2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #include <linux/cgroup.h>
30 #include <linux/cred.h>
31 #include <linux/ctype.h>
32 #include <linux/errno.h>
33 #include <linux/init_task.h>
34 #include <linux/kernel.h>
35 #include <linux/list.h>
37 #include <linux/mutex.h>
38 #include <linux/mount.h>
39 #include <linux/pagemap.h>
40 #include <linux/proc_fs.h>
41 #include <linux/rcupdate.h>
42 #include <linux/sched.h>
43 #include <linux/slab.h>
44 #include <linux/spinlock.h>
45 #include <linux/string.h>
46 #include <linux/sort.h>
47 #include <linux/kmod.h>
48 #include <linux/delayacct.h>
49 #include <linux/cgroupstats.h>
50 #include <linux/hashtable.h>
51 #include <linux/pid_namespace.h>
52 #include <linux/idr.h>
53 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
54 #include <linux/flex_array.h> /* used in cgroup_attach_task */
55 #include <linux/kthread.h>
57 #include <linux/atomic.h>
60 * pidlists linger the following amount before being destroyed. The goal
61 * is avoiding frequent destruction in the middle of consecutive read calls
62 * Expiring in the middle is a performance problem not a correctness one.
63 * 1 sec should be enough.
65 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
67 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
71 * cgroup_tree_mutex nests above cgroup_mutex and protects cftypes, file
72 * creation/removal and hierarchy changing operations including cgroup
73 * creation, removal, css association and controller rebinding. This outer
74 * lock is needed mainly to resolve the circular dependency between kernfs
75 * active ref and cgroup_mutex. cgroup_tree_mutex nests above both.
77 static DEFINE_MUTEX(cgroup_tree_mutex);
80 * cgroup_mutex is the master lock. Any modification to cgroup or its
81 * hierarchy must be performed while holding it.
83 #ifdef CONFIG_PROVE_RCU
84 DEFINE_MUTEX(cgroup_mutex);
85 EXPORT_SYMBOL_GPL(cgroup_mutex); /* only for lockdep */
87 static DEFINE_MUTEX(cgroup_mutex);
91 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
92 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
94 static DEFINE_SPINLOCK(release_agent_path_lock);
96 #define cgroup_assert_mutexes_or_rcu_locked() \
97 rcu_lockdep_assert(rcu_read_lock_held() || \
98 lockdep_is_held(&cgroup_tree_mutex) || \
99 lockdep_is_held(&cgroup_mutex), \
100 "cgroup_[tree_]mutex or RCU read lock required");
103 * cgroup destruction makes heavy use of work items and there can be a lot
104 * of concurrent destructions. Use a separate workqueue so that cgroup
105 * destruction work items don't end up filling up max_active of system_wq
106 * which may lead to deadlock.
108 static struct workqueue_struct *cgroup_destroy_wq;
111 * pidlist destructions need to be flushed on cgroup destruction. Use a
112 * separate workqueue as flush domain.
114 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
116 /* generate an array of cgroup subsystem pointers */
117 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
118 static struct cgroup_subsys *cgroup_subsys[] = {
119 #include <linux/cgroup_subsys.h>
123 /* array of cgroup subsystem names */
124 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
125 static const char *cgroup_subsys_name[] = {
126 #include <linux/cgroup_subsys.h>
131 * The dummy hierarchy, reserved for the subsystems that are otherwise
132 * unattached - it never has more than a single cgroup, and all tasks are
133 * part of that cgroup.
135 static struct cgroupfs_root cgroup_dummy_root;
137 /* dummy_top is a shorthand for the dummy hierarchy's top cgroup */
138 static struct cgroup * const cgroup_dummy_top = &cgroup_dummy_root.top_cgroup;
140 /* The list of hierarchy roots */
142 static LIST_HEAD(cgroup_roots);
143 static int cgroup_root_count;
145 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
146 static DEFINE_IDR(cgroup_hierarchy_idr);
148 static struct cgroup_name root_cgroup_name = { .name = "/" };
151 * Assign a monotonically increasing serial number to cgroups. It
152 * guarantees cgroups with bigger numbers are newer than those with smaller
153 * numbers. Also, as cgroups are always appended to the parent's
154 * ->children list, it guarantees that sibling cgroups are always sorted in
155 * the ascending serial number order on the list. Protected by
158 static u64 cgroup_serial_nr_next = 1;
160 /* This flag indicates whether tasks in the fork and exit paths should
161 * check for fork/exit handlers to call. This avoids us having to do
162 * extra work in the fork/exit path if none of the subsystems need to
165 static int need_forkexit_callback __read_mostly;
167 static struct cftype cgroup_base_files[];
169 static void cgroup_put(struct cgroup *cgrp);
170 static int rebind_subsystems(struct cgroupfs_root *root,
171 unsigned long added_mask, unsigned removed_mask);
172 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
173 static int cgroup_destroy_locked(struct cgroup *cgrp);
174 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
176 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
179 * cgroup_css - obtain a cgroup's css for the specified subsystem
180 * @cgrp: the cgroup of interest
181 * @ss: the subsystem of interest (%NULL returns the dummy_css)
183 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
184 * function must be called either under cgroup_mutex or rcu_read_lock() and
185 * the caller is responsible for pinning the returned css if it wants to
186 * keep accessing it outside the said locks. This function may return
187 * %NULL if @cgrp doesn't have @subsys_id enabled.
189 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
190 struct cgroup_subsys *ss)
193 return rcu_dereference_check(cgrp->subsys[ss->id],
194 lockdep_is_held(&cgroup_tree_mutex) ||
195 lockdep_is_held(&cgroup_mutex));
197 return &cgrp->dummy_css;
200 /* convenient tests for these bits */
201 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
203 return test_bit(CGRP_DEAD, &cgrp->flags);
206 struct cgroup_subsys_state *seq_css(struct seq_file *seq)
208 struct kernfs_open_file *of = seq->private;
209 struct cgroup *cgrp = of->kn->parent->priv;
210 struct cftype *cft = seq_cft(seq);
213 * This is open and unprotected implementation of cgroup_css().
214 * seq_css() is only called from a kernfs file operation which has
215 * an active reference on the file. Because all the subsystem
216 * files are drained before a css is disassociated with a cgroup,
217 * the matching css from the cgroup's subsys table is guaranteed to
218 * be and stay valid until the enclosing operation is complete.
221 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
223 return &cgrp->dummy_css;
225 EXPORT_SYMBOL_GPL(seq_css);
228 * cgroup_is_descendant - test ancestry
229 * @cgrp: the cgroup to be tested
230 * @ancestor: possible ancestor of @cgrp
232 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
233 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
234 * and @ancestor are accessible.
236 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
239 if (cgrp == ancestor)
245 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
247 static int cgroup_is_releasable(const struct cgroup *cgrp)
250 (1 << CGRP_RELEASABLE) |
251 (1 << CGRP_NOTIFY_ON_RELEASE);
252 return (cgrp->flags & bits) == bits;
255 static int notify_on_release(const struct cgroup *cgrp)
257 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
261 * for_each_css - iterate all css's of a cgroup
262 * @css: the iteration cursor
263 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
264 * @cgrp: the target cgroup to iterate css's of
266 * Should be called under cgroup_mutex.
268 #define for_each_css(css, ssid, cgrp) \
269 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
270 if (!((css) = rcu_dereference_check( \
271 (cgrp)->subsys[(ssid)], \
272 lockdep_is_held(&cgroup_tree_mutex) || \
273 lockdep_is_held(&cgroup_mutex)))) { } \
277 * for_each_subsys - iterate all enabled cgroup subsystems
278 * @ss: the iteration cursor
279 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
281 #define for_each_subsys(ss, ssid) \
282 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
283 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
285 /* iterate across the active hierarchies */
286 #define for_each_active_root(root) \
287 list_for_each_entry((root), &cgroup_roots, root_list)
290 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
291 * @cgrp: the cgroup to be checked for liveness
293 * On success, returns true; the mutex should be later unlocked. On
294 * failure returns false with no lock held.
296 static bool cgroup_lock_live_group(struct cgroup *cgrp)
298 mutex_lock(&cgroup_mutex);
299 if (cgroup_is_dead(cgrp)) {
300 mutex_unlock(&cgroup_mutex);
306 /* the list of cgroups eligible for automatic release. Protected by
307 * release_list_lock */
308 static LIST_HEAD(release_list);
309 static DEFINE_RAW_SPINLOCK(release_list_lock);
310 static void cgroup_release_agent(struct work_struct *work);
311 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
312 static void check_for_release(struct cgroup *cgrp);
315 * A cgroup can be associated with multiple css_sets as different tasks may
316 * belong to different cgroups on different hierarchies. In the other
317 * direction, a css_set is naturally associated with multiple cgroups.
318 * This M:N relationship is represented by the following link structure
319 * which exists for each association and allows traversing the associations
322 struct cgrp_cset_link {
323 /* the cgroup and css_set this link associates */
325 struct css_set *cset;
327 /* list of cgrp_cset_links anchored at cgrp->cset_links */
328 struct list_head cset_link;
330 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
331 struct list_head cgrp_link;
334 /* The default css_set - used by init and its children prior to any
335 * hierarchies being mounted. It contains a pointer to the root state
336 * for each subsystem. Also used to anchor the list of css_sets. Not
337 * reference-counted, to improve performance when child cgroups
338 * haven't been created.
341 static struct css_set init_css_set;
342 static struct cgrp_cset_link init_cgrp_cset_link;
345 * css_set_lock protects the list of css_set objects, and the chain of
346 * tasks off each css_set. Nests outside task->alloc_lock due to
347 * css_task_iter_start().
349 static DEFINE_RWLOCK(css_set_lock);
350 static int css_set_count;
353 * hash table for cgroup groups. This improves the performance to find
354 * an existing css_set. This hash doesn't (currently) take into
355 * account cgroups in empty hierarchies.
357 #define CSS_SET_HASH_BITS 7
358 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
360 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
362 unsigned long key = 0UL;
363 struct cgroup_subsys *ss;
366 for_each_subsys(ss, i)
367 key += (unsigned long)css[i];
368 key = (key >> 16) ^ key;
374 * We don't maintain the lists running through each css_set to its task
375 * until after the first call to css_task_iter_start(). This reduces the
376 * fork()/exit() overhead for people who have cgroups compiled into their
377 * kernel but not actually in use.
379 static int use_task_css_set_links __read_mostly;
381 static void __put_css_set(struct css_set *cset, int taskexit)
383 struct cgrp_cset_link *link, *tmp_link;
386 * Ensure that the refcount doesn't hit zero while any readers
387 * can see it. Similar to atomic_dec_and_lock(), but for an
390 if (atomic_add_unless(&cset->refcount, -1, 1))
392 write_lock(&css_set_lock);
393 if (!atomic_dec_and_test(&cset->refcount)) {
394 write_unlock(&css_set_lock);
398 /* This css_set is dead. unlink it and release cgroup refcounts */
399 hash_del(&cset->hlist);
402 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
403 struct cgroup *cgrp = link->cgrp;
405 list_del(&link->cset_link);
406 list_del(&link->cgrp_link);
408 /* @cgrp can't go away while we're holding css_set_lock */
409 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
411 set_bit(CGRP_RELEASABLE, &cgrp->flags);
412 check_for_release(cgrp);
418 write_unlock(&css_set_lock);
419 kfree_rcu(cset, rcu_head);
423 * refcounted get/put for css_set objects
425 static inline void get_css_set(struct css_set *cset)
427 atomic_inc(&cset->refcount);
430 static inline void put_css_set(struct css_set *cset)
432 __put_css_set(cset, 0);
435 static inline void put_css_set_taskexit(struct css_set *cset)
437 __put_css_set(cset, 1);
441 * compare_css_sets - helper function for find_existing_css_set().
442 * @cset: candidate css_set being tested
443 * @old_cset: existing css_set for a task
444 * @new_cgrp: cgroup that's being entered by the task
445 * @template: desired set of css pointers in css_set (pre-calculated)
447 * Returns true if "cset" matches "old_cset" except for the hierarchy
448 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
450 static bool compare_css_sets(struct css_set *cset,
451 struct css_set *old_cset,
452 struct cgroup *new_cgrp,
453 struct cgroup_subsys_state *template[])
455 struct list_head *l1, *l2;
457 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
458 /* Not all subsystems matched */
463 * Compare cgroup pointers in order to distinguish between
464 * different cgroups in heirarchies with no subsystems. We
465 * could get by with just this check alone (and skip the
466 * memcmp above) but on most setups the memcmp check will
467 * avoid the need for this more expensive check on almost all
471 l1 = &cset->cgrp_links;
472 l2 = &old_cset->cgrp_links;
474 struct cgrp_cset_link *link1, *link2;
475 struct cgroup *cgrp1, *cgrp2;
479 /* See if we reached the end - both lists are equal length. */
480 if (l1 == &cset->cgrp_links) {
481 BUG_ON(l2 != &old_cset->cgrp_links);
484 BUG_ON(l2 == &old_cset->cgrp_links);
486 /* Locate the cgroups associated with these links. */
487 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
488 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
491 /* Hierarchies should be linked in the same order. */
492 BUG_ON(cgrp1->root != cgrp2->root);
495 * If this hierarchy is the hierarchy of the cgroup
496 * that's changing, then we need to check that this
497 * css_set points to the new cgroup; if it's any other
498 * hierarchy, then this css_set should point to the
499 * same cgroup as the old css_set.
501 if (cgrp1->root == new_cgrp->root) {
502 if (cgrp1 != new_cgrp)
513 * find_existing_css_set - init css array and find the matching css_set
514 * @old_cset: the css_set that we're using before the cgroup transition
515 * @cgrp: the cgroup that we're moving into
516 * @template: out param for the new set of csses, should be clear on entry
518 static struct css_set *find_existing_css_set(struct css_set *old_cset,
520 struct cgroup_subsys_state *template[])
522 struct cgroupfs_root *root = cgrp->root;
523 struct cgroup_subsys *ss;
524 struct css_set *cset;
529 * Build the set of subsystem state objects that we want to see in the
530 * new css_set. while subsystems can change globally, the entries here
531 * won't change, so no need for locking.
533 for_each_subsys(ss, i) {
534 if (root->subsys_mask & (1UL << i)) {
535 /* Subsystem is in this hierarchy. So we want
536 * the subsystem state from the new
538 template[i] = cgroup_css(cgrp, ss);
540 /* Subsystem is not in this hierarchy, so we
541 * don't want to change the subsystem state */
542 template[i] = old_cset->subsys[i];
546 key = css_set_hash(template);
547 hash_for_each_possible(css_set_table, cset, hlist, key) {
548 if (!compare_css_sets(cset, old_cset, cgrp, template))
551 /* This css_set matches what we need */
555 /* No existing cgroup group matched */
559 static void free_cgrp_cset_links(struct list_head *links_to_free)
561 struct cgrp_cset_link *link, *tmp_link;
563 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
564 list_del(&link->cset_link);
570 * allocate_cgrp_cset_links - allocate cgrp_cset_links
571 * @count: the number of links to allocate
572 * @tmp_links: list_head the allocated links are put on
574 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
575 * through ->cset_link. Returns 0 on success or -errno.
577 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
579 struct cgrp_cset_link *link;
582 INIT_LIST_HEAD(tmp_links);
584 for (i = 0; i < count; i++) {
585 link = kzalloc(sizeof(*link), GFP_KERNEL);
587 free_cgrp_cset_links(tmp_links);
590 list_add(&link->cset_link, tmp_links);
596 * link_css_set - a helper function to link a css_set to a cgroup
597 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
598 * @cset: the css_set to be linked
599 * @cgrp: the destination cgroup
601 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
604 struct cgrp_cset_link *link;
606 BUG_ON(list_empty(tmp_links));
607 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
610 list_move(&link->cset_link, &cgrp->cset_links);
612 * Always add links to the tail of the list so that the list
613 * is sorted by order of hierarchy creation
615 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
619 * find_css_set - return a new css_set with one cgroup updated
620 * @old_cset: the baseline css_set
621 * @cgrp: the cgroup to be updated
623 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
624 * substituted into the appropriate hierarchy.
626 static struct css_set *find_css_set(struct css_set *old_cset,
629 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
630 struct css_set *cset;
631 struct list_head tmp_links;
632 struct cgrp_cset_link *link;
635 lockdep_assert_held(&cgroup_mutex);
637 /* First see if we already have a cgroup group that matches
639 read_lock(&css_set_lock);
640 cset = find_existing_css_set(old_cset, cgrp, template);
643 read_unlock(&css_set_lock);
648 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
652 /* Allocate all the cgrp_cset_link objects that we'll need */
653 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
658 atomic_set(&cset->refcount, 1);
659 INIT_LIST_HEAD(&cset->cgrp_links);
660 INIT_LIST_HEAD(&cset->tasks);
661 INIT_HLIST_NODE(&cset->hlist);
663 /* Copy the set of subsystem state objects generated in
664 * find_existing_css_set() */
665 memcpy(cset->subsys, template, sizeof(cset->subsys));
667 write_lock(&css_set_lock);
668 /* Add reference counts and links from the new css_set. */
669 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
670 struct cgroup *c = link->cgrp;
672 if (c->root == cgrp->root)
674 link_css_set(&tmp_links, cset, c);
677 BUG_ON(!list_empty(&tmp_links));
681 /* Add this cgroup group to the hash table */
682 key = css_set_hash(cset->subsys);
683 hash_add(css_set_table, &cset->hlist, key);
685 write_unlock(&css_set_lock);
690 static struct cgroupfs_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
692 struct cgroup *top_cgrp = kf_root->kn->priv;
694 return top_cgrp->root;
697 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
701 lockdep_assert_held(&cgroup_mutex);
703 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
708 root->hierarchy_id = id;
712 static void cgroup_exit_root_id(struct cgroupfs_root *root)
714 lockdep_assert_held(&cgroup_mutex);
716 if (root->hierarchy_id) {
717 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
718 root->hierarchy_id = 0;
722 static void cgroup_free_root(struct cgroupfs_root *root)
725 /* hierarhcy ID shoulid already have been released */
726 WARN_ON_ONCE(root->hierarchy_id);
728 idr_destroy(&root->cgroup_idr);
733 static void cgroup_get_root(struct cgroupfs_root *root)
736 * The caller must ensure that @root is alive, which can be
737 * achieved by holding a ref on one of the member cgroups or
738 * following a registered reference to @root while holding
741 WARN_ON_ONCE(atomic_read(&root->refcnt) <= 0);
742 atomic_inc(&root->refcnt);
745 static void cgroup_put_root(struct cgroupfs_root *root)
747 struct cgroup *cgrp = &root->top_cgroup;
748 struct cgrp_cset_link *link, *tmp_link;
752 * @root's refcnt reaching zero and its deregistration should be
753 * atomic w.r.t. cgroup_tree_mutex. This ensures that
754 * cgroup_get_root() is safe to invoke if @root is registered.
756 mutex_lock(&cgroup_tree_mutex);
757 if (!atomic_dec_and_test(&root->refcnt)) {
758 mutex_unlock(&cgroup_tree_mutex);
761 mutex_lock(&cgroup_mutex);
763 BUG_ON(root->number_of_cgroups != 1);
764 BUG_ON(!list_empty(&cgrp->children));
766 /* Rebind all subsystems back to the default hierarchy */
767 if (root->flags & CGRP_ROOT_SUBSYS_BOUND) {
768 ret = rebind_subsystems(root, 0, root->subsys_mask);
769 /* Shouldn't be able to fail ... */
774 * Release all the links from cset_links to this hierarchy's
777 write_lock(&css_set_lock);
779 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
780 list_del(&link->cset_link);
781 list_del(&link->cgrp_link);
784 write_unlock(&css_set_lock);
786 if (!list_empty(&root->root_list)) {
787 list_del(&root->root_list);
791 cgroup_exit_root_id(root);
793 mutex_unlock(&cgroup_mutex);
794 mutex_unlock(&cgroup_tree_mutex);
796 kernfs_destroy_root(root->kf_root);
797 cgroup_free_root(root);
801 * Return the cgroup for "task" from the given hierarchy. Must be
802 * called with cgroup_mutex held.
804 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
805 struct cgroupfs_root *root)
807 struct css_set *cset;
808 struct cgroup *res = NULL;
810 BUG_ON(!mutex_is_locked(&cgroup_mutex));
811 read_lock(&css_set_lock);
813 * No need to lock the task - since we hold cgroup_mutex the
814 * task can't change groups, so the only thing that can happen
815 * is that it exits and its css is set back to init_css_set.
817 cset = task_css_set(task);
818 if (cset == &init_css_set) {
819 res = &root->top_cgroup;
821 struct cgrp_cset_link *link;
823 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
824 struct cgroup *c = link->cgrp;
826 if (c->root == root) {
832 read_unlock(&css_set_lock);
838 * There is one global cgroup mutex. We also require taking
839 * task_lock() when dereferencing a task's cgroup subsys pointers.
840 * See "The task_lock() exception", at the end of this comment.
842 * A task must hold cgroup_mutex to modify cgroups.
844 * Any task can increment and decrement the count field without lock.
845 * So in general, code holding cgroup_mutex can't rely on the count
846 * field not changing. However, if the count goes to zero, then only
847 * cgroup_attach_task() can increment it again. Because a count of zero
848 * means that no tasks are currently attached, therefore there is no
849 * way a task attached to that cgroup can fork (the other way to
850 * increment the count). So code holding cgroup_mutex can safely
851 * assume that if the count is zero, it will stay zero. Similarly, if
852 * a task holds cgroup_mutex on a cgroup with zero count, it
853 * knows that the cgroup won't be removed, as cgroup_rmdir()
856 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
857 * (usually) take cgroup_mutex. These are the two most performance
858 * critical pieces of code here. The exception occurs on cgroup_exit(),
859 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
860 * is taken, and if the cgroup count is zero, a usermode call made
861 * to the release agent with the name of the cgroup (path relative to
862 * the root of cgroup file system) as the argument.
864 * A cgroup can only be deleted if both its 'count' of using tasks
865 * is zero, and its list of 'children' cgroups is empty. Since all
866 * tasks in the system use _some_ cgroup, and since there is always at
867 * least one task in the system (init, pid == 1), therefore, top_cgroup
868 * always has either children cgroups and/or using tasks. So we don't
869 * need a special hack to ensure that top_cgroup cannot be deleted.
871 * The task_lock() exception
873 * The need for this exception arises from the action of
874 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
875 * another. It does so using cgroup_mutex, however there are
876 * several performance critical places that need to reference
877 * task->cgroup without the expense of grabbing a system global
878 * mutex. Therefore except as noted below, when dereferencing or, as
879 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
880 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
881 * the task_struct routinely used for such matters.
883 * P.S. One more locking exception. RCU is used to guard the
884 * update of a tasks cgroup pointer by cgroup_attach_task()
887 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
888 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
889 static const struct file_operations proc_cgroupstats_operations;
891 static struct cgroup_name *cgroup_alloc_name(const char *name_str)
893 struct cgroup_name *name;
895 name = kmalloc(sizeof(*name) + strlen(name_str) + 1, GFP_KERNEL);
898 strcpy(name->name, name_str);
902 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
905 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
906 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
907 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
908 cft->ss->name, cft->name);
910 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
915 * cgroup_file_mode - deduce file mode of a control file
916 * @cft: the control file in question
918 * returns cft->mode if ->mode is not 0
919 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
920 * returns S_IRUGO if it has only a read handler
921 * returns S_IWUSR if it has only a write hander
923 static umode_t cgroup_file_mode(const struct cftype *cft)
930 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
933 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
940 static void cgroup_free_fn(struct work_struct *work)
942 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
944 mutex_lock(&cgroup_mutex);
945 cgrp->root->number_of_cgroups--;
946 mutex_unlock(&cgroup_mutex);
949 * We get a ref to the parent, and put the ref when this cgroup is
950 * being freed, so it's guaranteed that the parent won't be
951 * destroyed before its children.
953 cgroup_put(cgrp->parent);
955 /* put the root reference that we took when we created the cgroup */
956 cgroup_put_root(cgrp->root);
958 cgroup_pidlist_destroy_all(cgrp);
960 kernfs_put(cgrp->kn);
962 kfree(rcu_dereference_raw(cgrp->name));
966 static void cgroup_free_rcu(struct rcu_head *head)
968 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
970 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
971 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
974 static void cgroup_get(struct cgroup *cgrp)
976 WARN_ON_ONCE(cgroup_is_dead(cgrp));
977 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
978 atomic_inc(&cgrp->refcnt);
981 static void cgroup_put(struct cgroup *cgrp)
983 if (!atomic_dec_and_test(&cgrp->refcnt))
985 if (WARN_ON_ONCE(!cgroup_is_dead(cgrp)))
989 * XXX: cgrp->id is only used to look up css's. As cgroup and
990 * css's lifetimes will be decoupled, it should be made
991 * per-subsystem and moved to css->id so that lookups are
992 * successful until the target css is released.
994 mutex_lock(&cgroup_mutex);
995 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
996 mutex_unlock(&cgroup_mutex);
999 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
1002 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1004 char name[CGROUP_FILE_NAME_MAX];
1006 lockdep_assert_held(&cgroup_tree_mutex);
1007 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1011 * cgroup_clear_dir - remove subsys files in a cgroup directory
1012 * @cgrp: target cgroup
1013 * @subsys_mask: mask of the subsystem ids whose files should be removed
1015 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
1017 struct cgroup_subsys *ss;
1020 for_each_subsys(ss, i) {
1021 struct cftype *cfts;
1023 if (!test_bit(i, &subsys_mask))
1025 list_for_each_entry(cfts, &ss->cfts, node)
1026 cgroup_addrm_files(cgrp, cfts, false);
1030 static int rebind_subsystems(struct cgroupfs_root *root,
1031 unsigned long added_mask, unsigned removed_mask)
1033 struct cgroup *cgrp = &root->top_cgroup;
1034 struct cgroup_subsys *ss;
1037 lockdep_assert_held(&cgroup_tree_mutex);
1038 lockdep_assert_held(&cgroup_mutex);
1040 /* Check that any added subsystems are currently free */
1041 for_each_subsys(ss, i)
1042 if ((added_mask & (1 << i)) && ss->root != &cgroup_dummy_root)
1045 ret = cgroup_populate_dir(cgrp, added_mask);
1050 * Nothing can fail from this point on. Remove files for the
1051 * removed subsystems and rebind each subsystem.
1053 mutex_unlock(&cgroup_mutex);
1054 cgroup_clear_dir(cgrp, removed_mask);
1055 mutex_lock(&cgroup_mutex);
1057 for_each_subsys(ss, i) {
1058 unsigned long bit = 1UL << i;
1060 if (bit & added_mask) {
1061 /* We're binding this subsystem to this hierarchy */
1062 BUG_ON(cgroup_css(cgrp, ss));
1063 BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
1064 BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
1066 rcu_assign_pointer(cgrp->subsys[i],
1067 cgroup_css(cgroup_dummy_top, ss));
1068 cgroup_css(cgrp, ss)->cgroup = cgrp;
1072 ss->bind(cgroup_css(cgrp, ss));
1074 /* refcount was already taken, and we're keeping it */
1075 root->subsys_mask |= bit;
1076 } else if (bit & removed_mask) {
1077 /* We're removing this subsystem */
1078 BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1079 BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1082 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1084 cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1085 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1087 cgroup_subsys[i]->root = &cgroup_dummy_root;
1088 root->subsys_mask &= ~bit;
1093 * Mark @root has finished binding subsystems. @root->subsys_mask
1094 * now matches the bound subsystems.
1096 root->flags |= CGRP_ROOT_SUBSYS_BOUND;
1097 kernfs_activate(cgrp->kn);
1102 static int cgroup_show_options(struct seq_file *seq,
1103 struct kernfs_root *kf_root)
1105 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1106 struct cgroup_subsys *ss;
1109 for_each_subsys(ss, ssid)
1110 if (root->subsys_mask & (1 << ssid))
1111 seq_printf(seq, ",%s", ss->name);
1112 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1113 seq_puts(seq, ",sane_behavior");
1114 if (root->flags & CGRP_ROOT_NOPREFIX)
1115 seq_puts(seq, ",noprefix");
1116 if (root->flags & CGRP_ROOT_XATTR)
1117 seq_puts(seq, ",xattr");
1119 spin_lock(&release_agent_path_lock);
1120 if (strlen(root->release_agent_path))
1121 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1122 spin_unlock(&release_agent_path_lock);
1124 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1125 seq_puts(seq, ",clone_children");
1126 if (strlen(root->name))
1127 seq_printf(seq, ",name=%s", root->name);
1131 struct cgroup_sb_opts {
1132 unsigned long subsys_mask;
1133 unsigned long flags;
1134 char *release_agent;
1135 bool cpuset_clone_children;
1137 /* User explicitly requested empty subsystem */
1142 * Convert a hierarchy specifier into a bitmask of subsystems and
1143 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1144 * array. This function takes refcounts on subsystems to be used, unless it
1145 * returns error, in which case no refcounts are taken.
1147 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1149 char *token, *o = data;
1150 bool all_ss = false, one_ss = false;
1151 unsigned long mask = (unsigned long)-1;
1152 struct cgroup_subsys *ss;
1155 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1157 #ifdef CONFIG_CPUSETS
1158 mask = ~(1UL << cpuset_cgrp_id);
1161 memset(opts, 0, sizeof(*opts));
1163 while ((token = strsep(&o, ",")) != NULL) {
1166 if (!strcmp(token, "none")) {
1167 /* Explicitly have no subsystems */
1171 if (!strcmp(token, "all")) {
1172 /* Mutually exclusive option 'all' + subsystem name */
1178 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1179 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1182 if (!strcmp(token, "noprefix")) {
1183 opts->flags |= CGRP_ROOT_NOPREFIX;
1186 if (!strcmp(token, "clone_children")) {
1187 opts->cpuset_clone_children = true;
1190 if (!strcmp(token, "xattr")) {
1191 opts->flags |= CGRP_ROOT_XATTR;
1194 if (!strncmp(token, "release_agent=", 14)) {
1195 /* Specifying two release agents is forbidden */
1196 if (opts->release_agent)
1198 opts->release_agent =
1199 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1200 if (!opts->release_agent)
1204 if (!strncmp(token, "name=", 5)) {
1205 const char *name = token + 5;
1206 /* Can't specify an empty name */
1209 /* Must match [\w.-]+ */
1210 for (i = 0; i < strlen(name); i++) {
1214 if ((c == '.') || (c == '-') || (c == '_'))
1218 /* Specifying two names is forbidden */
1221 opts->name = kstrndup(name,
1222 MAX_CGROUP_ROOT_NAMELEN - 1,
1230 for_each_subsys(ss, i) {
1231 if (strcmp(token, ss->name))
1236 /* Mutually exclusive option 'all' + subsystem name */
1239 set_bit(i, &opts->subsys_mask);
1244 if (i == CGROUP_SUBSYS_COUNT)
1249 * If the 'all' option was specified select all the subsystems,
1250 * otherwise if 'none', 'name=' and a subsystem name options
1251 * were not specified, let's default to 'all'
1253 if (all_ss || (!one_ss && !opts->none && !opts->name))
1254 for_each_subsys(ss, i)
1256 set_bit(i, &opts->subsys_mask);
1258 /* Consistency checks */
1260 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1261 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1263 if (opts->flags & CGRP_ROOT_NOPREFIX) {
1264 pr_err("cgroup: sane_behavior: noprefix is not allowed\n");
1268 if (opts->cpuset_clone_children) {
1269 pr_err("cgroup: sane_behavior: clone_children is not allowed\n");
1273 if (opts->flags & CGRP_ROOT_XATTR)
1274 pr_warning("cgroup: sane_behavior: xattr is always available, flag unnecessary\n");
1278 * Option noprefix was introduced just for backward compatibility
1279 * with the old cpuset, so we allow noprefix only if mounting just
1280 * the cpuset subsystem.
1282 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1286 /* Can't specify "none" and some subsystems */
1287 if (opts->subsys_mask && opts->none)
1291 * We either have to specify by name or by subsystems. (So all
1292 * empty hierarchies must have a name).
1294 if (!opts->subsys_mask && !opts->name)
1300 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1303 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1304 struct cgroup_sb_opts opts;
1305 unsigned long added_mask, removed_mask;
1307 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1308 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1312 mutex_lock(&cgroup_tree_mutex);
1313 mutex_lock(&cgroup_mutex);
1315 /* See what subsystems are wanted */
1316 ret = parse_cgroupfs_options(data, &opts);
1320 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1321 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1322 task_tgid_nr(current), current->comm);
1324 added_mask = opts.subsys_mask & ~root->subsys_mask;
1325 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1327 /* Don't allow flags or name to change at remount */
1328 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1329 (opts.name && strcmp(opts.name, root->name))) {
1330 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1331 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1332 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1337 /* remounting is not allowed for populated hierarchies */
1338 if (root->number_of_cgroups > 1) {
1343 ret = rebind_subsystems(root, added_mask, removed_mask);
1347 if (opts.release_agent) {
1348 spin_lock(&release_agent_path_lock);
1349 strcpy(root->release_agent_path, opts.release_agent);
1350 spin_unlock(&release_agent_path_lock);
1353 kfree(opts.release_agent);
1355 mutex_unlock(&cgroup_mutex);
1356 mutex_unlock(&cgroup_tree_mutex);
1360 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1362 atomic_set(&cgrp->refcnt, 1);
1363 INIT_LIST_HEAD(&cgrp->sibling);
1364 INIT_LIST_HEAD(&cgrp->children);
1365 INIT_LIST_HEAD(&cgrp->cset_links);
1366 INIT_LIST_HEAD(&cgrp->release_list);
1367 INIT_LIST_HEAD(&cgrp->pidlists);
1368 mutex_init(&cgrp->pidlist_mutex);
1369 cgrp->dummy_css.cgroup = cgrp;
1372 static void init_cgroup_root(struct cgroupfs_root *root)
1374 struct cgroup *cgrp = &root->top_cgroup;
1376 atomic_set(&root->refcnt, 1);
1377 INIT_LIST_HEAD(&root->root_list);
1378 root->number_of_cgroups = 1;
1380 RCU_INIT_POINTER(cgrp->name, &root_cgroup_name);
1381 init_cgroup_housekeeping(cgrp);
1382 idr_init(&root->cgroup_idr);
1385 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1387 struct cgroupfs_root *root;
1389 if (!opts->subsys_mask && !opts->none)
1390 return ERR_PTR(-EINVAL);
1392 root = kzalloc(sizeof(*root), GFP_KERNEL);
1394 return ERR_PTR(-ENOMEM);
1396 init_cgroup_root(root);
1399 * We need to set @root->subsys_mask now so that @root can be
1400 * matched by cgroup_test_super() before it finishes
1401 * initialization; otherwise, competing mounts with the same
1402 * options may try to bind the same subsystems instead of waiting
1403 * for the first one leading to unexpected mount errors.
1404 * SUBSYS_BOUND will be set once actual binding is complete.
1406 root->subsys_mask = opts->subsys_mask;
1407 root->flags = opts->flags;
1408 if (opts->release_agent)
1409 strcpy(root->release_agent_path, opts->release_agent);
1411 strcpy(root->name, opts->name);
1412 if (opts->cpuset_clone_children)
1413 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1417 static int cgroup_setup_root(struct cgroupfs_root *root)
1419 LIST_HEAD(tmp_links);
1420 struct cgroup *root_cgrp = &root->top_cgroup;
1421 struct css_set *cset;
1424 lockdep_assert_held(&cgroup_tree_mutex);
1425 lockdep_assert_held(&cgroup_mutex);
1427 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1430 root_cgrp->id = ret;
1433 * We're accessing css_set_count without locking css_set_lock here,
1434 * but that's OK - it can only be increased by someone holding
1435 * cgroup_lock, and that's us. The worst that can happen is that we
1436 * have some link structures left over
1438 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1442 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1443 ret = cgroup_init_root_id(root, 2, 0);
1447 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1448 KERNFS_ROOT_CREATE_DEACTIVATED,
1450 if (IS_ERR(root->kf_root)) {
1451 ret = PTR_ERR(root->kf_root);
1454 root_cgrp->kn = root->kf_root->kn;
1456 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1460 ret = rebind_subsystems(root, root->subsys_mask, 0);
1465 * There must be no failure case after here, since rebinding takes
1466 * care of subsystems' refcounts, which are explicitly dropped in
1467 * the failure exit path.
1469 list_add(&root->root_list, &cgroup_roots);
1470 cgroup_root_count++;
1473 * Link the top cgroup in this hierarchy into all the css_set
1476 write_lock(&css_set_lock);
1477 hash_for_each(css_set_table, i, cset, hlist)
1478 link_css_set(&tmp_links, cset, root_cgrp);
1479 write_unlock(&css_set_lock);
1481 BUG_ON(!list_empty(&root_cgrp->children));
1482 BUG_ON(root->number_of_cgroups != 1);
1484 kernfs_activate(root_cgrp->kn);
1489 kernfs_destroy_root(root->kf_root);
1490 root->kf_root = NULL;
1492 cgroup_exit_root_id(root);
1494 free_cgrp_cset_links(&tmp_links);
1498 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1499 int flags, const char *unused_dev_name,
1502 struct cgroupfs_root *root;
1503 struct cgroup_sb_opts opts;
1504 struct dentry *dentry;
1507 mutex_lock(&cgroup_tree_mutex);
1508 mutex_lock(&cgroup_mutex);
1510 /* First find the desired set of subsystems */
1511 ret = parse_cgroupfs_options(data, &opts);
1515 /* look for a matching existing root */
1516 for_each_active_root(root) {
1517 bool name_match = false;
1520 * If we asked for a name then it must match. Also, if
1521 * name matches but sybsys_mask doesn't, we should fail.
1522 * Remember whether name matched.
1525 if (strcmp(opts.name, root->name))
1531 * If we asked for subsystems (or explicitly for no
1532 * subsystems) then they must match.
1534 if ((opts.subsys_mask || opts.none) &&
1535 (opts.subsys_mask != root->subsys_mask)) {
1542 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1543 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1544 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1548 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1552 cgroup_get_root(root);
1556 /* no such thing, create a new one */
1557 root = cgroup_root_from_opts(&opts);
1559 ret = PTR_ERR(root);
1563 ret = cgroup_setup_root(root);
1565 cgroup_free_root(root);
1568 mutex_unlock(&cgroup_mutex);
1569 mutex_unlock(&cgroup_tree_mutex);
1571 kfree(opts.release_agent);
1575 return ERR_PTR(ret);
1577 dentry = kernfs_mount(fs_type, flags, root->kf_root);
1579 cgroup_put_root(root);
1583 static void cgroup_kill_sb(struct super_block *sb)
1585 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1586 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1588 cgroup_put_root(root);
1592 static struct file_system_type cgroup_fs_type = {
1594 .mount = cgroup_mount,
1595 .kill_sb = cgroup_kill_sb,
1598 static struct kobject *cgroup_kobj;
1601 * cgroup_path - generate the path of a cgroup
1602 * @cgrp: the cgroup in question
1603 * @buf: the buffer to write the path into
1604 * @buflen: the length of the buffer
1606 * Writes path of cgroup into buf. Returns 0 on success, -errno on error.
1608 * We can't generate cgroup path using dentry->d_name, as accessing
1609 * dentry->name must be protected by irq-unsafe dentry->d_lock or parent
1610 * inode's i_mutex, while on the other hand cgroup_path() can be called
1611 * with some irq-safe spinlocks held.
1613 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1615 int ret = -ENAMETOOLONG;
1618 if (!cgrp->parent) {
1619 if (strlcpy(buf, "/", buflen) >= buflen)
1620 return -ENAMETOOLONG;
1624 start = buf + buflen - 1;
1629 const char *name = cgroup_name(cgrp);
1633 if ((start -= len) < buf)
1635 memcpy(start, name, len);
1641 cgrp = cgrp->parent;
1642 } while (cgrp->parent);
1644 memmove(buf, start, buf + buflen - start);
1649 EXPORT_SYMBOL_GPL(cgroup_path);
1652 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1653 * @task: target task
1654 * @buf: the buffer to write the path into
1655 * @buflen: the length of the buffer
1657 * Determine @task's cgroup on the first (the one with the lowest non-zero
1658 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1659 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1660 * cgroup controller callbacks.
1662 * Returns 0 on success, fails with -%ENAMETOOLONG if @buflen is too short.
1664 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1666 struct cgroupfs_root *root;
1667 struct cgroup *cgrp;
1668 int hierarchy_id = 1, ret = 0;
1671 return -ENAMETOOLONG;
1673 mutex_lock(&cgroup_mutex);
1675 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1678 cgrp = task_cgroup_from_root(task, root);
1679 ret = cgroup_path(cgrp, buf, buflen);
1681 /* if no hierarchy exists, everyone is in "/" */
1682 memcpy(buf, "/", 2);
1685 mutex_unlock(&cgroup_mutex);
1688 EXPORT_SYMBOL_GPL(task_cgroup_path);
1691 * Control Group taskset
1693 struct task_and_cgroup {
1694 struct task_struct *task;
1695 struct cgroup *cgrp;
1696 struct css_set *cset;
1699 struct cgroup_taskset {
1700 struct task_and_cgroup single;
1701 struct flex_array *tc_array;
1704 struct cgroup *cur_cgrp;
1708 * cgroup_taskset_first - reset taskset and return the first task
1709 * @tset: taskset of interest
1711 * @tset iteration is initialized and the first task is returned.
1713 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1715 if (tset->tc_array) {
1717 return cgroup_taskset_next(tset);
1719 tset->cur_cgrp = tset->single.cgrp;
1720 return tset->single.task;
1723 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1726 * cgroup_taskset_next - iterate to the next task in taskset
1727 * @tset: taskset of interest
1729 * Return the next task in @tset. Iteration must have been initialized
1730 * with cgroup_taskset_first().
1732 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1734 struct task_and_cgroup *tc;
1736 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1739 tc = flex_array_get(tset->tc_array, tset->idx++);
1740 tset->cur_cgrp = tc->cgrp;
1743 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1746 * cgroup_taskset_cur_css - return the matching css for the current task
1747 * @tset: taskset of interest
1748 * @subsys_id: the ID of the target subsystem
1750 * Return the css for the current (last returned) task of @tset for
1751 * subsystem specified by @subsys_id. This function must be preceded by
1752 * either cgroup_taskset_first() or cgroup_taskset_next().
1754 struct cgroup_subsys_state *cgroup_taskset_cur_css(struct cgroup_taskset *tset,
1757 return cgroup_css(tset->cur_cgrp, cgroup_subsys[subsys_id]);
1759 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_css);
1762 * cgroup_taskset_size - return the number of tasks in taskset
1763 * @tset: taskset of interest
1765 int cgroup_taskset_size(struct cgroup_taskset *tset)
1767 return tset->tc_array ? tset->tc_array_len : 1;
1769 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1773 * cgroup_task_migrate - move a task from one cgroup to another.
1775 * Must be called with cgroup_mutex and threadgroup locked.
1777 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1778 struct task_struct *tsk,
1779 struct css_set *new_cset)
1781 struct css_set *old_cset;
1784 * We are synchronized through threadgroup_lock() against PF_EXITING
1785 * setting such that we can't race against cgroup_exit() changing the
1786 * css_set to init_css_set and dropping the old one.
1788 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1789 old_cset = task_css_set(tsk);
1792 rcu_assign_pointer(tsk->cgroups, new_cset);
1795 /* Update the css_set linked lists if we're using them */
1796 write_lock(&css_set_lock);
1797 if (!list_empty(&tsk->cg_list))
1798 list_move(&tsk->cg_list, &new_cset->tasks);
1799 write_unlock(&css_set_lock);
1802 * We just gained a reference on old_cset by taking it from the
1803 * task. As trading it for new_cset is protected by cgroup_mutex,
1804 * we're safe to drop it here; it will be freed under RCU.
1806 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1807 put_css_set(old_cset);
1811 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1812 * @cgrp: the cgroup to attach to
1813 * @tsk: the task or the leader of the threadgroup to be attached
1814 * @threadgroup: attach the whole threadgroup?
1816 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1817 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1819 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1822 int retval, i, group_size;
1823 struct cgroupfs_root *root = cgrp->root;
1824 struct cgroup_subsys_state *css, *failed_css = NULL;
1825 /* threadgroup list cursor and array */
1826 struct task_struct *leader = tsk;
1827 struct task_and_cgroup *tc;
1828 struct flex_array *group;
1829 struct cgroup_taskset tset = { };
1832 * step 0: in order to do expensive, possibly blocking operations for
1833 * every thread, we cannot iterate the thread group list, since it needs
1834 * rcu or tasklist locked. instead, build an array of all threads in the
1835 * group - group_rwsem prevents new threads from appearing, and if
1836 * threads exit, this will just be an over-estimate.
1839 group_size = get_nr_threads(tsk);
1842 /* flex_array supports very large thread-groups better than kmalloc. */
1843 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1846 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1847 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1849 goto out_free_group_list;
1853 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1854 * already PF_EXITING could be freed from underneath us unless we
1855 * take an rcu_read_lock.
1859 struct task_and_cgroup ent;
1861 /* @tsk either already exited or can't exit until the end */
1862 if (tsk->flags & PF_EXITING)
1865 /* as per above, nr_threads may decrease, but not increase. */
1866 BUG_ON(i >= group_size);
1868 ent.cgrp = task_cgroup_from_root(tsk, root);
1869 /* nothing to do if this task is already in the cgroup */
1870 if (ent.cgrp == cgrp)
1873 * saying GFP_ATOMIC has no effect here because we did prealloc
1874 * earlier, but it's good form to communicate our expectations.
1876 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
1877 BUG_ON(retval != 0);
1882 } while_each_thread(leader, tsk);
1884 /* remember the number of threads in the array for later. */
1886 tset.tc_array = group;
1887 tset.tc_array_len = group_size;
1889 /* methods shouldn't be called if no task is actually migrating */
1892 goto out_free_group_list;
1895 * step 1: check that we can legitimately attach to the cgroup.
1897 for_each_css(css, i, cgrp) {
1898 if (css->ss->can_attach) {
1899 retval = css->ss->can_attach(css, &tset);
1902 goto out_cancel_attach;
1908 * step 2: make sure css_sets exist for all threads to be migrated.
1909 * we use find_css_set, which allocates a new one if necessary.
1911 for (i = 0; i < group_size; i++) {
1912 struct css_set *old_cset;
1914 tc = flex_array_get(group, i);
1915 old_cset = task_css_set(tc->task);
1916 tc->cset = find_css_set(old_cset, cgrp);
1919 goto out_put_css_set_refs;
1924 * step 3: now that we're guaranteed success wrt the css_sets,
1925 * proceed to move all tasks to the new cgroup. There are no
1926 * failure cases after here, so this is the commit point.
1928 for (i = 0; i < group_size; i++) {
1929 tc = flex_array_get(group, i);
1930 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
1932 /* nothing is sensitive to fork() after this point. */
1935 * step 4: do subsystem attach callbacks.
1937 for_each_css(css, i, cgrp)
1938 if (css->ss->attach)
1939 css->ss->attach(css, &tset);
1942 * step 5: success! and cleanup
1945 out_put_css_set_refs:
1947 for (i = 0; i < group_size; i++) {
1948 tc = flex_array_get(group, i);
1951 put_css_set(tc->cset);
1956 for_each_css(css, i, cgrp) {
1957 if (css == failed_css)
1959 if (css->ss->cancel_attach)
1960 css->ss->cancel_attach(css, &tset);
1963 out_free_group_list:
1964 flex_array_free(group);
1969 * Find the task_struct of the task to attach by vpid and pass it along to the
1970 * function to attach either it or all tasks in its threadgroup. Will lock
1971 * cgroup_mutex and threadgroup; may take task_lock of task.
1973 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
1975 struct task_struct *tsk;
1976 const struct cred *cred = current_cred(), *tcred;
1979 if (!cgroup_lock_live_group(cgrp))
1985 tsk = find_task_by_vpid(pid);
1989 goto out_unlock_cgroup;
1992 * even if we're attaching all tasks in the thread group, we
1993 * only need to check permissions on one of them.
1995 tcred = __task_cred(tsk);
1996 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
1997 !uid_eq(cred->euid, tcred->uid) &&
1998 !uid_eq(cred->euid, tcred->suid)) {
2001 goto out_unlock_cgroup;
2007 tsk = tsk->group_leader;
2010 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2011 * trapped in a cpuset, or RT worker may be born in a cgroup
2012 * with no rt_runtime allocated. Just say no.
2014 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2017 goto out_unlock_cgroup;
2020 get_task_struct(tsk);
2023 threadgroup_lock(tsk);
2025 if (!thread_group_leader(tsk)) {
2027 * a race with de_thread from another thread's exec()
2028 * may strip us of our leadership, if this happens,
2029 * there is no choice but to throw this task away and
2030 * try again; this is
2031 * "double-double-toil-and-trouble-check locking".
2033 threadgroup_unlock(tsk);
2034 put_task_struct(tsk);
2035 goto retry_find_task;
2039 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2041 threadgroup_unlock(tsk);
2043 put_task_struct(tsk);
2045 mutex_unlock(&cgroup_mutex);
2050 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2051 * @from: attach to all cgroups of a given task
2052 * @tsk: the task to be attached
2054 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2056 struct cgroupfs_root *root;
2059 mutex_lock(&cgroup_mutex);
2060 for_each_active_root(root) {
2061 struct cgroup *from_cgrp = task_cgroup_from_root(from, root);
2063 retval = cgroup_attach_task(from_cgrp, tsk, false);
2067 mutex_unlock(&cgroup_mutex);
2071 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2073 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2074 struct cftype *cft, u64 pid)
2076 return attach_task_by_pid(css->cgroup, pid, false);
2079 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2080 struct cftype *cft, u64 tgid)
2082 return attach_task_by_pid(css->cgroup, tgid, true);
2085 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2086 struct cftype *cft, const char *buffer)
2088 struct cgroupfs_root *root = css->cgroup->root;
2090 BUILD_BUG_ON(sizeof(root->release_agent_path) < PATH_MAX);
2091 if (!cgroup_lock_live_group(css->cgroup))
2093 spin_lock(&release_agent_path_lock);
2094 strlcpy(root->release_agent_path, buffer,
2095 sizeof(root->release_agent_path));
2096 spin_unlock(&release_agent_path_lock);
2097 mutex_unlock(&cgroup_mutex);
2101 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2103 struct cgroup *cgrp = seq_css(seq)->cgroup;
2105 if (!cgroup_lock_live_group(cgrp))
2107 seq_puts(seq, cgrp->root->release_agent_path);
2108 seq_putc(seq, '\n');
2109 mutex_unlock(&cgroup_mutex);
2113 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2115 struct cgroup *cgrp = seq_css(seq)->cgroup;
2117 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2121 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2122 size_t nbytes, loff_t off)
2124 struct cgroup *cgrp = of->kn->parent->priv;
2125 struct cftype *cft = of->kn->priv;
2126 struct cgroup_subsys_state *css;
2130 * kernfs guarantees that a file isn't deleted with operations in
2131 * flight, which means that the matching css is and stays alive and
2132 * doesn't need to be pinned. The RCU locking is not necessary
2133 * either. It's just for the convenience of using cgroup_css().
2136 css = cgroup_css(cgrp, cft->ss);
2139 if (cft->write_string) {
2140 ret = cft->write_string(css, cft, strstrip(buf));
2141 } else if (cft->write_u64) {
2142 unsigned long long v;
2143 ret = kstrtoull(buf, 0, &v);
2145 ret = cft->write_u64(css, cft, v);
2146 } else if (cft->write_s64) {
2148 ret = kstrtoll(buf, 0, &v);
2150 ret = cft->write_s64(css, cft, v);
2151 } else if (cft->trigger) {
2152 ret = cft->trigger(css, (unsigned int)cft->private);
2157 return ret ?: nbytes;
2160 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2162 return seq_cft(seq)->seq_start(seq, ppos);
2165 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2167 return seq_cft(seq)->seq_next(seq, v, ppos);
2170 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2172 seq_cft(seq)->seq_stop(seq, v);
2175 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2177 struct cftype *cft = seq_cft(m);
2178 struct cgroup_subsys_state *css = seq_css(m);
2181 return cft->seq_show(m, arg);
2184 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2185 else if (cft->read_s64)
2186 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2192 static struct kernfs_ops cgroup_kf_single_ops = {
2193 .atomic_write_len = PAGE_SIZE,
2194 .write = cgroup_file_write,
2195 .seq_show = cgroup_seqfile_show,
2198 static struct kernfs_ops cgroup_kf_ops = {
2199 .atomic_write_len = PAGE_SIZE,
2200 .write = cgroup_file_write,
2201 .seq_start = cgroup_seqfile_start,
2202 .seq_next = cgroup_seqfile_next,
2203 .seq_stop = cgroup_seqfile_stop,
2204 .seq_show = cgroup_seqfile_show,
2208 * cgroup_rename - Only allow simple rename of directories in place.
2210 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2211 const char *new_name_str)
2213 struct cgroup *cgrp = kn->priv;
2214 struct cgroup_name *name, *old_name;
2217 if (kernfs_type(kn) != KERNFS_DIR)
2219 if (kn->parent != new_parent)
2223 * This isn't a proper migration and its usefulness is very
2224 * limited. Disallow if sane_behavior.
2226 if (cgroup_sane_behavior(cgrp))
2229 name = cgroup_alloc_name(new_name_str);
2233 mutex_lock(&cgroup_tree_mutex);
2234 mutex_lock(&cgroup_mutex);
2236 ret = kernfs_rename(kn, new_parent, new_name_str);
2238 old_name = rcu_dereference_protected(cgrp->name, true);
2239 rcu_assign_pointer(cgrp->name, name);
2244 mutex_unlock(&cgroup_mutex);
2245 mutex_unlock(&cgroup_tree_mutex);
2247 kfree_rcu(old_name, rcu_head);
2251 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2253 char name[CGROUP_FILE_NAME_MAX];
2254 struct kernfs_node *kn;
2255 struct lock_class_key *key = NULL;
2257 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2258 key = &cft->lockdep_key;
2260 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2261 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2269 * cgroup_addrm_files - add or remove files to a cgroup directory
2270 * @cgrp: the target cgroup
2271 * @cfts: array of cftypes to be added
2272 * @is_add: whether to add or remove
2274 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2275 * For removals, this function never fails. If addition fails, this
2276 * function doesn't remove files already added. The caller is responsible
2279 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2285 lockdep_assert_held(&cgroup_tree_mutex);
2287 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2288 /* does cft->flags tell us to skip this file on @cgrp? */
2289 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2291 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2293 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2297 ret = cgroup_add_file(cgrp, cft);
2299 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2304 cgroup_rm_file(cgrp, cft);
2310 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2313 struct cgroup_subsys *ss = cfts[0].ss;
2314 struct cgroup *root = &ss->root->top_cgroup;
2315 struct cgroup_subsys_state *css;
2318 lockdep_assert_held(&cgroup_tree_mutex);
2320 /* don't bother if @ss isn't attached */
2321 if (ss->root == &cgroup_dummy_root)
2324 /* add/rm files for all cgroups created before */
2325 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2326 struct cgroup *cgrp = css->cgroup;
2328 if (cgroup_is_dead(cgrp))
2331 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2337 kernfs_activate(root->kn);
2341 static void cgroup_exit_cftypes(struct cftype *cfts)
2345 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2346 /* free copy for custom atomic_write_len, see init_cftypes() */
2347 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2354 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2358 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2359 struct kernfs_ops *kf_ops;
2361 WARN_ON(cft->ss || cft->kf_ops);
2364 kf_ops = &cgroup_kf_ops;
2366 kf_ops = &cgroup_kf_single_ops;
2369 * Ugh... if @cft wants a custom max_write_len, we need to
2370 * make a copy of kf_ops to set its atomic_write_len.
2372 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
2373 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
2375 cgroup_exit_cftypes(cfts);
2378 kf_ops->atomic_write_len = cft->max_write_len;
2381 cft->kf_ops = kf_ops;
2388 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
2390 lockdep_assert_held(&cgroup_tree_mutex);
2392 if (!cfts || !cfts[0].ss)
2395 list_del(&cfts->node);
2396 cgroup_apply_cftypes(cfts, false);
2397 cgroup_exit_cftypes(cfts);
2402 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2403 * @cfts: zero-length name terminated array of cftypes
2405 * Unregister @cfts. Files described by @cfts are removed from all
2406 * existing cgroups and all future cgroups won't have them either. This
2407 * function can be called anytime whether @cfts' subsys is attached or not.
2409 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2412 int cgroup_rm_cftypes(struct cftype *cfts)
2416 mutex_lock(&cgroup_tree_mutex);
2417 ret = cgroup_rm_cftypes_locked(cfts);
2418 mutex_unlock(&cgroup_tree_mutex);
2423 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2424 * @ss: target cgroup subsystem
2425 * @cfts: zero-length name terminated array of cftypes
2427 * Register @cfts to @ss. Files described by @cfts are created for all
2428 * existing cgroups to which @ss is attached and all future cgroups will
2429 * have them too. This function can be called anytime whether @ss is
2432 * Returns 0 on successful registration, -errno on failure. Note that this
2433 * function currently returns 0 as long as @cfts registration is successful
2434 * even if some file creation attempts on existing cgroups fail.
2436 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2440 ret = cgroup_init_cftypes(ss, cfts);
2444 mutex_lock(&cgroup_tree_mutex);
2446 list_add_tail(&cfts->node, &ss->cfts);
2447 ret = cgroup_apply_cftypes(cfts, true);
2449 cgroup_rm_cftypes_locked(cfts);
2451 mutex_unlock(&cgroup_tree_mutex);
2454 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2457 * cgroup_task_count - count the number of tasks in a cgroup.
2458 * @cgrp: the cgroup in question
2460 * Return the number of tasks in the cgroup.
2462 int cgroup_task_count(const struct cgroup *cgrp)
2465 struct cgrp_cset_link *link;
2467 read_lock(&css_set_lock);
2468 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2469 count += atomic_read(&link->cset->refcount);
2470 read_unlock(&css_set_lock);
2475 * To reduce the fork() overhead for systems that are not actually using
2476 * their cgroups capability, we don't maintain the lists running through
2477 * each css_set to its tasks until we see the list actually used - in other
2478 * words after the first call to css_task_iter_start().
2480 static void cgroup_enable_task_cg_lists(void)
2482 struct task_struct *p, *g;
2483 write_lock(&css_set_lock);
2484 use_task_css_set_links = 1;
2486 * We need tasklist_lock because RCU is not safe against
2487 * while_each_thread(). Besides, a forking task that has passed
2488 * cgroup_post_fork() without seeing use_task_css_set_links = 1
2489 * is not guaranteed to have its child immediately visible in the
2490 * tasklist if we walk through it with RCU.
2492 read_lock(&tasklist_lock);
2493 do_each_thread(g, p) {
2496 * We should check if the process is exiting, otherwise
2497 * it will race with cgroup_exit() in that the list
2498 * entry won't be deleted though the process has exited.
2500 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2501 list_add(&p->cg_list, &task_css_set(p)->tasks);
2503 } while_each_thread(g, p);
2504 read_unlock(&tasklist_lock);
2505 write_unlock(&css_set_lock);
2509 * css_next_child - find the next child of a given css
2510 * @pos_css: the current position (%NULL to initiate traversal)
2511 * @parent_css: css whose children to walk
2513 * This function returns the next child of @parent_css and should be called
2514 * under either cgroup_mutex or RCU read lock. The only requirement is
2515 * that @parent_css and @pos_css are accessible. The next sibling is
2516 * guaranteed to be returned regardless of their states.
2518 struct cgroup_subsys_state *
2519 css_next_child(struct cgroup_subsys_state *pos_css,
2520 struct cgroup_subsys_state *parent_css)
2522 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2523 struct cgroup *cgrp = parent_css->cgroup;
2524 struct cgroup *next;
2526 cgroup_assert_mutexes_or_rcu_locked();
2529 * @pos could already have been removed. Once a cgroup is removed,
2530 * its ->sibling.next is no longer updated when its next sibling
2531 * changes. As CGRP_DEAD assertion is serialized and happens
2532 * before the cgroup is taken off the ->sibling list, if we see it
2533 * unasserted, it's guaranteed that the next sibling hasn't
2534 * finished its grace period even if it's already removed, and thus
2535 * safe to dereference from this RCU critical section. If
2536 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2537 * to be visible as %true here.
2539 * If @pos is dead, its next pointer can't be dereferenced;
2540 * however, as each cgroup is given a monotonically increasing
2541 * unique serial number and always appended to the sibling list,
2542 * the next one can be found by walking the parent's children until
2543 * we see a cgroup with higher serial number than @pos's. While
2544 * this path can be slower, it's taken only when either the current
2545 * cgroup is removed or iteration and removal race.
2548 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2549 } else if (likely(!cgroup_is_dead(pos))) {
2550 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2552 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2553 if (next->serial_nr > pos->serial_nr)
2557 if (&next->sibling == &cgrp->children)
2560 return cgroup_css(next, parent_css->ss);
2562 EXPORT_SYMBOL_GPL(css_next_child);
2565 * css_next_descendant_pre - find the next descendant for pre-order walk
2566 * @pos: the current position (%NULL to initiate traversal)
2567 * @root: css whose descendants to walk
2569 * To be used by css_for_each_descendant_pre(). Find the next descendant
2570 * to visit for pre-order traversal of @root's descendants. @root is
2571 * included in the iteration and the first node to be visited.
2573 * While this function requires cgroup_mutex or RCU read locking, it
2574 * doesn't require the whole traversal to be contained in a single critical
2575 * section. This function will return the correct next descendant as long
2576 * as both @pos and @root are accessible and @pos is a descendant of @root.
2578 struct cgroup_subsys_state *
2579 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2580 struct cgroup_subsys_state *root)
2582 struct cgroup_subsys_state *next;
2584 cgroup_assert_mutexes_or_rcu_locked();
2586 /* if first iteration, visit @root */
2590 /* visit the first child if exists */
2591 next = css_next_child(NULL, pos);
2595 /* no child, visit my or the closest ancestor's next sibling */
2596 while (pos != root) {
2597 next = css_next_child(pos, css_parent(pos));
2600 pos = css_parent(pos);
2605 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
2608 * css_rightmost_descendant - return the rightmost descendant of a css
2609 * @pos: css of interest
2611 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2612 * is returned. This can be used during pre-order traversal to skip
2615 * While this function requires cgroup_mutex or RCU read locking, it
2616 * doesn't require the whole traversal to be contained in a single critical
2617 * section. This function will return the correct rightmost descendant as
2618 * long as @pos is accessible.
2620 struct cgroup_subsys_state *
2621 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2623 struct cgroup_subsys_state *last, *tmp;
2625 cgroup_assert_mutexes_or_rcu_locked();
2629 /* ->prev isn't RCU safe, walk ->next till the end */
2631 css_for_each_child(tmp, last)
2637 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
2639 static struct cgroup_subsys_state *
2640 css_leftmost_descendant(struct cgroup_subsys_state *pos)
2642 struct cgroup_subsys_state *last;
2646 pos = css_next_child(NULL, pos);
2653 * css_next_descendant_post - find the next descendant for post-order walk
2654 * @pos: the current position (%NULL to initiate traversal)
2655 * @root: css whose descendants to walk
2657 * To be used by css_for_each_descendant_post(). Find the next descendant
2658 * to visit for post-order traversal of @root's descendants. @root is
2659 * included in the iteration and the last node to be visited.
2661 * While this function requires cgroup_mutex or RCU read locking, it
2662 * doesn't require the whole traversal to be contained in a single critical
2663 * section. This function will return the correct next descendant as long
2664 * as both @pos and @cgroup are accessible and @pos is a descendant of
2667 struct cgroup_subsys_state *
2668 css_next_descendant_post(struct cgroup_subsys_state *pos,
2669 struct cgroup_subsys_state *root)
2671 struct cgroup_subsys_state *next;
2673 cgroup_assert_mutexes_or_rcu_locked();
2675 /* if first iteration, visit leftmost descendant which may be @root */
2677 return css_leftmost_descendant(root);
2679 /* if we visited @root, we're done */
2683 /* if there's an unvisited sibling, visit its leftmost descendant */
2684 next = css_next_child(pos, css_parent(pos));
2686 return css_leftmost_descendant(next);
2688 /* no sibling left, visit parent */
2689 return css_parent(pos);
2691 EXPORT_SYMBOL_GPL(css_next_descendant_post);
2694 * css_advance_task_iter - advance a task itererator to the next css_set
2695 * @it: the iterator to advance
2697 * Advance @it to the next css_set to walk.
2699 static void css_advance_task_iter(struct css_task_iter *it)
2701 struct list_head *l = it->cset_link;
2702 struct cgrp_cset_link *link;
2703 struct css_set *cset;
2705 /* Advance to the next non-empty css_set */
2708 if (l == &it->origin_css->cgroup->cset_links) {
2709 it->cset_link = NULL;
2712 link = list_entry(l, struct cgrp_cset_link, cset_link);
2714 } while (list_empty(&cset->tasks));
2716 it->task = cset->tasks.next;
2720 * css_task_iter_start - initiate task iteration
2721 * @css: the css to walk tasks of
2722 * @it: the task iterator to use
2724 * Initiate iteration through the tasks of @css. The caller can call
2725 * css_task_iter_next() to walk through the tasks until the function
2726 * returns NULL. On completion of iteration, css_task_iter_end() must be
2729 * Note that this function acquires a lock which is released when the
2730 * iteration finishes. The caller can't sleep while iteration is in
2733 void css_task_iter_start(struct cgroup_subsys_state *css,
2734 struct css_task_iter *it)
2735 __acquires(css_set_lock)
2738 * The first time anyone tries to iterate across a css, we need to
2739 * enable the list linking each css_set to its tasks, and fix up
2740 * all existing tasks.
2742 if (!use_task_css_set_links)
2743 cgroup_enable_task_cg_lists();
2745 read_lock(&css_set_lock);
2747 it->origin_css = css;
2748 it->cset_link = &css->cgroup->cset_links;
2750 css_advance_task_iter(it);
2754 * css_task_iter_next - return the next task for the iterator
2755 * @it: the task iterator being iterated
2757 * The "next" function for task iteration. @it should have been
2758 * initialized via css_task_iter_start(). Returns NULL when the iteration
2761 struct task_struct *css_task_iter_next(struct css_task_iter *it)
2763 struct task_struct *res;
2764 struct list_head *l = it->task;
2765 struct cgrp_cset_link *link;
2767 /* If the iterator cg is NULL, we have no tasks */
2770 res = list_entry(l, struct task_struct, cg_list);
2771 /* Advance iterator to find next entry */
2773 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
2774 if (l == &link->cset->tasks) {
2776 * We reached the end of this task list - move on to the
2777 * next cgrp_cset_link.
2779 css_advance_task_iter(it);
2787 * css_task_iter_end - finish task iteration
2788 * @it: the task iterator to finish
2790 * Finish task iteration started by css_task_iter_start().
2792 void css_task_iter_end(struct css_task_iter *it)
2793 __releases(css_set_lock)
2795 read_unlock(&css_set_lock);
2798 static inline int started_after_time(struct task_struct *t1,
2799 struct timespec *time,
2800 struct task_struct *t2)
2802 int start_diff = timespec_compare(&t1->start_time, time);
2803 if (start_diff > 0) {
2805 } else if (start_diff < 0) {
2809 * Arbitrarily, if two processes started at the same
2810 * time, we'll say that the lower pointer value
2811 * started first. Note that t2 may have exited by now
2812 * so this may not be a valid pointer any longer, but
2813 * that's fine - it still serves to distinguish
2814 * between two tasks started (effectively) simultaneously.
2821 * This function is a callback from heap_insert() and is used to order
2823 * In this case we order the heap in descending task start time.
2825 static inline int started_after(void *p1, void *p2)
2827 struct task_struct *t1 = p1;
2828 struct task_struct *t2 = p2;
2829 return started_after_time(t1, &t2->start_time, t2);
2833 * css_scan_tasks - iterate though all the tasks in a css
2834 * @css: the css to iterate tasks of
2835 * @test: optional test callback
2836 * @process: process callback
2837 * @data: data passed to @test and @process
2838 * @heap: optional pre-allocated heap used for task iteration
2840 * Iterate through all the tasks in @css, calling @test for each, and if it
2841 * returns %true, call @process for it also.
2843 * @test may be NULL, meaning always true (select all tasks), which
2844 * effectively duplicates css_task_iter_{start,next,end}() but does not
2845 * lock css_set_lock for the call to @process.
2847 * It is guaranteed that @process will act on every task that is a member
2848 * of @css for the duration of this call. This function may or may not
2849 * call @process for tasks that exit or move to a different css during the
2850 * call, or are forked or move into the css during the call.
2852 * Note that @test may be called with locks held, and may in some
2853 * situations be called multiple times for the same task, so it should be
2856 * If @heap is non-NULL, a heap has been pre-allocated and will be used for
2857 * heap operations (and its "gt" member will be overwritten), else a
2858 * temporary heap will be used (allocation of which may cause this function
2861 int css_scan_tasks(struct cgroup_subsys_state *css,
2862 bool (*test)(struct task_struct *, void *),
2863 void (*process)(struct task_struct *, void *),
2864 void *data, struct ptr_heap *heap)
2867 struct css_task_iter it;
2868 struct task_struct *p, *dropped;
2869 /* Never dereference latest_task, since it's not refcounted */
2870 struct task_struct *latest_task = NULL;
2871 struct ptr_heap tmp_heap;
2872 struct timespec latest_time = { 0, 0 };
2875 /* The caller supplied our heap and pre-allocated its memory */
2876 heap->gt = &started_after;
2878 /* We need to allocate our own heap memory */
2880 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2882 /* cannot allocate the heap */
2888 * Scan tasks in the css, using the @test callback to determine
2889 * which are of interest, and invoking @process callback on the
2890 * ones which need an update. Since we don't want to hold any
2891 * locks during the task updates, gather tasks to be processed in a
2892 * heap structure. The heap is sorted by descending task start
2893 * time. If the statically-sized heap fills up, we overflow tasks
2894 * that started later, and in future iterations only consider tasks
2895 * that started after the latest task in the previous pass. This
2896 * guarantees forward progress and that we don't miss any tasks.
2899 css_task_iter_start(css, &it);
2900 while ((p = css_task_iter_next(&it))) {
2902 * Only affect tasks that qualify per the caller's callback,
2903 * if he provided one
2905 if (test && !test(p, data))
2908 * Only process tasks that started after the last task
2911 if (!started_after_time(p, &latest_time, latest_task))
2913 dropped = heap_insert(heap, p);
2914 if (dropped == NULL) {
2916 * The new task was inserted; the heap wasn't
2920 } else if (dropped != p) {
2922 * The new task was inserted, and pushed out a
2926 put_task_struct(dropped);
2929 * Else the new task was newer than anything already in
2930 * the heap and wasn't inserted
2933 css_task_iter_end(&it);
2936 for (i = 0; i < heap->size; i++) {
2937 struct task_struct *q = heap->ptrs[i];
2939 latest_time = q->start_time;
2942 /* Process the task per the caller's callback */
2947 * If we had to process any tasks at all, scan again
2948 * in case some of them were in the middle of forking
2949 * children that didn't get processed.
2950 * Not the most efficient way to do it, but it avoids
2951 * having to take callback_mutex in the fork path
2955 if (heap == &tmp_heap)
2956 heap_free(&tmp_heap);
2960 static void cgroup_transfer_one_task(struct task_struct *task, void *data)
2962 struct cgroup *new_cgroup = data;
2964 mutex_lock(&cgroup_mutex);
2965 cgroup_attach_task(new_cgroup, task, false);
2966 mutex_unlock(&cgroup_mutex);
2970 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
2971 * @to: cgroup to which the tasks will be moved
2972 * @from: cgroup in which the tasks currently reside
2974 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
2976 return css_scan_tasks(&from->dummy_css, NULL, cgroup_transfer_one_task,
2981 * Stuff for reading the 'tasks'/'procs' files.
2983 * Reading this file can return large amounts of data if a cgroup has
2984 * *lots* of attached tasks. So it may need several calls to read(),
2985 * but we cannot guarantee that the information we produce is correct
2986 * unless we produce it entirely atomically.
2990 /* which pidlist file are we talking about? */
2991 enum cgroup_filetype {
2997 * A pidlist is a list of pids that virtually represents the contents of one
2998 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
2999 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3002 struct cgroup_pidlist {
3004 * used to find which pidlist is wanted. doesn't change as long as
3005 * this particular list stays in the list.
3007 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3010 /* how many elements the above list has */
3012 /* each of these stored in a list by its cgroup */
3013 struct list_head links;
3014 /* pointer to the cgroup we belong to, for list removal purposes */
3015 struct cgroup *owner;
3016 /* for delayed destruction */
3017 struct delayed_work destroy_dwork;
3021 * The following two functions "fix" the issue where there are more pids
3022 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3023 * TODO: replace with a kernel-wide solution to this problem
3025 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3026 static void *pidlist_allocate(int count)
3028 if (PIDLIST_TOO_LARGE(count))
3029 return vmalloc(count * sizeof(pid_t));
3031 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3034 static void pidlist_free(void *p)
3036 if (is_vmalloc_addr(p))
3043 * Used to destroy all pidlists lingering waiting for destroy timer. None
3044 * should be left afterwards.
3046 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3048 struct cgroup_pidlist *l, *tmp_l;
3050 mutex_lock(&cgrp->pidlist_mutex);
3051 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3052 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3053 mutex_unlock(&cgrp->pidlist_mutex);
3055 flush_workqueue(cgroup_pidlist_destroy_wq);
3056 BUG_ON(!list_empty(&cgrp->pidlists));
3059 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3061 struct delayed_work *dwork = to_delayed_work(work);
3062 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3064 struct cgroup_pidlist *tofree = NULL;
3066 mutex_lock(&l->owner->pidlist_mutex);
3069 * Destroy iff we didn't get queued again. The state won't change
3070 * as destroy_dwork can only be queued while locked.
3072 if (!delayed_work_pending(dwork)) {
3073 list_del(&l->links);
3074 pidlist_free(l->list);
3075 put_pid_ns(l->key.ns);
3079 mutex_unlock(&l->owner->pidlist_mutex);
3084 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3085 * Returns the number of unique elements.
3087 static int pidlist_uniq(pid_t *list, int length)
3092 * we presume the 0th element is unique, so i starts at 1. trivial
3093 * edge cases first; no work needs to be done for either
3095 if (length == 0 || length == 1)
3097 /* src and dest walk down the list; dest counts unique elements */
3098 for (src = 1; src < length; src++) {
3099 /* find next unique element */
3100 while (list[src] == list[src-1]) {
3105 /* dest always points to where the next unique element goes */
3106 list[dest] = list[src];
3114 * The two pid files - task and cgroup.procs - guaranteed that the result
3115 * is sorted, which forced this whole pidlist fiasco. As pid order is
3116 * different per namespace, each namespace needs differently sorted list,
3117 * making it impossible to use, for example, single rbtree of member tasks
3118 * sorted by task pointer. As pidlists can be fairly large, allocating one
3119 * per open file is dangerous, so cgroup had to implement shared pool of
3120 * pidlists keyed by cgroup and namespace.
3122 * All this extra complexity was caused by the original implementation
3123 * committing to an entirely unnecessary property. In the long term, we
3124 * want to do away with it. Explicitly scramble sort order if
3125 * sane_behavior so that no such expectation exists in the new interface.
3127 * Scrambling is done by swapping every two consecutive bits, which is
3128 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3130 static pid_t pid_fry(pid_t pid)
3132 unsigned a = pid & 0x55555555;
3133 unsigned b = pid & 0xAAAAAAAA;
3135 return (a << 1) | (b >> 1);
3138 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3140 if (cgroup_sane_behavior(cgrp))
3141 return pid_fry(pid);
3146 static int cmppid(const void *a, const void *b)
3148 return *(pid_t *)a - *(pid_t *)b;
3151 static int fried_cmppid(const void *a, const void *b)
3153 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3156 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3157 enum cgroup_filetype type)
3159 struct cgroup_pidlist *l;
3160 /* don't need task_nsproxy() if we're looking at ourself */
3161 struct pid_namespace *ns = task_active_pid_ns(current);
3163 lockdep_assert_held(&cgrp->pidlist_mutex);
3165 list_for_each_entry(l, &cgrp->pidlists, links)
3166 if (l->key.type == type && l->key.ns == ns)
3172 * find the appropriate pidlist for our purpose (given procs vs tasks)
3173 * returns with the lock on that pidlist already held, and takes care
3174 * of the use count, or returns NULL with no locks held if we're out of
3177 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3178 enum cgroup_filetype type)
3180 struct cgroup_pidlist *l;
3182 lockdep_assert_held(&cgrp->pidlist_mutex);
3184 l = cgroup_pidlist_find(cgrp, type);
3188 /* entry not found; create a new one */
3189 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3193 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3195 /* don't need task_nsproxy() if we're looking at ourself */
3196 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3198 list_add(&l->links, &cgrp->pidlists);
3203 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3205 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3206 struct cgroup_pidlist **lp)
3210 int pid, n = 0; /* used for populating the array */
3211 struct css_task_iter it;
3212 struct task_struct *tsk;
3213 struct cgroup_pidlist *l;
3215 lockdep_assert_held(&cgrp->pidlist_mutex);
3218 * If cgroup gets more users after we read count, we won't have
3219 * enough space - tough. This race is indistinguishable to the
3220 * caller from the case that the additional cgroup users didn't
3221 * show up until sometime later on.
3223 length = cgroup_task_count(cgrp);
3224 array = pidlist_allocate(length);
3227 /* now, populate the array */
3228 css_task_iter_start(&cgrp->dummy_css, &it);
3229 while ((tsk = css_task_iter_next(&it))) {
3230 if (unlikely(n == length))
3232 /* get tgid or pid for procs or tasks file respectively */
3233 if (type == CGROUP_FILE_PROCS)
3234 pid = task_tgid_vnr(tsk);
3236 pid = task_pid_vnr(tsk);
3237 if (pid > 0) /* make sure to only use valid results */
3240 css_task_iter_end(&it);
3242 /* now sort & (if procs) strip out duplicates */
3243 if (cgroup_sane_behavior(cgrp))
3244 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3246 sort(array, length, sizeof(pid_t), cmppid, NULL);
3247 if (type == CGROUP_FILE_PROCS)
3248 length = pidlist_uniq(array, length);
3250 l = cgroup_pidlist_find_create(cgrp, type);
3252 mutex_unlock(&cgrp->pidlist_mutex);
3253 pidlist_free(array);
3257 /* store array, freeing old if necessary */
3258 pidlist_free(l->list);
3266 * cgroupstats_build - build and fill cgroupstats
3267 * @stats: cgroupstats to fill information into
3268 * @dentry: A dentry entry belonging to the cgroup for which stats have
3271 * Build and fill cgroupstats so that taskstats can export it to user
3274 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3276 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3277 struct cgroup *cgrp;
3278 struct css_task_iter it;
3279 struct task_struct *tsk;
3281 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3282 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3283 kernfs_type(kn) != KERNFS_DIR)
3287 * We aren't being called from kernfs and there's no guarantee on
3288 * @kn->priv's validity. For this and css_tryget_from_dir(),
3289 * @kn->priv is RCU safe. Let's do the RCU dancing.
3292 cgrp = rcu_dereference(kn->priv);
3298 css_task_iter_start(&cgrp->dummy_css, &it);
3299 while ((tsk = css_task_iter_next(&it))) {
3300 switch (tsk->state) {
3302 stats->nr_running++;
3304 case TASK_INTERRUPTIBLE:
3305 stats->nr_sleeping++;
3307 case TASK_UNINTERRUPTIBLE:
3308 stats->nr_uninterruptible++;
3311 stats->nr_stopped++;
3314 if (delayacct_is_task_waiting_on_io(tsk))
3315 stats->nr_io_wait++;
3319 css_task_iter_end(&it);
3327 * seq_file methods for the tasks/procs files. The seq_file position is the
3328 * next pid to display; the seq_file iterator is a pointer to the pid
3329 * in the cgroup->l->list array.
3332 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3335 * Initially we receive a position value that corresponds to
3336 * one more than the last pid shown (or 0 on the first call or
3337 * after a seek to the start). Use a binary-search to find the
3338 * next pid to display, if any
3340 struct kernfs_open_file *of = s->private;
3341 struct cgroup *cgrp = seq_css(s)->cgroup;
3342 struct cgroup_pidlist *l;
3343 enum cgroup_filetype type = seq_cft(s)->private;
3344 int index = 0, pid = *pos;
3347 mutex_lock(&cgrp->pidlist_mutex);
3350 * !NULL @of->priv indicates that this isn't the first start()
3351 * after open. If the matching pidlist is around, we can use that.
3352 * Look for it. Note that @of->priv can't be used directly. It
3353 * could already have been destroyed.
3356 of->priv = cgroup_pidlist_find(cgrp, type);
3359 * Either this is the first start() after open or the matching
3360 * pidlist has been destroyed inbetween. Create a new one.
3363 ret = pidlist_array_load(cgrp, type,
3364 (struct cgroup_pidlist **)&of->priv);
3366 return ERR_PTR(ret);
3371 int end = l->length;
3373 while (index < end) {
3374 int mid = (index + end) / 2;
3375 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3378 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3384 /* If we're off the end of the array, we're done */
3385 if (index >= l->length)
3387 /* Update the abstract position to be the actual pid that we found */
3388 iter = l->list + index;
3389 *pos = cgroup_pid_fry(cgrp, *iter);
3393 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3395 struct kernfs_open_file *of = s->private;
3396 struct cgroup_pidlist *l = of->priv;
3399 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3400 CGROUP_PIDLIST_DESTROY_DELAY);
3401 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3404 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3406 struct kernfs_open_file *of = s->private;
3407 struct cgroup_pidlist *l = of->priv;
3409 pid_t *end = l->list + l->length;
3411 * Advance to the next pid in the array. If this goes off the
3418 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3423 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3425 return seq_printf(s, "%d\n", *(int *)v);
3429 * seq_operations functions for iterating on pidlists through seq_file -
3430 * independent of whether it's tasks or procs
3432 static const struct seq_operations cgroup_pidlist_seq_operations = {
3433 .start = cgroup_pidlist_start,
3434 .stop = cgroup_pidlist_stop,
3435 .next = cgroup_pidlist_next,
3436 .show = cgroup_pidlist_show,
3439 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3442 return notify_on_release(css->cgroup);
3445 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3446 struct cftype *cft, u64 val)
3448 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3450 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3452 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3456 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3459 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3462 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3463 struct cftype *cft, u64 val)
3466 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3468 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3472 static struct cftype cgroup_base_files[] = {
3474 .name = "cgroup.procs",
3475 .seq_start = cgroup_pidlist_start,
3476 .seq_next = cgroup_pidlist_next,
3477 .seq_stop = cgroup_pidlist_stop,
3478 .seq_show = cgroup_pidlist_show,
3479 .private = CGROUP_FILE_PROCS,
3480 .write_u64 = cgroup_procs_write,
3481 .mode = S_IRUGO | S_IWUSR,
3484 .name = "cgroup.clone_children",
3485 .flags = CFTYPE_INSANE,
3486 .read_u64 = cgroup_clone_children_read,
3487 .write_u64 = cgroup_clone_children_write,
3490 .name = "cgroup.sane_behavior",
3491 .flags = CFTYPE_ONLY_ON_ROOT,
3492 .seq_show = cgroup_sane_behavior_show,
3496 * Historical crazy stuff. These don't have "cgroup." prefix and
3497 * don't exist if sane_behavior. If you're depending on these, be
3498 * prepared to be burned.
3502 .flags = CFTYPE_INSANE, /* use "procs" instead */
3503 .seq_start = cgroup_pidlist_start,
3504 .seq_next = cgroup_pidlist_next,
3505 .seq_stop = cgroup_pidlist_stop,
3506 .seq_show = cgroup_pidlist_show,
3507 .private = CGROUP_FILE_TASKS,
3508 .write_u64 = cgroup_tasks_write,
3509 .mode = S_IRUGO | S_IWUSR,
3512 .name = "notify_on_release",
3513 .flags = CFTYPE_INSANE,
3514 .read_u64 = cgroup_read_notify_on_release,
3515 .write_u64 = cgroup_write_notify_on_release,
3518 .name = "release_agent",
3519 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3520 .seq_show = cgroup_release_agent_show,
3521 .write_string = cgroup_release_agent_write,
3522 .max_write_len = PATH_MAX - 1,
3528 * cgroup_populate_dir - create subsys files in a cgroup directory
3529 * @cgrp: target cgroup
3530 * @subsys_mask: mask of the subsystem ids whose files should be added
3532 * On failure, no file is added.
3534 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3536 struct cgroup_subsys *ss;
3539 /* process cftsets of each subsystem */
3540 for_each_subsys(ss, i) {
3541 struct cftype *cfts;
3543 if (!test_bit(i, &subsys_mask))
3546 list_for_each_entry(cfts, &ss->cfts, node) {
3547 ret = cgroup_addrm_files(cgrp, cfts, true);
3554 cgroup_clear_dir(cgrp, subsys_mask);
3559 * css destruction is four-stage process.
3561 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3562 * Implemented in kill_css().
3564 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3565 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3566 * by invoking offline_css(). After offlining, the base ref is put.
3567 * Implemented in css_killed_work_fn().
3569 * 3. When the percpu_ref reaches zero, the only possible remaining
3570 * accessors are inside RCU read sections. css_release() schedules the
3573 * 4. After the grace period, the css can be freed. Implemented in
3574 * css_free_work_fn().
3576 * It is actually hairier because both step 2 and 4 require process context
3577 * and thus involve punting to css->destroy_work adding two additional
3578 * steps to the already complex sequence.
3580 static void css_free_work_fn(struct work_struct *work)
3582 struct cgroup_subsys_state *css =
3583 container_of(work, struct cgroup_subsys_state, destroy_work);
3584 struct cgroup *cgrp = css->cgroup;
3587 css_put(css->parent);
3589 css->ss->css_free(css);
3593 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3595 struct cgroup_subsys_state *css =
3596 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3598 INIT_WORK(&css->destroy_work, css_free_work_fn);
3599 queue_work(cgroup_destroy_wq, &css->destroy_work);
3602 static void css_release(struct percpu_ref *ref)
3604 struct cgroup_subsys_state *css =
3605 container_of(ref, struct cgroup_subsys_state, refcnt);
3607 rcu_assign_pointer(css->cgroup->subsys[css->ss->id], NULL);
3608 call_rcu(&css->rcu_head, css_free_rcu_fn);
3611 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3612 struct cgroup *cgrp)
3619 css->parent = cgroup_css(cgrp->parent, ss);
3621 css->flags |= CSS_ROOT;
3623 BUG_ON(cgroup_css(cgrp, ss));
3626 /* invoke ->css_online() on a new CSS and mark it online if successful */
3627 static int online_css(struct cgroup_subsys_state *css)
3629 struct cgroup_subsys *ss = css->ss;
3632 lockdep_assert_held(&cgroup_tree_mutex);
3633 lockdep_assert_held(&cgroup_mutex);
3636 ret = ss->css_online(css);
3638 css->flags |= CSS_ONLINE;
3639 css->cgroup->nr_css++;
3640 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
3645 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
3646 static void offline_css(struct cgroup_subsys_state *css)
3648 struct cgroup_subsys *ss = css->ss;
3650 lockdep_assert_held(&cgroup_tree_mutex);
3651 lockdep_assert_held(&cgroup_mutex);
3653 if (!(css->flags & CSS_ONLINE))
3656 if (ss->css_offline)
3657 ss->css_offline(css);
3659 css->flags &= ~CSS_ONLINE;
3660 css->cgroup->nr_css--;
3661 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
3665 * create_css - create a cgroup_subsys_state
3666 * @cgrp: the cgroup new css will be associated with
3667 * @ss: the subsys of new css
3669 * Create a new css associated with @cgrp - @ss pair. On success, the new
3670 * css is online and installed in @cgrp with all interface files created.
3671 * Returns 0 on success, -errno on failure.
3673 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
3675 struct cgroup *parent = cgrp->parent;
3676 struct cgroup_subsys_state *css;
3679 lockdep_assert_held(&cgroup_mutex);
3681 css = ss->css_alloc(cgroup_css(parent, ss));
3683 return PTR_ERR(css);
3685 err = percpu_ref_init(&css->refcnt, css_release);
3689 init_css(css, ss, cgrp);
3691 err = cgroup_populate_dir(cgrp, 1 << ss->id);
3695 err = online_css(css);
3700 css_get(css->parent);
3702 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
3704 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",
3705 current->comm, current->pid, ss->name);
3706 if (!strcmp(ss->name, "memory"))
3707 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
3708 ss->warned_broken_hierarchy = true;
3714 percpu_ref_cancel_init(&css->refcnt);
3720 * cgroup_create - create a cgroup
3721 * @parent: cgroup that will be parent of the new cgroup
3722 * @name_str: name of the new cgroup
3723 * @mode: mode to set on new cgroup
3725 static long cgroup_create(struct cgroup *parent, const char *name_str,
3728 struct cgroup *cgrp;
3729 struct cgroup_name *name;
3730 struct cgroupfs_root *root = parent->root;
3732 struct cgroup_subsys *ss;
3733 struct kernfs_node *kn;
3735 /* allocate the cgroup and its ID, 0 is reserved for the root */
3736 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3740 name = cgroup_alloc_name(name_str);
3745 rcu_assign_pointer(cgrp->name, name);
3747 mutex_lock(&cgroup_tree_mutex);
3750 * Only live parents can have children. Note that the liveliness
3751 * check isn't strictly necessary because cgroup_mkdir() and
3752 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
3753 * anyway so that locking is contained inside cgroup proper and we
3754 * don't get nasty surprises if we ever grow another caller.
3756 if (!cgroup_lock_live_group(parent)) {
3758 goto err_unlock_tree;
3762 * Temporarily set the pointer to NULL, so idr_find() won't return
3763 * a half-baked cgroup.
3765 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
3771 init_cgroup_housekeeping(cgrp);
3773 cgrp->parent = parent;
3774 cgrp->dummy_css.parent = &parent->dummy_css;
3775 cgrp->root = parent->root;
3777 if (notify_on_release(parent))
3778 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3780 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
3781 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
3783 /* create the directory */
3784 kn = kernfs_create_dir(parent->kn, name->name, mode, cgrp);
3792 * This extra ref will be put in cgroup_free_fn() and guarantees
3793 * that @cgrp->kn is always accessible.
3797 cgrp->serial_nr = cgroup_serial_nr_next++;
3799 /* allocation complete, commit to creation */
3800 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
3801 root->number_of_cgroups++;
3804 * Grab a reference on the root and parent so that they don't get
3805 * deleted while there are child cgroups.
3807 cgroup_get_root(root);
3811 * @cgrp is now fully operational. If something fails after this
3812 * point, it'll be released via the normal destruction path.
3814 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
3816 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
3820 /* let's create and online css's */
3821 for_each_subsys(ss, ssid) {
3822 if (root->subsys_mask & (1 << ssid)) {
3823 err = create_css(cgrp, ss);
3829 kernfs_activate(kn);
3831 mutex_unlock(&cgroup_mutex);
3832 mutex_unlock(&cgroup_tree_mutex);
3837 idr_remove(&root->cgroup_idr, cgrp->id);
3839 mutex_unlock(&cgroup_mutex);
3841 mutex_unlock(&cgroup_tree_mutex);
3842 kfree(rcu_dereference_raw(cgrp->name));
3848 cgroup_destroy_locked(cgrp);
3849 mutex_unlock(&cgroup_mutex);
3850 mutex_unlock(&cgroup_tree_mutex);
3854 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3857 struct cgroup *parent = parent_kn->priv;
3859 return cgroup_create(parent, name, mode);
3863 * This is called when the refcnt of a css is confirmed to be killed.
3864 * css_tryget() is now guaranteed to fail.
3866 static void css_killed_work_fn(struct work_struct *work)
3868 struct cgroup_subsys_state *css =
3869 container_of(work, struct cgroup_subsys_state, destroy_work);
3870 struct cgroup *cgrp = css->cgroup;
3872 mutex_lock(&cgroup_tree_mutex);
3873 mutex_lock(&cgroup_mutex);
3876 * css_tryget() is guaranteed to fail now. Tell subsystems to
3877 * initate destruction.
3882 * If @cgrp is marked dead, it's waiting for refs of all css's to
3883 * be disabled before proceeding to the second phase of cgroup
3884 * destruction. If we are the last one, kick it off.
3886 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
3887 cgroup_destroy_css_killed(cgrp);
3889 mutex_unlock(&cgroup_mutex);
3890 mutex_unlock(&cgroup_tree_mutex);
3893 * Put the css refs from kill_css(). Each css holds an extra
3894 * reference to the cgroup's dentry and cgroup removal proceeds
3895 * regardless of css refs. On the last put of each css, whenever
3896 * that may be, the extra dentry ref is put so that dentry
3897 * destruction happens only after all css's are released.
3902 /* css kill confirmation processing requires process context, bounce */
3903 static void css_killed_ref_fn(struct percpu_ref *ref)
3905 struct cgroup_subsys_state *css =
3906 container_of(ref, struct cgroup_subsys_state, refcnt);
3908 INIT_WORK(&css->destroy_work, css_killed_work_fn);
3909 queue_work(cgroup_destroy_wq, &css->destroy_work);
3913 * kill_css - destroy a css
3914 * @css: css to destroy
3916 * This function initiates destruction of @css by removing cgroup interface
3917 * files and putting its base reference. ->css_offline() will be invoked
3918 * asynchronously once css_tryget() is guaranteed to fail and when the
3919 * reference count reaches zero, @css will be released.
3921 static void kill_css(struct cgroup_subsys_state *css)
3924 * This must happen before css is disassociated with its cgroup.
3925 * See seq_css() for details.
3927 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3930 * Killing would put the base ref, but we need to keep it alive
3931 * until after ->css_offline().
3936 * cgroup core guarantees that, by the time ->css_offline() is
3937 * invoked, no new css reference will be given out via
3938 * css_tryget(). We can't simply call percpu_ref_kill() and
3939 * proceed to offlining css's because percpu_ref_kill() doesn't
3940 * guarantee that the ref is seen as killed on all CPUs on return.
3942 * Use percpu_ref_kill_and_confirm() to get notifications as each
3943 * css is confirmed to be seen as killed on all CPUs.
3945 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
3949 * cgroup_destroy_locked - the first stage of cgroup destruction
3950 * @cgrp: cgroup to be destroyed
3952 * css's make use of percpu refcnts whose killing latency shouldn't be
3953 * exposed to userland and are RCU protected. Also, cgroup core needs to
3954 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
3955 * invoked. To satisfy all the requirements, destruction is implemented in
3956 * the following two steps.
3958 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
3959 * userland visible parts and start killing the percpu refcnts of
3960 * css's. Set up so that the next stage will be kicked off once all
3961 * the percpu refcnts are confirmed to be killed.
3963 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
3964 * rest of destruction. Once all cgroup references are gone, the
3965 * cgroup is RCU-freed.
3967 * This function implements s1. After this step, @cgrp is gone as far as
3968 * the userland is concerned and a new cgroup with the same name may be
3969 * created. As cgroup doesn't care about the names internally, this
3970 * doesn't cause any problem.
3972 static int cgroup_destroy_locked(struct cgroup *cgrp)
3973 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
3975 struct cgroup *child;
3976 struct cgroup_subsys_state *css;
3980 lockdep_assert_held(&cgroup_tree_mutex);
3981 lockdep_assert_held(&cgroup_mutex);
3984 * css_set_lock synchronizes access to ->cset_links and prevents
3985 * @cgrp from being removed while __put_css_set() is in progress.
3987 read_lock(&css_set_lock);
3988 empty = list_empty(&cgrp->cset_links);
3989 read_unlock(&css_set_lock);
3994 * Make sure there's no live children. We can't test ->children
3995 * emptiness as dead children linger on it while being destroyed;
3996 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
4000 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
4001 empty = cgroup_is_dead(child);
4010 * Initiate massacre of all css's. cgroup_destroy_css_killed()
4011 * will be invoked to perform the rest of destruction once the
4012 * percpu refs of all css's are confirmed to be killed. This
4013 * involves removing the subsystem's files, drop cgroup_mutex.
4015 mutex_unlock(&cgroup_mutex);
4016 for_each_css(css, ssid, cgrp)
4018 mutex_lock(&cgroup_mutex);
4021 * Mark @cgrp dead. This prevents further task migration and child
4022 * creation by disabling cgroup_lock_live_group(). Note that
4023 * CGRP_DEAD assertion is depended upon by css_next_child() to
4024 * resume iteration after dropping RCU read lock. See
4025 * css_next_child() for details.
4027 set_bit(CGRP_DEAD, &cgrp->flags);
4029 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4030 raw_spin_lock(&release_list_lock);
4031 if (!list_empty(&cgrp->release_list))
4032 list_del_init(&cgrp->release_list);
4033 raw_spin_unlock(&release_list_lock);
4036 * If @cgrp has css's attached, the second stage of cgroup
4037 * destruction is kicked off from css_killed_work_fn() after the
4038 * refs of all attached css's are killed. If @cgrp doesn't have
4039 * any css, we kick it off here.
4042 cgroup_destroy_css_killed(cgrp);
4044 /* remove @cgrp directory along with the base files */
4045 mutex_unlock(&cgroup_mutex);
4048 * There are two control paths which try to determine cgroup from
4049 * dentry without going through kernfs - cgroupstats_build() and
4050 * css_tryget_from_dir(). Those are supported by RCU protecting
4051 * clearing of cgrp->kn->priv backpointer, which should happen
4052 * after all files under it have been removed.
4054 kernfs_remove(cgrp->kn); /* @cgrp has an extra ref on its kn */
4055 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4057 mutex_lock(&cgroup_mutex);
4063 * cgroup_destroy_css_killed - the second step of cgroup destruction
4064 * @work: cgroup->destroy_free_work
4066 * This function is invoked from a work item for a cgroup which is being
4067 * destroyed after all css's are offlined and performs the rest of
4068 * destruction. This is the second step of destruction described in the
4069 * comment above cgroup_destroy_locked().
4071 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
4073 struct cgroup *parent = cgrp->parent;
4075 lockdep_assert_held(&cgroup_tree_mutex);
4076 lockdep_assert_held(&cgroup_mutex);
4078 /* delete this cgroup from parent->children */
4079 list_del_rcu(&cgrp->sibling);
4083 set_bit(CGRP_RELEASABLE, &parent->flags);
4084 check_for_release(parent);
4087 static int cgroup_rmdir(struct kernfs_node *kn)
4089 struct cgroup *cgrp = kn->priv;
4093 * This is self-destruction but @kn can't be removed while this
4094 * callback is in progress. Let's break active protection. Once
4095 * the protection is broken, @cgrp can be destroyed at any point.
4096 * Pin it so that it stays accessible.
4099 kernfs_break_active_protection(kn);
4101 mutex_lock(&cgroup_tree_mutex);
4102 mutex_lock(&cgroup_mutex);
4105 * @cgrp might already have been destroyed while we're trying to
4108 if (!cgroup_is_dead(cgrp))
4109 ret = cgroup_destroy_locked(cgrp);
4111 mutex_unlock(&cgroup_mutex);
4112 mutex_unlock(&cgroup_tree_mutex);
4114 kernfs_unbreak_active_protection(kn);
4119 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4120 .remount_fs = cgroup_remount,
4121 .show_options = cgroup_show_options,
4122 .mkdir = cgroup_mkdir,
4123 .rmdir = cgroup_rmdir,
4124 .rename = cgroup_rename,
4127 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4129 struct cgroup_subsys_state *css;
4131 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4133 mutex_lock(&cgroup_tree_mutex);
4134 mutex_lock(&cgroup_mutex);
4136 INIT_LIST_HEAD(&ss->cfts);
4138 /* Create the top cgroup state for this subsystem */
4139 ss->root = &cgroup_dummy_root;
4140 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4141 /* We don't handle early failures gracefully */
4142 BUG_ON(IS_ERR(css));
4143 init_css(css, ss, cgroup_dummy_top);
4145 /* Update the init_css_set to contain a subsys
4146 * pointer to this state - since the subsystem is
4147 * newly registered, all tasks and hence the
4148 * init_css_set is in the subsystem's top cgroup. */
4149 init_css_set.subsys[ss->id] = css;
4151 need_forkexit_callback |= ss->fork || ss->exit;
4153 /* At system boot, before all subsystems have been
4154 * registered, no tasks have been forked, so we don't
4155 * need to invoke fork callbacks here. */
4156 BUG_ON(!list_empty(&init_task.tasks));
4158 BUG_ON(online_css(css));
4160 mutex_unlock(&cgroup_mutex);
4161 mutex_unlock(&cgroup_tree_mutex);
4165 * cgroup_init_early - cgroup initialization at system boot
4167 * Initialize cgroups at system boot, and initialize any
4168 * subsystems that request early init.
4170 int __init cgroup_init_early(void)
4172 struct cgroup_subsys *ss;
4175 atomic_set(&init_css_set.refcount, 1);
4176 INIT_LIST_HEAD(&init_css_set.cgrp_links);
4177 INIT_LIST_HEAD(&init_css_set.tasks);
4178 INIT_HLIST_NODE(&init_css_set.hlist);
4180 init_cgroup_root(&cgroup_dummy_root);
4181 cgroup_root_count = 1;
4182 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4184 init_cgrp_cset_link.cset = &init_css_set;
4185 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4186 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4187 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4189 for_each_subsys(ss, i) {
4190 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4191 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4192 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4194 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4195 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4198 ss->name = cgroup_subsys_name[i];
4201 cgroup_init_subsys(ss);
4207 * cgroup_init - cgroup initialization
4209 * Register cgroup filesystem and /proc file, and initialize
4210 * any subsystems that didn't request early init.
4212 int __init cgroup_init(void)
4214 struct cgroup_subsys *ss;
4218 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4220 for_each_subsys(ss, i) {
4221 if (!ss->early_init)
4222 cgroup_init_subsys(ss);
4225 * cftype registration needs kmalloc and can't be done
4226 * during early_init. Register base cftypes separately.
4228 if (ss->base_cftypes)
4229 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4232 /* allocate id for the dummy hierarchy */
4233 mutex_lock(&cgroup_mutex);
4235 /* Add init_css_set to the hash table */
4236 key = css_set_hash(init_css_set.subsys);
4237 hash_add(css_set_table, &init_css_set.hlist, key);
4239 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4241 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
4245 mutex_unlock(&cgroup_mutex);
4247 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4251 err = register_filesystem(&cgroup_fs_type);
4253 kobject_put(cgroup_kobj);
4257 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4261 static int __init cgroup_wq_init(void)
4264 * There isn't much point in executing destruction path in
4265 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4267 * XXX: Must be ordered to make sure parent is offlined after
4268 * children. The ordering requirement is for memcg where a
4269 * parent's offline may wait for a child's leading to deadlock. In
4270 * the long term, this should be fixed from memcg side.
4272 * We would prefer to do this in cgroup_init() above, but that
4273 * is called before init_workqueues(): so leave this until after.
4275 cgroup_destroy_wq = alloc_ordered_workqueue("cgroup_destroy", 0);
4276 BUG_ON(!cgroup_destroy_wq);
4279 * Used to destroy pidlists and separate to serve as flush domain.
4280 * Cap @max_active to 1 too.
4282 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4284 BUG_ON(!cgroup_pidlist_destroy_wq);
4288 core_initcall(cgroup_wq_init);
4291 * proc_cgroup_show()
4292 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4293 * - Used for /proc/<pid>/cgroup.
4294 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4295 * doesn't really matter if tsk->cgroup changes after we read it,
4296 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4297 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4298 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4299 * cgroup to top_cgroup.
4302 /* TODO: Use a proper seq_file iterator */
4303 int proc_cgroup_show(struct seq_file *m, void *v)
4306 struct task_struct *tsk;
4309 struct cgroupfs_root *root;
4312 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4318 tsk = get_pid_task(pid, PIDTYPE_PID);
4324 mutex_lock(&cgroup_mutex);
4326 for_each_active_root(root) {
4327 struct cgroup_subsys *ss;
4328 struct cgroup *cgrp;
4329 int ssid, count = 0;
4331 seq_printf(m, "%d:", root->hierarchy_id);
4332 for_each_subsys(ss, ssid)
4333 if (root->subsys_mask & (1 << ssid))
4334 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4335 if (strlen(root->name))
4336 seq_printf(m, "%sname=%s", count ? "," : "",
4339 cgrp = task_cgroup_from_root(tsk, root);
4340 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4348 mutex_unlock(&cgroup_mutex);
4349 put_task_struct(tsk);
4356 /* Display information about each subsystem and each hierarchy */
4357 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4359 struct cgroup_subsys *ss;
4362 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4364 * ideally we don't want subsystems moving around while we do this.
4365 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4366 * subsys/hierarchy state.
4368 mutex_lock(&cgroup_mutex);
4370 for_each_subsys(ss, i)
4371 seq_printf(m, "%s\t%d\t%d\t%d\n",
4372 ss->name, ss->root->hierarchy_id,
4373 ss->root->number_of_cgroups, !ss->disabled);
4375 mutex_unlock(&cgroup_mutex);
4379 static int cgroupstats_open(struct inode *inode, struct file *file)
4381 return single_open(file, proc_cgroupstats_show, NULL);
4384 static const struct file_operations proc_cgroupstats_operations = {
4385 .open = cgroupstats_open,
4387 .llseek = seq_lseek,
4388 .release = single_release,
4392 * cgroup_fork - attach newly forked task to its parents cgroup.
4393 * @child: pointer to task_struct of forking parent process.
4395 * Description: A task inherits its parent's cgroup at fork().
4397 * A pointer to the shared css_set was automatically copied in
4398 * fork.c by dup_task_struct(). However, we ignore that copy, since
4399 * it was not made under the protection of RCU or cgroup_mutex, so
4400 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4401 * have already changed current->cgroups, allowing the previously
4402 * referenced cgroup group to be removed and freed.
4404 * At the point that cgroup_fork() is called, 'current' is the parent
4405 * task, and the passed argument 'child' points to the child task.
4407 void cgroup_fork(struct task_struct *child)
4410 get_css_set(task_css_set(current));
4411 child->cgroups = current->cgroups;
4412 task_unlock(current);
4413 INIT_LIST_HEAD(&child->cg_list);
4417 * cgroup_post_fork - called on a new task after adding it to the task list
4418 * @child: the task in question
4420 * Adds the task to the list running through its css_set if necessary and
4421 * call the subsystem fork() callbacks. Has to be after the task is
4422 * visible on the task list in case we race with the first call to
4423 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4426 void cgroup_post_fork(struct task_struct *child)
4428 struct cgroup_subsys *ss;
4432 * use_task_css_set_links is set to 1 before we walk the tasklist
4433 * under the tasklist_lock and we read it here after we added the child
4434 * to the tasklist under the tasklist_lock as well. If the child wasn't
4435 * yet in the tasklist when we walked through it from
4436 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4437 * should be visible now due to the paired locking and barriers implied
4438 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4439 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4442 if (use_task_css_set_links) {
4443 write_lock(&css_set_lock);
4445 if (list_empty(&child->cg_list))
4446 list_add(&child->cg_list, &task_css_set(child)->tasks);
4448 write_unlock(&css_set_lock);
4452 * Call ss->fork(). This must happen after @child is linked on
4453 * css_set; otherwise, @child might change state between ->fork()
4454 * and addition to css_set.
4456 if (need_forkexit_callback) {
4457 for_each_subsys(ss, i)
4464 * cgroup_exit - detach cgroup from exiting task
4465 * @tsk: pointer to task_struct of exiting process
4466 * @run_callback: run exit callbacks?
4468 * Description: Detach cgroup from @tsk and release it.
4470 * Note that cgroups marked notify_on_release force every task in
4471 * them to take the global cgroup_mutex mutex when exiting.
4472 * This could impact scaling on very large systems. Be reluctant to
4473 * use notify_on_release cgroups where very high task exit scaling
4474 * is required on large systems.
4476 * the_top_cgroup_hack:
4478 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4480 * We call cgroup_exit() while the task is still competent to
4481 * handle notify_on_release(), then leave the task attached to the
4482 * root cgroup in each hierarchy for the remainder of its exit.
4484 * To do this properly, we would increment the reference count on
4485 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4486 * code we would add a second cgroup function call, to drop that
4487 * reference. This would just create an unnecessary hot spot on
4488 * the top_cgroup reference count, to no avail.
4490 * Normally, holding a reference to a cgroup without bumping its
4491 * count is unsafe. The cgroup could go away, or someone could
4492 * attach us to a different cgroup, decrementing the count on
4493 * the first cgroup that we never incremented. But in this case,
4494 * top_cgroup isn't going away, and either task has PF_EXITING set,
4495 * which wards off any cgroup_attach_task() attempts, or task is a failed
4496 * fork, never visible to cgroup_attach_task.
4498 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4500 struct cgroup_subsys *ss;
4501 struct css_set *cset;
4505 * Unlink from the css_set task list if necessary.
4506 * Optimistically check cg_list before taking
4509 if (!list_empty(&tsk->cg_list)) {
4510 write_lock(&css_set_lock);
4511 if (!list_empty(&tsk->cg_list))
4512 list_del_init(&tsk->cg_list);
4513 write_unlock(&css_set_lock);
4516 /* Reassign the task to the init_css_set. */
4518 cset = task_css_set(tsk);
4519 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4521 if (run_callbacks && need_forkexit_callback) {
4522 /* see cgroup_post_fork() for details */
4523 for_each_subsys(ss, i) {
4525 struct cgroup_subsys_state *old_css = cset->subsys[i];
4526 struct cgroup_subsys_state *css = task_css(tsk, i);
4528 ss->exit(css, old_css, tsk);
4534 put_css_set_taskexit(cset);
4537 static void check_for_release(struct cgroup *cgrp)
4539 if (cgroup_is_releasable(cgrp) &&
4540 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4542 * Control Group is currently removeable. If it's not
4543 * already queued for a userspace notification, queue
4546 int need_schedule_work = 0;
4548 raw_spin_lock(&release_list_lock);
4549 if (!cgroup_is_dead(cgrp) &&
4550 list_empty(&cgrp->release_list)) {
4551 list_add(&cgrp->release_list, &release_list);
4552 need_schedule_work = 1;
4554 raw_spin_unlock(&release_list_lock);
4555 if (need_schedule_work)
4556 schedule_work(&release_agent_work);
4561 * Notify userspace when a cgroup is released, by running the
4562 * configured release agent with the name of the cgroup (path
4563 * relative to the root of cgroup file system) as the argument.
4565 * Most likely, this user command will try to rmdir this cgroup.
4567 * This races with the possibility that some other task will be
4568 * attached to this cgroup before it is removed, or that some other
4569 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4570 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4571 * unused, and this cgroup will be reprieved from its death sentence,
4572 * to continue to serve a useful existence. Next time it's released,
4573 * we will get notified again, if it still has 'notify_on_release' set.
4575 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4576 * means only wait until the task is successfully execve()'d. The
4577 * separate release agent task is forked by call_usermodehelper(),
4578 * then control in this thread returns here, without waiting for the
4579 * release agent task. We don't bother to wait because the caller of
4580 * this routine has no use for the exit status of the release agent
4581 * task, so no sense holding our caller up for that.
4583 static void cgroup_release_agent(struct work_struct *work)
4585 BUG_ON(work != &release_agent_work);
4586 mutex_lock(&cgroup_mutex);
4587 raw_spin_lock(&release_list_lock);
4588 while (!list_empty(&release_list)) {
4589 char *argv[3], *envp[3];
4591 char *pathbuf = NULL, *agentbuf = NULL;
4592 struct cgroup *cgrp = list_entry(release_list.next,
4595 list_del_init(&cgrp->release_list);
4596 raw_spin_unlock(&release_list_lock);
4597 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4600 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4602 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4607 argv[i++] = agentbuf;
4608 argv[i++] = pathbuf;
4612 /* minimal command environment */
4613 envp[i++] = "HOME=/";
4614 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4617 /* Drop the lock while we invoke the usermode helper,
4618 * since the exec could involve hitting disk and hence
4619 * be a slow process */
4620 mutex_unlock(&cgroup_mutex);
4621 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4622 mutex_lock(&cgroup_mutex);
4626 raw_spin_lock(&release_list_lock);
4628 raw_spin_unlock(&release_list_lock);
4629 mutex_unlock(&cgroup_mutex);
4632 static int __init cgroup_disable(char *str)
4634 struct cgroup_subsys *ss;
4638 while ((token = strsep(&str, ",")) != NULL) {
4642 for_each_subsys(ss, i) {
4643 if (!strcmp(token, ss->name)) {
4645 printk(KERN_INFO "Disabling %s control group"
4646 " subsystem\n", ss->name);
4653 __setup("cgroup_disable=", cgroup_disable);
4656 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
4657 * @dentry: directory dentry of interest
4658 * @ss: subsystem of interest
4660 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
4661 * to get the corresponding css and return it. If such css doesn't exist
4662 * or can't be pinned, an ERR_PTR value is returned.
4664 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
4665 struct cgroup_subsys *ss)
4667 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4668 struct cgroup_subsys_state *css = NULL;
4669 struct cgroup *cgrp;
4671 /* is @dentry a cgroup dir? */
4672 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4673 kernfs_type(kn) != KERNFS_DIR)
4674 return ERR_PTR(-EBADF);
4679 * This path doesn't originate from kernfs and @kn could already
4680 * have been or be removed at any point. @kn->priv is RCU
4681 * protected for this access. See destroy_locked() for details.
4683 cgrp = rcu_dereference(kn->priv);
4685 css = cgroup_css(cgrp, ss);
4687 if (!css || !css_tryget(css))
4688 css = ERR_PTR(-ENOENT);
4695 * css_from_id - lookup css by id
4696 * @id: the cgroup id
4697 * @ss: cgroup subsys to be looked into
4699 * Returns the css if there's valid one with @id, otherwise returns NULL.
4700 * Should be called under rcu_read_lock().
4702 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
4704 struct cgroup *cgrp;
4706 cgroup_assert_mutexes_or_rcu_locked();
4708 cgrp = idr_find(&ss->root->cgroup_idr, id);
4710 return cgroup_css(cgrp, ss);
4714 #ifdef CONFIG_CGROUP_DEBUG
4715 static struct cgroup_subsys_state *
4716 debug_css_alloc(struct cgroup_subsys_state *parent_css)
4718 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4721 return ERR_PTR(-ENOMEM);
4726 static void debug_css_free(struct cgroup_subsys_state *css)
4731 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
4734 return cgroup_task_count(css->cgroup);
4737 static u64 current_css_set_read(struct cgroup_subsys_state *css,
4740 return (u64)(unsigned long)current->cgroups;
4743 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
4749 count = atomic_read(&task_css_set(current)->refcount);
4754 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
4756 struct cgrp_cset_link *link;
4757 struct css_set *cset;
4759 read_lock(&css_set_lock);
4761 cset = rcu_dereference(current->cgroups);
4762 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
4763 struct cgroup *c = link->cgrp;
4764 const char *name = "?";
4766 if (c != cgroup_dummy_top)
4767 name = cgroup_name(c);
4769 seq_printf(seq, "Root %d group %s\n",
4770 c->root->hierarchy_id, name);
4773 read_unlock(&css_set_lock);
4777 #define MAX_TASKS_SHOWN_PER_CSS 25
4778 static int cgroup_css_links_read(struct seq_file *seq, void *v)
4780 struct cgroup_subsys_state *css = seq_css(seq);
4781 struct cgrp_cset_link *link;
4783 read_lock(&css_set_lock);
4784 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
4785 struct css_set *cset = link->cset;
4786 struct task_struct *task;
4788 seq_printf(seq, "css_set %p\n", cset);
4789 list_for_each_entry(task, &cset->tasks, cg_list) {
4790 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4791 seq_puts(seq, " ...\n");
4794 seq_printf(seq, " task %d\n",
4795 task_pid_vnr(task));
4799 read_unlock(&css_set_lock);
4803 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
4805 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4808 static struct cftype debug_files[] = {
4810 .name = "taskcount",
4811 .read_u64 = debug_taskcount_read,
4815 .name = "current_css_set",
4816 .read_u64 = current_css_set_read,
4820 .name = "current_css_set_refcount",
4821 .read_u64 = current_css_set_refcount_read,
4825 .name = "current_css_set_cg_links",
4826 .seq_show = current_css_set_cg_links_read,
4830 .name = "cgroup_css_links",
4831 .seq_show = cgroup_css_links_read,
4835 .name = "releasable",
4836 .read_u64 = releasable_read,
4842 struct cgroup_subsys debug_cgrp_subsys = {
4843 .css_alloc = debug_css_alloc,
4844 .css_free = debug_css_free,
4845 .base_cftypes = debug_files,
4847 #endif /* CONFIG_CGROUP_DEBUG */