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/backing-dev.h>
44 #include <linux/slab.h>
45 #include <linux/magic.h>
46 #include <linux/spinlock.h>
47 #include <linux/string.h>
48 #include <linux/sort.h>
49 #include <linux/kmod.h>
50 #include <linux/delayacct.h>
51 #include <linux/cgroupstats.h>
52 #include <linux/hashtable.h>
53 #include <linux/namei.h>
54 #include <linux/pid_namespace.h>
55 #include <linux/idr.h>
56 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
57 #include <linux/flex_array.h> /* used in cgroup_attach_task */
58 #include <linux/kthread.h>
60 #include <linux/atomic.h>
63 * pidlists linger the following amount before being destroyed. The goal
64 * is avoiding frequent destruction in the middle of consecutive read calls
65 * Expiring in the middle is a performance problem not a correctness one.
66 * 1 sec should be enough.
68 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
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_destroy_css_killed(struct cgroup *cgrp);
170 static int cgroup_destroy_locked(struct cgroup *cgrp);
171 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
173 static int cgroup_file_release(struct inode *inode, struct file *file);
174 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
177 * cgroup_css - obtain a cgroup's css for the specified subsystem
178 * @cgrp: the cgroup of interest
179 * @ss: the subsystem of interest (%NULL returns the dummy_css)
181 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
182 * function must be called either under cgroup_mutex or rcu_read_lock() and
183 * the caller is responsible for pinning the returned css if it wants to
184 * keep accessing it outside the said locks. This function may return
185 * %NULL if @cgrp doesn't have @subsys_id enabled.
187 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
188 struct cgroup_subsys *ss)
191 return rcu_dereference_check(cgrp->subsys[ss->id],
192 lockdep_is_held(&cgroup_tree_mutex) ||
193 lockdep_is_held(&cgroup_mutex));
195 return &cgrp->dummy_css;
198 /* convenient tests for these bits */
199 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
201 return test_bit(CGRP_DEAD, &cgrp->flags);
205 * cgroup_is_descendant - test ancestry
206 * @cgrp: the cgroup to be tested
207 * @ancestor: possible ancestor of @cgrp
209 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
210 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
211 * and @ancestor are accessible.
213 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
216 if (cgrp == ancestor)
222 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
224 static int cgroup_is_releasable(const struct cgroup *cgrp)
227 (1 << CGRP_RELEASABLE) |
228 (1 << CGRP_NOTIFY_ON_RELEASE);
229 return (cgrp->flags & bits) == bits;
232 static int notify_on_release(const struct cgroup *cgrp)
234 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
238 * for_each_css - iterate all css's of a cgroup
239 * @css: the iteration cursor
240 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
241 * @cgrp: the target cgroup to iterate css's of
243 * Should be called under cgroup_mutex.
245 #define for_each_css(css, ssid, cgrp) \
246 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
247 if (!((css) = rcu_dereference_check( \
248 (cgrp)->subsys[(ssid)], \
249 lockdep_is_held(&cgroup_tree_mutex) || \
250 lockdep_is_held(&cgroup_mutex)))) { } \
254 * for_each_subsys - iterate all enabled cgroup subsystems
255 * @ss: the iteration cursor
256 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
258 #define for_each_subsys(ss, ssid) \
259 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
260 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
262 /* iterate across the active hierarchies */
263 #define for_each_active_root(root) \
264 list_for_each_entry((root), &cgroup_roots, root_list)
266 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
268 return dentry->d_fsdata;
271 static inline struct cfent *__d_cfe(struct dentry *dentry)
273 return dentry->d_fsdata;
276 static inline struct cftype *__d_cft(struct dentry *dentry)
278 return __d_cfe(dentry)->type;
282 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
283 * @cgrp: the cgroup to be checked for liveness
285 * On success, returns true; the mutex should be later unlocked. On
286 * failure returns false with no lock held.
288 static bool cgroup_lock_live_group(struct cgroup *cgrp)
290 mutex_lock(&cgroup_mutex);
291 if (cgroup_is_dead(cgrp)) {
292 mutex_unlock(&cgroup_mutex);
298 /* the list of cgroups eligible for automatic release. Protected by
299 * release_list_lock */
300 static LIST_HEAD(release_list);
301 static DEFINE_RAW_SPINLOCK(release_list_lock);
302 static void cgroup_release_agent(struct work_struct *work);
303 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
304 static void check_for_release(struct cgroup *cgrp);
307 * A cgroup can be associated with multiple css_sets as different tasks may
308 * belong to different cgroups on different hierarchies. In the other
309 * direction, a css_set is naturally associated with multiple cgroups.
310 * This M:N relationship is represented by the following link structure
311 * which exists for each association and allows traversing the associations
314 struct cgrp_cset_link {
315 /* the cgroup and css_set this link associates */
317 struct css_set *cset;
319 /* list of cgrp_cset_links anchored at cgrp->cset_links */
320 struct list_head cset_link;
322 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
323 struct list_head cgrp_link;
326 /* The default css_set - used by init and its children prior to any
327 * hierarchies being mounted. It contains a pointer to the root state
328 * for each subsystem. Also used to anchor the list of css_sets. Not
329 * reference-counted, to improve performance when child cgroups
330 * haven't been created.
333 static struct css_set init_css_set;
334 static struct cgrp_cset_link init_cgrp_cset_link;
337 * css_set_lock protects the list of css_set objects, and the chain of
338 * tasks off each css_set. Nests outside task->alloc_lock due to
339 * css_task_iter_start().
341 static DEFINE_RWLOCK(css_set_lock);
342 static int css_set_count;
345 * hash table for cgroup groups. This improves the performance to find
346 * an existing css_set. This hash doesn't (currently) take into
347 * account cgroups in empty hierarchies.
349 #define CSS_SET_HASH_BITS 7
350 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
352 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
354 unsigned long key = 0UL;
355 struct cgroup_subsys *ss;
358 for_each_subsys(ss, i)
359 key += (unsigned long)css[i];
360 key = (key >> 16) ^ key;
366 * We don't maintain the lists running through each css_set to its task
367 * until after the first call to css_task_iter_start(). This reduces the
368 * fork()/exit() overhead for people who have cgroups compiled into their
369 * kernel but not actually in use.
371 static int use_task_css_set_links __read_mostly;
373 static void __put_css_set(struct css_set *cset, int taskexit)
375 struct cgrp_cset_link *link, *tmp_link;
378 * Ensure that the refcount doesn't hit zero while any readers
379 * can see it. Similar to atomic_dec_and_lock(), but for an
382 if (atomic_add_unless(&cset->refcount, -1, 1))
384 write_lock(&css_set_lock);
385 if (!atomic_dec_and_test(&cset->refcount)) {
386 write_unlock(&css_set_lock);
390 /* This css_set is dead. unlink it and release cgroup refcounts */
391 hash_del(&cset->hlist);
394 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
395 struct cgroup *cgrp = link->cgrp;
397 list_del(&link->cset_link);
398 list_del(&link->cgrp_link);
400 /* @cgrp can't go away while we're holding css_set_lock */
401 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
403 set_bit(CGRP_RELEASABLE, &cgrp->flags);
404 check_for_release(cgrp);
410 write_unlock(&css_set_lock);
411 kfree_rcu(cset, rcu_head);
415 * refcounted get/put for css_set objects
417 static inline void get_css_set(struct css_set *cset)
419 atomic_inc(&cset->refcount);
422 static inline void put_css_set(struct css_set *cset)
424 __put_css_set(cset, 0);
427 static inline void put_css_set_taskexit(struct css_set *cset)
429 __put_css_set(cset, 1);
433 * compare_css_sets - helper function for find_existing_css_set().
434 * @cset: candidate css_set being tested
435 * @old_cset: existing css_set for a task
436 * @new_cgrp: cgroup that's being entered by the task
437 * @template: desired set of css pointers in css_set (pre-calculated)
439 * Returns true if "cset" matches "old_cset" except for the hierarchy
440 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
442 static bool compare_css_sets(struct css_set *cset,
443 struct css_set *old_cset,
444 struct cgroup *new_cgrp,
445 struct cgroup_subsys_state *template[])
447 struct list_head *l1, *l2;
449 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
450 /* Not all subsystems matched */
455 * Compare cgroup pointers in order to distinguish between
456 * different cgroups in heirarchies with no subsystems. We
457 * could get by with just this check alone (and skip the
458 * memcmp above) but on most setups the memcmp check will
459 * avoid the need for this more expensive check on almost all
463 l1 = &cset->cgrp_links;
464 l2 = &old_cset->cgrp_links;
466 struct cgrp_cset_link *link1, *link2;
467 struct cgroup *cgrp1, *cgrp2;
471 /* See if we reached the end - both lists are equal length. */
472 if (l1 == &cset->cgrp_links) {
473 BUG_ON(l2 != &old_cset->cgrp_links);
476 BUG_ON(l2 == &old_cset->cgrp_links);
478 /* Locate the cgroups associated with these links. */
479 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
480 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
483 /* Hierarchies should be linked in the same order. */
484 BUG_ON(cgrp1->root != cgrp2->root);
487 * If this hierarchy is the hierarchy of the cgroup
488 * that's changing, then we need to check that this
489 * css_set points to the new cgroup; if it's any other
490 * hierarchy, then this css_set should point to the
491 * same cgroup as the old css_set.
493 if (cgrp1->root == new_cgrp->root) {
494 if (cgrp1 != new_cgrp)
505 * find_existing_css_set - init css array and find the matching css_set
506 * @old_cset: the css_set that we're using before the cgroup transition
507 * @cgrp: the cgroup that we're moving into
508 * @template: out param for the new set of csses, should be clear on entry
510 static struct css_set *find_existing_css_set(struct css_set *old_cset,
512 struct cgroup_subsys_state *template[])
514 struct cgroupfs_root *root = cgrp->root;
515 struct cgroup_subsys *ss;
516 struct css_set *cset;
521 * Build the set of subsystem state objects that we want to see in the
522 * new css_set. while subsystems can change globally, the entries here
523 * won't change, so no need for locking.
525 for_each_subsys(ss, i) {
526 if (root->subsys_mask & (1UL << i)) {
527 /* Subsystem is in this hierarchy. So we want
528 * the subsystem state from the new
530 template[i] = cgroup_css(cgrp, ss);
532 /* Subsystem is not in this hierarchy, so we
533 * don't want to change the subsystem state */
534 template[i] = old_cset->subsys[i];
538 key = css_set_hash(template);
539 hash_for_each_possible(css_set_table, cset, hlist, key) {
540 if (!compare_css_sets(cset, old_cset, cgrp, template))
543 /* This css_set matches what we need */
547 /* No existing cgroup group matched */
551 static void free_cgrp_cset_links(struct list_head *links_to_free)
553 struct cgrp_cset_link *link, *tmp_link;
555 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
556 list_del(&link->cset_link);
562 * allocate_cgrp_cset_links - allocate cgrp_cset_links
563 * @count: the number of links to allocate
564 * @tmp_links: list_head the allocated links are put on
566 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
567 * through ->cset_link. Returns 0 on success or -errno.
569 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
571 struct cgrp_cset_link *link;
574 INIT_LIST_HEAD(tmp_links);
576 for (i = 0; i < count; i++) {
577 link = kzalloc(sizeof(*link), GFP_KERNEL);
579 free_cgrp_cset_links(tmp_links);
582 list_add(&link->cset_link, tmp_links);
588 * link_css_set - a helper function to link a css_set to a cgroup
589 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
590 * @cset: the css_set to be linked
591 * @cgrp: the destination cgroup
593 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
596 struct cgrp_cset_link *link;
598 BUG_ON(list_empty(tmp_links));
599 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
602 list_move(&link->cset_link, &cgrp->cset_links);
604 * Always add links to the tail of the list so that the list
605 * is sorted by order of hierarchy creation
607 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
611 * find_css_set - return a new css_set with one cgroup updated
612 * @old_cset: the baseline css_set
613 * @cgrp: the cgroup to be updated
615 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
616 * substituted into the appropriate hierarchy.
618 static struct css_set *find_css_set(struct css_set *old_cset,
621 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
622 struct css_set *cset;
623 struct list_head tmp_links;
624 struct cgrp_cset_link *link;
627 lockdep_assert_held(&cgroup_mutex);
629 /* First see if we already have a cgroup group that matches
631 read_lock(&css_set_lock);
632 cset = find_existing_css_set(old_cset, cgrp, template);
635 read_unlock(&css_set_lock);
640 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
644 /* Allocate all the cgrp_cset_link objects that we'll need */
645 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
650 atomic_set(&cset->refcount, 1);
651 INIT_LIST_HEAD(&cset->cgrp_links);
652 INIT_LIST_HEAD(&cset->tasks);
653 INIT_HLIST_NODE(&cset->hlist);
655 /* Copy the set of subsystem state objects generated in
656 * find_existing_css_set() */
657 memcpy(cset->subsys, template, sizeof(cset->subsys));
659 write_lock(&css_set_lock);
660 /* Add reference counts and links from the new css_set. */
661 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
662 struct cgroup *c = link->cgrp;
664 if (c->root == cgrp->root)
666 link_css_set(&tmp_links, cset, c);
669 BUG_ON(!list_empty(&tmp_links));
673 /* Add this cgroup group to the hash table */
674 key = css_set_hash(cset->subsys);
675 hash_add(css_set_table, &cset->hlist, key);
677 write_unlock(&css_set_lock);
683 * Return the cgroup for "task" from the given hierarchy. Must be
684 * called with cgroup_mutex held.
686 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
687 struct cgroupfs_root *root)
689 struct css_set *cset;
690 struct cgroup *res = NULL;
692 BUG_ON(!mutex_is_locked(&cgroup_mutex));
693 read_lock(&css_set_lock);
695 * No need to lock the task - since we hold cgroup_mutex the
696 * task can't change groups, so the only thing that can happen
697 * is that it exits and its css is set back to init_css_set.
699 cset = task_css_set(task);
700 if (cset == &init_css_set) {
701 res = &root->top_cgroup;
703 struct cgrp_cset_link *link;
705 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
706 struct cgroup *c = link->cgrp;
708 if (c->root == root) {
714 read_unlock(&css_set_lock);
720 * There is one global cgroup mutex. We also require taking
721 * task_lock() when dereferencing a task's cgroup subsys pointers.
722 * See "The task_lock() exception", at the end of this comment.
724 * A task must hold cgroup_mutex to modify cgroups.
726 * Any task can increment and decrement the count field without lock.
727 * So in general, code holding cgroup_mutex can't rely on the count
728 * field not changing. However, if the count goes to zero, then only
729 * cgroup_attach_task() can increment it again. Because a count of zero
730 * means that no tasks are currently attached, therefore there is no
731 * way a task attached to that cgroup can fork (the other way to
732 * increment the count). So code holding cgroup_mutex can safely
733 * assume that if the count is zero, it will stay zero. Similarly, if
734 * a task holds cgroup_mutex on a cgroup with zero count, it
735 * knows that the cgroup won't be removed, as cgroup_rmdir()
738 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
739 * (usually) take cgroup_mutex. These are the two most performance
740 * critical pieces of code here. The exception occurs on cgroup_exit(),
741 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
742 * is taken, and if the cgroup count is zero, a usermode call made
743 * to the release agent with the name of the cgroup (path relative to
744 * the root of cgroup file system) as the argument.
746 * A cgroup can only be deleted if both its 'count' of using tasks
747 * is zero, and its list of 'children' cgroups is empty. Since all
748 * tasks in the system use _some_ cgroup, and since there is always at
749 * least one task in the system (init, pid == 1), therefore, top_cgroup
750 * always has either children cgroups and/or using tasks. So we don't
751 * need a special hack to ensure that top_cgroup cannot be deleted.
753 * The task_lock() exception
755 * The need for this exception arises from the action of
756 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
757 * another. It does so using cgroup_mutex, however there are
758 * several performance critical places that need to reference
759 * task->cgroup without the expense of grabbing a system global
760 * mutex. Therefore except as noted below, when dereferencing or, as
761 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
762 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
763 * the task_struct routinely used for such matters.
765 * P.S. One more locking exception. RCU is used to guard the
766 * update of a tasks cgroup pointer by cgroup_attach_task()
770 * A couple of forward declarations required, due to cyclic reference loop:
771 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
772 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
776 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode);
777 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
778 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
779 static const struct inode_operations cgroup_dir_inode_operations;
780 static const struct file_operations proc_cgroupstats_operations;
782 static struct backing_dev_info cgroup_backing_dev_info = {
784 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
787 static struct inode *cgroup_new_inode(umode_t mode, struct super_block *sb)
789 struct inode *inode = new_inode(sb);
792 inode->i_ino = get_next_ino();
793 inode->i_mode = mode;
794 inode->i_uid = current_fsuid();
795 inode->i_gid = current_fsgid();
796 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
797 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
802 static struct cgroup_name *cgroup_alloc_name(struct dentry *dentry)
804 struct cgroup_name *name;
806 name = kmalloc(sizeof(*name) + dentry->d_name.len + 1, GFP_KERNEL);
809 strcpy(name->name, dentry->d_name.name);
813 static void cgroup_free_fn(struct work_struct *work)
815 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
817 mutex_lock(&cgroup_mutex);
818 cgrp->root->number_of_cgroups--;
819 mutex_unlock(&cgroup_mutex);
822 * We get a ref to the parent's dentry, and put the ref when
823 * this cgroup is being freed, so it's guaranteed that the
824 * parent won't be destroyed before its children.
826 dput(cgrp->parent->dentry);
829 * Drop the active superblock reference that we took when we
830 * created the cgroup. This will free cgrp->root, if we are
831 * holding the last reference to @sb.
833 deactivate_super(cgrp->root->sb);
835 cgroup_pidlist_destroy_all(cgrp);
837 simple_xattrs_free(&cgrp->xattrs);
839 kfree(rcu_dereference_raw(cgrp->name));
843 static void cgroup_free_rcu(struct rcu_head *head)
845 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
847 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
848 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
851 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
853 /* is dentry a directory ? if so, kfree() associated cgroup */
854 if (S_ISDIR(inode->i_mode)) {
855 struct cgroup *cgrp = dentry->d_fsdata;
857 BUG_ON(!(cgroup_is_dead(cgrp)));
860 * XXX: cgrp->id is only used to look up css's. As cgroup
861 * and css's lifetimes will be decoupled, it should be made
862 * per-subsystem and moved to css->id so that lookups are
863 * successful until the target css is released.
865 mutex_lock(&cgroup_mutex);
866 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
867 mutex_unlock(&cgroup_mutex);
870 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
872 struct cfent *cfe = __d_cfe(dentry);
873 struct cgroup *cgrp = dentry->d_parent->d_fsdata;
875 WARN_ONCE(!list_empty(&cfe->node) &&
876 cgrp != &cgrp->root->top_cgroup,
877 "cfe still linked for %s\n", cfe->type->name);
878 simple_xattrs_free(&cfe->xattrs);
884 static void remove_dir(struct dentry *d)
886 struct dentry *parent = dget(d->d_parent);
889 simple_rmdir(parent->d_inode, d);
893 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
897 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
898 lockdep_assert_held(&cgroup_tree_mutex);
901 * If we're doing cleanup due to failure of cgroup_create(),
902 * the corresponding @cfe may not exist.
904 list_for_each_entry(cfe, &cgrp->files, node) {
905 struct dentry *d = cfe->dentry;
907 if (cft && cfe->type != cft)
912 simple_unlink(cgrp->dentry->d_inode, d);
913 list_del_init(&cfe->node);
921 * cgroup_clear_dir - remove subsys files in a cgroup directory
922 * @cgrp: target cgroup
923 * @subsys_mask: mask of the subsystem ids whose files should be removed
925 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
927 struct cgroup_subsys *ss;
930 for_each_subsys(ss, i) {
931 struct cftype_set *set;
933 if (!test_bit(i, &subsys_mask))
935 list_for_each_entry(set, &ss->cftsets, node)
936 cgroup_addrm_files(cgrp, set->cfts, false);
941 * NOTE : the dentry must have been dget()'ed
943 static void cgroup_d_remove_dir(struct dentry *dentry)
945 struct dentry *parent;
947 parent = dentry->d_parent;
948 spin_lock(&parent->d_lock);
949 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
950 list_del_init(&dentry->d_u.d_child);
951 spin_unlock(&dentry->d_lock);
952 spin_unlock(&parent->d_lock);
956 static int rebind_subsystems(struct cgroupfs_root *root,
957 unsigned long added_mask, unsigned removed_mask)
959 struct cgroup *cgrp = &root->top_cgroup;
960 struct cgroup_subsys *ss;
963 lockdep_assert_held(&cgroup_tree_mutex);
964 lockdep_assert_held(&cgroup_mutex);
966 /* Check that any added subsystems are currently free */
967 for_each_subsys(ss, i)
968 if ((added_mask & (1 << i)) && ss->root != &cgroup_dummy_root)
971 ret = cgroup_populate_dir(cgrp, added_mask);
976 * Nothing can fail from this point on. Remove files for the
977 * removed subsystems and rebind each subsystem.
979 mutex_unlock(&cgroup_mutex);
980 cgroup_clear_dir(cgrp, removed_mask);
981 mutex_lock(&cgroup_mutex);
983 for_each_subsys(ss, i) {
984 unsigned long bit = 1UL << i;
986 if (bit & added_mask) {
987 /* We're binding this subsystem to this hierarchy */
988 BUG_ON(cgroup_css(cgrp, ss));
989 BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
990 BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
992 rcu_assign_pointer(cgrp->subsys[i],
993 cgroup_css(cgroup_dummy_top, ss));
994 cgroup_css(cgrp, ss)->cgroup = cgrp;
998 ss->bind(cgroup_css(cgrp, ss));
1000 /* refcount was already taken, and we're keeping it */
1001 root->subsys_mask |= bit;
1002 } else if (bit & removed_mask) {
1003 /* We're removing this subsystem */
1004 BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1005 BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1008 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1010 cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1011 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1013 cgroup_subsys[i]->root = &cgroup_dummy_root;
1014 root->subsys_mask &= ~bit;
1019 * Mark @root has finished binding subsystems. @root->subsys_mask
1020 * now matches the bound subsystems.
1022 root->flags |= CGRP_ROOT_SUBSYS_BOUND;
1027 static int cgroup_show_options(struct seq_file *seq, struct dentry *dentry)
1029 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
1030 struct cgroup_subsys *ss;
1033 for_each_subsys(ss, ssid)
1034 if (root->subsys_mask & (1 << ssid))
1035 seq_printf(seq, ",%s", ss->name);
1036 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1037 seq_puts(seq, ",sane_behavior");
1038 if (root->flags & CGRP_ROOT_NOPREFIX)
1039 seq_puts(seq, ",noprefix");
1040 if (root->flags & CGRP_ROOT_XATTR)
1041 seq_puts(seq, ",xattr");
1043 spin_lock(&release_agent_path_lock);
1044 if (strlen(root->release_agent_path))
1045 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1046 spin_unlock(&release_agent_path_lock);
1048 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1049 seq_puts(seq, ",clone_children");
1050 if (strlen(root->name))
1051 seq_printf(seq, ",name=%s", root->name);
1055 struct cgroup_sb_opts {
1056 unsigned long subsys_mask;
1057 unsigned long flags;
1058 char *release_agent;
1059 bool cpuset_clone_children;
1061 /* User explicitly requested empty subsystem */
1064 struct cgroupfs_root *new_root;
1069 * Convert a hierarchy specifier into a bitmask of subsystems and
1070 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1071 * array. This function takes refcounts on subsystems to be used, unless it
1072 * returns error, in which case no refcounts are taken.
1074 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1076 char *token, *o = data;
1077 bool all_ss = false, one_ss = false;
1078 unsigned long mask = (unsigned long)-1;
1079 struct cgroup_subsys *ss;
1082 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1084 #ifdef CONFIG_CPUSETS
1085 mask = ~(1UL << cpuset_cgrp_id);
1088 memset(opts, 0, sizeof(*opts));
1090 while ((token = strsep(&o, ",")) != NULL) {
1093 if (!strcmp(token, "none")) {
1094 /* Explicitly have no subsystems */
1098 if (!strcmp(token, "all")) {
1099 /* Mutually exclusive option 'all' + subsystem name */
1105 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1106 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1109 if (!strcmp(token, "noprefix")) {
1110 opts->flags |= CGRP_ROOT_NOPREFIX;
1113 if (!strcmp(token, "clone_children")) {
1114 opts->cpuset_clone_children = true;
1117 if (!strcmp(token, "xattr")) {
1118 opts->flags |= CGRP_ROOT_XATTR;
1121 if (!strncmp(token, "release_agent=", 14)) {
1122 /* Specifying two release agents is forbidden */
1123 if (opts->release_agent)
1125 opts->release_agent =
1126 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1127 if (!opts->release_agent)
1131 if (!strncmp(token, "name=", 5)) {
1132 const char *name = token + 5;
1133 /* Can't specify an empty name */
1136 /* Must match [\w.-]+ */
1137 for (i = 0; i < strlen(name); i++) {
1141 if ((c == '.') || (c == '-') || (c == '_'))
1145 /* Specifying two names is forbidden */
1148 opts->name = kstrndup(name,
1149 MAX_CGROUP_ROOT_NAMELEN - 1,
1157 for_each_subsys(ss, i) {
1158 if (strcmp(token, ss->name))
1163 /* Mutually exclusive option 'all' + subsystem name */
1166 set_bit(i, &opts->subsys_mask);
1171 if (i == CGROUP_SUBSYS_COUNT)
1176 * If the 'all' option was specified select all the subsystems,
1177 * otherwise if 'none', 'name=' and a subsystem name options
1178 * were not specified, let's default to 'all'
1180 if (all_ss || (!one_ss && !opts->none && !opts->name))
1181 for_each_subsys(ss, i)
1183 set_bit(i, &opts->subsys_mask);
1185 /* Consistency checks */
1187 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1188 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1190 if (opts->flags & CGRP_ROOT_NOPREFIX) {
1191 pr_err("cgroup: sane_behavior: noprefix is not allowed\n");
1195 if (opts->cpuset_clone_children) {
1196 pr_err("cgroup: sane_behavior: clone_children is not allowed\n");
1202 * Option noprefix was introduced just for backward compatibility
1203 * with the old cpuset, so we allow noprefix only if mounting just
1204 * the cpuset subsystem.
1206 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1210 /* Can't specify "none" and some subsystems */
1211 if (opts->subsys_mask && opts->none)
1215 * We either have to specify by name or by subsystems. (So all
1216 * empty hierarchies must have a name).
1218 if (!opts->subsys_mask && !opts->name)
1224 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1227 struct cgroupfs_root *root = sb->s_fs_info;
1228 struct cgroup *cgrp = &root->top_cgroup;
1229 struct cgroup_sb_opts opts;
1230 unsigned long added_mask, removed_mask;
1232 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1233 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1237 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1238 mutex_lock(&cgroup_tree_mutex);
1239 mutex_lock(&cgroup_mutex);
1241 /* See what subsystems are wanted */
1242 ret = parse_cgroupfs_options(data, &opts);
1246 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1247 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1248 task_tgid_nr(current), current->comm);
1250 added_mask = opts.subsys_mask & ~root->subsys_mask;
1251 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1253 /* Don't allow flags or name to change at remount */
1254 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1255 (opts.name && strcmp(opts.name, root->name))) {
1256 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1257 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1258 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1263 /* remounting is not allowed for populated hierarchies */
1264 if (root->number_of_cgroups > 1) {
1269 ret = rebind_subsystems(root, added_mask, removed_mask);
1273 if (opts.release_agent) {
1274 spin_lock(&release_agent_path_lock);
1275 strcpy(root->release_agent_path, opts.release_agent);
1276 spin_unlock(&release_agent_path_lock);
1279 kfree(opts.release_agent);
1281 mutex_unlock(&cgroup_mutex);
1282 mutex_unlock(&cgroup_tree_mutex);
1283 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1287 static const struct super_operations cgroup_ops = {
1288 .statfs = simple_statfs,
1289 .drop_inode = generic_delete_inode,
1290 .show_options = cgroup_show_options,
1291 .remount_fs = cgroup_remount,
1294 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1296 INIT_LIST_HEAD(&cgrp->sibling);
1297 INIT_LIST_HEAD(&cgrp->children);
1298 INIT_LIST_HEAD(&cgrp->files);
1299 INIT_LIST_HEAD(&cgrp->cset_links);
1300 INIT_LIST_HEAD(&cgrp->release_list);
1301 INIT_LIST_HEAD(&cgrp->pidlists);
1302 mutex_init(&cgrp->pidlist_mutex);
1303 cgrp->dummy_css.cgroup = cgrp;
1304 simple_xattrs_init(&cgrp->xattrs);
1307 static void init_cgroup_root(struct cgroupfs_root *root)
1309 struct cgroup *cgrp = &root->top_cgroup;
1311 INIT_LIST_HEAD(&root->root_list);
1312 root->number_of_cgroups = 1;
1314 RCU_INIT_POINTER(cgrp->name, &root_cgroup_name);
1315 init_cgroup_housekeeping(cgrp);
1316 idr_init(&root->cgroup_idr);
1319 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
1323 lockdep_assert_held(&cgroup_mutex);
1325 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
1330 root->hierarchy_id = id;
1334 static void cgroup_exit_root_id(struct cgroupfs_root *root)
1336 lockdep_assert_held(&cgroup_mutex);
1338 if (root->hierarchy_id) {
1339 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1340 root->hierarchy_id = 0;
1344 static int cgroup_test_super(struct super_block *sb, void *data)
1346 struct cgroup_sb_opts *opts = data;
1347 struct cgroupfs_root *root = sb->s_fs_info;
1349 /* If we asked for a name then it must match */
1350 if (opts->name && strcmp(opts->name, root->name))
1354 * If we asked for subsystems (or explicitly for no
1355 * subsystems) then they must match
1357 if ((opts->subsys_mask || opts->none)
1358 && (opts->subsys_mask != root->subsys_mask))
1364 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1366 struct cgroupfs_root *root;
1368 if (!opts->subsys_mask && !opts->none)
1371 root = kzalloc(sizeof(*root), GFP_KERNEL);
1373 return ERR_PTR(-ENOMEM);
1375 init_cgroup_root(root);
1378 * We need to set @root->subsys_mask now so that @root can be
1379 * matched by cgroup_test_super() before it finishes
1380 * initialization; otherwise, competing mounts with the same
1381 * options may try to bind the same subsystems instead of waiting
1382 * for the first one leading to unexpected mount errors.
1383 * SUBSYS_BOUND will be set once actual binding is complete.
1385 root->subsys_mask = opts->subsys_mask;
1386 root->flags = opts->flags;
1387 if (opts->release_agent)
1388 strcpy(root->release_agent_path, opts->release_agent);
1390 strcpy(root->name, opts->name);
1391 if (opts->cpuset_clone_children)
1392 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1396 static void cgroup_free_root(struct cgroupfs_root *root)
1399 /* hierarhcy ID shoulid already have been released */
1400 WARN_ON_ONCE(root->hierarchy_id);
1402 idr_destroy(&root->cgroup_idr);
1407 static int cgroup_set_super(struct super_block *sb, void *data)
1410 struct cgroup_sb_opts *opts = data;
1412 /* If we don't have a new root, we can't set up a new sb */
1413 if (!opts->new_root)
1416 BUG_ON(!opts->subsys_mask && !opts->none);
1418 ret = set_anon_super(sb, NULL);
1422 sb->s_fs_info = opts->new_root;
1423 opts->new_root->sb = sb;
1425 sb->s_blocksize = PAGE_CACHE_SIZE;
1426 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1427 sb->s_magic = CGROUP_SUPER_MAGIC;
1428 sb->s_op = &cgroup_ops;
1433 static int cgroup_get_rootdir(struct super_block *sb)
1435 static const struct dentry_operations cgroup_dops = {
1436 .d_iput = cgroup_diput,
1437 .d_delete = always_delete_dentry,
1440 struct inode *inode =
1441 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1446 inode->i_fop = &simple_dir_operations;
1447 inode->i_op = &cgroup_dir_inode_operations;
1448 /* directories start off with i_nlink == 2 (for "." entry) */
1450 sb->s_root = d_make_root(inode);
1453 /* for everything else we want ->d_op set */
1454 sb->s_d_op = &cgroup_dops;
1458 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1459 int flags, const char *unused_dev_name,
1462 struct cgroup_sb_opts opts;
1463 struct cgroupfs_root *root;
1465 struct super_block *sb;
1466 struct cgroupfs_root *new_root;
1467 struct list_head tmp_links;
1468 struct inode *inode;
1469 const struct cred *cred;
1471 /* First find the desired set of subsystems */
1472 mutex_lock(&cgroup_mutex);
1473 ret = parse_cgroupfs_options(data, &opts);
1474 mutex_unlock(&cgroup_mutex);
1479 * Allocate a new cgroup root. We may not need it if we're
1480 * reusing an existing hierarchy.
1482 new_root = cgroup_root_from_opts(&opts);
1483 if (IS_ERR(new_root)) {
1484 ret = PTR_ERR(new_root);
1487 opts.new_root = new_root;
1489 /* Locate an existing or new sb for this hierarchy */
1490 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, 0, &opts);
1493 cgroup_free_root(opts.new_root);
1497 root = sb->s_fs_info;
1499 if (root == opts.new_root) {
1500 /* We used the new root structure, so this is a new hierarchy */
1501 struct cgroup *root_cgrp = &root->top_cgroup;
1502 struct cgroupfs_root *existing_root;
1504 struct css_set *cset;
1506 BUG_ON(sb->s_root != NULL);
1508 ret = cgroup_get_rootdir(sb);
1510 goto drop_new_super;
1511 inode = sb->s_root->d_inode;
1513 mutex_lock(&inode->i_mutex);
1514 mutex_lock(&cgroup_tree_mutex);
1515 mutex_lock(&cgroup_mutex);
1517 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1520 root_cgrp->id = ret;
1522 /* Check for name clashes with existing mounts */
1524 if (strlen(root->name))
1525 for_each_active_root(existing_root)
1526 if (!strcmp(existing_root->name, root->name))
1530 * We're accessing css_set_count without locking
1531 * css_set_lock here, but that's OK - it can only be
1532 * increased by someone holding cgroup_lock, and
1533 * that's us. The worst that can happen is that we
1534 * have some link structures left over
1536 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1540 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1541 ret = cgroup_init_root_id(root, 2, 0);
1545 sb->s_root->d_fsdata = root_cgrp;
1546 root_cgrp->dentry = sb->s_root;
1549 * We're inside get_sb() and will call lookup_one_len() to
1550 * create the root files, which doesn't work if SELinux is
1551 * in use. The following cred dancing somehow works around
1552 * it. See 2ce9738ba ("cgroupfs: use init_cred when
1553 * populating new cgroupfs mount") for more details.
1555 cred = override_creds(&init_cred);
1557 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1561 ret = rebind_subsystems(root, root->subsys_mask, 0);
1568 * There must be no failure case after here, since rebinding
1569 * takes care of subsystems' refcounts, which are explicitly
1570 * dropped in the failure exit path.
1573 list_add(&root->root_list, &cgroup_roots);
1574 cgroup_root_count++;
1576 /* Link the top cgroup in this hierarchy into all
1577 * the css_set objects */
1578 write_lock(&css_set_lock);
1579 hash_for_each(css_set_table, i, cset, hlist)
1580 link_css_set(&tmp_links, cset, root_cgrp);
1581 write_unlock(&css_set_lock);
1583 free_cgrp_cset_links(&tmp_links);
1585 BUG_ON(!list_empty(&root_cgrp->children));
1586 BUG_ON(root->number_of_cgroups != 1);
1588 mutex_unlock(&cgroup_mutex);
1589 mutex_unlock(&cgroup_tree_mutex);
1590 mutex_unlock(&inode->i_mutex);
1593 * We re-used an existing hierarchy - the new root (if
1594 * any) is not needed
1596 cgroup_free_root(opts.new_root);
1598 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1599 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1600 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1602 goto drop_new_super;
1604 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1609 kfree(opts.release_agent);
1611 return dget(sb->s_root);
1614 free_cgrp_cset_links(&tmp_links);
1615 cgroup_addrm_files(&root->top_cgroup, cgroup_base_files, false);
1618 cgroup_exit_root_id(root);
1619 mutex_unlock(&cgroup_mutex);
1620 mutex_unlock(&cgroup_tree_mutex);
1621 mutex_unlock(&inode->i_mutex);
1623 deactivate_locked_super(sb);
1625 kfree(opts.release_agent);
1627 return ERR_PTR(ret);
1630 static void cgroup_kill_sb(struct super_block *sb)
1632 struct cgroupfs_root *root = sb->s_fs_info;
1633 struct cgroup *cgrp = &root->top_cgroup;
1634 struct cgrp_cset_link *link, *tmp_link;
1639 BUG_ON(root->number_of_cgroups != 1);
1640 BUG_ON(!list_empty(&cgrp->children));
1642 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1643 mutex_lock(&cgroup_tree_mutex);
1644 mutex_lock(&cgroup_mutex);
1646 /* Rebind all subsystems back to the default hierarchy */
1647 if (root->flags & CGRP_ROOT_SUBSYS_BOUND) {
1648 ret = rebind_subsystems(root, 0, root->subsys_mask);
1649 /* Shouldn't be able to fail ... */
1654 * Release all the links from cset_links to this hierarchy's
1657 write_lock(&css_set_lock);
1659 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1660 list_del(&link->cset_link);
1661 list_del(&link->cgrp_link);
1664 write_unlock(&css_set_lock);
1666 if (!list_empty(&root->root_list)) {
1667 list_del(&root->root_list);
1668 cgroup_root_count--;
1671 cgroup_exit_root_id(root);
1673 mutex_unlock(&cgroup_mutex);
1674 mutex_unlock(&cgroup_tree_mutex);
1675 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1677 simple_xattrs_free(&cgrp->xattrs);
1679 kill_litter_super(sb);
1680 cgroup_free_root(root);
1683 static struct file_system_type cgroup_fs_type = {
1685 .mount = cgroup_mount,
1686 .kill_sb = cgroup_kill_sb,
1689 static struct kobject *cgroup_kobj;
1692 * cgroup_path - generate the path of a cgroup
1693 * @cgrp: the cgroup in question
1694 * @buf: the buffer to write the path into
1695 * @buflen: the length of the buffer
1697 * Writes path of cgroup into buf. Returns 0 on success, -errno on error.
1699 * We can't generate cgroup path using dentry->d_name, as accessing
1700 * dentry->name must be protected by irq-unsafe dentry->d_lock or parent
1701 * inode's i_mutex, while on the other hand cgroup_path() can be called
1702 * with some irq-safe spinlocks held.
1704 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1706 int ret = -ENAMETOOLONG;
1709 if (!cgrp->parent) {
1710 if (strlcpy(buf, "/", buflen) >= buflen)
1711 return -ENAMETOOLONG;
1715 start = buf + buflen - 1;
1720 const char *name = cgroup_name(cgrp);
1724 if ((start -= len) < buf)
1726 memcpy(start, name, len);
1732 cgrp = cgrp->parent;
1733 } while (cgrp->parent);
1735 memmove(buf, start, buf + buflen - start);
1740 EXPORT_SYMBOL_GPL(cgroup_path);
1743 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1744 * @task: target task
1745 * @buf: the buffer to write the path into
1746 * @buflen: the length of the buffer
1748 * Determine @task's cgroup on the first (the one with the lowest non-zero
1749 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1750 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1751 * cgroup controller callbacks.
1753 * Returns 0 on success, fails with -%ENAMETOOLONG if @buflen is too short.
1755 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1757 struct cgroupfs_root *root;
1758 struct cgroup *cgrp;
1759 int hierarchy_id = 1, ret = 0;
1762 return -ENAMETOOLONG;
1764 mutex_lock(&cgroup_mutex);
1766 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1769 cgrp = task_cgroup_from_root(task, root);
1770 ret = cgroup_path(cgrp, buf, buflen);
1772 /* if no hierarchy exists, everyone is in "/" */
1773 memcpy(buf, "/", 2);
1776 mutex_unlock(&cgroup_mutex);
1779 EXPORT_SYMBOL_GPL(task_cgroup_path);
1782 * Control Group taskset
1784 struct task_and_cgroup {
1785 struct task_struct *task;
1786 struct cgroup *cgrp;
1787 struct css_set *cset;
1790 struct cgroup_taskset {
1791 struct task_and_cgroup single;
1792 struct flex_array *tc_array;
1795 struct cgroup *cur_cgrp;
1799 * cgroup_taskset_first - reset taskset and return the first task
1800 * @tset: taskset of interest
1802 * @tset iteration is initialized and the first task is returned.
1804 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1806 if (tset->tc_array) {
1808 return cgroup_taskset_next(tset);
1810 tset->cur_cgrp = tset->single.cgrp;
1811 return tset->single.task;
1814 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1817 * cgroup_taskset_next - iterate to the next task in taskset
1818 * @tset: taskset of interest
1820 * Return the next task in @tset. Iteration must have been initialized
1821 * with cgroup_taskset_first().
1823 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1825 struct task_and_cgroup *tc;
1827 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1830 tc = flex_array_get(tset->tc_array, tset->idx++);
1831 tset->cur_cgrp = tc->cgrp;
1834 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1837 * cgroup_taskset_cur_css - return the matching css for the current task
1838 * @tset: taskset of interest
1839 * @subsys_id: the ID of the target subsystem
1841 * Return the css for the current (last returned) task of @tset for
1842 * subsystem specified by @subsys_id. This function must be preceded by
1843 * either cgroup_taskset_first() or cgroup_taskset_next().
1845 struct cgroup_subsys_state *cgroup_taskset_cur_css(struct cgroup_taskset *tset,
1848 return cgroup_css(tset->cur_cgrp, cgroup_subsys[subsys_id]);
1850 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_css);
1853 * cgroup_taskset_size - return the number of tasks in taskset
1854 * @tset: taskset of interest
1856 int cgroup_taskset_size(struct cgroup_taskset *tset)
1858 return tset->tc_array ? tset->tc_array_len : 1;
1860 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1864 * cgroup_task_migrate - move a task from one cgroup to another.
1866 * Must be called with cgroup_mutex and threadgroup locked.
1868 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1869 struct task_struct *tsk,
1870 struct css_set *new_cset)
1872 struct css_set *old_cset;
1875 * We are synchronized through threadgroup_lock() against PF_EXITING
1876 * setting such that we can't race against cgroup_exit() changing the
1877 * css_set to init_css_set and dropping the old one.
1879 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1880 old_cset = task_css_set(tsk);
1883 rcu_assign_pointer(tsk->cgroups, new_cset);
1886 /* Update the css_set linked lists if we're using them */
1887 write_lock(&css_set_lock);
1888 if (!list_empty(&tsk->cg_list))
1889 list_move(&tsk->cg_list, &new_cset->tasks);
1890 write_unlock(&css_set_lock);
1893 * We just gained a reference on old_cset by taking it from the
1894 * task. As trading it for new_cset is protected by cgroup_mutex,
1895 * we're safe to drop it here; it will be freed under RCU.
1897 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1898 put_css_set(old_cset);
1902 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1903 * @cgrp: the cgroup to attach to
1904 * @tsk: the task or the leader of the threadgroup to be attached
1905 * @threadgroup: attach the whole threadgroup?
1907 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1908 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1910 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1913 int retval, i, group_size;
1914 struct cgroupfs_root *root = cgrp->root;
1915 struct cgroup_subsys_state *css, *failed_css = NULL;
1916 /* threadgroup list cursor and array */
1917 struct task_struct *leader = tsk;
1918 struct task_and_cgroup *tc;
1919 struct flex_array *group;
1920 struct cgroup_taskset tset = { };
1923 * step 0: in order to do expensive, possibly blocking operations for
1924 * every thread, we cannot iterate the thread group list, since it needs
1925 * rcu or tasklist locked. instead, build an array of all threads in the
1926 * group - group_rwsem prevents new threads from appearing, and if
1927 * threads exit, this will just be an over-estimate.
1930 group_size = get_nr_threads(tsk);
1933 /* flex_array supports very large thread-groups better than kmalloc. */
1934 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1937 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1938 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1940 goto out_free_group_list;
1944 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1945 * already PF_EXITING could be freed from underneath us unless we
1946 * take an rcu_read_lock.
1950 struct task_and_cgroup ent;
1952 /* @tsk either already exited or can't exit until the end */
1953 if (tsk->flags & PF_EXITING)
1956 /* as per above, nr_threads may decrease, but not increase. */
1957 BUG_ON(i >= group_size);
1959 ent.cgrp = task_cgroup_from_root(tsk, root);
1960 /* nothing to do if this task is already in the cgroup */
1961 if (ent.cgrp == cgrp)
1964 * saying GFP_ATOMIC has no effect here because we did prealloc
1965 * earlier, but it's good form to communicate our expectations.
1967 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
1968 BUG_ON(retval != 0);
1973 } while_each_thread(leader, tsk);
1975 /* remember the number of threads in the array for later. */
1977 tset.tc_array = group;
1978 tset.tc_array_len = group_size;
1980 /* methods shouldn't be called if no task is actually migrating */
1983 goto out_free_group_list;
1986 * step 1: check that we can legitimately attach to the cgroup.
1988 for_each_css(css, i, cgrp) {
1989 if (css->ss->can_attach) {
1990 retval = css->ss->can_attach(css, &tset);
1993 goto out_cancel_attach;
1999 * step 2: make sure css_sets exist for all threads to be migrated.
2000 * we use find_css_set, which allocates a new one if necessary.
2002 for (i = 0; i < group_size; i++) {
2003 struct css_set *old_cset;
2005 tc = flex_array_get(group, i);
2006 old_cset = task_css_set(tc->task);
2007 tc->cset = find_css_set(old_cset, cgrp);
2010 goto out_put_css_set_refs;
2015 * step 3: now that we're guaranteed success wrt the css_sets,
2016 * proceed to move all tasks to the new cgroup. There are no
2017 * failure cases after here, so this is the commit point.
2019 for (i = 0; i < group_size; i++) {
2020 tc = flex_array_get(group, i);
2021 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
2023 /* nothing is sensitive to fork() after this point. */
2026 * step 4: do subsystem attach callbacks.
2028 for_each_css(css, i, cgrp)
2029 if (css->ss->attach)
2030 css->ss->attach(css, &tset);
2033 * step 5: success! and cleanup
2036 out_put_css_set_refs:
2038 for (i = 0; i < group_size; i++) {
2039 tc = flex_array_get(group, i);
2042 put_css_set(tc->cset);
2047 for_each_css(css, i, cgrp) {
2048 if (css == failed_css)
2050 if (css->ss->cancel_attach)
2051 css->ss->cancel_attach(css, &tset);
2054 out_free_group_list:
2055 flex_array_free(group);
2060 * Find the task_struct of the task to attach by vpid and pass it along to the
2061 * function to attach either it or all tasks in its threadgroup. Will lock
2062 * cgroup_mutex and threadgroup; may take task_lock of task.
2064 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2066 struct task_struct *tsk;
2067 const struct cred *cred = current_cred(), *tcred;
2070 if (!cgroup_lock_live_group(cgrp))
2076 tsk = find_task_by_vpid(pid);
2080 goto out_unlock_cgroup;
2083 * even if we're attaching all tasks in the thread group, we
2084 * only need to check permissions on one of them.
2086 tcred = __task_cred(tsk);
2087 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2088 !uid_eq(cred->euid, tcred->uid) &&
2089 !uid_eq(cred->euid, tcred->suid)) {
2092 goto out_unlock_cgroup;
2098 tsk = tsk->group_leader;
2101 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2102 * trapped in a cpuset, or RT worker may be born in a cgroup
2103 * with no rt_runtime allocated. Just say no.
2105 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2108 goto out_unlock_cgroup;
2111 get_task_struct(tsk);
2114 threadgroup_lock(tsk);
2116 if (!thread_group_leader(tsk)) {
2118 * a race with de_thread from another thread's exec()
2119 * may strip us of our leadership, if this happens,
2120 * there is no choice but to throw this task away and
2121 * try again; this is
2122 * "double-double-toil-and-trouble-check locking".
2124 threadgroup_unlock(tsk);
2125 put_task_struct(tsk);
2126 goto retry_find_task;
2130 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2132 threadgroup_unlock(tsk);
2134 put_task_struct(tsk);
2136 mutex_unlock(&cgroup_mutex);
2141 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2142 * @from: attach to all cgroups of a given task
2143 * @tsk: the task to be attached
2145 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2147 struct cgroupfs_root *root;
2150 mutex_lock(&cgroup_mutex);
2151 for_each_active_root(root) {
2152 struct cgroup *from_cgrp = task_cgroup_from_root(from, root);
2154 retval = cgroup_attach_task(from_cgrp, tsk, false);
2158 mutex_unlock(&cgroup_mutex);
2162 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2164 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2165 struct cftype *cft, u64 pid)
2167 return attach_task_by_pid(css->cgroup, pid, false);
2170 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2171 struct cftype *cft, u64 tgid)
2173 return attach_task_by_pid(css->cgroup, tgid, true);
2176 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2177 struct cftype *cft, const char *buffer)
2179 BUILD_BUG_ON(sizeof(css->cgroup->root->release_agent_path) < PATH_MAX);
2180 if (strlen(buffer) >= PATH_MAX)
2182 if (!cgroup_lock_live_group(css->cgroup))
2184 spin_lock(&release_agent_path_lock);
2185 strcpy(css->cgroup->root->release_agent_path, buffer);
2186 spin_unlock(&release_agent_path_lock);
2187 mutex_unlock(&cgroup_mutex);
2191 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2193 struct cgroup *cgrp = seq_css(seq)->cgroup;
2195 if (!cgroup_lock_live_group(cgrp))
2197 seq_puts(seq, cgrp->root->release_agent_path);
2198 seq_putc(seq, '\n');
2199 mutex_unlock(&cgroup_mutex);
2203 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2205 struct cgroup *cgrp = seq_css(seq)->cgroup;
2207 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2211 /* A buffer size big enough for numbers or short strings */
2212 #define CGROUP_LOCAL_BUFFER_SIZE 64
2214 static ssize_t cgroup_file_write(struct file *file, const char __user *userbuf,
2215 size_t nbytes, loff_t *ppos)
2217 struct cfent *cfe = __d_cfe(file->f_dentry);
2218 struct cftype *cft = __d_cft(file->f_dentry);
2219 struct cgroup_subsys_state *css = cfe->css;
2220 size_t max_bytes = cft->max_write_len ?: CGROUP_LOCAL_BUFFER_SIZE - 1;
2224 if (nbytes >= max_bytes)
2227 buf = kmalloc(nbytes + 1, GFP_KERNEL);
2231 if (copy_from_user(buf, userbuf, nbytes)) {
2238 if (cft->write_string) {
2239 ret = cft->write_string(css, cft, strstrip(buf));
2240 } else if (cft->write_u64) {
2241 unsigned long long v;
2242 ret = kstrtoull(buf, 0, &v);
2244 ret = cft->write_u64(css, cft, v);
2245 } else if (cft->write_s64) {
2247 ret = kstrtoll(buf, 0, &v);
2249 ret = cft->write_s64(css, cft, v);
2250 } else if (cft->trigger) {
2251 ret = cft->trigger(css, (unsigned int)cft->private);
2257 return ret ?: nbytes;
2261 * seqfile ops/methods for returning structured data. Currently just
2262 * supports string->u64 maps, but can be extended in future.
2265 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2267 struct cftype *cft = seq_cft(seq);
2269 if (cft->seq_start) {
2270 return cft->seq_start(seq, ppos);
2273 * The same behavior and code as single_open(). Returns
2274 * !NULL if pos is at the beginning; otherwise, NULL.
2276 return NULL + !*ppos;
2280 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2282 struct cftype *cft = seq_cft(seq);
2284 if (cft->seq_next) {
2285 return cft->seq_next(seq, v, ppos);
2288 * The same behavior and code as single_open(), always
2289 * terminate after the initial read.
2296 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2298 struct cftype *cft = seq_cft(seq);
2301 cft->seq_stop(seq, v);
2304 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2306 struct cftype *cft = seq_cft(m);
2307 struct cgroup_subsys_state *css = seq_css(m);
2310 return cft->seq_show(m, arg);
2313 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2314 else if (cft->read_s64)
2315 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2321 static struct seq_operations cgroup_seq_operations = {
2322 .start = cgroup_seqfile_start,
2323 .next = cgroup_seqfile_next,
2324 .stop = cgroup_seqfile_stop,
2325 .show = cgroup_seqfile_show,
2328 static int cgroup_file_open(struct inode *inode, struct file *file)
2330 struct cfent *cfe = __d_cfe(file->f_dentry);
2331 struct cftype *cft = __d_cft(file->f_dentry);
2332 struct cgroup *cgrp = __d_cgrp(cfe->dentry->d_parent);
2333 struct cgroup_subsys_state *css;
2334 struct cgroup_open_file *of;
2337 err = generic_file_open(inode, file);
2342 * If the file belongs to a subsystem, pin the css. Will be
2343 * unpinned either on open failure or release. This ensures that
2344 * @css stays alive for all file operations.
2347 css = cgroup_css(cgrp, cft->ss);
2348 if (cft->ss && !css_tryget(css))
2356 * @cfe->css is used by read/write/close to determine the
2357 * associated css. @file->private_data would be a better place but
2358 * that's already used by seqfile. Multiple accessors may use it
2359 * simultaneously which is okay as the association never changes.
2361 WARN_ON_ONCE(cfe->css && cfe->css != css);
2364 of = __seq_open_private(file, &cgroup_seq_operations,
2365 sizeof(struct cgroup_open_file));
2376 static int cgroup_file_release(struct inode *inode, struct file *file)
2378 struct cfent *cfe = __d_cfe(file->f_dentry);
2379 struct cgroup_subsys_state *css = cfe->css;
2383 return seq_release_private(inode, file);
2387 * cgroup_rename - Only allow simple rename of directories in place.
2389 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2390 struct inode *new_dir, struct dentry *new_dentry)
2393 struct cgroup_name *name, *old_name;
2394 struct cgroup *cgrp;
2397 * It's convinient to use parent dir's i_mutex to protected
2400 lockdep_assert_held(&old_dir->i_mutex);
2402 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2404 if (new_dentry->d_inode)
2406 if (old_dir != new_dir)
2409 cgrp = __d_cgrp(old_dentry);
2412 * This isn't a proper migration and its usefulness is very
2413 * limited. Disallow if sane_behavior.
2415 if (cgroup_sane_behavior(cgrp))
2418 name = cgroup_alloc_name(new_dentry);
2422 ret = simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2428 old_name = rcu_dereference_protected(cgrp->name, true);
2429 rcu_assign_pointer(cgrp->name, name);
2431 kfree_rcu(old_name, rcu_head);
2435 static struct simple_xattrs *__d_xattrs(struct dentry *dentry)
2437 if (S_ISDIR(dentry->d_inode->i_mode))
2438 return &__d_cgrp(dentry)->xattrs;
2440 return &__d_cfe(dentry)->xattrs;
2443 static inline int xattr_enabled(struct dentry *dentry)
2445 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
2446 return root->flags & CGRP_ROOT_XATTR;
2449 static bool is_valid_xattr(const char *name)
2451 if (!strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
2452 !strncmp(name, XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN))
2457 static int cgroup_setxattr(struct dentry *dentry, const char *name,
2458 const void *val, size_t size, int flags)
2460 if (!xattr_enabled(dentry))
2462 if (!is_valid_xattr(name))
2464 return simple_xattr_set(__d_xattrs(dentry), name, val, size, flags);
2467 static int cgroup_removexattr(struct dentry *dentry, const char *name)
2469 if (!xattr_enabled(dentry))
2471 if (!is_valid_xattr(name))
2473 return simple_xattr_remove(__d_xattrs(dentry), name);
2476 static ssize_t cgroup_getxattr(struct dentry *dentry, const char *name,
2477 void *buf, size_t size)
2479 if (!xattr_enabled(dentry))
2481 if (!is_valid_xattr(name))
2483 return simple_xattr_get(__d_xattrs(dentry), name, buf, size);
2486 static ssize_t cgroup_listxattr(struct dentry *dentry, char *buf, size_t size)
2488 if (!xattr_enabled(dentry))
2490 return simple_xattr_list(__d_xattrs(dentry), buf, size);
2493 static const struct file_operations cgroup_file_operations = {
2495 .write = cgroup_file_write,
2496 .llseek = generic_file_llseek,
2497 .open = cgroup_file_open,
2498 .release = cgroup_file_release,
2501 static const struct inode_operations cgroup_file_inode_operations = {
2502 .setxattr = cgroup_setxattr,
2503 .getxattr = cgroup_getxattr,
2504 .listxattr = cgroup_listxattr,
2505 .removexattr = cgroup_removexattr,
2508 static const struct inode_operations cgroup_dir_inode_operations = {
2509 .lookup = simple_lookup,
2510 .mkdir = cgroup_mkdir,
2511 .rmdir = cgroup_rmdir,
2512 .rename = cgroup_rename,
2513 .setxattr = cgroup_setxattr,
2514 .getxattr = cgroup_getxattr,
2515 .listxattr = cgroup_listxattr,
2516 .removexattr = cgroup_removexattr,
2519 static int cgroup_create_file(struct dentry *dentry, umode_t mode,
2520 struct super_block *sb)
2522 struct inode *inode;
2526 if (dentry->d_inode)
2529 inode = cgroup_new_inode(mode, sb);
2533 if (S_ISDIR(mode)) {
2534 inode->i_op = &cgroup_dir_inode_operations;
2535 inode->i_fop = &simple_dir_operations;
2537 /* start off with i_nlink == 2 (for "." entry) */
2539 inc_nlink(dentry->d_parent->d_inode);
2542 * Control reaches here with cgroup_mutex held.
2543 * @inode->i_mutex should nest outside cgroup_mutex but we
2544 * want to populate it immediately without releasing
2545 * cgroup_mutex. As @inode isn't visible to anyone else
2546 * yet, trylock will always succeed without affecting
2549 WARN_ON_ONCE(!mutex_trylock(&inode->i_mutex));
2550 } else if (S_ISREG(mode)) {
2552 inode->i_fop = &cgroup_file_operations;
2553 inode->i_op = &cgroup_file_inode_operations;
2555 d_instantiate(dentry, inode);
2556 dget(dentry); /* Extra count - pin the dentry in core */
2561 * cgroup_file_mode - deduce file mode of a control file
2562 * @cft: the control file in question
2564 * returns cft->mode if ->mode is not 0
2565 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2566 * returns S_IRUGO if it has only a read handler
2567 * returns S_IWUSR if it has only a write hander
2569 static umode_t cgroup_file_mode(const struct cftype *cft)
2576 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
2579 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
2586 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2588 struct dentry *dir = cgrp->dentry;
2589 struct cgroup *parent = __d_cgrp(dir);
2590 struct dentry *dentry;
2594 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2596 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
2597 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
2598 strcpy(name, cft->ss->name);
2601 strcat(name, cft->name);
2603 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2605 cfe = kzalloc(sizeof(*cfe), GFP_KERNEL);
2609 dentry = lookup_one_len(name, dir, strlen(name));
2610 if (IS_ERR(dentry)) {
2611 error = PTR_ERR(dentry);
2615 cfe->type = (void *)cft;
2616 cfe->dentry = dentry;
2617 dentry->d_fsdata = cfe;
2618 simple_xattrs_init(&cfe->xattrs);
2620 mode = cgroup_file_mode(cft);
2621 error = cgroup_create_file(dentry, mode | S_IFREG, cgrp->root->sb);
2623 list_add_tail(&cfe->node, &parent->files);
2633 * cgroup_addrm_files - add or remove files to a cgroup directory
2634 * @cgrp: the target cgroup
2635 * @cfts: array of cftypes to be added
2636 * @is_add: whether to add or remove
2638 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2639 * For removals, this function never fails. If addition fails, this
2640 * function doesn't remove files already added. The caller is responsible
2643 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2649 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
2650 lockdep_assert_held(&cgroup_tree_mutex);
2652 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2653 /* does cft->flags tell us to skip this file on @cgrp? */
2654 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2656 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2658 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2662 ret = cgroup_add_file(cgrp, cft);
2664 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2669 cgroup_rm_file(cgrp, cft);
2675 static void cgroup_cfts_prepare(void)
2676 __acquires(&cgroup_mutex)
2679 * Thanks to the entanglement with vfs inode locking, we can't walk
2680 * the existing cgroups under cgroup_mutex and create files.
2681 * Instead, we use css_for_each_descendant_pre() and drop RCU read
2682 * lock before calling cgroup_addrm_files().
2684 mutex_lock(&cgroup_tree_mutex);
2685 mutex_lock(&cgroup_mutex);
2688 static int cgroup_cfts_commit(struct cftype *cfts, bool is_add)
2689 __releases(&cgroup_mutex)
2692 struct cgroup_subsys *ss = cfts[0].ss;
2693 struct cgroup *root = &ss->root->top_cgroup;
2694 struct super_block *sb = ss->root->sb;
2695 struct dentry *prev = NULL;
2696 struct inode *inode;
2697 struct cgroup_subsys_state *css;
2701 mutex_unlock(&cgroup_mutex);
2703 /* %NULL @cfts indicates abort and don't bother if @ss isn't attached */
2704 if (!cfts || ss->root == &cgroup_dummy_root ||
2705 !atomic_inc_not_zero(&sb->s_active)) {
2706 mutex_unlock(&cgroup_tree_mutex);
2711 * All cgroups which are created after we drop cgroup_mutex will
2712 * have the updated set of files, so we only need to update the
2713 * cgroups created before the current @cgroup_serial_nr_next.
2715 update_before = cgroup_serial_nr_next;
2717 /* add/rm files for all cgroups created before */
2718 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2719 struct cgroup *cgrp = css->cgroup;
2721 if (cgroup_is_dead(cgrp))
2724 inode = cgrp->dentry->d_inode;
2727 prev = cgrp->dentry;
2729 mutex_unlock(&cgroup_tree_mutex);
2730 mutex_lock(&inode->i_mutex);
2731 mutex_lock(&cgroup_tree_mutex);
2732 if (cgrp->serial_nr < update_before && !cgroup_is_dead(cgrp))
2733 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2734 mutex_unlock(&inode->i_mutex);
2738 mutex_unlock(&cgroup_tree_mutex);
2740 deactivate_super(sb);
2745 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2746 * @ss: target cgroup subsystem
2747 * @cfts: zero-length name terminated array of cftypes
2749 * Register @cfts to @ss. Files described by @cfts are created for all
2750 * existing cgroups to which @ss is attached and all future cgroups will
2751 * have them too. This function can be called anytime whether @ss is
2754 * Returns 0 on successful registration, -errno on failure. Note that this
2755 * function currently returns 0 as long as @cfts registration is successful
2756 * even if some file creation attempts on existing cgroups fail.
2758 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2760 struct cftype_set *set;
2764 set = kzalloc(sizeof(*set), GFP_KERNEL);
2768 for (cft = cfts; cft->name[0] != '\0'; cft++)
2771 cgroup_cfts_prepare();
2773 list_add_tail(&set->node, &ss->cftsets);
2774 ret = cgroup_cfts_commit(cfts, true);
2776 cgroup_rm_cftypes(cfts);
2779 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2782 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2783 * @cfts: zero-length name terminated array of cftypes
2785 * Unregister @cfts. Files described by @cfts are removed from all
2786 * existing cgroups and all future cgroups won't have them either. This
2787 * function can be called anytime whether @cfts' subsys is attached or not.
2789 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2792 int cgroup_rm_cftypes(struct cftype *cfts)
2794 struct cftype_set *set;
2796 if (!cfts || !cfts[0].ss)
2799 cgroup_cfts_prepare();
2801 list_for_each_entry(set, &cfts[0].ss->cftsets, node) {
2802 if (set->cfts == cfts) {
2803 list_del(&set->node);
2805 cgroup_cfts_commit(cfts, false);
2810 cgroup_cfts_commit(NULL, false);
2815 * cgroup_task_count - count the number of tasks in a cgroup.
2816 * @cgrp: the cgroup in question
2818 * Return the number of tasks in the cgroup.
2820 int cgroup_task_count(const struct cgroup *cgrp)
2823 struct cgrp_cset_link *link;
2825 read_lock(&css_set_lock);
2826 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2827 count += atomic_read(&link->cset->refcount);
2828 read_unlock(&css_set_lock);
2833 * To reduce the fork() overhead for systems that are not actually using
2834 * their cgroups capability, we don't maintain the lists running through
2835 * each css_set to its tasks until we see the list actually used - in other
2836 * words after the first call to css_task_iter_start().
2838 static void cgroup_enable_task_cg_lists(void)
2840 struct task_struct *p, *g;
2841 write_lock(&css_set_lock);
2842 use_task_css_set_links = 1;
2844 * We need tasklist_lock because RCU is not safe against
2845 * while_each_thread(). Besides, a forking task that has passed
2846 * cgroup_post_fork() without seeing use_task_css_set_links = 1
2847 * is not guaranteed to have its child immediately visible in the
2848 * tasklist if we walk through it with RCU.
2850 read_lock(&tasklist_lock);
2851 do_each_thread(g, p) {
2854 * We should check if the process is exiting, otherwise
2855 * it will race with cgroup_exit() in that the list
2856 * entry won't be deleted though the process has exited.
2858 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2859 list_add(&p->cg_list, &task_css_set(p)->tasks);
2861 } while_each_thread(g, p);
2862 read_unlock(&tasklist_lock);
2863 write_unlock(&css_set_lock);
2867 * css_next_child - find the next child of a given css
2868 * @pos_css: the current position (%NULL to initiate traversal)
2869 * @parent_css: css whose children to walk
2871 * This function returns the next child of @parent_css and should be called
2872 * under either cgroup_mutex or RCU read lock. The only requirement is
2873 * that @parent_css and @pos_css are accessible. The next sibling is
2874 * guaranteed to be returned regardless of their states.
2876 struct cgroup_subsys_state *
2877 css_next_child(struct cgroup_subsys_state *pos_css,
2878 struct cgroup_subsys_state *parent_css)
2880 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2881 struct cgroup *cgrp = parent_css->cgroup;
2882 struct cgroup *next;
2884 cgroup_assert_mutexes_or_rcu_locked();
2887 * @pos could already have been removed. Once a cgroup is removed,
2888 * its ->sibling.next is no longer updated when its next sibling
2889 * changes. As CGRP_DEAD assertion is serialized and happens
2890 * before the cgroup is taken off the ->sibling list, if we see it
2891 * unasserted, it's guaranteed that the next sibling hasn't
2892 * finished its grace period even if it's already removed, and thus
2893 * safe to dereference from this RCU critical section. If
2894 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2895 * to be visible as %true here.
2897 * If @pos is dead, its next pointer can't be dereferenced;
2898 * however, as each cgroup is given a monotonically increasing
2899 * unique serial number and always appended to the sibling list,
2900 * the next one can be found by walking the parent's children until
2901 * we see a cgroup with higher serial number than @pos's. While
2902 * this path can be slower, it's taken only when either the current
2903 * cgroup is removed or iteration and removal race.
2906 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2907 } else if (likely(!cgroup_is_dead(pos))) {
2908 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2910 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2911 if (next->serial_nr > pos->serial_nr)
2915 if (&next->sibling == &cgrp->children)
2918 return cgroup_css(next, parent_css->ss);
2920 EXPORT_SYMBOL_GPL(css_next_child);
2923 * css_next_descendant_pre - find the next descendant for pre-order walk
2924 * @pos: the current position (%NULL to initiate traversal)
2925 * @root: css whose descendants to walk
2927 * To be used by css_for_each_descendant_pre(). Find the next descendant
2928 * to visit for pre-order traversal of @root's descendants. @root is
2929 * included in the iteration and the first node to be visited.
2931 * While this function requires cgroup_mutex or RCU read locking, it
2932 * doesn't require the whole traversal to be contained in a single critical
2933 * section. This function will return the correct next descendant as long
2934 * as both @pos and @root are accessible and @pos is a descendant of @root.
2936 struct cgroup_subsys_state *
2937 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2938 struct cgroup_subsys_state *root)
2940 struct cgroup_subsys_state *next;
2942 cgroup_assert_mutexes_or_rcu_locked();
2944 /* if first iteration, visit @root */
2948 /* visit the first child if exists */
2949 next = css_next_child(NULL, pos);
2953 /* no child, visit my or the closest ancestor's next sibling */
2954 while (pos != root) {
2955 next = css_next_child(pos, css_parent(pos));
2958 pos = css_parent(pos);
2963 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
2966 * css_rightmost_descendant - return the rightmost descendant of a css
2967 * @pos: css of interest
2969 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2970 * is returned. This can be used during pre-order traversal to skip
2973 * While this function requires cgroup_mutex or RCU read locking, it
2974 * doesn't require the whole traversal to be contained in a single critical
2975 * section. This function will return the correct rightmost descendant as
2976 * long as @pos is accessible.
2978 struct cgroup_subsys_state *
2979 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2981 struct cgroup_subsys_state *last, *tmp;
2983 cgroup_assert_mutexes_or_rcu_locked();
2987 /* ->prev isn't RCU safe, walk ->next till the end */
2989 css_for_each_child(tmp, last)
2995 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
2997 static struct cgroup_subsys_state *
2998 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3000 struct cgroup_subsys_state *last;
3004 pos = css_next_child(NULL, pos);
3011 * css_next_descendant_post - find the next descendant for post-order walk
3012 * @pos: the current position (%NULL to initiate traversal)
3013 * @root: css whose descendants to walk
3015 * To be used by css_for_each_descendant_post(). Find the next descendant
3016 * to visit for post-order traversal of @root's descendants. @root is
3017 * included in the iteration and the last node to be visited.
3019 * While this function requires cgroup_mutex or RCU read locking, it
3020 * doesn't require the whole traversal to be contained in a single critical
3021 * section. This function will return the correct next descendant as long
3022 * as both @pos and @cgroup are accessible and @pos is a descendant of
3025 struct cgroup_subsys_state *
3026 css_next_descendant_post(struct cgroup_subsys_state *pos,
3027 struct cgroup_subsys_state *root)
3029 struct cgroup_subsys_state *next;
3031 cgroup_assert_mutexes_or_rcu_locked();
3033 /* if first iteration, visit leftmost descendant which may be @root */
3035 return css_leftmost_descendant(root);
3037 /* if we visited @root, we're done */
3041 /* if there's an unvisited sibling, visit its leftmost descendant */
3042 next = css_next_child(pos, css_parent(pos));
3044 return css_leftmost_descendant(next);
3046 /* no sibling left, visit parent */
3047 return css_parent(pos);
3049 EXPORT_SYMBOL_GPL(css_next_descendant_post);
3052 * css_advance_task_iter - advance a task itererator to the next css_set
3053 * @it: the iterator to advance
3055 * Advance @it to the next css_set to walk.
3057 static void css_advance_task_iter(struct css_task_iter *it)
3059 struct list_head *l = it->cset_link;
3060 struct cgrp_cset_link *link;
3061 struct css_set *cset;
3063 /* Advance to the next non-empty css_set */
3066 if (l == &it->origin_css->cgroup->cset_links) {
3067 it->cset_link = NULL;
3070 link = list_entry(l, struct cgrp_cset_link, cset_link);
3072 } while (list_empty(&cset->tasks));
3074 it->task = cset->tasks.next;
3078 * css_task_iter_start - initiate task iteration
3079 * @css: the css to walk tasks of
3080 * @it: the task iterator to use
3082 * Initiate iteration through the tasks of @css. The caller can call
3083 * css_task_iter_next() to walk through the tasks until the function
3084 * returns NULL. On completion of iteration, css_task_iter_end() must be
3087 * Note that this function acquires a lock which is released when the
3088 * iteration finishes. The caller can't sleep while iteration is in
3091 void css_task_iter_start(struct cgroup_subsys_state *css,
3092 struct css_task_iter *it)
3093 __acquires(css_set_lock)
3096 * The first time anyone tries to iterate across a css, we need to
3097 * enable the list linking each css_set to its tasks, and fix up
3098 * all existing tasks.
3100 if (!use_task_css_set_links)
3101 cgroup_enable_task_cg_lists();
3103 read_lock(&css_set_lock);
3105 it->origin_css = css;
3106 it->cset_link = &css->cgroup->cset_links;
3108 css_advance_task_iter(it);
3112 * css_task_iter_next - return the next task for the iterator
3113 * @it: the task iterator being iterated
3115 * The "next" function for task iteration. @it should have been
3116 * initialized via css_task_iter_start(). Returns NULL when the iteration
3119 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3121 struct task_struct *res;
3122 struct list_head *l = it->task;
3123 struct cgrp_cset_link *link;
3125 /* If the iterator cg is NULL, we have no tasks */
3128 res = list_entry(l, struct task_struct, cg_list);
3129 /* Advance iterator to find next entry */
3131 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
3132 if (l == &link->cset->tasks) {
3134 * We reached the end of this task list - move on to the
3135 * next cgrp_cset_link.
3137 css_advance_task_iter(it);
3145 * css_task_iter_end - finish task iteration
3146 * @it: the task iterator to finish
3148 * Finish task iteration started by css_task_iter_start().
3150 void css_task_iter_end(struct css_task_iter *it)
3151 __releases(css_set_lock)
3153 read_unlock(&css_set_lock);
3156 static inline int started_after_time(struct task_struct *t1,
3157 struct timespec *time,
3158 struct task_struct *t2)
3160 int start_diff = timespec_compare(&t1->start_time, time);
3161 if (start_diff > 0) {
3163 } else if (start_diff < 0) {
3167 * Arbitrarily, if two processes started at the same
3168 * time, we'll say that the lower pointer value
3169 * started first. Note that t2 may have exited by now
3170 * so this may not be a valid pointer any longer, but
3171 * that's fine - it still serves to distinguish
3172 * between two tasks started (effectively) simultaneously.
3179 * This function is a callback from heap_insert() and is used to order
3181 * In this case we order the heap in descending task start time.
3183 static inline int started_after(void *p1, void *p2)
3185 struct task_struct *t1 = p1;
3186 struct task_struct *t2 = p2;
3187 return started_after_time(t1, &t2->start_time, t2);
3191 * css_scan_tasks - iterate though all the tasks in a css
3192 * @css: the css to iterate tasks of
3193 * @test: optional test callback
3194 * @process: process callback
3195 * @data: data passed to @test and @process
3196 * @heap: optional pre-allocated heap used for task iteration
3198 * Iterate through all the tasks in @css, calling @test for each, and if it
3199 * returns %true, call @process for it also.
3201 * @test may be NULL, meaning always true (select all tasks), which
3202 * effectively duplicates css_task_iter_{start,next,end}() but does not
3203 * lock css_set_lock for the call to @process.
3205 * It is guaranteed that @process will act on every task that is a member
3206 * of @css for the duration of this call. This function may or may not
3207 * call @process for tasks that exit or move to a different css during the
3208 * call, or are forked or move into the css during the call.
3210 * Note that @test may be called with locks held, and may in some
3211 * situations be called multiple times for the same task, so it should be
3214 * If @heap is non-NULL, a heap has been pre-allocated and will be used for
3215 * heap operations (and its "gt" member will be overwritten), else a
3216 * temporary heap will be used (allocation of which may cause this function
3219 int css_scan_tasks(struct cgroup_subsys_state *css,
3220 bool (*test)(struct task_struct *, void *),
3221 void (*process)(struct task_struct *, void *),
3222 void *data, struct ptr_heap *heap)
3225 struct css_task_iter it;
3226 struct task_struct *p, *dropped;
3227 /* Never dereference latest_task, since it's not refcounted */
3228 struct task_struct *latest_task = NULL;
3229 struct ptr_heap tmp_heap;
3230 struct timespec latest_time = { 0, 0 };
3233 /* The caller supplied our heap and pre-allocated its memory */
3234 heap->gt = &started_after;
3236 /* We need to allocate our own heap memory */
3238 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
3240 /* cannot allocate the heap */
3246 * Scan tasks in the css, using the @test callback to determine
3247 * which are of interest, and invoking @process callback on the
3248 * ones which need an update. Since we don't want to hold any
3249 * locks during the task updates, gather tasks to be processed in a
3250 * heap structure. The heap is sorted by descending task start
3251 * time. If the statically-sized heap fills up, we overflow tasks
3252 * that started later, and in future iterations only consider tasks
3253 * that started after the latest task in the previous pass. This
3254 * guarantees forward progress and that we don't miss any tasks.
3257 css_task_iter_start(css, &it);
3258 while ((p = css_task_iter_next(&it))) {
3260 * Only affect tasks that qualify per the caller's callback,
3261 * if he provided one
3263 if (test && !test(p, data))
3266 * Only process tasks that started after the last task
3269 if (!started_after_time(p, &latest_time, latest_task))
3271 dropped = heap_insert(heap, p);
3272 if (dropped == NULL) {
3274 * The new task was inserted; the heap wasn't
3278 } else if (dropped != p) {
3280 * The new task was inserted, and pushed out a
3284 put_task_struct(dropped);
3287 * Else the new task was newer than anything already in
3288 * the heap and wasn't inserted
3291 css_task_iter_end(&it);
3294 for (i = 0; i < heap->size; i++) {
3295 struct task_struct *q = heap->ptrs[i];
3297 latest_time = q->start_time;
3300 /* Process the task per the caller's callback */
3305 * If we had to process any tasks at all, scan again
3306 * in case some of them were in the middle of forking
3307 * children that didn't get processed.
3308 * Not the most efficient way to do it, but it avoids
3309 * having to take callback_mutex in the fork path
3313 if (heap == &tmp_heap)
3314 heap_free(&tmp_heap);
3318 static void cgroup_transfer_one_task(struct task_struct *task, void *data)
3320 struct cgroup *new_cgroup = data;
3322 mutex_lock(&cgroup_mutex);
3323 cgroup_attach_task(new_cgroup, task, false);
3324 mutex_unlock(&cgroup_mutex);
3328 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3329 * @to: cgroup to which the tasks will be moved
3330 * @from: cgroup in which the tasks currently reside
3332 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3334 return css_scan_tasks(&from->dummy_css, NULL, cgroup_transfer_one_task,
3339 * Stuff for reading the 'tasks'/'procs' files.
3341 * Reading this file can return large amounts of data if a cgroup has
3342 * *lots* of attached tasks. So it may need several calls to read(),
3343 * but we cannot guarantee that the information we produce is correct
3344 * unless we produce it entirely atomically.
3348 /* which pidlist file are we talking about? */
3349 enum cgroup_filetype {
3355 * A pidlist is a list of pids that virtually represents the contents of one
3356 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3357 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3360 struct cgroup_pidlist {
3362 * used to find which pidlist is wanted. doesn't change as long as
3363 * this particular list stays in the list.
3365 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3368 /* how many elements the above list has */
3370 /* each of these stored in a list by its cgroup */
3371 struct list_head links;
3372 /* pointer to the cgroup we belong to, for list removal purposes */
3373 struct cgroup *owner;
3374 /* for delayed destruction */
3375 struct delayed_work destroy_dwork;
3379 * The following two functions "fix" the issue where there are more pids
3380 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3381 * TODO: replace with a kernel-wide solution to this problem
3383 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3384 static void *pidlist_allocate(int count)
3386 if (PIDLIST_TOO_LARGE(count))
3387 return vmalloc(count * sizeof(pid_t));
3389 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3392 static void pidlist_free(void *p)
3394 if (is_vmalloc_addr(p))
3401 * Used to destroy all pidlists lingering waiting for destroy timer. None
3402 * should be left afterwards.
3404 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3406 struct cgroup_pidlist *l, *tmp_l;
3408 mutex_lock(&cgrp->pidlist_mutex);
3409 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3410 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3411 mutex_unlock(&cgrp->pidlist_mutex);
3413 flush_workqueue(cgroup_pidlist_destroy_wq);
3414 BUG_ON(!list_empty(&cgrp->pidlists));
3417 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3419 struct delayed_work *dwork = to_delayed_work(work);
3420 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3422 struct cgroup_pidlist *tofree = NULL;
3424 mutex_lock(&l->owner->pidlist_mutex);
3427 * Destroy iff we didn't get queued again. The state won't change
3428 * as destroy_dwork can only be queued while locked.
3430 if (!delayed_work_pending(dwork)) {
3431 list_del(&l->links);
3432 pidlist_free(l->list);
3433 put_pid_ns(l->key.ns);
3437 mutex_unlock(&l->owner->pidlist_mutex);
3442 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3443 * Returns the number of unique elements.
3445 static int pidlist_uniq(pid_t *list, int length)
3450 * we presume the 0th element is unique, so i starts at 1. trivial
3451 * edge cases first; no work needs to be done for either
3453 if (length == 0 || length == 1)
3455 /* src and dest walk down the list; dest counts unique elements */
3456 for (src = 1; src < length; src++) {
3457 /* find next unique element */
3458 while (list[src] == list[src-1]) {
3463 /* dest always points to where the next unique element goes */
3464 list[dest] = list[src];
3472 * The two pid files - task and cgroup.procs - guaranteed that the result
3473 * is sorted, which forced this whole pidlist fiasco. As pid order is
3474 * different per namespace, each namespace needs differently sorted list,
3475 * making it impossible to use, for example, single rbtree of member tasks
3476 * sorted by task pointer. As pidlists can be fairly large, allocating one
3477 * per open file is dangerous, so cgroup had to implement shared pool of
3478 * pidlists keyed by cgroup and namespace.
3480 * All this extra complexity was caused by the original implementation
3481 * committing to an entirely unnecessary property. In the long term, we
3482 * want to do away with it. Explicitly scramble sort order if
3483 * sane_behavior so that no such expectation exists in the new interface.
3485 * Scrambling is done by swapping every two consecutive bits, which is
3486 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3488 static pid_t pid_fry(pid_t pid)
3490 unsigned a = pid & 0x55555555;
3491 unsigned b = pid & 0xAAAAAAAA;
3493 return (a << 1) | (b >> 1);
3496 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3498 if (cgroup_sane_behavior(cgrp))
3499 return pid_fry(pid);
3504 static int cmppid(const void *a, const void *b)
3506 return *(pid_t *)a - *(pid_t *)b;
3509 static int fried_cmppid(const void *a, const void *b)
3511 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3514 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3515 enum cgroup_filetype type)
3517 struct cgroup_pidlist *l;
3518 /* don't need task_nsproxy() if we're looking at ourself */
3519 struct pid_namespace *ns = task_active_pid_ns(current);
3521 lockdep_assert_held(&cgrp->pidlist_mutex);
3523 list_for_each_entry(l, &cgrp->pidlists, links)
3524 if (l->key.type == type && l->key.ns == ns)
3530 * find the appropriate pidlist for our purpose (given procs vs tasks)
3531 * returns with the lock on that pidlist already held, and takes care
3532 * of the use count, or returns NULL with no locks held if we're out of
3535 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3536 enum cgroup_filetype type)
3538 struct cgroup_pidlist *l;
3540 lockdep_assert_held(&cgrp->pidlist_mutex);
3542 l = cgroup_pidlist_find(cgrp, type);
3546 /* entry not found; create a new one */
3547 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3551 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3553 /* don't need task_nsproxy() if we're looking at ourself */
3554 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3556 list_add(&l->links, &cgrp->pidlists);
3561 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3563 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3564 struct cgroup_pidlist **lp)
3568 int pid, n = 0; /* used for populating the array */
3569 struct css_task_iter it;
3570 struct task_struct *tsk;
3571 struct cgroup_pidlist *l;
3573 lockdep_assert_held(&cgrp->pidlist_mutex);
3576 * If cgroup gets more users after we read count, we won't have
3577 * enough space - tough. This race is indistinguishable to the
3578 * caller from the case that the additional cgroup users didn't
3579 * show up until sometime later on.
3581 length = cgroup_task_count(cgrp);
3582 array = pidlist_allocate(length);
3585 /* now, populate the array */
3586 css_task_iter_start(&cgrp->dummy_css, &it);
3587 while ((tsk = css_task_iter_next(&it))) {
3588 if (unlikely(n == length))
3590 /* get tgid or pid for procs or tasks file respectively */
3591 if (type == CGROUP_FILE_PROCS)
3592 pid = task_tgid_vnr(tsk);
3594 pid = task_pid_vnr(tsk);
3595 if (pid > 0) /* make sure to only use valid results */
3598 css_task_iter_end(&it);
3600 /* now sort & (if procs) strip out duplicates */
3601 if (cgroup_sane_behavior(cgrp))
3602 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3604 sort(array, length, sizeof(pid_t), cmppid, NULL);
3605 if (type == CGROUP_FILE_PROCS)
3606 length = pidlist_uniq(array, length);
3608 l = cgroup_pidlist_find_create(cgrp, type);
3610 mutex_unlock(&cgrp->pidlist_mutex);
3611 pidlist_free(array);
3615 /* store array, freeing old if necessary */
3616 pidlist_free(l->list);
3624 * cgroupstats_build - build and fill cgroupstats
3625 * @stats: cgroupstats to fill information into
3626 * @dentry: A dentry entry belonging to the cgroup for which stats have
3629 * Build and fill cgroupstats so that taskstats can export it to user
3632 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3635 struct cgroup *cgrp;
3636 struct css_task_iter it;
3637 struct task_struct *tsk;
3640 * Validate dentry by checking the superblock operations,
3641 * and make sure it's a directory.
3643 if (dentry->d_sb->s_op != &cgroup_ops ||
3644 !S_ISDIR(dentry->d_inode->i_mode))
3648 cgrp = dentry->d_fsdata;
3650 css_task_iter_start(&cgrp->dummy_css, &it);
3651 while ((tsk = css_task_iter_next(&it))) {
3652 switch (tsk->state) {
3654 stats->nr_running++;
3656 case TASK_INTERRUPTIBLE:
3657 stats->nr_sleeping++;
3659 case TASK_UNINTERRUPTIBLE:
3660 stats->nr_uninterruptible++;
3663 stats->nr_stopped++;
3666 if (delayacct_is_task_waiting_on_io(tsk))
3667 stats->nr_io_wait++;
3671 css_task_iter_end(&it);
3679 * seq_file methods for the tasks/procs files. The seq_file position is the
3680 * next pid to display; the seq_file iterator is a pointer to the pid
3681 * in the cgroup->l->list array.
3684 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3687 * Initially we receive a position value that corresponds to
3688 * one more than the last pid shown (or 0 on the first call or
3689 * after a seek to the start). Use a binary-search to find the
3690 * next pid to display, if any
3692 struct cgroup_open_file *of = s->private;
3693 struct cgroup *cgrp = seq_css(s)->cgroup;
3694 struct cgroup_pidlist *l;
3695 enum cgroup_filetype type = seq_cft(s)->private;
3696 int index = 0, pid = *pos;
3699 mutex_lock(&cgrp->pidlist_mutex);
3702 * !NULL @of->priv indicates that this isn't the first start()
3703 * after open. If the matching pidlist is around, we can use that.
3704 * Look for it. Note that @of->priv can't be used directly. It
3705 * could already have been destroyed.
3708 of->priv = cgroup_pidlist_find(cgrp, type);
3711 * Either this is the first start() after open or the matching
3712 * pidlist has been destroyed inbetween. Create a new one.
3715 ret = pidlist_array_load(cgrp, type,
3716 (struct cgroup_pidlist **)&of->priv);
3718 return ERR_PTR(ret);
3723 int end = l->length;
3725 while (index < end) {
3726 int mid = (index + end) / 2;
3727 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3730 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3736 /* If we're off the end of the array, we're done */
3737 if (index >= l->length)
3739 /* Update the abstract position to be the actual pid that we found */
3740 iter = l->list + index;
3741 *pos = cgroup_pid_fry(cgrp, *iter);
3745 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3747 struct cgroup_open_file *of = s->private;
3748 struct cgroup_pidlist *l = of->priv;
3751 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3752 CGROUP_PIDLIST_DESTROY_DELAY);
3753 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3756 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3758 struct cgroup_open_file *of = s->private;
3759 struct cgroup_pidlist *l = of->priv;
3761 pid_t *end = l->list + l->length;
3763 * Advance to the next pid in the array. If this goes off the
3770 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3775 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3777 return seq_printf(s, "%d\n", *(int *)v);
3781 * seq_operations functions for iterating on pidlists through seq_file -
3782 * independent of whether it's tasks or procs
3784 static const struct seq_operations cgroup_pidlist_seq_operations = {
3785 .start = cgroup_pidlist_start,
3786 .stop = cgroup_pidlist_stop,
3787 .next = cgroup_pidlist_next,
3788 .show = cgroup_pidlist_show,
3791 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3794 return notify_on_release(css->cgroup);
3797 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3798 struct cftype *cft, u64 val)
3800 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3802 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3804 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3809 * When dput() is called asynchronously, if umount has been done and
3810 * then deactivate_super() in cgroup_free_fn() kills the superblock,
3811 * there's a small window that vfs will see the root dentry with non-zero
3812 * refcnt and trigger BUG().
3814 * That's why we hold a reference before dput() and drop it right after.
3816 static void cgroup_dput(struct cgroup *cgrp)
3818 struct super_block *sb = cgrp->root->sb;
3820 atomic_inc(&sb->s_active);
3822 deactivate_super(sb);
3825 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3828 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3831 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3832 struct cftype *cft, u64 val)
3835 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3837 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3841 static struct cftype cgroup_base_files[] = {
3843 .name = "cgroup.procs",
3844 .seq_start = cgroup_pidlist_start,
3845 .seq_next = cgroup_pidlist_next,
3846 .seq_stop = cgroup_pidlist_stop,
3847 .seq_show = cgroup_pidlist_show,
3848 .private = CGROUP_FILE_PROCS,
3849 .write_u64 = cgroup_procs_write,
3850 .mode = S_IRUGO | S_IWUSR,
3853 .name = "cgroup.clone_children",
3854 .flags = CFTYPE_INSANE,
3855 .read_u64 = cgroup_clone_children_read,
3856 .write_u64 = cgroup_clone_children_write,
3859 .name = "cgroup.sane_behavior",
3860 .flags = CFTYPE_ONLY_ON_ROOT,
3861 .seq_show = cgroup_sane_behavior_show,
3865 * Historical crazy stuff. These don't have "cgroup." prefix and
3866 * don't exist if sane_behavior. If you're depending on these, be
3867 * prepared to be burned.
3871 .flags = CFTYPE_INSANE, /* use "procs" instead */
3872 .seq_start = cgroup_pidlist_start,
3873 .seq_next = cgroup_pidlist_next,
3874 .seq_stop = cgroup_pidlist_stop,
3875 .seq_show = cgroup_pidlist_show,
3876 .private = CGROUP_FILE_TASKS,
3877 .write_u64 = cgroup_tasks_write,
3878 .mode = S_IRUGO | S_IWUSR,
3881 .name = "notify_on_release",
3882 .flags = CFTYPE_INSANE,
3883 .read_u64 = cgroup_read_notify_on_release,
3884 .write_u64 = cgroup_write_notify_on_release,
3887 .name = "release_agent",
3888 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3889 .seq_show = cgroup_release_agent_show,
3890 .write_string = cgroup_release_agent_write,
3891 .max_write_len = PATH_MAX,
3897 * cgroup_populate_dir - create subsys files in a cgroup directory
3898 * @cgrp: target cgroup
3899 * @subsys_mask: mask of the subsystem ids whose files should be added
3901 * On failure, no file is added.
3903 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3905 struct cgroup_subsys *ss;
3908 /* process cftsets of each subsystem */
3909 for_each_subsys(ss, i) {
3910 struct cftype_set *set;
3912 if (!test_bit(i, &subsys_mask))
3915 list_for_each_entry(set, &ss->cftsets, node) {
3916 ret = cgroup_addrm_files(cgrp, set->cfts, true);
3923 cgroup_clear_dir(cgrp, subsys_mask);
3928 * css destruction is four-stage process.
3930 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3931 * Implemented in kill_css().
3933 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3934 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3935 * by invoking offline_css(). After offlining, the base ref is put.
3936 * Implemented in css_killed_work_fn().
3938 * 3. When the percpu_ref reaches zero, the only possible remaining
3939 * accessors are inside RCU read sections. css_release() schedules the
3942 * 4. After the grace period, the css can be freed. Implemented in
3943 * css_free_work_fn().
3945 * It is actually hairier because both step 2 and 4 require process context
3946 * and thus involve punting to css->destroy_work adding two additional
3947 * steps to the already complex sequence.
3949 static void css_free_work_fn(struct work_struct *work)
3951 struct cgroup_subsys_state *css =
3952 container_of(work, struct cgroup_subsys_state, destroy_work);
3953 struct cgroup *cgrp = css->cgroup;
3956 css_put(css->parent);
3958 css->ss->css_free(css);
3962 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3964 struct cgroup_subsys_state *css =
3965 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3968 * css holds an extra ref to @cgrp->dentry which is put on the last
3969 * css_put(). dput() requires process context which we don't have.
3971 INIT_WORK(&css->destroy_work, css_free_work_fn);
3972 queue_work(cgroup_destroy_wq, &css->destroy_work);
3975 static void css_release(struct percpu_ref *ref)
3977 struct cgroup_subsys_state *css =
3978 container_of(ref, struct cgroup_subsys_state, refcnt);
3980 rcu_assign_pointer(css->cgroup->subsys[css->ss->id], NULL);
3981 call_rcu(&css->rcu_head, css_free_rcu_fn);
3984 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3985 struct cgroup *cgrp)
3992 css->parent = cgroup_css(cgrp->parent, ss);
3994 css->flags |= CSS_ROOT;
3996 BUG_ON(cgroup_css(cgrp, ss));
3999 /* invoke ->css_online() on a new CSS and mark it online if successful */
4000 static int online_css(struct cgroup_subsys_state *css)
4002 struct cgroup_subsys *ss = css->ss;
4005 lockdep_assert_held(&cgroup_tree_mutex);
4006 lockdep_assert_held(&cgroup_mutex);
4009 ret = ss->css_online(css);
4011 css->flags |= CSS_ONLINE;
4012 css->cgroup->nr_css++;
4013 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4018 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4019 static void offline_css(struct cgroup_subsys_state *css)
4021 struct cgroup_subsys *ss = css->ss;
4023 lockdep_assert_held(&cgroup_tree_mutex);
4024 lockdep_assert_held(&cgroup_mutex);
4026 if (!(css->flags & CSS_ONLINE))
4029 if (ss->css_offline)
4030 ss->css_offline(css);
4032 css->flags &= ~CSS_ONLINE;
4033 css->cgroup->nr_css--;
4034 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
4038 * create_css - create a cgroup_subsys_state
4039 * @cgrp: the cgroup new css will be associated with
4040 * @ss: the subsys of new css
4042 * Create a new css associated with @cgrp - @ss pair. On success, the new
4043 * css is online and installed in @cgrp with all interface files created.
4044 * Returns 0 on success, -errno on failure.
4046 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
4048 struct cgroup *parent = cgrp->parent;
4049 struct cgroup_subsys_state *css;
4052 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
4053 lockdep_assert_held(&cgroup_mutex);
4055 css = ss->css_alloc(cgroup_css(parent, ss));
4057 return PTR_ERR(css);
4059 err = percpu_ref_init(&css->refcnt, css_release);
4063 init_css(css, ss, cgrp);
4065 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4069 err = online_css(css);
4074 css_get(css->parent);
4076 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4078 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",
4079 current->comm, current->pid, ss->name);
4080 if (!strcmp(ss->name, "memory"))
4081 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
4082 ss->warned_broken_hierarchy = true;
4088 percpu_ref_cancel_init(&css->refcnt);
4094 * cgroup_create - create a cgroup
4095 * @parent: cgroup that will be parent of the new cgroup
4096 * @dentry: dentry of the new cgroup
4097 * @mode: mode to set on new inode
4099 * Must be called with the mutex on the parent inode held
4101 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
4104 struct cgroup *cgrp;
4105 struct cgroup_name *name;
4106 struct cgroupfs_root *root = parent->root;
4108 struct cgroup_subsys *ss;
4109 struct super_block *sb = root->sb;
4111 /* allocate the cgroup and its ID, 0 is reserved for the root */
4112 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4116 name = cgroup_alloc_name(dentry);
4121 rcu_assign_pointer(cgrp->name, name);
4123 mutex_lock(&cgroup_tree_mutex);
4126 * Only live parents can have children. Note that the liveliness
4127 * check isn't strictly necessary because cgroup_mkdir() and
4128 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
4129 * anyway so that locking is contained inside cgroup proper and we
4130 * don't get nasty surprises if we ever grow another caller.
4132 if (!cgroup_lock_live_group(parent)) {
4134 goto err_unlock_tree;
4138 * Temporarily set the pointer to NULL, so idr_find() won't return
4139 * a half-baked cgroup.
4141 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
4147 /* Grab a reference on the superblock so the hierarchy doesn't
4148 * get deleted on unmount if there are child cgroups. This
4149 * can be done outside cgroup_mutex, since the sb can't
4150 * disappear while someone has an open control file on the
4152 atomic_inc(&sb->s_active);
4154 init_cgroup_housekeeping(cgrp);
4156 dentry->d_fsdata = cgrp;
4157 cgrp->dentry = dentry;
4159 cgrp->parent = parent;
4160 cgrp->dummy_css.parent = &parent->dummy_css;
4161 cgrp->root = parent->root;
4163 if (notify_on_release(parent))
4164 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4166 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4167 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4170 * Create directory. cgroup_create_file() returns with the new
4171 * directory locked on success so that it can be populated without
4172 * dropping cgroup_mutex.
4174 err = cgroup_create_file(dentry, S_IFDIR | mode, sb);
4177 lockdep_assert_held(&dentry->d_inode->i_mutex);
4179 cgrp->serial_nr = cgroup_serial_nr_next++;
4181 /* allocation complete, commit to creation */
4182 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
4183 root->number_of_cgroups++;
4185 /* hold a ref to the parent's dentry */
4186 dget(parent->dentry);
4189 * @cgrp is now fully operational. If something fails after this
4190 * point, it'll be released via the normal destruction path.
4192 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4194 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
4198 /* let's create and online css's */
4199 for_each_subsys(ss, ssid) {
4200 if (root->subsys_mask & (1 << ssid)) {
4201 err = create_css(cgrp, ss);
4207 mutex_unlock(&cgroup_mutex);
4208 mutex_unlock(&cgroup_tree_mutex);
4209 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
4214 idr_remove(&root->cgroup_idr, cgrp->id);
4215 /* Release the reference count that we took on the superblock */
4216 deactivate_super(sb);
4218 mutex_unlock(&cgroup_mutex);
4220 mutex_unlock(&cgroup_tree_mutex);
4221 kfree(rcu_dereference_raw(cgrp->name));
4227 cgroup_destroy_locked(cgrp);
4228 mutex_unlock(&cgroup_mutex);
4229 mutex_unlock(&cgroup_tree_mutex);
4230 mutex_unlock(&dentry->d_inode->i_mutex);
4234 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
4236 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
4238 /* the vfs holds inode->i_mutex already */
4239 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
4243 * This is called when the refcnt of a css is confirmed to be killed.
4244 * css_tryget() is now guaranteed to fail.
4246 static void css_killed_work_fn(struct work_struct *work)
4248 struct cgroup_subsys_state *css =
4249 container_of(work, struct cgroup_subsys_state, destroy_work);
4250 struct cgroup *cgrp = css->cgroup;
4252 mutex_lock(&cgroup_tree_mutex);
4253 mutex_lock(&cgroup_mutex);
4256 * css_tryget() is guaranteed to fail now. Tell subsystems to
4257 * initate destruction.
4262 * If @cgrp is marked dead, it's waiting for refs of all css's to
4263 * be disabled before proceeding to the second phase of cgroup
4264 * destruction. If we are the last one, kick it off.
4266 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
4267 cgroup_destroy_css_killed(cgrp);
4269 mutex_unlock(&cgroup_mutex);
4270 mutex_unlock(&cgroup_tree_mutex);
4273 * Put the css refs from kill_css(). Each css holds an extra
4274 * reference to the cgroup's dentry and cgroup removal proceeds
4275 * regardless of css refs. On the last put of each css, whenever
4276 * that may be, the extra dentry ref is put so that dentry
4277 * destruction happens only after all css's are released.
4282 /* css kill confirmation processing requires process context, bounce */
4283 static void css_killed_ref_fn(struct percpu_ref *ref)
4285 struct cgroup_subsys_state *css =
4286 container_of(ref, struct cgroup_subsys_state, refcnt);
4288 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4289 queue_work(cgroup_destroy_wq, &css->destroy_work);
4293 * kill_css - destroy a css
4294 * @css: css to destroy
4296 * This function initiates destruction of @css by removing cgroup interface
4297 * files and putting its base reference. ->css_offline() will be invoked
4298 * asynchronously once css_tryget() is guaranteed to fail and when the
4299 * reference count reaches zero, @css will be released.
4301 static void kill_css(struct cgroup_subsys_state *css)
4303 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4306 * Killing would put the base ref, but we need to keep it alive
4307 * until after ->css_offline().
4312 * cgroup core guarantees that, by the time ->css_offline() is
4313 * invoked, no new css reference will be given out via
4314 * css_tryget(). We can't simply call percpu_ref_kill() and
4315 * proceed to offlining css's because percpu_ref_kill() doesn't
4316 * guarantee that the ref is seen as killed on all CPUs on return.
4318 * Use percpu_ref_kill_and_confirm() to get notifications as each
4319 * css is confirmed to be seen as killed on all CPUs.
4321 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4325 * cgroup_destroy_locked - the first stage of cgroup destruction
4326 * @cgrp: cgroup to be destroyed
4328 * css's make use of percpu refcnts whose killing latency shouldn't be
4329 * exposed to userland and are RCU protected. Also, cgroup core needs to
4330 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
4331 * invoked. To satisfy all the requirements, destruction is implemented in
4332 * the following two steps.
4334 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4335 * userland visible parts and start killing the percpu refcnts of
4336 * css's. Set up so that the next stage will be kicked off once all
4337 * the percpu refcnts are confirmed to be killed.
4339 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4340 * rest of destruction. Once all cgroup references are gone, the
4341 * cgroup is RCU-freed.
4343 * This function implements s1. After this step, @cgrp is gone as far as
4344 * the userland is concerned and a new cgroup with the same name may be
4345 * created. As cgroup doesn't care about the names internally, this
4346 * doesn't cause any problem.
4348 static int cgroup_destroy_locked(struct cgroup *cgrp)
4349 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4351 struct dentry *d = cgrp->dentry;
4352 struct cgroup_subsys_state *css;
4353 struct cgroup *child;
4357 lockdep_assert_held(&d->d_inode->i_mutex);
4358 lockdep_assert_held(&cgroup_tree_mutex);
4359 lockdep_assert_held(&cgroup_mutex);
4362 * css_set_lock synchronizes access to ->cset_links and prevents
4363 * @cgrp from being removed while __put_css_set() is in progress.
4365 read_lock(&css_set_lock);
4366 empty = list_empty(&cgrp->cset_links);
4367 read_unlock(&css_set_lock);
4372 * Make sure there's no live children. We can't test ->children
4373 * emptiness as dead children linger on it while being destroyed;
4374 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
4378 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
4379 empty = cgroup_is_dead(child);
4388 * Initiate massacre of all css's. cgroup_destroy_css_killed()
4389 * will be invoked to perform the rest of destruction once the
4390 * percpu refs of all css's are confirmed to be killed. This
4391 * involves removing the subsystem's files, drop cgroup_mutex.
4393 mutex_unlock(&cgroup_mutex);
4394 for_each_css(css, ssid, cgrp)
4396 mutex_lock(&cgroup_mutex);
4399 * Mark @cgrp dead. This prevents further task migration and child
4400 * creation by disabling cgroup_lock_live_group(). Note that
4401 * CGRP_DEAD assertion is depended upon by css_next_child() to
4402 * resume iteration after dropping RCU read lock. See
4403 * css_next_child() for details.
4405 set_bit(CGRP_DEAD, &cgrp->flags);
4407 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4408 raw_spin_lock(&release_list_lock);
4409 if (!list_empty(&cgrp->release_list))
4410 list_del_init(&cgrp->release_list);
4411 raw_spin_unlock(&release_list_lock);
4414 * If @cgrp has css's attached, the second stage of cgroup
4415 * destruction is kicked off from css_killed_work_fn() after the
4416 * refs of all attached css's are killed. If @cgrp doesn't have
4417 * any css, we kick it off here.
4420 cgroup_destroy_css_killed(cgrp);
4423 * Clear the base files and remove @cgrp directory. The removal
4424 * puts the base ref but we aren't quite done with @cgrp yet, so
4427 mutex_unlock(&cgroup_mutex);
4428 cgroup_addrm_files(cgrp, cgroup_base_files, false);
4430 cgroup_d_remove_dir(d);
4431 mutex_lock(&cgroup_mutex);
4437 * cgroup_destroy_css_killed - the second step of cgroup destruction
4438 * @work: cgroup->destroy_free_work
4440 * This function is invoked from a work item for a cgroup which is being
4441 * destroyed after all css's are offlined and performs the rest of
4442 * destruction. This is the second step of destruction described in the
4443 * comment above cgroup_destroy_locked().
4445 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
4447 struct cgroup *parent = cgrp->parent;
4448 struct dentry *d = cgrp->dentry;
4450 lockdep_assert_held(&cgroup_tree_mutex);
4451 lockdep_assert_held(&cgroup_mutex);
4453 /* delete this cgroup from parent->children */
4454 list_del_rcu(&cgrp->sibling);
4458 set_bit(CGRP_RELEASABLE, &parent->flags);
4459 check_for_release(parent);
4462 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
4466 mutex_lock(&cgroup_tree_mutex);
4467 mutex_lock(&cgroup_mutex);
4468 ret = cgroup_destroy_locked(dentry->d_fsdata);
4469 mutex_unlock(&cgroup_mutex);
4470 mutex_unlock(&cgroup_tree_mutex);
4475 static void __init cgroup_init_cftsets(struct cgroup_subsys *ss)
4477 INIT_LIST_HEAD(&ss->cftsets);
4480 * base_cftset is embedded in subsys itself, no need to worry about
4483 if (ss->base_cftypes) {
4486 for (cft = ss->base_cftypes; cft->name[0] != '\0'; cft++)
4489 ss->base_cftset.cfts = ss->base_cftypes;
4490 list_add_tail(&ss->base_cftset.node, &ss->cftsets);
4494 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4496 struct cgroup_subsys_state *css;
4498 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4500 mutex_lock(&cgroup_tree_mutex);
4501 mutex_lock(&cgroup_mutex);
4503 /* init base cftset */
4504 cgroup_init_cftsets(ss);
4506 /* Create the top cgroup state for this subsystem */
4507 ss->root = &cgroup_dummy_root;
4508 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4509 /* We don't handle early failures gracefully */
4510 BUG_ON(IS_ERR(css));
4511 init_css(css, ss, cgroup_dummy_top);
4513 /* Update the init_css_set to contain a subsys
4514 * pointer to this state - since the subsystem is
4515 * newly registered, all tasks and hence the
4516 * init_css_set is in the subsystem's top cgroup. */
4517 init_css_set.subsys[ss->id] = css;
4519 need_forkexit_callback |= ss->fork || ss->exit;
4521 /* At system boot, before all subsystems have been
4522 * registered, no tasks have been forked, so we don't
4523 * need to invoke fork callbacks here. */
4524 BUG_ON(!list_empty(&init_task.tasks));
4526 BUG_ON(online_css(css));
4528 mutex_unlock(&cgroup_mutex);
4529 mutex_unlock(&cgroup_tree_mutex);
4533 * cgroup_init_early - cgroup initialization at system boot
4535 * Initialize cgroups at system boot, and initialize any
4536 * subsystems that request early init.
4538 int __init cgroup_init_early(void)
4540 struct cgroup_subsys *ss;
4543 atomic_set(&init_css_set.refcount, 1);
4544 INIT_LIST_HEAD(&init_css_set.cgrp_links);
4545 INIT_LIST_HEAD(&init_css_set.tasks);
4546 INIT_HLIST_NODE(&init_css_set.hlist);
4548 init_cgroup_root(&cgroup_dummy_root);
4549 cgroup_root_count = 1;
4550 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4552 init_cgrp_cset_link.cset = &init_css_set;
4553 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4554 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4555 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4557 for_each_subsys(ss, i) {
4558 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4559 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4560 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4562 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4563 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4566 ss->name = cgroup_subsys_name[i];
4569 cgroup_init_subsys(ss);
4575 * cgroup_init - cgroup initialization
4577 * Register cgroup filesystem and /proc file, and initialize
4578 * any subsystems that didn't request early init.
4580 int __init cgroup_init(void)
4582 struct cgroup_subsys *ss;
4586 err = bdi_init(&cgroup_backing_dev_info);
4590 for_each_subsys(ss, i) {
4591 if (!ss->early_init)
4592 cgroup_init_subsys(ss);
4595 /* allocate id for the dummy hierarchy */
4596 mutex_lock(&cgroup_mutex);
4598 /* Add init_css_set to the hash table */
4599 key = css_set_hash(init_css_set.subsys);
4600 hash_add(css_set_table, &init_css_set.hlist, key);
4602 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4604 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
4608 mutex_unlock(&cgroup_mutex);
4610 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4616 err = register_filesystem(&cgroup_fs_type);
4618 kobject_put(cgroup_kobj);
4622 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4626 bdi_destroy(&cgroup_backing_dev_info);
4631 static int __init cgroup_wq_init(void)
4634 * There isn't much point in executing destruction path in
4635 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4637 * XXX: Must be ordered to make sure parent is offlined after
4638 * children. The ordering requirement is for memcg where a
4639 * parent's offline may wait for a child's leading to deadlock. In
4640 * the long term, this should be fixed from memcg side.
4642 * We would prefer to do this in cgroup_init() above, but that
4643 * is called before init_workqueues(): so leave this until after.
4645 cgroup_destroy_wq = alloc_ordered_workqueue("cgroup_destroy", 0);
4646 BUG_ON(!cgroup_destroy_wq);
4649 * Used to destroy pidlists and separate to serve as flush domain.
4650 * Cap @max_active to 1 too.
4652 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4654 BUG_ON(!cgroup_pidlist_destroy_wq);
4658 core_initcall(cgroup_wq_init);
4661 * proc_cgroup_show()
4662 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4663 * - Used for /proc/<pid>/cgroup.
4664 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4665 * doesn't really matter if tsk->cgroup changes after we read it,
4666 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4667 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4668 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4669 * cgroup to top_cgroup.
4672 /* TODO: Use a proper seq_file iterator */
4673 int proc_cgroup_show(struct seq_file *m, void *v)
4676 struct task_struct *tsk;
4679 struct cgroupfs_root *root;
4682 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4688 tsk = get_pid_task(pid, PIDTYPE_PID);
4694 mutex_lock(&cgroup_mutex);
4696 for_each_active_root(root) {
4697 struct cgroup_subsys *ss;
4698 struct cgroup *cgrp;
4699 int ssid, count = 0;
4701 seq_printf(m, "%d:", root->hierarchy_id);
4702 for_each_subsys(ss, ssid)
4703 if (root->subsys_mask & (1 << ssid))
4704 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4705 if (strlen(root->name))
4706 seq_printf(m, "%sname=%s", count ? "," : "",
4709 cgrp = task_cgroup_from_root(tsk, root);
4710 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4718 mutex_unlock(&cgroup_mutex);
4719 put_task_struct(tsk);
4726 /* Display information about each subsystem and each hierarchy */
4727 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4729 struct cgroup_subsys *ss;
4732 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4734 * ideally we don't want subsystems moving around while we do this.
4735 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4736 * subsys/hierarchy state.
4738 mutex_lock(&cgroup_mutex);
4740 for_each_subsys(ss, i)
4741 seq_printf(m, "%s\t%d\t%d\t%d\n",
4742 ss->name, ss->root->hierarchy_id,
4743 ss->root->number_of_cgroups, !ss->disabled);
4745 mutex_unlock(&cgroup_mutex);
4749 static int cgroupstats_open(struct inode *inode, struct file *file)
4751 return single_open(file, proc_cgroupstats_show, NULL);
4754 static const struct file_operations proc_cgroupstats_operations = {
4755 .open = cgroupstats_open,
4757 .llseek = seq_lseek,
4758 .release = single_release,
4762 * cgroup_fork - attach newly forked task to its parents cgroup.
4763 * @child: pointer to task_struct of forking parent process.
4765 * Description: A task inherits its parent's cgroup at fork().
4767 * A pointer to the shared css_set was automatically copied in
4768 * fork.c by dup_task_struct(). However, we ignore that copy, since
4769 * it was not made under the protection of RCU or cgroup_mutex, so
4770 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4771 * have already changed current->cgroups, allowing the previously
4772 * referenced cgroup group to be removed and freed.
4774 * At the point that cgroup_fork() is called, 'current' is the parent
4775 * task, and the passed argument 'child' points to the child task.
4777 void cgroup_fork(struct task_struct *child)
4780 get_css_set(task_css_set(current));
4781 child->cgroups = current->cgroups;
4782 task_unlock(current);
4783 INIT_LIST_HEAD(&child->cg_list);
4787 * cgroup_post_fork - called on a new task after adding it to the task list
4788 * @child: the task in question
4790 * Adds the task to the list running through its css_set if necessary and
4791 * call the subsystem fork() callbacks. Has to be after the task is
4792 * visible on the task list in case we race with the first call to
4793 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4796 void cgroup_post_fork(struct task_struct *child)
4798 struct cgroup_subsys *ss;
4802 * use_task_css_set_links is set to 1 before we walk the tasklist
4803 * under the tasklist_lock and we read it here after we added the child
4804 * to the tasklist under the tasklist_lock as well. If the child wasn't
4805 * yet in the tasklist when we walked through it from
4806 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4807 * should be visible now due to the paired locking and barriers implied
4808 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4809 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4812 if (use_task_css_set_links) {
4813 write_lock(&css_set_lock);
4815 if (list_empty(&child->cg_list))
4816 list_add(&child->cg_list, &task_css_set(child)->tasks);
4818 write_unlock(&css_set_lock);
4822 * Call ss->fork(). This must happen after @child is linked on
4823 * css_set; otherwise, @child might change state between ->fork()
4824 * and addition to css_set.
4826 if (need_forkexit_callback) {
4827 for_each_subsys(ss, i)
4834 * cgroup_exit - detach cgroup from exiting task
4835 * @tsk: pointer to task_struct of exiting process
4836 * @run_callback: run exit callbacks?
4838 * Description: Detach cgroup from @tsk and release it.
4840 * Note that cgroups marked notify_on_release force every task in
4841 * them to take the global cgroup_mutex mutex when exiting.
4842 * This could impact scaling on very large systems. Be reluctant to
4843 * use notify_on_release cgroups where very high task exit scaling
4844 * is required on large systems.
4846 * the_top_cgroup_hack:
4848 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4850 * We call cgroup_exit() while the task is still competent to
4851 * handle notify_on_release(), then leave the task attached to the
4852 * root cgroup in each hierarchy for the remainder of its exit.
4854 * To do this properly, we would increment the reference count on
4855 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4856 * code we would add a second cgroup function call, to drop that
4857 * reference. This would just create an unnecessary hot spot on
4858 * the top_cgroup reference count, to no avail.
4860 * Normally, holding a reference to a cgroup without bumping its
4861 * count is unsafe. The cgroup could go away, or someone could
4862 * attach us to a different cgroup, decrementing the count on
4863 * the first cgroup that we never incremented. But in this case,
4864 * top_cgroup isn't going away, and either task has PF_EXITING set,
4865 * which wards off any cgroup_attach_task() attempts, or task is a failed
4866 * fork, never visible to cgroup_attach_task.
4868 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4870 struct cgroup_subsys *ss;
4871 struct css_set *cset;
4875 * Unlink from the css_set task list if necessary.
4876 * Optimistically check cg_list before taking
4879 if (!list_empty(&tsk->cg_list)) {
4880 write_lock(&css_set_lock);
4881 if (!list_empty(&tsk->cg_list))
4882 list_del_init(&tsk->cg_list);
4883 write_unlock(&css_set_lock);
4886 /* Reassign the task to the init_css_set. */
4888 cset = task_css_set(tsk);
4889 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4891 if (run_callbacks && need_forkexit_callback) {
4892 /* see cgroup_post_fork() for details */
4893 for_each_subsys(ss, i) {
4895 struct cgroup_subsys_state *old_css = cset->subsys[i];
4896 struct cgroup_subsys_state *css = task_css(tsk, i);
4898 ss->exit(css, old_css, tsk);
4904 put_css_set_taskexit(cset);
4907 static void check_for_release(struct cgroup *cgrp)
4909 if (cgroup_is_releasable(cgrp) &&
4910 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4912 * Control Group is currently removeable. If it's not
4913 * already queued for a userspace notification, queue
4916 int need_schedule_work = 0;
4918 raw_spin_lock(&release_list_lock);
4919 if (!cgroup_is_dead(cgrp) &&
4920 list_empty(&cgrp->release_list)) {
4921 list_add(&cgrp->release_list, &release_list);
4922 need_schedule_work = 1;
4924 raw_spin_unlock(&release_list_lock);
4925 if (need_schedule_work)
4926 schedule_work(&release_agent_work);
4931 * Notify userspace when a cgroup is released, by running the
4932 * configured release agent with the name of the cgroup (path
4933 * relative to the root of cgroup file system) as the argument.
4935 * Most likely, this user command will try to rmdir this cgroup.
4937 * This races with the possibility that some other task will be
4938 * attached to this cgroup before it is removed, or that some other
4939 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4940 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4941 * unused, and this cgroup will be reprieved from its death sentence,
4942 * to continue to serve a useful existence. Next time it's released,
4943 * we will get notified again, if it still has 'notify_on_release' set.
4945 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4946 * means only wait until the task is successfully execve()'d. The
4947 * separate release agent task is forked by call_usermodehelper(),
4948 * then control in this thread returns here, without waiting for the
4949 * release agent task. We don't bother to wait because the caller of
4950 * this routine has no use for the exit status of the release agent
4951 * task, so no sense holding our caller up for that.
4953 static void cgroup_release_agent(struct work_struct *work)
4955 BUG_ON(work != &release_agent_work);
4956 mutex_lock(&cgroup_mutex);
4957 raw_spin_lock(&release_list_lock);
4958 while (!list_empty(&release_list)) {
4959 char *argv[3], *envp[3];
4961 char *pathbuf = NULL, *agentbuf = NULL;
4962 struct cgroup *cgrp = list_entry(release_list.next,
4965 list_del_init(&cgrp->release_list);
4966 raw_spin_unlock(&release_list_lock);
4967 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4970 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4972 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4977 argv[i++] = agentbuf;
4978 argv[i++] = pathbuf;
4982 /* minimal command environment */
4983 envp[i++] = "HOME=/";
4984 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4987 /* Drop the lock while we invoke the usermode helper,
4988 * since the exec could involve hitting disk and hence
4989 * be a slow process */
4990 mutex_unlock(&cgroup_mutex);
4991 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4992 mutex_lock(&cgroup_mutex);
4996 raw_spin_lock(&release_list_lock);
4998 raw_spin_unlock(&release_list_lock);
4999 mutex_unlock(&cgroup_mutex);
5002 static int __init cgroup_disable(char *str)
5004 struct cgroup_subsys *ss;
5008 while ((token = strsep(&str, ",")) != NULL) {
5012 for_each_subsys(ss, i) {
5013 if (!strcmp(token, ss->name)) {
5015 printk(KERN_INFO "Disabling %s control group"
5016 " subsystem\n", ss->name);
5023 __setup("cgroup_disable=", cgroup_disable);
5026 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
5027 * @dentry: directory dentry of interest
5028 * @ss: subsystem of interest
5030 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5031 * to get the corresponding css and return it. If such css doesn't exist
5032 * or can't be pinned, an ERR_PTR value is returned.
5034 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
5035 struct cgroup_subsys *ss)
5037 struct cgroup *cgrp;
5038 struct cgroup_subsys_state *css;
5040 /* is @dentry a cgroup dir? */
5041 if (!dentry->d_inode ||
5042 dentry->d_inode->i_op != &cgroup_dir_inode_operations)
5043 return ERR_PTR(-EBADF);
5047 cgrp = __d_cgrp(dentry);
5048 css = cgroup_css(cgrp, ss);
5050 if (!css || !css_tryget(css))
5051 css = ERR_PTR(-ENOENT);
5058 * css_from_id - lookup css by id
5059 * @id: the cgroup id
5060 * @ss: cgroup subsys to be looked into
5062 * Returns the css if there's valid one with @id, otherwise returns NULL.
5063 * Should be called under rcu_read_lock().
5065 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5067 struct cgroup *cgrp;
5069 cgroup_assert_mutexes_or_rcu_locked();
5071 cgrp = idr_find(&ss->root->cgroup_idr, id);
5073 return cgroup_css(cgrp, ss);
5077 #ifdef CONFIG_CGROUP_DEBUG
5078 static struct cgroup_subsys_state *
5079 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5081 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5084 return ERR_PTR(-ENOMEM);
5089 static void debug_css_free(struct cgroup_subsys_state *css)
5094 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5097 return cgroup_task_count(css->cgroup);
5100 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5103 return (u64)(unsigned long)current->cgroups;
5106 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5112 count = atomic_read(&task_css_set(current)->refcount);
5117 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5119 struct cgrp_cset_link *link;
5120 struct css_set *cset;
5122 read_lock(&css_set_lock);
5124 cset = rcu_dereference(current->cgroups);
5125 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5126 struct cgroup *c = link->cgrp;
5130 name = c->dentry->d_name.name;
5133 seq_printf(seq, "Root %d group %s\n",
5134 c->root->hierarchy_id, name);
5137 read_unlock(&css_set_lock);
5141 #define MAX_TASKS_SHOWN_PER_CSS 25
5142 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5144 struct cgroup_subsys_state *css = seq_css(seq);
5145 struct cgrp_cset_link *link;
5147 read_lock(&css_set_lock);
5148 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5149 struct css_set *cset = link->cset;
5150 struct task_struct *task;
5152 seq_printf(seq, "css_set %p\n", cset);
5153 list_for_each_entry(task, &cset->tasks, cg_list) {
5154 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
5155 seq_puts(seq, " ...\n");
5158 seq_printf(seq, " task %d\n",
5159 task_pid_vnr(task));
5163 read_unlock(&css_set_lock);
5167 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5169 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5172 static struct cftype debug_files[] = {
5174 .name = "taskcount",
5175 .read_u64 = debug_taskcount_read,
5179 .name = "current_css_set",
5180 .read_u64 = current_css_set_read,
5184 .name = "current_css_set_refcount",
5185 .read_u64 = current_css_set_refcount_read,
5189 .name = "current_css_set_cg_links",
5190 .seq_show = current_css_set_cg_links_read,
5194 .name = "cgroup_css_links",
5195 .seq_show = cgroup_css_links_read,
5199 .name = "releasable",
5200 .read_u64 = releasable_read,
5206 struct cgroup_subsys debug_cgrp_subsys = {
5207 .css_alloc = debug_css_alloc,
5208 .css_free = debug_css_free,
5209 .base_cftypes = debug_files,
5211 #endif /* CONFIG_CGROUP_DEBUG */