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>
37 #include <linux/mutex.h>
38 #include <linux/mount.h>
39 #include <linux/pagemap.h>
40 #include <linux/proc_fs.h>
41 #include <linux/rcupdate.h>
42 #include <linux/sched.h>
43 #include <linux/slab.h>
44 #include <linux/spinlock.h>
45 #include <linux/string.h>
46 #include <linux/sort.h>
47 #include <linux/kmod.h>
48 #include <linux/delayacct.h>
49 #include <linux/cgroupstats.h>
50 #include <linux/hashtable.h>
51 #include <linux/pid_namespace.h>
52 #include <linux/idr.h>
53 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
54 #include <linux/flex_array.h> /* used in cgroup_attach_task */
55 #include <linux/kthread.h>
56 #include <linux/delay.h>
58 #include <linux/atomic.h>
61 * pidlists linger the following amount before being destroyed. The goal
62 * is avoiding frequent destruction in the middle of consecutive read calls
63 * Expiring in the middle is a performance problem not a correctness one.
64 * 1 sec should be enough.
66 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
68 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
72 * cgroup_tree_mutex nests above cgroup_mutex and protects cftypes, file
73 * creation/removal and hierarchy changing operations including cgroup
74 * creation, removal, css association and controller rebinding. This outer
75 * lock is needed mainly to resolve the circular dependency between kernfs
76 * active ref and cgroup_mutex. cgroup_tree_mutex nests above both.
78 static DEFINE_MUTEX(cgroup_tree_mutex);
81 * cgroup_mutex is the master lock. Any modification to cgroup or its
82 * hierarchy must be performed while holding it.
84 #ifdef CONFIG_PROVE_RCU
85 DEFINE_MUTEX(cgroup_mutex);
86 EXPORT_SYMBOL_GPL(cgroup_mutex); /* only for lockdep */
88 static DEFINE_MUTEX(cgroup_mutex);
92 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
93 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
95 static DEFINE_SPINLOCK(release_agent_path_lock);
97 #define cgroup_assert_mutexes_or_rcu_locked() \
98 rcu_lockdep_assert(rcu_read_lock_held() || \
99 lockdep_is_held(&cgroup_tree_mutex) || \
100 lockdep_is_held(&cgroup_mutex), \
101 "cgroup_[tree_]mutex or RCU read lock required");
104 * cgroup destruction makes heavy use of work items and there can be a lot
105 * of concurrent destructions. Use a separate workqueue so that cgroup
106 * destruction work items don't end up filling up max_active of system_wq
107 * which may lead to deadlock.
109 static struct workqueue_struct *cgroup_destroy_wq;
112 * pidlist destructions need to be flushed on cgroup destruction. Use a
113 * separate workqueue as flush domain.
115 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
117 /* generate an array of cgroup subsystem pointers */
118 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
119 static struct cgroup_subsys *cgroup_subsys[] = {
120 #include <linux/cgroup_subsys.h>
124 /* array of cgroup subsystem names */
125 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
126 static const char *cgroup_subsys_name[] = {
127 #include <linux/cgroup_subsys.h>
132 * The dummy hierarchy, reserved for the subsystems that are otherwise
133 * unattached - it never has more than a single cgroup, and all tasks are
134 * part of that cgroup.
136 static struct cgroupfs_root cgroup_dummy_root;
138 /* dummy_top is a shorthand for the dummy hierarchy's top cgroup */
139 static struct cgroup * const cgroup_dummy_top = &cgroup_dummy_root.top_cgroup;
141 /* The list of hierarchy roots */
143 static LIST_HEAD(cgroup_roots);
144 static int cgroup_root_count;
146 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
147 static DEFINE_IDR(cgroup_hierarchy_idr);
150 * Assign a monotonically increasing serial number to cgroups. It
151 * guarantees cgroups with bigger numbers are newer than those with smaller
152 * numbers. Also, as cgroups are always appended to the parent's
153 * ->children list, it guarantees that sibling cgroups are always sorted in
154 * the ascending serial number order on the list. Protected by
157 static u64 cgroup_serial_nr_next = 1;
159 /* This flag indicates whether tasks in the fork and exit paths should
160 * check for fork/exit handlers to call. This avoids us having to do
161 * extra work in the fork/exit path if none of the subsystems need to
164 static int need_forkexit_callback __read_mostly;
166 static struct cftype cgroup_base_files[];
168 static void cgroup_put(struct cgroup *cgrp);
169 static int rebind_subsystems(struct cgroupfs_root *root,
170 unsigned long added_mask, unsigned removed_mask);
171 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
172 static int cgroup_destroy_locked(struct cgroup *cgrp);
173 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
175 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
178 * cgroup_css - obtain a cgroup's css for the specified subsystem
179 * @cgrp: the cgroup of interest
180 * @ss: the subsystem of interest (%NULL returns the dummy_css)
182 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
183 * function must be called either under cgroup_mutex or rcu_read_lock() and
184 * the caller is responsible for pinning the returned css if it wants to
185 * keep accessing it outside the said locks. This function may return
186 * %NULL if @cgrp doesn't have @subsys_id enabled.
188 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
189 struct cgroup_subsys *ss)
192 return rcu_dereference_check(cgrp->subsys[ss->id],
193 lockdep_is_held(&cgroup_tree_mutex) ||
194 lockdep_is_held(&cgroup_mutex));
196 return &cgrp->dummy_css;
199 /* convenient tests for these bits */
200 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
202 return test_bit(CGRP_DEAD, &cgrp->flags);
205 struct cgroup_subsys_state *seq_css(struct seq_file *seq)
207 struct kernfs_open_file *of = seq->private;
208 struct cgroup *cgrp = of->kn->parent->priv;
209 struct cftype *cft = seq_cft(seq);
212 * This is open and unprotected implementation of cgroup_css().
213 * seq_css() is only called from a kernfs file operation which has
214 * an active reference on the file. Because all the subsystem
215 * files are drained before a css is disassociated with a cgroup,
216 * the matching css from the cgroup's subsys table is guaranteed to
217 * be and stay valid until the enclosing operation is complete.
220 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
222 return &cgrp->dummy_css;
224 EXPORT_SYMBOL_GPL(seq_css);
227 * cgroup_is_descendant - test ancestry
228 * @cgrp: the cgroup to be tested
229 * @ancestor: possible ancestor of @cgrp
231 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
232 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
233 * and @ancestor are accessible.
235 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
238 if (cgrp == ancestor)
244 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
246 static int cgroup_is_releasable(const struct cgroup *cgrp)
249 (1 << CGRP_RELEASABLE) |
250 (1 << CGRP_NOTIFY_ON_RELEASE);
251 return (cgrp->flags & bits) == bits;
254 static int notify_on_release(const struct cgroup *cgrp)
256 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
260 * for_each_css - iterate all css's of a cgroup
261 * @css: the iteration cursor
262 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
263 * @cgrp: the target cgroup to iterate css's of
265 * Should be called under cgroup_mutex.
267 #define for_each_css(css, ssid, cgrp) \
268 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
269 if (!((css) = rcu_dereference_check( \
270 (cgrp)->subsys[(ssid)], \
271 lockdep_is_held(&cgroup_tree_mutex) || \
272 lockdep_is_held(&cgroup_mutex)))) { } \
276 * for_each_subsys - iterate all enabled cgroup subsystems
277 * @ss: the iteration cursor
278 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
280 #define for_each_subsys(ss, ssid) \
281 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
282 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
284 /* iterate across the active hierarchies */
285 #define for_each_active_root(root) \
286 list_for_each_entry((root), &cgroup_roots, root_list)
289 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
290 * @cgrp: the cgroup to be checked for liveness
292 * On success, returns true; the mutex should be later unlocked. On
293 * failure returns false with no lock held.
295 static bool cgroup_lock_live_group(struct cgroup *cgrp)
297 mutex_lock(&cgroup_mutex);
298 if (cgroup_is_dead(cgrp)) {
299 mutex_unlock(&cgroup_mutex);
305 /* the list of cgroups eligible for automatic release. Protected by
306 * release_list_lock */
307 static LIST_HEAD(release_list);
308 static DEFINE_RAW_SPINLOCK(release_list_lock);
309 static void cgroup_release_agent(struct work_struct *work);
310 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
311 static void check_for_release(struct cgroup *cgrp);
314 * A cgroup can be associated with multiple css_sets as different tasks may
315 * belong to different cgroups on different hierarchies. In the other
316 * direction, a css_set is naturally associated with multiple cgroups.
317 * This M:N relationship is represented by the following link structure
318 * which exists for each association and allows traversing the associations
321 struct cgrp_cset_link {
322 /* the cgroup and css_set this link associates */
324 struct css_set *cset;
326 /* list of cgrp_cset_links anchored at cgrp->cset_links */
327 struct list_head cset_link;
329 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
330 struct list_head cgrp_link;
333 /* The default css_set - used by init and its children prior to any
334 * hierarchies being mounted. It contains a pointer to the root state
335 * for each subsystem. Also used to anchor the list of css_sets. Not
336 * reference-counted, to improve performance when child cgroups
337 * haven't been created.
340 static struct css_set init_css_set;
341 static struct cgrp_cset_link init_cgrp_cset_link;
344 * css_set_lock protects the list of css_set objects, and the chain of
345 * tasks off each css_set. Nests outside task->alloc_lock due to
346 * css_task_iter_start().
348 static DEFINE_RWLOCK(css_set_lock);
349 static int css_set_count;
352 * hash table for cgroup groups. This improves the performance to find
353 * an existing css_set. This hash doesn't (currently) take into
354 * account cgroups in empty hierarchies.
356 #define CSS_SET_HASH_BITS 7
357 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
359 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
361 unsigned long key = 0UL;
362 struct cgroup_subsys *ss;
365 for_each_subsys(ss, i)
366 key += (unsigned long)css[i];
367 key = (key >> 16) ^ key;
373 * We don't maintain the lists running through each css_set to its task
374 * until after the first call to css_task_iter_start(). This reduces the
375 * fork()/exit() overhead for people who have cgroups compiled into their
376 * kernel but not actually in use.
378 static int use_task_css_set_links __read_mostly;
380 static void __put_css_set(struct css_set *cset, int taskexit)
382 struct cgrp_cset_link *link, *tmp_link;
385 * Ensure that the refcount doesn't hit zero while any readers
386 * can see it. Similar to atomic_dec_and_lock(), but for an
389 if (atomic_add_unless(&cset->refcount, -1, 1))
391 write_lock(&css_set_lock);
392 if (!atomic_dec_and_test(&cset->refcount)) {
393 write_unlock(&css_set_lock);
397 /* This css_set is dead. unlink it and release cgroup refcounts */
398 hash_del(&cset->hlist);
401 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
402 struct cgroup *cgrp = link->cgrp;
404 list_del(&link->cset_link);
405 list_del(&link->cgrp_link);
407 /* @cgrp can't go away while we're holding css_set_lock */
408 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
410 set_bit(CGRP_RELEASABLE, &cgrp->flags);
411 check_for_release(cgrp);
417 write_unlock(&css_set_lock);
418 kfree_rcu(cset, rcu_head);
422 * refcounted get/put for css_set objects
424 static inline void get_css_set(struct css_set *cset)
426 atomic_inc(&cset->refcount);
429 static inline void put_css_set(struct css_set *cset)
431 __put_css_set(cset, 0);
434 static inline void put_css_set_taskexit(struct css_set *cset)
436 __put_css_set(cset, 1);
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 cgroupfs_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->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 read_lock(&css_set_lock);
639 cset = find_existing_css_set(old_cset, cgrp, template);
642 read_unlock(&css_set_lock);
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_HLIST_NODE(&cset->hlist);
662 /* Copy the set of subsystem state objects generated in
663 * find_existing_css_set() */
664 memcpy(cset->subsys, template, sizeof(cset->subsys));
666 write_lock(&css_set_lock);
667 /* Add reference counts and links from the new css_set. */
668 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
669 struct cgroup *c = link->cgrp;
671 if (c->root == cgrp->root)
673 link_css_set(&tmp_links, cset, c);
676 BUG_ON(!list_empty(&tmp_links));
680 /* Add this cgroup group to the hash table */
681 key = css_set_hash(cset->subsys);
682 hash_add(css_set_table, &cset->hlist, key);
684 write_unlock(&css_set_lock);
689 static struct cgroupfs_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
691 struct cgroup *top_cgrp = kf_root->kn->priv;
693 return top_cgrp->root;
696 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
700 lockdep_assert_held(&cgroup_mutex);
702 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
707 root->hierarchy_id = id;
711 static void cgroup_exit_root_id(struct cgroupfs_root *root)
713 lockdep_assert_held(&cgroup_mutex);
715 if (root->hierarchy_id) {
716 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
717 root->hierarchy_id = 0;
721 static void cgroup_free_root(struct cgroupfs_root *root)
724 /* hierarhcy ID shoulid already have been released */
725 WARN_ON_ONCE(root->hierarchy_id);
727 idr_destroy(&root->cgroup_idr);
732 static void cgroup_destroy_root(struct cgroupfs_root *root)
734 struct cgroup *cgrp = &root->top_cgroup;
735 struct cgrp_cset_link *link, *tmp_link;
738 mutex_lock(&cgroup_tree_mutex);
739 mutex_lock(&cgroup_mutex);
741 BUG_ON(atomic_read(&root->nr_cgrps));
742 BUG_ON(!list_empty(&cgrp->children));
744 /* Rebind all subsystems back to the default hierarchy */
745 if (root->flags & CGRP_ROOT_SUBSYS_BOUND) {
746 ret = rebind_subsystems(root, 0, root->subsys_mask);
747 /* Shouldn't be able to fail ... */
752 * Release all the links from cset_links to this hierarchy's
755 write_lock(&css_set_lock);
757 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
758 list_del(&link->cset_link);
759 list_del(&link->cgrp_link);
762 write_unlock(&css_set_lock);
764 if (!list_empty(&root->root_list)) {
765 list_del(&root->root_list);
769 cgroup_exit_root_id(root);
771 mutex_unlock(&cgroup_mutex);
772 mutex_unlock(&cgroup_tree_mutex);
774 kernfs_destroy_root(root->kf_root);
775 cgroup_free_root(root);
779 * Return the cgroup for "task" from the given hierarchy. Must be
780 * called with cgroup_mutex held.
782 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
783 struct cgroupfs_root *root)
785 struct css_set *cset;
786 struct cgroup *res = NULL;
788 BUG_ON(!mutex_is_locked(&cgroup_mutex));
789 read_lock(&css_set_lock);
791 * No need to lock the task - since we hold cgroup_mutex the
792 * task can't change groups, so the only thing that can happen
793 * is that it exits and its css is set back to init_css_set.
795 cset = task_css_set(task);
796 if (cset == &init_css_set) {
797 res = &root->top_cgroup;
799 struct cgrp_cset_link *link;
801 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
802 struct cgroup *c = link->cgrp;
804 if (c->root == root) {
810 read_unlock(&css_set_lock);
816 * There is one global cgroup mutex. We also require taking
817 * task_lock() when dereferencing a task's cgroup subsys pointers.
818 * See "The task_lock() exception", at the end of this comment.
820 * A task must hold cgroup_mutex to modify cgroups.
822 * Any task can increment and decrement the count field without lock.
823 * So in general, code holding cgroup_mutex can't rely on the count
824 * field not changing. However, if the count goes to zero, then only
825 * cgroup_attach_task() can increment it again. Because a count of zero
826 * means that no tasks are currently attached, therefore there is no
827 * way a task attached to that cgroup can fork (the other way to
828 * increment the count). So code holding cgroup_mutex can safely
829 * assume that if the count is zero, it will stay zero. Similarly, if
830 * a task holds cgroup_mutex on a cgroup with zero count, it
831 * knows that the cgroup won't be removed, as cgroup_rmdir()
834 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
835 * (usually) take cgroup_mutex. These are the two most performance
836 * critical pieces of code here. The exception occurs on cgroup_exit(),
837 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
838 * is taken, and if the cgroup count is zero, a usermode call made
839 * to the release agent with the name of the cgroup (path relative to
840 * the root of cgroup file system) as the argument.
842 * A cgroup can only be deleted if both its 'count' of using tasks
843 * is zero, and its list of 'children' cgroups is empty. Since all
844 * tasks in the system use _some_ cgroup, and since there is always at
845 * least one task in the system (init, pid == 1), therefore, top_cgroup
846 * always has either children cgroups and/or using tasks. So we don't
847 * need a special hack to ensure that top_cgroup cannot be deleted.
849 * The task_lock() exception
851 * The need for this exception arises from the action of
852 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
853 * another. It does so using cgroup_mutex, however there are
854 * several performance critical places that need to reference
855 * task->cgroup without the expense of grabbing a system global
856 * mutex. Therefore except as noted below, when dereferencing or, as
857 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
858 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
859 * the task_struct routinely used for such matters.
861 * P.S. One more locking exception. RCU is used to guard the
862 * update of a tasks cgroup pointer by cgroup_attach_task()
865 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
866 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
867 static const struct file_operations proc_cgroupstats_operations;
869 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
872 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
873 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
874 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
875 cft->ss->name, cft->name);
877 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
882 * cgroup_file_mode - deduce file mode of a control file
883 * @cft: the control file in question
885 * returns cft->mode if ->mode is not 0
886 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
887 * returns S_IRUGO if it has only a read handler
888 * returns S_IWUSR if it has only a write hander
890 static umode_t cgroup_file_mode(const struct cftype *cft)
897 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
900 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
907 static void cgroup_free_fn(struct work_struct *work)
909 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
911 atomic_dec(&cgrp->root->nr_cgrps);
912 cgroup_pidlist_destroy_all(cgrp);
916 * We get a ref to the parent, and put the ref when this
917 * cgroup is being freed, so it's guaranteed that the
918 * parent won't be destroyed before its children.
920 cgroup_put(cgrp->parent);
921 kernfs_put(cgrp->kn);
925 * This is top cgroup's refcnt reaching zero, which
926 * indicates that the root should be released.
928 cgroup_destroy_root(cgrp->root);
932 static void cgroup_free_rcu(struct rcu_head *head)
934 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
936 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
937 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
940 static void cgroup_get(struct cgroup *cgrp)
942 WARN_ON_ONCE(cgroup_is_dead(cgrp));
943 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
944 atomic_inc(&cgrp->refcnt);
947 static void cgroup_put(struct cgroup *cgrp)
949 if (!atomic_dec_and_test(&cgrp->refcnt))
951 if (WARN_ON_ONCE(cgrp->parent && !cgroup_is_dead(cgrp)))
955 * XXX: cgrp->id is only used to look up css's. As cgroup and
956 * css's lifetimes will be decoupled, it should be made
957 * per-subsystem and moved to css->id so that lookups are
958 * successful until the target css is released.
960 mutex_lock(&cgroup_mutex);
961 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
962 mutex_unlock(&cgroup_mutex);
965 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
968 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
970 char name[CGROUP_FILE_NAME_MAX];
972 lockdep_assert_held(&cgroup_tree_mutex);
973 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
977 * cgroup_clear_dir - remove subsys files in a cgroup directory
978 * @cgrp: target cgroup
979 * @subsys_mask: mask of the subsystem ids whose files should be removed
981 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
983 struct cgroup_subsys *ss;
986 for_each_subsys(ss, i) {
989 if (!test_bit(i, &subsys_mask))
991 list_for_each_entry(cfts, &ss->cfts, node)
992 cgroup_addrm_files(cgrp, cfts, false);
996 static int rebind_subsystems(struct cgroupfs_root *root,
997 unsigned long added_mask, unsigned removed_mask)
999 struct cgroup *cgrp = &root->top_cgroup;
1000 struct cgroup_subsys *ss;
1003 lockdep_assert_held(&cgroup_tree_mutex);
1004 lockdep_assert_held(&cgroup_mutex);
1006 /* Check that any added subsystems are currently free */
1007 for_each_subsys(ss, i)
1008 if ((added_mask & (1 << i)) && ss->root != &cgroup_dummy_root)
1011 ret = cgroup_populate_dir(cgrp, added_mask);
1016 * Nothing can fail from this point on. Remove files for the
1017 * removed subsystems and rebind each subsystem.
1019 mutex_unlock(&cgroup_mutex);
1020 cgroup_clear_dir(cgrp, removed_mask);
1021 mutex_lock(&cgroup_mutex);
1023 for_each_subsys(ss, i) {
1024 unsigned long bit = 1UL << i;
1026 if (bit & added_mask) {
1027 /* We're binding this subsystem to this hierarchy */
1028 BUG_ON(cgroup_css(cgrp, ss));
1029 BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
1030 BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
1032 rcu_assign_pointer(cgrp->subsys[i],
1033 cgroup_css(cgroup_dummy_top, ss));
1034 cgroup_css(cgrp, ss)->cgroup = cgrp;
1038 ss->bind(cgroup_css(cgrp, ss));
1040 /* refcount was already taken, and we're keeping it */
1041 root->subsys_mask |= bit;
1042 } else if (bit & removed_mask) {
1043 /* We're removing this subsystem */
1044 BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1045 BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1048 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1050 cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1051 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1053 cgroup_subsys[i]->root = &cgroup_dummy_root;
1054 root->subsys_mask &= ~bit;
1059 * Mark @root has finished binding subsystems. @root->subsys_mask
1060 * now matches the bound subsystems.
1062 root->flags |= CGRP_ROOT_SUBSYS_BOUND;
1063 kernfs_activate(cgrp->kn);
1068 static int cgroup_show_options(struct seq_file *seq,
1069 struct kernfs_root *kf_root)
1071 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1072 struct cgroup_subsys *ss;
1075 for_each_subsys(ss, ssid)
1076 if (root->subsys_mask & (1 << ssid))
1077 seq_printf(seq, ",%s", ss->name);
1078 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1079 seq_puts(seq, ",sane_behavior");
1080 if (root->flags & CGRP_ROOT_NOPREFIX)
1081 seq_puts(seq, ",noprefix");
1082 if (root->flags & CGRP_ROOT_XATTR)
1083 seq_puts(seq, ",xattr");
1085 spin_lock(&release_agent_path_lock);
1086 if (strlen(root->release_agent_path))
1087 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1088 spin_unlock(&release_agent_path_lock);
1090 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1091 seq_puts(seq, ",clone_children");
1092 if (strlen(root->name))
1093 seq_printf(seq, ",name=%s", root->name);
1097 struct cgroup_sb_opts {
1098 unsigned long subsys_mask;
1099 unsigned long flags;
1100 char *release_agent;
1101 bool cpuset_clone_children;
1103 /* User explicitly requested empty subsystem */
1108 * Convert a hierarchy specifier into a bitmask of subsystems and
1109 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1110 * array. This function takes refcounts on subsystems to be used, unless it
1111 * returns error, in which case no refcounts are taken.
1113 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1115 char *token, *o = data;
1116 bool all_ss = false, one_ss = false;
1117 unsigned long mask = (unsigned long)-1;
1118 struct cgroup_subsys *ss;
1121 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1123 #ifdef CONFIG_CPUSETS
1124 mask = ~(1UL << cpuset_cgrp_id);
1127 memset(opts, 0, sizeof(*opts));
1129 while ((token = strsep(&o, ",")) != NULL) {
1132 if (!strcmp(token, "none")) {
1133 /* Explicitly have no subsystems */
1137 if (!strcmp(token, "all")) {
1138 /* Mutually exclusive option 'all' + subsystem name */
1144 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1145 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1148 if (!strcmp(token, "noprefix")) {
1149 opts->flags |= CGRP_ROOT_NOPREFIX;
1152 if (!strcmp(token, "clone_children")) {
1153 opts->cpuset_clone_children = true;
1156 if (!strcmp(token, "xattr")) {
1157 opts->flags |= CGRP_ROOT_XATTR;
1160 if (!strncmp(token, "release_agent=", 14)) {
1161 /* Specifying two release agents is forbidden */
1162 if (opts->release_agent)
1164 opts->release_agent =
1165 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1166 if (!opts->release_agent)
1170 if (!strncmp(token, "name=", 5)) {
1171 const char *name = token + 5;
1172 /* Can't specify an empty name */
1175 /* Must match [\w.-]+ */
1176 for (i = 0; i < strlen(name); i++) {
1180 if ((c == '.') || (c == '-') || (c == '_'))
1184 /* Specifying two names is forbidden */
1187 opts->name = kstrndup(name,
1188 MAX_CGROUP_ROOT_NAMELEN - 1,
1196 for_each_subsys(ss, i) {
1197 if (strcmp(token, ss->name))
1202 /* Mutually exclusive option 'all' + subsystem name */
1205 set_bit(i, &opts->subsys_mask);
1210 if (i == CGROUP_SUBSYS_COUNT)
1215 * If the 'all' option was specified select all the subsystems,
1216 * otherwise if 'none', 'name=' and a subsystem name options
1217 * were not specified, let's default to 'all'
1219 if (all_ss || (!one_ss && !opts->none && !opts->name))
1220 for_each_subsys(ss, i)
1222 set_bit(i, &opts->subsys_mask);
1224 /* Consistency checks */
1226 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1227 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1229 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1230 opts->cpuset_clone_children || opts->release_agent ||
1232 pr_err("cgroup: sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1238 * Option noprefix was introduced just for backward compatibility
1239 * with the old cpuset, so we allow noprefix only if mounting just
1240 * the cpuset subsystem.
1242 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1246 /* Can't specify "none" and some subsystems */
1247 if (opts->subsys_mask && opts->none)
1251 * We either have to specify by name or by subsystems. (So all
1252 * empty hierarchies must have a name).
1254 if (!opts->subsys_mask && !opts->name)
1260 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1263 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1264 struct cgroup_sb_opts opts;
1265 unsigned long added_mask, removed_mask;
1267 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1268 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1272 mutex_lock(&cgroup_tree_mutex);
1273 mutex_lock(&cgroup_mutex);
1275 /* See what subsystems are wanted */
1276 ret = parse_cgroupfs_options(data, &opts);
1280 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1281 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1282 task_tgid_nr(current), current->comm);
1284 added_mask = opts.subsys_mask & ~root->subsys_mask;
1285 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1287 /* Don't allow flags or name to change at remount */
1288 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1289 (opts.name && strcmp(opts.name, root->name))) {
1290 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1291 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1292 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1297 /* remounting is not allowed for populated hierarchies */
1298 if (!list_empty(&root->top_cgroup.children)) {
1303 ret = rebind_subsystems(root, added_mask, removed_mask);
1307 if (opts.release_agent) {
1308 spin_lock(&release_agent_path_lock);
1309 strcpy(root->release_agent_path, opts.release_agent);
1310 spin_unlock(&release_agent_path_lock);
1313 kfree(opts.release_agent);
1315 mutex_unlock(&cgroup_mutex);
1316 mutex_unlock(&cgroup_tree_mutex);
1320 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1322 atomic_set(&cgrp->refcnt, 1);
1323 INIT_LIST_HEAD(&cgrp->sibling);
1324 INIT_LIST_HEAD(&cgrp->children);
1325 INIT_LIST_HEAD(&cgrp->cset_links);
1326 INIT_LIST_HEAD(&cgrp->release_list);
1327 INIT_LIST_HEAD(&cgrp->pidlists);
1328 mutex_init(&cgrp->pidlist_mutex);
1329 cgrp->dummy_css.cgroup = cgrp;
1332 static void init_cgroup_root(struct cgroupfs_root *root)
1334 struct cgroup *cgrp = &root->top_cgroup;
1336 INIT_LIST_HEAD(&root->root_list);
1337 atomic_set(&root->nr_cgrps, 1);
1339 init_cgroup_housekeeping(cgrp);
1340 idr_init(&root->cgroup_idr);
1343 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1345 struct cgroupfs_root *root;
1347 if (!opts->subsys_mask && !opts->none)
1348 return ERR_PTR(-EINVAL);
1350 root = kzalloc(sizeof(*root), GFP_KERNEL);
1352 return ERR_PTR(-ENOMEM);
1354 init_cgroup_root(root);
1357 * We need to set @root->subsys_mask now so that @root can be
1358 * matched by cgroup_test_super() before it finishes
1359 * initialization; otherwise, competing mounts with the same
1360 * options may try to bind the same subsystems instead of waiting
1361 * for the first one leading to unexpected mount errors.
1362 * SUBSYS_BOUND will be set once actual binding is complete.
1364 root->subsys_mask = opts->subsys_mask;
1365 root->flags = opts->flags;
1366 if (opts->release_agent)
1367 strcpy(root->release_agent_path, opts->release_agent);
1369 strcpy(root->name, opts->name);
1370 if (opts->cpuset_clone_children)
1371 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1375 static int cgroup_setup_root(struct cgroupfs_root *root)
1377 LIST_HEAD(tmp_links);
1378 struct cgroup *root_cgrp = &root->top_cgroup;
1379 struct css_set *cset;
1382 lockdep_assert_held(&cgroup_tree_mutex);
1383 lockdep_assert_held(&cgroup_mutex);
1385 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1388 root_cgrp->id = ret;
1391 * We're accessing css_set_count without locking css_set_lock here,
1392 * but that's OK - it can only be increased by someone holding
1393 * cgroup_lock, and that's us. The worst that can happen is that we
1394 * have some link structures left over
1396 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1400 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1401 ret = cgroup_init_root_id(root, 2, 0);
1405 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1406 KERNFS_ROOT_CREATE_DEACTIVATED,
1408 if (IS_ERR(root->kf_root)) {
1409 ret = PTR_ERR(root->kf_root);
1412 root_cgrp->kn = root->kf_root->kn;
1414 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1418 ret = rebind_subsystems(root, root->subsys_mask, 0);
1423 * There must be no failure case after here, since rebinding takes
1424 * care of subsystems' refcounts, which are explicitly dropped in
1425 * the failure exit path.
1427 list_add(&root->root_list, &cgroup_roots);
1428 cgroup_root_count++;
1431 * Link the top cgroup in this hierarchy into all the css_set
1434 write_lock(&css_set_lock);
1435 hash_for_each(css_set_table, i, cset, hlist)
1436 link_css_set(&tmp_links, cset, root_cgrp);
1437 write_unlock(&css_set_lock);
1439 BUG_ON(!list_empty(&root_cgrp->children));
1440 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1442 kernfs_activate(root_cgrp->kn);
1447 kernfs_destroy_root(root->kf_root);
1448 root->kf_root = NULL;
1450 cgroup_exit_root_id(root);
1452 free_cgrp_cset_links(&tmp_links);
1456 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1457 int flags, const char *unused_dev_name,
1460 struct cgroupfs_root *root;
1461 struct cgroup_sb_opts opts;
1462 struct dentry *dentry;
1465 mutex_lock(&cgroup_tree_mutex);
1466 mutex_lock(&cgroup_mutex);
1468 /* First find the desired set of subsystems */
1469 ret = parse_cgroupfs_options(data, &opts);
1473 /* look for a matching existing root */
1474 for_each_active_root(root) {
1475 bool name_match = false;
1478 * If we asked for a name then it must match. Also, if
1479 * name matches but sybsys_mask doesn't, we should fail.
1480 * Remember whether name matched.
1483 if (strcmp(opts.name, root->name))
1489 * If we asked for subsystems (or explicitly for no
1490 * subsystems) then they must match.
1492 if ((opts.subsys_mask || opts.none) &&
1493 (opts.subsys_mask != root->subsys_mask)) {
1500 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1501 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1502 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1506 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1511 * A root's lifetime is governed by its top cgroup. Zero
1512 * ref indicate that the root is being destroyed. Wait for
1513 * destruction to complete so that the subsystems are free.
1514 * We can use wait_queue for the wait but this path is
1515 * super cold. Let's just sleep for a bit and retry.
1517 if (!atomic_inc_not_zero(&root->top_cgroup.refcnt)) {
1518 mutex_unlock(&cgroup_mutex);
1519 mutex_unlock(&cgroup_tree_mutex);
1528 /* no such thing, create a new one */
1529 root = cgroup_root_from_opts(&opts);
1531 ret = PTR_ERR(root);
1535 ret = cgroup_setup_root(root);
1537 cgroup_free_root(root);
1540 mutex_unlock(&cgroup_mutex);
1541 mutex_unlock(&cgroup_tree_mutex);
1543 kfree(opts.release_agent);
1547 return ERR_PTR(ret);
1549 dentry = kernfs_mount(fs_type, flags, root->kf_root);
1551 cgroup_put(&root->top_cgroup);
1555 static void cgroup_kill_sb(struct super_block *sb)
1557 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1558 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1560 cgroup_put(&root->top_cgroup);
1564 static struct file_system_type cgroup_fs_type = {
1566 .mount = cgroup_mount,
1567 .kill_sb = cgroup_kill_sb,
1570 static struct kobject *cgroup_kobj;
1573 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1574 * @task: target task
1575 * @buf: the buffer to write the path into
1576 * @buflen: the length of the buffer
1578 * Determine @task's cgroup on the first (the one with the lowest non-zero
1579 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1580 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1581 * cgroup controller callbacks.
1583 * Return value is the same as kernfs_path().
1585 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1587 struct cgroupfs_root *root;
1588 struct cgroup *cgrp;
1589 int hierarchy_id = 1;
1592 mutex_lock(&cgroup_mutex);
1594 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1597 cgrp = task_cgroup_from_root(task, root);
1598 path = cgroup_path(cgrp, buf, buflen);
1600 /* if no hierarchy exists, everyone is in "/" */
1601 if (strlcpy(buf, "/", buflen) < buflen)
1605 mutex_unlock(&cgroup_mutex);
1608 EXPORT_SYMBOL_GPL(task_cgroup_path);
1611 * Control Group taskset
1613 struct task_and_cgroup {
1614 struct task_struct *task;
1615 struct cgroup *cgrp;
1616 struct css_set *cset;
1619 struct cgroup_taskset {
1620 struct task_and_cgroup single;
1621 struct flex_array *tc_array;
1624 struct cgroup *cur_cgrp;
1628 * cgroup_taskset_first - reset taskset and return the first task
1629 * @tset: taskset of interest
1631 * @tset iteration is initialized and the first task is returned.
1633 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1635 if (tset->tc_array) {
1637 return cgroup_taskset_next(tset);
1639 tset->cur_cgrp = tset->single.cgrp;
1640 return tset->single.task;
1643 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1646 * cgroup_taskset_next - iterate to the next task in taskset
1647 * @tset: taskset of interest
1649 * Return the next task in @tset. Iteration must have been initialized
1650 * with cgroup_taskset_first().
1652 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1654 struct task_and_cgroup *tc;
1656 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1659 tc = flex_array_get(tset->tc_array, tset->idx++);
1660 tset->cur_cgrp = tc->cgrp;
1663 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1666 * cgroup_taskset_cur_css - return the matching css for the current task
1667 * @tset: taskset of interest
1668 * @subsys_id: the ID of the target subsystem
1670 * Return the css for the current (last returned) task of @tset for
1671 * subsystem specified by @subsys_id. This function must be preceded by
1672 * either cgroup_taskset_first() or cgroup_taskset_next().
1674 struct cgroup_subsys_state *cgroup_taskset_cur_css(struct cgroup_taskset *tset,
1677 return cgroup_css(tset->cur_cgrp, cgroup_subsys[subsys_id]);
1679 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_css);
1682 * cgroup_taskset_size - return the number of tasks in taskset
1683 * @tset: taskset of interest
1685 int cgroup_taskset_size(struct cgroup_taskset *tset)
1687 return tset->tc_array ? tset->tc_array_len : 1;
1689 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1693 * cgroup_task_migrate - move a task from one cgroup to another.
1695 * Must be called with cgroup_mutex and threadgroup locked.
1697 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1698 struct task_struct *tsk,
1699 struct css_set *new_cset)
1701 struct css_set *old_cset;
1704 * We are synchronized through threadgroup_lock() against PF_EXITING
1705 * setting such that we can't race against cgroup_exit() changing the
1706 * css_set to init_css_set and dropping the old one.
1708 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1709 old_cset = task_css_set(tsk);
1712 rcu_assign_pointer(tsk->cgroups, new_cset);
1715 /* Update the css_set linked lists if we're using them */
1716 write_lock(&css_set_lock);
1717 if (!list_empty(&tsk->cg_list))
1718 list_move(&tsk->cg_list, &new_cset->tasks);
1719 write_unlock(&css_set_lock);
1722 * We just gained a reference on old_cset by taking it from the
1723 * task. As trading it for new_cset is protected by cgroup_mutex,
1724 * we're safe to drop it here; it will be freed under RCU.
1726 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1727 put_css_set(old_cset);
1731 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1732 * @cgrp: the cgroup to attach to
1733 * @tsk: the task or the leader of the threadgroup to be attached
1734 * @threadgroup: attach the whole threadgroup?
1736 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1737 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1739 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1742 int retval, i, group_size;
1743 struct cgroupfs_root *root = cgrp->root;
1744 struct cgroup_subsys_state *css, *failed_css = NULL;
1745 /* threadgroup list cursor and array */
1746 struct task_struct *leader = tsk;
1747 struct task_and_cgroup *tc;
1748 struct flex_array *group;
1749 struct cgroup_taskset tset = { };
1752 * step 0: in order to do expensive, possibly blocking operations for
1753 * every thread, we cannot iterate the thread group list, since it needs
1754 * rcu or tasklist locked. instead, build an array of all threads in the
1755 * group - group_rwsem prevents new threads from appearing, and if
1756 * threads exit, this will just be an over-estimate.
1759 group_size = get_nr_threads(tsk);
1762 /* flex_array supports very large thread-groups better than kmalloc. */
1763 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1766 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1767 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1769 goto out_free_group_list;
1773 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1774 * already PF_EXITING could be freed from underneath us unless we
1775 * take an rcu_read_lock.
1779 struct task_and_cgroup ent;
1781 /* @tsk either already exited or can't exit until the end */
1782 if (tsk->flags & PF_EXITING)
1785 /* as per above, nr_threads may decrease, but not increase. */
1786 BUG_ON(i >= group_size);
1788 ent.cgrp = task_cgroup_from_root(tsk, root);
1789 /* nothing to do if this task is already in the cgroup */
1790 if (ent.cgrp == cgrp)
1793 * saying GFP_ATOMIC has no effect here because we did prealloc
1794 * earlier, but it's good form to communicate our expectations.
1796 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
1797 BUG_ON(retval != 0);
1802 } while_each_thread(leader, tsk);
1804 /* remember the number of threads in the array for later. */
1806 tset.tc_array = group;
1807 tset.tc_array_len = group_size;
1809 /* methods shouldn't be called if no task is actually migrating */
1812 goto out_free_group_list;
1815 * step 1: check that we can legitimately attach to the cgroup.
1817 for_each_css(css, i, cgrp) {
1818 if (css->ss->can_attach) {
1819 retval = css->ss->can_attach(css, &tset);
1822 goto out_cancel_attach;
1828 * step 2: make sure css_sets exist for all threads to be migrated.
1829 * we use find_css_set, which allocates a new one if necessary.
1831 for (i = 0; i < group_size; i++) {
1832 struct css_set *old_cset;
1834 tc = flex_array_get(group, i);
1835 old_cset = task_css_set(tc->task);
1836 tc->cset = find_css_set(old_cset, cgrp);
1839 goto out_put_css_set_refs;
1844 * step 3: now that we're guaranteed success wrt the css_sets,
1845 * proceed to move all tasks to the new cgroup. There are no
1846 * failure cases after here, so this is the commit point.
1848 for (i = 0; i < group_size; i++) {
1849 tc = flex_array_get(group, i);
1850 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
1852 /* nothing is sensitive to fork() after this point. */
1855 * step 4: do subsystem attach callbacks.
1857 for_each_css(css, i, cgrp)
1858 if (css->ss->attach)
1859 css->ss->attach(css, &tset);
1862 * step 5: success! and cleanup
1865 out_put_css_set_refs:
1867 for (i = 0; i < group_size; i++) {
1868 tc = flex_array_get(group, i);
1871 put_css_set(tc->cset);
1876 for_each_css(css, i, cgrp) {
1877 if (css == failed_css)
1879 if (css->ss->cancel_attach)
1880 css->ss->cancel_attach(css, &tset);
1883 out_free_group_list:
1884 flex_array_free(group);
1889 * Find the task_struct of the task to attach by vpid and pass it along to the
1890 * function to attach either it or all tasks in its threadgroup. Will lock
1891 * cgroup_mutex and threadgroup; may take task_lock of task.
1893 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
1895 struct task_struct *tsk;
1896 const struct cred *cred = current_cred(), *tcred;
1899 if (!cgroup_lock_live_group(cgrp))
1905 tsk = find_task_by_vpid(pid);
1909 goto out_unlock_cgroup;
1912 * even if we're attaching all tasks in the thread group, we
1913 * only need to check permissions on one of them.
1915 tcred = __task_cred(tsk);
1916 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
1917 !uid_eq(cred->euid, tcred->uid) &&
1918 !uid_eq(cred->euid, tcred->suid)) {
1921 goto out_unlock_cgroup;
1927 tsk = tsk->group_leader;
1930 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
1931 * trapped in a cpuset, or RT worker may be born in a cgroup
1932 * with no rt_runtime allocated. Just say no.
1934 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
1937 goto out_unlock_cgroup;
1940 get_task_struct(tsk);
1943 threadgroup_lock(tsk);
1945 if (!thread_group_leader(tsk)) {
1947 * a race with de_thread from another thread's exec()
1948 * may strip us of our leadership, if this happens,
1949 * there is no choice but to throw this task away and
1950 * try again; this is
1951 * "double-double-toil-and-trouble-check locking".
1953 threadgroup_unlock(tsk);
1954 put_task_struct(tsk);
1955 goto retry_find_task;
1959 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
1961 threadgroup_unlock(tsk);
1963 put_task_struct(tsk);
1965 mutex_unlock(&cgroup_mutex);
1970 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1971 * @from: attach to all cgroups of a given task
1972 * @tsk: the task to be attached
1974 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1976 struct cgroupfs_root *root;
1979 mutex_lock(&cgroup_mutex);
1980 for_each_active_root(root) {
1981 struct cgroup *from_cgrp = task_cgroup_from_root(from, root);
1983 retval = cgroup_attach_task(from_cgrp, tsk, false);
1987 mutex_unlock(&cgroup_mutex);
1991 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
1993 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
1994 struct cftype *cft, u64 pid)
1996 return attach_task_by_pid(css->cgroup, pid, false);
1999 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2000 struct cftype *cft, u64 tgid)
2002 return attach_task_by_pid(css->cgroup, tgid, true);
2005 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2006 struct cftype *cft, const char *buffer)
2008 struct cgroupfs_root *root = css->cgroup->root;
2010 BUILD_BUG_ON(sizeof(root->release_agent_path) < PATH_MAX);
2011 if (!cgroup_lock_live_group(css->cgroup))
2013 spin_lock(&release_agent_path_lock);
2014 strlcpy(root->release_agent_path, buffer,
2015 sizeof(root->release_agent_path));
2016 spin_unlock(&release_agent_path_lock);
2017 mutex_unlock(&cgroup_mutex);
2021 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2023 struct cgroup *cgrp = seq_css(seq)->cgroup;
2025 if (!cgroup_lock_live_group(cgrp))
2027 seq_puts(seq, cgrp->root->release_agent_path);
2028 seq_putc(seq, '\n');
2029 mutex_unlock(&cgroup_mutex);
2033 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2035 struct cgroup *cgrp = seq_css(seq)->cgroup;
2037 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2041 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2042 size_t nbytes, loff_t off)
2044 struct cgroup *cgrp = of->kn->parent->priv;
2045 struct cftype *cft = of->kn->priv;
2046 struct cgroup_subsys_state *css;
2050 * kernfs guarantees that a file isn't deleted with operations in
2051 * flight, which means that the matching css is and stays alive and
2052 * doesn't need to be pinned. The RCU locking is not necessary
2053 * either. It's just for the convenience of using cgroup_css().
2056 css = cgroup_css(cgrp, cft->ss);
2059 if (cft->write_string) {
2060 ret = cft->write_string(css, cft, strstrip(buf));
2061 } else if (cft->write_u64) {
2062 unsigned long long v;
2063 ret = kstrtoull(buf, 0, &v);
2065 ret = cft->write_u64(css, cft, v);
2066 } else if (cft->write_s64) {
2068 ret = kstrtoll(buf, 0, &v);
2070 ret = cft->write_s64(css, cft, v);
2071 } else if (cft->trigger) {
2072 ret = cft->trigger(css, (unsigned int)cft->private);
2077 return ret ?: nbytes;
2080 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2082 return seq_cft(seq)->seq_start(seq, ppos);
2085 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2087 return seq_cft(seq)->seq_next(seq, v, ppos);
2090 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2092 seq_cft(seq)->seq_stop(seq, v);
2095 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2097 struct cftype *cft = seq_cft(m);
2098 struct cgroup_subsys_state *css = seq_css(m);
2101 return cft->seq_show(m, arg);
2104 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2105 else if (cft->read_s64)
2106 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2112 static struct kernfs_ops cgroup_kf_single_ops = {
2113 .atomic_write_len = PAGE_SIZE,
2114 .write = cgroup_file_write,
2115 .seq_show = cgroup_seqfile_show,
2118 static struct kernfs_ops cgroup_kf_ops = {
2119 .atomic_write_len = PAGE_SIZE,
2120 .write = cgroup_file_write,
2121 .seq_start = cgroup_seqfile_start,
2122 .seq_next = cgroup_seqfile_next,
2123 .seq_stop = cgroup_seqfile_stop,
2124 .seq_show = cgroup_seqfile_show,
2128 * cgroup_rename - Only allow simple rename of directories in place.
2130 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2131 const char *new_name_str)
2133 struct cgroup *cgrp = kn->priv;
2136 if (kernfs_type(kn) != KERNFS_DIR)
2138 if (kn->parent != new_parent)
2142 * This isn't a proper migration and its usefulness is very
2143 * limited. Disallow if sane_behavior.
2145 if (cgroup_sane_behavior(cgrp))
2148 mutex_lock(&cgroup_tree_mutex);
2149 mutex_lock(&cgroup_mutex);
2151 ret = kernfs_rename(kn, new_parent, new_name_str);
2153 mutex_unlock(&cgroup_mutex);
2154 mutex_unlock(&cgroup_tree_mutex);
2158 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2160 char name[CGROUP_FILE_NAME_MAX];
2161 struct kernfs_node *kn;
2162 struct lock_class_key *key = NULL;
2164 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2165 key = &cft->lockdep_key;
2167 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2168 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2176 * cgroup_addrm_files - add or remove files to a cgroup directory
2177 * @cgrp: the target cgroup
2178 * @cfts: array of cftypes to be added
2179 * @is_add: whether to add or remove
2181 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2182 * For removals, this function never fails. If addition fails, this
2183 * function doesn't remove files already added. The caller is responsible
2186 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2192 lockdep_assert_held(&cgroup_tree_mutex);
2194 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2195 /* does cft->flags tell us to skip this file on @cgrp? */
2196 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2198 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2200 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2204 ret = cgroup_add_file(cgrp, cft);
2206 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2211 cgroup_rm_file(cgrp, cft);
2217 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2220 struct cgroup_subsys *ss = cfts[0].ss;
2221 struct cgroup *root = &ss->root->top_cgroup;
2222 struct cgroup_subsys_state *css;
2225 lockdep_assert_held(&cgroup_tree_mutex);
2227 /* don't bother if @ss isn't attached */
2228 if (ss->root == &cgroup_dummy_root)
2231 /* add/rm files for all cgroups created before */
2232 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2233 struct cgroup *cgrp = css->cgroup;
2235 if (cgroup_is_dead(cgrp))
2238 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2244 kernfs_activate(root->kn);
2248 static void cgroup_exit_cftypes(struct cftype *cfts)
2252 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2253 /* free copy for custom atomic_write_len, see init_cftypes() */
2254 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2261 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2265 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2266 struct kernfs_ops *kf_ops;
2268 WARN_ON(cft->ss || cft->kf_ops);
2271 kf_ops = &cgroup_kf_ops;
2273 kf_ops = &cgroup_kf_single_ops;
2276 * Ugh... if @cft wants a custom max_write_len, we need to
2277 * make a copy of kf_ops to set its atomic_write_len.
2279 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
2280 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
2282 cgroup_exit_cftypes(cfts);
2285 kf_ops->atomic_write_len = cft->max_write_len;
2288 cft->kf_ops = kf_ops;
2295 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
2297 lockdep_assert_held(&cgroup_tree_mutex);
2299 if (!cfts || !cfts[0].ss)
2302 list_del(&cfts->node);
2303 cgroup_apply_cftypes(cfts, false);
2304 cgroup_exit_cftypes(cfts);
2309 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2310 * @cfts: zero-length name terminated array of cftypes
2312 * Unregister @cfts. Files described by @cfts are removed from all
2313 * existing cgroups and all future cgroups won't have them either. This
2314 * function can be called anytime whether @cfts' subsys is attached or not.
2316 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2319 int cgroup_rm_cftypes(struct cftype *cfts)
2323 mutex_lock(&cgroup_tree_mutex);
2324 ret = cgroup_rm_cftypes_locked(cfts);
2325 mutex_unlock(&cgroup_tree_mutex);
2330 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2331 * @ss: target cgroup subsystem
2332 * @cfts: zero-length name terminated array of cftypes
2334 * Register @cfts to @ss. Files described by @cfts are created for all
2335 * existing cgroups to which @ss is attached and all future cgroups will
2336 * have them too. This function can be called anytime whether @ss is
2339 * Returns 0 on successful registration, -errno on failure. Note that this
2340 * function currently returns 0 as long as @cfts registration is successful
2341 * even if some file creation attempts on existing cgroups fail.
2343 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2347 ret = cgroup_init_cftypes(ss, cfts);
2351 mutex_lock(&cgroup_tree_mutex);
2353 list_add_tail(&cfts->node, &ss->cfts);
2354 ret = cgroup_apply_cftypes(cfts, true);
2356 cgroup_rm_cftypes_locked(cfts);
2358 mutex_unlock(&cgroup_tree_mutex);
2361 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2364 * cgroup_task_count - count the number of tasks in a cgroup.
2365 * @cgrp: the cgroup in question
2367 * Return the number of tasks in the cgroup.
2369 int cgroup_task_count(const struct cgroup *cgrp)
2372 struct cgrp_cset_link *link;
2374 read_lock(&css_set_lock);
2375 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2376 count += atomic_read(&link->cset->refcount);
2377 read_unlock(&css_set_lock);
2382 * To reduce the fork() overhead for systems that are not actually using
2383 * their cgroups capability, we don't maintain the lists running through
2384 * each css_set to its tasks until we see the list actually used - in other
2385 * words after the first call to css_task_iter_start().
2387 static void cgroup_enable_task_cg_lists(void)
2389 struct task_struct *p, *g;
2390 write_lock(&css_set_lock);
2391 use_task_css_set_links = 1;
2393 * We need tasklist_lock because RCU is not safe against
2394 * while_each_thread(). Besides, a forking task that has passed
2395 * cgroup_post_fork() without seeing use_task_css_set_links = 1
2396 * is not guaranteed to have its child immediately visible in the
2397 * tasklist if we walk through it with RCU.
2399 read_lock(&tasklist_lock);
2400 do_each_thread(g, p) {
2403 * We should check if the process is exiting, otherwise
2404 * it will race with cgroup_exit() in that the list
2405 * entry won't be deleted though the process has exited.
2407 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2408 list_add(&p->cg_list, &task_css_set(p)->tasks);
2410 } while_each_thread(g, p);
2411 read_unlock(&tasklist_lock);
2412 write_unlock(&css_set_lock);
2416 * css_next_child - find the next child of a given css
2417 * @pos_css: the current position (%NULL to initiate traversal)
2418 * @parent_css: css whose children to walk
2420 * This function returns the next child of @parent_css and should be called
2421 * under either cgroup_mutex or RCU read lock. The only requirement is
2422 * that @parent_css and @pos_css are accessible. The next sibling is
2423 * guaranteed to be returned regardless of their states.
2425 struct cgroup_subsys_state *
2426 css_next_child(struct cgroup_subsys_state *pos_css,
2427 struct cgroup_subsys_state *parent_css)
2429 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2430 struct cgroup *cgrp = parent_css->cgroup;
2431 struct cgroup *next;
2433 cgroup_assert_mutexes_or_rcu_locked();
2436 * @pos could already have been removed. Once a cgroup is removed,
2437 * its ->sibling.next is no longer updated when its next sibling
2438 * changes. As CGRP_DEAD assertion is serialized and happens
2439 * before the cgroup is taken off the ->sibling list, if we see it
2440 * unasserted, it's guaranteed that the next sibling hasn't
2441 * finished its grace period even if it's already removed, and thus
2442 * safe to dereference from this RCU critical section. If
2443 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2444 * to be visible as %true here.
2446 * If @pos is dead, its next pointer can't be dereferenced;
2447 * however, as each cgroup is given a monotonically increasing
2448 * unique serial number and always appended to the sibling list,
2449 * the next one can be found by walking the parent's children until
2450 * we see a cgroup with higher serial number than @pos's. While
2451 * this path can be slower, it's taken only when either the current
2452 * cgroup is removed or iteration and removal race.
2455 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2456 } else if (likely(!cgroup_is_dead(pos))) {
2457 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2459 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2460 if (next->serial_nr > pos->serial_nr)
2464 if (&next->sibling == &cgrp->children)
2467 return cgroup_css(next, parent_css->ss);
2469 EXPORT_SYMBOL_GPL(css_next_child);
2472 * css_next_descendant_pre - find the next descendant for pre-order walk
2473 * @pos: the current position (%NULL to initiate traversal)
2474 * @root: css whose descendants to walk
2476 * To be used by css_for_each_descendant_pre(). Find the next descendant
2477 * to visit for pre-order traversal of @root's descendants. @root is
2478 * included in the iteration and the first node to be visited.
2480 * While this function requires cgroup_mutex or RCU read locking, it
2481 * doesn't require the whole traversal to be contained in a single critical
2482 * section. This function will return the correct next descendant as long
2483 * as both @pos and @root are accessible and @pos is a descendant of @root.
2485 struct cgroup_subsys_state *
2486 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2487 struct cgroup_subsys_state *root)
2489 struct cgroup_subsys_state *next;
2491 cgroup_assert_mutexes_or_rcu_locked();
2493 /* if first iteration, visit @root */
2497 /* visit the first child if exists */
2498 next = css_next_child(NULL, pos);
2502 /* no child, visit my or the closest ancestor's next sibling */
2503 while (pos != root) {
2504 next = css_next_child(pos, css_parent(pos));
2507 pos = css_parent(pos);
2512 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
2515 * css_rightmost_descendant - return the rightmost descendant of a css
2516 * @pos: css of interest
2518 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2519 * is returned. This can be used during pre-order traversal to skip
2522 * While this function requires cgroup_mutex or RCU read locking, it
2523 * doesn't require the whole traversal to be contained in a single critical
2524 * section. This function will return the correct rightmost descendant as
2525 * long as @pos is accessible.
2527 struct cgroup_subsys_state *
2528 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2530 struct cgroup_subsys_state *last, *tmp;
2532 cgroup_assert_mutexes_or_rcu_locked();
2536 /* ->prev isn't RCU safe, walk ->next till the end */
2538 css_for_each_child(tmp, last)
2544 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
2546 static struct cgroup_subsys_state *
2547 css_leftmost_descendant(struct cgroup_subsys_state *pos)
2549 struct cgroup_subsys_state *last;
2553 pos = css_next_child(NULL, pos);
2560 * css_next_descendant_post - find the next descendant for post-order walk
2561 * @pos: the current position (%NULL to initiate traversal)
2562 * @root: css whose descendants to walk
2564 * To be used by css_for_each_descendant_post(). Find the next descendant
2565 * to visit for post-order traversal of @root's descendants. @root is
2566 * included in the iteration and the last node to be visited.
2568 * While this function requires cgroup_mutex or RCU read locking, it
2569 * doesn't require the whole traversal to be contained in a single critical
2570 * section. This function will return the correct next descendant as long
2571 * as both @pos and @cgroup are accessible and @pos is a descendant of
2574 struct cgroup_subsys_state *
2575 css_next_descendant_post(struct cgroup_subsys_state *pos,
2576 struct cgroup_subsys_state *root)
2578 struct cgroup_subsys_state *next;
2580 cgroup_assert_mutexes_or_rcu_locked();
2582 /* if first iteration, visit leftmost descendant which may be @root */
2584 return css_leftmost_descendant(root);
2586 /* if we visited @root, we're done */
2590 /* if there's an unvisited sibling, visit its leftmost descendant */
2591 next = css_next_child(pos, css_parent(pos));
2593 return css_leftmost_descendant(next);
2595 /* no sibling left, visit parent */
2596 return css_parent(pos);
2598 EXPORT_SYMBOL_GPL(css_next_descendant_post);
2601 * css_advance_task_iter - advance a task itererator to the next css_set
2602 * @it: the iterator to advance
2604 * Advance @it to the next css_set to walk.
2606 static void css_advance_task_iter(struct css_task_iter *it)
2608 struct list_head *l = it->cset_link;
2609 struct cgrp_cset_link *link;
2610 struct css_set *cset;
2612 /* Advance to the next non-empty css_set */
2615 if (l == &it->origin_css->cgroup->cset_links) {
2616 it->cset_link = NULL;
2619 link = list_entry(l, struct cgrp_cset_link, cset_link);
2621 } while (list_empty(&cset->tasks));
2623 it->task = cset->tasks.next;
2627 * css_task_iter_start - initiate task iteration
2628 * @css: the css to walk tasks of
2629 * @it: the task iterator to use
2631 * Initiate iteration through the tasks of @css. The caller can call
2632 * css_task_iter_next() to walk through the tasks until the function
2633 * returns NULL. On completion of iteration, css_task_iter_end() must be
2636 * Note that this function acquires a lock which is released when the
2637 * iteration finishes. The caller can't sleep while iteration is in
2640 void css_task_iter_start(struct cgroup_subsys_state *css,
2641 struct css_task_iter *it)
2642 __acquires(css_set_lock)
2645 * The first time anyone tries to iterate across a css, we need to
2646 * enable the list linking each css_set to its tasks, and fix up
2647 * all existing tasks.
2649 if (!use_task_css_set_links)
2650 cgroup_enable_task_cg_lists();
2652 read_lock(&css_set_lock);
2654 it->origin_css = css;
2655 it->cset_link = &css->cgroup->cset_links;
2657 css_advance_task_iter(it);
2661 * css_task_iter_next - return the next task for the iterator
2662 * @it: the task iterator being iterated
2664 * The "next" function for task iteration. @it should have been
2665 * initialized via css_task_iter_start(). Returns NULL when the iteration
2668 struct task_struct *css_task_iter_next(struct css_task_iter *it)
2670 struct task_struct *res;
2671 struct list_head *l = it->task;
2672 struct cgrp_cset_link *link;
2674 /* If the iterator cg is NULL, we have no tasks */
2677 res = list_entry(l, struct task_struct, cg_list);
2678 /* Advance iterator to find next entry */
2680 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
2681 if (l == &link->cset->tasks) {
2683 * We reached the end of this task list - move on to the
2684 * next cgrp_cset_link.
2686 css_advance_task_iter(it);
2694 * css_task_iter_end - finish task iteration
2695 * @it: the task iterator to finish
2697 * Finish task iteration started by css_task_iter_start().
2699 void css_task_iter_end(struct css_task_iter *it)
2700 __releases(css_set_lock)
2702 read_unlock(&css_set_lock);
2705 static inline int started_after_time(struct task_struct *t1,
2706 struct timespec *time,
2707 struct task_struct *t2)
2709 int start_diff = timespec_compare(&t1->start_time, time);
2710 if (start_diff > 0) {
2712 } else if (start_diff < 0) {
2716 * Arbitrarily, if two processes started at the same
2717 * time, we'll say that the lower pointer value
2718 * started first. Note that t2 may have exited by now
2719 * so this may not be a valid pointer any longer, but
2720 * that's fine - it still serves to distinguish
2721 * between two tasks started (effectively) simultaneously.
2728 * This function is a callback from heap_insert() and is used to order
2730 * In this case we order the heap in descending task start time.
2732 static inline int started_after(void *p1, void *p2)
2734 struct task_struct *t1 = p1;
2735 struct task_struct *t2 = p2;
2736 return started_after_time(t1, &t2->start_time, t2);
2740 * css_scan_tasks - iterate though all the tasks in a css
2741 * @css: the css to iterate tasks of
2742 * @test: optional test callback
2743 * @process: process callback
2744 * @data: data passed to @test and @process
2745 * @heap: optional pre-allocated heap used for task iteration
2747 * Iterate through all the tasks in @css, calling @test for each, and if it
2748 * returns %true, call @process for it also.
2750 * @test may be NULL, meaning always true (select all tasks), which
2751 * effectively duplicates css_task_iter_{start,next,end}() but does not
2752 * lock css_set_lock for the call to @process.
2754 * It is guaranteed that @process will act on every task that is a member
2755 * of @css for the duration of this call. This function may or may not
2756 * call @process for tasks that exit or move to a different css during the
2757 * call, or are forked or move into the css during the call.
2759 * Note that @test may be called with locks held, and may in some
2760 * situations be called multiple times for the same task, so it should be
2763 * If @heap is non-NULL, a heap has been pre-allocated and will be used for
2764 * heap operations (and its "gt" member will be overwritten), else a
2765 * temporary heap will be used (allocation of which may cause this function
2768 int css_scan_tasks(struct cgroup_subsys_state *css,
2769 bool (*test)(struct task_struct *, void *),
2770 void (*process)(struct task_struct *, void *),
2771 void *data, struct ptr_heap *heap)
2774 struct css_task_iter it;
2775 struct task_struct *p, *dropped;
2776 /* Never dereference latest_task, since it's not refcounted */
2777 struct task_struct *latest_task = NULL;
2778 struct ptr_heap tmp_heap;
2779 struct timespec latest_time = { 0, 0 };
2782 /* The caller supplied our heap and pre-allocated its memory */
2783 heap->gt = &started_after;
2785 /* We need to allocate our own heap memory */
2787 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2789 /* cannot allocate the heap */
2795 * Scan tasks in the css, using the @test callback to determine
2796 * which are of interest, and invoking @process callback on the
2797 * ones which need an update. Since we don't want to hold any
2798 * locks during the task updates, gather tasks to be processed in a
2799 * heap structure. The heap is sorted by descending task start
2800 * time. If the statically-sized heap fills up, we overflow tasks
2801 * that started later, and in future iterations only consider tasks
2802 * that started after the latest task in the previous pass. This
2803 * guarantees forward progress and that we don't miss any tasks.
2806 css_task_iter_start(css, &it);
2807 while ((p = css_task_iter_next(&it))) {
2809 * Only affect tasks that qualify per the caller's callback,
2810 * if he provided one
2812 if (test && !test(p, data))
2815 * Only process tasks that started after the last task
2818 if (!started_after_time(p, &latest_time, latest_task))
2820 dropped = heap_insert(heap, p);
2821 if (dropped == NULL) {
2823 * The new task was inserted; the heap wasn't
2827 } else if (dropped != p) {
2829 * The new task was inserted, and pushed out a
2833 put_task_struct(dropped);
2836 * Else the new task was newer than anything already in
2837 * the heap and wasn't inserted
2840 css_task_iter_end(&it);
2843 for (i = 0; i < heap->size; i++) {
2844 struct task_struct *q = heap->ptrs[i];
2846 latest_time = q->start_time;
2849 /* Process the task per the caller's callback */
2854 * If we had to process any tasks at all, scan again
2855 * in case some of them were in the middle of forking
2856 * children that didn't get processed.
2857 * Not the most efficient way to do it, but it avoids
2858 * having to take callback_mutex in the fork path
2862 if (heap == &tmp_heap)
2863 heap_free(&tmp_heap);
2867 static void cgroup_transfer_one_task(struct task_struct *task, void *data)
2869 struct cgroup *new_cgroup = data;
2871 mutex_lock(&cgroup_mutex);
2872 cgroup_attach_task(new_cgroup, task, false);
2873 mutex_unlock(&cgroup_mutex);
2877 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
2878 * @to: cgroup to which the tasks will be moved
2879 * @from: cgroup in which the tasks currently reside
2881 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
2883 return css_scan_tasks(&from->dummy_css, NULL, cgroup_transfer_one_task,
2888 * Stuff for reading the 'tasks'/'procs' files.
2890 * Reading this file can return large amounts of data if a cgroup has
2891 * *lots* of attached tasks. So it may need several calls to read(),
2892 * but we cannot guarantee that the information we produce is correct
2893 * unless we produce it entirely atomically.
2897 /* which pidlist file are we talking about? */
2898 enum cgroup_filetype {
2904 * A pidlist is a list of pids that virtually represents the contents of one
2905 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
2906 * a pair (one each for procs, tasks) for each pid namespace that's relevant
2909 struct cgroup_pidlist {
2911 * used to find which pidlist is wanted. doesn't change as long as
2912 * this particular list stays in the list.
2914 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
2917 /* how many elements the above list has */
2919 /* each of these stored in a list by its cgroup */
2920 struct list_head links;
2921 /* pointer to the cgroup we belong to, for list removal purposes */
2922 struct cgroup *owner;
2923 /* for delayed destruction */
2924 struct delayed_work destroy_dwork;
2928 * The following two functions "fix" the issue where there are more pids
2929 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2930 * TODO: replace with a kernel-wide solution to this problem
2932 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2933 static void *pidlist_allocate(int count)
2935 if (PIDLIST_TOO_LARGE(count))
2936 return vmalloc(count * sizeof(pid_t));
2938 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2941 static void pidlist_free(void *p)
2943 if (is_vmalloc_addr(p))
2950 * Used to destroy all pidlists lingering waiting for destroy timer. None
2951 * should be left afterwards.
2953 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
2955 struct cgroup_pidlist *l, *tmp_l;
2957 mutex_lock(&cgrp->pidlist_mutex);
2958 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
2959 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
2960 mutex_unlock(&cgrp->pidlist_mutex);
2962 flush_workqueue(cgroup_pidlist_destroy_wq);
2963 BUG_ON(!list_empty(&cgrp->pidlists));
2966 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
2968 struct delayed_work *dwork = to_delayed_work(work);
2969 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
2971 struct cgroup_pidlist *tofree = NULL;
2973 mutex_lock(&l->owner->pidlist_mutex);
2976 * Destroy iff we didn't get queued again. The state won't change
2977 * as destroy_dwork can only be queued while locked.
2979 if (!delayed_work_pending(dwork)) {
2980 list_del(&l->links);
2981 pidlist_free(l->list);
2982 put_pid_ns(l->key.ns);
2986 mutex_unlock(&l->owner->pidlist_mutex);
2991 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2992 * Returns the number of unique elements.
2994 static int pidlist_uniq(pid_t *list, int length)
2999 * we presume the 0th element is unique, so i starts at 1. trivial
3000 * edge cases first; no work needs to be done for either
3002 if (length == 0 || length == 1)
3004 /* src and dest walk down the list; dest counts unique elements */
3005 for (src = 1; src < length; src++) {
3006 /* find next unique element */
3007 while (list[src] == list[src-1]) {
3012 /* dest always points to where the next unique element goes */
3013 list[dest] = list[src];
3021 * The two pid files - task and cgroup.procs - guaranteed that the result
3022 * is sorted, which forced this whole pidlist fiasco. As pid order is
3023 * different per namespace, each namespace needs differently sorted list,
3024 * making it impossible to use, for example, single rbtree of member tasks
3025 * sorted by task pointer. As pidlists can be fairly large, allocating one
3026 * per open file is dangerous, so cgroup had to implement shared pool of
3027 * pidlists keyed by cgroup and namespace.
3029 * All this extra complexity was caused by the original implementation
3030 * committing to an entirely unnecessary property. In the long term, we
3031 * want to do away with it. Explicitly scramble sort order if
3032 * sane_behavior so that no such expectation exists in the new interface.
3034 * Scrambling is done by swapping every two consecutive bits, which is
3035 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3037 static pid_t pid_fry(pid_t pid)
3039 unsigned a = pid & 0x55555555;
3040 unsigned b = pid & 0xAAAAAAAA;
3042 return (a << 1) | (b >> 1);
3045 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3047 if (cgroup_sane_behavior(cgrp))
3048 return pid_fry(pid);
3053 static int cmppid(const void *a, const void *b)
3055 return *(pid_t *)a - *(pid_t *)b;
3058 static int fried_cmppid(const void *a, const void *b)
3060 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3063 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3064 enum cgroup_filetype type)
3066 struct cgroup_pidlist *l;
3067 /* don't need task_nsproxy() if we're looking at ourself */
3068 struct pid_namespace *ns = task_active_pid_ns(current);
3070 lockdep_assert_held(&cgrp->pidlist_mutex);
3072 list_for_each_entry(l, &cgrp->pidlists, links)
3073 if (l->key.type == type && l->key.ns == ns)
3079 * find the appropriate pidlist for our purpose (given procs vs tasks)
3080 * returns with the lock on that pidlist already held, and takes care
3081 * of the use count, or returns NULL with no locks held if we're out of
3084 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3085 enum cgroup_filetype type)
3087 struct cgroup_pidlist *l;
3089 lockdep_assert_held(&cgrp->pidlist_mutex);
3091 l = cgroup_pidlist_find(cgrp, type);
3095 /* entry not found; create a new one */
3096 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3100 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3102 /* don't need task_nsproxy() if we're looking at ourself */
3103 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3105 list_add(&l->links, &cgrp->pidlists);
3110 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3112 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3113 struct cgroup_pidlist **lp)
3117 int pid, n = 0; /* used for populating the array */
3118 struct css_task_iter it;
3119 struct task_struct *tsk;
3120 struct cgroup_pidlist *l;
3122 lockdep_assert_held(&cgrp->pidlist_mutex);
3125 * If cgroup gets more users after we read count, we won't have
3126 * enough space - tough. This race is indistinguishable to the
3127 * caller from the case that the additional cgroup users didn't
3128 * show up until sometime later on.
3130 length = cgroup_task_count(cgrp);
3131 array = pidlist_allocate(length);
3134 /* now, populate the array */
3135 css_task_iter_start(&cgrp->dummy_css, &it);
3136 while ((tsk = css_task_iter_next(&it))) {
3137 if (unlikely(n == length))
3139 /* get tgid or pid for procs or tasks file respectively */
3140 if (type == CGROUP_FILE_PROCS)
3141 pid = task_tgid_vnr(tsk);
3143 pid = task_pid_vnr(tsk);
3144 if (pid > 0) /* make sure to only use valid results */
3147 css_task_iter_end(&it);
3149 /* now sort & (if procs) strip out duplicates */
3150 if (cgroup_sane_behavior(cgrp))
3151 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3153 sort(array, length, sizeof(pid_t), cmppid, NULL);
3154 if (type == CGROUP_FILE_PROCS)
3155 length = pidlist_uniq(array, length);
3157 l = cgroup_pidlist_find_create(cgrp, type);
3159 mutex_unlock(&cgrp->pidlist_mutex);
3160 pidlist_free(array);
3164 /* store array, freeing old if necessary */
3165 pidlist_free(l->list);
3173 * cgroupstats_build - build and fill cgroupstats
3174 * @stats: cgroupstats to fill information into
3175 * @dentry: A dentry entry belonging to the cgroup for which stats have
3178 * Build and fill cgroupstats so that taskstats can export it to user
3181 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3183 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3184 struct cgroup *cgrp;
3185 struct css_task_iter it;
3186 struct task_struct *tsk;
3188 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3189 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3190 kernfs_type(kn) != KERNFS_DIR)
3194 * We aren't being called from kernfs and there's no guarantee on
3195 * @kn->priv's validity. For this and css_tryget_from_dir(),
3196 * @kn->priv is RCU safe. Let's do the RCU dancing.
3199 cgrp = rcu_dereference(kn->priv);
3205 css_task_iter_start(&cgrp->dummy_css, &it);
3206 while ((tsk = css_task_iter_next(&it))) {
3207 switch (tsk->state) {
3209 stats->nr_running++;
3211 case TASK_INTERRUPTIBLE:
3212 stats->nr_sleeping++;
3214 case TASK_UNINTERRUPTIBLE:
3215 stats->nr_uninterruptible++;
3218 stats->nr_stopped++;
3221 if (delayacct_is_task_waiting_on_io(tsk))
3222 stats->nr_io_wait++;
3226 css_task_iter_end(&it);
3234 * seq_file methods for the tasks/procs files. The seq_file position is the
3235 * next pid to display; the seq_file iterator is a pointer to the pid
3236 * in the cgroup->l->list array.
3239 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3242 * Initially we receive a position value that corresponds to
3243 * one more than the last pid shown (or 0 on the first call or
3244 * after a seek to the start). Use a binary-search to find the
3245 * next pid to display, if any
3247 struct kernfs_open_file *of = s->private;
3248 struct cgroup *cgrp = seq_css(s)->cgroup;
3249 struct cgroup_pidlist *l;
3250 enum cgroup_filetype type = seq_cft(s)->private;
3251 int index = 0, pid = *pos;
3254 mutex_lock(&cgrp->pidlist_mutex);
3257 * !NULL @of->priv indicates that this isn't the first start()
3258 * after open. If the matching pidlist is around, we can use that.
3259 * Look for it. Note that @of->priv can't be used directly. It
3260 * could already have been destroyed.
3263 of->priv = cgroup_pidlist_find(cgrp, type);
3266 * Either this is the first start() after open or the matching
3267 * pidlist has been destroyed inbetween. Create a new one.
3270 ret = pidlist_array_load(cgrp, type,
3271 (struct cgroup_pidlist **)&of->priv);
3273 return ERR_PTR(ret);
3278 int end = l->length;
3280 while (index < end) {
3281 int mid = (index + end) / 2;
3282 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3285 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3291 /* If we're off the end of the array, we're done */
3292 if (index >= l->length)
3294 /* Update the abstract position to be the actual pid that we found */
3295 iter = l->list + index;
3296 *pos = cgroup_pid_fry(cgrp, *iter);
3300 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3302 struct kernfs_open_file *of = s->private;
3303 struct cgroup_pidlist *l = of->priv;
3306 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3307 CGROUP_PIDLIST_DESTROY_DELAY);
3308 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3311 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3313 struct kernfs_open_file *of = s->private;
3314 struct cgroup_pidlist *l = of->priv;
3316 pid_t *end = l->list + l->length;
3318 * Advance to the next pid in the array. If this goes off the
3325 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3330 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3332 return seq_printf(s, "%d\n", *(int *)v);
3336 * seq_operations functions for iterating on pidlists through seq_file -
3337 * independent of whether it's tasks or procs
3339 static const struct seq_operations cgroup_pidlist_seq_operations = {
3340 .start = cgroup_pidlist_start,
3341 .stop = cgroup_pidlist_stop,
3342 .next = cgroup_pidlist_next,
3343 .show = cgroup_pidlist_show,
3346 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3349 return notify_on_release(css->cgroup);
3352 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3353 struct cftype *cft, u64 val)
3355 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3357 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3359 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3363 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3366 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3369 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3370 struct cftype *cft, u64 val)
3373 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3375 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3379 static struct cftype cgroup_base_files[] = {
3381 .name = "cgroup.procs",
3382 .seq_start = cgroup_pidlist_start,
3383 .seq_next = cgroup_pidlist_next,
3384 .seq_stop = cgroup_pidlist_stop,
3385 .seq_show = cgroup_pidlist_show,
3386 .private = CGROUP_FILE_PROCS,
3387 .write_u64 = cgroup_procs_write,
3388 .mode = S_IRUGO | S_IWUSR,
3391 .name = "cgroup.clone_children",
3392 .flags = CFTYPE_INSANE,
3393 .read_u64 = cgroup_clone_children_read,
3394 .write_u64 = cgroup_clone_children_write,
3397 .name = "cgroup.sane_behavior",
3398 .flags = CFTYPE_ONLY_ON_ROOT,
3399 .seq_show = cgroup_sane_behavior_show,
3403 * Historical crazy stuff. These don't have "cgroup." prefix and
3404 * don't exist if sane_behavior. If you're depending on these, be
3405 * prepared to be burned.
3409 .flags = CFTYPE_INSANE, /* use "procs" instead */
3410 .seq_start = cgroup_pidlist_start,
3411 .seq_next = cgroup_pidlist_next,
3412 .seq_stop = cgroup_pidlist_stop,
3413 .seq_show = cgroup_pidlist_show,
3414 .private = CGROUP_FILE_TASKS,
3415 .write_u64 = cgroup_tasks_write,
3416 .mode = S_IRUGO | S_IWUSR,
3419 .name = "notify_on_release",
3420 .flags = CFTYPE_INSANE,
3421 .read_u64 = cgroup_read_notify_on_release,
3422 .write_u64 = cgroup_write_notify_on_release,
3425 .name = "release_agent",
3426 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3427 .seq_show = cgroup_release_agent_show,
3428 .write_string = cgroup_release_agent_write,
3429 .max_write_len = PATH_MAX - 1,
3435 * cgroup_populate_dir - create subsys files in a cgroup directory
3436 * @cgrp: target cgroup
3437 * @subsys_mask: mask of the subsystem ids whose files should be added
3439 * On failure, no file is added.
3441 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3443 struct cgroup_subsys *ss;
3446 /* process cftsets of each subsystem */
3447 for_each_subsys(ss, i) {
3448 struct cftype *cfts;
3450 if (!test_bit(i, &subsys_mask))
3453 list_for_each_entry(cfts, &ss->cfts, node) {
3454 ret = cgroup_addrm_files(cgrp, cfts, true);
3461 cgroup_clear_dir(cgrp, subsys_mask);
3466 * css destruction is four-stage process.
3468 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3469 * Implemented in kill_css().
3471 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3472 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3473 * by invoking offline_css(). After offlining, the base ref is put.
3474 * Implemented in css_killed_work_fn().
3476 * 3. When the percpu_ref reaches zero, the only possible remaining
3477 * accessors are inside RCU read sections. css_release() schedules the
3480 * 4. After the grace period, the css can be freed. Implemented in
3481 * css_free_work_fn().
3483 * It is actually hairier because both step 2 and 4 require process context
3484 * and thus involve punting to css->destroy_work adding two additional
3485 * steps to the already complex sequence.
3487 static void css_free_work_fn(struct work_struct *work)
3489 struct cgroup_subsys_state *css =
3490 container_of(work, struct cgroup_subsys_state, destroy_work);
3491 struct cgroup *cgrp = css->cgroup;
3494 css_put(css->parent);
3496 css->ss->css_free(css);
3500 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3502 struct cgroup_subsys_state *css =
3503 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3505 INIT_WORK(&css->destroy_work, css_free_work_fn);
3506 queue_work(cgroup_destroy_wq, &css->destroy_work);
3509 static void css_release(struct percpu_ref *ref)
3511 struct cgroup_subsys_state *css =
3512 container_of(ref, struct cgroup_subsys_state, refcnt);
3514 rcu_assign_pointer(css->cgroup->subsys[css->ss->id], NULL);
3515 call_rcu(&css->rcu_head, css_free_rcu_fn);
3518 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3519 struct cgroup *cgrp)
3526 css->parent = cgroup_css(cgrp->parent, ss);
3528 css->flags |= CSS_ROOT;
3530 BUG_ON(cgroup_css(cgrp, ss));
3533 /* invoke ->css_online() on a new CSS and mark it online if successful */
3534 static int online_css(struct cgroup_subsys_state *css)
3536 struct cgroup_subsys *ss = css->ss;
3539 lockdep_assert_held(&cgroup_tree_mutex);
3540 lockdep_assert_held(&cgroup_mutex);
3543 ret = ss->css_online(css);
3545 css->flags |= CSS_ONLINE;
3546 css->cgroup->nr_css++;
3547 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
3552 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
3553 static void offline_css(struct cgroup_subsys_state *css)
3555 struct cgroup_subsys *ss = css->ss;
3557 lockdep_assert_held(&cgroup_tree_mutex);
3558 lockdep_assert_held(&cgroup_mutex);
3560 if (!(css->flags & CSS_ONLINE))
3563 if (ss->css_offline)
3564 ss->css_offline(css);
3566 css->flags &= ~CSS_ONLINE;
3567 css->cgroup->nr_css--;
3568 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
3572 * create_css - create a cgroup_subsys_state
3573 * @cgrp: the cgroup new css will be associated with
3574 * @ss: the subsys of new css
3576 * Create a new css associated with @cgrp - @ss pair. On success, the new
3577 * css is online and installed in @cgrp with all interface files created.
3578 * Returns 0 on success, -errno on failure.
3580 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
3582 struct cgroup *parent = cgrp->parent;
3583 struct cgroup_subsys_state *css;
3586 lockdep_assert_held(&cgroup_mutex);
3588 css = ss->css_alloc(cgroup_css(parent, ss));
3590 return PTR_ERR(css);
3592 err = percpu_ref_init(&css->refcnt, css_release);
3596 init_css(css, ss, cgrp);
3598 err = cgroup_populate_dir(cgrp, 1 << ss->id);
3602 err = online_css(css);
3607 css_get(css->parent);
3609 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
3611 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",
3612 current->comm, current->pid, ss->name);
3613 if (!strcmp(ss->name, "memory"))
3614 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
3615 ss->warned_broken_hierarchy = true;
3621 percpu_ref_cancel_init(&css->refcnt);
3627 * cgroup_create - create a cgroup
3628 * @parent: cgroup that will be parent of the new cgroup
3629 * @name: name of the new cgroup
3630 * @mode: mode to set on new cgroup
3632 static long cgroup_create(struct cgroup *parent, const char *name,
3635 struct cgroup *cgrp;
3636 struct cgroupfs_root *root = parent->root;
3638 struct cgroup_subsys *ss;
3639 struct kernfs_node *kn;
3641 /* allocate the cgroup and its ID, 0 is reserved for the root */
3642 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3646 mutex_lock(&cgroup_tree_mutex);
3649 * Only live parents can have children. Note that the liveliness
3650 * check isn't strictly necessary because cgroup_mkdir() and
3651 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
3652 * anyway so that locking is contained inside cgroup proper and we
3653 * don't get nasty surprises if we ever grow another caller.
3655 if (!cgroup_lock_live_group(parent)) {
3657 goto err_unlock_tree;
3661 * Temporarily set the pointer to NULL, so idr_find() won't return
3662 * a half-baked cgroup.
3664 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
3670 init_cgroup_housekeeping(cgrp);
3672 cgrp->parent = parent;
3673 cgrp->dummy_css.parent = &parent->dummy_css;
3674 cgrp->root = parent->root;
3676 if (notify_on_release(parent))
3677 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3679 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
3680 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
3682 /* create the directory */
3683 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
3691 * This extra ref will be put in cgroup_free_fn() and guarantees
3692 * that @cgrp->kn is always accessible.
3696 cgrp->serial_nr = cgroup_serial_nr_next++;
3698 /* allocation complete, commit to creation */
3699 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
3700 atomic_inc(&root->nr_cgrps);
3704 * @cgrp is now fully operational. If something fails after this
3705 * point, it'll be released via the normal destruction path.
3707 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
3709 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
3713 /* let's create and online css's */
3714 for_each_subsys(ss, ssid) {
3715 if (root->subsys_mask & (1 << ssid)) {
3716 err = create_css(cgrp, ss);
3722 kernfs_activate(kn);
3724 mutex_unlock(&cgroup_mutex);
3725 mutex_unlock(&cgroup_tree_mutex);
3730 idr_remove(&root->cgroup_idr, cgrp->id);
3732 mutex_unlock(&cgroup_mutex);
3734 mutex_unlock(&cgroup_tree_mutex);
3739 cgroup_destroy_locked(cgrp);
3740 mutex_unlock(&cgroup_mutex);
3741 mutex_unlock(&cgroup_tree_mutex);
3745 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3748 struct cgroup *parent = parent_kn->priv;
3750 return cgroup_create(parent, name, mode);
3754 * This is called when the refcnt of a css is confirmed to be killed.
3755 * css_tryget() is now guaranteed to fail.
3757 static void css_killed_work_fn(struct work_struct *work)
3759 struct cgroup_subsys_state *css =
3760 container_of(work, struct cgroup_subsys_state, destroy_work);
3761 struct cgroup *cgrp = css->cgroup;
3763 mutex_lock(&cgroup_tree_mutex);
3764 mutex_lock(&cgroup_mutex);
3767 * css_tryget() is guaranteed to fail now. Tell subsystems to
3768 * initate destruction.
3773 * If @cgrp is marked dead, it's waiting for refs of all css's to
3774 * be disabled before proceeding to the second phase of cgroup
3775 * destruction. If we are the last one, kick it off.
3777 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
3778 cgroup_destroy_css_killed(cgrp);
3780 mutex_unlock(&cgroup_mutex);
3781 mutex_unlock(&cgroup_tree_mutex);
3784 * Put the css refs from kill_css(). Each css holds an extra
3785 * reference to the cgroup's dentry and cgroup removal proceeds
3786 * regardless of css refs. On the last put of each css, whenever
3787 * that may be, the extra dentry ref is put so that dentry
3788 * destruction happens only after all css's are released.
3793 /* css kill confirmation processing requires process context, bounce */
3794 static void css_killed_ref_fn(struct percpu_ref *ref)
3796 struct cgroup_subsys_state *css =
3797 container_of(ref, struct cgroup_subsys_state, refcnt);
3799 INIT_WORK(&css->destroy_work, css_killed_work_fn);
3800 queue_work(cgroup_destroy_wq, &css->destroy_work);
3804 * kill_css - destroy a css
3805 * @css: css to destroy
3807 * This function initiates destruction of @css by removing cgroup interface
3808 * files and putting its base reference. ->css_offline() will be invoked
3809 * asynchronously once css_tryget() is guaranteed to fail and when the
3810 * reference count reaches zero, @css will be released.
3812 static void kill_css(struct cgroup_subsys_state *css)
3815 * This must happen before css is disassociated with its cgroup.
3816 * See seq_css() for details.
3818 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3821 * Killing would put the base ref, but we need to keep it alive
3822 * until after ->css_offline().
3827 * cgroup core guarantees that, by the time ->css_offline() is
3828 * invoked, no new css reference will be given out via
3829 * css_tryget(). We can't simply call percpu_ref_kill() and
3830 * proceed to offlining css's because percpu_ref_kill() doesn't
3831 * guarantee that the ref is seen as killed on all CPUs on return.
3833 * Use percpu_ref_kill_and_confirm() to get notifications as each
3834 * css is confirmed to be seen as killed on all CPUs.
3836 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
3840 * cgroup_destroy_locked - the first stage of cgroup destruction
3841 * @cgrp: cgroup to be destroyed
3843 * css's make use of percpu refcnts whose killing latency shouldn't be
3844 * exposed to userland and are RCU protected. Also, cgroup core needs to
3845 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
3846 * invoked. To satisfy all the requirements, destruction is implemented in
3847 * the following two steps.
3849 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
3850 * userland visible parts and start killing the percpu refcnts of
3851 * css's. Set up so that the next stage will be kicked off once all
3852 * the percpu refcnts are confirmed to be killed.
3854 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
3855 * rest of destruction. Once all cgroup references are gone, the
3856 * cgroup is RCU-freed.
3858 * This function implements s1. After this step, @cgrp is gone as far as
3859 * the userland is concerned and a new cgroup with the same name may be
3860 * created. As cgroup doesn't care about the names internally, this
3861 * doesn't cause any problem.
3863 static int cgroup_destroy_locked(struct cgroup *cgrp)
3864 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
3866 struct cgroup *child;
3867 struct cgroup_subsys_state *css;
3871 lockdep_assert_held(&cgroup_tree_mutex);
3872 lockdep_assert_held(&cgroup_mutex);
3875 * css_set_lock synchronizes access to ->cset_links and prevents
3876 * @cgrp from being removed while __put_css_set() is in progress.
3878 read_lock(&css_set_lock);
3879 empty = list_empty(&cgrp->cset_links);
3880 read_unlock(&css_set_lock);
3885 * Make sure there's no live children. We can't test ->children
3886 * emptiness as dead children linger on it while being destroyed;
3887 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
3891 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
3892 empty = cgroup_is_dead(child);
3901 * Initiate massacre of all css's. cgroup_destroy_css_killed()
3902 * will be invoked to perform the rest of destruction once the
3903 * percpu refs of all css's are confirmed to be killed. This
3904 * involves removing the subsystem's files, drop cgroup_mutex.
3906 mutex_unlock(&cgroup_mutex);
3907 for_each_css(css, ssid, cgrp)
3909 mutex_lock(&cgroup_mutex);
3912 * Mark @cgrp dead. This prevents further task migration and child
3913 * creation by disabling cgroup_lock_live_group(). Note that
3914 * CGRP_DEAD assertion is depended upon by css_next_child() to
3915 * resume iteration after dropping RCU read lock. See
3916 * css_next_child() for details.
3918 set_bit(CGRP_DEAD, &cgrp->flags);
3920 /* CGRP_DEAD is set, remove from ->release_list for the last time */
3921 raw_spin_lock(&release_list_lock);
3922 if (!list_empty(&cgrp->release_list))
3923 list_del_init(&cgrp->release_list);
3924 raw_spin_unlock(&release_list_lock);
3927 * If @cgrp has css's attached, the second stage of cgroup
3928 * destruction is kicked off from css_killed_work_fn() after the
3929 * refs of all attached css's are killed. If @cgrp doesn't have
3930 * any css, we kick it off here.
3933 cgroup_destroy_css_killed(cgrp);
3935 /* remove @cgrp directory along with the base files */
3936 mutex_unlock(&cgroup_mutex);
3939 * There are two control paths which try to determine cgroup from
3940 * dentry without going through kernfs - cgroupstats_build() and
3941 * css_tryget_from_dir(). Those are supported by RCU protecting
3942 * clearing of cgrp->kn->priv backpointer, which should happen
3943 * after all files under it have been removed.
3945 kernfs_remove(cgrp->kn); /* @cgrp has an extra ref on its kn */
3946 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
3948 mutex_lock(&cgroup_mutex);
3954 * cgroup_destroy_css_killed - the second step of cgroup destruction
3955 * @work: cgroup->destroy_free_work
3957 * This function is invoked from a work item for a cgroup which is being
3958 * destroyed after all css's are offlined and performs the rest of
3959 * destruction. This is the second step of destruction described in the
3960 * comment above cgroup_destroy_locked().
3962 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
3964 struct cgroup *parent = cgrp->parent;
3966 lockdep_assert_held(&cgroup_tree_mutex);
3967 lockdep_assert_held(&cgroup_mutex);
3969 /* delete this cgroup from parent->children */
3970 list_del_rcu(&cgrp->sibling);
3974 set_bit(CGRP_RELEASABLE, &parent->flags);
3975 check_for_release(parent);
3978 static int cgroup_rmdir(struct kernfs_node *kn)
3980 struct cgroup *cgrp = kn->priv;
3984 * This is self-destruction but @kn can't be removed while this
3985 * callback is in progress. Let's break active protection. Once
3986 * the protection is broken, @cgrp can be destroyed at any point.
3987 * Pin it so that it stays accessible.
3990 kernfs_break_active_protection(kn);
3992 mutex_lock(&cgroup_tree_mutex);
3993 mutex_lock(&cgroup_mutex);
3996 * @cgrp might already have been destroyed while we're trying to
3999 if (!cgroup_is_dead(cgrp))
4000 ret = cgroup_destroy_locked(cgrp);
4002 mutex_unlock(&cgroup_mutex);
4003 mutex_unlock(&cgroup_tree_mutex);
4005 kernfs_unbreak_active_protection(kn);
4010 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4011 .remount_fs = cgroup_remount,
4012 .show_options = cgroup_show_options,
4013 .mkdir = cgroup_mkdir,
4014 .rmdir = cgroup_rmdir,
4015 .rename = cgroup_rename,
4018 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4020 struct cgroup_subsys_state *css;
4022 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4024 mutex_lock(&cgroup_tree_mutex);
4025 mutex_lock(&cgroup_mutex);
4027 INIT_LIST_HEAD(&ss->cfts);
4029 /* Create the top cgroup state for this subsystem */
4030 ss->root = &cgroup_dummy_root;
4031 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4032 /* We don't handle early failures gracefully */
4033 BUG_ON(IS_ERR(css));
4034 init_css(css, ss, cgroup_dummy_top);
4036 /* Update the init_css_set to contain a subsys
4037 * pointer to this state - since the subsystem is
4038 * newly registered, all tasks and hence the
4039 * init_css_set is in the subsystem's top cgroup. */
4040 init_css_set.subsys[ss->id] = css;
4042 need_forkexit_callback |= ss->fork || ss->exit;
4044 /* At system boot, before all subsystems have been
4045 * registered, no tasks have been forked, so we don't
4046 * need to invoke fork callbacks here. */
4047 BUG_ON(!list_empty(&init_task.tasks));
4049 BUG_ON(online_css(css));
4051 mutex_unlock(&cgroup_mutex);
4052 mutex_unlock(&cgroup_tree_mutex);
4056 * cgroup_init_early - cgroup initialization at system boot
4058 * Initialize cgroups at system boot, and initialize any
4059 * subsystems that request early init.
4061 int __init cgroup_init_early(void)
4063 struct cgroup_subsys *ss;
4066 atomic_set(&init_css_set.refcount, 1);
4067 INIT_LIST_HEAD(&init_css_set.cgrp_links);
4068 INIT_LIST_HEAD(&init_css_set.tasks);
4069 INIT_HLIST_NODE(&init_css_set.hlist);
4071 init_cgroup_root(&cgroup_dummy_root);
4072 cgroup_root_count = 1;
4073 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4075 init_cgrp_cset_link.cset = &init_css_set;
4076 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4077 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4078 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4080 for_each_subsys(ss, i) {
4081 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4082 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4083 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4085 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4086 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4089 ss->name = cgroup_subsys_name[i];
4092 cgroup_init_subsys(ss);
4098 * cgroup_init - cgroup initialization
4100 * Register cgroup filesystem and /proc file, and initialize
4101 * any subsystems that didn't request early init.
4103 int __init cgroup_init(void)
4105 struct cgroup_subsys *ss;
4109 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4111 for_each_subsys(ss, i) {
4112 if (!ss->early_init)
4113 cgroup_init_subsys(ss);
4116 * cftype registration needs kmalloc and can't be done
4117 * during early_init. Register base cftypes separately.
4119 if (ss->base_cftypes)
4120 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4123 /* allocate id for the dummy hierarchy */
4124 mutex_lock(&cgroup_mutex);
4126 /* Add init_css_set to the hash table */
4127 key = css_set_hash(init_css_set.subsys);
4128 hash_add(css_set_table, &init_css_set.hlist, key);
4130 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4132 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
4136 mutex_unlock(&cgroup_mutex);
4138 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4142 err = register_filesystem(&cgroup_fs_type);
4144 kobject_put(cgroup_kobj);
4148 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4152 static int __init cgroup_wq_init(void)
4155 * There isn't much point in executing destruction path in
4156 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4158 * XXX: Must be ordered to make sure parent is offlined after
4159 * children. The ordering requirement is for memcg where a
4160 * parent's offline may wait for a child's leading to deadlock. In
4161 * the long term, this should be fixed from memcg side.
4163 * We would prefer to do this in cgroup_init() above, but that
4164 * is called before init_workqueues(): so leave this until after.
4166 cgroup_destroy_wq = alloc_ordered_workqueue("cgroup_destroy", 0);
4167 BUG_ON(!cgroup_destroy_wq);
4170 * Used to destroy pidlists and separate to serve as flush domain.
4171 * Cap @max_active to 1 too.
4173 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4175 BUG_ON(!cgroup_pidlist_destroy_wq);
4179 core_initcall(cgroup_wq_init);
4182 * proc_cgroup_show()
4183 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4184 * - Used for /proc/<pid>/cgroup.
4185 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4186 * doesn't really matter if tsk->cgroup changes after we read it,
4187 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4188 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4189 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4190 * cgroup to top_cgroup.
4193 /* TODO: Use a proper seq_file iterator */
4194 int proc_cgroup_show(struct seq_file *m, void *v)
4197 struct task_struct *tsk;
4200 struct cgroupfs_root *root;
4203 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4209 tsk = get_pid_task(pid, PIDTYPE_PID);
4215 mutex_lock(&cgroup_mutex);
4217 for_each_active_root(root) {
4218 struct cgroup_subsys *ss;
4219 struct cgroup *cgrp;
4220 int ssid, count = 0;
4222 seq_printf(m, "%d:", root->hierarchy_id);
4223 for_each_subsys(ss, ssid)
4224 if (root->subsys_mask & (1 << ssid))
4225 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4226 if (strlen(root->name))
4227 seq_printf(m, "%sname=%s", count ? "," : "",
4230 cgrp = task_cgroup_from_root(tsk, root);
4231 path = cgroup_path(cgrp, buf, PATH_MAX);
4233 retval = -ENAMETOOLONG;
4241 mutex_unlock(&cgroup_mutex);
4242 put_task_struct(tsk);
4249 /* Display information about each subsystem and each hierarchy */
4250 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4252 struct cgroup_subsys *ss;
4255 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4257 * ideally we don't want subsystems moving around while we do this.
4258 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4259 * subsys/hierarchy state.
4261 mutex_lock(&cgroup_mutex);
4263 for_each_subsys(ss, i)
4264 seq_printf(m, "%s\t%d\t%d\t%d\n",
4265 ss->name, ss->root->hierarchy_id,
4266 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4268 mutex_unlock(&cgroup_mutex);
4272 static int cgroupstats_open(struct inode *inode, struct file *file)
4274 return single_open(file, proc_cgroupstats_show, NULL);
4277 static const struct file_operations proc_cgroupstats_operations = {
4278 .open = cgroupstats_open,
4280 .llseek = seq_lseek,
4281 .release = single_release,
4285 * cgroup_fork - attach newly forked task to its parents cgroup.
4286 * @child: pointer to task_struct of forking parent process.
4288 * Description: A task inherits its parent's cgroup at fork().
4290 * A pointer to the shared css_set was automatically copied in
4291 * fork.c by dup_task_struct(). However, we ignore that copy, since
4292 * it was not made under the protection of RCU or cgroup_mutex, so
4293 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4294 * have already changed current->cgroups, allowing the previously
4295 * referenced cgroup group to be removed and freed.
4297 * At the point that cgroup_fork() is called, 'current' is the parent
4298 * task, and the passed argument 'child' points to the child task.
4300 void cgroup_fork(struct task_struct *child)
4303 get_css_set(task_css_set(current));
4304 child->cgroups = current->cgroups;
4305 task_unlock(current);
4306 INIT_LIST_HEAD(&child->cg_list);
4310 * cgroup_post_fork - called on a new task after adding it to the task list
4311 * @child: the task in question
4313 * Adds the task to the list running through its css_set if necessary and
4314 * call the subsystem fork() callbacks. Has to be after the task is
4315 * visible on the task list in case we race with the first call to
4316 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4319 void cgroup_post_fork(struct task_struct *child)
4321 struct cgroup_subsys *ss;
4325 * use_task_css_set_links is set to 1 before we walk the tasklist
4326 * under the tasklist_lock and we read it here after we added the child
4327 * to the tasklist under the tasklist_lock as well. If the child wasn't
4328 * yet in the tasklist when we walked through it from
4329 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4330 * should be visible now due to the paired locking and barriers implied
4331 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4332 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4335 if (use_task_css_set_links) {
4336 write_lock(&css_set_lock);
4338 if (list_empty(&child->cg_list))
4339 list_add(&child->cg_list, &task_css_set(child)->tasks);
4341 write_unlock(&css_set_lock);
4345 * Call ss->fork(). This must happen after @child is linked on
4346 * css_set; otherwise, @child might change state between ->fork()
4347 * and addition to css_set.
4349 if (need_forkexit_callback) {
4350 for_each_subsys(ss, i)
4357 * cgroup_exit - detach cgroup from exiting task
4358 * @tsk: pointer to task_struct of exiting process
4359 * @run_callback: run exit callbacks?
4361 * Description: Detach cgroup from @tsk and release it.
4363 * Note that cgroups marked notify_on_release force every task in
4364 * them to take the global cgroup_mutex mutex when exiting.
4365 * This could impact scaling on very large systems. Be reluctant to
4366 * use notify_on_release cgroups where very high task exit scaling
4367 * is required on large systems.
4369 * the_top_cgroup_hack:
4371 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4373 * We call cgroup_exit() while the task is still competent to
4374 * handle notify_on_release(), then leave the task attached to the
4375 * root cgroup in each hierarchy for the remainder of its exit.
4377 * To do this properly, we would increment the reference count on
4378 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4379 * code we would add a second cgroup function call, to drop that
4380 * reference. This would just create an unnecessary hot spot on
4381 * the top_cgroup reference count, to no avail.
4383 * Normally, holding a reference to a cgroup without bumping its
4384 * count is unsafe. The cgroup could go away, or someone could
4385 * attach us to a different cgroup, decrementing the count on
4386 * the first cgroup that we never incremented. But in this case,
4387 * top_cgroup isn't going away, and either task has PF_EXITING set,
4388 * which wards off any cgroup_attach_task() attempts, or task is a failed
4389 * fork, never visible to cgroup_attach_task.
4391 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4393 struct cgroup_subsys *ss;
4394 struct css_set *cset;
4398 * Unlink from the css_set task list if necessary.
4399 * Optimistically check cg_list before taking
4402 if (!list_empty(&tsk->cg_list)) {
4403 write_lock(&css_set_lock);
4404 if (!list_empty(&tsk->cg_list))
4405 list_del_init(&tsk->cg_list);
4406 write_unlock(&css_set_lock);
4409 /* Reassign the task to the init_css_set. */
4411 cset = task_css_set(tsk);
4412 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4414 if (run_callbacks && need_forkexit_callback) {
4415 /* see cgroup_post_fork() for details */
4416 for_each_subsys(ss, i) {
4418 struct cgroup_subsys_state *old_css = cset->subsys[i];
4419 struct cgroup_subsys_state *css = task_css(tsk, i);
4421 ss->exit(css, old_css, tsk);
4427 put_css_set_taskexit(cset);
4430 static void check_for_release(struct cgroup *cgrp)
4432 if (cgroup_is_releasable(cgrp) &&
4433 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4435 * Control Group is currently removeable. If it's not
4436 * already queued for a userspace notification, queue
4439 int need_schedule_work = 0;
4441 raw_spin_lock(&release_list_lock);
4442 if (!cgroup_is_dead(cgrp) &&
4443 list_empty(&cgrp->release_list)) {
4444 list_add(&cgrp->release_list, &release_list);
4445 need_schedule_work = 1;
4447 raw_spin_unlock(&release_list_lock);
4448 if (need_schedule_work)
4449 schedule_work(&release_agent_work);
4454 * Notify userspace when a cgroup is released, by running the
4455 * configured release agent with the name of the cgroup (path
4456 * relative to the root of cgroup file system) as the argument.
4458 * Most likely, this user command will try to rmdir this cgroup.
4460 * This races with the possibility that some other task will be
4461 * attached to this cgroup before it is removed, or that some other
4462 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4463 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4464 * unused, and this cgroup will be reprieved from its death sentence,
4465 * to continue to serve a useful existence. Next time it's released,
4466 * we will get notified again, if it still has 'notify_on_release' set.
4468 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4469 * means only wait until the task is successfully execve()'d. The
4470 * separate release agent task is forked by call_usermodehelper(),
4471 * then control in this thread returns here, without waiting for the
4472 * release agent task. We don't bother to wait because the caller of
4473 * this routine has no use for the exit status of the release agent
4474 * task, so no sense holding our caller up for that.
4476 static void cgroup_release_agent(struct work_struct *work)
4478 BUG_ON(work != &release_agent_work);
4479 mutex_lock(&cgroup_mutex);
4480 raw_spin_lock(&release_list_lock);
4481 while (!list_empty(&release_list)) {
4482 char *argv[3], *envp[3];
4484 char *pathbuf = NULL, *agentbuf = NULL, *path;
4485 struct cgroup *cgrp = list_entry(release_list.next,
4488 list_del_init(&cgrp->release_list);
4489 raw_spin_unlock(&release_list_lock);
4490 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
4493 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
4496 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4501 argv[i++] = agentbuf;
4506 /* minimal command environment */
4507 envp[i++] = "HOME=/";
4508 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4511 /* Drop the lock while we invoke the usermode helper,
4512 * since the exec could involve hitting disk and hence
4513 * be a slow process */
4514 mutex_unlock(&cgroup_mutex);
4515 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4516 mutex_lock(&cgroup_mutex);
4520 raw_spin_lock(&release_list_lock);
4522 raw_spin_unlock(&release_list_lock);
4523 mutex_unlock(&cgroup_mutex);
4526 static int __init cgroup_disable(char *str)
4528 struct cgroup_subsys *ss;
4532 while ((token = strsep(&str, ",")) != NULL) {
4536 for_each_subsys(ss, i) {
4537 if (!strcmp(token, ss->name)) {
4539 printk(KERN_INFO "Disabling %s control group"
4540 " subsystem\n", ss->name);
4547 __setup("cgroup_disable=", cgroup_disable);
4550 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
4551 * @dentry: directory dentry of interest
4552 * @ss: subsystem of interest
4554 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
4555 * to get the corresponding css and return it. If such css doesn't exist
4556 * or can't be pinned, an ERR_PTR value is returned.
4558 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
4559 struct cgroup_subsys *ss)
4561 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4562 struct cgroup_subsys_state *css = NULL;
4563 struct cgroup *cgrp;
4565 /* is @dentry a cgroup dir? */
4566 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4567 kernfs_type(kn) != KERNFS_DIR)
4568 return ERR_PTR(-EBADF);
4573 * This path doesn't originate from kernfs and @kn could already
4574 * have been or be removed at any point. @kn->priv is RCU
4575 * protected for this access. See destroy_locked() for details.
4577 cgrp = rcu_dereference(kn->priv);
4579 css = cgroup_css(cgrp, ss);
4581 if (!css || !css_tryget(css))
4582 css = ERR_PTR(-ENOENT);
4589 * css_from_id - lookup css by id
4590 * @id: the cgroup id
4591 * @ss: cgroup subsys to be looked into
4593 * Returns the css if there's valid one with @id, otherwise returns NULL.
4594 * Should be called under rcu_read_lock().
4596 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
4598 struct cgroup *cgrp;
4600 cgroup_assert_mutexes_or_rcu_locked();
4602 cgrp = idr_find(&ss->root->cgroup_idr, id);
4604 return cgroup_css(cgrp, ss);
4608 #ifdef CONFIG_CGROUP_DEBUG
4609 static struct cgroup_subsys_state *
4610 debug_css_alloc(struct cgroup_subsys_state *parent_css)
4612 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4615 return ERR_PTR(-ENOMEM);
4620 static void debug_css_free(struct cgroup_subsys_state *css)
4625 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
4628 return cgroup_task_count(css->cgroup);
4631 static u64 current_css_set_read(struct cgroup_subsys_state *css,
4634 return (u64)(unsigned long)current->cgroups;
4637 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
4643 count = atomic_read(&task_css_set(current)->refcount);
4648 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
4650 struct cgrp_cset_link *link;
4651 struct css_set *cset;
4654 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
4658 read_lock(&css_set_lock);
4660 cset = rcu_dereference(current->cgroups);
4661 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
4662 struct cgroup *c = link->cgrp;
4663 const char *name = "?";
4665 if (c != cgroup_dummy_top) {
4666 cgroup_name(c, name_buf, NAME_MAX + 1);
4670 seq_printf(seq, "Root %d group %s\n",
4671 c->root->hierarchy_id, name);
4674 read_unlock(&css_set_lock);
4679 #define MAX_TASKS_SHOWN_PER_CSS 25
4680 static int cgroup_css_links_read(struct seq_file *seq, void *v)
4682 struct cgroup_subsys_state *css = seq_css(seq);
4683 struct cgrp_cset_link *link;
4685 read_lock(&css_set_lock);
4686 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
4687 struct css_set *cset = link->cset;
4688 struct task_struct *task;
4690 seq_printf(seq, "css_set %p\n", cset);
4691 list_for_each_entry(task, &cset->tasks, cg_list) {
4692 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4693 seq_puts(seq, " ...\n");
4696 seq_printf(seq, " task %d\n",
4697 task_pid_vnr(task));
4701 read_unlock(&css_set_lock);
4705 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
4707 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4710 static struct cftype debug_files[] = {
4712 .name = "taskcount",
4713 .read_u64 = debug_taskcount_read,
4717 .name = "current_css_set",
4718 .read_u64 = current_css_set_read,
4722 .name = "current_css_set_refcount",
4723 .read_u64 = current_css_set_refcount_read,
4727 .name = "current_css_set_cg_links",
4728 .seq_show = current_css_set_cg_links_read,
4732 .name = "cgroup_css_links",
4733 .seq_show = cgroup_css_links_read,
4737 .name = "releasable",
4738 .read_u64 = releasable_read,
4744 struct cgroup_subsys debug_cgrp_subsys = {
4745 .css_alloc = debug_css_alloc,
4746 .css_free = debug_css_free,
4747 .base_cftypes = debug_files,
4749 #endif /* CONFIG_CGROUP_DEBUG */