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>
57 #include <linux/atomic.h>
60 * pidlists linger the following amount before being destroyed. The goal
61 * is avoiding frequent destruction in the middle of consecutive read calls
62 * Expiring in the middle is a performance problem not a correctness one.
63 * 1 sec should be enough.
65 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
67 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
71 * cgroup_tree_mutex nests above cgroup_mutex and protects cftypes, file
72 * creation/removal and hierarchy changing operations including cgroup
73 * creation, removal, css association and controller rebinding. This outer
74 * lock is needed mainly to resolve the circular dependency between kernfs
75 * active ref and cgroup_mutex. cgroup_tree_mutex nests above both.
77 static DEFINE_MUTEX(cgroup_tree_mutex);
80 * cgroup_mutex is the master lock. Any modification to cgroup or its
81 * hierarchy must be performed while holding it.
83 #ifdef CONFIG_PROVE_RCU
84 DEFINE_MUTEX(cgroup_mutex);
85 EXPORT_SYMBOL_GPL(cgroup_mutex); /* only for lockdep */
87 static DEFINE_MUTEX(cgroup_mutex);
91 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
92 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
94 static DEFINE_SPINLOCK(release_agent_path_lock);
96 #define cgroup_assert_mutexes_or_rcu_locked() \
97 rcu_lockdep_assert(rcu_read_lock_held() || \
98 lockdep_is_held(&cgroup_tree_mutex) || \
99 lockdep_is_held(&cgroup_mutex), \
100 "cgroup_[tree_]mutex or RCU read lock required");
103 * cgroup destruction makes heavy use of work items and there can be a lot
104 * of concurrent destructions. Use a separate workqueue so that cgroup
105 * destruction work items don't end up filling up max_active of system_wq
106 * which may lead to deadlock.
108 static struct workqueue_struct *cgroup_destroy_wq;
111 * pidlist destructions need to be flushed on cgroup destruction. Use a
112 * separate workqueue as flush domain.
114 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
116 /* generate an array of cgroup subsystem pointers */
117 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
118 static struct cgroup_subsys *cgroup_subsys[] = {
119 #include <linux/cgroup_subsys.h>
123 /* array of cgroup subsystem names */
124 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
125 static const char *cgroup_subsys_name[] = {
126 #include <linux/cgroup_subsys.h>
131 * The dummy hierarchy, reserved for the subsystems that are otherwise
132 * unattached - it never has more than a single cgroup, and all tasks are
133 * part of that cgroup.
135 static struct cgroupfs_root cgroup_dummy_root;
137 /* dummy_top is a shorthand for the dummy hierarchy's top cgroup */
138 static struct cgroup * const cgroup_dummy_top = &cgroup_dummy_root.top_cgroup;
140 /* The list of hierarchy roots */
142 static LIST_HEAD(cgroup_roots);
143 static int cgroup_root_count;
145 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
146 static DEFINE_IDR(cgroup_hierarchy_idr);
149 * Assign a monotonically increasing serial number to cgroups. It
150 * guarantees cgroups with bigger numbers are newer than those with smaller
151 * numbers. Also, as cgroups are always appended to the parent's
152 * ->children list, it guarantees that sibling cgroups are always sorted in
153 * the ascending serial number order on the list. Protected by
156 static u64 cgroup_serial_nr_next = 1;
158 /* This flag indicates whether tasks in the fork and exit paths should
159 * check for fork/exit handlers to call. This avoids us having to do
160 * extra work in the fork/exit path if none of the subsystems need to
163 static int need_forkexit_callback __read_mostly;
165 static struct cftype cgroup_base_files[];
167 static void cgroup_put(struct cgroup *cgrp);
168 static int rebind_subsystems(struct cgroupfs_root *root,
169 unsigned long added_mask, unsigned removed_mask);
170 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
171 static int cgroup_destroy_locked(struct cgroup *cgrp);
172 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
174 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
177 * cgroup_css - obtain a cgroup's css for the specified subsystem
178 * @cgrp: the cgroup of interest
179 * @ss: the subsystem of interest (%NULL returns the dummy_css)
181 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
182 * function must be called either under cgroup_mutex or rcu_read_lock() and
183 * the caller is responsible for pinning the returned css if it wants to
184 * keep accessing it outside the said locks. This function may return
185 * %NULL if @cgrp doesn't have @subsys_id enabled.
187 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
188 struct cgroup_subsys *ss)
191 return rcu_dereference_check(cgrp->subsys[ss->id],
192 lockdep_is_held(&cgroup_tree_mutex) ||
193 lockdep_is_held(&cgroup_mutex));
195 return &cgrp->dummy_css;
198 /* convenient tests for these bits */
199 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
201 return test_bit(CGRP_DEAD, &cgrp->flags);
204 struct cgroup_subsys_state *seq_css(struct seq_file *seq)
206 struct kernfs_open_file *of = seq->private;
207 struct cgroup *cgrp = of->kn->parent->priv;
208 struct cftype *cft = seq_cft(seq);
211 * This is open and unprotected implementation of cgroup_css().
212 * seq_css() is only called from a kernfs file operation which has
213 * an active reference on the file. Because all the subsystem
214 * files are drained before a css is disassociated with a cgroup,
215 * the matching css from the cgroup's subsys table is guaranteed to
216 * be and stay valid until the enclosing operation is complete.
219 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
221 return &cgrp->dummy_css;
223 EXPORT_SYMBOL_GPL(seq_css);
226 * cgroup_is_descendant - test ancestry
227 * @cgrp: the cgroup to be tested
228 * @ancestor: possible ancestor of @cgrp
230 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
231 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
232 * and @ancestor are accessible.
234 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
237 if (cgrp == ancestor)
243 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
245 static int cgroup_is_releasable(const struct cgroup *cgrp)
248 (1 << CGRP_RELEASABLE) |
249 (1 << CGRP_NOTIFY_ON_RELEASE);
250 return (cgrp->flags & bits) == bits;
253 static int notify_on_release(const struct cgroup *cgrp)
255 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
259 * for_each_css - iterate all css's of a cgroup
260 * @css: the iteration cursor
261 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
262 * @cgrp: the target cgroup to iterate css's of
264 * Should be called under cgroup_mutex.
266 #define for_each_css(css, ssid, cgrp) \
267 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
268 if (!((css) = rcu_dereference_check( \
269 (cgrp)->subsys[(ssid)], \
270 lockdep_is_held(&cgroup_tree_mutex) || \
271 lockdep_is_held(&cgroup_mutex)))) { } \
275 * for_each_subsys - iterate all enabled cgroup subsystems
276 * @ss: the iteration cursor
277 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
279 #define for_each_subsys(ss, ssid) \
280 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
281 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
283 /* iterate across the active hierarchies */
284 #define for_each_active_root(root) \
285 list_for_each_entry((root), &cgroup_roots, root_list)
288 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
289 * @cgrp: the cgroup to be checked for liveness
291 * On success, returns true; the mutex should be later unlocked. On
292 * failure returns false with no lock held.
294 static bool cgroup_lock_live_group(struct cgroup *cgrp)
296 mutex_lock(&cgroup_mutex);
297 if (cgroup_is_dead(cgrp)) {
298 mutex_unlock(&cgroup_mutex);
304 /* the list of cgroups eligible for automatic release. Protected by
305 * release_list_lock */
306 static LIST_HEAD(release_list);
307 static DEFINE_RAW_SPINLOCK(release_list_lock);
308 static void cgroup_release_agent(struct work_struct *work);
309 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
310 static void check_for_release(struct cgroup *cgrp);
313 * A cgroup can be associated with multiple css_sets as different tasks may
314 * belong to different cgroups on different hierarchies. In the other
315 * direction, a css_set is naturally associated with multiple cgroups.
316 * This M:N relationship is represented by the following link structure
317 * which exists for each association and allows traversing the associations
320 struct cgrp_cset_link {
321 /* the cgroup and css_set this link associates */
323 struct css_set *cset;
325 /* list of cgrp_cset_links anchored at cgrp->cset_links */
326 struct list_head cset_link;
328 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
329 struct list_head cgrp_link;
332 /* The default css_set - used by init and its children prior to any
333 * hierarchies being mounted. It contains a pointer to the root state
334 * for each subsystem. Also used to anchor the list of css_sets. Not
335 * reference-counted, to improve performance when child cgroups
336 * haven't been created.
339 static struct css_set init_css_set;
340 static struct cgrp_cset_link init_cgrp_cset_link;
343 * css_set_lock protects the list of css_set objects, and the chain of
344 * tasks off each css_set. Nests outside task->alloc_lock due to
345 * css_task_iter_start().
347 static DEFINE_RWLOCK(css_set_lock);
348 static int css_set_count;
351 * hash table for cgroup groups. This improves the performance to find
352 * an existing css_set. This hash doesn't (currently) take into
353 * account cgroups in empty hierarchies.
355 #define CSS_SET_HASH_BITS 7
356 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
358 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
360 unsigned long key = 0UL;
361 struct cgroup_subsys *ss;
364 for_each_subsys(ss, i)
365 key += (unsigned long)css[i];
366 key = (key >> 16) ^ key;
372 * We don't maintain the lists running through each css_set to its task
373 * until after the first call to css_task_iter_start(). This reduces the
374 * fork()/exit() overhead for people who have cgroups compiled into their
375 * kernel but not actually in use.
377 static int use_task_css_set_links __read_mostly;
379 static void __put_css_set(struct css_set *cset, int taskexit)
381 struct cgrp_cset_link *link, *tmp_link;
384 * Ensure that the refcount doesn't hit zero while any readers
385 * can see it. Similar to atomic_dec_and_lock(), but for an
388 if (atomic_add_unless(&cset->refcount, -1, 1))
390 write_lock(&css_set_lock);
391 if (!atomic_dec_and_test(&cset->refcount)) {
392 write_unlock(&css_set_lock);
396 /* This css_set is dead. unlink it and release cgroup refcounts */
397 hash_del(&cset->hlist);
400 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
401 struct cgroup *cgrp = link->cgrp;
403 list_del(&link->cset_link);
404 list_del(&link->cgrp_link);
406 /* @cgrp can't go away while we're holding css_set_lock */
407 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
409 set_bit(CGRP_RELEASABLE, &cgrp->flags);
410 check_for_release(cgrp);
416 write_unlock(&css_set_lock);
417 kfree_rcu(cset, rcu_head);
421 * refcounted get/put for css_set objects
423 static inline void get_css_set(struct css_set *cset)
425 atomic_inc(&cset->refcount);
428 static inline void put_css_set(struct css_set *cset)
430 __put_css_set(cset, 0);
433 static inline void put_css_set_taskexit(struct css_set *cset)
435 __put_css_set(cset, 1);
439 * compare_css_sets - helper function for find_existing_css_set().
440 * @cset: candidate css_set being tested
441 * @old_cset: existing css_set for a task
442 * @new_cgrp: cgroup that's being entered by the task
443 * @template: desired set of css pointers in css_set (pre-calculated)
445 * Returns true if "cset" matches "old_cset" except for the hierarchy
446 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
448 static bool compare_css_sets(struct css_set *cset,
449 struct css_set *old_cset,
450 struct cgroup *new_cgrp,
451 struct cgroup_subsys_state *template[])
453 struct list_head *l1, *l2;
455 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
456 /* Not all subsystems matched */
461 * Compare cgroup pointers in order to distinguish between
462 * different cgroups in heirarchies with no subsystems. We
463 * could get by with just this check alone (and skip the
464 * memcmp above) but on most setups the memcmp check will
465 * avoid the need for this more expensive check on almost all
469 l1 = &cset->cgrp_links;
470 l2 = &old_cset->cgrp_links;
472 struct cgrp_cset_link *link1, *link2;
473 struct cgroup *cgrp1, *cgrp2;
477 /* See if we reached the end - both lists are equal length. */
478 if (l1 == &cset->cgrp_links) {
479 BUG_ON(l2 != &old_cset->cgrp_links);
482 BUG_ON(l2 == &old_cset->cgrp_links);
484 /* Locate the cgroups associated with these links. */
485 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
486 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
489 /* Hierarchies should be linked in the same order. */
490 BUG_ON(cgrp1->root != cgrp2->root);
493 * If this hierarchy is the hierarchy of the cgroup
494 * that's changing, then we need to check that this
495 * css_set points to the new cgroup; if it's any other
496 * hierarchy, then this css_set should point to the
497 * same cgroup as the old css_set.
499 if (cgrp1->root == new_cgrp->root) {
500 if (cgrp1 != new_cgrp)
511 * find_existing_css_set - init css array and find the matching css_set
512 * @old_cset: the css_set that we're using before the cgroup transition
513 * @cgrp: the cgroup that we're moving into
514 * @template: out param for the new set of csses, should be clear on entry
516 static struct css_set *find_existing_css_set(struct css_set *old_cset,
518 struct cgroup_subsys_state *template[])
520 struct cgroupfs_root *root = cgrp->root;
521 struct cgroup_subsys *ss;
522 struct css_set *cset;
527 * Build the set of subsystem state objects that we want to see in the
528 * new css_set. while subsystems can change globally, the entries here
529 * won't change, so no need for locking.
531 for_each_subsys(ss, i) {
532 if (root->subsys_mask & (1UL << i)) {
533 /* Subsystem is in this hierarchy. So we want
534 * the subsystem state from the new
536 template[i] = cgroup_css(cgrp, ss);
538 /* Subsystem is not in this hierarchy, so we
539 * don't want to change the subsystem state */
540 template[i] = old_cset->subsys[i];
544 key = css_set_hash(template);
545 hash_for_each_possible(css_set_table, cset, hlist, key) {
546 if (!compare_css_sets(cset, old_cset, cgrp, template))
549 /* This css_set matches what we need */
553 /* No existing cgroup group matched */
557 static void free_cgrp_cset_links(struct list_head *links_to_free)
559 struct cgrp_cset_link *link, *tmp_link;
561 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
562 list_del(&link->cset_link);
568 * allocate_cgrp_cset_links - allocate cgrp_cset_links
569 * @count: the number of links to allocate
570 * @tmp_links: list_head the allocated links are put on
572 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
573 * through ->cset_link. Returns 0 on success or -errno.
575 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
577 struct cgrp_cset_link *link;
580 INIT_LIST_HEAD(tmp_links);
582 for (i = 0; i < count; i++) {
583 link = kzalloc(sizeof(*link), GFP_KERNEL);
585 free_cgrp_cset_links(tmp_links);
588 list_add(&link->cset_link, tmp_links);
594 * link_css_set - a helper function to link a css_set to a cgroup
595 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
596 * @cset: the css_set to be linked
597 * @cgrp: the destination cgroup
599 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
602 struct cgrp_cset_link *link;
604 BUG_ON(list_empty(tmp_links));
605 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
608 list_move(&link->cset_link, &cgrp->cset_links);
610 * Always add links to the tail of the list so that the list
611 * is sorted by order of hierarchy creation
613 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
617 * find_css_set - return a new css_set with one cgroup updated
618 * @old_cset: the baseline css_set
619 * @cgrp: the cgroup to be updated
621 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
622 * substituted into the appropriate hierarchy.
624 static struct css_set *find_css_set(struct css_set *old_cset,
627 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
628 struct css_set *cset;
629 struct list_head tmp_links;
630 struct cgrp_cset_link *link;
633 lockdep_assert_held(&cgroup_mutex);
635 /* First see if we already have a cgroup group that matches
637 read_lock(&css_set_lock);
638 cset = find_existing_css_set(old_cset, cgrp, template);
641 read_unlock(&css_set_lock);
646 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
650 /* Allocate all the cgrp_cset_link objects that we'll need */
651 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
656 atomic_set(&cset->refcount, 1);
657 INIT_LIST_HEAD(&cset->cgrp_links);
658 INIT_LIST_HEAD(&cset->tasks);
659 INIT_HLIST_NODE(&cset->hlist);
661 /* Copy the set of subsystem state objects generated in
662 * find_existing_css_set() */
663 memcpy(cset->subsys, template, sizeof(cset->subsys));
665 write_lock(&css_set_lock);
666 /* Add reference counts and links from the new css_set. */
667 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
668 struct cgroup *c = link->cgrp;
670 if (c->root == cgrp->root)
672 link_css_set(&tmp_links, cset, c);
675 BUG_ON(!list_empty(&tmp_links));
679 /* Add this cgroup group to the hash table */
680 key = css_set_hash(cset->subsys);
681 hash_add(css_set_table, &cset->hlist, key);
683 write_unlock(&css_set_lock);
688 static struct cgroupfs_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
690 struct cgroup *top_cgrp = kf_root->kn->priv;
692 return top_cgrp->root;
695 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
699 lockdep_assert_held(&cgroup_mutex);
701 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
706 root->hierarchy_id = id;
710 static void cgroup_exit_root_id(struct cgroupfs_root *root)
712 lockdep_assert_held(&cgroup_mutex);
714 if (root->hierarchy_id) {
715 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
716 root->hierarchy_id = 0;
720 static void cgroup_free_root(struct cgroupfs_root *root)
723 /* hierarhcy ID shoulid already have been released */
724 WARN_ON_ONCE(root->hierarchy_id);
726 idr_destroy(&root->cgroup_idr);
731 static void cgroup_get_root(struct cgroupfs_root *root)
734 * The caller must ensure that @root is alive, which can be
735 * achieved by holding a ref on one of the member cgroups or
736 * following a registered reference to @root while holding
739 WARN_ON_ONCE(atomic_read(&root->refcnt) <= 0);
740 atomic_inc(&root->refcnt);
743 static void cgroup_put_root(struct cgroupfs_root *root)
745 struct cgroup *cgrp = &root->top_cgroup;
746 struct cgrp_cset_link *link, *tmp_link;
750 * @root's refcnt reaching zero and its deregistration should be
751 * atomic w.r.t. cgroup_tree_mutex. This ensures that
752 * cgroup_get_root() is safe to invoke if @root is registered.
754 mutex_lock(&cgroup_tree_mutex);
755 if (!atomic_dec_and_test(&root->refcnt)) {
756 mutex_unlock(&cgroup_tree_mutex);
759 mutex_lock(&cgroup_mutex);
761 BUG_ON(root->number_of_cgroups != 1);
762 BUG_ON(!list_empty(&cgrp->children));
764 /* Rebind all subsystems back to the default hierarchy */
765 if (root->flags & CGRP_ROOT_SUBSYS_BOUND) {
766 ret = rebind_subsystems(root, 0, root->subsys_mask);
767 /* Shouldn't be able to fail ... */
772 * Release all the links from cset_links to this hierarchy's
775 write_lock(&css_set_lock);
777 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
778 list_del(&link->cset_link);
779 list_del(&link->cgrp_link);
782 write_unlock(&css_set_lock);
784 if (!list_empty(&root->root_list)) {
785 list_del(&root->root_list);
789 cgroup_exit_root_id(root);
791 mutex_unlock(&cgroup_mutex);
792 mutex_unlock(&cgroup_tree_mutex);
794 kernfs_destroy_root(root->kf_root);
795 cgroup_free_root(root);
799 * Return the cgroup for "task" from the given hierarchy. Must be
800 * called with cgroup_mutex held.
802 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
803 struct cgroupfs_root *root)
805 struct css_set *cset;
806 struct cgroup *res = NULL;
808 BUG_ON(!mutex_is_locked(&cgroup_mutex));
809 read_lock(&css_set_lock);
811 * No need to lock the task - since we hold cgroup_mutex the
812 * task can't change groups, so the only thing that can happen
813 * is that it exits and its css is set back to init_css_set.
815 cset = task_css_set(task);
816 if (cset == &init_css_set) {
817 res = &root->top_cgroup;
819 struct cgrp_cset_link *link;
821 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
822 struct cgroup *c = link->cgrp;
824 if (c->root == root) {
830 read_unlock(&css_set_lock);
836 * There is one global cgroup mutex. We also require taking
837 * task_lock() when dereferencing a task's cgroup subsys pointers.
838 * See "The task_lock() exception", at the end of this comment.
840 * A task must hold cgroup_mutex to modify cgroups.
842 * Any task can increment and decrement the count field without lock.
843 * So in general, code holding cgroup_mutex can't rely on the count
844 * field not changing. However, if the count goes to zero, then only
845 * cgroup_attach_task() can increment it again. Because a count of zero
846 * means that no tasks are currently attached, therefore there is no
847 * way a task attached to that cgroup can fork (the other way to
848 * increment the count). So code holding cgroup_mutex can safely
849 * assume that if the count is zero, it will stay zero. Similarly, if
850 * a task holds cgroup_mutex on a cgroup with zero count, it
851 * knows that the cgroup won't be removed, as cgroup_rmdir()
854 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
855 * (usually) take cgroup_mutex. These are the two most performance
856 * critical pieces of code here. The exception occurs on cgroup_exit(),
857 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
858 * is taken, and if the cgroup count is zero, a usermode call made
859 * to the release agent with the name of the cgroup (path relative to
860 * the root of cgroup file system) as the argument.
862 * A cgroup can only be deleted if both its 'count' of using tasks
863 * is zero, and its list of 'children' cgroups is empty. Since all
864 * tasks in the system use _some_ cgroup, and since there is always at
865 * least one task in the system (init, pid == 1), therefore, top_cgroup
866 * always has either children cgroups and/or using tasks. So we don't
867 * need a special hack to ensure that top_cgroup cannot be deleted.
869 * The task_lock() exception
871 * The need for this exception arises from the action of
872 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
873 * another. It does so using cgroup_mutex, however there are
874 * several performance critical places that need to reference
875 * task->cgroup without the expense of grabbing a system global
876 * mutex. Therefore except as noted below, when dereferencing or, as
877 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
878 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
879 * the task_struct routinely used for such matters.
881 * P.S. One more locking exception. RCU is used to guard the
882 * update of a tasks cgroup pointer by cgroup_attach_task()
885 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
886 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
887 static const struct file_operations proc_cgroupstats_operations;
889 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
892 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
893 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
894 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
895 cft->ss->name, cft->name);
897 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
902 * cgroup_file_mode - deduce file mode of a control file
903 * @cft: the control file in question
905 * returns cft->mode if ->mode is not 0
906 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
907 * returns S_IRUGO if it has only a read handler
908 * returns S_IWUSR if it has only a write hander
910 static umode_t cgroup_file_mode(const struct cftype *cft)
917 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
920 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
927 static void cgroup_free_fn(struct work_struct *work)
929 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
931 mutex_lock(&cgroup_mutex);
932 cgrp->root->number_of_cgroups--;
933 mutex_unlock(&cgroup_mutex);
936 * We get a ref to the parent, and put the ref when this cgroup is
937 * being freed, so it's guaranteed that the parent won't be
938 * destroyed before its children.
940 cgroup_put(cgrp->parent);
942 /* put the root reference that we took when we created the cgroup */
943 cgroup_put_root(cgrp->root);
945 cgroup_pidlist_destroy_all(cgrp);
947 kernfs_put(cgrp->kn);
951 static void cgroup_free_rcu(struct rcu_head *head)
953 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
955 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
956 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
959 static void cgroup_get(struct cgroup *cgrp)
961 WARN_ON_ONCE(cgroup_is_dead(cgrp));
962 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
963 atomic_inc(&cgrp->refcnt);
966 static void cgroup_put(struct cgroup *cgrp)
968 if (!atomic_dec_and_test(&cgrp->refcnt))
970 if (WARN_ON_ONCE(!cgroup_is_dead(cgrp)))
974 * XXX: cgrp->id is only used to look up css's. As cgroup and
975 * css's lifetimes will be decoupled, it should be made
976 * per-subsystem and moved to css->id so that lookups are
977 * successful until the target css is released.
979 mutex_lock(&cgroup_mutex);
980 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
981 mutex_unlock(&cgroup_mutex);
984 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
987 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
989 char name[CGROUP_FILE_NAME_MAX];
991 lockdep_assert_held(&cgroup_tree_mutex);
992 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
996 * cgroup_clear_dir - remove subsys files in a cgroup directory
997 * @cgrp: target cgroup
998 * @subsys_mask: mask of the subsystem ids whose files should be removed
1000 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
1002 struct cgroup_subsys *ss;
1005 for_each_subsys(ss, i) {
1006 struct cftype *cfts;
1008 if (!test_bit(i, &subsys_mask))
1010 list_for_each_entry(cfts, &ss->cfts, node)
1011 cgroup_addrm_files(cgrp, cfts, false);
1015 static int rebind_subsystems(struct cgroupfs_root *root,
1016 unsigned long added_mask, unsigned removed_mask)
1018 struct cgroup *cgrp = &root->top_cgroup;
1019 struct cgroup_subsys *ss;
1022 lockdep_assert_held(&cgroup_tree_mutex);
1023 lockdep_assert_held(&cgroup_mutex);
1025 /* Check that any added subsystems are currently free */
1026 for_each_subsys(ss, i)
1027 if ((added_mask & (1 << i)) && ss->root != &cgroup_dummy_root)
1030 ret = cgroup_populate_dir(cgrp, added_mask);
1035 * Nothing can fail from this point on. Remove files for the
1036 * removed subsystems and rebind each subsystem.
1038 mutex_unlock(&cgroup_mutex);
1039 cgroup_clear_dir(cgrp, removed_mask);
1040 mutex_lock(&cgroup_mutex);
1042 for_each_subsys(ss, i) {
1043 unsigned long bit = 1UL << i;
1045 if (bit & added_mask) {
1046 /* We're binding this subsystem to this hierarchy */
1047 BUG_ON(cgroup_css(cgrp, ss));
1048 BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
1049 BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
1051 rcu_assign_pointer(cgrp->subsys[i],
1052 cgroup_css(cgroup_dummy_top, ss));
1053 cgroup_css(cgrp, ss)->cgroup = cgrp;
1057 ss->bind(cgroup_css(cgrp, ss));
1059 /* refcount was already taken, and we're keeping it */
1060 root->subsys_mask |= bit;
1061 } else if (bit & removed_mask) {
1062 /* We're removing this subsystem */
1063 BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1064 BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1067 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1069 cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1070 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1072 cgroup_subsys[i]->root = &cgroup_dummy_root;
1073 root->subsys_mask &= ~bit;
1078 * Mark @root has finished binding subsystems. @root->subsys_mask
1079 * now matches the bound subsystems.
1081 root->flags |= CGRP_ROOT_SUBSYS_BOUND;
1082 kernfs_activate(cgrp->kn);
1087 static int cgroup_show_options(struct seq_file *seq,
1088 struct kernfs_root *kf_root)
1090 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1091 struct cgroup_subsys *ss;
1094 for_each_subsys(ss, ssid)
1095 if (root->subsys_mask & (1 << ssid))
1096 seq_printf(seq, ",%s", ss->name);
1097 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1098 seq_puts(seq, ",sane_behavior");
1099 if (root->flags & CGRP_ROOT_NOPREFIX)
1100 seq_puts(seq, ",noprefix");
1101 if (root->flags & CGRP_ROOT_XATTR)
1102 seq_puts(seq, ",xattr");
1104 spin_lock(&release_agent_path_lock);
1105 if (strlen(root->release_agent_path))
1106 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1107 spin_unlock(&release_agent_path_lock);
1109 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1110 seq_puts(seq, ",clone_children");
1111 if (strlen(root->name))
1112 seq_printf(seq, ",name=%s", root->name);
1116 struct cgroup_sb_opts {
1117 unsigned long subsys_mask;
1118 unsigned long flags;
1119 char *release_agent;
1120 bool cpuset_clone_children;
1122 /* User explicitly requested empty subsystem */
1127 * Convert a hierarchy specifier into a bitmask of subsystems and
1128 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1129 * array. This function takes refcounts on subsystems to be used, unless it
1130 * returns error, in which case no refcounts are taken.
1132 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1134 char *token, *o = data;
1135 bool all_ss = false, one_ss = false;
1136 unsigned long mask = (unsigned long)-1;
1137 struct cgroup_subsys *ss;
1140 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1142 #ifdef CONFIG_CPUSETS
1143 mask = ~(1UL << cpuset_cgrp_id);
1146 memset(opts, 0, sizeof(*opts));
1148 while ((token = strsep(&o, ",")) != NULL) {
1151 if (!strcmp(token, "none")) {
1152 /* Explicitly have no subsystems */
1156 if (!strcmp(token, "all")) {
1157 /* Mutually exclusive option 'all' + subsystem name */
1163 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1164 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1167 if (!strcmp(token, "noprefix")) {
1168 opts->flags |= CGRP_ROOT_NOPREFIX;
1171 if (!strcmp(token, "clone_children")) {
1172 opts->cpuset_clone_children = true;
1175 if (!strcmp(token, "xattr")) {
1176 opts->flags |= CGRP_ROOT_XATTR;
1179 if (!strncmp(token, "release_agent=", 14)) {
1180 /* Specifying two release agents is forbidden */
1181 if (opts->release_agent)
1183 opts->release_agent =
1184 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1185 if (!opts->release_agent)
1189 if (!strncmp(token, "name=", 5)) {
1190 const char *name = token + 5;
1191 /* Can't specify an empty name */
1194 /* Must match [\w.-]+ */
1195 for (i = 0; i < strlen(name); i++) {
1199 if ((c == '.') || (c == '-') || (c == '_'))
1203 /* Specifying two names is forbidden */
1206 opts->name = kstrndup(name,
1207 MAX_CGROUP_ROOT_NAMELEN - 1,
1215 for_each_subsys(ss, i) {
1216 if (strcmp(token, ss->name))
1221 /* Mutually exclusive option 'all' + subsystem name */
1224 set_bit(i, &opts->subsys_mask);
1229 if (i == CGROUP_SUBSYS_COUNT)
1234 * If the 'all' option was specified select all the subsystems,
1235 * otherwise if 'none', 'name=' and a subsystem name options
1236 * were not specified, let's default to 'all'
1238 if (all_ss || (!one_ss && !opts->none && !opts->name))
1239 for_each_subsys(ss, i)
1241 set_bit(i, &opts->subsys_mask);
1243 /* Consistency checks */
1245 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1246 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1248 if (opts->flags & CGRP_ROOT_NOPREFIX) {
1249 pr_err("cgroup: sane_behavior: noprefix is not allowed\n");
1253 if (opts->cpuset_clone_children) {
1254 pr_err("cgroup: sane_behavior: clone_children is not allowed\n");
1258 if (opts->flags & CGRP_ROOT_XATTR)
1259 pr_warning("cgroup: sane_behavior: xattr is always available, flag unnecessary\n");
1263 * Option noprefix was introduced just for backward compatibility
1264 * with the old cpuset, so we allow noprefix only if mounting just
1265 * the cpuset subsystem.
1267 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1271 /* Can't specify "none" and some subsystems */
1272 if (opts->subsys_mask && opts->none)
1276 * We either have to specify by name or by subsystems. (So all
1277 * empty hierarchies must have a name).
1279 if (!opts->subsys_mask && !opts->name)
1285 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1288 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1289 struct cgroup_sb_opts opts;
1290 unsigned long added_mask, removed_mask;
1292 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1293 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1297 mutex_lock(&cgroup_tree_mutex);
1298 mutex_lock(&cgroup_mutex);
1300 /* See what subsystems are wanted */
1301 ret = parse_cgroupfs_options(data, &opts);
1305 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1306 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1307 task_tgid_nr(current), current->comm);
1309 added_mask = opts.subsys_mask & ~root->subsys_mask;
1310 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1312 /* Don't allow flags or name to change at remount */
1313 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1314 (opts.name && strcmp(opts.name, root->name))) {
1315 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1316 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1317 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1322 /* remounting is not allowed for populated hierarchies */
1323 if (root->number_of_cgroups > 1) {
1328 ret = rebind_subsystems(root, added_mask, removed_mask);
1332 if (opts.release_agent) {
1333 spin_lock(&release_agent_path_lock);
1334 strcpy(root->release_agent_path, opts.release_agent);
1335 spin_unlock(&release_agent_path_lock);
1338 kfree(opts.release_agent);
1340 mutex_unlock(&cgroup_mutex);
1341 mutex_unlock(&cgroup_tree_mutex);
1345 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1347 atomic_set(&cgrp->refcnt, 1);
1348 INIT_LIST_HEAD(&cgrp->sibling);
1349 INIT_LIST_HEAD(&cgrp->children);
1350 INIT_LIST_HEAD(&cgrp->cset_links);
1351 INIT_LIST_HEAD(&cgrp->release_list);
1352 INIT_LIST_HEAD(&cgrp->pidlists);
1353 mutex_init(&cgrp->pidlist_mutex);
1354 cgrp->dummy_css.cgroup = cgrp;
1357 static void init_cgroup_root(struct cgroupfs_root *root)
1359 struct cgroup *cgrp = &root->top_cgroup;
1361 atomic_set(&root->refcnt, 1);
1362 INIT_LIST_HEAD(&root->root_list);
1363 root->number_of_cgroups = 1;
1365 init_cgroup_housekeeping(cgrp);
1366 idr_init(&root->cgroup_idr);
1369 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1371 struct cgroupfs_root *root;
1373 if (!opts->subsys_mask && !opts->none)
1374 return ERR_PTR(-EINVAL);
1376 root = kzalloc(sizeof(*root), GFP_KERNEL);
1378 return ERR_PTR(-ENOMEM);
1380 init_cgroup_root(root);
1383 * We need to set @root->subsys_mask now so that @root can be
1384 * matched by cgroup_test_super() before it finishes
1385 * initialization; otherwise, competing mounts with the same
1386 * options may try to bind the same subsystems instead of waiting
1387 * for the first one leading to unexpected mount errors.
1388 * SUBSYS_BOUND will be set once actual binding is complete.
1390 root->subsys_mask = opts->subsys_mask;
1391 root->flags = opts->flags;
1392 if (opts->release_agent)
1393 strcpy(root->release_agent_path, opts->release_agent);
1395 strcpy(root->name, opts->name);
1396 if (opts->cpuset_clone_children)
1397 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1401 static int cgroup_setup_root(struct cgroupfs_root *root)
1403 LIST_HEAD(tmp_links);
1404 struct cgroup *root_cgrp = &root->top_cgroup;
1405 struct css_set *cset;
1408 lockdep_assert_held(&cgroup_tree_mutex);
1409 lockdep_assert_held(&cgroup_mutex);
1411 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1414 root_cgrp->id = ret;
1417 * We're accessing css_set_count without locking css_set_lock here,
1418 * but that's OK - it can only be increased by someone holding
1419 * cgroup_lock, and that's us. The worst that can happen is that we
1420 * have some link structures left over
1422 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1426 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1427 ret = cgroup_init_root_id(root, 2, 0);
1431 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1432 KERNFS_ROOT_CREATE_DEACTIVATED,
1434 if (IS_ERR(root->kf_root)) {
1435 ret = PTR_ERR(root->kf_root);
1438 root_cgrp->kn = root->kf_root->kn;
1440 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1444 ret = rebind_subsystems(root, root->subsys_mask, 0);
1449 * There must be no failure case after here, since rebinding takes
1450 * care of subsystems' refcounts, which are explicitly dropped in
1451 * the failure exit path.
1453 list_add(&root->root_list, &cgroup_roots);
1454 cgroup_root_count++;
1457 * Link the top cgroup in this hierarchy into all the css_set
1460 write_lock(&css_set_lock);
1461 hash_for_each(css_set_table, i, cset, hlist)
1462 link_css_set(&tmp_links, cset, root_cgrp);
1463 write_unlock(&css_set_lock);
1465 BUG_ON(!list_empty(&root_cgrp->children));
1466 BUG_ON(root->number_of_cgroups != 1);
1468 kernfs_activate(root_cgrp->kn);
1473 kernfs_destroy_root(root->kf_root);
1474 root->kf_root = NULL;
1476 cgroup_exit_root_id(root);
1478 free_cgrp_cset_links(&tmp_links);
1482 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1483 int flags, const char *unused_dev_name,
1486 struct cgroupfs_root *root;
1487 struct cgroup_sb_opts opts;
1488 struct dentry *dentry;
1491 mutex_lock(&cgroup_tree_mutex);
1492 mutex_lock(&cgroup_mutex);
1494 /* First find the desired set of subsystems */
1495 ret = parse_cgroupfs_options(data, &opts);
1499 /* look for a matching existing root */
1500 for_each_active_root(root) {
1501 bool name_match = false;
1504 * If we asked for a name then it must match. Also, if
1505 * name matches but sybsys_mask doesn't, we should fail.
1506 * Remember whether name matched.
1509 if (strcmp(opts.name, root->name))
1515 * If we asked for subsystems (or explicitly for no
1516 * subsystems) then they must match.
1518 if ((opts.subsys_mask || opts.none) &&
1519 (opts.subsys_mask != root->subsys_mask)) {
1526 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1527 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1528 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1532 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1536 cgroup_get_root(root);
1540 /* no such thing, create a new one */
1541 root = cgroup_root_from_opts(&opts);
1543 ret = PTR_ERR(root);
1547 ret = cgroup_setup_root(root);
1549 cgroup_free_root(root);
1552 mutex_unlock(&cgroup_mutex);
1553 mutex_unlock(&cgroup_tree_mutex);
1555 kfree(opts.release_agent);
1559 return ERR_PTR(ret);
1561 dentry = kernfs_mount(fs_type, flags, root->kf_root);
1563 cgroup_put_root(root);
1567 static void cgroup_kill_sb(struct super_block *sb)
1569 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1570 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1572 cgroup_put_root(root);
1576 static struct file_system_type cgroup_fs_type = {
1578 .mount = cgroup_mount,
1579 .kill_sb = cgroup_kill_sb,
1582 static struct kobject *cgroup_kobj;
1585 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1586 * @task: target task
1587 * @buf: the buffer to write the path into
1588 * @buflen: the length of the buffer
1590 * Determine @task's cgroup on the first (the one with the lowest non-zero
1591 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1592 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1593 * cgroup controller callbacks.
1595 * Return value is the same as kernfs_path().
1597 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1599 struct cgroupfs_root *root;
1600 struct cgroup *cgrp;
1601 int hierarchy_id = 1;
1604 mutex_lock(&cgroup_mutex);
1606 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1609 cgrp = task_cgroup_from_root(task, root);
1610 path = cgroup_path(cgrp, buf, buflen);
1612 /* if no hierarchy exists, everyone is in "/" */
1613 if (strlcpy(buf, "/", buflen) < buflen)
1617 mutex_unlock(&cgroup_mutex);
1620 EXPORT_SYMBOL_GPL(task_cgroup_path);
1623 * Control Group taskset
1625 struct task_and_cgroup {
1626 struct task_struct *task;
1627 struct cgroup *cgrp;
1628 struct css_set *cset;
1631 struct cgroup_taskset {
1632 struct task_and_cgroup single;
1633 struct flex_array *tc_array;
1636 struct cgroup *cur_cgrp;
1640 * cgroup_taskset_first - reset taskset and return the first task
1641 * @tset: taskset of interest
1643 * @tset iteration is initialized and the first task is returned.
1645 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1647 if (tset->tc_array) {
1649 return cgroup_taskset_next(tset);
1651 tset->cur_cgrp = tset->single.cgrp;
1652 return tset->single.task;
1655 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1658 * cgroup_taskset_next - iterate to the next task in taskset
1659 * @tset: taskset of interest
1661 * Return the next task in @tset. Iteration must have been initialized
1662 * with cgroup_taskset_first().
1664 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1666 struct task_and_cgroup *tc;
1668 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1671 tc = flex_array_get(tset->tc_array, tset->idx++);
1672 tset->cur_cgrp = tc->cgrp;
1675 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1678 * cgroup_taskset_cur_css - return the matching css for the current task
1679 * @tset: taskset of interest
1680 * @subsys_id: the ID of the target subsystem
1682 * Return the css for the current (last returned) task of @tset for
1683 * subsystem specified by @subsys_id. This function must be preceded by
1684 * either cgroup_taskset_first() or cgroup_taskset_next().
1686 struct cgroup_subsys_state *cgroup_taskset_cur_css(struct cgroup_taskset *tset,
1689 return cgroup_css(tset->cur_cgrp, cgroup_subsys[subsys_id]);
1691 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_css);
1694 * cgroup_taskset_size - return the number of tasks in taskset
1695 * @tset: taskset of interest
1697 int cgroup_taskset_size(struct cgroup_taskset *tset)
1699 return tset->tc_array ? tset->tc_array_len : 1;
1701 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1705 * cgroup_task_migrate - move a task from one cgroup to another.
1707 * Must be called with cgroup_mutex and threadgroup locked.
1709 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1710 struct task_struct *tsk,
1711 struct css_set *new_cset)
1713 struct css_set *old_cset;
1716 * We are synchronized through threadgroup_lock() against PF_EXITING
1717 * setting such that we can't race against cgroup_exit() changing the
1718 * css_set to init_css_set and dropping the old one.
1720 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1721 old_cset = task_css_set(tsk);
1724 rcu_assign_pointer(tsk->cgroups, new_cset);
1727 /* Update the css_set linked lists if we're using them */
1728 write_lock(&css_set_lock);
1729 if (!list_empty(&tsk->cg_list))
1730 list_move(&tsk->cg_list, &new_cset->tasks);
1731 write_unlock(&css_set_lock);
1734 * We just gained a reference on old_cset by taking it from the
1735 * task. As trading it for new_cset is protected by cgroup_mutex,
1736 * we're safe to drop it here; it will be freed under RCU.
1738 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1739 put_css_set(old_cset);
1743 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1744 * @cgrp: the cgroup to attach to
1745 * @tsk: the task or the leader of the threadgroup to be attached
1746 * @threadgroup: attach the whole threadgroup?
1748 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1749 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1751 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1754 int retval, i, group_size;
1755 struct cgroupfs_root *root = cgrp->root;
1756 struct cgroup_subsys_state *css, *failed_css = NULL;
1757 /* threadgroup list cursor and array */
1758 struct task_struct *leader = tsk;
1759 struct task_and_cgroup *tc;
1760 struct flex_array *group;
1761 struct cgroup_taskset tset = { };
1764 * step 0: in order to do expensive, possibly blocking operations for
1765 * every thread, we cannot iterate the thread group list, since it needs
1766 * rcu or tasklist locked. instead, build an array of all threads in the
1767 * group - group_rwsem prevents new threads from appearing, and if
1768 * threads exit, this will just be an over-estimate.
1771 group_size = get_nr_threads(tsk);
1774 /* flex_array supports very large thread-groups better than kmalloc. */
1775 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1778 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1779 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1781 goto out_free_group_list;
1785 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1786 * already PF_EXITING could be freed from underneath us unless we
1787 * take an rcu_read_lock.
1791 struct task_and_cgroup ent;
1793 /* @tsk either already exited or can't exit until the end */
1794 if (tsk->flags & PF_EXITING)
1797 /* as per above, nr_threads may decrease, but not increase. */
1798 BUG_ON(i >= group_size);
1800 ent.cgrp = task_cgroup_from_root(tsk, root);
1801 /* nothing to do if this task is already in the cgroup */
1802 if (ent.cgrp == cgrp)
1805 * saying GFP_ATOMIC has no effect here because we did prealloc
1806 * earlier, but it's good form to communicate our expectations.
1808 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
1809 BUG_ON(retval != 0);
1814 } while_each_thread(leader, tsk);
1816 /* remember the number of threads in the array for later. */
1818 tset.tc_array = group;
1819 tset.tc_array_len = group_size;
1821 /* methods shouldn't be called if no task is actually migrating */
1824 goto out_free_group_list;
1827 * step 1: check that we can legitimately attach to the cgroup.
1829 for_each_css(css, i, cgrp) {
1830 if (css->ss->can_attach) {
1831 retval = css->ss->can_attach(css, &tset);
1834 goto out_cancel_attach;
1840 * step 2: make sure css_sets exist for all threads to be migrated.
1841 * we use find_css_set, which allocates a new one if necessary.
1843 for (i = 0; i < group_size; i++) {
1844 struct css_set *old_cset;
1846 tc = flex_array_get(group, i);
1847 old_cset = task_css_set(tc->task);
1848 tc->cset = find_css_set(old_cset, cgrp);
1851 goto out_put_css_set_refs;
1856 * step 3: now that we're guaranteed success wrt the css_sets,
1857 * proceed to move all tasks to the new cgroup. There are no
1858 * failure cases after here, so this is the commit point.
1860 for (i = 0; i < group_size; i++) {
1861 tc = flex_array_get(group, i);
1862 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
1864 /* nothing is sensitive to fork() after this point. */
1867 * step 4: do subsystem attach callbacks.
1869 for_each_css(css, i, cgrp)
1870 if (css->ss->attach)
1871 css->ss->attach(css, &tset);
1874 * step 5: success! and cleanup
1877 out_put_css_set_refs:
1879 for (i = 0; i < group_size; i++) {
1880 tc = flex_array_get(group, i);
1883 put_css_set(tc->cset);
1888 for_each_css(css, i, cgrp) {
1889 if (css == failed_css)
1891 if (css->ss->cancel_attach)
1892 css->ss->cancel_attach(css, &tset);
1895 out_free_group_list:
1896 flex_array_free(group);
1901 * Find the task_struct of the task to attach by vpid and pass it along to the
1902 * function to attach either it or all tasks in its threadgroup. Will lock
1903 * cgroup_mutex and threadgroup; may take task_lock of task.
1905 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
1907 struct task_struct *tsk;
1908 const struct cred *cred = current_cred(), *tcred;
1911 if (!cgroup_lock_live_group(cgrp))
1917 tsk = find_task_by_vpid(pid);
1921 goto out_unlock_cgroup;
1924 * even if we're attaching all tasks in the thread group, we
1925 * only need to check permissions on one of them.
1927 tcred = __task_cred(tsk);
1928 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
1929 !uid_eq(cred->euid, tcred->uid) &&
1930 !uid_eq(cred->euid, tcred->suid)) {
1933 goto out_unlock_cgroup;
1939 tsk = tsk->group_leader;
1942 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
1943 * trapped in a cpuset, or RT worker may be born in a cgroup
1944 * with no rt_runtime allocated. Just say no.
1946 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
1949 goto out_unlock_cgroup;
1952 get_task_struct(tsk);
1955 threadgroup_lock(tsk);
1957 if (!thread_group_leader(tsk)) {
1959 * a race with de_thread from another thread's exec()
1960 * may strip us of our leadership, if this happens,
1961 * there is no choice but to throw this task away and
1962 * try again; this is
1963 * "double-double-toil-and-trouble-check locking".
1965 threadgroup_unlock(tsk);
1966 put_task_struct(tsk);
1967 goto retry_find_task;
1971 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
1973 threadgroup_unlock(tsk);
1975 put_task_struct(tsk);
1977 mutex_unlock(&cgroup_mutex);
1982 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1983 * @from: attach to all cgroups of a given task
1984 * @tsk: the task to be attached
1986 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1988 struct cgroupfs_root *root;
1991 mutex_lock(&cgroup_mutex);
1992 for_each_active_root(root) {
1993 struct cgroup *from_cgrp = task_cgroup_from_root(from, root);
1995 retval = cgroup_attach_task(from_cgrp, tsk, false);
1999 mutex_unlock(&cgroup_mutex);
2003 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2005 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2006 struct cftype *cft, u64 pid)
2008 return attach_task_by_pid(css->cgroup, pid, false);
2011 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2012 struct cftype *cft, u64 tgid)
2014 return attach_task_by_pid(css->cgroup, tgid, true);
2017 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2018 struct cftype *cft, const char *buffer)
2020 struct cgroupfs_root *root = css->cgroup->root;
2022 BUILD_BUG_ON(sizeof(root->release_agent_path) < PATH_MAX);
2023 if (!cgroup_lock_live_group(css->cgroup))
2025 spin_lock(&release_agent_path_lock);
2026 strlcpy(root->release_agent_path, buffer,
2027 sizeof(root->release_agent_path));
2028 spin_unlock(&release_agent_path_lock);
2029 mutex_unlock(&cgroup_mutex);
2033 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2035 struct cgroup *cgrp = seq_css(seq)->cgroup;
2037 if (!cgroup_lock_live_group(cgrp))
2039 seq_puts(seq, cgrp->root->release_agent_path);
2040 seq_putc(seq, '\n');
2041 mutex_unlock(&cgroup_mutex);
2045 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2047 struct cgroup *cgrp = seq_css(seq)->cgroup;
2049 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2053 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2054 size_t nbytes, loff_t off)
2056 struct cgroup *cgrp = of->kn->parent->priv;
2057 struct cftype *cft = of->kn->priv;
2058 struct cgroup_subsys_state *css;
2062 * kernfs guarantees that a file isn't deleted with operations in
2063 * flight, which means that the matching css is and stays alive and
2064 * doesn't need to be pinned. The RCU locking is not necessary
2065 * either. It's just for the convenience of using cgroup_css().
2068 css = cgroup_css(cgrp, cft->ss);
2071 if (cft->write_string) {
2072 ret = cft->write_string(css, cft, strstrip(buf));
2073 } else if (cft->write_u64) {
2074 unsigned long long v;
2075 ret = kstrtoull(buf, 0, &v);
2077 ret = cft->write_u64(css, cft, v);
2078 } else if (cft->write_s64) {
2080 ret = kstrtoll(buf, 0, &v);
2082 ret = cft->write_s64(css, cft, v);
2083 } else if (cft->trigger) {
2084 ret = cft->trigger(css, (unsigned int)cft->private);
2089 return ret ?: nbytes;
2092 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2094 return seq_cft(seq)->seq_start(seq, ppos);
2097 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2099 return seq_cft(seq)->seq_next(seq, v, ppos);
2102 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2104 seq_cft(seq)->seq_stop(seq, v);
2107 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2109 struct cftype *cft = seq_cft(m);
2110 struct cgroup_subsys_state *css = seq_css(m);
2113 return cft->seq_show(m, arg);
2116 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2117 else if (cft->read_s64)
2118 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2124 static struct kernfs_ops cgroup_kf_single_ops = {
2125 .atomic_write_len = PAGE_SIZE,
2126 .write = cgroup_file_write,
2127 .seq_show = cgroup_seqfile_show,
2130 static struct kernfs_ops cgroup_kf_ops = {
2131 .atomic_write_len = PAGE_SIZE,
2132 .write = cgroup_file_write,
2133 .seq_start = cgroup_seqfile_start,
2134 .seq_next = cgroup_seqfile_next,
2135 .seq_stop = cgroup_seqfile_stop,
2136 .seq_show = cgroup_seqfile_show,
2140 * cgroup_rename - Only allow simple rename of directories in place.
2142 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2143 const char *new_name_str)
2145 struct cgroup *cgrp = kn->priv;
2148 if (kernfs_type(kn) != KERNFS_DIR)
2150 if (kn->parent != new_parent)
2154 * This isn't a proper migration and its usefulness is very
2155 * limited. Disallow if sane_behavior.
2157 if (cgroup_sane_behavior(cgrp))
2160 mutex_lock(&cgroup_tree_mutex);
2161 mutex_lock(&cgroup_mutex);
2163 ret = kernfs_rename(kn, new_parent, new_name_str);
2165 mutex_unlock(&cgroup_mutex);
2166 mutex_unlock(&cgroup_tree_mutex);
2170 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2172 char name[CGROUP_FILE_NAME_MAX];
2173 struct kernfs_node *kn;
2174 struct lock_class_key *key = NULL;
2176 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2177 key = &cft->lockdep_key;
2179 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2180 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2188 * cgroup_addrm_files - add or remove files to a cgroup directory
2189 * @cgrp: the target cgroup
2190 * @cfts: array of cftypes to be added
2191 * @is_add: whether to add or remove
2193 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2194 * For removals, this function never fails. If addition fails, this
2195 * function doesn't remove files already added. The caller is responsible
2198 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2204 lockdep_assert_held(&cgroup_tree_mutex);
2206 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2207 /* does cft->flags tell us to skip this file on @cgrp? */
2208 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2210 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2212 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2216 ret = cgroup_add_file(cgrp, cft);
2218 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2223 cgroup_rm_file(cgrp, cft);
2229 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2232 struct cgroup_subsys *ss = cfts[0].ss;
2233 struct cgroup *root = &ss->root->top_cgroup;
2234 struct cgroup_subsys_state *css;
2237 lockdep_assert_held(&cgroup_tree_mutex);
2239 /* don't bother if @ss isn't attached */
2240 if (ss->root == &cgroup_dummy_root)
2243 /* add/rm files for all cgroups created before */
2244 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2245 struct cgroup *cgrp = css->cgroup;
2247 if (cgroup_is_dead(cgrp))
2250 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2256 kernfs_activate(root->kn);
2260 static void cgroup_exit_cftypes(struct cftype *cfts)
2264 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2265 /* free copy for custom atomic_write_len, see init_cftypes() */
2266 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2273 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2277 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2278 struct kernfs_ops *kf_ops;
2280 WARN_ON(cft->ss || cft->kf_ops);
2283 kf_ops = &cgroup_kf_ops;
2285 kf_ops = &cgroup_kf_single_ops;
2288 * Ugh... if @cft wants a custom max_write_len, we need to
2289 * make a copy of kf_ops to set its atomic_write_len.
2291 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
2292 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
2294 cgroup_exit_cftypes(cfts);
2297 kf_ops->atomic_write_len = cft->max_write_len;
2300 cft->kf_ops = kf_ops;
2307 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
2309 lockdep_assert_held(&cgroup_tree_mutex);
2311 if (!cfts || !cfts[0].ss)
2314 list_del(&cfts->node);
2315 cgroup_apply_cftypes(cfts, false);
2316 cgroup_exit_cftypes(cfts);
2321 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2322 * @cfts: zero-length name terminated array of cftypes
2324 * Unregister @cfts. Files described by @cfts are removed from all
2325 * existing cgroups and all future cgroups won't have them either. This
2326 * function can be called anytime whether @cfts' subsys is attached or not.
2328 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2331 int cgroup_rm_cftypes(struct cftype *cfts)
2335 mutex_lock(&cgroup_tree_mutex);
2336 ret = cgroup_rm_cftypes_locked(cfts);
2337 mutex_unlock(&cgroup_tree_mutex);
2342 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2343 * @ss: target cgroup subsystem
2344 * @cfts: zero-length name terminated array of cftypes
2346 * Register @cfts to @ss. Files described by @cfts are created for all
2347 * existing cgroups to which @ss is attached and all future cgroups will
2348 * have them too. This function can be called anytime whether @ss is
2351 * Returns 0 on successful registration, -errno on failure. Note that this
2352 * function currently returns 0 as long as @cfts registration is successful
2353 * even if some file creation attempts on existing cgroups fail.
2355 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2359 ret = cgroup_init_cftypes(ss, cfts);
2363 mutex_lock(&cgroup_tree_mutex);
2365 list_add_tail(&cfts->node, &ss->cfts);
2366 ret = cgroup_apply_cftypes(cfts, true);
2368 cgroup_rm_cftypes_locked(cfts);
2370 mutex_unlock(&cgroup_tree_mutex);
2373 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2376 * cgroup_task_count - count the number of tasks in a cgroup.
2377 * @cgrp: the cgroup in question
2379 * Return the number of tasks in the cgroup.
2381 int cgroup_task_count(const struct cgroup *cgrp)
2384 struct cgrp_cset_link *link;
2386 read_lock(&css_set_lock);
2387 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2388 count += atomic_read(&link->cset->refcount);
2389 read_unlock(&css_set_lock);
2394 * To reduce the fork() overhead for systems that are not actually using
2395 * their cgroups capability, we don't maintain the lists running through
2396 * each css_set to its tasks until we see the list actually used - in other
2397 * words after the first call to css_task_iter_start().
2399 static void cgroup_enable_task_cg_lists(void)
2401 struct task_struct *p, *g;
2402 write_lock(&css_set_lock);
2403 use_task_css_set_links = 1;
2405 * We need tasklist_lock because RCU is not safe against
2406 * while_each_thread(). Besides, a forking task that has passed
2407 * cgroup_post_fork() without seeing use_task_css_set_links = 1
2408 * is not guaranteed to have its child immediately visible in the
2409 * tasklist if we walk through it with RCU.
2411 read_lock(&tasklist_lock);
2412 do_each_thread(g, p) {
2415 * We should check if the process is exiting, otherwise
2416 * it will race with cgroup_exit() in that the list
2417 * entry won't be deleted though the process has exited.
2419 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2420 list_add(&p->cg_list, &task_css_set(p)->tasks);
2422 } while_each_thread(g, p);
2423 read_unlock(&tasklist_lock);
2424 write_unlock(&css_set_lock);
2428 * css_next_child - find the next child of a given css
2429 * @pos_css: the current position (%NULL to initiate traversal)
2430 * @parent_css: css whose children to walk
2432 * This function returns the next child of @parent_css and should be called
2433 * under either cgroup_mutex or RCU read lock. The only requirement is
2434 * that @parent_css and @pos_css are accessible. The next sibling is
2435 * guaranteed to be returned regardless of their states.
2437 struct cgroup_subsys_state *
2438 css_next_child(struct cgroup_subsys_state *pos_css,
2439 struct cgroup_subsys_state *parent_css)
2441 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2442 struct cgroup *cgrp = parent_css->cgroup;
2443 struct cgroup *next;
2445 cgroup_assert_mutexes_or_rcu_locked();
2448 * @pos could already have been removed. Once a cgroup is removed,
2449 * its ->sibling.next is no longer updated when its next sibling
2450 * changes. As CGRP_DEAD assertion is serialized and happens
2451 * before the cgroup is taken off the ->sibling list, if we see it
2452 * unasserted, it's guaranteed that the next sibling hasn't
2453 * finished its grace period even if it's already removed, and thus
2454 * safe to dereference from this RCU critical section. If
2455 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2456 * to be visible as %true here.
2458 * If @pos is dead, its next pointer can't be dereferenced;
2459 * however, as each cgroup is given a monotonically increasing
2460 * unique serial number and always appended to the sibling list,
2461 * the next one can be found by walking the parent's children until
2462 * we see a cgroup with higher serial number than @pos's. While
2463 * this path can be slower, it's taken only when either the current
2464 * cgroup is removed or iteration and removal race.
2467 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2468 } else if (likely(!cgroup_is_dead(pos))) {
2469 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2471 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2472 if (next->serial_nr > pos->serial_nr)
2476 if (&next->sibling == &cgrp->children)
2479 return cgroup_css(next, parent_css->ss);
2481 EXPORT_SYMBOL_GPL(css_next_child);
2484 * css_next_descendant_pre - find the next descendant for pre-order walk
2485 * @pos: the current position (%NULL to initiate traversal)
2486 * @root: css whose descendants to walk
2488 * To be used by css_for_each_descendant_pre(). Find the next descendant
2489 * to visit for pre-order traversal of @root's descendants. @root is
2490 * included in the iteration and the first node to be visited.
2492 * While this function requires cgroup_mutex or RCU read locking, it
2493 * doesn't require the whole traversal to be contained in a single critical
2494 * section. This function will return the correct next descendant as long
2495 * as both @pos and @root are accessible and @pos is a descendant of @root.
2497 struct cgroup_subsys_state *
2498 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2499 struct cgroup_subsys_state *root)
2501 struct cgroup_subsys_state *next;
2503 cgroup_assert_mutexes_or_rcu_locked();
2505 /* if first iteration, visit @root */
2509 /* visit the first child if exists */
2510 next = css_next_child(NULL, pos);
2514 /* no child, visit my or the closest ancestor's next sibling */
2515 while (pos != root) {
2516 next = css_next_child(pos, css_parent(pos));
2519 pos = css_parent(pos);
2524 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
2527 * css_rightmost_descendant - return the rightmost descendant of a css
2528 * @pos: css of interest
2530 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2531 * is returned. This can be used during pre-order traversal to skip
2534 * While this function requires cgroup_mutex or RCU read locking, it
2535 * doesn't require the whole traversal to be contained in a single critical
2536 * section. This function will return the correct rightmost descendant as
2537 * long as @pos is accessible.
2539 struct cgroup_subsys_state *
2540 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2542 struct cgroup_subsys_state *last, *tmp;
2544 cgroup_assert_mutexes_or_rcu_locked();
2548 /* ->prev isn't RCU safe, walk ->next till the end */
2550 css_for_each_child(tmp, last)
2556 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
2558 static struct cgroup_subsys_state *
2559 css_leftmost_descendant(struct cgroup_subsys_state *pos)
2561 struct cgroup_subsys_state *last;
2565 pos = css_next_child(NULL, pos);
2572 * css_next_descendant_post - find the next descendant for post-order walk
2573 * @pos: the current position (%NULL to initiate traversal)
2574 * @root: css whose descendants to walk
2576 * To be used by css_for_each_descendant_post(). Find the next descendant
2577 * to visit for post-order traversal of @root's descendants. @root is
2578 * included in the iteration and the last node to be visited.
2580 * While this function requires cgroup_mutex or RCU read locking, it
2581 * doesn't require the whole traversal to be contained in a single critical
2582 * section. This function will return the correct next descendant as long
2583 * as both @pos and @cgroup are accessible and @pos is a descendant of
2586 struct cgroup_subsys_state *
2587 css_next_descendant_post(struct cgroup_subsys_state *pos,
2588 struct cgroup_subsys_state *root)
2590 struct cgroup_subsys_state *next;
2592 cgroup_assert_mutexes_or_rcu_locked();
2594 /* if first iteration, visit leftmost descendant which may be @root */
2596 return css_leftmost_descendant(root);
2598 /* if we visited @root, we're done */
2602 /* if there's an unvisited sibling, visit its leftmost descendant */
2603 next = css_next_child(pos, css_parent(pos));
2605 return css_leftmost_descendant(next);
2607 /* no sibling left, visit parent */
2608 return css_parent(pos);
2610 EXPORT_SYMBOL_GPL(css_next_descendant_post);
2613 * css_advance_task_iter - advance a task itererator to the next css_set
2614 * @it: the iterator to advance
2616 * Advance @it to the next css_set to walk.
2618 static void css_advance_task_iter(struct css_task_iter *it)
2620 struct list_head *l = it->cset_link;
2621 struct cgrp_cset_link *link;
2622 struct css_set *cset;
2624 /* Advance to the next non-empty css_set */
2627 if (l == &it->origin_css->cgroup->cset_links) {
2628 it->cset_link = NULL;
2631 link = list_entry(l, struct cgrp_cset_link, cset_link);
2633 } while (list_empty(&cset->tasks));
2635 it->task = cset->tasks.next;
2639 * css_task_iter_start - initiate task iteration
2640 * @css: the css to walk tasks of
2641 * @it: the task iterator to use
2643 * Initiate iteration through the tasks of @css. The caller can call
2644 * css_task_iter_next() to walk through the tasks until the function
2645 * returns NULL. On completion of iteration, css_task_iter_end() must be
2648 * Note that this function acquires a lock which is released when the
2649 * iteration finishes. The caller can't sleep while iteration is in
2652 void css_task_iter_start(struct cgroup_subsys_state *css,
2653 struct css_task_iter *it)
2654 __acquires(css_set_lock)
2657 * The first time anyone tries to iterate across a css, we need to
2658 * enable the list linking each css_set to its tasks, and fix up
2659 * all existing tasks.
2661 if (!use_task_css_set_links)
2662 cgroup_enable_task_cg_lists();
2664 read_lock(&css_set_lock);
2666 it->origin_css = css;
2667 it->cset_link = &css->cgroup->cset_links;
2669 css_advance_task_iter(it);
2673 * css_task_iter_next - return the next task for the iterator
2674 * @it: the task iterator being iterated
2676 * The "next" function for task iteration. @it should have been
2677 * initialized via css_task_iter_start(). Returns NULL when the iteration
2680 struct task_struct *css_task_iter_next(struct css_task_iter *it)
2682 struct task_struct *res;
2683 struct list_head *l = it->task;
2684 struct cgrp_cset_link *link;
2686 /* If the iterator cg is NULL, we have no tasks */
2689 res = list_entry(l, struct task_struct, cg_list);
2690 /* Advance iterator to find next entry */
2692 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
2693 if (l == &link->cset->tasks) {
2695 * We reached the end of this task list - move on to the
2696 * next cgrp_cset_link.
2698 css_advance_task_iter(it);
2706 * css_task_iter_end - finish task iteration
2707 * @it: the task iterator to finish
2709 * Finish task iteration started by css_task_iter_start().
2711 void css_task_iter_end(struct css_task_iter *it)
2712 __releases(css_set_lock)
2714 read_unlock(&css_set_lock);
2717 static inline int started_after_time(struct task_struct *t1,
2718 struct timespec *time,
2719 struct task_struct *t2)
2721 int start_diff = timespec_compare(&t1->start_time, time);
2722 if (start_diff > 0) {
2724 } else if (start_diff < 0) {
2728 * Arbitrarily, if two processes started at the same
2729 * time, we'll say that the lower pointer value
2730 * started first. Note that t2 may have exited by now
2731 * so this may not be a valid pointer any longer, but
2732 * that's fine - it still serves to distinguish
2733 * between two tasks started (effectively) simultaneously.
2740 * This function is a callback from heap_insert() and is used to order
2742 * In this case we order the heap in descending task start time.
2744 static inline int started_after(void *p1, void *p2)
2746 struct task_struct *t1 = p1;
2747 struct task_struct *t2 = p2;
2748 return started_after_time(t1, &t2->start_time, t2);
2752 * css_scan_tasks - iterate though all the tasks in a css
2753 * @css: the css to iterate tasks of
2754 * @test: optional test callback
2755 * @process: process callback
2756 * @data: data passed to @test and @process
2757 * @heap: optional pre-allocated heap used for task iteration
2759 * Iterate through all the tasks in @css, calling @test for each, and if it
2760 * returns %true, call @process for it also.
2762 * @test may be NULL, meaning always true (select all tasks), which
2763 * effectively duplicates css_task_iter_{start,next,end}() but does not
2764 * lock css_set_lock for the call to @process.
2766 * It is guaranteed that @process will act on every task that is a member
2767 * of @css for the duration of this call. This function may or may not
2768 * call @process for tasks that exit or move to a different css during the
2769 * call, or are forked or move into the css during the call.
2771 * Note that @test may be called with locks held, and may in some
2772 * situations be called multiple times for the same task, so it should be
2775 * If @heap is non-NULL, a heap has been pre-allocated and will be used for
2776 * heap operations (and its "gt" member will be overwritten), else a
2777 * temporary heap will be used (allocation of which may cause this function
2780 int css_scan_tasks(struct cgroup_subsys_state *css,
2781 bool (*test)(struct task_struct *, void *),
2782 void (*process)(struct task_struct *, void *),
2783 void *data, struct ptr_heap *heap)
2786 struct css_task_iter it;
2787 struct task_struct *p, *dropped;
2788 /* Never dereference latest_task, since it's not refcounted */
2789 struct task_struct *latest_task = NULL;
2790 struct ptr_heap tmp_heap;
2791 struct timespec latest_time = { 0, 0 };
2794 /* The caller supplied our heap and pre-allocated its memory */
2795 heap->gt = &started_after;
2797 /* We need to allocate our own heap memory */
2799 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2801 /* cannot allocate the heap */
2807 * Scan tasks in the css, using the @test callback to determine
2808 * which are of interest, and invoking @process callback on the
2809 * ones which need an update. Since we don't want to hold any
2810 * locks during the task updates, gather tasks to be processed in a
2811 * heap structure. The heap is sorted by descending task start
2812 * time. If the statically-sized heap fills up, we overflow tasks
2813 * that started later, and in future iterations only consider tasks
2814 * that started after the latest task in the previous pass. This
2815 * guarantees forward progress and that we don't miss any tasks.
2818 css_task_iter_start(css, &it);
2819 while ((p = css_task_iter_next(&it))) {
2821 * Only affect tasks that qualify per the caller's callback,
2822 * if he provided one
2824 if (test && !test(p, data))
2827 * Only process tasks that started after the last task
2830 if (!started_after_time(p, &latest_time, latest_task))
2832 dropped = heap_insert(heap, p);
2833 if (dropped == NULL) {
2835 * The new task was inserted; the heap wasn't
2839 } else if (dropped != p) {
2841 * The new task was inserted, and pushed out a
2845 put_task_struct(dropped);
2848 * Else the new task was newer than anything already in
2849 * the heap and wasn't inserted
2852 css_task_iter_end(&it);
2855 for (i = 0; i < heap->size; i++) {
2856 struct task_struct *q = heap->ptrs[i];
2858 latest_time = q->start_time;
2861 /* Process the task per the caller's callback */
2866 * If we had to process any tasks at all, scan again
2867 * in case some of them were in the middle of forking
2868 * children that didn't get processed.
2869 * Not the most efficient way to do it, but it avoids
2870 * having to take callback_mutex in the fork path
2874 if (heap == &tmp_heap)
2875 heap_free(&tmp_heap);
2879 static void cgroup_transfer_one_task(struct task_struct *task, void *data)
2881 struct cgroup *new_cgroup = data;
2883 mutex_lock(&cgroup_mutex);
2884 cgroup_attach_task(new_cgroup, task, false);
2885 mutex_unlock(&cgroup_mutex);
2889 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
2890 * @to: cgroup to which the tasks will be moved
2891 * @from: cgroup in which the tasks currently reside
2893 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
2895 return css_scan_tasks(&from->dummy_css, NULL, cgroup_transfer_one_task,
2900 * Stuff for reading the 'tasks'/'procs' files.
2902 * Reading this file can return large amounts of data if a cgroup has
2903 * *lots* of attached tasks. So it may need several calls to read(),
2904 * but we cannot guarantee that the information we produce is correct
2905 * unless we produce it entirely atomically.
2909 /* which pidlist file are we talking about? */
2910 enum cgroup_filetype {
2916 * A pidlist is a list of pids that virtually represents the contents of one
2917 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
2918 * a pair (one each for procs, tasks) for each pid namespace that's relevant
2921 struct cgroup_pidlist {
2923 * used to find which pidlist is wanted. doesn't change as long as
2924 * this particular list stays in the list.
2926 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
2929 /* how many elements the above list has */
2931 /* each of these stored in a list by its cgroup */
2932 struct list_head links;
2933 /* pointer to the cgroup we belong to, for list removal purposes */
2934 struct cgroup *owner;
2935 /* for delayed destruction */
2936 struct delayed_work destroy_dwork;
2940 * The following two functions "fix" the issue where there are more pids
2941 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2942 * TODO: replace with a kernel-wide solution to this problem
2944 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2945 static void *pidlist_allocate(int count)
2947 if (PIDLIST_TOO_LARGE(count))
2948 return vmalloc(count * sizeof(pid_t));
2950 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2953 static void pidlist_free(void *p)
2955 if (is_vmalloc_addr(p))
2962 * Used to destroy all pidlists lingering waiting for destroy timer. None
2963 * should be left afterwards.
2965 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
2967 struct cgroup_pidlist *l, *tmp_l;
2969 mutex_lock(&cgrp->pidlist_mutex);
2970 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
2971 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
2972 mutex_unlock(&cgrp->pidlist_mutex);
2974 flush_workqueue(cgroup_pidlist_destroy_wq);
2975 BUG_ON(!list_empty(&cgrp->pidlists));
2978 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
2980 struct delayed_work *dwork = to_delayed_work(work);
2981 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
2983 struct cgroup_pidlist *tofree = NULL;
2985 mutex_lock(&l->owner->pidlist_mutex);
2988 * Destroy iff we didn't get queued again. The state won't change
2989 * as destroy_dwork can only be queued while locked.
2991 if (!delayed_work_pending(dwork)) {
2992 list_del(&l->links);
2993 pidlist_free(l->list);
2994 put_pid_ns(l->key.ns);
2998 mutex_unlock(&l->owner->pidlist_mutex);
3003 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3004 * Returns the number of unique elements.
3006 static int pidlist_uniq(pid_t *list, int length)
3011 * we presume the 0th element is unique, so i starts at 1. trivial
3012 * edge cases first; no work needs to be done for either
3014 if (length == 0 || length == 1)
3016 /* src and dest walk down the list; dest counts unique elements */
3017 for (src = 1; src < length; src++) {
3018 /* find next unique element */
3019 while (list[src] == list[src-1]) {
3024 /* dest always points to where the next unique element goes */
3025 list[dest] = list[src];
3033 * The two pid files - task and cgroup.procs - guaranteed that the result
3034 * is sorted, which forced this whole pidlist fiasco. As pid order is
3035 * different per namespace, each namespace needs differently sorted list,
3036 * making it impossible to use, for example, single rbtree of member tasks
3037 * sorted by task pointer. As pidlists can be fairly large, allocating one
3038 * per open file is dangerous, so cgroup had to implement shared pool of
3039 * pidlists keyed by cgroup and namespace.
3041 * All this extra complexity was caused by the original implementation
3042 * committing to an entirely unnecessary property. In the long term, we
3043 * want to do away with it. Explicitly scramble sort order if
3044 * sane_behavior so that no such expectation exists in the new interface.
3046 * Scrambling is done by swapping every two consecutive bits, which is
3047 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3049 static pid_t pid_fry(pid_t pid)
3051 unsigned a = pid & 0x55555555;
3052 unsigned b = pid & 0xAAAAAAAA;
3054 return (a << 1) | (b >> 1);
3057 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3059 if (cgroup_sane_behavior(cgrp))
3060 return pid_fry(pid);
3065 static int cmppid(const void *a, const void *b)
3067 return *(pid_t *)a - *(pid_t *)b;
3070 static int fried_cmppid(const void *a, const void *b)
3072 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3075 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3076 enum cgroup_filetype type)
3078 struct cgroup_pidlist *l;
3079 /* don't need task_nsproxy() if we're looking at ourself */
3080 struct pid_namespace *ns = task_active_pid_ns(current);
3082 lockdep_assert_held(&cgrp->pidlist_mutex);
3084 list_for_each_entry(l, &cgrp->pidlists, links)
3085 if (l->key.type == type && l->key.ns == ns)
3091 * find the appropriate pidlist for our purpose (given procs vs tasks)
3092 * returns with the lock on that pidlist already held, and takes care
3093 * of the use count, or returns NULL with no locks held if we're out of
3096 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3097 enum cgroup_filetype type)
3099 struct cgroup_pidlist *l;
3101 lockdep_assert_held(&cgrp->pidlist_mutex);
3103 l = cgroup_pidlist_find(cgrp, type);
3107 /* entry not found; create a new one */
3108 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3112 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3114 /* don't need task_nsproxy() if we're looking at ourself */
3115 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3117 list_add(&l->links, &cgrp->pidlists);
3122 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3124 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3125 struct cgroup_pidlist **lp)
3129 int pid, n = 0; /* used for populating the array */
3130 struct css_task_iter it;
3131 struct task_struct *tsk;
3132 struct cgroup_pidlist *l;
3134 lockdep_assert_held(&cgrp->pidlist_mutex);
3137 * If cgroup gets more users after we read count, we won't have
3138 * enough space - tough. This race is indistinguishable to the
3139 * caller from the case that the additional cgroup users didn't
3140 * show up until sometime later on.
3142 length = cgroup_task_count(cgrp);
3143 array = pidlist_allocate(length);
3146 /* now, populate the array */
3147 css_task_iter_start(&cgrp->dummy_css, &it);
3148 while ((tsk = css_task_iter_next(&it))) {
3149 if (unlikely(n == length))
3151 /* get tgid or pid for procs or tasks file respectively */
3152 if (type == CGROUP_FILE_PROCS)
3153 pid = task_tgid_vnr(tsk);
3155 pid = task_pid_vnr(tsk);
3156 if (pid > 0) /* make sure to only use valid results */
3159 css_task_iter_end(&it);
3161 /* now sort & (if procs) strip out duplicates */
3162 if (cgroup_sane_behavior(cgrp))
3163 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3165 sort(array, length, sizeof(pid_t), cmppid, NULL);
3166 if (type == CGROUP_FILE_PROCS)
3167 length = pidlist_uniq(array, length);
3169 l = cgroup_pidlist_find_create(cgrp, type);
3171 mutex_unlock(&cgrp->pidlist_mutex);
3172 pidlist_free(array);
3176 /* store array, freeing old if necessary */
3177 pidlist_free(l->list);
3185 * cgroupstats_build - build and fill cgroupstats
3186 * @stats: cgroupstats to fill information into
3187 * @dentry: A dentry entry belonging to the cgroup for which stats have
3190 * Build and fill cgroupstats so that taskstats can export it to user
3193 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3195 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3196 struct cgroup *cgrp;
3197 struct css_task_iter it;
3198 struct task_struct *tsk;
3200 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3201 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3202 kernfs_type(kn) != KERNFS_DIR)
3206 * We aren't being called from kernfs and there's no guarantee on
3207 * @kn->priv's validity. For this and css_tryget_from_dir(),
3208 * @kn->priv is RCU safe. Let's do the RCU dancing.
3211 cgrp = rcu_dereference(kn->priv);
3217 css_task_iter_start(&cgrp->dummy_css, &it);
3218 while ((tsk = css_task_iter_next(&it))) {
3219 switch (tsk->state) {
3221 stats->nr_running++;
3223 case TASK_INTERRUPTIBLE:
3224 stats->nr_sleeping++;
3226 case TASK_UNINTERRUPTIBLE:
3227 stats->nr_uninterruptible++;
3230 stats->nr_stopped++;
3233 if (delayacct_is_task_waiting_on_io(tsk))
3234 stats->nr_io_wait++;
3238 css_task_iter_end(&it);
3246 * seq_file methods for the tasks/procs files. The seq_file position is the
3247 * next pid to display; the seq_file iterator is a pointer to the pid
3248 * in the cgroup->l->list array.
3251 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3254 * Initially we receive a position value that corresponds to
3255 * one more than the last pid shown (or 0 on the first call or
3256 * after a seek to the start). Use a binary-search to find the
3257 * next pid to display, if any
3259 struct kernfs_open_file *of = s->private;
3260 struct cgroup *cgrp = seq_css(s)->cgroup;
3261 struct cgroup_pidlist *l;
3262 enum cgroup_filetype type = seq_cft(s)->private;
3263 int index = 0, pid = *pos;
3266 mutex_lock(&cgrp->pidlist_mutex);
3269 * !NULL @of->priv indicates that this isn't the first start()
3270 * after open. If the matching pidlist is around, we can use that.
3271 * Look for it. Note that @of->priv can't be used directly. It
3272 * could already have been destroyed.
3275 of->priv = cgroup_pidlist_find(cgrp, type);
3278 * Either this is the first start() after open or the matching
3279 * pidlist has been destroyed inbetween. Create a new one.
3282 ret = pidlist_array_load(cgrp, type,
3283 (struct cgroup_pidlist **)&of->priv);
3285 return ERR_PTR(ret);
3290 int end = l->length;
3292 while (index < end) {
3293 int mid = (index + end) / 2;
3294 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3297 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3303 /* If we're off the end of the array, we're done */
3304 if (index >= l->length)
3306 /* Update the abstract position to be the actual pid that we found */
3307 iter = l->list + index;
3308 *pos = cgroup_pid_fry(cgrp, *iter);
3312 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3314 struct kernfs_open_file *of = s->private;
3315 struct cgroup_pidlist *l = of->priv;
3318 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3319 CGROUP_PIDLIST_DESTROY_DELAY);
3320 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3323 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3325 struct kernfs_open_file *of = s->private;
3326 struct cgroup_pidlist *l = of->priv;
3328 pid_t *end = l->list + l->length;
3330 * Advance to the next pid in the array. If this goes off the
3337 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3342 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3344 return seq_printf(s, "%d\n", *(int *)v);
3348 * seq_operations functions for iterating on pidlists through seq_file -
3349 * independent of whether it's tasks or procs
3351 static const struct seq_operations cgroup_pidlist_seq_operations = {
3352 .start = cgroup_pidlist_start,
3353 .stop = cgroup_pidlist_stop,
3354 .next = cgroup_pidlist_next,
3355 .show = cgroup_pidlist_show,
3358 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3361 return notify_on_release(css->cgroup);
3364 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3365 struct cftype *cft, u64 val)
3367 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3369 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3371 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3375 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3378 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3381 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3382 struct cftype *cft, u64 val)
3385 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3387 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3391 static struct cftype cgroup_base_files[] = {
3393 .name = "cgroup.procs",
3394 .seq_start = cgroup_pidlist_start,
3395 .seq_next = cgroup_pidlist_next,
3396 .seq_stop = cgroup_pidlist_stop,
3397 .seq_show = cgroup_pidlist_show,
3398 .private = CGROUP_FILE_PROCS,
3399 .write_u64 = cgroup_procs_write,
3400 .mode = S_IRUGO | S_IWUSR,
3403 .name = "cgroup.clone_children",
3404 .flags = CFTYPE_INSANE,
3405 .read_u64 = cgroup_clone_children_read,
3406 .write_u64 = cgroup_clone_children_write,
3409 .name = "cgroup.sane_behavior",
3410 .flags = CFTYPE_ONLY_ON_ROOT,
3411 .seq_show = cgroup_sane_behavior_show,
3415 * Historical crazy stuff. These don't have "cgroup." prefix and
3416 * don't exist if sane_behavior. If you're depending on these, be
3417 * prepared to be burned.
3421 .flags = CFTYPE_INSANE, /* use "procs" instead */
3422 .seq_start = cgroup_pidlist_start,
3423 .seq_next = cgroup_pidlist_next,
3424 .seq_stop = cgroup_pidlist_stop,
3425 .seq_show = cgroup_pidlist_show,
3426 .private = CGROUP_FILE_TASKS,
3427 .write_u64 = cgroup_tasks_write,
3428 .mode = S_IRUGO | S_IWUSR,
3431 .name = "notify_on_release",
3432 .flags = CFTYPE_INSANE,
3433 .read_u64 = cgroup_read_notify_on_release,
3434 .write_u64 = cgroup_write_notify_on_release,
3437 .name = "release_agent",
3438 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3439 .seq_show = cgroup_release_agent_show,
3440 .write_string = cgroup_release_agent_write,
3441 .max_write_len = PATH_MAX - 1,
3447 * cgroup_populate_dir - create subsys files in a cgroup directory
3448 * @cgrp: target cgroup
3449 * @subsys_mask: mask of the subsystem ids whose files should be added
3451 * On failure, no file is added.
3453 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3455 struct cgroup_subsys *ss;
3458 /* process cftsets of each subsystem */
3459 for_each_subsys(ss, i) {
3460 struct cftype *cfts;
3462 if (!test_bit(i, &subsys_mask))
3465 list_for_each_entry(cfts, &ss->cfts, node) {
3466 ret = cgroup_addrm_files(cgrp, cfts, true);
3473 cgroup_clear_dir(cgrp, subsys_mask);
3478 * css destruction is four-stage process.
3480 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3481 * Implemented in kill_css().
3483 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3484 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3485 * by invoking offline_css(). After offlining, the base ref is put.
3486 * Implemented in css_killed_work_fn().
3488 * 3. When the percpu_ref reaches zero, the only possible remaining
3489 * accessors are inside RCU read sections. css_release() schedules the
3492 * 4. After the grace period, the css can be freed. Implemented in
3493 * css_free_work_fn().
3495 * It is actually hairier because both step 2 and 4 require process context
3496 * and thus involve punting to css->destroy_work adding two additional
3497 * steps to the already complex sequence.
3499 static void css_free_work_fn(struct work_struct *work)
3501 struct cgroup_subsys_state *css =
3502 container_of(work, struct cgroup_subsys_state, destroy_work);
3503 struct cgroup *cgrp = css->cgroup;
3506 css_put(css->parent);
3508 css->ss->css_free(css);
3512 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3514 struct cgroup_subsys_state *css =
3515 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3517 INIT_WORK(&css->destroy_work, css_free_work_fn);
3518 queue_work(cgroup_destroy_wq, &css->destroy_work);
3521 static void css_release(struct percpu_ref *ref)
3523 struct cgroup_subsys_state *css =
3524 container_of(ref, struct cgroup_subsys_state, refcnt);
3526 rcu_assign_pointer(css->cgroup->subsys[css->ss->id], NULL);
3527 call_rcu(&css->rcu_head, css_free_rcu_fn);
3530 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3531 struct cgroup *cgrp)
3538 css->parent = cgroup_css(cgrp->parent, ss);
3540 css->flags |= CSS_ROOT;
3542 BUG_ON(cgroup_css(cgrp, ss));
3545 /* invoke ->css_online() on a new CSS and mark it online if successful */
3546 static int online_css(struct cgroup_subsys_state *css)
3548 struct cgroup_subsys *ss = css->ss;
3551 lockdep_assert_held(&cgroup_tree_mutex);
3552 lockdep_assert_held(&cgroup_mutex);
3555 ret = ss->css_online(css);
3557 css->flags |= CSS_ONLINE;
3558 css->cgroup->nr_css++;
3559 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
3564 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
3565 static void offline_css(struct cgroup_subsys_state *css)
3567 struct cgroup_subsys *ss = css->ss;
3569 lockdep_assert_held(&cgroup_tree_mutex);
3570 lockdep_assert_held(&cgroup_mutex);
3572 if (!(css->flags & CSS_ONLINE))
3575 if (ss->css_offline)
3576 ss->css_offline(css);
3578 css->flags &= ~CSS_ONLINE;
3579 css->cgroup->nr_css--;
3580 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
3584 * create_css - create a cgroup_subsys_state
3585 * @cgrp: the cgroup new css will be associated with
3586 * @ss: the subsys of new css
3588 * Create a new css associated with @cgrp - @ss pair. On success, the new
3589 * css is online and installed in @cgrp with all interface files created.
3590 * Returns 0 on success, -errno on failure.
3592 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
3594 struct cgroup *parent = cgrp->parent;
3595 struct cgroup_subsys_state *css;
3598 lockdep_assert_held(&cgroup_mutex);
3600 css = ss->css_alloc(cgroup_css(parent, ss));
3602 return PTR_ERR(css);
3604 err = percpu_ref_init(&css->refcnt, css_release);
3608 init_css(css, ss, cgrp);
3610 err = cgroup_populate_dir(cgrp, 1 << ss->id);
3614 err = online_css(css);
3619 css_get(css->parent);
3621 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
3623 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",
3624 current->comm, current->pid, ss->name);
3625 if (!strcmp(ss->name, "memory"))
3626 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
3627 ss->warned_broken_hierarchy = true;
3633 percpu_ref_cancel_init(&css->refcnt);
3639 * cgroup_create - create a cgroup
3640 * @parent: cgroup that will be parent of the new cgroup
3641 * @name: name of the new cgroup
3642 * @mode: mode to set on new cgroup
3644 static long cgroup_create(struct cgroup *parent, const char *name,
3647 struct cgroup *cgrp;
3648 struct cgroupfs_root *root = parent->root;
3650 struct cgroup_subsys *ss;
3651 struct kernfs_node *kn;
3653 /* allocate the cgroup and its ID, 0 is reserved for the root */
3654 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3658 mutex_lock(&cgroup_tree_mutex);
3661 * Only live parents can have children. Note that the liveliness
3662 * check isn't strictly necessary because cgroup_mkdir() and
3663 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
3664 * anyway so that locking is contained inside cgroup proper and we
3665 * don't get nasty surprises if we ever grow another caller.
3667 if (!cgroup_lock_live_group(parent)) {
3669 goto err_unlock_tree;
3673 * Temporarily set the pointer to NULL, so idr_find() won't return
3674 * a half-baked cgroup.
3676 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
3682 init_cgroup_housekeeping(cgrp);
3684 cgrp->parent = parent;
3685 cgrp->dummy_css.parent = &parent->dummy_css;
3686 cgrp->root = parent->root;
3688 if (notify_on_release(parent))
3689 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3691 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
3692 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
3694 /* create the directory */
3695 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
3703 * This extra ref will be put in cgroup_free_fn() and guarantees
3704 * that @cgrp->kn is always accessible.
3708 cgrp->serial_nr = cgroup_serial_nr_next++;
3710 /* allocation complete, commit to creation */
3711 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
3712 root->number_of_cgroups++;
3715 * Grab a reference on the root and parent so that they don't get
3716 * deleted while there are child cgroups.
3718 cgroup_get_root(root);
3722 * @cgrp is now fully operational. If something fails after this
3723 * point, it'll be released via the normal destruction path.
3725 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
3727 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
3731 /* let's create and online css's */
3732 for_each_subsys(ss, ssid) {
3733 if (root->subsys_mask & (1 << ssid)) {
3734 err = create_css(cgrp, ss);
3740 kernfs_activate(kn);
3742 mutex_unlock(&cgroup_mutex);
3743 mutex_unlock(&cgroup_tree_mutex);
3748 idr_remove(&root->cgroup_idr, cgrp->id);
3750 mutex_unlock(&cgroup_mutex);
3752 mutex_unlock(&cgroup_tree_mutex);
3757 cgroup_destroy_locked(cgrp);
3758 mutex_unlock(&cgroup_mutex);
3759 mutex_unlock(&cgroup_tree_mutex);
3763 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3766 struct cgroup *parent = parent_kn->priv;
3768 return cgroup_create(parent, name, mode);
3772 * This is called when the refcnt of a css is confirmed to be killed.
3773 * css_tryget() is now guaranteed to fail.
3775 static void css_killed_work_fn(struct work_struct *work)
3777 struct cgroup_subsys_state *css =
3778 container_of(work, struct cgroup_subsys_state, destroy_work);
3779 struct cgroup *cgrp = css->cgroup;
3781 mutex_lock(&cgroup_tree_mutex);
3782 mutex_lock(&cgroup_mutex);
3785 * css_tryget() is guaranteed to fail now. Tell subsystems to
3786 * initate destruction.
3791 * If @cgrp is marked dead, it's waiting for refs of all css's to
3792 * be disabled before proceeding to the second phase of cgroup
3793 * destruction. If we are the last one, kick it off.
3795 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
3796 cgroup_destroy_css_killed(cgrp);
3798 mutex_unlock(&cgroup_mutex);
3799 mutex_unlock(&cgroup_tree_mutex);
3802 * Put the css refs from kill_css(). Each css holds an extra
3803 * reference to the cgroup's dentry and cgroup removal proceeds
3804 * regardless of css refs. On the last put of each css, whenever
3805 * that may be, the extra dentry ref is put so that dentry
3806 * destruction happens only after all css's are released.
3811 /* css kill confirmation processing requires process context, bounce */
3812 static void css_killed_ref_fn(struct percpu_ref *ref)
3814 struct cgroup_subsys_state *css =
3815 container_of(ref, struct cgroup_subsys_state, refcnt);
3817 INIT_WORK(&css->destroy_work, css_killed_work_fn);
3818 queue_work(cgroup_destroy_wq, &css->destroy_work);
3822 * kill_css - destroy a css
3823 * @css: css to destroy
3825 * This function initiates destruction of @css by removing cgroup interface
3826 * files and putting its base reference. ->css_offline() will be invoked
3827 * asynchronously once css_tryget() is guaranteed to fail and when the
3828 * reference count reaches zero, @css will be released.
3830 static void kill_css(struct cgroup_subsys_state *css)
3833 * This must happen before css is disassociated with its cgroup.
3834 * See seq_css() for details.
3836 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3839 * Killing would put the base ref, but we need to keep it alive
3840 * until after ->css_offline().
3845 * cgroup core guarantees that, by the time ->css_offline() is
3846 * invoked, no new css reference will be given out via
3847 * css_tryget(). We can't simply call percpu_ref_kill() and
3848 * proceed to offlining css's because percpu_ref_kill() doesn't
3849 * guarantee that the ref is seen as killed on all CPUs on return.
3851 * Use percpu_ref_kill_and_confirm() to get notifications as each
3852 * css is confirmed to be seen as killed on all CPUs.
3854 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
3858 * cgroup_destroy_locked - the first stage of cgroup destruction
3859 * @cgrp: cgroup to be destroyed
3861 * css's make use of percpu refcnts whose killing latency shouldn't be
3862 * exposed to userland and are RCU protected. Also, cgroup core needs to
3863 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
3864 * invoked. To satisfy all the requirements, destruction is implemented in
3865 * the following two steps.
3867 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
3868 * userland visible parts and start killing the percpu refcnts of
3869 * css's. Set up so that the next stage will be kicked off once all
3870 * the percpu refcnts are confirmed to be killed.
3872 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
3873 * rest of destruction. Once all cgroup references are gone, the
3874 * cgroup is RCU-freed.
3876 * This function implements s1. After this step, @cgrp is gone as far as
3877 * the userland is concerned and a new cgroup with the same name may be
3878 * created. As cgroup doesn't care about the names internally, this
3879 * doesn't cause any problem.
3881 static int cgroup_destroy_locked(struct cgroup *cgrp)
3882 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
3884 struct cgroup *child;
3885 struct cgroup_subsys_state *css;
3889 lockdep_assert_held(&cgroup_tree_mutex);
3890 lockdep_assert_held(&cgroup_mutex);
3893 * css_set_lock synchronizes access to ->cset_links and prevents
3894 * @cgrp from being removed while __put_css_set() is in progress.
3896 read_lock(&css_set_lock);
3897 empty = list_empty(&cgrp->cset_links);
3898 read_unlock(&css_set_lock);
3903 * Make sure there's no live children. We can't test ->children
3904 * emptiness as dead children linger on it while being destroyed;
3905 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
3909 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
3910 empty = cgroup_is_dead(child);
3919 * Initiate massacre of all css's. cgroup_destroy_css_killed()
3920 * will be invoked to perform the rest of destruction once the
3921 * percpu refs of all css's are confirmed to be killed. This
3922 * involves removing the subsystem's files, drop cgroup_mutex.
3924 mutex_unlock(&cgroup_mutex);
3925 for_each_css(css, ssid, cgrp)
3927 mutex_lock(&cgroup_mutex);
3930 * Mark @cgrp dead. This prevents further task migration and child
3931 * creation by disabling cgroup_lock_live_group(). Note that
3932 * CGRP_DEAD assertion is depended upon by css_next_child() to
3933 * resume iteration after dropping RCU read lock. See
3934 * css_next_child() for details.
3936 set_bit(CGRP_DEAD, &cgrp->flags);
3938 /* CGRP_DEAD is set, remove from ->release_list for the last time */
3939 raw_spin_lock(&release_list_lock);
3940 if (!list_empty(&cgrp->release_list))
3941 list_del_init(&cgrp->release_list);
3942 raw_spin_unlock(&release_list_lock);
3945 * If @cgrp has css's attached, the second stage of cgroup
3946 * destruction is kicked off from css_killed_work_fn() after the
3947 * refs of all attached css's are killed. If @cgrp doesn't have
3948 * any css, we kick it off here.
3951 cgroup_destroy_css_killed(cgrp);
3953 /* remove @cgrp directory along with the base files */
3954 mutex_unlock(&cgroup_mutex);
3957 * There are two control paths which try to determine cgroup from
3958 * dentry without going through kernfs - cgroupstats_build() and
3959 * css_tryget_from_dir(). Those are supported by RCU protecting
3960 * clearing of cgrp->kn->priv backpointer, which should happen
3961 * after all files under it have been removed.
3963 kernfs_remove(cgrp->kn); /* @cgrp has an extra ref on its kn */
3964 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
3966 mutex_lock(&cgroup_mutex);
3972 * cgroup_destroy_css_killed - the second step of cgroup destruction
3973 * @work: cgroup->destroy_free_work
3975 * This function is invoked from a work item for a cgroup which is being
3976 * destroyed after all css's are offlined and performs the rest of
3977 * destruction. This is the second step of destruction described in the
3978 * comment above cgroup_destroy_locked().
3980 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
3982 struct cgroup *parent = cgrp->parent;
3984 lockdep_assert_held(&cgroup_tree_mutex);
3985 lockdep_assert_held(&cgroup_mutex);
3987 /* delete this cgroup from parent->children */
3988 list_del_rcu(&cgrp->sibling);
3992 set_bit(CGRP_RELEASABLE, &parent->flags);
3993 check_for_release(parent);
3996 static int cgroup_rmdir(struct kernfs_node *kn)
3998 struct cgroup *cgrp = kn->priv;
4002 * This is self-destruction but @kn can't be removed while this
4003 * callback is in progress. Let's break active protection. Once
4004 * the protection is broken, @cgrp can be destroyed at any point.
4005 * Pin it so that it stays accessible.
4008 kernfs_break_active_protection(kn);
4010 mutex_lock(&cgroup_tree_mutex);
4011 mutex_lock(&cgroup_mutex);
4014 * @cgrp might already have been destroyed while we're trying to
4017 if (!cgroup_is_dead(cgrp))
4018 ret = cgroup_destroy_locked(cgrp);
4020 mutex_unlock(&cgroup_mutex);
4021 mutex_unlock(&cgroup_tree_mutex);
4023 kernfs_unbreak_active_protection(kn);
4028 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4029 .remount_fs = cgroup_remount,
4030 .show_options = cgroup_show_options,
4031 .mkdir = cgroup_mkdir,
4032 .rmdir = cgroup_rmdir,
4033 .rename = cgroup_rename,
4036 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4038 struct cgroup_subsys_state *css;
4040 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4042 mutex_lock(&cgroup_tree_mutex);
4043 mutex_lock(&cgroup_mutex);
4045 INIT_LIST_HEAD(&ss->cfts);
4047 /* Create the top cgroup state for this subsystem */
4048 ss->root = &cgroup_dummy_root;
4049 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4050 /* We don't handle early failures gracefully */
4051 BUG_ON(IS_ERR(css));
4052 init_css(css, ss, cgroup_dummy_top);
4054 /* Update the init_css_set to contain a subsys
4055 * pointer to this state - since the subsystem is
4056 * newly registered, all tasks and hence the
4057 * init_css_set is in the subsystem's top cgroup. */
4058 init_css_set.subsys[ss->id] = css;
4060 need_forkexit_callback |= ss->fork || ss->exit;
4062 /* At system boot, before all subsystems have been
4063 * registered, no tasks have been forked, so we don't
4064 * need to invoke fork callbacks here. */
4065 BUG_ON(!list_empty(&init_task.tasks));
4067 BUG_ON(online_css(css));
4069 mutex_unlock(&cgroup_mutex);
4070 mutex_unlock(&cgroup_tree_mutex);
4074 * cgroup_init_early - cgroup initialization at system boot
4076 * Initialize cgroups at system boot, and initialize any
4077 * subsystems that request early init.
4079 int __init cgroup_init_early(void)
4081 struct cgroup_subsys *ss;
4084 atomic_set(&init_css_set.refcount, 1);
4085 INIT_LIST_HEAD(&init_css_set.cgrp_links);
4086 INIT_LIST_HEAD(&init_css_set.tasks);
4087 INIT_HLIST_NODE(&init_css_set.hlist);
4089 init_cgroup_root(&cgroup_dummy_root);
4090 cgroup_root_count = 1;
4091 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4093 init_cgrp_cset_link.cset = &init_css_set;
4094 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4095 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4096 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4098 for_each_subsys(ss, i) {
4099 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4100 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4101 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4103 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4104 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4107 ss->name = cgroup_subsys_name[i];
4110 cgroup_init_subsys(ss);
4116 * cgroup_init - cgroup initialization
4118 * Register cgroup filesystem and /proc file, and initialize
4119 * any subsystems that didn't request early init.
4121 int __init cgroup_init(void)
4123 struct cgroup_subsys *ss;
4127 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4129 for_each_subsys(ss, i) {
4130 if (!ss->early_init)
4131 cgroup_init_subsys(ss);
4134 * cftype registration needs kmalloc and can't be done
4135 * during early_init. Register base cftypes separately.
4137 if (ss->base_cftypes)
4138 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4141 /* allocate id for the dummy hierarchy */
4142 mutex_lock(&cgroup_mutex);
4144 /* Add init_css_set to the hash table */
4145 key = css_set_hash(init_css_set.subsys);
4146 hash_add(css_set_table, &init_css_set.hlist, key);
4148 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4150 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
4154 mutex_unlock(&cgroup_mutex);
4156 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4160 err = register_filesystem(&cgroup_fs_type);
4162 kobject_put(cgroup_kobj);
4166 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4170 static int __init cgroup_wq_init(void)
4173 * There isn't much point in executing destruction path in
4174 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4176 * XXX: Must be ordered to make sure parent is offlined after
4177 * children. The ordering requirement is for memcg where a
4178 * parent's offline may wait for a child's leading to deadlock. In
4179 * the long term, this should be fixed from memcg side.
4181 * We would prefer to do this in cgroup_init() above, but that
4182 * is called before init_workqueues(): so leave this until after.
4184 cgroup_destroy_wq = alloc_ordered_workqueue("cgroup_destroy", 0);
4185 BUG_ON(!cgroup_destroy_wq);
4188 * Used to destroy pidlists and separate to serve as flush domain.
4189 * Cap @max_active to 1 too.
4191 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4193 BUG_ON(!cgroup_pidlist_destroy_wq);
4197 core_initcall(cgroup_wq_init);
4200 * proc_cgroup_show()
4201 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4202 * - Used for /proc/<pid>/cgroup.
4203 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4204 * doesn't really matter if tsk->cgroup changes after we read it,
4205 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4206 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4207 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4208 * cgroup to top_cgroup.
4211 /* TODO: Use a proper seq_file iterator */
4212 int proc_cgroup_show(struct seq_file *m, void *v)
4215 struct task_struct *tsk;
4218 struct cgroupfs_root *root;
4221 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4227 tsk = get_pid_task(pid, PIDTYPE_PID);
4233 mutex_lock(&cgroup_mutex);
4235 for_each_active_root(root) {
4236 struct cgroup_subsys *ss;
4237 struct cgroup *cgrp;
4238 int ssid, count = 0;
4240 seq_printf(m, "%d:", root->hierarchy_id);
4241 for_each_subsys(ss, ssid)
4242 if (root->subsys_mask & (1 << ssid))
4243 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4244 if (strlen(root->name))
4245 seq_printf(m, "%sname=%s", count ? "," : "",
4248 cgrp = task_cgroup_from_root(tsk, root);
4249 path = cgroup_path(cgrp, buf, PATH_MAX);
4251 retval = -ENAMETOOLONG;
4259 mutex_unlock(&cgroup_mutex);
4260 put_task_struct(tsk);
4267 /* Display information about each subsystem and each hierarchy */
4268 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4270 struct cgroup_subsys *ss;
4273 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4275 * ideally we don't want subsystems moving around while we do this.
4276 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4277 * subsys/hierarchy state.
4279 mutex_lock(&cgroup_mutex);
4281 for_each_subsys(ss, i)
4282 seq_printf(m, "%s\t%d\t%d\t%d\n",
4283 ss->name, ss->root->hierarchy_id,
4284 ss->root->number_of_cgroups, !ss->disabled);
4286 mutex_unlock(&cgroup_mutex);
4290 static int cgroupstats_open(struct inode *inode, struct file *file)
4292 return single_open(file, proc_cgroupstats_show, NULL);
4295 static const struct file_operations proc_cgroupstats_operations = {
4296 .open = cgroupstats_open,
4298 .llseek = seq_lseek,
4299 .release = single_release,
4303 * cgroup_fork - attach newly forked task to its parents cgroup.
4304 * @child: pointer to task_struct of forking parent process.
4306 * Description: A task inherits its parent's cgroup at fork().
4308 * A pointer to the shared css_set was automatically copied in
4309 * fork.c by dup_task_struct(). However, we ignore that copy, since
4310 * it was not made under the protection of RCU or cgroup_mutex, so
4311 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4312 * have already changed current->cgroups, allowing the previously
4313 * referenced cgroup group to be removed and freed.
4315 * At the point that cgroup_fork() is called, 'current' is the parent
4316 * task, and the passed argument 'child' points to the child task.
4318 void cgroup_fork(struct task_struct *child)
4321 get_css_set(task_css_set(current));
4322 child->cgroups = current->cgroups;
4323 task_unlock(current);
4324 INIT_LIST_HEAD(&child->cg_list);
4328 * cgroup_post_fork - called on a new task after adding it to the task list
4329 * @child: the task in question
4331 * Adds the task to the list running through its css_set if necessary and
4332 * call the subsystem fork() callbacks. Has to be after the task is
4333 * visible on the task list in case we race with the first call to
4334 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4337 void cgroup_post_fork(struct task_struct *child)
4339 struct cgroup_subsys *ss;
4343 * use_task_css_set_links is set to 1 before we walk the tasklist
4344 * under the tasklist_lock and we read it here after we added the child
4345 * to the tasklist under the tasklist_lock as well. If the child wasn't
4346 * yet in the tasklist when we walked through it from
4347 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4348 * should be visible now due to the paired locking and barriers implied
4349 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4350 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4353 if (use_task_css_set_links) {
4354 write_lock(&css_set_lock);
4356 if (list_empty(&child->cg_list))
4357 list_add(&child->cg_list, &task_css_set(child)->tasks);
4359 write_unlock(&css_set_lock);
4363 * Call ss->fork(). This must happen after @child is linked on
4364 * css_set; otherwise, @child might change state between ->fork()
4365 * and addition to css_set.
4367 if (need_forkexit_callback) {
4368 for_each_subsys(ss, i)
4375 * cgroup_exit - detach cgroup from exiting task
4376 * @tsk: pointer to task_struct of exiting process
4377 * @run_callback: run exit callbacks?
4379 * Description: Detach cgroup from @tsk and release it.
4381 * Note that cgroups marked notify_on_release force every task in
4382 * them to take the global cgroup_mutex mutex when exiting.
4383 * This could impact scaling on very large systems. Be reluctant to
4384 * use notify_on_release cgroups where very high task exit scaling
4385 * is required on large systems.
4387 * the_top_cgroup_hack:
4389 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4391 * We call cgroup_exit() while the task is still competent to
4392 * handle notify_on_release(), then leave the task attached to the
4393 * root cgroup in each hierarchy for the remainder of its exit.
4395 * To do this properly, we would increment the reference count on
4396 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4397 * code we would add a second cgroup function call, to drop that
4398 * reference. This would just create an unnecessary hot spot on
4399 * the top_cgroup reference count, to no avail.
4401 * Normally, holding a reference to a cgroup without bumping its
4402 * count is unsafe. The cgroup could go away, or someone could
4403 * attach us to a different cgroup, decrementing the count on
4404 * the first cgroup that we never incremented. But in this case,
4405 * top_cgroup isn't going away, and either task has PF_EXITING set,
4406 * which wards off any cgroup_attach_task() attempts, or task is a failed
4407 * fork, never visible to cgroup_attach_task.
4409 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4411 struct cgroup_subsys *ss;
4412 struct css_set *cset;
4416 * Unlink from the css_set task list if necessary.
4417 * Optimistically check cg_list before taking
4420 if (!list_empty(&tsk->cg_list)) {
4421 write_lock(&css_set_lock);
4422 if (!list_empty(&tsk->cg_list))
4423 list_del_init(&tsk->cg_list);
4424 write_unlock(&css_set_lock);
4427 /* Reassign the task to the init_css_set. */
4429 cset = task_css_set(tsk);
4430 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4432 if (run_callbacks && need_forkexit_callback) {
4433 /* see cgroup_post_fork() for details */
4434 for_each_subsys(ss, i) {
4436 struct cgroup_subsys_state *old_css = cset->subsys[i];
4437 struct cgroup_subsys_state *css = task_css(tsk, i);
4439 ss->exit(css, old_css, tsk);
4445 put_css_set_taskexit(cset);
4448 static void check_for_release(struct cgroup *cgrp)
4450 if (cgroup_is_releasable(cgrp) &&
4451 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4453 * Control Group is currently removeable. If it's not
4454 * already queued for a userspace notification, queue
4457 int need_schedule_work = 0;
4459 raw_spin_lock(&release_list_lock);
4460 if (!cgroup_is_dead(cgrp) &&
4461 list_empty(&cgrp->release_list)) {
4462 list_add(&cgrp->release_list, &release_list);
4463 need_schedule_work = 1;
4465 raw_spin_unlock(&release_list_lock);
4466 if (need_schedule_work)
4467 schedule_work(&release_agent_work);
4472 * Notify userspace when a cgroup is released, by running the
4473 * configured release agent with the name of the cgroup (path
4474 * relative to the root of cgroup file system) as the argument.
4476 * Most likely, this user command will try to rmdir this cgroup.
4478 * This races with the possibility that some other task will be
4479 * attached to this cgroup before it is removed, or that some other
4480 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4481 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4482 * unused, and this cgroup will be reprieved from its death sentence,
4483 * to continue to serve a useful existence. Next time it's released,
4484 * we will get notified again, if it still has 'notify_on_release' set.
4486 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4487 * means only wait until the task is successfully execve()'d. The
4488 * separate release agent task is forked by call_usermodehelper(),
4489 * then control in this thread returns here, without waiting for the
4490 * release agent task. We don't bother to wait because the caller of
4491 * this routine has no use for the exit status of the release agent
4492 * task, so no sense holding our caller up for that.
4494 static void cgroup_release_agent(struct work_struct *work)
4496 BUG_ON(work != &release_agent_work);
4497 mutex_lock(&cgroup_mutex);
4498 raw_spin_lock(&release_list_lock);
4499 while (!list_empty(&release_list)) {
4500 char *argv[3], *envp[3];
4502 char *pathbuf = NULL, *agentbuf = NULL, *path;
4503 struct cgroup *cgrp = list_entry(release_list.next,
4506 list_del_init(&cgrp->release_list);
4507 raw_spin_unlock(&release_list_lock);
4508 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
4511 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
4514 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4519 argv[i++] = agentbuf;
4524 /* minimal command environment */
4525 envp[i++] = "HOME=/";
4526 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4529 /* Drop the lock while we invoke the usermode helper,
4530 * since the exec could involve hitting disk and hence
4531 * be a slow process */
4532 mutex_unlock(&cgroup_mutex);
4533 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4534 mutex_lock(&cgroup_mutex);
4538 raw_spin_lock(&release_list_lock);
4540 raw_spin_unlock(&release_list_lock);
4541 mutex_unlock(&cgroup_mutex);
4544 static int __init cgroup_disable(char *str)
4546 struct cgroup_subsys *ss;
4550 while ((token = strsep(&str, ",")) != NULL) {
4554 for_each_subsys(ss, i) {
4555 if (!strcmp(token, ss->name)) {
4557 printk(KERN_INFO "Disabling %s control group"
4558 " subsystem\n", ss->name);
4565 __setup("cgroup_disable=", cgroup_disable);
4568 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
4569 * @dentry: directory dentry of interest
4570 * @ss: subsystem of interest
4572 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
4573 * to get the corresponding css and return it. If such css doesn't exist
4574 * or can't be pinned, an ERR_PTR value is returned.
4576 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
4577 struct cgroup_subsys *ss)
4579 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4580 struct cgroup_subsys_state *css = NULL;
4581 struct cgroup *cgrp;
4583 /* is @dentry a cgroup dir? */
4584 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4585 kernfs_type(kn) != KERNFS_DIR)
4586 return ERR_PTR(-EBADF);
4591 * This path doesn't originate from kernfs and @kn could already
4592 * have been or be removed at any point. @kn->priv is RCU
4593 * protected for this access. See destroy_locked() for details.
4595 cgrp = rcu_dereference(kn->priv);
4597 css = cgroup_css(cgrp, ss);
4599 if (!css || !css_tryget(css))
4600 css = ERR_PTR(-ENOENT);
4607 * css_from_id - lookup css by id
4608 * @id: the cgroup id
4609 * @ss: cgroup subsys to be looked into
4611 * Returns the css if there's valid one with @id, otherwise returns NULL.
4612 * Should be called under rcu_read_lock().
4614 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
4616 struct cgroup *cgrp;
4618 cgroup_assert_mutexes_or_rcu_locked();
4620 cgrp = idr_find(&ss->root->cgroup_idr, id);
4622 return cgroup_css(cgrp, ss);
4626 #ifdef CONFIG_CGROUP_DEBUG
4627 static struct cgroup_subsys_state *
4628 debug_css_alloc(struct cgroup_subsys_state *parent_css)
4630 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4633 return ERR_PTR(-ENOMEM);
4638 static void debug_css_free(struct cgroup_subsys_state *css)
4643 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
4646 return cgroup_task_count(css->cgroup);
4649 static u64 current_css_set_read(struct cgroup_subsys_state *css,
4652 return (u64)(unsigned long)current->cgroups;
4655 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
4661 count = atomic_read(&task_css_set(current)->refcount);
4666 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
4668 struct cgrp_cset_link *link;
4669 struct css_set *cset;
4672 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
4676 read_lock(&css_set_lock);
4678 cset = rcu_dereference(current->cgroups);
4679 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
4680 struct cgroup *c = link->cgrp;
4681 const char *name = "?";
4683 if (c != cgroup_dummy_top) {
4684 cgroup_name(c, name_buf, NAME_MAX + 1);
4688 seq_printf(seq, "Root %d group %s\n",
4689 c->root->hierarchy_id, name);
4692 read_unlock(&css_set_lock);
4697 #define MAX_TASKS_SHOWN_PER_CSS 25
4698 static int cgroup_css_links_read(struct seq_file *seq, void *v)
4700 struct cgroup_subsys_state *css = seq_css(seq);
4701 struct cgrp_cset_link *link;
4703 read_lock(&css_set_lock);
4704 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
4705 struct css_set *cset = link->cset;
4706 struct task_struct *task;
4708 seq_printf(seq, "css_set %p\n", cset);
4709 list_for_each_entry(task, &cset->tasks, cg_list) {
4710 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4711 seq_puts(seq, " ...\n");
4714 seq_printf(seq, " task %d\n",
4715 task_pid_vnr(task));
4719 read_unlock(&css_set_lock);
4723 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
4725 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4728 static struct cftype debug_files[] = {
4730 .name = "taskcount",
4731 .read_u64 = debug_taskcount_read,
4735 .name = "current_css_set",
4736 .read_u64 = current_css_set_read,
4740 .name = "current_css_set_refcount",
4741 .read_u64 = current_css_set_refcount_read,
4745 .name = "current_css_set_cg_links",
4746 .seq_show = current_css_set_cg_links_read,
4750 .name = "cgroup_css_links",
4751 .seq_show = cgroup_css_links_read,
4755 .name = "releasable",
4756 .read_u64 = releasable_read,
4762 struct cgroup_subsys debug_cgrp_subsys = {
4763 .css_alloc = debug_css_alloc,
4764 .css_free = debug_css_free,
4765 .base_cftypes = debug_files,
4767 #endif /* CONFIG_CGROUP_DEBUG */