if (PIDLIST_TOO_LARGE(count))
2990 return vmalloc(count * sizeof(pid_t));
2992 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2995 static void pidlist_free(void *p)
2997 if (is_vmalloc_addr(p))
3004 * Used to destroy all pidlists lingering waiting for destroy timer. None
3005 * should be left afterwards.
3007 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3009 struct cgroup_pidlist *l, *tmp_l;
3011 mutex_lock(&cgrp->pidlist_mutex);
3012 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3013 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3014 mutex_unlock(&cgrp->pidlist_mutex);
3016 flush_workqueue(cgroup_pidlist_destroy_wq);
3017 BUG_ON(!list_empty(&cgrp->pidlists));
3020 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3022 struct delayed_work *dwork = to_delayed_work(work);
3023 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3025 struct cgroup_pidlist *tofree = NULL;
3027 mutex_lock(&l->owner->pidlist_mutex);
3030 * Destroy iff we didn't get queued again. The state won't change
3031 * as destroy_dwork can only be queued while locked.
3033 if (!delayed_work_pending(dwork)) {
3034 list_del(&l->links);
3035 pidlist_free(l->list);
3036 put_pid_ns(l->key.ns);
3040 mutex_unlock(&l->owner->pidlist_mutex);
3045 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3046 * Returns the number of unique elements.
3048 static int pidlist_uniq(pid_t *list, int length)
3053 * we presume the 0th element is unique, so i starts at 1. trivial
3054 * edge cases first; no work needs to be done for either
3056 if (length == 0 || length == 1)
3058 /* src and dest walk down the list; dest counts unique elements */
3059 for (src = 1; src < length; src++) {
3060 /* find next unique element */
3061 while (list[src] == list[src-1]) {
3066 /* dest always points to where the next unique element goes */
3067 list[dest] = list[src];
3075 * The two pid files - task and cgroup.procs - guaranteed that the result
3076 * is sorted, which forced this whole pidlist fiasco. As pid order is
3077 * different per namespace, each namespace needs differently sorted list,
3078 * making it impossible to use, for example, single rbtree of member tasks
3079 * sorted by task pointer. As pidlists can be fairly large, allocating one
3080 * per open file is dangerous, so cgroup had to implement shared pool of
3081 * pidlists keyed by cgroup and namespace.
3083 * All this extra complexity was caused by the original implementation
3084 * committing to an entirely unnecessary property. In the long term, we
3085 * want to do away with it. Explicitly scramble sort order if
3086 * sane_behavior so that no such expectation exists in the new interface.
3088 * Scrambling is done by swapping every two consecutive bits, which is
3089 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3091 static pid_t pid_fry(pid_t pid)
3093 unsigned a = pid & 0x55555555;
3094 unsigned b = pid & 0xAAAAAAAA;
3096 return (a << 1) | (b >> 1);
3099 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3101 if (cgroup_sane_behavior(cgrp))
3102 return pid_fry(pid);
3107 static int cmppid(const void *a, const void *b)
3109 return *(pid_t *)a - *(pid_t *)b;
3112 static int fried_cmppid(const void *a, const void *b)
3114 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3117 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3118 enum cgroup_filetype type)
3120 struct cgroup_pidlist *l;
3121 /* don't need task_nsproxy() if we're looking at ourself */
3122 struct pid_namespace *ns = task_active_pid_ns(current);
3124 lockdep_assert_held(&cgrp->pidlist_mutex);
3126 list_for_each_entry(l, &cgrp->pidlists, links)
3127 if (l->key.type == type && l->key.ns == ns)
3133 * find the appropriate pidlist for our purpose (given procs vs tasks)
3134 * returns with the lock on that pidlist already held, and takes care
3135 * of the use count, or returns NULL with no locks held if we're out of
3138 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3139 enum cgroup_filetype type)
3141 struct cgroup_pidlist *l;
3143 lockdep_assert_held(&cgrp->pidlist_mutex);
3145 l = cgroup_pidlist_find(cgrp, type);
3149 /* entry not found; create a new one */
3150 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3154 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3156 /* don't need task_nsproxy() if we're looking at ourself */
3157 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3159 list_add(&l->links, &cgrp->pidlists);
3164 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3166 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3167 struct cgroup_pidlist **lp)
3171 int pid, n = 0; /* used for populating the array */
3172 struct css_task_iter it;
3173 struct task_struct *tsk;
3174 struct cgroup_pidlist *l;
3176 lockdep_assert_held(&cgrp->pidlist_mutex);
3179 * If cgroup gets more users after we read count, we won't have
3180 * enough space - tough. This race is indistinguishable to the
3181 * caller from the case that the additional cgroup users didn't
3182 * show up until sometime later on.
3184 length = cgroup_task_count(cgrp);
3185 array = pidlist_allocate(length);
3188 /* now, populate the array */
3189 css_task_iter_start(&cgrp->dummy_css, &it);
3190 while ((tsk = css_task_iter_next(&it))) {
3191 if (unlikely(n == length))
3193 /* get tgid or pid for procs or tasks file respectively */
3194 if (type == CGROUP_FILE_PROCS)
3195 pid = task_tgid_vnr(tsk);
3197 pid = task_pid_vnr(tsk);
3198 if (pid > 0) /* make sure to only use valid results */
3201 css_task_iter_end(&it);
3203 /* now sort & (if procs) strip out duplicates */
3204 if (cgroup_sane_behavior(cgrp))
3205 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3207 sort(array, length, sizeof(pid_t), cmppid, NULL);
3208 if (type == CGROUP_FILE_PROCS)
3209 length = pidlist_uniq(array, length);
3211 l = cgroup_pidlist_find_create(cgrp, type);
3213 mutex_unlock(&cgrp->pidlist_mutex);
3214 pidlist_free(array);
3218 /* store array, freeing old if necessary */
3219 pidlist_free(l->list);
3227 * cgroupstats_build - build and fill cgroupstats
3228 * @stats: cgroupstats to fill information into
3229 * @dentry: A dentry entry belonging to the cgroup for which stats have
3232 * Build and fill cgroupstats so that taskstats can export it to user
3235 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3237 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3238 struct cgroup *cgrp;
3239 struct css_task_iter it;
3240 struct task_struct *tsk;
3242 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3243 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3244 kernfs_type(kn) != KERNFS_DIR)
3247 mutex_lock(&cgroup_mutex);
3250 * We aren't being called from kernfs and there's no guarantee on
3251 * @kn->priv's validity. For this and css_tryget_from_dir(),
3252 * @kn->priv is RCU safe. Let's do the RCU dancing.
3255 cgrp = rcu_dereference(kn->priv);
3256 if (!cgrp || cgroup_is_dead(cgrp)) {
3258 mutex_unlock(&cgroup_mutex);
3263 css_task_iter_start(&cgrp->dummy_css, &it);
3264 while ((tsk = css_task_iter_next(&it))) {
3265 switch (tsk->state) {
3267 stats->nr_running++;
3269 case TASK_INTERRUPTIBLE:
3270 stats->nr_sleeping++;
3272 case TASK_UNINTERRUPTIBLE:
3273 stats->nr_uninterruptible++;
3276 stats->nr_stopped++;
3279 if (delayacct_is_task_waiting_on_io(tsk))
3280 stats->nr_io_wait++;
3284 css_task_iter_end(&it);
3286 mutex_unlock(&cgroup_mutex);
3292 * seq_file methods for the tasks/procs files. The seq_file position is the
3293 * next pid to display; the seq_file iterator is a pointer to the pid
3294 * in the cgroup->l->list array.
3297 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3300 * Initially we receive a position value that corresponds to
3301 * one more than the last pid shown (or 0 on the first call or
3302 * after a seek to the start). Use a binary-search to find the
3303 * next pid to display, if any
3305 struct kernfs_open_file *of = s->private;
3306 struct cgroup *cgrp = seq_css(s)->cgroup;
3307 struct cgroup_pidlist *l;
3308 enum cgroup_filetype type = seq_cft(s)->private;
3309 int index = 0, pid = *pos;
3312 mutex_lock(&cgrp->pidlist_mutex);
3315 * !NULL @of->priv indicates that this isn't the first start()
3316 * after open. If the matching pidlist is around, we can use that.
3317 * Look for it. Note that @of->priv can't be used directly. It
3318 * could already have been destroyed.
3321 of->priv = cgroup_pidlist_find(cgrp, type);
3324 * Either this is the first start() after open or the matching
3325 * pidlist has been destroyed inbetween. Create a new one.
3328 ret = pidlist_array_load(cgrp, type,
3329 (struct cgroup_pidlist **)&of->priv);
3331 return ERR_PTR(ret);
3336 int end = l->length;
3338 while (index < end) {
3339 int mid = (index + end) / 2;
3340 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3343 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3349 /* If we're off the end of the array, we're done */
3350 if (index >= l->length)
3352 /* Update the abstract position to be the actual pid that we found */
3353 iter = l->list + index;
3354 *pos = cgroup_pid_fry(cgrp, *iter);
3358 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3360 struct kernfs_open_file *of = s->private;
3361 struct cgroup_pidlist *l = of->priv;
3364 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3365 CGROUP_PIDLIST_DESTROY_DELAY);
3366 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3369 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3371 struct kernfs_open_file *of = s->private;
3372 struct cgroup_pidlist *l = of->priv;
3374 pid_t *end = l->list + l->length;
3376 * Advance to the next pid in the array. If this goes off the
3383 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3388 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3390 return seq_printf(s, "%d\n", *(int *)v);
3394 * seq_operations functions for iterating on pidlists through seq_file -
3395 * independent of whether it's tasks or procs
3397 static const struct seq_operations cgroup_pidlist_seq_operations = {
3398 .start = cgroup_pidlist_start,
3399 .stop = cgroup_pidlist_stop,
3400 .next = cgroup_pidlist_next,
3401 .show = cgroup_pidlist_show,
3404 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3407 return notify_on_release(css->cgroup);
3410 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3411 struct cftype *cft, u64 val)
3413 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3415 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3417 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3421 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3424 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3427 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3428 struct cftype *cft, u64 val)
3431 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3433 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3437 static struct cftype cgroup_base_files[] = {
3439 .name = "cgroup.procs",
3440 .seq_start = cgroup_pidlist_start,
3441 .seq_next = cgroup_pidlist_next,
3442 .seq_stop = cgroup_pidlist_stop,
3443 .seq_show = cgroup_pidlist_show,
3444 .private = CGROUP_FILE_PROCS,
3445 .write_u64 = cgroup_procs_write,
3446 .mode = S_IRUGO | S_IWUSR,
3449 .name = "cgroup.clone_children",
3450 .flags = CFTYPE_INSANE,
3451 .read_u64 = cgroup_clone_children_read,
3452 .write_u64 = cgroup_clone_children_write,
3455 .name = "cgroup.sane_behavior",
3456 .flags = CFTYPE_ONLY_ON_ROOT,
3457 .seq_show = cgroup_sane_behavior_show,
3461 * Historical crazy stuff. These don't have "cgroup." prefix and
3462 * don't exist if sane_behavior. If you're depending on these, be
3463 * prepared to be burned.
3467 .flags = CFTYPE_INSANE, /* use "procs" instead */
3468 .seq_start = cgroup_pidlist_start,
3469 .seq_next = cgroup_pidlist_next,
3470 .seq_stop = cgroup_pidlist_stop,
3471 .seq_show = cgroup_pidlist_show,
3472 .private = CGROUP_FILE_TASKS,
3473 .write_u64 = cgroup_tasks_write,
3474 .mode = S_IRUGO | S_IWUSR,
3477 .name = "notify_on_release",
3478 .flags = CFTYPE_INSANE,
3479 .read_u64 = cgroup_read_notify_on_release,
3480 .write_u64 = cgroup_write_notify_on_release,
3483 .name = "release_agent",
3484 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3485 .seq_show = cgroup_release_agent_show,
3486 .write_string = cgroup_release_agent_write,
3487 .max_write_len = PATH_MAX - 1,
3493 * cgroup_populate_dir - create subsys files in a cgroup directory
3494 * @cgrp: target cgroup
3495 * @subsys_mask: mask of the subsystem ids whose files should be added
3497 * On failure, no file is added.
3499 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3501 struct cgroup_subsys *ss;
3504 /* process cftsets of each subsystem */
3505 for_each_subsys(ss, i) {
3506 struct cftype *cfts;
3508 if (!test_bit(i, &subsys_mask))
3511 list_for_each_entry(cfts, &ss->cfts, node) {
3512 ret = cgroup_addrm_files(cgrp, cfts, true);
3519 cgroup_clear_dir(cgrp, subsys_mask);
3524 * css destruction is four-stage process.
3526 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3527 * Implemented in kill_css().
3529 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3530 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3531 * by invoking offline_css(). After offlining, the base ref is put.
3532 * Implemented in css_killed_work_fn().
3534 * 3. When the percpu_ref reaches zero, the only possible remaining
3535 * accessors are inside RCU read sections. css_release() schedules the
3538 * 4. After the grace period, the css can be freed. Implemented in
3539 * css_free_work_fn().
3541 * It is actually hairier because both step 2 and 4 require process context
3542 * and thus involve punting to css->destroy_work adding two additional
3543 * steps to the already complex sequence.
3545 static void css_free_work_fn(struct work_struct *work)
3547 struct cgroup_subsys_state *css =
3548 container_of(work, struct cgroup_subsys_state, destroy_work);
3549 struct cgroup *cgrp = css->cgroup;
3552 css_put(css->parent);
3554 css->ss->css_free(css);
3558 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3560 struct cgroup_subsys_state *css =
3561 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3563 INIT_WORK(&css->destroy_work, css_free_work_fn);
3564 queue_work(cgroup_destroy_wq, &css->destroy_work);
3567 static void css_release(struct percpu_ref *ref)
3569 struct cgroup_subsys_state *css =
3570 container_of(ref, struct cgroup_subsys_state, refcnt);
3572 RCU_INIT_POINTER(css->cgroup->subsys[css->ss->id], NULL);
3573 call_rcu(&css->rcu_head, css_free_rcu_fn);
3576 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3577 struct cgroup *cgrp)
3584 css->parent = cgroup_css(cgrp->parent, ss);
3586 css->flags |= CSS_ROOT;
3588 BUG_ON(cgroup_css(cgrp, ss));
3591 /* invoke ->css_online() on a new CSS and mark it online if successful */
3592 static int online_css(struct cgroup_subsys_state *css)
3594 struct cgroup_subsys *ss = css->ss;
3597 lockdep_assert_held(&cgroup_tree_mutex);
3598 lockdep_assert_held(&cgroup_mutex);
3601 ret = ss->css_online(css);
3603 css->flags |= CSS_ONLINE;
3604 css->cgroup->nr_css++;
3605 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
3610 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
3611 static void offline_css(struct cgroup_subsys_state *css)
3613 struct cgroup_subsys *ss = css->ss;
3615 lockdep_assert_held(&cgroup_tree_mutex);
3616 lockdep_assert_held(&cgroup_mutex);
3618 if (!(css->flags & CSS_ONLINE))
3621 if (ss->css_offline)
3622 ss->css_offline(css);
3624 css->flags &= ~CSS_ONLINE;
3625 css->cgroup->nr_css--;
3626 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
3630 * create_css - create a cgroup_subsys_state
3631 * @cgrp: the cgroup new css will be associated with
3632 * @ss: the subsys of new css
3634 * Create a new css associated with @cgrp - @ss pair. On success, the new
3635 * css is online and installed in @cgrp with all interface files created.
3636 * Returns 0 on success, -errno on failure.
3638 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
3640 struct cgroup *parent = cgrp->parent;
3641 struct cgroup_subsys_state *css;
3644 lockdep_assert_held(&cgroup_mutex);
3646 css = ss->css_alloc(cgroup_css(parent, ss));
3648 return PTR_ERR(css);
3650 err = percpu_ref_init(&css->refcnt, css_release);
3654 init_css(css, ss, cgrp);
3656 err = cgroup_populate_dir(cgrp, 1 << ss->id);
3658 goto err_free_percpu_ref;
3660 err = online_css(css);
3665 css_get(css->parent);
3667 cgrp->subsys_mask |= 1 << ss->id;
3669 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
3671 pr_warning("cgroup: %s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
3672 current->comm, current->pid, ss->name);
3673 if (!strcmp(ss->name, "memory"))
3674 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
3675 ss->warned_broken_hierarchy = true;
3681 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3682 err_free_percpu_ref:
3683 percpu_ref_cancel_init(&css->refcnt);
3690 * cgroup_create - create a cgroup
3691 * @parent: cgroup that will be parent of the new cgroup
3692 * @name: name of the new cgroup
3693 * @mode: mode to set on new cgroup
3695 static long cgroup_create(struct cgroup *parent, const char *name,
3698 struct cgroup *cgrp;
3699 struct cgroup_root *root = parent->root;
3701 struct cgroup_subsys *ss;
3702 struct kernfs_node *kn;
3705 * XXX: The default hierarchy isn't fully implemented yet. Block
3706 * !root cgroup creation on it for now.
3708 if (root == &cgrp_dfl_root)
3711 /* allocate the cgroup and its ID, 0 is reserved for the root */
3712 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3716 mutex_lock(&cgroup_tree_mutex);
3719 * Only live parents can have children. Note that the liveliness
3720 * check isn't strictly necessary because cgroup_mkdir() and
3721 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
3722 * anyway so that locking is contained inside cgroup proper and we
3723 * don't get nasty surprises if we ever grow another caller.
3725 if (!cgroup_lock_live_group(parent)) {
3727 goto err_unlock_tree;
3731 * Temporarily set the pointer to NULL, so idr_find() won't return
3732 * a half-baked cgroup.
3734 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
3740 init_cgroup_housekeeping(cgrp);
3742 cgrp->parent = parent;
3743 cgrp->dummy_css.parent = &parent->dummy_css;
3744 cgrp->root = parent->root;
3746 if (notify_on_release(parent))
3747 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3749 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
3750 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
3752 /* create the directory */
3753 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
3761 * This extra ref will be put in cgroup_free_fn() and guarantees
3762 * that @cgrp->kn is always accessible.
3766 cgrp->serial_nr = cgroup_serial_nr_next++;
3768 /* allocation complete, commit to creation */
3769 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
3770 atomic_inc(&root->nr_cgrps);
3774 * @cgrp is now fully operational. If something fails after this
3775 * point, it'll be released via the normal destruction path.
3777 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
3779 err = cgroup_kn_set_ugid(kn);
3783 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
3787 /* let's create and online css's */
3788 for_each_subsys(ss, ssid) {
3789 if (root->cgrp.subsys_mask & (1 << ssid)) {
3790 err = create_css(cgrp, ss);
3796 kernfs_activate(kn);
3798 mutex_unlock(&cgroup_mutex);
3799 mutex_unlock(&cgroup_tree_mutex);
3804 idr_remove(&root->cgroup_idr, cgrp->id);
3806 mutex_unlock(&cgroup_mutex);
3808 mutex_unlock(&cgroup_tree_mutex);
3813 cgroup_destroy_locked(cgrp);
3814 mutex_unlock(&cgroup_mutex);
3815 mutex_unlock(&cgroup_tree_mutex);
3819 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3822 struct cgroup *parent = parent_kn->priv;
3826 * cgroup_create() grabs cgroup_tree_mutex which nests outside
3827 * kernfs active_ref and cgroup_create() already synchronizes
3828 * properly against removal through cgroup_lock_live_group().
3829 * Break it before calling cgroup_create().
3832 kernfs_break_active_protection(parent_kn);
3834 ret = cgroup_create(parent, name, mode);
3836 kernfs_unbreak_active_protection(parent_kn);
3842 * This is called when the refcnt of a css is confirmed to be killed.
3843 * css_tryget() is now guaranteed to fail.
3845 static void css_killed_work_fn(struct work_struct *work)
3847 struct cgroup_subsys_state *css =
3848 container_of(work, struct cgroup_subsys_state, destroy_work);
3849 struct cgroup *cgrp = css->cgroup;
3851 mutex_lock(&cgroup_tree_mutex);
3852 mutex_lock(&cgroup_mutex);
3855 * css_tryget() is guaranteed to fail now. Tell subsystems to
3856 * initate destruction.
3861 * If @cgrp is marked dead, it's waiting for refs of all css's to
3862 * be disabled before proceeding to the second phase of cgroup
3863 * destruction. If we are the last one, kick it off.
3865 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
3866 cgroup_destroy_css_killed(cgrp);
3868 mutex_unlock(&cgroup_mutex);
3869 mutex_unlock(&cgroup_tree_mutex);
3872 * Put the css refs from kill_css(). Each css holds an extra
3873 * reference to the cgroup's dentry and cgroup removal proceeds
3874 * regardless of css refs. On the last put of each css, whenever
3875 * that may be, the extra dentry ref is put so that dentry
3876 * destruction happens only after all css's are released.
3881 /* css kill confirmation processing requires process context, bounce */
3882 static void css_killed_ref_fn(struct percpu_ref *ref)
3884 struct cgroup_subsys_state *css =
3885 container_of(ref, struct cgroup_subsys_state, refcnt);
3887 INIT_WORK(&css->destroy_work, css_killed_work_fn);
3888 queue_work(cgroup_destroy_wq, &css->destroy_work);
3891 static void __kill_css(struct cgroup_subsys_state *css)
3893 lockdep_assert_held(&cgroup_tree_mutex);
3896 * This must happen before css is disassociated with its cgroup.
3897 * See seq_css() for details.
3899 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3902 * Killing would put the base ref, but we need to keep it alive
3903 * until after ->css_offline().
3908 * cgroup core guarantees that, by the time ->css_offline() is
3909 * invoked, no new css reference will be given out via
3910 * css_tryget(). We can't simply call percpu_ref_kill() and
3911 * proceed to offlining css's because percpu_ref_kill() doesn't
3912 * guarantee that the ref is seen as killed on all CPUs on return.
3914 * Use percpu_ref_kill_and_confirm() to get notifications as each
3915 * css is confirmed to be seen as killed on all CPUs.
3917 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
3921 * kill_css - destroy a css
3922 * @css: css to destroy
3924 * This function initiates destruction of @css by removing cgroup interface
3925 * files and putting its base reference. ->css_offline() will be invoked
3926 * asynchronously once css_tryget() is guaranteed to fail and when the
3927 * reference count reaches zero, @css will be released.
3929 static void kill_css(struct cgroup_subsys_state *css)
3931 struct cgroup *cgrp = css->cgroup;
3933 lockdep_assert_held(&cgroup_tree_mutex);
3935 /* if already killed, noop */
3936 if (cgrp->subsys_mask & (1 << css->ss->id)) {
3937 cgrp->subsys_mask &= ~(1 << css->ss->id);
3943 * cgroup_destroy_locked - the first stage of cgroup destruction
3944 * @cgrp: cgroup to be destroyed
3946 * css's make use of percpu refcnts whose killing latency shouldn't be
3947 * exposed to userland and are RCU protected. Also, cgroup core needs to
3948 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
3949 * invoked. To satisfy all the requirements, destruction is implemented in
3950 * the following two steps.
3952 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
3953 * userland visible parts and start killing the percpu refcnts of
3954 * css's. Set up so that the next stage will be kicked off once all
3955 * the percpu refcnts are confirmed to be killed.
3957 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
3958 * rest of destruction. Once all cgroup references are gone, the
3959 * cgroup is RCU-freed.
3961 * This function implements s1. After this step, @cgrp is gone as far as
3962 * the userland is concerned and a new cgroup with the same name may be
3963 * created. As cgroup doesn't care about the names internally, this
3964 * doesn't cause any problem.
3966 static int cgroup_destroy_locked(struct cgroup *cgrp)
3967 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
3969 struct cgroup *child;
3970 struct cgroup_subsys_state *css;
3974 lockdep_assert_held(&cgroup_tree_mutex);
3975 lockdep_assert_held(&cgroup_mutex);
3978 * css_set_rwsem synchronizes access to ->cset_links and prevents
3979 * @cgrp from being removed while put_css_set() is in progress.
3981 down_read(&css_set_rwsem);
3982 empty = list_empty(&cgrp->cset_links);
3983 up_read(&css_set_rwsem);
3988 * Make sure there's no live children. We can't test ->children
3989 * emptiness as dead children linger on it while being destroyed;
3990 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
3994 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
3995 empty = cgroup_is_dead(child);
4004 * Mark @cgrp dead. This prevents further task migration and child
4005 * creation by disabling cgroup_lock_live_group(). Note that
4006 * CGRP_DEAD assertion is depended upon by css_next_child() to
4007 * resume iteration after dropping RCU read lock. See
4008 * css_next_child() for details.
4010 set_bit(CGRP_DEAD, &cgrp->flags);
4013 * Initiate massacre of all css's. cgroup_destroy_css_killed()
4014 * will be invoked to perform the rest of destruction once the
4015 * percpu refs of all css's are confirmed to be killed. This
4016 * involves removing the subsystem's files, drop cgroup_mutex.
4018 mutex_unlock(&cgroup_mutex);
4019 for_each_css(css, ssid, cgrp)
4021 mutex_lock(&cgroup_mutex);
4023 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4024 raw_spin_lock(&release_list_lock);
4025 if (!list_empty(&cgrp->release_list))
4026 list_del_init(&cgrp->release_list);
4027 raw_spin_unlock(&release_list_lock);
4030 * If @cgrp has css's attached, the second stage of cgroup
4031 * destruction is kicked off from css_killed_work_fn() after the
4032 * refs of all attached css's are killed. If @cgrp doesn't have
4033 * any css, we kick it off here.
4036 cgroup_destroy_css_killed(cgrp);
4038 /* remove @cgrp directory along with the base files */
4039 mutex_unlock(&cgroup_mutex);
4042 * There are two control paths which try to determine cgroup from
4043 * dentry without going through kernfs - cgroupstats_build() and
4044 * css_tryget_from_dir(). Those are supported by RCU protecting
4045 * clearing of cgrp->kn->priv backpointer, which should happen
4046 * after all files under it have been removed.
4048 kernfs_remove(cgrp->kn); /* @cgrp has an extra ref on its kn */
4049 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4051 mutex_lock(&cgroup_mutex);
4057 * cgroup_destroy_css_killed - the second step of cgroup destruction
4058 * @work: cgroup->destroy_free_work
4060 * This function is invoked from a work item for a cgroup which is being
4061 * destroyed after all css's are offlined and performs the rest of
4062 * destruction. This is the second step of destruction described in the
4063 * comment above cgroup_destroy_locked().
4065 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
4067 struct cgroup *parent = cgrp->parent;
4069 lockdep_assert_held(&cgroup_tree_mutex);
4070 lockdep_assert_held(&cgroup_mutex);
4072 /* delete this cgroup from parent->children */
4073 list_del_rcu(&cgrp->sibling);
4077 set_bit(CGRP_RELEASABLE, &parent->flags);
4078 check_for_release(parent);
4081 static int cgroup_rmdir(struct kernfs_node *kn)
4083 struct cgroup *cgrp = kn->priv;
4087 * This is self-destruction but @kn can't be removed while this
4088 * callback is in progress. Let's break active protection. Once
4089 * the protection is broken, @cgrp can be destroyed at any point.
4090 * Pin it so that it stays accessible.
4093 kernfs_break_active_protection(kn);
4095 mutex_lock(&cgroup_tree_mutex);
4096 mutex_lock(&cgroup_mutex);
4099 * @cgrp might already have been destroyed while we're trying to
4102 if (!cgroup_is_dead(cgrp))
4103 ret = cgroup_destroy_locked(cgrp);
4105 mutex_unlock(&cgroup_mutex);
4106 mutex_unlock(&cgroup_tree_mutex);
4108 kernfs_unbreak_active_protection(kn);
4113 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4114 .remount_fs = cgroup_remount,
4115 .show_options = cgroup_show_options,
4116 .mkdir = cgroup_mkdir,
4117 .rmdir = cgroup_rmdir,
4118 .rename = cgroup_rename,
4121 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4123 struct cgroup_subsys_state *css;
4125 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4127 mutex_lock(&cgroup_tree_mutex);
4128 mutex_lock(&cgroup_mutex);
4130 INIT_LIST_HEAD(&ss->cfts);
4132 /* Create the root cgroup state for this subsystem */
4133 ss->root = &cgrp_dfl_root;
4134 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4135 /* We don't handle early failures gracefully */
4136 BUG_ON(IS_ERR(css));
4137 init_css(css, ss, &cgrp_dfl_root.cgrp);
4139 /* Update the init_css_set to contain a subsys
4140 * pointer to this state - since the subsystem is
4141 * newly registered, all tasks and hence the
4142 * init_css_set is in the subsystem's root cgroup. */
4143 init_css_set.subsys[ss->id] = css;
4145 need_forkexit_callback |= ss->fork || ss->exit;
4147 /* At system boot, before all subsystems have been
4148 * registered, no tasks have been forked, so we don't
4149 * need to invoke fork callbacks here. */
4150 BUG_ON(!list_empty(&init_task.tasks));
4152 BUG_ON(online_css(css));
4154 cgrp_dfl_root.cgrp.subsys_mask |= 1 << ss->id;
4156 mutex_unlock(&cgroup_mutex);
4157 mutex_unlock(&cgroup_tree_mutex);
4161 * cgroup_init_early - cgroup initialization at system boot
4163 * Initialize cgroups at system boot, and initialize any
4164 * subsystems that request early init.
4166 int __init cgroup_init_early(void)
4168 static struct cgroup_sb_opts __initdata opts =
4169 { .flags = CGRP_ROOT_SANE_BEHAVIOR };
4170 struct cgroup_subsys *ss;
4173 init_cgroup_root(&cgrp_dfl_root, &opts);
4174 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4176 for_each_subsys(ss, i) {
4177 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4178 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4179 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4181 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4182 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4185 ss->name = cgroup_subsys_name[i];
4188 cgroup_init_subsys(ss);
4194 * cgroup_init - cgroup initialization
4196 * Register cgroup filesystem and /proc file, and initialize
4197 * any subsystems that didn't request early init.
4199 int __init cgroup_init(void)
4201 struct cgroup_subsys *ss;
4205 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4207 mutex_lock(&cgroup_tree_mutex);
4208 mutex_lock(&cgroup_mutex);
4210 /* Add init_css_set to the hash table */
4211 key = css_set_hash(init_css_set.subsys);
4212 hash_add(css_set_table, &init_css_set.hlist, key);
4214 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4216 mutex_unlock(&cgroup_mutex);
4217 mutex_unlock(&cgroup_tree_mutex);
4219 for_each_subsys(ss, ssid) {
4220 if (!ss->early_init)
4221 cgroup_init_subsys(ss);
4224 * cftype registration needs kmalloc and can't be done
4225 * during early_init. Register base cftypes separately.
4227 if (ss->base_cftypes)
4228 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4231 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4235 err = register_filesystem(&cgroup_fs_type);
4237 kobject_put(cgroup_kobj);
4241 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4245 static int __init cgroup_wq_init(void)
4248 * There isn't much point in executing destruction path in
4249 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4250 * Use 1 for @max_active.
4252 * We would prefer to do this in cgroup_init() above, but that
4253 * is called before init_workqueues(): so leave this until after.
4255 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
4256 BUG_ON(!cgroup_destroy_wq);
4259 * Used to destroy pidlists and separate to serve as flush domain.
4260 * Cap @max_active to 1 too.
4262 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4264 BUG_ON(!cgroup_pidlist_destroy_wq);
4268 core_initcall(cgroup_wq_init);
4271 * proc_cgroup_show()
4272 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4273 * - Used for /proc/<pid>/cgroup.
4276 /* TODO: Use a proper seq_file iterator */
4277 int proc_cgroup_show(struct seq_file *m, void *v)
4280 struct task_struct *tsk;
4283 struct cgroup_root *root;
4286 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4292 tsk = get_pid_task(pid, PIDTYPE_PID);
4298 mutex_lock(&cgroup_mutex);
4299 down_read(&css_set_rwsem);
4301 for_each_root(root) {
4302 struct cgroup_subsys *ss;
4303 struct cgroup *cgrp;
4304 int ssid, count = 0;
4306 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
4309 seq_printf(m, "%d:", root->hierarchy_id);
4310 for_each_subsys(ss, ssid)
4311 if (root->cgrp.subsys_mask & (1 << ssid))
4312 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4313 if (strlen(root->name))
4314 seq_printf(m, "%sname=%s", count ? "," : "",
4317 cgrp = task_cgroup_from_root(tsk, root);
4318 path = cgroup_path(cgrp, buf, PATH_MAX);
4320 retval = -ENAMETOOLONG;
4328 up_read(&css_set_rwsem);
4329 mutex_unlock(&cgroup_mutex);
4330 put_task_struct(tsk);
4337 /* Display information about each subsystem and each hierarchy */
4338 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4340 struct cgroup_subsys *ss;
4343 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4345 * ideally we don't want subsystems moving around while we do this.
4346 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4347 * subsys/hierarchy state.
4349 mutex_lock(&cgroup_mutex);
4351 for_each_subsys(ss, i)
4352 seq_printf(m, "%s\t%d\t%d\t%d\n",
4353 ss->name, ss->root->hierarchy_id,
4354 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4356 mutex_unlock(&cgroup_mutex);
4360 static int cgroupstats_open(struct inode *inode, struct file *file)
4362 return single_open(file, proc_cgroupstats_show, NULL);
4365 static const struct file_operations proc_cgroupstats_operations = {
4366 .open = cgroupstats_open,
4368 .llseek = seq_lseek,
4369 .release = single_release,
4373 * cgroup_fork - initialize cgroup related fields during copy_process()
4374 * @child: pointer to task_struct of forking parent process.
4376 * A task is associated with the init_css_set until cgroup_post_fork()
4377 * attaches it to the parent's css_set. Empty cg_list indicates that
4378 * @child isn't holding reference to its css_set.
4380 void cgroup_fork(struct task_struct *child)
4382 RCU_INIT_POINTER(child->cgroups, &init_css_set);
4383 INIT_LIST_HEAD(&child->cg_list);
4387 * cgroup_post_fork - called on a new task after adding it to the task list
4388 * @child: the task in question
4390 * Adds the task to the list running through its css_set if necessary and
4391 * call the subsystem fork() callbacks. Has to be after the task is
4392 * visible on the task list in case we race with the first call to
4393 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4396 void cgroup_post_fork(struct task_struct *child)
4398 struct cgroup_subsys *ss;
4402 * This may race against cgroup_enable_task_cg_links(). As that
4403 * function sets use_task_css_set_links before grabbing
4404 * tasklist_lock and we just went through tasklist_lock to add
4405 * @child, it's guaranteed that either we see the set
4406 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
4407 * @child during its iteration.
4409 * If we won the race, @child is associated with %current's
4410 * css_set. Grabbing css_set_rwsem guarantees both that the
4411 * association is stable, and, on completion of the parent's
4412 * migration, @child is visible in the source of migration or
4413 * already in the destination cgroup. This guarantee is necessary
4414 * when implementing operations which need to migrate all tasks of
4415 * a cgroup to another.
4417 * Note that if we lose to cgroup_enable_task_cg_links(), @child
4418 * will remain in init_css_set. This is safe because all tasks are
4419 * in the init_css_set before cg_links is enabled and there's no
4420 * operation which transfers all tasks out of init_css_set.
4422 if (use_task_css_set_links) {
4423 struct css_set *cset;
4425 down_write(&css_set_rwsem);
4426 cset = task_css_set(current);
4427 if (list_empty(&child->cg_list)) {
4428 rcu_assign_pointer(child->cgroups, cset);
4429 list_add(&child->cg_list, &cset->tasks);
4432 up_write(&css_set_rwsem);
4436 * Call ss->fork(). This must happen after @child is linked on
4437 * css_set; otherwise, @child might change state between ->fork()
4438 * and addition to css_set.
4440 if (need_forkexit_callback) {
4441 for_each_subsys(ss, i)
4448 * cgroup_exit - detach cgroup from exiting task
4449 * @tsk: pointer to task_struct of exiting process
4451 * Description: Detach cgroup from @tsk and release it.
4453 * Note that cgroups marked notify_on_release force every task in
4454 * them to take the global cgroup_mutex mutex when exiting.
4455 * This could impact scaling on very large systems. Be reluctant to
4456 * use notify_on_release cgroups where very high task exit scaling
4457 * is required on large systems.
4459 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
4460 * call cgroup_exit() while the task is still competent to handle
4461 * notify_on_release(), then leave the task attached to the root cgroup in
4462 * each hierarchy for the remainder of its exit. No need to bother with
4463 * init_css_set refcnting. init_css_set never goes away and we can't race
4464 * with migration path - PF_EXITING is visible to migration path.
4466 void cgroup_exit(struct task_struct *tsk)
4468 struct cgroup_subsys *ss;
4469 struct css_set *cset;
4470 bool put_cset = false;
4474 * Unlink from @tsk from its css_set. As migration path can't race
4475 * with us, we can check cg_list without grabbing css_set_rwsem.
4477 if (!list_empty(&tsk->cg_list)) {
4478 down_write(&css_set_rwsem);
4479 list_del_init(&tsk->cg_list);
4480 up_write(&css_set_rwsem);
4484 /* Reassign the task to the init_css_set. */
4485 cset = task_css_set(tsk);
4486 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4488 if (need_forkexit_callback) {
4489 /* see cgroup_post_fork() for details */
4490 for_each_subsys(ss, i) {
4492 struct cgroup_subsys_state *old_css = cset->subsys[i];
4493 struct cgroup_subsys_state *css = task_css(tsk, i);
4495 ss->exit(css, old_css, tsk);
4501 put_css_set(cset, true);
4504 static void check_for_release(struct cgroup *cgrp)
4506 if (cgroup_is_releasable(cgrp) &&
4507 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4509 * Control Group is currently removeable. If it's not
4510 * already queued for a userspace notification, queue
4513 int need_schedule_work = 0;
4515 raw_spin_lock(&release_list_lock);
4516 if (!cgroup_is_dead(cgrp) &&
4517 list_empty(&cgrp->release_list)) {
4518 list_add(&cgrp->release_list, &release_list);
4519 need_schedule_work = 1;
4521 raw_spin_unlock(&release_list_lock);
4522 if (need_schedule_work)
4523 schedule_work(&release_agent_work);
4528 * Notify userspace when a cgroup is released, by running the
4529 * configured release agent with the name of the cgroup (path
4530 * relative to the root of cgroup file system) as the argument.
4532 * Most likely, this user command will try to rmdir this cgroup.
4534 * This races with the possibility that some other task will be
4535 * attached to this cgroup before it is removed, or that some other
4536 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4537 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4538 * unused, and this cgroup will be reprieved from its death sentence,
4539 * to continue to serve a useful existence. Next time it's released,
4540 * we will get notified again, if it still has 'notify_on_release' set.
4542 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4543 * means only wait until the task is successfully execve()'d. The
4544 * separate release agent task is forked by call_usermodehelper(),
4545 * then control in this thread returns here, without waiting for the
4546 * release agent task. We don't bother to wait because the caller of
4547 * this routine has no use for the exit status of the release agent
4548 * task, so no sense holding our caller up for that.
4550 static void cgroup_release_agent(struct work_struct *work)
4552 BUG_ON(work != &release_agent_work);
4553 mutex_lock(&cgroup_mutex);
4554 raw_spin_lock(&release_list_lock);
4555 while (!list_empty(&release_list)) {
4556 char *argv[3], *envp[3];
4558 char *pathbuf = NULL, *agentbuf = NULL, *path;
4559 struct cgroup *cgrp = list_entry(release_list.next,
4562 list_del_init(&cgrp->release_list);
4563 raw_spin_unlock(&release_list_lock);
4564 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
4567 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
4570 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4575 argv[i++] = agentbuf;
4580 /* minimal command environment */
4581 envp[i++] = "HOME=/";
4582 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4585 /* Drop the lock while we invoke the usermode helper,
4586 * since the exec could involve hitting disk and hence
4587 * be a slow process */
4588 mutex_unlock(&cgroup_mutex);
4589 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4590 mutex_lock(&cgroup_mutex);
4594 raw_spin_lock(&release_list_lock);
4596 raw_spin_unlock(&release_list_lock);
4597 mutex_unlock(&cgroup_mutex);
4600 static int __init cgroup_disable(char *str)
4602 struct cgroup_subsys *ss;
4606 while ((token = strsep(&str, ",")) != NULL) {
4610 for_each_subsys(ss, i) {
4611 if (!strcmp(token, ss->name)) {
4613 printk(KERN_INFO "Disabling %s control group"
4614 " subsystem\n", ss->name);
4621 __setup("cgroup_disable=", cgroup_disable);
4624 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
4625 * @dentry: directory dentry of interest
4626 * @ss: subsystem of interest
4628 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
4629 * to get the corresponding css and return it. If such css doesn't exist
4630 * or can't be pinned, an ERR_PTR value is returned.
4632 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
4633 struct cgroup_subsys *ss)
4635 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4636 struct cgroup_subsys_state *css = NULL;
4637 struct cgroup *cgrp;
4639 /* is @dentry a cgroup dir? */
4640 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4641 kernfs_type(kn) != KERNFS_DIR)
4642 return ERR_PTR(-EBADF);
4647 * This path doesn't originate from kernfs and @kn could already
4648 * have been or be removed at any point. @kn->priv is RCU
4649 * protected for this access. See destroy_locked() for details.
4651 cgrp = rcu_dereference(kn->priv);
4653 css = cgroup_css(cgrp, ss);
4655 if (!css || !css_tryget(css))
4656 css = ERR_PTR(-ENOENT);
4663 * css_from_id - lookup css by id
4664 * @id: the cgroup id
4665 * @ss: cgroup subsys to be looked into
4667 * Returns the css if there's valid one with @id, otherwise returns NULL.
4668 * Should be called under rcu_read_lock().
4670 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
4672 struct cgroup *cgrp;
4674 cgroup_assert_mutexes_or_rcu_locked();
4676 cgrp = idr_find(&ss->root->cgroup_idr, id);
4678 return cgroup_css(cgrp, ss);
4682 #ifdef CONFIG_CGROUP_DEBUG
4683 static struct cgroup_subsys_state *
4684 debug_css_alloc(struct cgroup_subsys_state *parent_css)
4686 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4689 return ERR_PTR(-ENOMEM);
4694 static void debug_css_free(struct cgroup_subsys_state *css)
4699 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
4702 return cgroup_task_count(css->cgroup);
4705 static u64 current_css_set_read(struct cgroup_subsys_state *css,
4708 return (u64)(unsigned long)current->cgroups;
4711 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
4717 count = atomic_read(&task_css_set(current)->refcount);
4722 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
4724 struct cgrp_cset_link *link;
4725 struct css_set *cset;
4728 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
4732 down_read(&css_set_rwsem);
4734 cset = rcu_dereference(current->cgroups);
4735 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
4736 struct cgroup *c = link->cgrp;
4738 cgroup_name(c, name_buf, NAME_MAX + 1);
4739 seq_printf(seq, "Root %d group %s\n",
4740 c->root->hierarchy_id, name_buf);
4743 up_read(&css_set_rwsem);
4748 #define MAX_TASKS_SHOWN_PER_CSS 25
4749 static int cgroup_css_links_read(struct seq_file *seq, void *v)
4751 struct cgroup_subsys_state *css = seq_css(seq);
4752 struct cgrp_cset_link *link;
4754 down_read(&css_set_rwsem);
4755 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
4756 struct css_set *cset = link->cset;
4757 struct task_struct *task;
4760 seq_printf(seq, "css_set %p\n", cset);
4762 list_for_each_entry(task, &cset->tasks, cg_list) {
4763 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
4765 seq_printf(seq, " task %d\n", task_pid_vnr(task));
4768 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
4769 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
4771 seq_printf(seq, " task %d\n", task_pid_vnr(task));
4775 seq_puts(seq, " ...\n");
4777 up_read(&css_set_rwsem);
4781 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
4783 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4786 static struct cftype debug_files[] = {
4788 .name = "taskcount",
4789 .read_u64 = debug_taskcount_read,
4793 .name = "current_css_set",
4794 .read_u64 = current_css_set_read,
4798 .name = "current_css_set_refcount",
4799 .read_u64 = current_css_set_refcount_read,
4803 .name = "current_css_set_cg_links",
4804 .seq_show = current_css_set_cg_links_read,
4808 .name = "cgroup_css_links",
4809 .seq_show = cgroup_css_links_read,
4813 .name = "releasable",
4814 .read_u64 = releasable_read,
4820 struct cgroup_subsys debug_cgrp_subsys = {
4821 .css_alloc = debug_css_alloc,
4822 .css_free = debug_css_free,
4823 .base_cftypes = debug_files,
4825 #endif /* CONFIG_CGROUP_DEBUG */