2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #include <linux/cgroup.h>
30 #include <linux/cred.h>
31 #include <linux/ctype.h>
32 #include <linux/errno.h>
33 #include <linux/init_task.h>
34 #include <linux/kernel.h>
35 #include <linux/list.h>
37 #include <linux/mutex.h>
38 #include <linux/mount.h>
39 #include <linux/pagemap.h>
40 #include <linux/proc_fs.h>
41 #include <linux/rcupdate.h>
42 #include <linux/sched.h>
43 #include <linux/slab.h>
44 #include <linux/spinlock.h>
45 #include <linux/string.h>
46 #include <linux/sort.h>
47 #include <linux/kmod.h>
48 #include <linux/delayacct.h>
49 #include <linux/cgroupstats.h>
50 #include <linux/hashtable.h>
51 #include <linux/pid_namespace.h>
52 #include <linux/idr.h>
53 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
54 #include <linux/flex_array.h> /* used in cgroup_attach_task */
55 #include <linux/kthread.h>
56 #include <linux/delay.h>
58 #include <linux/atomic.h>
61 * pidlists linger the following amount before being destroyed. The goal
62 * is avoiding frequent destruction in the middle of consecutive read calls
63 * Expiring in the middle is a performance problem not a correctness one.
64 * 1 sec should be enough.
66 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
68 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
72 * cgroup_tree_mutex nests above cgroup_mutex and protects cftypes, file
73 * creation/removal and hierarchy changing operations including cgroup
74 * creation, removal, css association and controller rebinding. This outer
75 * lock is needed mainly to resolve the circular dependency between kernfs
76 * active ref and cgroup_mutex. cgroup_tree_mutex nests above both.
78 static DEFINE_MUTEX(cgroup_tree_mutex);
81 * cgroup_mutex is the master lock. Any modification to cgroup or its
82 * hierarchy must be performed while holding it.
84 #ifdef CONFIG_PROVE_RCU
85 DEFINE_MUTEX(cgroup_mutex);
86 EXPORT_SYMBOL_GPL(cgroup_mutex); /* only for lockdep */
88 static DEFINE_MUTEX(cgroup_mutex);
92 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
93 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
95 static DEFINE_SPINLOCK(release_agent_path_lock);
97 #define cgroup_assert_mutexes_or_rcu_locked() \
98 rcu_lockdep_assert(rcu_read_lock_held() || \
99 lockdep_is_held(&cgroup_tree_mutex) || \
100 lockdep_is_held(&cgroup_mutex), \
101 "cgroup_[tree_]mutex or RCU read lock required");
104 * cgroup destruction makes heavy use of work items and there can be a lot
105 * of concurrent destructions. Use a separate workqueue so that cgroup
106 * destruction work items don't end up filling up max_active of system_wq
107 * which may lead to deadlock.
109 static struct workqueue_struct *cgroup_destroy_wq;
112 * pidlist destructions need to be flushed on cgroup destruction. Use a
113 * separate workqueue as flush domain.
115 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
117 /* generate an array of cgroup subsystem pointers */
118 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
119 static struct cgroup_subsys *cgroup_subsys[] = {
120 #include <linux/cgroup_subsys.h>
124 /* array of cgroup subsystem names */
125 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
126 static const char *cgroup_subsys_name[] = {
127 #include <linux/cgroup_subsys.h>
132 * The dummy hierarchy, reserved for the subsystems that are otherwise
133 * unattached - it never has more than a single cgroup, and all tasks are
134 * part of that cgroup.
136 static struct cgroupfs_root cgroup_dummy_root;
138 /* dummy_top is a shorthand for the dummy hierarchy's top cgroup */
139 static struct cgroup * const cgroup_dummy_top = &cgroup_dummy_root.top_cgroup;
141 /* The list of hierarchy roots */
143 static LIST_HEAD(cgroup_roots);
144 static int cgroup_root_count;
146 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
147 static DEFINE_IDR(cgroup_hierarchy_idr);
150 * Assign a monotonically increasing serial number to cgroups. It
151 * guarantees cgroups with bigger numbers are newer than those with smaller
152 * numbers. Also, as cgroups are always appended to the parent's
153 * ->children list, it guarantees that sibling cgroups are always sorted in
154 * the ascending serial number order on the list. Protected by
157 static u64 cgroup_serial_nr_next = 1;
159 /* This flag indicates whether tasks in the fork and exit paths should
160 * check for fork/exit handlers to call. This avoids us having to do
161 * extra work in the fork/exit path if none of the subsystems need to
164 static int need_forkexit_callback __read_mostly;
166 static struct cftype cgroup_base_files[];
168 static void cgroup_put(struct cgroup *cgrp);
169 static int rebind_subsystems(struct cgroupfs_root *root,
170 unsigned long added_mask, unsigned removed_mask);
171 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
172 static int cgroup_destroy_locked(struct cgroup *cgrp);
173 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
175 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
178 * cgroup_css - obtain a cgroup's css for the specified subsystem
179 * @cgrp: the cgroup of interest
180 * @ss: the subsystem of interest (%NULL returns the dummy_css)
182 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
183 * function must be called either under cgroup_mutex or rcu_read_lock() and
184 * the caller is responsible for pinning the returned css if it wants to
185 * keep accessing it outside the said locks. This function may return
186 * %NULL if @cgrp doesn't have @subsys_id enabled.
188 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
189 struct cgroup_subsys *ss)
192 return rcu_dereference_check(cgrp->subsys[ss->id],
193 lockdep_is_held(&cgroup_tree_mutex) ||
194 lockdep_is_held(&cgroup_mutex));
196 return &cgrp->dummy_css;
199 /* convenient tests for these bits */
200 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
202 return test_bit(CGRP_DEAD, &cgrp->flags);
205 struct cgroup_subsys_state *seq_css(struct seq_file *seq)
207 struct kernfs_open_file *of = seq->private;
208 struct cgroup *cgrp = of->kn->parent->priv;
209 struct cftype *cft = seq_cft(seq);
212 * This is open and unprotected implementation of cgroup_css().
213 * seq_css() is only called from a kernfs file operation which has
214 * an active reference on the file. Because all the subsystem
215 * files are drained before a css is disassociated with a cgroup,
216 * the matching css from the cgroup's subsys table is guaranteed to
217 * be and stay valid until the enclosing operation is complete.
220 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
222 return &cgrp->dummy_css;
224 EXPORT_SYMBOL_GPL(seq_css);
227 * cgroup_is_descendant - test ancestry
228 * @cgrp: the cgroup to be tested
229 * @ancestor: possible ancestor of @cgrp
231 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
232 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
233 * and @ancestor are accessible.
235 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
238 if (cgrp == ancestor)
244 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
246 static int cgroup_is_releasable(const struct cgroup *cgrp)
249 (1 << CGRP_RELEASABLE) |
250 (1 << CGRP_NOTIFY_ON_RELEASE);
251 return (cgrp->flags & bits) == bits;
254 static int notify_on_release(const struct cgroup *cgrp)
256 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
260 * for_each_css - iterate all css's of a cgroup
261 * @css: the iteration cursor
262 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
263 * @cgrp: the target cgroup to iterate css's of
265 * Should be called under cgroup_mutex.
267 #define for_each_css(css, ssid, cgrp) \
268 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
269 if (!((css) = rcu_dereference_check( \
270 (cgrp)->subsys[(ssid)], \
271 lockdep_is_held(&cgroup_tree_mutex) || \
272 lockdep_is_held(&cgroup_mutex)))) { } \
276 * for_each_subsys - iterate all enabled cgroup subsystems
277 * @ss: the iteration cursor
278 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
280 #define for_each_subsys(ss, ssid) \
281 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
282 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
284 /* iterate across the active hierarchies */
285 #define for_each_active_root(root) \
286 list_for_each_entry((root), &cgroup_roots, root_list)
289 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
290 * @cgrp: the cgroup to be checked for liveness
292 * On success, returns true; the mutex should be later unlocked. On
293 * failure returns false with no lock held.
295 static bool cgroup_lock_live_group(struct cgroup *cgrp)
297 mutex_lock(&cgroup_mutex);
298 if (cgroup_is_dead(cgrp)) {
299 mutex_unlock(&cgroup_mutex);
305 /* the list of cgroups eligible for automatic release. Protected by
306 * release_list_lock */
307 static LIST_HEAD(release_list);
308 static DEFINE_RAW_SPINLOCK(release_list_lock);
309 static void cgroup_release_agent(struct work_struct *work);
310 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
311 static void check_for_release(struct cgroup *cgrp);
314 * A cgroup can be associated with multiple css_sets as different tasks may
315 * belong to different cgroups on different hierarchies. In the other
316 * direction, a css_set is naturally associated with multiple cgroups.
317 * This M:N relationship is represented by the following link structure
318 * which exists for each association and allows traversing the associations
321 struct cgrp_cset_link {
322 /* the cgroup and css_set this link associates */
324 struct css_set *cset;
326 /* list of cgrp_cset_links anchored at cgrp->cset_links */
327 struct list_head cset_link;
329 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
330 struct list_head cgrp_link;
333 /* The default css_set - used by init and its children prior to any
334 * hierarchies being mounted. It contains a pointer to the root state
335 * for each subsystem. Also used to anchor the list of css_sets. Not
336 * reference-counted, to improve performance when child cgroups
337 * haven't been created.
340 static struct css_set init_css_set;
341 static struct cgrp_cset_link init_cgrp_cset_link;
344 * css_set_lock protects the list of css_set objects, and the chain of
345 * tasks off each css_set. Nests outside task->alloc_lock due to
346 * css_task_iter_start().
348 static DEFINE_RWLOCK(css_set_lock);
349 static int css_set_count;
352 * hash table for cgroup groups. This improves the performance to find
353 * an existing css_set. This hash doesn't (currently) take into
354 * account cgroups in empty hierarchies.
356 #define CSS_SET_HASH_BITS 7
357 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
359 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
361 unsigned long key = 0UL;
362 struct cgroup_subsys *ss;
365 for_each_subsys(ss, i)
366 key += (unsigned long)css[i];
367 key = (key >> 16) ^ key;
373 * We don't maintain the lists running through each css_set to its task
374 * until after the first call to css_task_iter_start(). This reduces the
375 * fork()/exit() overhead for people who have cgroups compiled into their
376 * kernel but not actually in use.
378 static int use_task_css_set_links __read_mostly;
380 static void __put_css_set(struct css_set *cset, int taskexit)
382 struct cgrp_cset_link *link, *tmp_link;
385 * Ensure that the refcount doesn't hit zero while any readers
386 * can see it. Similar to atomic_dec_and_lock(), but for an
389 if (atomic_add_unless(&cset->refcount, -1, 1))
391 write_lock(&css_set_lock);
392 if (!atomic_dec_and_test(&cset->refcount)) {
393 write_unlock(&css_set_lock);
397 /* This css_set is dead. unlink it and release cgroup refcounts */
398 hash_del(&cset->hlist);
401 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
402 struct cgroup *cgrp = link->cgrp;
404 list_del(&link->cset_link);
405 list_del(&link->cgrp_link);
407 /* @cgrp can't go away while we're holding css_set_lock */
408 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
410 set_bit(CGRP_RELEASABLE, &cgrp->flags);
411 check_for_release(cgrp);
417 write_unlock(&css_set_lock);
418 kfree_rcu(cset, rcu_head);
422 * refcounted get/put for css_set objects
424 static inline void get_css_set(struct css_set *cset)
426 atomic_inc(&cset->refcount);
429 static inline void put_css_set(struct css_set *cset)
431 __put_css_set(cset, 0);
434 static inline void put_css_set_taskexit(struct css_set *cset)
436 __put_css_set(cset, 1);
440 * compare_css_sets - helper function for find_existing_css_set().
441 * @cset: candidate css_set being tested
442 * @old_cset: existing css_set for a task
443 * @new_cgrp: cgroup that's being entered by the task
444 * @template: desired set of css pointers in css_set (pre-calculated)
446 * Returns true if "cset" matches "old_cset" except for the hierarchy
447 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
449 static bool compare_css_sets(struct css_set *cset,
450 struct css_set *old_cset,
451 struct cgroup *new_cgrp,
452 struct cgroup_subsys_state *template[])
454 struct list_head *l1, *l2;
456 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
457 /* Not all subsystems matched */
462 * Compare cgroup pointers in order to distinguish between
463 * different cgroups in heirarchies with no subsystems. We
464 * could get by with just this check alone (and skip the
465 * memcmp above) but on most setups the memcmp check will
466 * avoid the need for this more expensive check on almost all
470 l1 = &cset->cgrp_links;
471 l2 = &old_cset->cgrp_links;
473 struct cgrp_cset_link *link1, *link2;
474 struct cgroup *cgrp1, *cgrp2;
478 /* See if we reached the end - both lists are equal length. */
479 if (l1 == &cset->cgrp_links) {
480 BUG_ON(l2 != &old_cset->cgrp_links);
483 BUG_ON(l2 == &old_cset->cgrp_links);
485 /* Locate the cgroups associated with these links. */
486 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
487 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
490 /* Hierarchies should be linked in the same order. */
491 BUG_ON(cgrp1->root != cgrp2->root);
494 * If this hierarchy is the hierarchy of the cgroup
495 * that's changing, then we need to check that this
496 * css_set points to the new cgroup; if it's any other
497 * hierarchy, then this css_set should point to the
498 * same cgroup as the old css_set.
500 if (cgrp1->root == new_cgrp->root) {
501 if (cgrp1 != new_cgrp)
512 * find_existing_css_set - init css array and find the matching css_set
513 * @old_cset: the css_set that we're using before the cgroup transition
514 * @cgrp: the cgroup that we're moving into
515 * @template: out param for the new set of csses, should be clear on entry
517 static struct css_set *find_existing_css_set(struct css_set *old_cset,
519 struct cgroup_subsys_state *template[])
521 struct cgroupfs_root *root = cgrp->root;
522 struct cgroup_subsys *ss;
523 struct css_set *cset;
528 * Build the set of subsystem state objects that we want to see in the
529 * new css_set. while subsystems can change globally, the entries here
530 * won't change, so no need for locking.
532 for_each_subsys(ss, i) {
533 if (root->subsys_mask & (1UL << i)) {
534 /* Subsystem is in this hierarchy. So we want
535 * the subsystem state from the new
537 template[i] = cgroup_css(cgrp, ss);
539 /* Subsystem is not in this hierarchy, so we
540 * don't want to change the subsystem state */
541 template[i] = old_cset->subsys[i];
545 key = css_set_hash(template);
546 hash_for_each_possible(css_set_table, cset, hlist, key) {
547 if (!compare_css_sets(cset, old_cset, cgrp, template))
550 /* This css_set matches what we need */
554 /* No existing cgroup group matched */
558 static void free_cgrp_cset_links(struct list_head *links_to_free)
560 struct cgrp_cset_link *link, *tmp_link;
562 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
563 list_del(&link->cset_link);
569 * allocate_cgrp_cset_links - allocate cgrp_cset_links
570 * @count: the number of links to allocate
571 * @tmp_links: list_head the allocated links are put on
573 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
574 * through ->cset_link. Returns 0 on success or -errno.
576 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
578 struct cgrp_cset_link *link;
581 INIT_LIST_HEAD(tmp_links);
583 for (i = 0; i < count; i++) {
584 link = kzalloc(sizeof(*link), GFP_KERNEL);
586 free_cgrp_cset_links(tmp_links);
589 list_add(&link->cset_link, tmp_links);
595 * link_css_set - a helper function to link a css_set to a cgroup
596 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
597 * @cset: the css_set to be linked
598 * @cgrp: the destination cgroup
600 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
603 struct cgrp_cset_link *link;
605 BUG_ON(list_empty(tmp_links));
606 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
609 list_move(&link->cset_link, &cgrp->cset_links);
611 * Always add links to the tail of the list so that the list
612 * is sorted by order of hierarchy creation
614 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
618 * find_css_set - return a new css_set with one cgroup updated
619 * @old_cset: the baseline css_set
620 * @cgrp: the cgroup to be updated
622 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
623 * substituted into the appropriate hierarchy.
625 static struct css_set *find_css_set(struct css_set *old_cset,
628 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
629 struct css_set *cset;
630 struct list_head tmp_links;
631 struct cgrp_cset_link *link;
634 lockdep_assert_held(&cgroup_mutex);
636 /* First see if we already have a cgroup group that matches
638 read_lock(&css_set_lock);
639 cset = find_existing_css_set(old_cset, cgrp, template);
642 read_unlock(&css_set_lock);
647 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
651 /* Allocate all the cgrp_cset_link objects that we'll need */
652 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
657 atomic_set(&cset->refcount, 1);
658 INIT_LIST_HEAD(&cset->cgrp_links);
659 INIT_LIST_HEAD(&cset->tasks);
660 INIT_HLIST_NODE(&cset->hlist);
662 /* Copy the set of subsystem state objects generated in
663 * find_existing_css_set() */
664 memcpy(cset->subsys, template, sizeof(cset->subsys));
666 write_lock(&css_set_lock);
667 /* Add reference counts and links from the new css_set. */
668 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
669 struct cgroup *c = link->cgrp;
671 if (c->root == cgrp->root)
673 link_css_set(&tmp_links, cset, c);
676 BUG_ON(!list_empty(&tmp_links));
680 /* Add this cgroup group to the hash table */
681 key = css_set_hash(cset->subsys);
682 hash_add(css_set_table, &cset->hlist, key);
684 write_unlock(&css_set_lock);
689 static struct cgroupfs_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
691 struct cgroup *top_cgrp = kf_root->kn->priv;
693 return top_cgrp->root;
696 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
700 lockdep_assert_held(&cgroup_mutex);
702 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
707 root->hierarchy_id = id;
711 static void cgroup_exit_root_id(struct cgroupfs_root *root)
713 lockdep_assert_held(&cgroup_mutex);
715 if (root->hierarchy_id) {
716 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
717 root->hierarchy_id = 0;
721 static void cgroup_free_root(struct cgroupfs_root *root)
724 /* hierarhcy ID shoulid already have been released */
725 WARN_ON_ONCE(root->hierarchy_id);
727 idr_destroy(&root->cgroup_idr);
732 static void cgroup_destroy_root(struct cgroupfs_root *root)
734 struct cgroup *cgrp = &root->top_cgroup;
735 struct cgrp_cset_link *link, *tmp_link;
737 mutex_lock(&cgroup_tree_mutex);
738 mutex_lock(&cgroup_mutex);
740 BUG_ON(atomic_read(&root->nr_cgrps));
741 BUG_ON(!list_empty(&cgrp->children));
743 /* Rebind all subsystems back to the default hierarchy */
744 WARN_ON(rebind_subsystems(root, 0, root->subsys_mask));
747 * Release all the links from cset_links to this hierarchy's
750 write_lock(&css_set_lock);
752 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
753 list_del(&link->cset_link);
754 list_del(&link->cgrp_link);
757 write_unlock(&css_set_lock);
759 if (!list_empty(&root->root_list)) {
760 list_del(&root->root_list);
764 cgroup_exit_root_id(root);
766 mutex_unlock(&cgroup_mutex);
767 mutex_unlock(&cgroup_tree_mutex);
769 kernfs_destroy_root(root->kf_root);
770 cgroup_free_root(root);
774 * Return the cgroup for "task" from the given hierarchy. Must be
775 * called with cgroup_mutex held.
777 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
778 struct cgroupfs_root *root)
780 struct css_set *cset;
781 struct cgroup *res = NULL;
783 BUG_ON(!mutex_is_locked(&cgroup_mutex));
784 read_lock(&css_set_lock);
786 * No need to lock the task - since we hold cgroup_mutex the
787 * task can't change groups, so the only thing that can happen
788 * is that it exits and its css is set back to init_css_set.
790 cset = task_css_set(task);
791 if (cset == &init_css_set) {
792 res = &root->top_cgroup;
794 struct cgrp_cset_link *link;
796 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
797 struct cgroup *c = link->cgrp;
799 if (c->root == root) {
805 read_unlock(&css_set_lock);
811 * There is one global cgroup mutex. We also require taking
812 * task_lock() when dereferencing a task's cgroup subsys pointers.
813 * See "The task_lock() exception", at the end of this comment.
815 * A task must hold cgroup_mutex to modify cgroups.
817 * Any task can increment and decrement the count field without lock.
818 * So in general, code holding cgroup_mutex can't rely on the count
819 * field not changing. However, if the count goes to zero, then only
820 * cgroup_attach_task() can increment it again. Because a count of zero
821 * means that no tasks are currently attached, therefore there is no
822 * way a task attached to that cgroup can fork (the other way to
823 * increment the count). So code holding cgroup_mutex can safely
824 * assume that if the count is zero, it will stay zero. Similarly, if
825 * a task holds cgroup_mutex on a cgroup with zero count, it
826 * knows that the cgroup won't be removed, as cgroup_rmdir()
829 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
830 * (usually) take cgroup_mutex. These are the two most performance
831 * critical pieces of code here. The exception occurs on cgroup_exit(),
832 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
833 * is taken, and if the cgroup count is zero, a usermode call made
834 * to the release agent with the name of the cgroup (path relative to
835 * the root of cgroup file system) as the argument.
837 * A cgroup can only be deleted if both its 'count' of using tasks
838 * is zero, and its list of 'children' cgroups is empty. Since all
839 * tasks in the system use _some_ cgroup, and since there is always at
840 * least one task in the system (init, pid == 1), therefore, top_cgroup
841 * always has either children cgroups and/or using tasks. So we don't
842 * need a special hack to ensure that top_cgroup cannot be deleted.
844 * The task_lock() exception
846 * The need for this exception arises from the action of
847 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
848 * another. It does so using cgroup_mutex, however there are
849 * several performance critical places that need to reference
850 * task->cgroup without the expense of grabbing a system global
851 * mutex. Therefore except as noted below, when dereferencing or, as
852 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
853 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
854 * the task_struct routinely used for such matters.
856 * P.S. One more locking exception. RCU is used to guard the
857 * update of a tasks cgroup pointer by cgroup_attach_task()
860 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
861 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
862 static const struct file_operations proc_cgroupstats_operations;
864 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
867 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
868 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
869 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
870 cft->ss->name, cft->name);
872 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
877 * cgroup_file_mode - deduce file mode of a control file
878 * @cft: the control file in question
880 * returns cft->mode if ->mode is not 0
881 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
882 * returns S_IRUGO if it has only a read handler
883 * returns S_IWUSR if it has only a write hander
885 static umode_t cgroup_file_mode(const struct cftype *cft)
892 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
895 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
902 static void cgroup_free_fn(struct work_struct *work)
904 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
906 atomic_dec(&cgrp->root->nr_cgrps);
907 cgroup_pidlist_destroy_all(cgrp);
911 * We get a ref to the parent, and put the ref when this
912 * cgroup is being freed, so it's guaranteed that the
913 * parent won't be destroyed before its children.
915 cgroup_put(cgrp->parent);
916 kernfs_put(cgrp->kn);
920 * This is top cgroup's refcnt reaching zero, which
921 * indicates that the root should be released.
923 cgroup_destroy_root(cgrp->root);
927 static void cgroup_free_rcu(struct rcu_head *head)
929 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
931 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
932 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
935 static void cgroup_get(struct cgroup *cgrp)
937 WARN_ON_ONCE(cgroup_is_dead(cgrp));
938 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
939 atomic_inc(&cgrp->refcnt);
942 static void cgroup_put(struct cgroup *cgrp)
944 if (!atomic_dec_and_test(&cgrp->refcnt))
946 if (WARN_ON_ONCE(cgrp->parent && !cgroup_is_dead(cgrp)))
950 * XXX: cgrp->id is only used to look up css's. As cgroup and
951 * css's lifetimes will be decoupled, it should be made
952 * per-subsystem and moved to css->id so that lookups are
953 * successful until the target css is released.
955 mutex_lock(&cgroup_mutex);
956 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
957 mutex_unlock(&cgroup_mutex);
960 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
963 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
965 char name[CGROUP_FILE_NAME_MAX];
967 lockdep_assert_held(&cgroup_tree_mutex);
968 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
972 * cgroup_clear_dir - remove subsys files in a cgroup directory
973 * @cgrp: target cgroup
974 * @subsys_mask: mask of the subsystem ids whose files should be removed
976 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
978 struct cgroup_subsys *ss;
981 for_each_subsys(ss, i) {
984 if (!test_bit(i, &subsys_mask))
986 list_for_each_entry(cfts, &ss->cfts, node)
987 cgroup_addrm_files(cgrp, cfts, false);
991 static int rebind_subsystems(struct cgroupfs_root *root,
992 unsigned long added_mask, unsigned removed_mask)
994 struct cgroup *cgrp = &root->top_cgroup;
995 struct cgroup_subsys *ss;
998 lockdep_assert_held(&cgroup_tree_mutex);
999 lockdep_assert_held(&cgroup_mutex);
1001 /* Check that any added subsystems are currently free */
1002 for_each_subsys(ss, i)
1003 if ((added_mask & (1 << i)) && ss->root != &cgroup_dummy_root)
1006 ret = cgroup_populate_dir(cgrp, added_mask);
1011 * Nothing can fail from this point on. Remove files for the
1012 * removed subsystems and rebind each subsystem.
1014 mutex_unlock(&cgroup_mutex);
1015 cgroup_clear_dir(cgrp, removed_mask);
1016 mutex_lock(&cgroup_mutex);
1018 for_each_subsys(ss, i) {
1019 unsigned long bit = 1UL << i;
1021 if (bit & added_mask) {
1022 /* We're binding this subsystem to this hierarchy */
1023 BUG_ON(cgroup_css(cgrp, ss));
1024 BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
1025 BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
1027 rcu_assign_pointer(cgrp->subsys[i],
1028 cgroup_css(cgroup_dummy_top, ss));
1029 cgroup_css(cgrp, ss)->cgroup = cgrp;
1033 ss->bind(cgroup_css(cgrp, ss));
1035 /* refcount was already taken, and we're keeping it */
1036 root->subsys_mask |= bit;
1037 } else if (bit & removed_mask) {
1038 /* We're removing this subsystem */
1039 BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1040 BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1043 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1045 cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1046 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1048 cgroup_subsys[i]->root = &cgroup_dummy_root;
1049 root->subsys_mask &= ~bit;
1053 kernfs_activate(cgrp->kn);
1057 static int cgroup_show_options(struct seq_file *seq,
1058 struct kernfs_root *kf_root)
1060 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1061 struct cgroup_subsys *ss;
1064 for_each_subsys(ss, ssid)
1065 if (root->subsys_mask & (1 << ssid))
1066 seq_printf(seq, ",%s", ss->name);
1067 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1068 seq_puts(seq, ",sane_behavior");
1069 if (root->flags & CGRP_ROOT_NOPREFIX)
1070 seq_puts(seq, ",noprefix");
1071 if (root->flags & CGRP_ROOT_XATTR)
1072 seq_puts(seq, ",xattr");
1074 spin_lock(&release_agent_path_lock);
1075 if (strlen(root->release_agent_path))
1076 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1077 spin_unlock(&release_agent_path_lock);
1079 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1080 seq_puts(seq, ",clone_children");
1081 if (strlen(root->name))
1082 seq_printf(seq, ",name=%s", root->name);
1086 struct cgroup_sb_opts {
1087 unsigned long subsys_mask;
1088 unsigned long flags;
1089 char *release_agent;
1090 bool cpuset_clone_children;
1092 /* User explicitly requested empty subsystem */
1097 * Convert a hierarchy specifier into a bitmask of subsystems and
1098 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1099 * array. This function takes refcounts on subsystems to be used, unless it
1100 * returns error, in which case no refcounts are taken.
1102 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1104 char *token, *o = data;
1105 bool all_ss = false, one_ss = false;
1106 unsigned long mask = (unsigned long)-1;
1107 struct cgroup_subsys *ss;
1110 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1112 #ifdef CONFIG_CPUSETS
1113 mask = ~(1UL << cpuset_cgrp_id);
1116 memset(opts, 0, sizeof(*opts));
1118 while ((token = strsep(&o, ",")) != NULL) {
1121 if (!strcmp(token, "none")) {
1122 /* Explicitly have no subsystems */
1126 if (!strcmp(token, "all")) {
1127 /* Mutually exclusive option 'all' + subsystem name */
1133 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1134 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1137 if (!strcmp(token, "noprefix")) {
1138 opts->flags |= CGRP_ROOT_NOPREFIX;
1141 if (!strcmp(token, "clone_children")) {
1142 opts->cpuset_clone_children = true;
1145 if (!strcmp(token, "xattr")) {
1146 opts->flags |= CGRP_ROOT_XATTR;
1149 if (!strncmp(token, "release_agent=", 14)) {
1150 /* Specifying two release agents is forbidden */
1151 if (opts->release_agent)
1153 opts->release_agent =
1154 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1155 if (!opts->release_agent)
1159 if (!strncmp(token, "name=", 5)) {
1160 const char *name = token + 5;
1161 /* Can't specify an empty name */
1164 /* Must match [\w.-]+ */
1165 for (i = 0; i < strlen(name); i++) {
1169 if ((c == '.') || (c == '-') || (c == '_'))
1173 /* Specifying two names is forbidden */
1176 opts->name = kstrndup(name,
1177 MAX_CGROUP_ROOT_NAMELEN - 1,
1185 for_each_subsys(ss, i) {
1186 if (strcmp(token, ss->name))
1191 /* Mutually exclusive option 'all' + subsystem name */
1194 set_bit(i, &opts->subsys_mask);
1199 if (i == CGROUP_SUBSYS_COUNT)
1204 * If the 'all' option was specified select all the subsystems,
1205 * otherwise if 'none', 'name=' and a subsystem name options
1206 * were not specified, let's default to 'all'
1208 if (all_ss || (!one_ss && !opts->none && !opts->name))
1209 for_each_subsys(ss, i)
1211 set_bit(i, &opts->subsys_mask);
1213 /* Consistency checks */
1215 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1216 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1218 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1219 opts->cpuset_clone_children || opts->release_agent ||
1221 pr_err("cgroup: sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1227 * Option noprefix was introduced just for backward compatibility
1228 * with the old cpuset, so we allow noprefix only if mounting just
1229 * the cpuset subsystem.
1231 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1235 /* Can't specify "none" and some subsystems */
1236 if (opts->subsys_mask && opts->none)
1240 * We either have to specify by name or by subsystems. (So all
1241 * empty hierarchies must have a name).
1243 if (!opts->subsys_mask && !opts->name)
1249 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1252 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1253 struct cgroup_sb_opts opts;
1254 unsigned long added_mask, removed_mask;
1256 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1257 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1261 mutex_lock(&cgroup_tree_mutex);
1262 mutex_lock(&cgroup_mutex);
1264 /* See what subsystems are wanted */
1265 ret = parse_cgroupfs_options(data, &opts);
1269 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1270 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1271 task_tgid_nr(current), current->comm);
1273 added_mask = opts.subsys_mask & ~root->subsys_mask;
1274 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1276 /* Don't allow flags or name to change at remount */
1277 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1278 (opts.name && strcmp(opts.name, root->name))) {
1279 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1280 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1281 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1286 /* remounting is not allowed for populated hierarchies */
1287 if (!list_empty(&root->top_cgroup.children)) {
1292 ret = rebind_subsystems(root, added_mask, removed_mask);
1296 if (opts.release_agent) {
1297 spin_lock(&release_agent_path_lock);
1298 strcpy(root->release_agent_path, opts.release_agent);
1299 spin_unlock(&release_agent_path_lock);
1302 kfree(opts.release_agent);
1304 mutex_unlock(&cgroup_mutex);
1305 mutex_unlock(&cgroup_tree_mutex);
1309 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1311 atomic_set(&cgrp->refcnt, 1);
1312 INIT_LIST_HEAD(&cgrp->sibling);
1313 INIT_LIST_HEAD(&cgrp->children);
1314 INIT_LIST_HEAD(&cgrp->cset_links);
1315 INIT_LIST_HEAD(&cgrp->release_list);
1316 INIT_LIST_HEAD(&cgrp->pidlists);
1317 mutex_init(&cgrp->pidlist_mutex);
1318 cgrp->dummy_css.cgroup = cgrp;
1321 static void init_cgroup_root(struct cgroupfs_root *root)
1323 struct cgroup *cgrp = &root->top_cgroup;
1325 INIT_LIST_HEAD(&root->root_list);
1326 atomic_set(&root->nr_cgrps, 1);
1328 init_cgroup_housekeeping(cgrp);
1329 idr_init(&root->cgroup_idr);
1332 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1334 struct cgroupfs_root *root;
1336 if (!opts->subsys_mask && !opts->none)
1337 return ERR_PTR(-EINVAL);
1339 root = kzalloc(sizeof(*root), GFP_KERNEL);
1341 return ERR_PTR(-ENOMEM);
1343 init_cgroup_root(root);
1345 root->flags = opts->flags;
1346 if (opts->release_agent)
1347 strcpy(root->release_agent_path, opts->release_agent);
1349 strcpy(root->name, opts->name);
1350 if (opts->cpuset_clone_children)
1351 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1355 static int cgroup_setup_root(struct cgroupfs_root *root, unsigned long ss_mask)
1357 LIST_HEAD(tmp_links);
1358 struct cgroup *root_cgrp = &root->top_cgroup;
1359 struct css_set *cset;
1362 lockdep_assert_held(&cgroup_tree_mutex);
1363 lockdep_assert_held(&cgroup_mutex);
1365 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1368 root_cgrp->id = ret;
1371 * We're accessing css_set_count without locking css_set_lock here,
1372 * but that's OK - it can only be increased by someone holding
1373 * cgroup_lock, and that's us. The worst that can happen is that we
1374 * have some link structures left over
1376 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1380 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1381 ret = cgroup_init_root_id(root, 2, 0);
1385 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1386 KERNFS_ROOT_CREATE_DEACTIVATED,
1388 if (IS_ERR(root->kf_root)) {
1389 ret = PTR_ERR(root->kf_root);
1392 root_cgrp->kn = root->kf_root->kn;
1394 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1398 ret = rebind_subsystems(root, ss_mask, 0);
1403 * There must be no failure case after here, since rebinding takes
1404 * care of subsystems' refcounts, which are explicitly dropped in
1405 * the failure exit path.
1407 list_add(&root->root_list, &cgroup_roots);
1408 cgroup_root_count++;
1411 * Link the top cgroup in this hierarchy into all the css_set
1414 write_lock(&css_set_lock);
1415 hash_for_each(css_set_table, i, cset, hlist)
1416 link_css_set(&tmp_links, cset, root_cgrp);
1417 write_unlock(&css_set_lock);
1419 BUG_ON(!list_empty(&root_cgrp->children));
1420 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1422 kernfs_activate(root_cgrp->kn);
1427 kernfs_destroy_root(root->kf_root);
1428 root->kf_root = NULL;
1430 cgroup_exit_root_id(root);
1432 free_cgrp_cset_links(&tmp_links);
1436 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1437 int flags, const char *unused_dev_name,
1440 struct cgroupfs_root *root;
1441 struct cgroup_sb_opts opts;
1442 struct dentry *dentry;
1445 mutex_lock(&cgroup_tree_mutex);
1446 mutex_lock(&cgroup_mutex);
1448 /* First find the desired set of subsystems */
1449 ret = parse_cgroupfs_options(data, &opts);
1453 /* look for a matching existing root */
1454 for_each_active_root(root) {
1455 bool name_match = false;
1458 * If we asked for a name then it must match. Also, if
1459 * name matches but sybsys_mask doesn't, we should fail.
1460 * Remember whether name matched.
1463 if (strcmp(opts.name, root->name))
1469 * If we asked for subsystems (or explicitly for no
1470 * subsystems) then they must match.
1472 if ((opts.subsys_mask || opts.none) &&
1473 (opts.subsys_mask != root->subsys_mask)) {
1480 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1481 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1482 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1486 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1491 * A root's lifetime is governed by its top cgroup. Zero
1492 * ref indicate that the root is being destroyed. Wait for
1493 * destruction to complete so that the subsystems are free.
1494 * We can use wait_queue for the wait but this path is
1495 * super cold. Let's just sleep for a bit and retry.
1497 if (!atomic_inc_not_zero(&root->top_cgroup.refcnt)) {
1498 mutex_unlock(&cgroup_mutex);
1499 mutex_unlock(&cgroup_tree_mutex);
1508 /* no such thing, create a new one */
1509 root = cgroup_root_from_opts(&opts);
1511 ret = PTR_ERR(root);
1515 ret = cgroup_setup_root(root, opts.subsys_mask);
1517 cgroup_free_root(root);
1520 mutex_unlock(&cgroup_mutex);
1521 mutex_unlock(&cgroup_tree_mutex);
1523 kfree(opts.release_agent);
1527 return ERR_PTR(ret);
1529 dentry = kernfs_mount(fs_type, flags, root->kf_root);
1531 cgroup_put(&root->top_cgroup);
1535 static void cgroup_kill_sb(struct super_block *sb)
1537 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1538 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1540 cgroup_put(&root->top_cgroup);
1544 static struct file_system_type cgroup_fs_type = {
1546 .mount = cgroup_mount,
1547 .kill_sb = cgroup_kill_sb,
1550 static struct kobject *cgroup_kobj;
1553 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1554 * @task: target task
1555 * @buf: the buffer to write the path into
1556 * @buflen: the length of the buffer
1558 * Determine @task's cgroup on the first (the one with the lowest non-zero
1559 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1560 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1561 * cgroup controller callbacks.
1563 * Return value is the same as kernfs_path().
1565 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1567 struct cgroupfs_root *root;
1568 struct cgroup *cgrp;
1569 int hierarchy_id = 1;
1572 mutex_lock(&cgroup_mutex);
1574 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1577 cgrp = task_cgroup_from_root(task, root);
1578 path = cgroup_path(cgrp, buf, buflen);
1580 /* if no hierarchy exists, everyone is in "/" */
1581 if (strlcpy(buf, "/", buflen) < buflen)
1585 mutex_unlock(&cgroup_mutex);
1588 EXPORT_SYMBOL_GPL(task_cgroup_path);
1591 * Control Group taskset
1593 struct task_and_cgroup {
1594 struct task_struct *task;
1595 struct cgroup *cgrp;
1596 struct css_set *cset;
1599 struct cgroup_taskset {
1600 struct task_and_cgroup single;
1601 struct flex_array *tc_array;
1604 struct cgroup *cur_cgrp;
1608 * cgroup_taskset_first - reset taskset and return the first task
1609 * @tset: taskset of interest
1611 * @tset iteration is initialized and the first task is returned.
1613 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1615 if (tset->tc_array) {
1617 return cgroup_taskset_next(tset);
1619 tset->cur_cgrp = tset->single.cgrp;
1620 return tset->single.task;
1623 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1626 * cgroup_taskset_next - iterate to the next task in taskset
1627 * @tset: taskset of interest
1629 * Return the next task in @tset. Iteration must have been initialized
1630 * with cgroup_taskset_first().
1632 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1634 struct task_and_cgroup *tc;
1636 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1639 tc = flex_array_get(tset->tc_array, tset->idx++);
1640 tset->cur_cgrp = tc->cgrp;
1643 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1646 * cgroup_taskset_cur_css - return the matching css for the current task
1647 * @tset: taskset of interest
1648 * @subsys_id: the ID of the target subsystem
1650 * Return the css for the current (last returned) task of @tset for
1651 * subsystem specified by @subsys_id. This function must be preceded by
1652 * either cgroup_taskset_first() or cgroup_taskset_next().
1654 struct cgroup_subsys_state *cgroup_taskset_cur_css(struct cgroup_taskset *tset,
1657 return cgroup_css(tset->cur_cgrp, cgroup_subsys[subsys_id]);
1659 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_css);
1662 * cgroup_taskset_size - return the number of tasks in taskset
1663 * @tset: taskset of interest
1665 int cgroup_taskset_size(struct cgroup_taskset *tset)
1667 return tset->tc_array ? tset->tc_array_len : 1;
1669 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1673 * cgroup_task_migrate - move a task from one cgroup to another.
1675 * Must be called with cgroup_mutex and threadgroup locked.
1677 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1678 struct task_struct *tsk,
1679 struct css_set *new_cset)
1681 struct css_set *old_cset;
1684 * We are synchronized through threadgroup_lock() against PF_EXITING
1685 * setting such that we can't race against cgroup_exit() changing the
1686 * css_set to init_css_set and dropping the old one.
1688 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1689 old_cset = task_css_set(tsk);
1692 rcu_assign_pointer(tsk->cgroups, new_cset);
1695 /* Update the css_set linked lists if we're using them */
1696 write_lock(&css_set_lock);
1697 if (!list_empty(&tsk->cg_list))
1698 list_move(&tsk->cg_list, &new_cset->tasks);
1699 write_unlock(&css_set_lock);
1702 * We just gained a reference on old_cset by taking it from the
1703 * task. As trading it for new_cset is protected by cgroup_mutex,
1704 * we're safe to drop it here; it will be freed under RCU.
1706 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1707 put_css_set(old_cset);
1711 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1712 * @cgrp: the cgroup to attach to
1713 * @tsk: the task or the leader of the threadgroup to be attached
1714 * @threadgroup: attach the whole threadgroup?
1716 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1717 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1719 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1722 int retval, i, group_size;
1723 struct cgroupfs_root *root = cgrp->root;
1724 struct cgroup_subsys_state *css, *failed_css = NULL;
1725 /* threadgroup list cursor and array */
1726 struct task_struct *leader = tsk;
1727 struct task_and_cgroup *tc;
1728 struct flex_array *group;
1729 struct cgroup_taskset tset = { };
1732 * step 0: in order to do expensive, possibly blocking operations for
1733 * every thread, we cannot iterate the thread group list, since it needs
1734 * rcu or tasklist locked. instead, build an array of all threads in the
1735 * group - group_rwsem prevents new threads from appearing, and if
1736 * threads exit, this will just be an over-estimate.
1739 group_size = get_nr_threads(tsk);
1742 /* flex_array supports very large thread-groups better than kmalloc. */
1743 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1746 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1747 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1749 goto out_free_group_list;
1753 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1754 * already PF_EXITING could be freed from underneath us unless we
1755 * take an rcu_read_lock.
1759 struct task_and_cgroup ent;
1761 /* @tsk either already exited or can't exit until the end */
1762 if (tsk->flags & PF_EXITING)
1765 /* as per above, nr_threads may decrease, but not increase. */
1766 BUG_ON(i >= group_size);
1768 ent.cgrp = task_cgroup_from_root(tsk, root);
1769 /* nothing to do if this task is already in the cgroup */
1770 if (ent.cgrp == cgrp)
1773 * saying GFP_ATOMIC has no effect here because we did prealloc
1774 * earlier, but it's good form to communicate our expectations.
1776 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
1777 BUG_ON(retval != 0);
1782 } while_each_thread(leader, tsk);
1784 /* remember the number of threads in the array for later. */
1786 tset.tc_array = group;
1787 tset.tc_array_len = group_size;
1789 /* methods shouldn't be called if no task is actually migrating */
1792 goto out_free_group_list;
1795 * step 1: check that we can legitimately attach to the cgroup.
1797 for_each_css(css, i, cgrp) {
1798 if (css->ss->can_attach) {
1799 retval = css->ss->can_attach(css, &tset);
1802 goto out_cancel_attach;
1808 * step 2: make sure css_sets exist for all threads to be migrated.
1809 * we use find_css_set, which allocates a new one if necessary.
1811 for (i = 0; i < group_size; i++) {
1812 struct css_set *old_cset;
1814 tc = flex_array_get(group, i);
1815 old_cset = task_css_set(tc->task);
1816 tc->cset = find_css_set(old_cset, cgrp);
1819 goto out_put_css_set_refs;
1824 * step 3: now that we're guaranteed success wrt the css_sets,
1825 * proceed to move all tasks to the new cgroup. There are no
1826 * failure cases after here, so this is the commit point.
1828 for (i = 0; i < group_size; i++) {
1829 tc = flex_array_get(group, i);
1830 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
1832 /* nothing is sensitive to fork() after this point. */
1835 * step 4: do subsystem attach callbacks.
1837 for_each_css(css, i, cgrp)
1838 if (css->ss->attach)
1839 css->ss->attach(css, &tset);
1842 * step 5: success! and cleanup
1845 out_put_css_set_refs:
1847 for (i = 0; i < group_size; i++) {
1848 tc = flex_array_get(group, i);
1851 put_css_set(tc->cset);
1856 for_each_css(css, i, cgrp) {
1857 if (css == failed_css)
1859 if (css->ss->cancel_attach)
1860 css->ss->cancel_attach(css, &tset);
1863 out_free_group_list:
1864 flex_array_free(group);
1869 * Find the task_struct of the task to attach by vpid and pass it along to the
1870 * function to attach either it or all tasks in its threadgroup. Will lock
1871 * cgroup_mutex and threadgroup; may take task_lock of task.
1873 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
1875 struct task_struct *tsk;
1876 const struct cred *cred = current_cred(), *tcred;
1879 if (!cgroup_lock_live_group(cgrp))
1885 tsk = find_task_by_vpid(pid);
1889 goto out_unlock_cgroup;
1892 * even if we're attaching all tasks in the thread group, we
1893 * only need to check permissions on one of them.
1895 tcred = __task_cred(tsk);
1896 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
1897 !uid_eq(cred->euid, tcred->uid) &&
1898 !uid_eq(cred->euid, tcred->suid)) {
1901 goto out_unlock_cgroup;
1907 tsk = tsk->group_leader;
1910 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
1911 * trapped in a cpuset, or RT worker may be born in a cgroup
1912 * with no rt_runtime allocated. Just say no.
1914 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
1917 goto out_unlock_cgroup;
1920 get_task_struct(tsk);
1923 threadgroup_lock(tsk);
1925 if (!thread_group_leader(tsk)) {
1927 * a race with de_thread from another thread's exec()
1928 * may strip us of our leadership, if this happens,
1929 * there is no choice but to throw this task away and
1930 * try again; this is
1931 * "double-double-toil-and-trouble-check locking".
1933 threadgroup_unlock(tsk);
1934 put_task_struct(tsk);
1935 goto retry_find_task;
1939 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
1941 threadgroup_unlock(tsk);
1943 put_task_struct(tsk);
1945 mutex_unlock(&cgroup_mutex);
1950 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1951 * @from: attach to all cgroups of a given task
1952 * @tsk: the task to be attached
1954 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1956 struct cgroupfs_root *root;
1959 mutex_lock(&cgroup_mutex);
1960 for_each_active_root(root) {
1961 struct cgroup *from_cgrp = task_cgroup_from_root(from, root);
1963 retval = cgroup_attach_task(from_cgrp, tsk, false);
1967 mutex_unlock(&cgroup_mutex);
1971 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
1973 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
1974 struct cftype *cft, u64 pid)
1976 return attach_task_by_pid(css->cgroup, pid, false);
1979 static int cgroup_procs_write(struct cgroup_subsys_state *css,
1980 struct cftype *cft, u64 tgid)
1982 return attach_task_by_pid(css->cgroup, tgid, true);
1985 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
1986 struct cftype *cft, const char *buffer)
1988 struct cgroupfs_root *root = css->cgroup->root;
1990 BUILD_BUG_ON(sizeof(root->release_agent_path) < PATH_MAX);
1991 if (!cgroup_lock_live_group(css->cgroup))
1993 spin_lock(&release_agent_path_lock);
1994 strlcpy(root->release_agent_path, buffer,
1995 sizeof(root->release_agent_path));
1996 spin_unlock(&release_agent_path_lock);
1997 mutex_unlock(&cgroup_mutex);
2001 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2003 struct cgroup *cgrp = seq_css(seq)->cgroup;
2005 if (!cgroup_lock_live_group(cgrp))
2007 seq_puts(seq, cgrp->root->release_agent_path);
2008 seq_putc(seq, '\n');
2009 mutex_unlock(&cgroup_mutex);
2013 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2015 struct cgroup *cgrp = seq_css(seq)->cgroup;
2017 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2021 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2022 size_t nbytes, loff_t off)
2024 struct cgroup *cgrp = of->kn->parent->priv;
2025 struct cftype *cft = of->kn->priv;
2026 struct cgroup_subsys_state *css;
2030 * kernfs guarantees that a file isn't deleted with operations in
2031 * flight, which means that the matching css is and stays alive and
2032 * doesn't need to be pinned. The RCU locking is not necessary
2033 * either. It's just for the convenience of using cgroup_css().
2036 css = cgroup_css(cgrp, cft->ss);
2039 if (cft->write_string) {
2040 ret = cft->write_string(css, cft, strstrip(buf));
2041 } else if (cft->write_u64) {
2042 unsigned long long v;
2043 ret = kstrtoull(buf, 0, &v);
2045 ret = cft->write_u64(css, cft, v);
2046 } else if (cft->write_s64) {
2048 ret = kstrtoll(buf, 0, &v);
2050 ret = cft->write_s64(css, cft, v);
2051 } else if (cft->trigger) {
2052 ret = cft->trigger(css, (unsigned int)cft->private);
2057 return ret ?: nbytes;
2060 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2062 return seq_cft(seq)->seq_start(seq, ppos);
2065 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2067 return seq_cft(seq)->seq_next(seq, v, ppos);
2070 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2072 seq_cft(seq)->seq_stop(seq, v);
2075 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2077 struct cftype *cft = seq_cft(m);
2078 struct cgroup_subsys_state *css = seq_css(m);
2081 return cft->seq_show(m, arg);
2084 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2085 else if (cft->read_s64)
2086 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2092 static struct kernfs_ops cgroup_kf_single_ops = {
2093 .atomic_write_len = PAGE_SIZE,
2094 .write = cgroup_file_write,
2095 .seq_show = cgroup_seqfile_show,
2098 static struct kernfs_ops cgroup_kf_ops = {
2099 .atomic_write_len = PAGE_SIZE,
2100 .write = cgroup_file_write,
2101 .seq_start = cgroup_seqfile_start,
2102 .seq_next = cgroup_seqfile_next,
2103 .seq_stop = cgroup_seqfile_stop,
2104 .seq_show = cgroup_seqfile_show,
2108 * cgroup_rename - Only allow simple rename of directories in place.
2110 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2111 const char *new_name_str)
2113 struct cgroup *cgrp = kn->priv;
2116 if (kernfs_type(kn) != KERNFS_DIR)
2118 if (kn->parent != new_parent)
2122 * This isn't a proper migration and its usefulness is very
2123 * limited. Disallow if sane_behavior.
2125 if (cgroup_sane_behavior(cgrp))
2128 mutex_lock(&cgroup_tree_mutex);
2129 mutex_lock(&cgroup_mutex);
2131 ret = kernfs_rename(kn, new_parent, new_name_str);
2133 mutex_unlock(&cgroup_mutex);
2134 mutex_unlock(&cgroup_tree_mutex);
2138 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2140 char name[CGROUP_FILE_NAME_MAX];
2141 struct kernfs_node *kn;
2142 struct lock_class_key *key = NULL;
2144 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2145 key = &cft->lockdep_key;
2147 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2148 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2156 * cgroup_addrm_files - add or remove files to a cgroup directory
2157 * @cgrp: the target cgroup
2158 * @cfts: array of cftypes to be added
2159 * @is_add: whether to add or remove
2161 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2162 * For removals, this function never fails. If addition fails, this
2163 * function doesn't remove files already added. The caller is responsible
2166 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2172 lockdep_assert_held(&cgroup_tree_mutex);
2174 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2175 /* does cft->flags tell us to skip this file on @cgrp? */
2176 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2178 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2180 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2184 ret = cgroup_add_file(cgrp, cft);
2186 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2191 cgroup_rm_file(cgrp, cft);
2197 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2200 struct cgroup_subsys *ss = cfts[0].ss;
2201 struct cgroup *root = &ss->root->top_cgroup;
2202 struct cgroup_subsys_state *css;
2205 lockdep_assert_held(&cgroup_tree_mutex);
2207 /* don't bother if @ss isn't attached */
2208 if (ss->root == &cgroup_dummy_root)
2211 /* add/rm files for all cgroups created before */
2212 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2213 struct cgroup *cgrp = css->cgroup;
2215 if (cgroup_is_dead(cgrp))
2218 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2224 kernfs_activate(root->kn);
2228 static void cgroup_exit_cftypes(struct cftype *cfts)
2232 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2233 /* free copy for custom atomic_write_len, see init_cftypes() */
2234 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2241 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2245 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2246 struct kernfs_ops *kf_ops;
2248 WARN_ON(cft->ss || cft->kf_ops);
2251 kf_ops = &cgroup_kf_ops;
2253 kf_ops = &cgroup_kf_single_ops;
2256 * Ugh... if @cft wants a custom max_write_len, we need to
2257 * make a copy of kf_ops to set its atomic_write_len.
2259 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
2260 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
2262 cgroup_exit_cftypes(cfts);
2265 kf_ops->atomic_write_len = cft->max_write_len;
2268 cft->kf_ops = kf_ops;
2275 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
2277 lockdep_assert_held(&cgroup_tree_mutex);
2279 if (!cfts || !cfts[0].ss)
2282 list_del(&cfts->node);
2283 cgroup_apply_cftypes(cfts, false);
2284 cgroup_exit_cftypes(cfts);
2289 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2290 * @cfts: zero-length name terminated array of cftypes
2292 * Unregister @cfts. Files described by @cfts are removed from all
2293 * existing cgroups and all future cgroups won't have them either. This
2294 * function can be called anytime whether @cfts' subsys is attached or not.
2296 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2299 int cgroup_rm_cftypes(struct cftype *cfts)
2303 mutex_lock(&cgroup_tree_mutex);
2304 ret = cgroup_rm_cftypes_locked(cfts);
2305 mutex_unlock(&cgroup_tree_mutex);
2310 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2311 * @ss: target cgroup subsystem
2312 * @cfts: zero-length name terminated array of cftypes
2314 * Register @cfts to @ss. Files described by @cfts are created for all
2315 * existing cgroups to which @ss is attached and all future cgroups will
2316 * have them too. This function can be called anytime whether @ss is
2319 * Returns 0 on successful registration, -errno on failure. Note that this
2320 * function currently returns 0 as long as @cfts registration is successful
2321 * even if some file creation attempts on existing cgroups fail.
2323 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2327 ret = cgroup_init_cftypes(ss, cfts);
2331 mutex_lock(&cgroup_tree_mutex);
2333 list_add_tail(&cfts->node, &ss->cfts);
2334 ret = cgroup_apply_cftypes(cfts, true);
2336 cgroup_rm_cftypes_locked(cfts);
2338 mutex_unlock(&cgroup_tree_mutex);
2341 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2344 * cgroup_task_count - count the number of tasks in a cgroup.
2345 * @cgrp: the cgroup in question
2347 * Return the number of tasks in the cgroup.
2349 int cgroup_task_count(const struct cgroup *cgrp)
2352 struct cgrp_cset_link *link;
2354 read_lock(&css_set_lock);
2355 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2356 count += atomic_read(&link->cset->refcount);
2357 read_unlock(&css_set_lock);
2362 * To reduce the fork() overhead for systems that are not actually using
2363 * their cgroups capability, we don't maintain the lists running through
2364 * each css_set to its tasks until we see the list actually used - in other
2365 * words after the first call to css_task_iter_start().
2367 static void cgroup_enable_task_cg_lists(void)
2369 struct task_struct *p, *g;
2370 write_lock(&css_set_lock);
2371 use_task_css_set_links = 1;
2373 * We need tasklist_lock because RCU is not safe against
2374 * while_each_thread(). Besides, a forking task that has passed
2375 * cgroup_post_fork() without seeing use_task_css_set_links = 1
2376 * is not guaranteed to have its child immediately visible in the
2377 * tasklist if we walk through it with RCU.
2379 read_lock(&tasklist_lock);
2380 do_each_thread(g, p) {
2383 * We should check if the process is exiting, otherwise
2384 * it will race with cgroup_exit() in that the list
2385 * entry won't be deleted though the process has exited.
2387 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2388 list_add(&p->cg_list, &task_css_set(p)->tasks);
2390 } while_each_thread(g, p);
2391 read_unlock(&tasklist_lock);
2392 write_unlock(&css_set_lock);
2396 * css_next_child - find the next child of a given css
2397 * @pos_css: the current position (%NULL to initiate traversal)
2398 * @parent_css: css whose children to walk
2400 * This function returns the next child of @parent_css and should be called
2401 * under either cgroup_mutex or RCU read lock. The only requirement is
2402 * that @parent_css and @pos_css are accessible. The next sibling is
2403 * guaranteed to be returned regardless of their states.
2405 struct cgroup_subsys_state *
2406 css_next_child(struct cgroup_subsys_state *pos_css,
2407 struct cgroup_subsys_state *parent_css)
2409 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2410 struct cgroup *cgrp = parent_css->cgroup;
2411 struct cgroup *next;
2413 cgroup_assert_mutexes_or_rcu_locked();
2416 * @pos could already have been removed. Once a cgroup is removed,
2417 * its ->sibling.next is no longer updated when its next sibling
2418 * changes. As CGRP_DEAD assertion is serialized and happens
2419 * before the cgroup is taken off the ->sibling list, if we see it
2420 * unasserted, it's guaranteed that the next sibling hasn't
2421 * finished its grace period even if it's already removed, and thus
2422 * safe to dereference from this RCU critical section. If
2423 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2424 * to be visible as %true here.
2426 * If @pos is dead, its next pointer can't be dereferenced;
2427 * however, as each cgroup is given a monotonically increasing
2428 * unique serial number and always appended to the sibling list,
2429 * the next one can be found by walking the parent's children until
2430 * we see a cgroup with higher serial number than @pos's. While
2431 * this path can be slower, it's taken only when either the current
2432 * cgroup is removed or iteration and removal race.
2435 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2436 } else if (likely(!cgroup_is_dead(pos))) {
2437 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2439 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2440 if (next->serial_nr > pos->serial_nr)
2444 if (&next->sibling == &cgrp->children)
2447 return cgroup_css(next, parent_css->ss);
2449 EXPORT_SYMBOL_GPL(css_next_child);
2452 * css_next_descendant_pre - find the next descendant for pre-order walk
2453 * @pos: the current position (%NULL to initiate traversal)
2454 * @root: css whose descendants to walk
2456 * To be used by css_for_each_descendant_pre(). Find the next descendant
2457 * to visit for pre-order traversal of @root's descendants. @root is
2458 * included in the iteration and the first node to be visited.
2460 * While this function requires cgroup_mutex or RCU read locking, it
2461 * doesn't require the whole traversal to be contained in a single critical
2462 * section. This function will return the correct next descendant as long
2463 * as both @pos and @root are accessible and @pos is a descendant of @root.
2465 struct cgroup_subsys_state *
2466 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2467 struct cgroup_subsys_state *root)
2469 struct cgroup_subsys_state *next;
2471 cgroup_assert_mutexes_or_rcu_locked();
2473 /* if first iteration, visit @root */
2477 /* visit the first child if exists */
2478 next = css_next_child(NULL, pos);
2482 /* no child, visit my or the closest ancestor's next sibling */
2483 while (pos != root) {
2484 next = css_next_child(pos, css_parent(pos));
2487 pos = css_parent(pos);
2492 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
2495 * css_rightmost_descendant - return the rightmost descendant of a css
2496 * @pos: css of interest
2498 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2499 * is returned. This can be used during pre-order traversal to skip
2502 * While this function requires cgroup_mutex or RCU read locking, it
2503 * doesn't require the whole traversal to be contained in a single critical
2504 * section. This function will return the correct rightmost descendant as
2505 * long as @pos is accessible.
2507 struct cgroup_subsys_state *
2508 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2510 struct cgroup_subsys_state *last, *tmp;
2512 cgroup_assert_mutexes_or_rcu_locked();
2516 /* ->prev isn't RCU safe, walk ->next till the end */
2518 css_for_each_child(tmp, last)
2524 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
2526 static struct cgroup_subsys_state *
2527 css_leftmost_descendant(struct cgroup_subsys_state *pos)
2529 struct cgroup_subsys_state *last;
2533 pos = css_next_child(NULL, pos);
2540 * css_next_descendant_post - find the next descendant for post-order walk
2541 * @pos: the current position (%NULL to initiate traversal)
2542 * @root: css whose descendants to walk
2544 * To be used by css_for_each_descendant_post(). Find the next descendant
2545 * to visit for post-order traversal of @root's descendants. @root is
2546 * included in the iteration and the last node to be visited.
2548 * While this function requires cgroup_mutex or RCU read locking, it
2549 * doesn't require the whole traversal to be contained in a single critical
2550 * section. This function will return the correct next descendant as long
2551 * as both @pos and @cgroup are accessible and @pos is a descendant of
2554 struct cgroup_subsys_state *
2555 css_next_descendant_post(struct cgroup_subsys_state *pos,
2556 struct cgroup_subsys_state *root)
2558 struct cgroup_subsys_state *next;
2560 cgroup_assert_mutexes_or_rcu_locked();
2562 /* if first iteration, visit leftmost descendant which may be @root */
2564 return css_leftmost_descendant(root);
2566 /* if we visited @root, we're done */
2570 /* if there's an unvisited sibling, visit its leftmost descendant */
2571 next = css_next_child(pos, css_parent(pos));
2573 return css_leftmost_descendant(next);
2575 /* no sibling left, visit parent */
2576 return css_parent(pos);
2578 EXPORT_SYMBOL_GPL(css_next_descendant_post);
2581 * css_advance_task_iter - advance a task itererator to the next css_set
2582 * @it: the iterator to advance
2584 * Advance @it to the next css_set to walk.
2586 static void css_advance_task_iter(struct css_task_iter *it)
2588 struct list_head *l = it->cset_link;
2589 struct cgrp_cset_link *link;
2590 struct css_set *cset;
2592 /* Advance to the next non-empty css_set */
2595 if (l == &it->origin_css->cgroup->cset_links) {
2596 it->cset_link = NULL;
2599 link = list_entry(l, struct cgrp_cset_link, cset_link);
2601 } while (list_empty(&cset->tasks));
2603 it->task = cset->tasks.next;
2607 * css_task_iter_start - initiate task iteration
2608 * @css: the css to walk tasks of
2609 * @it: the task iterator to use
2611 * Initiate iteration through the tasks of @css. The caller can call
2612 * css_task_iter_next() to walk through the tasks until the function
2613 * returns NULL. On completion of iteration, css_task_iter_end() must be
2616 * Note that this function acquires a lock which is released when the
2617 * iteration finishes. The caller can't sleep while iteration is in
2620 void css_task_iter_start(struct cgroup_subsys_state *css,
2621 struct css_task_iter *it)
2622 __acquires(css_set_lock)
2625 * The first time anyone tries to iterate across a css, we need to
2626 * enable the list linking each css_set to its tasks, and fix up
2627 * all existing tasks.
2629 if (!use_task_css_set_links)
2630 cgroup_enable_task_cg_lists();
2632 read_lock(&css_set_lock);
2634 it->origin_css = css;
2635 it->cset_link = &css->cgroup->cset_links;
2637 css_advance_task_iter(it);
2641 * css_task_iter_next - return the next task for the iterator
2642 * @it: the task iterator being iterated
2644 * The "next" function for task iteration. @it should have been
2645 * initialized via css_task_iter_start(). Returns NULL when the iteration
2648 struct task_struct *css_task_iter_next(struct css_task_iter *it)
2650 struct task_struct *res;
2651 struct list_head *l = it->task;
2652 struct cgrp_cset_link *link;
2654 /* If the iterator cg is NULL, we have no tasks */
2657 res = list_entry(l, struct task_struct, cg_list);
2658 /* Advance iterator to find next entry */
2660 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
2661 if (l == &link->cset->tasks) {
2663 * We reached the end of this task list - move on to the
2664 * next cgrp_cset_link.
2666 css_advance_task_iter(it);
2674 * css_task_iter_end - finish task iteration
2675 * @it: the task iterator to finish
2677 * Finish task iteration started by css_task_iter_start().
2679 void css_task_iter_end(struct css_task_iter *it)
2680 __releases(css_set_lock)
2682 read_unlock(&css_set_lock);
2685 static inline int started_after_time(struct task_struct *t1,
2686 struct timespec *time,
2687 struct task_struct *t2)
2689 int start_diff = timespec_compare(&t1->start_time, time);
2690 if (start_diff > 0) {
2692 } else if (start_diff < 0) {
2696 * Arbitrarily, if two processes started at the same
2697 * time, we'll say that the lower pointer value
2698 * started first. Note that t2 may have exited by now
2699 * so this may not be a valid pointer any longer, but
2700 * that's fine - it still serves to distinguish
2701 * between two tasks started (effectively) simultaneously.
2708 * This function is a callback from heap_insert() and is used to order
2710 * In this case we order the heap in descending task start time.
2712 static inline int started_after(void *p1, void *p2)
2714 struct task_struct *t1 = p1;
2715 struct task_struct *t2 = p2;
2716 return started_after_time(t1, &t2->start_time, t2);
2720 * css_scan_tasks - iterate though all the tasks in a css
2721 * @css: the css to iterate tasks of
2722 * @test: optional test callback
2723 * @process: process callback
2724 * @data: data passed to @test and @process
2725 * @heap: optional pre-allocated heap used for task iteration
2727 * Iterate through all the tasks in @css, calling @test for each, and if it
2728 * returns %true, call @process for it also.
2730 * @test may be NULL, meaning always true (select all tasks), which
2731 * effectively duplicates css_task_iter_{start,next,end}() but does not
2732 * lock css_set_lock for the call to @process.
2734 * It is guaranteed that @process will act on every task that is a member
2735 * of @css for the duration of this call. This function may or may not
2736 * call @process for tasks that exit or move to a different css during the
2737 * call, or are forked or move into the css during the call.
2739 * Note that @test may be called with locks held, and may in some
2740 * situations be called multiple times for the same task, so it should be
2743 * If @heap is non-NULL, a heap has been pre-allocated and will be used for
2744 * heap operations (and its "gt" member will be overwritten), else a
2745 * temporary heap will be used (allocation of which may cause this function
2748 int css_scan_tasks(struct cgroup_subsys_state *css,
2749 bool (*test)(struct task_struct *, void *),
2750 void (*process)(struct task_struct *, void *),
2751 void *data, struct ptr_heap *heap)
2754 struct css_task_iter it;
2755 struct task_struct *p, *dropped;
2756 /* Never dereference latest_task, since it's not refcounted */
2757 struct task_struct *latest_task = NULL;
2758 struct ptr_heap tmp_heap;
2759 struct timespec latest_time = { 0, 0 };
2762 /* The caller supplied our heap and pre-allocated its memory */
2763 heap->gt = &started_after;
2765 /* We need to allocate our own heap memory */
2767 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2769 /* cannot allocate the heap */
2775 * Scan tasks in the css, using the @test callback to determine
2776 * which are of interest, and invoking @process callback on the
2777 * ones which need an update. Since we don't want to hold any
2778 * locks during the task updates, gather tasks to be processed in a
2779 * heap structure. The heap is sorted by descending task start
2780 * time. If the statically-sized heap fills up, we overflow tasks
2781 * that started later, and in future iterations only consider tasks
2782 * that started after the latest task in the previous pass. This
2783 * guarantees forward progress and that we don't miss any tasks.
2786 css_task_iter_start(css, &it);
2787 while ((p = css_task_iter_next(&it))) {
2789 * Only affect tasks that qualify per the caller's callback,
2790 * if he provided one
2792 if (test && !test(p, data))
2795 * Only process tasks that started after the last task
2798 if (!started_after_time(p, &latest_time, latest_task))
2800 dropped = heap_insert(heap, p);
2801 if (dropped == NULL) {
2803 * The new task was inserted; the heap wasn't
2807 } else if (dropped != p) {
2809 * The new task was inserted, and pushed out a
2813 put_task_struct(dropped);
2816 * Else the new task was newer than anything already in
2817 * the heap and wasn't inserted
2820 css_task_iter_end(&it);
2823 for (i = 0; i < heap->size; i++) {
2824 struct task_struct *q = heap->ptrs[i];
2826 latest_time = q->start_time;
2829 /* Process the task per the caller's callback */
2834 * If we had to process any tasks at all, scan again
2835 * in case some of them were in the middle of forking
2836 * children that didn't get processed.
2837 * Not the most efficient way to do it, but it avoids
2838 * having to take callback_mutex in the fork path
2842 if (heap == &tmp_heap)
2843 heap_free(&tmp_heap);
2847 static void cgroup_transfer_one_task(struct task_struct *task, void *data)
2849 struct cgroup *new_cgroup = data;
2851 mutex_lock(&cgroup_mutex);
2852 cgroup_attach_task(new_cgroup, task, false);
2853 mutex_unlock(&cgroup_mutex);
2857 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
2858 * @to: cgroup to which the tasks will be moved
2859 * @from: cgroup in which the tasks currently reside
2861 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
2863 return css_scan_tasks(&from->dummy_css, NULL, cgroup_transfer_one_task,
2868 * Stuff for reading the 'tasks'/'procs' files.
2870 * Reading this file can return large amounts of data if a cgroup has
2871 * *lots* of attached tasks. So it may need several calls to read(),
2872 * but we cannot guarantee that the information we produce is correct
2873 * unless we produce it entirely atomically.
2877 /* which pidlist file are we talking about? */
2878 enum cgroup_filetype {
2884 * A pidlist is a list of pids that virtually represents the contents of one
2885 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
2886 * a pair (one each for procs, tasks) for each pid namespace that's relevant
2889 struct cgroup_pidlist {
2891 * used to find which pidlist is wanted. doesn't change as long as
2892 * this particular list stays in the list.
2894 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
2897 /* how many elements the above list has */
2899 /* each of these stored in a list by its cgroup */
2900 struct list_head links;
2901 /* pointer to the cgroup we belong to, for list removal purposes */
2902 struct cgroup *owner;
2903 /* for delayed destruction */
2904 struct delayed_work destroy_dwork;
2908 * The following two functions "fix" the issue where there are more pids
2909 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2910 * TODO: replace with a kernel-wide solution to this problem
2912 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2913 static void *pidlist_allocate(int count)
2915 if (PIDLIST_TOO_LARGE(count))
2916 return vmalloc(count * sizeof(pid_t));
2918 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2921 static void pidlist_free(void *p)
2923 if (is_vmalloc_addr(p))
2930 * Used to destroy all pidlists lingering waiting for destroy timer. None
2931 * should be left afterwards.
2933 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
2935 struct cgroup_pidlist *l, *tmp_l;
2937 mutex_lock(&cgrp->pidlist_mutex);
2938 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
2939 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
2940 mutex_unlock(&cgrp->pidlist_mutex);
2942 flush_workqueue(cgroup_pidlist_destroy_wq);
2943 BUG_ON(!list_empty(&cgrp->pidlists));
2946 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
2948 struct delayed_work *dwork = to_delayed_work(work);
2949 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
2951 struct cgroup_pidlist *tofree = NULL;
2953 mutex_lock(&l->owner->pidlist_mutex);
2956 * Destroy iff we didn't get queued again. The state won't change
2957 * as destroy_dwork can only be queued while locked.
2959 if (!delayed_work_pending(dwork)) {
2960 list_del(&l->links);
2961 pidlist_free(l->list);
2962 put_pid_ns(l->key.ns);
2966 mutex_unlock(&l->owner->pidlist_mutex);
2971 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2972 * Returns the number of unique elements.
2974 static int pidlist_uniq(pid_t *list, int length)
2979 * we presume the 0th element is unique, so i starts at 1. trivial
2980 * edge cases first; no work needs to be done for either
2982 if (length == 0 || length == 1)
2984 /* src and dest walk down the list; dest counts unique elements */
2985 for (src = 1; src < length; src++) {
2986 /* find next unique element */
2987 while (list[src] == list[src-1]) {
2992 /* dest always points to where the next unique element goes */
2993 list[dest] = list[src];
3001 * The two pid files - task and cgroup.procs - guaranteed that the result
3002 * is sorted, which forced this whole pidlist fiasco. As pid order is
3003 * different per namespace, each namespace needs differently sorted list,
3004 * making it impossible to use, for example, single rbtree of member tasks
3005 * sorted by task pointer. As pidlists can be fairly large, allocating one
3006 * per open file is dangerous, so cgroup had to implement shared pool of
3007 * pidlists keyed by cgroup and namespace.
3009 * All this extra complexity was caused by the original implementation
3010 * committing to an entirely unnecessary property. In the long term, we
3011 * want to do away with it. Explicitly scramble sort order if
3012 * sane_behavior so that no such expectation exists in the new interface.
3014 * Scrambling is done by swapping every two consecutive bits, which is
3015 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3017 static pid_t pid_fry(pid_t pid)
3019 unsigned a = pid & 0x55555555;
3020 unsigned b = pid & 0xAAAAAAAA;
3022 return (a << 1) | (b >> 1);
3025 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3027 if (cgroup_sane_behavior(cgrp))
3028 return pid_fry(pid);
3033 static int cmppid(const void *a, const void *b)
3035 return *(pid_t *)a - *(pid_t *)b;
3038 static int fried_cmppid(const void *a, const void *b)
3040 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3043 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3044 enum cgroup_filetype type)
3046 struct cgroup_pidlist *l;
3047 /* don't need task_nsproxy() if we're looking at ourself */
3048 struct pid_namespace *ns = task_active_pid_ns(current);
3050 lockdep_assert_held(&cgrp->pidlist_mutex);
3052 list_for_each_entry(l, &cgrp->pidlists, links)
3053 if (l->key.type == type && l->key.ns == ns)
3059 * find the appropriate pidlist for our purpose (given procs vs tasks)
3060 * returns with the lock on that pidlist already held, and takes care
3061 * of the use count, or returns NULL with no locks held if we're out of
3064 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3065 enum cgroup_filetype type)
3067 struct cgroup_pidlist *l;
3069 lockdep_assert_held(&cgrp->pidlist_mutex);
3071 l = cgroup_pidlist_find(cgrp, type);
3075 /* entry not found; create a new one */
3076 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3080 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3082 /* don't need task_nsproxy() if we're looking at ourself */
3083 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3085 list_add(&l->links, &cgrp->pidlists);
3090 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3092 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3093 struct cgroup_pidlist **lp)
3097 int pid, n = 0; /* used for populating the array */
3098 struct css_task_iter it;
3099 struct task_struct *tsk;
3100 struct cgroup_pidlist *l;
3102 lockdep_assert_held(&cgrp->pidlist_mutex);
3105 * If cgroup gets more users after we read count, we won't have
3106 * enough space - tough. This race is indistinguishable to the
3107 * caller from the case that the additional cgroup users didn't
3108 * show up until sometime later on.
3110 length = cgroup_task_count(cgrp);
3111 array = pidlist_allocate(length);
3114 /* now, populate the array */
3115 css_task_iter_start(&cgrp->dummy_css, &it);
3116 while ((tsk = css_task_iter_next(&it))) {
3117 if (unlikely(n == length))
3119 /* get tgid or pid for procs or tasks file respectively */
3120 if (type == CGROUP_FILE_PROCS)
3121 pid = task_tgid_vnr(tsk);
3123 pid = task_pid_vnr(tsk);
3124 if (pid > 0) /* make sure to only use valid results */
3127 css_task_iter_end(&it);
3129 /* now sort & (if procs) strip out duplicates */
3130 if (cgroup_sane_behavior(cgrp))
3131 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3133 sort(array, length, sizeof(pid_t), cmppid, NULL);
3134 if (type == CGROUP_FILE_PROCS)
3135 length = pidlist_uniq(array, length);
3137 l = cgroup_pidlist_find_create(cgrp, type);
3139 mutex_unlock(&cgrp->pidlist_mutex);
3140 pidlist_free(array);
3144 /* store array, freeing old if necessary */
3145 pidlist_free(l->list);
3153 * cgroupstats_build - build and fill cgroupstats
3154 * @stats: cgroupstats to fill information into
3155 * @dentry: A dentry entry belonging to the cgroup for which stats have
3158 * Build and fill cgroupstats so that taskstats can export it to user
3161 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3163 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3164 struct cgroup *cgrp;
3165 struct css_task_iter it;
3166 struct task_struct *tsk;
3168 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3169 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3170 kernfs_type(kn) != KERNFS_DIR)
3174 * We aren't being called from kernfs and there's no guarantee on
3175 * @kn->priv's validity. For this and css_tryget_from_dir(),
3176 * @kn->priv is RCU safe. Let's do the RCU dancing.
3179 cgrp = rcu_dereference(kn->priv);
3185 css_task_iter_start(&cgrp->dummy_css, &it);
3186 while ((tsk = css_task_iter_next(&it))) {
3187 switch (tsk->state) {
3189 stats->nr_running++;
3191 case TASK_INTERRUPTIBLE:
3192 stats->nr_sleeping++;
3194 case TASK_UNINTERRUPTIBLE:
3195 stats->nr_uninterruptible++;
3198 stats->nr_stopped++;
3201 if (delayacct_is_task_waiting_on_io(tsk))
3202 stats->nr_io_wait++;
3206 css_task_iter_end(&it);
3214 * seq_file methods for the tasks/procs files. The seq_file position is the
3215 * next pid to display; the seq_file iterator is a pointer to the pid
3216 * in the cgroup->l->list array.
3219 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3222 * Initially we receive a position value that corresponds to
3223 * one more than the last pid shown (or 0 on the first call or
3224 * after a seek to the start). Use a binary-search to find the
3225 * next pid to display, if any
3227 struct kernfs_open_file *of = s->private;
3228 struct cgroup *cgrp = seq_css(s)->cgroup;
3229 struct cgroup_pidlist *l;
3230 enum cgroup_filetype type = seq_cft(s)->private;
3231 int index = 0, pid = *pos;
3234 mutex_lock(&cgrp->pidlist_mutex);
3237 * !NULL @of->priv indicates that this isn't the first start()
3238 * after open. If the matching pidlist is around, we can use that.
3239 * Look for it. Note that @of->priv can't be used directly. It
3240 * could already have been destroyed.
3243 of->priv = cgroup_pidlist_find(cgrp, type);
3246 * Either this is the first start() after open or the matching
3247 * pidlist has been destroyed inbetween. Create a new one.
3250 ret = pidlist_array_load(cgrp, type,
3251 (struct cgroup_pidlist **)&of->priv);
3253 return ERR_PTR(ret);
3258 int end = l->length;
3260 while (index < end) {
3261 int mid = (index + end) / 2;
3262 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3265 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3271 /* If we're off the end of the array, we're done */
3272 if (index >= l->length)
3274 /* Update the abstract position to be the actual pid that we found */
3275 iter = l->list + index;
3276 *pos = cgroup_pid_fry(cgrp, *iter);
3280 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3282 struct kernfs_open_file *of = s->private;
3283 struct cgroup_pidlist *l = of->priv;
3286 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3287 CGROUP_PIDLIST_DESTROY_DELAY);
3288 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3291 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3293 struct kernfs_open_file *of = s->private;
3294 struct cgroup_pidlist *l = of->priv;
3296 pid_t *end = l->list + l->length;
3298 * Advance to the next pid in the array. If this goes off the
3305 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3310 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3312 return seq_printf(s, "%d\n", *(int *)v);
3316 * seq_operations functions for iterating on pidlists through seq_file -
3317 * independent of whether it's tasks or procs
3319 static const struct seq_operations cgroup_pidlist_seq_operations = {
3320 .start = cgroup_pidlist_start,
3321 .stop = cgroup_pidlist_stop,
3322 .next = cgroup_pidlist_next,
3323 .show = cgroup_pidlist_show,
3326 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3329 return notify_on_release(css->cgroup);
3332 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3333 struct cftype *cft, u64 val)
3335 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3337 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3339 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3343 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3346 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3349 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3350 struct cftype *cft, u64 val)
3353 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3355 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3359 static struct cftype cgroup_base_files[] = {
3361 .name = "cgroup.procs",
3362 .seq_start = cgroup_pidlist_start,
3363 .seq_next = cgroup_pidlist_next,
3364 .seq_stop = cgroup_pidlist_stop,
3365 .seq_show = cgroup_pidlist_show,
3366 .private = CGROUP_FILE_PROCS,
3367 .write_u64 = cgroup_procs_write,
3368 .mode = S_IRUGO | S_IWUSR,
3371 .name = "cgroup.clone_children",
3372 .flags = CFTYPE_INSANE,
3373 .read_u64 = cgroup_clone_children_read,
3374 .write_u64 = cgroup_clone_children_write,
3377 .name = "cgroup.sane_behavior",
3378 .flags = CFTYPE_ONLY_ON_ROOT,
3379 .seq_show = cgroup_sane_behavior_show,
3383 * Historical crazy stuff. These don't have "cgroup." prefix and
3384 * don't exist if sane_behavior. If you're depending on these, be
3385 * prepared to be burned.
3389 .flags = CFTYPE_INSANE, /* use "procs" instead */
3390 .seq_start = cgroup_pidlist_start,
3391 .seq_next = cgroup_pidlist_next,
3392 .seq_stop = cgroup_pidlist_stop,
3393 .seq_show = cgroup_pidlist_show,
3394 .private = CGROUP_FILE_TASKS,
3395 .write_u64 = cgroup_tasks_write,
3396 .mode = S_IRUGO | S_IWUSR,
3399 .name = "notify_on_release",
3400 .flags = CFTYPE_INSANE,
3401 .read_u64 = cgroup_read_notify_on_release,
3402 .write_u64 = cgroup_write_notify_on_release,
3405 .name = "release_agent",
3406 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3407 .seq_show = cgroup_release_agent_show,
3408 .write_string = cgroup_release_agent_write,
3409 .max_write_len = PATH_MAX - 1,
3415 * cgroup_populate_dir - create subsys files in a cgroup directory
3416 * @cgrp: target cgroup
3417 * @subsys_mask: mask of the subsystem ids whose files should be added
3419 * On failure, no file is added.
3421 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3423 struct cgroup_subsys *ss;
3426 /* process cftsets of each subsystem */
3427 for_each_subsys(ss, i) {
3428 struct cftype *cfts;
3430 if (!test_bit(i, &subsys_mask))
3433 list_for_each_entry(cfts, &ss->cfts, node) {
3434 ret = cgroup_addrm_files(cgrp, cfts, true);
3441 cgroup_clear_dir(cgrp, subsys_mask);
3446 * css destruction is four-stage process.
3448 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3449 * Implemented in kill_css().
3451 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3452 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3453 * by invoking offline_css(). After offlining, the base ref is put.
3454 * Implemented in css_killed_work_fn().
3456 * 3. When the percpu_ref reaches zero, the only possible remaining
3457 * accessors are inside RCU read sections. css_release() schedules the
3460 * 4. After the grace period, the css can be freed. Implemented in
3461 * css_free_work_fn().
3463 * It is actually hairier because both step 2 and 4 require process context
3464 * and thus involve punting to css->destroy_work adding two additional
3465 * steps to the already complex sequence.
3467 static void css_free_work_fn(struct work_struct *work)
3469 struct cgroup_subsys_state *css =
3470 container_of(work, struct cgroup_subsys_state, destroy_work);
3471 struct cgroup *cgrp = css->cgroup;
3474 css_put(css->parent);
3476 css->ss->css_free(css);
3480 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3482 struct cgroup_subsys_state *css =
3483 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3485 INIT_WORK(&css->destroy_work, css_free_work_fn);
3486 queue_work(cgroup_destroy_wq, &css->destroy_work);
3489 static void css_release(struct percpu_ref *ref)
3491 struct cgroup_subsys_state *css =
3492 container_of(ref, struct cgroup_subsys_state, refcnt);
3494 rcu_assign_pointer(css->cgroup->subsys[css->ss->id], NULL);
3495 call_rcu(&css->rcu_head, css_free_rcu_fn);
3498 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3499 struct cgroup *cgrp)
3506 css->parent = cgroup_css(cgrp->parent, ss);
3508 css->flags |= CSS_ROOT;
3510 BUG_ON(cgroup_css(cgrp, ss));
3513 /* invoke ->css_online() on a new CSS and mark it online if successful */
3514 static int online_css(struct cgroup_subsys_state *css)
3516 struct cgroup_subsys *ss = css->ss;
3519 lockdep_assert_held(&cgroup_tree_mutex);
3520 lockdep_assert_held(&cgroup_mutex);
3523 ret = ss->css_online(css);
3525 css->flags |= CSS_ONLINE;
3526 css->cgroup->nr_css++;
3527 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
3532 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
3533 static void offline_css(struct cgroup_subsys_state *css)
3535 struct cgroup_subsys *ss = css->ss;
3537 lockdep_assert_held(&cgroup_tree_mutex);
3538 lockdep_assert_held(&cgroup_mutex);
3540 if (!(css->flags & CSS_ONLINE))
3543 if (ss->css_offline)
3544 ss->css_offline(css);
3546 css->flags &= ~CSS_ONLINE;
3547 css->cgroup->nr_css--;
3548 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
3552 * create_css - create a cgroup_subsys_state
3553 * @cgrp: the cgroup new css will be associated with
3554 * @ss: the subsys of new css
3556 * Create a new css associated with @cgrp - @ss pair. On success, the new
3557 * css is online and installed in @cgrp with all interface files created.
3558 * Returns 0 on success, -errno on failure.
3560 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
3562 struct cgroup *parent = cgrp->parent;
3563 struct cgroup_subsys_state *css;
3566 lockdep_assert_held(&cgroup_mutex);
3568 css = ss->css_alloc(cgroup_css(parent, ss));
3570 return PTR_ERR(css);
3572 err = percpu_ref_init(&css->refcnt, css_release);
3576 init_css(css, ss, cgrp);
3578 err = cgroup_populate_dir(cgrp, 1 << ss->id);
3582 err = online_css(css);
3587 css_get(css->parent);
3589 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
3591 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",
3592 current->comm, current->pid, ss->name);
3593 if (!strcmp(ss->name, "memory"))
3594 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
3595 ss->warned_broken_hierarchy = true;
3601 percpu_ref_cancel_init(&css->refcnt);
3607 * cgroup_create - create a cgroup
3608 * @parent: cgroup that will be parent of the new cgroup
3609 * @name: name of the new cgroup
3610 * @mode: mode to set on new cgroup
3612 static long cgroup_create(struct cgroup *parent, const char *name,
3615 struct cgroup *cgrp;
3616 struct cgroupfs_root *root = parent->root;
3618 struct cgroup_subsys *ss;
3619 struct kernfs_node *kn;
3621 /* allocate the cgroup and its ID, 0 is reserved for the root */
3622 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3626 mutex_lock(&cgroup_tree_mutex);
3629 * Only live parents can have children. Note that the liveliness
3630 * check isn't strictly necessary because cgroup_mkdir() and
3631 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
3632 * anyway so that locking is contained inside cgroup proper and we
3633 * don't get nasty surprises if we ever grow another caller.
3635 if (!cgroup_lock_live_group(parent)) {
3637 goto err_unlock_tree;
3641 * Temporarily set the pointer to NULL, so idr_find() won't return
3642 * a half-baked cgroup.
3644 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
3650 init_cgroup_housekeeping(cgrp);
3652 cgrp->parent = parent;
3653 cgrp->dummy_css.parent = &parent->dummy_css;
3654 cgrp->root = parent->root;
3656 if (notify_on_release(parent))
3657 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3659 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
3660 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
3662 /* create the directory */
3663 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
3671 * This extra ref will be put in cgroup_free_fn() and guarantees
3672 * that @cgrp->kn is always accessible.
3676 cgrp->serial_nr = cgroup_serial_nr_next++;
3678 /* allocation complete, commit to creation */
3679 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
3680 atomic_inc(&root->nr_cgrps);
3684 * @cgrp is now fully operational. If something fails after this
3685 * point, it'll be released via the normal destruction path.
3687 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
3689 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
3693 /* let's create and online css's */
3694 for_each_subsys(ss, ssid) {
3695 if (root->subsys_mask & (1 << ssid)) {
3696 err = create_css(cgrp, ss);
3702 kernfs_activate(kn);
3704 mutex_unlock(&cgroup_mutex);
3705 mutex_unlock(&cgroup_tree_mutex);
3710 idr_remove(&root->cgroup_idr, cgrp->id);
3712 mutex_unlock(&cgroup_mutex);
3714 mutex_unlock(&cgroup_tree_mutex);
3719 cgroup_destroy_locked(cgrp);
3720 mutex_unlock(&cgroup_mutex);
3721 mutex_unlock(&cgroup_tree_mutex);
3725 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3728 struct cgroup *parent = parent_kn->priv;
3730 return cgroup_create(parent, name, mode);
3734 * This is called when the refcnt of a css is confirmed to be killed.
3735 * css_tryget() is now guaranteed to fail.
3737 static void css_killed_work_fn(struct work_struct *work)
3739 struct cgroup_subsys_state *css =
3740 container_of(work, struct cgroup_subsys_state, destroy_work);
3741 struct cgroup *cgrp = css->cgroup;
3743 mutex_lock(&cgroup_tree_mutex);
3744 mutex_lock(&cgroup_mutex);
3747 * css_tryget() is guaranteed to fail now. Tell subsystems to
3748 * initate destruction.
3753 * If @cgrp is marked dead, it's waiting for refs of all css's to
3754 * be disabled before proceeding to the second phase of cgroup
3755 * destruction. If we are the last one, kick it off.
3757 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
3758 cgroup_destroy_css_killed(cgrp);
3760 mutex_unlock(&cgroup_mutex);
3761 mutex_unlock(&cgroup_tree_mutex);
3764 * Put the css refs from kill_css(). Each css holds an extra
3765 * reference to the cgroup's dentry and cgroup removal proceeds
3766 * regardless of css refs. On the last put of each css, whenever
3767 * that may be, the extra dentry ref is put so that dentry
3768 * destruction happens only after all css's are released.
3773 /* css kill confirmation processing requires process context, bounce */
3774 static void css_killed_ref_fn(struct percpu_ref *ref)
3776 struct cgroup_subsys_state *css =
3777 container_of(ref, struct cgroup_subsys_state, refcnt);
3779 INIT_WORK(&css->destroy_work, css_killed_work_fn);
3780 queue_work(cgroup_destroy_wq, &css->destroy_work);
3784 * kill_css - destroy a css
3785 * @css: css to destroy
3787 * This function initiates destruction of @css by removing cgroup interface
3788 * files and putting its base reference. ->css_offline() will be invoked
3789 * asynchronously once css_tryget() is guaranteed to fail and when the
3790 * reference count reaches zero, @css will be released.
3792 static void kill_css(struct cgroup_subsys_state *css)
3795 * This must happen before css is disassociated with its cgroup.
3796 * See seq_css() for details.
3798 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3801 * Killing would put the base ref, but we need to keep it alive
3802 * until after ->css_offline().
3807 * cgroup core guarantees that, by the time ->css_offline() is
3808 * invoked, no new css reference will be given out via
3809 * css_tryget(). We can't simply call percpu_ref_kill() and
3810 * proceed to offlining css's because percpu_ref_kill() doesn't
3811 * guarantee that the ref is seen as killed on all CPUs on return.
3813 * Use percpu_ref_kill_and_confirm() to get notifications as each
3814 * css is confirmed to be seen as killed on all CPUs.
3816 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
3820 * cgroup_destroy_locked - the first stage of cgroup destruction
3821 * @cgrp: cgroup to be destroyed
3823 * css's make use of percpu refcnts whose killing latency shouldn't be
3824 * exposed to userland and are RCU protected. Also, cgroup core needs to
3825 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
3826 * invoked. To satisfy all the requirements, destruction is implemented in
3827 * the following two steps.
3829 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
3830 * userland visible parts and start killing the percpu refcnts of
3831 * css's. Set up so that the next stage will be kicked off once all
3832 * the percpu refcnts are confirmed to be killed.
3834 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
3835 * rest of destruction. Once all cgroup references are gone, the
3836 * cgroup is RCU-freed.
3838 * This function implements s1. After this step, @cgrp is gone as far as
3839 * the userland is concerned and a new cgroup with the same name may be
3840 * created. As cgroup doesn't care about the names internally, this
3841 * doesn't cause any problem.
3843 static int cgroup_destroy_locked(struct cgroup *cgrp)
3844 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
3846 struct cgroup *child;
3847 struct cgroup_subsys_state *css;
3851 lockdep_assert_held(&cgroup_tree_mutex);
3852 lockdep_assert_held(&cgroup_mutex);
3855 * css_set_lock synchronizes access to ->cset_links and prevents
3856 * @cgrp from being removed while __put_css_set() is in progress.
3858 read_lock(&css_set_lock);
3859 empty = list_empty(&cgrp->cset_links);
3860 read_unlock(&css_set_lock);
3865 * Make sure there's no live children. We can't test ->children
3866 * emptiness as dead children linger on it while being destroyed;
3867 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
3871 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
3872 empty = cgroup_is_dead(child);
3881 * Initiate massacre of all css's. cgroup_destroy_css_killed()
3882 * will be invoked to perform the rest of destruction once the
3883 * percpu refs of all css's are confirmed to be killed. This
3884 * involves removing the subsystem's files, drop cgroup_mutex.
3886 mutex_unlock(&cgroup_mutex);
3887 for_each_css(css, ssid, cgrp)
3889 mutex_lock(&cgroup_mutex);
3892 * Mark @cgrp dead. This prevents further task migration and child
3893 * creation by disabling cgroup_lock_live_group(). Note that
3894 * CGRP_DEAD assertion is depended upon by css_next_child() to
3895 * resume iteration after dropping RCU read lock. See
3896 * css_next_child() for details.
3898 set_bit(CGRP_DEAD, &cgrp->flags);
3900 /* CGRP_DEAD is set, remove from ->release_list for the last time */
3901 raw_spin_lock(&release_list_lock);
3902 if (!list_empty(&cgrp->release_list))
3903 list_del_init(&cgrp->release_list);
3904 raw_spin_unlock(&release_list_lock);
3907 * If @cgrp has css's attached, the second stage of cgroup
3908 * destruction is kicked off from css_killed_work_fn() after the
3909 * refs of all attached css's are killed. If @cgrp doesn't have
3910 * any css, we kick it off here.
3913 cgroup_destroy_css_killed(cgrp);
3915 /* remove @cgrp directory along with the base files */
3916 mutex_unlock(&cgroup_mutex);
3919 * There are two control paths which try to determine cgroup from
3920 * dentry without going through kernfs - cgroupstats_build() and
3921 * css_tryget_from_dir(). Those are supported by RCU protecting
3922 * clearing of cgrp->kn->priv backpointer, which should happen
3923 * after all files under it have been removed.
3925 kernfs_remove(cgrp->kn); /* @cgrp has an extra ref on its kn */
3926 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
3928 mutex_lock(&cgroup_mutex);
3934 * cgroup_destroy_css_killed - the second step of cgroup destruction
3935 * @work: cgroup->destroy_free_work
3937 * This function is invoked from a work item for a cgroup which is being
3938 * destroyed after all css's are offlined and performs the rest of
3939 * destruction. This is the second step of destruction described in the
3940 * comment above cgroup_destroy_locked().
3942 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
3944 struct cgroup *parent = cgrp->parent;
3946 lockdep_assert_held(&cgroup_tree_mutex);
3947 lockdep_assert_held(&cgroup_mutex);
3949 /* delete this cgroup from parent->children */
3950 list_del_rcu(&cgrp->sibling);
3954 set_bit(CGRP_RELEASABLE, &parent->flags);
3955 check_for_release(parent);
3958 static int cgroup_rmdir(struct kernfs_node *kn)
3960 struct cgroup *cgrp = kn->priv;
3964 * This is self-destruction but @kn can't be removed while this
3965 * callback is in progress. Let's break active protection. Once
3966 * the protection is broken, @cgrp can be destroyed at any point.
3967 * Pin it so that it stays accessible.
3970 kernfs_break_active_protection(kn);
3972 mutex_lock(&cgroup_tree_mutex);
3973 mutex_lock(&cgroup_mutex);
3976 * @cgrp might already have been destroyed while we're trying to
3979 if (!cgroup_is_dead(cgrp))
3980 ret = cgroup_destroy_locked(cgrp);
3982 mutex_unlock(&cgroup_mutex);
3983 mutex_unlock(&cgroup_tree_mutex);
3985 kernfs_unbreak_active_protection(kn);
3990 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
3991 .remount_fs = cgroup_remount,
3992 .show_options = cgroup_show_options,
3993 .mkdir = cgroup_mkdir,
3994 .rmdir = cgroup_rmdir,
3995 .rename = cgroup_rename,
3998 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4000 struct cgroup_subsys_state *css;
4002 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4004 mutex_lock(&cgroup_tree_mutex);
4005 mutex_lock(&cgroup_mutex);
4007 INIT_LIST_HEAD(&ss->cfts);
4009 /* Create the top cgroup state for this subsystem */
4010 ss->root = &cgroup_dummy_root;
4011 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4012 /* We don't handle early failures gracefully */
4013 BUG_ON(IS_ERR(css));
4014 init_css(css, ss, cgroup_dummy_top);
4016 /* Update the init_css_set to contain a subsys
4017 * pointer to this state - since the subsystem is
4018 * newly registered, all tasks and hence the
4019 * init_css_set is in the subsystem's top cgroup. */
4020 init_css_set.subsys[ss->id] = css;
4022 need_forkexit_callback |= ss->fork || ss->exit;
4024 /* At system boot, before all subsystems have been
4025 * registered, no tasks have been forked, so we don't
4026 * need to invoke fork callbacks here. */
4027 BUG_ON(!list_empty(&init_task.tasks));
4029 BUG_ON(online_css(css));
4031 mutex_unlock(&cgroup_mutex);
4032 mutex_unlock(&cgroup_tree_mutex);
4036 * cgroup_init_early - cgroup initialization at system boot
4038 * Initialize cgroups at system boot, and initialize any
4039 * subsystems that request early init.
4041 int __init cgroup_init_early(void)
4043 struct cgroup_subsys *ss;
4046 atomic_set(&init_css_set.refcount, 1);
4047 INIT_LIST_HEAD(&init_css_set.cgrp_links);
4048 INIT_LIST_HEAD(&init_css_set.tasks);
4049 INIT_HLIST_NODE(&init_css_set.hlist);
4051 init_cgroup_root(&cgroup_dummy_root);
4052 cgroup_root_count = 1;
4053 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4055 init_cgrp_cset_link.cset = &init_css_set;
4056 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4057 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4058 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4060 for_each_subsys(ss, i) {
4061 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4062 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4063 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4065 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4066 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4069 ss->name = cgroup_subsys_name[i];
4072 cgroup_init_subsys(ss);
4078 * cgroup_init - cgroup initialization
4080 * Register cgroup filesystem and /proc file, and initialize
4081 * any subsystems that didn't request early init.
4083 int __init cgroup_init(void)
4085 struct cgroup_subsys *ss;
4089 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4091 for_each_subsys(ss, i) {
4092 if (!ss->early_init)
4093 cgroup_init_subsys(ss);
4096 * cftype registration needs kmalloc and can't be done
4097 * during early_init. Register base cftypes separately.
4099 if (ss->base_cftypes)
4100 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4103 /* allocate id for the dummy hierarchy */
4104 mutex_lock(&cgroup_mutex);
4106 /* Add init_css_set to the hash table */
4107 key = css_set_hash(init_css_set.subsys);
4108 hash_add(css_set_table, &init_css_set.hlist, key);
4110 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4112 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
4116 mutex_unlock(&cgroup_mutex);
4118 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4122 err = register_filesystem(&cgroup_fs_type);
4124 kobject_put(cgroup_kobj);
4128 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4132 static int __init cgroup_wq_init(void)
4135 * There isn't much point in executing destruction path in
4136 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4138 * XXX: Must be ordered to make sure parent is offlined after
4139 * children. The ordering requirement is for memcg where a
4140 * parent's offline may wait for a child's leading to deadlock. In
4141 * the long term, this should be fixed from memcg side.
4143 * We would prefer to do this in cgroup_init() above, but that
4144 * is called before init_workqueues(): so leave this until after.
4146 cgroup_destroy_wq = alloc_ordered_workqueue("cgroup_destroy", 0);
4147 BUG_ON(!cgroup_destroy_wq);
4150 * Used to destroy pidlists and separate to serve as flush domain.
4151 * Cap @max_active to 1 too.
4153 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4155 BUG_ON(!cgroup_pidlist_destroy_wq);
4159 core_initcall(cgroup_wq_init);
4162 * proc_cgroup_show()
4163 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4164 * - Used for /proc/<pid>/cgroup.
4165 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4166 * doesn't really matter if tsk->cgroup changes after we read it,
4167 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4168 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4169 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4170 * cgroup to top_cgroup.
4173 /* TODO: Use a proper seq_file iterator */
4174 int proc_cgroup_show(struct seq_file *m, void *v)
4177 struct task_struct *tsk;
4180 struct cgroupfs_root *root;
4183 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4189 tsk = get_pid_task(pid, PIDTYPE_PID);
4195 mutex_lock(&cgroup_mutex);
4197 for_each_active_root(root) {
4198 struct cgroup_subsys *ss;
4199 struct cgroup *cgrp;
4200 int ssid, count = 0;
4202 seq_printf(m, "%d:", root->hierarchy_id);
4203 for_each_subsys(ss, ssid)
4204 if (root->subsys_mask & (1 << ssid))
4205 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4206 if (strlen(root->name))
4207 seq_printf(m, "%sname=%s", count ? "," : "",
4210 cgrp = task_cgroup_from_root(tsk, root);
4211 path = cgroup_path(cgrp, buf, PATH_MAX);
4213 retval = -ENAMETOOLONG;
4221 mutex_unlock(&cgroup_mutex);
4222 put_task_struct(tsk);
4229 /* Display information about each subsystem and each hierarchy */
4230 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4232 struct cgroup_subsys *ss;
4235 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4237 * ideally we don't want subsystems moving around while we do this.
4238 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4239 * subsys/hierarchy state.
4241 mutex_lock(&cgroup_mutex);
4243 for_each_subsys(ss, i)
4244 seq_printf(m, "%s\t%d\t%d\t%d\n",
4245 ss->name, ss->root->hierarchy_id,
4246 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4248 mutex_unlock(&cgroup_mutex);
4252 static int cgroupstats_open(struct inode *inode, struct file *file)
4254 return single_open(file, proc_cgroupstats_show, NULL);
4257 static const struct file_operations proc_cgroupstats_operations = {
4258 .open = cgroupstats_open,
4260 .llseek = seq_lseek,
4261 .release = single_release,
4265 * cgroup_fork - attach newly forked task to its parents cgroup.
4266 * @child: pointer to task_struct of forking parent process.
4268 * Description: A task inherits its parent's cgroup at fork().
4270 * A pointer to the shared css_set was automatically copied in
4271 * fork.c by dup_task_struct(). However, we ignore that copy, since
4272 * it was not made under the protection of RCU or cgroup_mutex, so
4273 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4274 * have already changed current->cgroups, allowing the previously
4275 * referenced cgroup group to be removed and freed.
4277 * At the point that cgroup_fork() is called, 'current' is the parent
4278 * task, and the passed argument 'child' points to the child task.
4280 void cgroup_fork(struct task_struct *child)
4283 get_css_set(task_css_set(current));
4284 child->cgroups = current->cgroups;
4285 task_unlock(current);
4286 INIT_LIST_HEAD(&child->cg_list);
4290 * cgroup_post_fork - called on a new task after adding it to the task list
4291 * @child: the task in question
4293 * Adds the task to the list running through its css_set if necessary and
4294 * call the subsystem fork() callbacks. Has to be after the task is
4295 * visible on the task list in case we race with the first call to
4296 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4299 void cgroup_post_fork(struct task_struct *child)
4301 struct cgroup_subsys *ss;
4305 * use_task_css_set_links is set to 1 before we walk the tasklist
4306 * under the tasklist_lock and we read it here after we added the child
4307 * to the tasklist under the tasklist_lock as well. If the child wasn't
4308 * yet in the tasklist when we walked through it from
4309 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4310 * should be visible now due to the paired locking and barriers implied
4311 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4312 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4315 if (use_task_css_set_links) {
4316 write_lock(&css_set_lock);
4318 if (list_empty(&child->cg_list))
4319 list_add(&child->cg_list, &task_css_set(child)->tasks);
4321 write_unlock(&css_set_lock);
4325 * Call ss->fork(). This must happen after @child is linked on
4326 * css_set; otherwise, @child might change state between ->fork()
4327 * and addition to css_set.
4329 if (need_forkexit_callback) {
4330 for_each_subsys(ss, i)
4337 * cgroup_exit - detach cgroup from exiting task
4338 * @tsk: pointer to task_struct of exiting process
4339 * @run_callback: run exit callbacks?
4341 * Description: Detach cgroup from @tsk and release it.
4343 * Note that cgroups marked notify_on_release force every task in
4344 * them to take the global cgroup_mutex mutex when exiting.
4345 * This could impact scaling on very large systems. Be reluctant to
4346 * use notify_on_release cgroups where very high task exit scaling
4347 * is required on large systems.
4349 * the_top_cgroup_hack:
4351 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4353 * We call cgroup_exit() while the task is still competent to
4354 * handle notify_on_release(), then leave the task attached to the
4355 * root cgroup in each hierarchy for the remainder of its exit.
4357 * To do this properly, we would increment the reference count on
4358 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4359 * code we would add a second cgroup function call, to drop that
4360 * reference. This would just create an unnecessary hot spot on
4361 * the top_cgroup reference count, to no avail.
4363 * Normally, holding a reference to a cgroup without bumping its
4364 * count is unsafe. The cgroup could go away, or someone could
4365 * attach us to a different cgroup, decrementing the count on
4366 * the first cgroup that we never incremented. But in this case,
4367 * top_cgroup isn't going away, and either task has PF_EXITING set,
4368 * which wards off any cgroup_attach_task() attempts, or task is a failed
4369 * fork, never visible to cgroup_attach_task.
4371 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4373 struct cgroup_subsys *ss;
4374 struct css_set *cset;
4378 * Unlink from the css_set task list if necessary.
4379 * Optimistically check cg_list before taking
4382 if (!list_empty(&tsk->cg_list)) {
4383 write_lock(&css_set_lock);
4384 if (!list_empty(&tsk->cg_list))
4385 list_del_init(&tsk->cg_list);
4386 write_unlock(&css_set_lock);
4389 /* Reassign the task to the init_css_set. */
4391 cset = task_css_set(tsk);
4392 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4394 if (run_callbacks && need_forkexit_callback) {
4395 /* see cgroup_post_fork() for details */
4396 for_each_subsys(ss, i) {
4398 struct cgroup_subsys_state *old_css = cset->subsys[i];
4399 struct cgroup_subsys_state *css = task_css(tsk, i);
4401 ss->exit(css, old_css, tsk);
4407 put_css_set_taskexit(cset);
4410 static void check_for_release(struct cgroup *cgrp)
4412 if (cgroup_is_releasable(cgrp) &&
4413 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4415 * Control Group is currently removeable. If it's not
4416 * already queued for a userspace notification, queue
4419 int need_schedule_work = 0;
4421 raw_spin_lock(&release_list_lock);
4422 if (!cgroup_is_dead(cgrp) &&
4423 list_empty(&cgrp->release_list)) {
4424 list_add(&cgrp->release_list, &release_list);
4425 need_schedule_work = 1;
4427 raw_spin_unlock(&release_list_lock);
4428 if (need_schedule_work)
4429 schedule_work(&release_agent_work);
4434 * Notify userspace when a cgroup is released, by running the
4435 * configured release agent with the name of the cgroup (path
4436 * relative to the root of cgroup file system) as the argument.
4438 * Most likely, this user command will try to rmdir this cgroup.
4440 * This races with the possibility that some other task will be
4441 * attached to this cgroup before it is removed, or that some other
4442 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4443 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4444 * unused, and this cgroup will be reprieved from its death sentence,
4445 * to continue to serve a useful existence. Next time it's released,
4446 * we will get notified again, if it still has 'notify_on_release' set.
4448 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4449 * means only wait until the task is successfully execve()'d. The
4450 * separate release agent task is forked by call_usermodehelper(),
4451 * then control in this thread returns here, without waiting for the
4452 * release agent task. We don't bother to wait because the caller of
4453 * this routine has no use for the exit status of the release agent
4454 * task, so no sense holding our caller up for that.
4456 static void cgroup_release_agent(struct work_struct *work)
4458 BUG_ON(work != &release_agent_work);
4459 mutex_lock(&cgroup_mutex);
4460 raw_spin_lock(&release_list_lock);
4461 while (!list_empty(&release_list)) {
4462 char *argv[3], *envp[3];
4464 char *pathbuf = NULL, *agentbuf = NULL, *path;
4465 struct cgroup *cgrp = list_entry(release_list.next,
4468 list_del_init(&cgrp->release_list);
4469 raw_spin_unlock(&release_list_lock);
4470 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
4473 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
4476 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4481 argv[i++] = agentbuf;
4486 /* minimal command environment */
4487 envp[i++] = "HOME=/";
4488 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4491 /* Drop the lock while we invoke the usermode helper,
4492 * since the exec could involve hitting disk and hence
4493 * be a slow process */
4494 mutex_unlock(&cgroup_mutex);
4495 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4496 mutex_lock(&cgroup_mutex);
4500 raw_spin_lock(&release_list_lock);
4502 raw_spin_unlock(&release_list_lock);
4503 mutex_unlock(&cgroup_mutex);
4506 static int __init cgroup_disable(char *str)
4508 struct cgroup_subsys *ss;
4512 while ((token = strsep(&str, ",")) != NULL) {
4516 for_each_subsys(ss, i) {
4517 if (!strcmp(token, ss->name)) {
4519 printk(KERN_INFO "Disabling %s control group"
4520 " subsystem\n", ss->name);
4527 __setup("cgroup_disable=", cgroup_disable);
4530 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
4531 * @dentry: directory dentry of interest
4532 * @ss: subsystem of interest
4534 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
4535 * to get the corresponding css and return it. If such css doesn't exist
4536 * or can't be pinned, an ERR_PTR value is returned.
4538 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
4539 struct cgroup_subsys *ss)
4541 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4542 struct cgroup_subsys_state *css = NULL;
4543 struct cgroup *cgrp;
4545 /* is @dentry a cgroup dir? */
4546 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4547 kernfs_type(kn) != KERNFS_DIR)
4548 return ERR_PTR(-EBADF);
4553 * This path doesn't originate from kernfs and @kn could already
4554 * have been or be removed at any point. @kn->priv is RCU
4555 * protected for this access. See destroy_locked() for details.
4557 cgrp = rcu_dereference(kn->priv);
4559 css = cgroup_css(cgrp, ss);
4561 if (!css || !css_tryget(css))
4562 css = ERR_PTR(-ENOENT);
4569 * css_from_id - lookup css by id
4570 * @id: the cgroup id
4571 * @ss: cgroup subsys to be looked into
4573 * Returns the css if there's valid one with @id, otherwise returns NULL.
4574 * Should be called under rcu_read_lock().
4576 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
4578 struct cgroup *cgrp;
4580 cgroup_assert_mutexes_or_rcu_locked();
4582 cgrp = idr_find(&ss->root->cgroup_idr, id);
4584 return cgroup_css(cgrp, ss);
4588 #ifdef CONFIG_CGROUP_DEBUG
4589 static struct cgroup_subsys_state *
4590 debug_css_alloc(struct cgroup_subsys_state *parent_css)
4592 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4595 return ERR_PTR(-ENOMEM);
4600 static void debug_css_free(struct cgroup_subsys_state *css)
4605 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
4608 return cgroup_task_count(css->cgroup);
4611 static u64 current_css_set_read(struct cgroup_subsys_state *css,
4614 return (u64)(unsigned long)current->cgroups;
4617 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
4623 count = atomic_read(&task_css_set(current)->refcount);
4628 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
4630 struct cgrp_cset_link *link;
4631 struct css_set *cset;
4634 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
4638 read_lock(&css_set_lock);
4640 cset = rcu_dereference(current->cgroups);
4641 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
4642 struct cgroup *c = link->cgrp;
4643 const char *name = "?";
4645 if (c != cgroup_dummy_top) {
4646 cgroup_name(c, name_buf, NAME_MAX + 1);
4650 seq_printf(seq, "Root %d group %s\n",
4651 c->root->hierarchy_id, name);
4654 read_unlock(&css_set_lock);
4659 #define MAX_TASKS_SHOWN_PER_CSS 25
4660 static int cgroup_css_links_read(struct seq_file *seq, void *v)
4662 struct cgroup_subsys_state *css = seq_css(seq);
4663 struct cgrp_cset_link *link;
4665 read_lock(&css_set_lock);
4666 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
4667 struct css_set *cset = link->cset;
4668 struct task_struct *task;
4670 seq_printf(seq, "css_set %p\n", cset);
4671 list_for_each_entry(task, &cset->tasks, cg_list) {
4672 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4673 seq_puts(seq, " ...\n");
4676 seq_printf(seq, " task %d\n",
4677 task_pid_vnr(task));
4681 read_unlock(&css_set_lock);
4685 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
4687 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4690 static struct cftype debug_files[] = {
4692 .name = "taskcount",
4693 .read_u64 = debug_taskcount_read,
4697 .name = "current_css_set",
4698 .read_u64 = current_css_set_read,
4702 .name = "current_css_set_refcount",
4703 .read_u64 = current_css_set_refcount_read,
4707 .name = "current_css_set_cg_links",
4708 .seq_show = current_css_set_cg_links_read,
4712 .name = "cgroup_css_links",
4713 .seq_show = cgroup_css_links_read,
4717 .name = "releasable",
4718 .read_u64 = releasable_read,
4724 struct cgroup_subsys debug_cgrp_subsys = {
4725 .css_alloc = debug_css_alloc,
4726 .css_free = debug_css_free,
4727 .base_cftypes = debug_files,
4729 #endif /* CONFIG_CGROUP_DEBUG */