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/ctype.h>
31 #include <linux/errno.h>
33 #include <linux/kernel.h>
34 #include <linux/list.h>
36 #include <linux/mutex.h>
37 #include <linux/mount.h>
38 #include <linux/pagemap.h>
39 #include <linux/proc_fs.h>
40 #include <linux/rcupdate.h>
41 #include <linux/sched.h>
42 #include <linux/backing-dev.h>
43 #include <linux/seq_file.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/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/cgroupstats.h>
53 #include <linux/hash.h>
54 #include <linux/namei.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/eventfd.h>
59 #include <linux/poll.h>
60 #include <linux/capability.h>
62 #include <asm/atomic.h>
64 static DEFINE_MUTEX(cgroup_mutex);
67 * Generate an array of cgroup subsystem pointers. At boot time, this is
68 * populated up to CGROUP_BUILTIN_SUBSYS_COUNT, and modular subsystems are
69 * registered after that. The mutable section of this array is protected by
72 #define SUBSYS(_x) &_x ## _subsys,
73 static struct cgroup_subsys *subsys[CGROUP_SUBSYS_COUNT] = {
74 #include <linux/cgroup_subsys.h>
77 #define MAX_CGROUP_ROOT_NAMELEN 64
80 * A cgroupfs_root represents the root of a cgroup hierarchy,
81 * and may be associated with a superblock to form an active
84 struct cgroupfs_root {
85 struct super_block *sb;
88 * The bitmask of subsystems intended to be attached to this
91 unsigned long subsys_bits;
93 /* Unique id for this hierarchy. */
96 /* The bitmask of subsystems currently attached to this hierarchy */
97 unsigned long actual_subsys_bits;
99 /* A list running through the attached subsystems */
100 struct list_head subsys_list;
102 /* The root cgroup for this hierarchy */
103 struct cgroup top_cgroup;
105 /* Tracks how many cgroups are currently defined in hierarchy.*/
106 int number_of_cgroups;
108 /* A list running through the active hierarchies */
109 struct list_head root_list;
111 /* Hierarchy-specific flags */
114 /* The path to use for release notifications. */
115 char release_agent_path[PATH_MAX];
117 /* The name for this hierarchy - may be empty */
118 char name[MAX_CGROUP_ROOT_NAMELEN];
122 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
123 * subsystems that are otherwise unattached - it never has more than a
124 * single cgroup, and all tasks are part of that cgroup.
126 static struct cgroupfs_root rootnode;
129 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
130 * cgroup_subsys->use_id != 0.
132 #define CSS_ID_MAX (65535)
135 * The css to which this ID points. This pointer is set to valid value
136 * after cgroup is populated. If cgroup is removed, this will be NULL.
137 * This pointer is expected to be RCU-safe because destroy()
138 * is called after synchronize_rcu(). But for safe use, css_is_removed()
139 * css_tryget() should be used for avoiding race.
141 struct cgroup_subsys_state __rcu *css;
147 * Depth in hierarchy which this ID belongs to.
149 unsigned short depth;
151 * ID is freed by RCU. (and lookup routine is RCU safe.)
153 struct rcu_head rcu_head;
155 * Hierarchy of CSS ID belongs to.
157 unsigned short stack[0]; /* Array of Length (depth+1) */
161 * cgroup_event represents events which userspace want to receive.
163 struct cgroup_event {
165 * Cgroup which the event belongs to.
169 * Control file which the event associated.
173 * eventfd to signal userspace about the event.
175 struct eventfd_ctx *eventfd;
177 * Each of these stored in a list by the cgroup.
179 struct list_head list;
181 * All fields below needed to unregister event when
182 * userspace closes eventfd.
185 wait_queue_head_t *wqh;
187 struct work_struct remove;
190 /* The list of hierarchy roots */
192 static LIST_HEAD(roots);
193 static int root_count;
195 static DEFINE_IDA(hierarchy_ida);
196 static int next_hierarchy_id;
197 static DEFINE_SPINLOCK(hierarchy_id_lock);
199 /* dummytop is a shorthand for the dummy hierarchy's top cgroup */
200 #define dummytop (&rootnode.top_cgroup)
202 /* This flag indicates whether tasks in the fork and exit paths should
203 * check for fork/exit handlers to call. This avoids us having to do
204 * extra work in the fork/exit path if none of the subsystems need to
207 static int need_forkexit_callback __read_mostly;
209 #ifdef CONFIG_PROVE_LOCKING
210 int cgroup_lock_is_held(void)
212 return lockdep_is_held(&cgroup_mutex);
214 #else /* #ifdef CONFIG_PROVE_LOCKING */
215 int cgroup_lock_is_held(void)
217 return mutex_is_locked(&cgroup_mutex);
219 #endif /* #else #ifdef CONFIG_PROVE_LOCKING */
221 EXPORT_SYMBOL_GPL(cgroup_lock_is_held);
223 /* convenient tests for these bits */
224 inline int cgroup_is_removed(const struct cgroup *cgrp)
226 return test_bit(CGRP_REMOVED, &cgrp->flags);
229 /* bits in struct cgroupfs_root flags field */
231 ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
234 static int cgroup_is_releasable(const struct cgroup *cgrp)
237 (1 << CGRP_RELEASABLE) |
238 (1 << CGRP_NOTIFY_ON_RELEASE);
239 return (cgrp->flags & bits) == bits;
242 static int notify_on_release(const struct cgroup *cgrp)
244 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
247 static int clone_children(const struct cgroup *cgrp)
249 return test_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
253 * for_each_subsys() allows you to iterate on each subsystem attached to
254 * an active hierarchy
256 #define for_each_subsys(_root, _ss) \
257 list_for_each_entry(_ss, &_root->subsys_list, sibling)
259 /* for_each_active_root() allows you to iterate across the active hierarchies */
260 #define for_each_active_root(_root) \
261 list_for_each_entry(_root, &roots, root_list)
263 /* the list of cgroups eligible for automatic release. Protected by
264 * release_list_lock */
265 static LIST_HEAD(release_list);
266 static DEFINE_SPINLOCK(release_list_lock);
267 static void cgroup_release_agent(struct work_struct *work);
268 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
269 static void check_for_release(struct cgroup *cgrp);
272 * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
273 * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
274 * reference to css->refcnt. In general, this refcnt is expected to goes down
277 * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
279 DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
281 static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
283 if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
284 wake_up_all(&cgroup_rmdir_waitq);
287 void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
292 void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
294 cgroup_wakeup_rmdir_waiter(css->cgroup);
298 /* Link structure for associating css_set objects with cgroups */
299 struct cg_cgroup_link {
301 * List running through cg_cgroup_links associated with a
302 * cgroup, anchored on cgroup->css_sets
304 struct list_head cgrp_link_list;
307 * List running through cg_cgroup_links pointing at a
308 * single css_set object, anchored on css_set->cg_links
310 struct list_head cg_link_list;
314 /* The default css_set - used by init and its children prior to any
315 * hierarchies being mounted. It contains a pointer to the root state
316 * for each subsystem. Also used to anchor the list of css_sets. Not
317 * reference-counted, to improve performance when child cgroups
318 * haven't been created.
321 static struct css_set init_css_set;
322 static struct cg_cgroup_link init_css_set_link;
324 static int cgroup_init_idr(struct cgroup_subsys *ss,
325 struct cgroup_subsys_state *css);
327 /* css_set_lock protects the list of css_set objects, and the
328 * chain of tasks off each css_set. Nests outside task->alloc_lock
329 * due to cgroup_iter_start() */
330 static DEFINE_RWLOCK(css_set_lock);
331 static int css_set_count;
334 * hash table for cgroup groups. This improves the performance to find
335 * an existing css_set. This hash doesn't (currently) take into
336 * account cgroups in empty hierarchies.
338 #define CSS_SET_HASH_BITS 7
339 #define CSS_SET_TABLE_SIZE (1 << CSS_SET_HASH_BITS)
340 static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
342 static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
346 unsigned long tmp = 0UL;
348 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
349 tmp += (unsigned long)css[i];
350 tmp = (tmp >> 16) ^ tmp;
352 index = hash_long(tmp, CSS_SET_HASH_BITS);
354 return &css_set_table[index];
357 static void free_css_set_work(struct work_struct *work)
359 struct css_set *cg = container_of(work, struct css_set, work);
360 struct cg_cgroup_link *link;
361 struct cg_cgroup_link *saved_link;
363 write_lock(&css_set_lock);
364 list_for_each_entry_safe(link, saved_link, &cg->cg_links,
366 struct cgroup *cgrp = link->cgrp;
367 list_del(&link->cg_link_list);
368 list_del(&link->cgrp_link_list);
369 if (atomic_dec_and_test(&cgrp->count)) {
370 check_for_release(cgrp);
371 cgroup_wakeup_rmdir_waiter(cgrp);
375 write_unlock(&css_set_lock);
380 static void free_css_set_rcu(struct rcu_head *obj)
382 struct css_set *cg = container_of(obj, struct css_set, rcu_head);
384 INIT_WORK(&cg->work, free_css_set_work);
385 schedule_work(&cg->work);
388 /* We don't maintain the lists running through each css_set to its
389 * task until after the first call to cgroup_iter_start(). This
390 * reduces the fork()/exit() overhead for people who have cgroups
391 * compiled into their kernel but not actually in use */
392 static int use_task_css_set_links __read_mostly;
395 * refcounted get/put for css_set objects
397 static inline void get_css_set(struct css_set *cg)
399 atomic_inc(&cg->refcount);
402 static void put_css_set(struct css_set *cg)
405 * Ensure that the refcount doesn't hit zero while any readers
406 * can see it. Similar to atomic_dec_and_lock(), but for an
409 if (atomic_add_unless(&cg->refcount, -1, 1))
411 write_lock(&css_set_lock);
412 if (!atomic_dec_and_test(&cg->refcount)) {
413 write_unlock(&css_set_lock);
417 hlist_del(&cg->hlist);
420 write_unlock(&css_set_lock);
421 call_rcu(&cg->rcu_head, free_css_set_rcu);
425 * compare_css_sets - helper function for find_existing_css_set().
426 * @cg: candidate css_set being tested
427 * @old_cg: existing css_set for a task
428 * @new_cgrp: cgroup that's being entered by the task
429 * @template: desired set of css pointers in css_set (pre-calculated)
431 * Returns true if "cg" matches "old_cg" except for the hierarchy
432 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
434 static bool compare_css_sets(struct css_set *cg,
435 struct css_set *old_cg,
436 struct cgroup *new_cgrp,
437 struct cgroup_subsys_state *template[])
439 struct list_head *l1, *l2;
441 if (memcmp(template, cg->subsys, sizeof(cg->subsys))) {
442 /* Not all subsystems matched */
447 * Compare cgroup pointers in order to distinguish between
448 * different cgroups in heirarchies with no subsystems. We
449 * could get by with just this check alone (and skip the
450 * memcmp above) but on most setups the memcmp check will
451 * avoid the need for this more expensive check on almost all
456 l2 = &old_cg->cg_links;
458 struct cg_cgroup_link *cgl1, *cgl2;
459 struct cgroup *cg1, *cg2;
463 /* See if we reached the end - both lists are equal length. */
464 if (l1 == &cg->cg_links) {
465 BUG_ON(l2 != &old_cg->cg_links);
468 BUG_ON(l2 == &old_cg->cg_links);
470 /* Locate the cgroups associated with these links. */
471 cgl1 = list_entry(l1, struct cg_cgroup_link, cg_link_list);
472 cgl2 = list_entry(l2, struct cg_cgroup_link, cg_link_list);
475 /* Hierarchies should be linked in the same order. */
476 BUG_ON(cg1->root != cg2->root);
479 * If this hierarchy is the hierarchy of the cgroup
480 * that's changing, then we need to check that this
481 * css_set points to the new cgroup; if it's any other
482 * hierarchy, then this css_set should point to the
483 * same cgroup as the old css_set.
485 if (cg1->root == new_cgrp->root) {
497 * find_existing_css_set() is a helper for
498 * find_css_set(), and checks to see whether an existing
499 * css_set is suitable.
501 * oldcg: the cgroup group that we're using before the cgroup
504 * cgrp: the cgroup that we're moving into
506 * template: location in which to build the desired set of subsystem
507 * state objects for the new cgroup group
509 static struct css_set *find_existing_css_set(
510 struct css_set *oldcg,
512 struct cgroup_subsys_state *template[])
515 struct cgroupfs_root *root = cgrp->root;
516 struct hlist_head *hhead;
517 struct hlist_node *node;
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 (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
526 if (root->subsys_bits & (1UL << i)) {
527 /* Subsystem is in this hierarchy. So we want
528 * the subsystem state from the new
530 template[i] = cgrp->subsys[i];
532 /* Subsystem is not in this hierarchy, so we
533 * don't want to change the subsystem state */
534 template[i] = oldcg->subsys[i];
538 hhead = css_set_hash(template);
539 hlist_for_each_entry(cg, node, hhead, hlist) {
540 if (!compare_css_sets(cg, oldcg, cgrp, template))
543 /* This css_set matches what we need */
547 /* No existing cgroup group matched */
551 static void free_cg_links(struct list_head *tmp)
553 struct cg_cgroup_link *link;
554 struct cg_cgroup_link *saved_link;
556 list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
557 list_del(&link->cgrp_link_list);
563 * allocate_cg_links() allocates "count" cg_cgroup_link structures
564 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
565 * success or a negative error
567 static int allocate_cg_links(int count, struct list_head *tmp)
569 struct cg_cgroup_link *link;
572 for (i = 0; i < count; i++) {
573 link = kmalloc(sizeof(*link), GFP_KERNEL);
578 list_add(&link->cgrp_link_list, tmp);
584 * link_css_set - a helper function to link a css_set to a cgroup
585 * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
586 * @cg: the css_set to be linked
587 * @cgrp: the destination cgroup
589 static void link_css_set(struct list_head *tmp_cg_links,
590 struct css_set *cg, struct cgroup *cgrp)
592 struct cg_cgroup_link *link;
594 BUG_ON(list_empty(tmp_cg_links));
595 link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
599 atomic_inc(&cgrp->count);
600 list_move(&link->cgrp_link_list, &cgrp->css_sets);
602 * Always add links to the tail of the list so that the list
603 * is sorted by order of hierarchy creation
605 list_add_tail(&link->cg_link_list, &cg->cg_links);
609 * find_css_set() takes an existing cgroup group and a
610 * cgroup object, and returns a css_set object that's
611 * equivalent to the old group, but with the given cgroup
612 * substituted into the appropriate hierarchy. Must be called with
615 static struct css_set *find_css_set(
616 struct css_set *oldcg, struct cgroup *cgrp)
619 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
621 struct list_head tmp_cg_links;
623 struct hlist_head *hhead;
624 struct cg_cgroup_link *link;
626 /* First see if we already have a cgroup group that matches
628 read_lock(&css_set_lock);
629 res = find_existing_css_set(oldcg, cgrp, template);
632 read_unlock(&css_set_lock);
637 res = kmalloc(sizeof(*res), GFP_KERNEL);
641 /* Allocate all the cg_cgroup_link objects that we'll need */
642 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
647 atomic_set(&res->refcount, 1);
648 INIT_LIST_HEAD(&res->cg_links);
649 INIT_LIST_HEAD(&res->tasks);
650 INIT_HLIST_NODE(&res->hlist);
652 /* Copy the set of subsystem state objects generated in
653 * find_existing_css_set() */
654 memcpy(res->subsys, template, sizeof(res->subsys));
656 write_lock(&css_set_lock);
657 /* Add reference counts and links from the new css_set. */
658 list_for_each_entry(link, &oldcg->cg_links, cg_link_list) {
659 struct cgroup *c = link->cgrp;
660 if (c->root == cgrp->root)
662 link_css_set(&tmp_cg_links, res, c);
665 BUG_ON(!list_empty(&tmp_cg_links));
669 /* Add this cgroup group to the hash table */
670 hhead = css_set_hash(res->subsys);
671 hlist_add_head(&res->hlist, hhead);
673 write_unlock(&css_set_lock);
679 * Return the cgroup for "task" from the given hierarchy. Must be
680 * called with cgroup_mutex held.
682 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
683 struct cgroupfs_root *root)
686 struct cgroup *res = NULL;
688 BUG_ON(!mutex_is_locked(&cgroup_mutex));
689 read_lock(&css_set_lock);
691 * No need to lock the task - since we hold cgroup_mutex the
692 * task can't change groups, so the only thing that can happen
693 * is that it exits and its css is set back to init_css_set.
696 if (css == &init_css_set) {
697 res = &root->top_cgroup;
699 struct cg_cgroup_link *link;
700 list_for_each_entry(link, &css->cg_links, cg_link_list) {
701 struct cgroup *c = link->cgrp;
702 if (c->root == root) {
708 read_unlock(&css_set_lock);
714 * There is one global cgroup mutex. We also require taking
715 * task_lock() when dereferencing a task's cgroup subsys pointers.
716 * See "The task_lock() exception", at the end of this comment.
718 * A task must hold cgroup_mutex to modify cgroups.
720 * Any task can increment and decrement the count field without lock.
721 * So in general, code holding cgroup_mutex can't rely on the count
722 * field not changing. However, if the count goes to zero, then only
723 * cgroup_attach_task() can increment it again. Because a count of zero
724 * means that no tasks are currently attached, therefore there is no
725 * way a task attached to that cgroup can fork (the other way to
726 * increment the count). So code holding cgroup_mutex can safely
727 * assume that if the count is zero, it will stay zero. Similarly, if
728 * a task holds cgroup_mutex on a cgroup with zero count, it
729 * knows that the cgroup won't be removed, as cgroup_rmdir()
732 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
733 * (usually) take cgroup_mutex. These are the two most performance
734 * critical pieces of code here. The exception occurs on cgroup_exit(),
735 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
736 * is taken, and if the cgroup count is zero, a usermode call made
737 * to the release agent with the name of the cgroup (path relative to
738 * the root of cgroup file system) as the argument.
740 * A cgroup can only be deleted if both its 'count' of using tasks
741 * is zero, and its list of 'children' cgroups is empty. Since all
742 * tasks in the system use _some_ cgroup, and since there is always at
743 * least one task in the system (init, pid == 1), therefore, top_cgroup
744 * always has either children cgroups and/or using tasks. So we don't
745 * need a special hack to ensure that top_cgroup cannot be deleted.
747 * The task_lock() exception
749 * The need for this exception arises from the action of
750 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
751 * another. It does so using cgroup_mutex, however there are
752 * several performance critical places that need to reference
753 * task->cgroups without the expense of grabbing a system global
754 * mutex. Therefore except as noted below, when dereferencing or, as
755 * in cgroup_attach_task(), modifying a task's cgroups pointer we use
756 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
757 * the task_struct routinely used for such matters.
759 * P.S. One more locking exception. RCU is used to guard the
760 * update of a tasks cgroup pointer by cgroup_attach_task()
764 * cgroup_lock - lock out any changes to cgroup structures
767 void cgroup_lock(void)
769 mutex_lock(&cgroup_mutex);
771 EXPORT_SYMBOL_GPL(cgroup_lock);
774 * cgroup_unlock - release lock on cgroup changes
776 * Undo the lock taken in a previous cgroup_lock() call.
778 void cgroup_unlock(void)
780 mutex_unlock(&cgroup_mutex);
782 EXPORT_SYMBOL_GPL(cgroup_unlock);
785 * A couple of forward declarations required, due to cyclic reference loop:
786 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
787 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
791 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
792 static struct dentry *cgroup_lookup(struct inode *, struct dentry *, struct nameidata *);
793 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
794 static int cgroup_populate_dir(struct cgroup *cgrp);
795 static const struct inode_operations cgroup_dir_inode_operations;
796 static const struct file_operations proc_cgroupstats_operations;
798 static struct backing_dev_info cgroup_backing_dev_info = {
800 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
803 static int alloc_css_id(struct cgroup_subsys *ss,
804 struct cgroup *parent, struct cgroup *child);
806 static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
808 struct inode *inode = new_inode(sb);
811 inode->i_ino = get_next_ino();
812 inode->i_mode = mode;
813 inode->i_uid = current_fsuid();
814 inode->i_gid = current_fsgid();
815 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
816 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
822 * Call subsys's pre_destroy handler.
823 * This is called before css refcnt check.
825 static int cgroup_call_pre_destroy(struct cgroup *cgrp)
827 struct cgroup_subsys *ss;
830 for_each_subsys(cgrp->root, ss)
831 if (ss->pre_destroy) {
832 ret = ss->pre_destroy(ss, cgrp);
840 static void free_cgroup_rcu(struct rcu_head *obj)
842 struct cgroup *cgrp = container_of(obj, struct cgroup, rcu_head);
847 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
849 /* is dentry a directory ? if so, kfree() associated cgroup */
850 if (S_ISDIR(inode->i_mode)) {
851 struct cgroup *cgrp = dentry->d_fsdata;
852 struct cgroup_subsys *ss;
853 BUG_ON(!(cgroup_is_removed(cgrp)));
854 /* It's possible for external users to be holding css
855 * reference counts on a cgroup; css_put() needs to
856 * be able to access the cgroup after decrementing
857 * the reference count in order to know if it needs to
858 * queue the cgroup to be handled by the release
862 mutex_lock(&cgroup_mutex);
864 * Release the subsystem state objects.
866 for_each_subsys(cgrp->root, ss)
867 ss->destroy(ss, cgrp);
869 cgrp->root->number_of_cgroups--;
870 mutex_unlock(&cgroup_mutex);
873 * Drop the active superblock reference that we took when we
876 deactivate_super(cgrp->root->sb);
879 * if we're getting rid of the cgroup, refcount should ensure
880 * that there are no pidlists left.
882 BUG_ON(!list_empty(&cgrp->pidlists));
884 call_rcu(&cgrp->rcu_head, free_cgroup_rcu);
889 static int cgroup_delete(const struct dentry *d)
894 static void remove_dir(struct dentry *d)
896 struct dentry *parent = dget(d->d_parent);
899 simple_rmdir(parent->d_inode, d);
903 static void cgroup_clear_directory(struct dentry *dentry)
905 struct list_head *node;
907 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
908 spin_lock(&dentry->d_lock);
909 node = dentry->d_subdirs.next;
910 while (node != &dentry->d_subdirs) {
911 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
913 spin_lock_nested(&d->d_lock, DENTRY_D_LOCK_NESTED);
916 /* This should never be called on a cgroup
917 * directory with child cgroups */
918 BUG_ON(d->d_inode->i_mode & S_IFDIR);
920 spin_unlock(&d->d_lock);
921 spin_unlock(&dentry->d_lock);
923 simple_unlink(dentry->d_inode, d);
925 spin_lock(&dentry->d_lock);
927 spin_unlock(&d->d_lock);
928 node = dentry->d_subdirs.next;
930 spin_unlock(&dentry->d_lock);
934 * NOTE : the dentry must have been dget()'ed
936 static void cgroup_d_remove_dir(struct dentry *dentry)
938 struct dentry *parent;
940 cgroup_clear_directory(dentry);
942 parent = dentry->d_parent;
943 spin_lock(&parent->d_lock);
944 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
945 list_del_init(&dentry->d_u.d_child);
946 spin_unlock(&dentry->d_lock);
947 spin_unlock(&parent->d_lock);
952 * Call with cgroup_mutex held. Drops reference counts on modules, including
953 * any duplicate ones that parse_cgroupfs_options took. If this function
954 * returns an error, no reference counts are touched.
956 static int rebind_subsystems(struct cgroupfs_root *root,
957 unsigned long final_bits)
959 unsigned long added_bits, removed_bits;
960 struct cgroup *cgrp = &root->top_cgroup;
963 BUG_ON(!mutex_is_locked(&cgroup_mutex));
965 removed_bits = root->actual_subsys_bits & ~final_bits;
966 added_bits = final_bits & ~root->actual_subsys_bits;
967 /* Check that any added subsystems are currently free */
968 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
969 unsigned long bit = 1UL << i;
970 struct cgroup_subsys *ss = subsys[i];
971 if (!(bit & added_bits))
974 * Nobody should tell us to do a subsys that doesn't exist:
975 * parse_cgroupfs_options should catch that case and refcounts
976 * ensure that subsystems won't disappear once selected.
979 if (ss->root != &rootnode) {
980 /* Subsystem isn't free */
985 /* Currently we don't handle adding/removing subsystems when
986 * any child cgroups exist. This is theoretically supportable
987 * but involves complex error handling, so it's being left until
989 if (root->number_of_cgroups > 1)
992 /* Process each subsystem */
993 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
994 struct cgroup_subsys *ss = subsys[i];
995 unsigned long bit = 1UL << i;
996 if (bit & added_bits) {
997 /* We're binding this subsystem to this hierarchy */
999 BUG_ON(cgrp->subsys[i]);
1000 BUG_ON(!dummytop->subsys[i]);
1001 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
1002 mutex_lock(&ss->hierarchy_mutex);
1003 cgrp->subsys[i] = dummytop->subsys[i];
1004 cgrp->subsys[i]->cgroup = cgrp;
1005 list_move(&ss->sibling, &root->subsys_list);
1009 mutex_unlock(&ss->hierarchy_mutex);
1010 /* refcount was already taken, and we're keeping it */
1011 } else if (bit & removed_bits) {
1012 /* We're removing this subsystem */
1014 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
1015 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
1016 mutex_lock(&ss->hierarchy_mutex);
1018 ss->bind(ss, dummytop);
1019 dummytop->subsys[i]->cgroup = dummytop;
1020 cgrp->subsys[i] = NULL;
1021 subsys[i]->root = &rootnode;
1022 list_move(&ss->sibling, &rootnode.subsys_list);
1023 mutex_unlock(&ss->hierarchy_mutex);
1024 /* subsystem is now free - drop reference on module */
1025 module_put(ss->module);
1026 } else if (bit & final_bits) {
1027 /* Subsystem state should already exist */
1029 BUG_ON(!cgrp->subsys[i]);
1031 * a refcount was taken, but we already had one, so
1032 * drop the extra reference.
1034 module_put(ss->module);
1035 #ifdef CONFIG_MODULE_UNLOAD
1036 BUG_ON(ss->module && !module_refcount(ss->module));
1039 /* Subsystem state shouldn't exist */
1040 BUG_ON(cgrp->subsys[i]);
1043 root->subsys_bits = root->actual_subsys_bits = final_bits;
1049 static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
1051 struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
1052 struct cgroup_subsys *ss;
1054 mutex_lock(&cgroup_mutex);
1055 for_each_subsys(root, ss)
1056 seq_printf(seq, ",%s", ss->name);
1057 if (test_bit(ROOT_NOPREFIX, &root->flags))
1058 seq_puts(seq, ",noprefix");
1059 if (strlen(root->release_agent_path))
1060 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1061 if (clone_children(&root->top_cgroup))
1062 seq_puts(seq, ",clone_children");
1063 if (strlen(root->name))
1064 seq_printf(seq, ",name=%s", root->name);
1065 mutex_unlock(&cgroup_mutex);
1069 struct cgroup_sb_opts {
1070 unsigned long subsys_bits;
1071 unsigned long flags;
1072 char *release_agent;
1073 bool clone_children;
1075 /* User explicitly requested empty subsystem */
1078 struct cgroupfs_root *new_root;
1083 * Convert a hierarchy specifier into a bitmask of subsystems and flags. Call
1084 * with cgroup_mutex held to protect the subsys[] array. This function takes
1085 * refcounts on subsystems to be used, unless it returns error, in which case
1086 * no refcounts are taken.
1088 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1090 char *token, *o = data;
1091 bool all_ss = false, one_ss = false;
1092 unsigned long mask = (unsigned long)-1;
1094 bool module_pin_failed = false;
1096 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1098 #ifdef CONFIG_CPUSETS
1099 mask = ~(1UL << cpuset_subsys_id);
1102 memset(opts, 0, sizeof(*opts));
1104 while ((token = strsep(&o, ",")) != NULL) {
1107 if (!strcmp(token, "none")) {
1108 /* Explicitly have no subsystems */
1112 if (!strcmp(token, "all")) {
1113 /* Mutually exclusive option 'all' + subsystem name */
1119 if (!strcmp(token, "noprefix")) {
1120 set_bit(ROOT_NOPREFIX, &opts->flags);
1123 if (!strcmp(token, "clone_children")) {
1124 opts->clone_children = true;
1127 if (!strncmp(token, "release_agent=", 14)) {
1128 /* Specifying two release agents is forbidden */
1129 if (opts->release_agent)
1131 opts->release_agent =
1132 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1133 if (!opts->release_agent)
1137 if (!strncmp(token, "name=", 5)) {
1138 const char *name = token + 5;
1139 /* Can't specify an empty name */
1142 /* Must match [\w.-]+ */
1143 for (i = 0; i < strlen(name); i++) {
1147 if ((c == '.') || (c == '-') || (c == '_'))
1151 /* Specifying two names is forbidden */
1154 opts->name = kstrndup(name,
1155 MAX_CGROUP_ROOT_NAMELEN - 1,
1163 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1164 struct cgroup_subsys *ss = subsys[i];
1167 if (strcmp(token, ss->name))
1172 /* Mutually exclusive option 'all' + subsystem name */
1175 set_bit(i, &opts->subsys_bits);
1180 if (i == CGROUP_SUBSYS_COUNT)
1185 * If the 'all' option was specified select all the subsystems,
1186 * otherwise 'all, 'none' and a subsystem name options were not
1187 * specified, let's default to 'all'
1189 if (all_ss || (!all_ss && !one_ss && !opts->none)) {
1190 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1191 struct cgroup_subsys *ss = subsys[i];
1196 set_bit(i, &opts->subsys_bits);
1200 /* Consistency checks */
1203 * Option noprefix was introduced just for backward compatibility
1204 * with the old cpuset, so we allow noprefix only if mounting just
1205 * the cpuset subsystem.
1207 if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
1208 (opts->subsys_bits & mask))
1212 /* Can't specify "none" and some subsystems */
1213 if (opts->subsys_bits && opts->none)
1217 * We either have to specify by name or by subsystems. (So all
1218 * empty hierarchies must have a name).
1220 if (!opts->subsys_bits && !opts->name)
1224 * Grab references on all the modules we'll need, so the subsystems
1225 * don't dance around before rebind_subsystems attaches them. This may
1226 * take duplicate reference counts on a subsystem that's already used,
1227 * but rebind_subsystems handles this case.
1229 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1230 unsigned long bit = 1UL << i;
1232 if (!(bit & opts->subsys_bits))
1234 if (!try_module_get(subsys[i]->module)) {
1235 module_pin_failed = true;
1239 if (module_pin_failed) {
1241 * oops, one of the modules was going away. this means that we
1242 * raced with a module_delete call, and to the user this is
1243 * essentially a "subsystem doesn't exist" case.
1245 for (i--; i >= CGROUP_BUILTIN_SUBSYS_COUNT; i--) {
1246 /* drop refcounts only on the ones we took */
1247 unsigned long bit = 1UL << i;
1249 if (!(bit & opts->subsys_bits))
1251 module_put(subsys[i]->module);
1259 static void drop_parsed_module_refcounts(unsigned long subsys_bits)
1262 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1263 unsigned long bit = 1UL << i;
1265 if (!(bit & subsys_bits))
1267 module_put(subsys[i]->module);
1271 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1274 struct cgroupfs_root *root = sb->s_fs_info;
1275 struct cgroup *cgrp = &root->top_cgroup;
1276 struct cgroup_sb_opts opts;
1278 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1279 mutex_lock(&cgroup_mutex);
1281 /* See what subsystems are wanted */
1282 ret = parse_cgroupfs_options(data, &opts);
1286 /* Don't allow flags or name to change at remount */
1287 if (opts.flags != root->flags ||
1288 (opts.name && strcmp(opts.name, root->name))) {
1290 drop_parsed_module_refcounts(opts.subsys_bits);
1294 ret = rebind_subsystems(root, opts.subsys_bits);
1296 drop_parsed_module_refcounts(opts.subsys_bits);
1300 /* (re)populate subsystem files */
1301 cgroup_populate_dir(cgrp);
1303 if (opts.release_agent)
1304 strcpy(root->release_agent_path, opts.release_agent);
1306 kfree(opts.release_agent);
1308 mutex_unlock(&cgroup_mutex);
1309 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1313 static const struct super_operations cgroup_ops = {
1314 .statfs = simple_statfs,
1315 .drop_inode = generic_delete_inode,
1316 .show_options = cgroup_show_options,
1317 .remount_fs = cgroup_remount,
1320 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1322 INIT_LIST_HEAD(&cgrp->sibling);
1323 INIT_LIST_HEAD(&cgrp->children);
1324 INIT_LIST_HEAD(&cgrp->css_sets);
1325 INIT_LIST_HEAD(&cgrp->release_list);
1326 INIT_LIST_HEAD(&cgrp->pidlists);
1327 mutex_init(&cgrp->pidlist_mutex);
1328 INIT_LIST_HEAD(&cgrp->event_list);
1329 spin_lock_init(&cgrp->event_list_lock);
1332 static void init_cgroup_root(struct cgroupfs_root *root)
1334 struct cgroup *cgrp = &root->top_cgroup;
1335 INIT_LIST_HEAD(&root->subsys_list);
1336 INIT_LIST_HEAD(&root->root_list);
1337 root->number_of_cgroups = 1;
1339 cgrp->top_cgroup = cgrp;
1340 init_cgroup_housekeeping(cgrp);
1343 static bool init_root_id(struct cgroupfs_root *root)
1348 if (!ida_pre_get(&hierarchy_ida, GFP_KERNEL))
1350 spin_lock(&hierarchy_id_lock);
1351 /* Try to allocate the next unused ID */
1352 ret = ida_get_new_above(&hierarchy_ida, next_hierarchy_id,
1353 &root->hierarchy_id);
1355 /* Try again starting from 0 */
1356 ret = ida_get_new(&hierarchy_ida, &root->hierarchy_id);
1358 next_hierarchy_id = root->hierarchy_id + 1;
1359 } else if (ret != -EAGAIN) {
1360 /* Can only get here if the 31-bit IDR is full ... */
1363 spin_unlock(&hierarchy_id_lock);
1368 static int cgroup_test_super(struct super_block *sb, void *data)
1370 struct cgroup_sb_opts *opts = data;
1371 struct cgroupfs_root *root = sb->s_fs_info;
1373 /* If we asked for a name then it must match */
1374 if (opts->name && strcmp(opts->name, root->name))
1378 * If we asked for subsystems (or explicitly for no
1379 * subsystems) then they must match
1381 if ((opts->subsys_bits || opts->none)
1382 && (opts->subsys_bits != root->subsys_bits))
1388 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1390 struct cgroupfs_root *root;
1392 if (!opts->subsys_bits && !opts->none)
1395 root = kzalloc(sizeof(*root), GFP_KERNEL);
1397 return ERR_PTR(-ENOMEM);
1399 if (!init_root_id(root)) {
1401 return ERR_PTR(-ENOMEM);
1403 init_cgroup_root(root);
1405 root->subsys_bits = opts->subsys_bits;
1406 root->flags = opts->flags;
1407 if (opts->release_agent)
1408 strcpy(root->release_agent_path, opts->release_agent);
1410 strcpy(root->name, opts->name);
1411 if (opts->clone_children)
1412 set_bit(CGRP_CLONE_CHILDREN, &root->top_cgroup.flags);
1416 static void cgroup_drop_root(struct cgroupfs_root *root)
1421 BUG_ON(!root->hierarchy_id);
1422 spin_lock(&hierarchy_id_lock);
1423 ida_remove(&hierarchy_ida, root->hierarchy_id);
1424 spin_unlock(&hierarchy_id_lock);
1428 static int cgroup_set_super(struct super_block *sb, void *data)
1431 struct cgroup_sb_opts *opts = data;
1433 /* If we don't have a new root, we can't set up a new sb */
1434 if (!opts->new_root)
1437 BUG_ON(!opts->subsys_bits && !opts->none);
1439 ret = set_anon_super(sb, NULL);
1443 sb->s_fs_info = opts->new_root;
1444 opts->new_root->sb = sb;
1446 sb->s_blocksize = PAGE_CACHE_SIZE;
1447 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1448 sb->s_magic = CGROUP_SUPER_MAGIC;
1449 sb->s_op = &cgroup_ops;
1454 static int cgroup_get_rootdir(struct super_block *sb)
1456 static const struct dentry_operations cgroup_dops = {
1457 .d_iput = cgroup_diput,
1458 .d_delete = cgroup_delete,
1461 struct inode *inode =
1462 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1463 struct dentry *dentry;
1468 inode->i_fop = &simple_dir_operations;
1469 inode->i_op = &cgroup_dir_inode_operations;
1470 /* directories start off with i_nlink == 2 (for "." entry) */
1472 dentry = d_alloc_root(inode);
1477 sb->s_root = dentry;
1478 /* for everything else we want ->d_op set */
1479 sb->s_d_op = &cgroup_dops;
1483 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1484 int flags, const char *unused_dev_name,
1487 struct cgroup_sb_opts opts;
1488 struct cgroupfs_root *root;
1490 struct super_block *sb;
1491 struct cgroupfs_root *new_root;
1493 /* First find the desired set of subsystems */
1494 mutex_lock(&cgroup_mutex);
1495 ret = parse_cgroupfs_options(data, &opts);
1496 mutex_unlock(&cgroup_mutex);
1501 * Allocate a new cgroup root. We may not need it if we're
1502 * reusing an existing hierarchy.
1504 new_root = cgroup_root_from_opts(&opts);
1505 if (IS_ERR(new_root)) {
1506 ret = PTR_ERR(new_root);
1509 opts.new_root = new_root;
1511 /* Locate an existing or new sb for this hierarchy */
1512 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
1515 cgroup_drop_root(opts.new_root);
1519 root = sb->s_fs_info;
1521 if (root == opts.new_root) {
1522 /* We used the new root structure, so this is a new hierarchy */
1523 struct list_head tmp_cg_links;
1524 struct cgroup *root_cgrp = &root->top_cgroup;
1525 struct inode *inode;
1526 struct cgroupfs_root *existing_root;
1529 BUG_ON(sb->s_root != NULL);
1531 ret = cgroup_get_rootdir(sb);
1533 goto drop_new_super;
1534 inode = sb->s_root->d_inode;
1536 mutex_lock(&inode->i_mutex);
1537 mutex_lock(&cgroup_mutex);
1539 if (strlen(root->name)) {
1540 /* Check for name clashes with existing mounts */
1541 for_each_active_root(existing_root) {
1542 if (!strcmp(existing_root->name, root->name)) {
1544 mutex_unlock(&cgroup_mutex);
1545 mutex_unlock(&inode->i_mutex);
1546 goto drop_new_super;
1552 * We're accessing css_set_count without locking
1553 * css_set_lock here, but that's OK - it can only be
1554 * increased by someone holding cgroup_lock, and
1555 * that's us. The worst that can happen is that we
1556 * have some link structures left over
1558 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1560 mutex_unlock(&cgroup_mutex);
1561 mutex_unlock(&inode->i_mutex);
1562 goto drop_new_super;
1565 ret = rebind_subsystems(root, root->subsys_bits);
1566 if (ret == -EBUSY) {
1567 mutex_unlock(&cgroup_mutex);
1568 mutex_unlock(&inode->i_mutex);
1569 free_cg_links(&tmp_cg_links);
1570 goto drop_new_super;
1573 * There must be no failure case after here, since rebinding
1574 * takes care of subsystems' refcounts, which are explicitly
1575 * dropped in the failure exit path.
1578 /* EBUSY should be the only error here */
1581 list_add(&root->root_list, &roots);
1584 sb->s_root->d_fsdata = root_cgrp;
1585 root->top_cgroup.dentry = sb->s_root;
1587 /* Link the top cgroup in this hierarchy into all
1588 * the css_set objects */
1589 write_lock(&css_set_lock);
1590 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1591 struct hlist_head *hhead = &css_set_table[i];
1592 struct hlist_node *node;
1595 hlist_for_each_entry(cg, node, hhead, hlist)
1596 link_css_set(&tmp_cg_links, cg, root_cgrp);
1598 write_unlock(&css_set_lock);
1600 free_cg_links(&tmp_cg_links);
1602 BUG_ON(!list_empty(&root_cgrp->sibling));
1603 BUG_ON(!list_empty(&root_cgrp->children));
1604 BUG_ON(root->number_of_cgroups != 1);
1606 cgroup_populate_dir(root_cgrp);
1607 mutex_unlock(&cgroup_mutex);
1608 mutex_unlock(&inode->i_mutex);
1611 * We re-used an existing hierarchy - the new root (if
1612 * any) is not needed
1614 cgroup_drop_root(opts.new_root);
1615 /* no subsys rebinding, so refcounts don't change */
1616 drop_parsed_module_refcounts(opts.subsys_bits);
1619 kfree(opts.release_agent);
1621 return dget(sb->s_root);
1624 deactivate_locked_super(sb);
1626 drop_parsed_module_refcounts(opts.subsys_bits);
1628 kfree(opts.release_agent);
1630 return ERR_PTR(ret);
1633 static void cgroup_kill_sb(struct super_block *sb) {
1634 struct cgroupfs_root *root = sb->s_fs_info;
1635 struct cgroup *cgrp = &root->top_cgroup;
1637 struct cg_cgroup_link *link;
1638 struct cg_cgroup_link *saved_link;
1642 BUG_ON(root->number_of_cgroups != 1);
1643 BUG_ON(!list_empty(&cgrp->children));
1644 BUG_ON(!list_empty(&cgrp->sibling));
1646 mutex_lock(&cgroup_mutex);
1648 /* Rebind all subsystems back to the default hierarchy */
1649 ret = rebind_subsystems(root, 0);
1650 /* Shouldn't be able to fail ... */
1654 * Release all the links from css_sets to this hierarchy's
1657 write_lock(&css_set_lock);
1659 list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1661 list_del(&link->cg_link_list);
1662 list_del(&link->cgrp_link_list);
1665 write_unlock(&css_set_lock);
1667 if (!list_empty(&root->root_list)) {
1668 list_del(&root->root_list);
1672 mutex_unlock(&cgroup_mutex);
1674 kill_litter_super(sb);
1675 cgroup_drop_root(root);
1678 static struct file_system_type cgroup_fs_type = {
1680 .mount = cgroup_mount,
1681 .kill_sb = cgroup_kill_sb,
1684 static struct kobject *cgroup_kobj;
1686 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1688 return dentry->d_fsdata;
1691 static inline struct cftype *__d_cft(struct dentry *dentry)
1693 return dentry->d_fsdata;
1697 * cgroup_path - generate the path of a cgroup
1698 * @cgrp: the cgroup in question
1699 * @buf: the buffer to write the path into
1700 * @buflen: the length of the buffer
1702 * Called with cgroup_mutex held or else with an RCU-protected cgroup
1703 * reference. Writes path of cgroup into buf. Returns 0 on success,
1706 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1709 struct dentry *dentry = rcu_dereference_check(cgrp->dentry,
1710 rcu_read_lock_held() ||
1711 cgroup_lock_is_held());
1713 if (!dentry || cgrp == dummytop) {
1715 * Inactive subsystems have no dentry for their root
1722 start = buf + buflen;
1726 int len = dentry->d_name.len;
1728 if ((start -= len) < buf)
1729 return -ENAMETOOLONG;
1730 memcpy(start, dentry->d_name.name, len);
1731 cgrp = cgrp->parent;
1735 dentry = rcu_dereference_check(cgrp->dentry,
1736 rcu_read_lock_held() ||
1737 cgroup_lock_is_held());
1741 return -ENAMETOOLONG;
1744 memmove(buf, start, buf + buflen - start);
1747 EXPORT_SYMBOL_GPL(cgroup_path);
1750 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1751 * @cgrp: the cgroup the task is attaching to
1752 * @tsk: the task to be attached
1754 * Call holding cgroup_mutex. May take task_lock of
1755 * the task 'tsk' during call.
1757 int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1760 struct cgroup_subsys *ss, *failed_ss = NULL;
1761 struct cgroup *oldcgrp;
1763 struct css_set *newcg;
1764 struct cgroupfs_root *root = cgrp->root;
1766 /* Nothing to do if the task is already in that cgroup */
1767 oldcgrp = task_cgroup_from_root(tsk, root);
1768 if (cgrp == oldcgrp)
1771 for_each_subsys(root, ss) {
1772 if (ss->can_attach) {
1773 retval = ss->can_attach(ss, cgrp, tsk, false);
1776 * Remember on which subsystem the can_attach()
1777 * failed, so that we only call cancel_attach()
1778 * against the subsystems whose can_attach()
1779 * succeeded. (See below)
1784 } else if (!capable(CAP_SYS_ADMIN)) {
1785 const struct cred *cred = current_cred(), *tcred;
1787 /* No can_attach() - check perms generically */
1788 tcred = __task_cred(tsk);
1789 if (cred->euid != tcred->uid &&
1790 cred->euid != tcred->suid) {
1801 * Locate or allocate a new css_set for this task,
1802 * based on its final set of cgroups
1804 newcg = find_css_set(cg, cgrp);
1812 if (tsk->flags & PF_EXITING) {
1818 rcu_assign_pointer(tsk->cgroups, newcg);
1821 /* Update the css_set linked lists if we're using them */
1822 write_lock(&css_set_lock);
1823 if (!list_empty(&tsk->cg_list))
1824 list_move(&tsk->cg_list, &newcg->tasks);
1825 write_unlock(&css_set_lock);
1827 for_each_subsys(root, ss) {
1829 ss->attach(ss, cgrp, oldcgrp, tsk, false);
1831 set_bit(CGRP_RELEASABLE, &cgrp->flags);
1832 /* put_css_set will not destroy cg until after an RCU grace period */
1836 * wake up rmdir() waiter. the rmdir should fail since the cgroup
1837 * is no longer empty.
1839 cgroup_wakeup_rmdir_waiter(cgrp);
1842 for_each_subsys(root, ss) {
1843 if (ss == failed_ss)
1845 * This subsystem was the one that failed the
1846 * can_attach() check earlier, so we don't need
1847 * to call cancel_attach() against it or any
1848 * remaining subsystems.
1851 if (ss->cancel_attach)
1852 ss->cancel_attach(ss, cgrp, tsk, false);
1859 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1860 * @from: attach to all cgroups of a given task
1861 * @tsk: the task to be attached
1863 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1865 struct cgroupfs_root *root;
1869 for_each_active_root(root) {
1870 struct cgroup *from_cg = task_cgroup_from_root(from, root);
1872 retval = cgroup_attach_task(from_cg, tsk);
1880 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
1883 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
1884 * held. May take task_lock of task
1886 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid)
1888 struct task_struct *tsk;
1893 tsk = find_task_by_vpid(pid);
1894 if (!tsk || tsk->flags & PF_EXITING) {
1898 get_task_struct(tsk);
1902 get_task_struct(tsk);
1905 ret = cgroup_attach_task(cgrp, tsk);
1906 put_task_struct(tsk);
1910 static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
1913 if (!cgroup_lock_live_group(cgrp))
1915 ret = attach_task_by_pid(cgrp, pid);
1921 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
1922 * @cgrp: the cgroup to be checked for liveness
1924 * On success, returns true; the lock should be later released with
1925 * cgroup_unlock(). On failure returns false with no lock held.
1927 bool cgroup_lock_live_group(struct cgroup *cgrp)
1929 mutex_lock(&cgroup_mutex);
1930 if (cgroup_is_removed(cgrp)) {
1931 mutex_unlock(&cgroup_mutex);
1936 EXPORT_SYMBOL_GPL(cgroup_lock_live_group);
1938 static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
1941 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1942 if (strlen(buffer) >= PATH_MAX)
1944 if (!cgroup_lock_live_group(cgrp))
1946 strcpy(cgrp->root->release_agent_path, buffer);
1951 static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
1952 struct seq_file *seq)
1954 if (!cgroup_lock_live_group(cgrp))
1956 seq_puts(seq, cgrp->root->release_agent_path);
1957 seq_putc(seq, '\n');
1962 /* A buffer size big enough for numbers or short strings */
1963 #define CGROUP_LOCAL_BUFFER_SIZE 64
1965 static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
1967 const char __user *userbuf,
1968 size_t nbytes, loff_t *unused_ppos)
1970 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
1976 if (nbytes >= sizeof(buffer))
1978 if (copy_from_user(buffer, userbuf, nbytes))
1981 buffer[nbytes] = 0; /* nul-terminate */
1982 if (cft->write_u64) {
1983 u64 val = simple_strtoull(strstrip(buffer), &end, 0);
1986 retval = cft->write_u64(cgrp, cft, val);
1988 s64 val = simple_strtoll(strstrip(buffer), &end, 0);
1991 retval = cft->write_s64(cgrp, cft, val);
1998 static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
2000 const char __user *userbuf,
2001 size_t nbytes, loff_t *unused_ppos)
2003 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
2005 size_t max_bytes = cft->max_write_len;
2006 char *buffer = local_buffer;
2009 max_bytes = sizeof(local_buffer) - 1;
2010 if (nbytes >= max_bytes)
2012 /* Allocate a dynamic buffer if we need one */
2013 if (nbytes >= sizeof(local_buffer)) {
2014 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
2018 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
2023 buffer[nbytes] = 0; /* nul-terminate */
2024 retval = cft->write_string(cgrp, cft, strstrip(buffer));
2028 if (buffer != local_buffer)
2033 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
2034 size_t nbytes, loff_t *ppos)
2036 struct cftype *cft = __d_cft(file->f_dentry);
2037 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2039 if (cgroup_is_removed(cgrp))
2042 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
2043 if (cft->write_u64 || cft->write_s64)
2044 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
2045 if (cft->write_string)
2046 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
2048 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
2049 return ret ? ret : nbytes;
2054 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
2056 char __user *buf, size_t nbytes,
2059 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2060 u64 val = cft->read_u64(cgrp, cft);
2061 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
2063 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2066 static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
2068 char __user *buf, size_t nbytes,
2071 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2072 s64 val = cft->read_s64(cgrp, cft);
2073 int len = sprintf(tmp, "%lld\n", (long long) val);
2075 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2078 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2079 size_t nbytes, loff_t *ppos)
2081 struct cftype *cft = __d_cft(file->f_dentry);
2082 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2084 if (cgroup_is_removed(cgrp))
2088 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2090 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2092 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2097 * seqfile ops/methods for returning structured data. Currently just
2098 * supports string->u64 maps, but can be extended in future.
2101 struct cgroup_seqfile_state {
2103 struct cgroup *cgroup;
2106 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2108 struct seq_file *sf = cb->state;
2109 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2112 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2114 struct cgroup_seqfile_state *state = m->private;
2115 struct cftype *cft = state->cft;
2116 if (cft->read_map) {
2117 struct cgroup_map_cb cb = {
2118 .fill = cgroup_map_add,
2121 return cft->read_map(state->cgroup, cft, &cb);
2123 return cft->read_seq_string(state->cgroup, cft, m);
2126 static int cgroup_seqfile_release(struct inode *inode, struct file *file)
2128 struct seq_file *seq = file->private_data;
2129 kfree(seq->private);
2130 return single_release(inode, file);
2133 static const struct file_operations cgroup_seqfile_operations = {
2135 .write = cgroup_file_write,
2136 .llseek = seq_lseek,
2137 .release = cgroup_seqfile_release,
2140 static int cgroup_file_open(struct inode *inode, struct file *file)
2145 err = generic_file_open(inode, file);
2148 cft = __d_cft(file->f_dentry);
2150 if (cft->read_map || cft->read_seq_string) {
2151 struct cgroup_seqfile_state *state =
2152 kzalloc(sizeof(*state), GFP_USER);
2156 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
2157 file->f_op = &cgroup_seqfile_operations;
2158 err = single_open(file, cgroup_seqfile_show, state);
2161 } else if (cft->open)
2162 err = cft->open(inode, file);
2169 static int cgroup_file_release(struct inode *inode, struct file *file)
2171 struct cftype *cft = __d_cft(file->f_dentry);
2173 return cft->release(inode, file);
2178 * cgroup_rename - Only allow simple rename of directories in place.
2180 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2181 struct inode *new_dir, struct dentry *new_dentry)
2183 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2185 if (new_dentry->d_inode)
2187 if (old_dir != new_dir)
2189 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2192 static const struct file_operations cgroup_file_operations = {
2193 .read = cgroup_file_read,
2194 .write = cgroup_file_write,
2195 .llseek = generic_file_llseek,
2196 .open = cgroup_file_open,
2197 .release = cgroup_file_release,
2200 static const struct inode_operations cgroup_dir_inode_operations = {
2201 .lookup = cgroup_lookup,
2202 .mkdir = cgroup_mkdir,
2203 .rmdir = cgroup_rmdir,
2204 .rename = cgroup_rename,
2207 static struct dentry *cgroup_lookup(struct inode *dir, struct dentry *dentry, struct nameidata *nd)
2209 if (dentry->d_name.len > NAME_MAX)
2210 return ERR_PTR(-ENAMETOOLONG);
2211 d_add(dentry, NULL);
2216 * Check if a file is a control file
2218 static inline struct cftype *__file_cft(struct file *file)
2220 if (file->f_dentry->d_inode->i_fop != &cgroup_file_operations)
2221 return ERR_PTR(-EINVAL);
2222 return __d_cft(file->f_dentry);
2225 static int cgroup_create_file(struct dentry *dentry, mode_t mode,
2226 struct super_block *sb)
2228 struct inode *inode;
2232 if (dentry->d_inode)
2235 inode = cgroup_new_inode(mode, sb);
2239 if (S_ISDIR(mode)) {
2240 inode->i_op = &cgroup_dir_inode_operations;
2241 inode->i_fop = &simple_dir_operations;
2243 /* start off with i_nlink == 2 (for "." entry) */
2246 /* start with the directory inode held, so that we can
2247 * populate it without racing with another mkdir */
2248 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2249 } else if (S_ISREG(mode)) {
2251 inode->i_fop = &cgroup_file_operations;
2253 d_instantiate(dentry, inode);
2254 dget(dentry); /* Extra count - pin the dentry in core */
2259 * cgroup_create_dir - create a directory for an object.
2260 * @cgrp: the cgroup we create the directory for. It must have a valid
2261 * ->parent field. And we are going to fill its ->dentry field.
2262 * @dentry: dentry of the new cgroup
2263 * @mode: mode to set on new directory.
2265 static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
2268 struct dentry *parent;
2271 parent = cgrp->parent->dentry;
2272 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
2274 dentry->d_fsdata = cgrp;
2275 inc_nlink(parent->d_inode);
2276 rcu_assign_pointer(cgrp->dentry, dentry);
2285 * cgroup_file_mode - deduce file mode of a control file
2286 * @cft: the control file in question
2288 * returns cft->mode if ->mode is not 0
2289 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2290 * returns S_IRUGO if it has only a read handler
2291 * returns S_IWUSR if it has only a write hander
2293 static mode_t cgroup_file_mode(const struct cftype *cft)
2300 if (cft->read || cft->read_u64 || cft->read_s64 ||
2301 cft->read_map || cft->read_seq_string)
2304 if (cft->write || cft->write_u64 || cft->write_s64 ||
2305 cft->write_string || cft->trigger)
2311 int cgroup_add_file(struct cgroup *cgrp,
2312 struct cgroup_subsys *subsys,
2313 const struct cftype *cft)
2315 struct dentry *dir = cgrp->dentry;
2316 struct dentry *dentry;
2320 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2321 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
2322 strcpy(name, subsys->name);
2325 strcat(name, cft->name);
2326 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2327 dentry = lookup_one_len(name, dir, strlen(name));
2328 if (!IS_ERR(dentry)) {
2329 mode = cgroup_file_mode(cft);
2330 error = cgroup_create_file(dentry, mode | S_IFREG,
2333 dentry->d_fsdata = (void *)cft;
2336 error = PTR_ERR(dentry);
2339 EXPORT_SYMBOL_GPL(cgroup_add_file);
2341 int cgroup_add_files(struct cgroup *cgrp,
2342 struct cgroup_subsys *subsys,
2343 const struct cftype cft[],
2347 for (i = 0; i < count; i++) {
2348 err = cgroup_add_file(cgrp, subsys, &cft[i]);
2354 EXPORT_SYMBOL_GPL(cgroup_add_files);
2357 * cgroup_task_count - count the number of tasks in a cgroup.
2358 * @cgrp: the cgroup in question
2360 * Return the number of tasks in the cgroup.
2362 int cgroup_task_count(const struct cgroup *cgrp)
2365 struct cg_cgroup_link *link;
2367 read_lock(&css_set_lock);
2368 list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
2369 count += atomic_read(&link->cg->refcount);
2371 read_unlock(&css_set_lock);
2376 * Advance a list_head iterator. The iterator should be positioned at
2377 * the start of a css_set
2379 static void cgroup_advance_iter(struct cgroup *cgrp,
2380 struct cgroup_iter *it)
2382 struct list_head *l = it->cg_link;
2383 struct cg_cgroup_link *link;
2386 /* Advance to the next non-empty css_set */
2389 if (l == &cgrp->css_sets) {
2393 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
2395 } while (list_empty(&cg->tasks));
2397 it->task = cg->tasks.next;
2401 * To reduce the fork() overhead for systems that are not actually
2402 * using their cgroups capability, we don't maintain the lists running
2403 * through each css_set to its tasks until we see the list actually
2404 * used - in other words after the first call to cgroup_iter_start().
2406 * The tasklist_lock is not held here, as do_each_thread() and
2407 * while_each_thread() are protected by RCU.
2409 static void cgroup_enable_task_cg_lists(void)
2411 struct task_struct *p, *g;
2412 write_lock(&css_set_lock);
2413 use_task_css_set_links = 1;
2414 do_each_thread(g, p) {
2417 * We should check if the process is exiting, otherwise
2418 * it will race with cgroup_exit() in that the list
2419 * entry won't be deleted though the process has exited.
2421 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2422 list_add(&p->cg_list, &p->cgroups->tasks);
2424 } while_each_thread(g, p);
2425 write_unlock(&css_set_lock);
2428 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
2431 * The first time anyone tries to iterate across a cgroup,
2432 * we need to enable the list linking each css_set to its
2433 * tasks, and fix up all existing tasks.
2435 if (!use_task_css_set_links)
2436 cgroup_enable_task_cg_lists();
2438 read_lock(&css_set_lock);
2439 it->cg_link = &cgrp->css_sets;
2440 cgroup_advance_iter(cgrp, it);
2443 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
2444 struct cgroup_iter *it)
2446 struct task_struct *res;
2447 struct list_head *l = it->task;
2448 struct cg_cgroup_link *link;
2450 /* If the iterator cg is NULL, we have no tasks */
2453 res = list_entry(l, struct task_struct, cg_list);
2454 /* Advance iterator to find next entry */
2456 link = list_entry(it->cg_link, struct cg_cgroup_link, cgrp_link_list);
2457 if (l == &link->cg->tasks) {
2458 /* We reached the end of this task list - move on to
2459 * the next cg_cgroup_link */
2460 cgroup_advance_iter(cgrp, it);
2467 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
2469 read_unlock(&css_set_lock);
2472 static inline int started_after_time(struct task_struct *t1,
2473 struct timespec *time,
2474 struct task_struct *t2)
2476 int start_diff = timespec_compare(&t1->start_time, time);
2477 if (start_diff > 0) {
2479 } else if (start_diff < 0) {
2483 * Arbitrarily, if two processes started at the same
2484 * time, we'll say that the lower pointer value
2485 * started first. Note that t2 may have exited by now
2486 * so this may not be a valid pointer any longer, but
2487 * that's fine - it still serves to distinguish
2488 * between two tasks started (effectively) simultaneously.
2495 * This function is a callback from heap_insert() and is used to order
2497 * In this case we order the heap in descending task start time.
2499 static inline int started_after(void *p1, void *p2)
2501 struct task_struct *t1 = p1;
2502 struct task_struct *t2 = p2;
2503 return started_after_time(t1, &t2->start_time, t2);
2507 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
2508 * @scan: struct cgroup_scanner containing arguments for the scan
2510 * Arguments include pointers to callback functions test_task() and
2512 * Iterate through all the tasks in a cgroup, calling test_task() for each,
2513 * and if it returns true, call process_task() for it also.
2514 * The test_task pointer may be NULL, meaning always true (select all tasks).
2515 * Effectively duplicates cgroup_iter_{start,next,end}()
2516 * but does not lock css_set_lock for the call to process_task().
2517 * The struct cgroup_scanner may be embedded in any structure of the caller's
2519 * It is guaranteed that process_task() will act on every task that
2520 * is a member of the cgroup for the duration of this call. This
2521 * function may or may not call process_task() for tasks that exit
2522 * or move to a different cgroup during the call, or are forked or
2523 * move into the cgroup during the call.
2525 * Note that test_task() may be called with locks held, and may in some
2526 * situations be called multiple times for the same task, so it should
2528 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
2529 * pre-allocated and will be used for heap operations (and its "gt" member will
2530 * be overwritten), else a temporary heap will be used (allocation of which
2531 * may cause this function to fail).
2533 int cgroup_scan_tasks(struct cgroup_scanner *scan)
2536 struct cgroup_iter it;
2537 struct task_struct *p, *dropped;
2538 /* Never dereference latest_task, since it's not refcounted */
2539 struct task_struct *latest_task = NULL;
2540 struct ptr_heap tmp_heap;
2541 struct ptr_heap *heap;
2542 struct timespec latest_time = { 0, 0 };
2545 /* The caller supplied our heap and pre-allocated its memory */
2547 heap->gt = &started_after;
2549 /* We need to allocate our own heap memory */
2551 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2553 /* cannot allocate the heap */
2559 * Scan tasks in the cgroup, using the scanner's "test_task" callback
2560 * to determine which are of interest, and using the scanner's
2561 * "process_task" callback to process any of them that need an update.
2562 * Since we don't want to hold any locks during the task updates,
2563 * gather tasks to be processed in a heap structure.
2564 * The heap is sorted by descending task start time.
2565 * If the statically-sized heap fills up, we overflow tasks that
2566 * started later, and in future iterations only consider tasks that
2567 * started after the latest task in the previous pass. This
2568 * guarantees forward progress and that we don't miss any tasks.
2571 cgroup_iter_start(scan->cg, &it);
2572 while ((p = cgroup_iter_next(scan->cg, &it))) {
2574 * Only affect tasks that qualify per the caller's callback,
2575 * if he provided one
2577 if (scan->test_task && !scan->test_task(p, scan))
2580 * Only process tasks that started after the last task
2583 if (!started_after_time(p, &latest_time, latest_task))
2585 dropped = heap_insert(heap, p);
2586 if (dropped == NULL) {
2588 * The new task was inserted; the heap wasn't
2592 } else if (dropped != p) {
2594 * The new task was inserted, and pushed out a
2598 put_task_struct(dropped);
2601 * Else the new task was newer than anything already in
2602 * the heap and wasn't inserted
2605 cgroup_iter_end(scan->cg, &it);
2608 for (i = 0; i < heap->size; i++) {
2609 struct task_struct *q = heap->ptrs[i];
2611 latest_time = q->start_time;
2614 /* Process the task per the caller's callback */
2615 scan->process_task(q, scan);
2619 * If we had to process any tasks at all, scan again
2620 * in case some of them were in the middle of forking
2621 * children that didn't get processed.
2622 * Not the most efficient way to do it, but it avoids
2623 * having to take callback_mutex in the fork path
2627 if (heap == &tmp_heap)
2628 heap_free(&tmp_heap);
2633 * Stuff for reading the 'tasks'/'procs' files.
2635 * Reading this file can return large amounts of data if a cgroup has
2636 * *lots* of attached tasks. So it may need several calls to read(),
2637 * but we cannot guarantee that the information we produce is correct
2638 * unless we produce it entirely atomically.
2643 * The following two functions "fix" the issue where there are more pids
2644 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2645 * TODO: replace with a kernel-wide solution to this problem
2647 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2648 static void *pidlist_allocate(int count)
2650 if (PIDLIST_TOO_LARGE(count))
2651 return vmalloc(count * sizeof(pid_t));
2653 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2655 static void pidlist_free(void *p)
2657 if (is_vmalloc_addr(p))
2662 static void *pidlist_resize(void *p, int newcount)
2665 /* note: if new alloc fails, old p will still be valid either way */
2666 if (is_vmalloc_addr(p)) {
2667 newlist = vmalloc(newcount * sizeof(pid_t));
2670 memcpy(newlist, p, newcount * sizeof(pid_t));
2673 newlist = krealloc(p, newcount * sizeof(pid_t), GFP_KERNEL);
2679 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2680 * If the new stripped list is sufficiently smaller and there's enough memory
2681 * to allocate a new buffer, will let go of the unneeded memory. Returns the
2682 * number of unique elements.
2684 /* is the size difference enough that we should re-allocate the array? */
2685 #define PIDLIST_REALLOC_DIFFERENCE(old, new) ((old) - PAGE_SIZE >= (new))
2686 static int pidlist_uniq(pid_t **p, int length)
2693 * we presume the 0th element is unique, so i starts at 1. trivial
2694 * edge cases first; no work needs to be done for either
2696 if (length == 0 || length == 1)
2698 /* src and dest walk down the list; dest counts unique elements */
2699 for (src = 1; src < length; src++) {
2700 /* find next unique element */
2701 while (list[src] == list[src-1]) {
2706 /* dest always points to where the next unique element goes */
2707 list[dest] = list[src];
2712 * if the length difference is large enough, we want to allocate a
2713 * smaller buffer to save memory. if this fails due to out of memory,
2714 * we'll just stay with what we've got.
2716 if (PIDLIST_REALLOC_DIFFERENCE(length, dest)) {
2717 newlist = pidlist_resize(list, dest);
2724 static int cmppid(const void *a, const void *b)
2726 return *(pid_t *)a - *(pid_t *)b;
2730 * find the appropriate pidlist for our purpose (given procs vs tasks)
2731 * returns with the lock on that pidlist already held, and takes care
2732 * of the use count, or returns NULL with no locks held if we're out of
2735 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
2736 enum cgroup_filetype type)
2738 struct cgroup_pidlist *l;
2739 /* don't need task_nsproxy() if we're looking at ourself */
2740 struct pid_namespace *ns = current->nsproxy->pid_ns;
2743 * We can't drop the pidlist_mutex before taking the l->mutex in case
2744 * the last ref-holder is trying to remove l from the list at the same
2745 * time. Holding the pidlist_mutex precludes somebody taking whichever
2746 * list we find out from under us - compare release_pid_array().
2748 mutex_lock(&cgrp->pidlist_mutex);
2749 list_for_each_entry(l, &cgrp->pidlists, links) {
2750 if (l->key.type == type && l->key.ns == ns) {
2751 /* make sure l doesn't vanish out from under us */
2752 down_write(&l->mutex);
2753 mutex_unlock(&cgrp->pidlist_mutex);
2757 /* entry not found; create a new one */
2758 l = kmalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
2760 mutex_unlock(&cgrp->pidlist_mutex);
2763 init_rwsem(&l->mutex);
2764 down_write(&l->mutex);
2766 l->key.ns = get_pid_ns(ns);
2767 l->use_count = 0; /* don't increment here */
2770 list_add(&l->links, &cgrp->pidlists);
2771 mutex_unlock(&cgrp->pidlist_mutex);
2776 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
2778 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
2779 struct cgroup_pidlist **lp)
2783 int pid, n = 0; /* used for populating the array */
2784 struct cgroup_iter it;
2785 struct task_struct *tsk;
2786 struct cgroup_pidlist *l;
2789 * If cgroup gets more users after we read count, we won't have
2790 * enough space - tough. This race is indistinguishable to the
2791 * caller from the case that the additional cgroup users didn't
2792 * show up until sometime later on.
2794 length = cgroup_task_count(cgrp);
2795 array = pidlist_allocate(length);
2798 /* now, populate the array */
2799 cgroup_iter_start(cgrp, &it);
2800 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2801 if (unlikely(n == length))
2803 /* get tgid or pid for procs or tasks file respectively */
2804 if (type == CGROUP_FILE_PROCS)
2805 pid = task_tgid_vnr(tsk);
2807 pid = task_pid_vnr(tsk);
2808 if (pid > 0) /* make sure to only use valid results */
2811 cgroup_iter_end(cgrp, &it);
2813 /* now sort & (if procs) strip out duplicates */
2814 sort(array, length, sizeof(pid_t), cmppid, NULL);
2815 if (type == CGROUP_FILE_PROCS)
2816 length = pidlist_uniq(&array, length);
2817 l = cgroup_pidlist_find(cgrp, type);
2819 pidlist_free(array);
2822 /* store array, freeing old if necessary - lock already held */
2823 pidlist_free(l->list);
2827 up_write(&l->mutex);
2833 * cgroupstats_build - build and fill cgroupstats
2834 * @stats: cgroupstats to fill information into
2835 * @dentry: A dentry entry belonging to the cgroup for which stats have
2838 * Build and fill cgroupstats so that taskstats can export it to user
2841 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2844 struct cgroup *cgrp;
2845 struct cgroup_iter it;
2846 struct task_struct *tsk;
2849 * Validate dentry by checking the superblock operations,
2850 * and make sure it's a directory.
2852 if (dentry->d_sb->s_op != &cgroup_ops ||
2853 !S_ISDIR(dentry->d_inode->i_mode))
2857 cgrp = dentry->d_fsdata;
2859 cgroup_iter_start(cgrp, &it);
2860 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2861 switch (tsk->state) {
2863 stats->nr_running++;
2865 case TASK_INTERRUPTIBLE:
2866 stats->nr_sleeping++;
2868 case TASK_UNINTERRUPTIBLE:
2869 stats->nr_uninterruptible++;
2872 stats->nr_stopped++;
2875 if (delayacct_is_task_waiting_on_io(tsk))
2876 stats->nr_io_wait++;
2880 cgroup_iter_end(cgrp, &it);
2888 * seq_file methods for the tasks/procs files. The seq_file position is the
2889 * next pid to display; the seq_file iterator is a pointer to the pid
2890 * in the cgroup->l->list array.
2893 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
2896 * Initially we receive a position value that corresponds to
2897 * one more than the last pid shown (or 0 on the first call or
2898 * after a seek to the start). Use a binary-search to find the
2899 * next pid to display, if any
2901 struct cgroup_pidlist *l = s->private;
2902 int index = 0, pid = *pos;
2905 down_read(&l->mutex);
2907 int end = l->length;
2909 while (index < end) {
2910 int mid = (index + end) / 2;
2911 if (l->list[mid] == pid) {
2914 } else if (l->list[mid] <= pid)
2920 /* If we're off the end of the array, we're done */
2921 if (index >= l->length)
2923 /* Update the abstract position to be the actual pid that we found */
2924 iter = l->list + index;
2929 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
2931 struct cgroup_pidlist *l = s->private;
2935 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
2937 struct cgroup_pidlist *l = s->private;
2939 pid_t *end = l->list + l->length;
2941 * Advance to the next pid in the array. If this goes off the
2953 static int cgroup_pidlist_show(struct seq_file *s, void *v)
2955 return seq_printf(s, "%d\n", *(int *)v);
2959 * seq_operations functions for iterating on pidlists through seq_file -
2960 * independent of whether it's tasks or procs
2962 static const struct seq_operations cgroup_pidlist_seq_operations = {
2963 .start = cgroup_pidlist_start,
2964 .stop = cgroup_pidlist_stop,
2965 .next = cgroup_pidlist_next,
2966 .show = cgroup_pidlist_show,
2969 static void cgroup_release_pid_array(struct cgroup_pidlist *l)
2972 * the case where we're the last user of this particular pidlist will
2973 * have us remove it from the cgroup's list, which entails taking the
2974 * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
2975 * pidlist_mutex, we have to take pidlist_mutex first.
2977 mutex_lock(&l->owner->pidlist_mutex);
2978 down_write(&l->mutex);
2979 BUG_ON(!l->use_count);
2980 if (!--l->use_count) {
2981 /* we're the last user if refcount is 0; remove and free */
2982 list_del(&l->links);
2983 mutex_unlock(&l->owner->pidlist_mutex);
2984 pidlist_free(l->list);
2985 put_pid_ns(l->key.ns);
2986 up_write(&l->mutex);
2990 mutex_unlock(&l->owner->pidlist_mutex);
2991 up_write(&l->mutex);
2994 static int cgroup_pidlist_release(struct inode *inode, struct file *file)
2996 struct cgroup_pidlist *l;
2997 if (!(file->f_mode & FMODE_READ))
3000 * the seq_file will only be initialized if the file was opened for
3001 * reading; hence we check if it's not null only in that case.
3003 l = ((struct seq_file *)file->private_data)->private;
3004 cgroup_release_pid_array(l);
3005 return seq_release(inode, file);
3008 static const struct file_operations cgroup_pidlist_operations = {
3010 .llseek = seq_lseek,
3011 .write = cgroup_file_write,
3012 .release = cgroup_pidlist_release,
3016 * The following functions handle opens on a file that displays a pidlist
3017 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
3020 /* helper function for the two below it */
3021 static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
3023 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
3024 struct cgroup_pidlist *l;
3027 /* Nothing to do for write-only files */
3028 if (!(file->f_mode & FMODE_READ))
3031 /* have the array populated */
3032 retval = pidlist_array_load(cgrp, type, &l);
3035 /* configure file information */
3036 file->f_op = &cgroup_pidlist_operations;
3038 retval = seq_open(file, &cgroup_pidlist_seq_operations);
3040 cgroup_release_pid_array(l);
3043 ((struct seq_file *)file->private_data)->private = l;
3046 static int cgroup_tasks_open(struct inode *unused, struct file *file)
3048 return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
3050 static int cgroup_procs_open(struct inode *unused, struct file *file)
3052 return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
3055 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
3058 return notify_on_release(cgrp);
3061 static int cgroup_write_notify_on_release(struct cgroup *cgrp,
3065 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
3067 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3069 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3074 * Unregister event and free resources.
3076 * Gets called from workqueue.
3078 static void cgroup_event_remove(struct work_struct *work)
3080 struct cgroup_event *event = container_of(work, struct cgroup_event,
3082 struct cgroup *cgrp = event->cgrp;
3084 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3086 eventfd_ctx_put(event->eventfd);
3092 * Gets called on POLLHUP on eventfd when user closes it.
3094 * Called with wqh->lock held and interrupts disabled.
3096 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3097 int sync, void *key)
3099 struct cgroup_event *event = container_of(wait,
3100 struct cgroup_event, wait);
3101 struct cgroup *cgrp = event->cgrp;
3102 unsigned long flags = (unsigned long)key;
3104 if (flags & POLLHUP) {
3105 __remove_wait_queue(event->wqh, &event->wait);
3106 spin_lock(&cgrp->event_list_lock);
3107 list_del(&event->list);
3108 spin_unlock(&cgrp->event_list_lock);
3110 * We are in atomic context, but cgroup_event_remove() may
3111 * sleep, so we have to call it in workqueue.
3113 schedule_work(&event->remove);
3119 static void cgroup_event_ptable_queue_proc(struct file *file,
3120 wait_queue_head_t *wqh, poll_table *pt)
3122 struct cgroup_event *event = container_of(pt,
3123 struct cgroup_event, pt);
3126 add_wait_queue(wqh, &event->wait);
3130 * Parse input and register new cgroup event handler.
3132 * Input must be in format '<event_fd> <control_fd> <args>'.
3133 * Interpretation of args is defined by control file implementation.
3135 static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3138 struct cgroup_event *event = NULL;
3139 unsigned int efd, cfd;
3140 struct file *efile = NULL;
3141 struct file *cfile = NULL;
3145 efd = simple_strtoul(buffer, &endp, 10);
3150 cfd = simple_strtoul(buffer, &endp, 10);
3151 if ((*endp != ' ') && (*endp != '\0'))
3155 event = kzalloc(sizeof(*event), GFP_KERNEL);
3159 INIT_LIST_HEAD(&event->list);
3160 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
3161 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
3162 INIT_WORK(&event->remove, cgroup_event_remove);
3164 efile = eventfd_fget(efd);
3165 if (IS_ERR(efile)) {
3166 ret = PTR_ERR(efile);
3170 event->eventfd = eventfd_ctx_fileget(efile);
3171 if (IS_ERR(event->eventfd)) {
3172 ret = PTR_ERR(event->eventfd);
3182 /* the process need read permission on control file */
3183 ret = file_permission(cfile, MAY_READ);
3187 event->cft = __file_cft(cfile);
3188 if (IS_ERR(event->cft)) {
3189 ret = PTR_ERR(event->cft);
3193 if (!event->cft->register_event || !event->cft->unregister_event) {
3198 ret = event->cft->register_event(cgrp, event->cft,
3199 event->eventfd, buffer);
3203 if (efile->f_op->poll(efile, &event->pt) & POLLHUP) {
3204 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3210 * Events should be removed after rmdir of cgroup directory, but before
3211 * destroying subsystem state objects. Let's take reference to cgroup
3212 * directory dentry to do that.
3216 spin_lock(&cgrp->event_list_lock);
3217 list_add(&event->list, &cgrp->event_list);
3218 spin_unlock(&cgrp->event_list_lock);
3229 if (event && event->eventfd && !IS_ERR(event->eventfd))
3230 eventfd_ctx_put(event->eventfd);
3232 if (!IS_ERR_OR_NULL(efile))
3240 static u64 cgroup_clone_children_read(struct cgroup *cgrp,
3243 return clone_children(cgrp);
3246 static int cgroup_clone_children_write(struct cgroup *cgrp,
3251 set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3253 clear_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3258 * for the common functions, 'private' gives the type of file
3260 /* for hysterical raisins, we can't put this on the older files */
3261 #define CGROUP_FILE_GENERIC_PREFIX "cgroup."
3262 static struct cftype files[] = {
3265 .open = cgroup_tasks_open,
3266 .write_u64 = cgroup_tasks_write,
3267 .release = cgroup_pidlist_release,
3268 .mode = S_IRUGO | S_IWUSR,
3271 .name = CGROUP_FILE_GENERIC_PREFIX "procs",
3272 .open = cgroup_procs_open,
3273 /* .write_u64 = cgroup_procs_write, TODO */
3274 .release = cgroup_pidlist_release,
3278 .name = "notify_on_release",
3279 .read_u64 = cgroup_read_notify_on_release,
3280 .write_u64 = cgroup_write_notify_on_release,
3283 .name = CGROUP_FILE_GENERIC_PREFIX "event_control",
3284 .write_string = cgroup_write_event_control,
3288 .name = "cgroup.clone_children",
3289 .read_u64 = cgroup_clone_children_read,
3290 .write_u64 = cgroup_clone_children_write,
3294 static struct cftype cft_release_agent = {
3295 .name = "release_agent",
3296 .read_seq_string = cgroup_release_agent_show,
3297 .write_string = cgroup_release_agent_write,
3298 .max_write_len = PATH_MAX,
3301 static int cgroup_populate_dir(struct cgroup *cgrp)
3304 struct cgroup_subsys *ss;
3306 /* First clear out any existing files */
3307 cgroup_clear_directory(cgrp->dentry);
3309 err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
3313 if (cgrp == cgrp->top_cgroup) {
3314 if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
3318 for_each_subsys(cgrp->root, ss) {
3319 if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
3322 /* This cgroup is ready now */
3323 for_each_subsys(cgrp->root, ss) {
3324 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3326 * Update id->css pointer and make this css visible from
3327 * CSS ID functions. This pointer will be dereferened
3328 * from RCU-read-side without locks.
3331 rcu_assign_pointer(css->id->css, css);
3337 static void init_cgroup_css(struct cgroup_subsys_state *css,
3338 struct cgroup_subsys *ss,
3339 struct cgroup *cgrp)
3342 atomic_set(&css->refcnt, 1);
3345 if (cgrp == dummytop)
3346 set_bit(CSS_ROOT, &css->flags);
3347 BUG_ON(cgrp->subsys[ss->subsys_id]);
3348 cgrp->subsys[ss->subsys_id] = css;
3351 static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
3353 /* We need to take each hierarchy_mutex in a consistent order */
3357 * No worry about a race with rebind_subsystems that might mess up the
3358 * locking order, since both parties are under cgroup_mutex.
3360 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3361 struct cgroup_subsys *ss = subsys[i];
3364 if (ss->root == root)
3365 mutex_lock(&ss->hierarchy_mutex);
3369 static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
3373 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3374 struct cgroup_subsys *ss = subsys[i];
3377 if (ss->root == root)
3378 mutex_unlock(&ss->hierarchy_mutex);
3383 * cgroup_create - create a cgroup
3384 * @parent: cgroup that will be parent of the new cgroup
3385 * @dentry: dentry of the new cgroup
3386 * @mode: mode to set on new inode
3388 * Must be called with the mutex on the parent inode held
3390 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
3393 struct cgroup *cgrp;
3394 struct cgroupfs_root *root = parent->root;
3396 struct cgroup_subsys *ss;
3397 struct super_block *sb = root->sb;
3399 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3403 /* Grab a reference on the superblock so the hierarchy doesn't
3404 * get deleted on unmount if there are child cgroups. This
3405 * can be done outside cgroup_mutex, since the sb can't
3406 * disappear while someone has an open control file on the
3408 atomic_inc(&sb->s_active);
3410 mutex_lock(&cgroup_mutex);
3412 init_cgroup_housekeeping(cgrp);
3414 cgrp->parent = parent;
3415 cgrp->root = parent->root;
3416 cgrp->top_cgroup = parent->top_cgroup;
3418 if (notify_on_release(parent))
3419 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3421 if (clone_children(parent))
3422 set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3424 for_each_subsys(root, ss) {
3425 struct cgroup_subsys_state *css = ss->create(ss, cgrp);
3431 init_cgroup_css(css, ss, cgrp);
3433 err = alloc_css_id(ss, parent, cgrp);
3437 /* At error, ->destroy() callback has to free assigned ID. */
3438 if (clone_children(parent) && ss->post_clone)
3439 ss->post_clone(ss, cgrp);
3442 cgroup_lock_hierarchy(root);
3443 list_add(&cgrp->sibling, &cgrp->parent->children);
3444 cgroup_unlock_hierarchy(root);
3445 root->number_of_cgroups++;
3447 err = cgroup_create_dir(cgrp, dentry, mode);
3451 set_bit(CGRP_RELEASABLE, &parent->flags);
3453 /* The cgroup directory was pre-locked for us */
3454 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
3456 err = cgroup_populate_dir(cgrp);
3457 /* If err < 0, we have a half-filled directory - oh well ;) */
3459 mutex_unlock(&cgroup_mutex);
3460 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
3466 cgroup_lock_hierarchy(root);
3467 list_del(&cgrp->sibling);
3468 cgroup_unlock_hierarchy(root);
3469 root->number_of_cgroups--;
3473 for_each_subsys(root, ss) {
3474 if (cgrp->subsys[ss->subsys_id])
3475 ss->destroy(ss, cgrp);
3478 mutex_unlock(&cgroup_mutex);
3480 /* Release the reference count that we took on the superblock */
3481 deactivate_super(sb);
3487 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
3489 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
3491 /* the vfs holds inode->i_mutex already */
3492 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
3495 static int cgroup_has_css_refs(struct cgroup *cgrp)
3497 /* Check the reference count on each subsystem. Since we
3498 * already established that there are no tasks in the
3499 * cgroup, if the css refcount is also 1, then there should
3500 * be no outstanding references, so the subsystem is safe to
3501 * destroy. We scan across all subsystems rather than using
3502 * the per-hierarchy linked list of mounted subsystems since
3503 * we can be called via check_for_release() with no
3504 * synchronization other than RCU, and the subsystem linked
3505 * list isn't RCU-safe */
3508 * We won't need to lock the subsys array, because the subsystems
3509 * we're concerned about aren't going anywhere since our cgroup root
3510 * has a reference on them.
3512 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3513 struct cgroup_subsys *ss = subsys[i];
3514 struct cgroup_subsys_state *css;
3515 /* Skip subsystems not present or not in this hierarchy */
3516 if (ss == NULL || ss->root != cgrp->root)
3518 css = cgrp->subsys[ss->subsys_id];
3519 /* When called from check_for_release() it's possible
3520 * that by this point the cgroup has been removed
3521 * and the css deleted. But a false-positive doesn't
3522 * matter, since it can only happen if the cgroup
3523 * has been deleted and hence no longer needs the
3524 * release agent to be called anyway. */
3525 if (css && (atomic_read(&css->refcnt) > 1))
3532 * Atomically mark all (or else none) of the cgroup's CSS objects as
3533 * CSS_REMOVED. Return true on success, or false if the cgroup has
3534 * busy subsystems. Call with cgroup_mutex held
3537 static int cgroup_clear_css_refs(struct cgroup *cgrp)
3539 struct cgroup_subsys *ss;
3540 unsigned long flags;
3541 bool failed = false;
3542 local_irq_save(flags);
3543 for_each_subsys(cgrp->root, ss) {
3544 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3547 /* We can only remove a CSS with a refcnt==1 */
3548 refcnt = atomic_read(&css->refcnt);
3555 * Drop the refcnt to 0 while we check other
3556 * subsystems. This will cause any racing
3557 * css_tryget() to spin until we set the
3558 * CSS_REMOVED bits or abort
3560 if (atomic_cmpxchg(&css->refcnt, refcnt, 0) == refcnt)
3566 for_each_subsys(cgrp->root, ss) {
3567 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3570 * Restore old refcnt if we previously managed
3571 * to clear it from 1 to 0
3573 if (!atomic_read(&css->refcnt))
3574 atomic_set(&css->refcnt, 1);
3576 /* Commit the fact that the CSS is removed */
3577 set_bit(CSS_REMOVED, &css->flags);
3580 local_irq_restore(flags);
3584 /* checks if all of the css_sets attached to a cgroup have a refcount of 0.
3585 * Must be called with css_set_lock held */
3586 static int cgroup_css_sets_empty(struct cgroup *cgrp)
3588 struct cg_cgroup_link *link;
3590 list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
3591 struct css_set *cg = link->cg;
3592 if (atomic_read(&cg->refcount) > 0)
3599 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
3601 struct cgroup *cgrp = dentry->d_fsdata;
3603 struct cgroup *parent;
3605 struct cgroup_event *event, *tmp;
3608 /* the vfs holds both inode->i_mutex already */
3610 mutex_lock(&cgroup_mutex);
3611 if (!cgroup_css_sets_empty(cgrp)) {
3612 mutex_unlock(&cgroup_mutex);
3615 if (!list_empty(&cgrp->children)) {
3616 mutex_unlock(&cgroup_mutex);
3619 mutex_unlock(&cgroup_mutex);
3622 * In general, subsystem has no css->refcnt after pre_destroy(). But
3623 * in racy cases, subsystem may have to get css->refcnt after
3624 * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
3625 * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
3626 * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
3627 * and subsystem's reference count handling. Please see css_get/put
3628 * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
3630 set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3633 * Call pre_destroy handlers of subsys. Notify subsystems
3634 * that rmdir() request comes.
3636 ret = cgroup_call_pre_destroy(cgrp);
3638 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3642 mutex_lock(&cgroup_mutex);
3643 parent = cgrp->parent;
3644 if (!cgroup_css_sets_empty(cgrp) || !list_empty(&cgrp->children)) {
3645 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3646 mutex_unlock(&cgroup_mutex);
3649 prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
3650 if (!cgroup_clear_css_refs(cgrp)) {
3651 mutex_unlock(&cgroup_mutex);
3653 * Because someone may call cgroup_wakeup_rmdir_waiter() before
3654 * prepare_to_wait(), we need to check this flag.
3656 if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
3658 finish_wait(&cgroup_rmdir_waitq, &wait);
3659 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3660 if (signal_pending(current))
3664 /* NO css_tryget() can success after here. */
3665 finish_wait(&cgroup_rmdir_waitq, &wait);
3666 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3668 spin_lock(&release_list_lock);
3669 set_bit(CGRP_REMOVED, &cgrp->flags);
3670 if (!list_empty(&cgrp->release_list))
3671 list_del_init(&cgrp->release_list);
3672 spin_unlock(&release_list_lock);
3674 cgroup_lock_hierarchy(cgrp->root);
3675 /* delete this cgroup from parent->children */
3676 list_del_init(&cgrp->sibling);
3677 cgroup_unlock_hierarchy(cgrp->root);
3679 d = dget(cgrp->dentry);
3681 cgroup_d_remove_dir(d);
3684 check_for_release(parent);
3687 * Unregister events and notify userspace.
3688 * Notify userspace about cgroup removing only after rmdir of cgroup
3689 * directory to avoid race between userspace and kernelspace
3691 spin_lock(&cgrp->event_list_lock);
3692 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
3693 list_del(&event->list);
3694 remove_wait_queue(event->wqh, &event->wait);
3695 eventfd_signal(event->eventfd, 1);
3696 schedule_work(&event->remove);
3698 spin_unlock(&cgrp->event_list_lock);
3700 mutex_unlock(&cgroup_mutex);
3704 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
3706 struct cgroup_subsys_state *css;
3708 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
3710 /* Create the top cgroup state for this subsystem */
3711 list_add(&ss->sibling, &rootnode.subsys_list);
3712 ss->root = &rootnode;
3713 css = ss->create(ss, dummytop);
3714 /* We don't handle early failures gracefully */
3715 BUG_ON(IS_ERR(css));
3716 init_cgroup_css(css, ss, dummytop);
3718 /* Update the init_css_set to contain a subsys
3719 * pointer to this state - since the subsystem is
3720 * newly registered, all tasks and hence the
3721 * init_css_set is in the subsystem's top cgroup. */
3722 init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
3724 need_forkexit_callback |= ss->fork || ss->exit;
3726 /* At system boot, before all subsystems have been
3727 * registered, no tasks have been forked, so we don't
3728 * need to invoke fork callbacks here. */
3729 BUG_ON(!list_empty(&init_task.tasks));
3731 mutex_init(&ss->hierarchy_mutex);
3732 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3735 /* this function shouldn't be used with modular subsystems, since they
3736 * need to register a subsys_id, among other things */
3741 * cgroup_load_subsys: load and register a modular subsystem at runtime
3742 * @ss: the subsystem to load
3744 * This function should be called in a modular subsystem's initcall. If the
3745 * subsystem is built as a module, it will be assigned a new subsys_id and set
3746 * up for use. If the subsystem is built-in anyway, work is delegated to the
3747 * simpler cgroup_init_subsys.
3749 int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
3752 struct cgroup_subsys_state *css;
3754 /* check name and function validity */
3755 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
3756 ss->create == NULL || ss->destroy == NULL)
3760 * we don't support callbacks in modular subsystems. this check is
3761 * before the ss->module check for consistency; a subsystem that could
3762 * be a module should still have no callbacks even if the user isn't
3763 * compiling it as one.
3765 if (ss->fork || ss->exit)
3769 * an optionally modular subsystem is built-in: we want to do nothing,
3770 * since cgroup_init_subsys will have already taken care of it.
3772 if (ss->module == NULL) {
3773 /* a few sanity checks */
3774 BUG_ON(ss->subsys_id >= CGROUP_BUILTIN_SUBSYS_COUNT);
3775 BUG_ON(subsys[ss->subsys_id] != ss);
3780 * need to register a subsys id before anything else - for example,
3781 * init_cgroup_css needs it.
3783 mutex_lock(&cgroup_mutex);
3784 /* find the first empty slot in the array */
3785 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
3786 if (subsys[i] == NULL)
3789 if (i == CGROUP_SUBSYS_COUNT) {
3790 /* maximum number of subsystems already registered! */
3791 mutex_unlock(&cgroup_mutex);
3794 /* assign ourselves the subsys_id */
3799 * no ss->create seems to need anything important in the ss struct, so
3800 * this can happen first (i.e. before the rootnode attachment).
3802 css = ss->create(ss, dummytop);
3804 /* failure case - need to deassign the subsys[] slot. */
3806 mutex_unlock(&cgroup_mutex);
3807 return PTR_ERR(css);
3810 list_add(&ss->sibling, &rootnode.subsys_list);
3811 ss->root = &rootnode;
3813 /* our new subsystem will be attached to the dummy hierarchy. */
3814 init_cgroup_css(css, ss, dummytop);
3815 /* init_idr must be after init_cgroup_css because it sets css->id. */
3817 int ret = cgroup_init_idr(ss, css);
3819 dummytop->subsys[ss->subsys_id] = NULL;
3820 ss->destroy(ss, dummytop);
3822 mutex_unlock(&cgroup_mutex);
3828 * Now we need to entangle the css into the existing css_sets. unlike
3829 * in cgroup_init_subsys, there are now multiple css_sets, so each one
3830 * will need a new pointer to it; done by iterating the css_set_table.
3831 * furthermore, modifying the existing css_sets will corrupt the hash
3832 * table state, so each changed css_set will need its hash recomputed.
3833 * this is all done under the css_set_lock.
3835 write_lock(&css_set_lock);
3836 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
3838 struct hlist_node *node, *tmp;
3839 struct hlist_head *bucket = &css_set_table[i], *new_bucket;
3841 hlist_for_each_entry_safe(cg, node, tmp, bucket, hlist) {
3842 /* skip entries that we already rehashed */
3843 if (cg->subsys[ss->subsys_id])
3845 /* remove existing entry */
3846 hlist_del(&cg->hlist);
3848 cg->subsys[ss->subsys_id] = css;
3849 /* recompute hash and restore entry */
3850 new_bucket = css_set_hash(cg->subsys);
3851 hlist_add_head(&cg->hlist, new_bucket);
3854 write_unlock(&css_set_lock);
3856 mutex_init(&ss->hierarchy_mutex);
3857 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3861 mutex_unlock(&cgroup_mutex);
3864 EXPORT_SYMBOL_GPL(cgroup_load_subsys);
3867 * cgroup_unload_subsys: unload a modular subsystem
3868 * @ss: the subsystem to unload
3870 * This function should be called in a modular subsystem's exitcall. When this
3871 * function is invoked, the refcount on the subsystem's module will be 0, so
3872 * the subsystem will not be attached to any hierarchy.
3874 void cgroup_unload_subsys(struct cgroup_subsys *ss)
3876 struct cg_cgroup_link *link;
3877 struct hlist_head *hhead;
3879 BUG_ON(ss->module == NULL);
3882 * we shouldn't be called if the subsystem is in use, and the use of
3883 * try_module_get in parse_cgroupfs_options should ensure that it
3884 * doesn't start being used while we're killing it off.
3886 BUG_ON(ss->root != &rootnode);
3888 mutex_lock(&cgroup_mutex);
3889 /* deassign the subsys_id */
3890 BUG_ON(ss->subsys_id < CGROUP_BUILTIN_SUBSYS_COUNT);
3891 subsys[ss->subsys_id] = NULL;
3893 /* remove subsystem from rootnode's list of subsystems */
3894 list_del_init(&ss->sibling);
3897 * disentangle the css from all css_sets attached to the dummytop. as
3898 * in loading, we need to pay our respects to the hashtable gods.
3900 write_lock(&css_set_lock);
3901 list_for_each_entry(link, &dummytop->css_sets, cgrp_link_list) {
3902 struct css_set *cg = link->cg;
3904 hlist_del(&cg->hlist);
3905 BUG_ON(!cg->subsys[ss->subsys_id]);
3906 cg->subsys[ss->subsys_id] = NULL;
3907 hhead = css_set_hash(cg->subsys);
3908 hlist_add_head(&cg->hlist, hhead);
3910 write_unlock(&css_set_lock);
3913 * remove subsystem's css from the dummytop and free it - need to free
3914 * before marking as null because ss->destroy needs the cgrp->subsys
3915 * pointer to find their state. note that this also takes care of
3916 * freeing the css_id.
3918 ss->destroy(ss, dummytop);
3919 dummytop->subsys[ss->subsys_id] = NULL;
3921 mutex_unlock(&cgroup_mutex);
3923 EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
3926 * cgroup_init_early - cgroup initialization at system boot
3928 * Initialize cgroups at system boot, and initialize any
3929 * subsystems that request early init.
3931 int __init cgroup_init_early(void)
3934 atomic_set(&init_css_set.refcount, 1);
3935 INIT_LIST_HEAD(&init_css_set.cg_links);
3936 INIT_LIST_HEAD(&init_css_set.tasks);
3937 INIT_HLIST_NODE(&init_css_set.hlist);
3939 init_cgroup_root(&rootnode);
3941 init_task.cgroups = &init_css_set;
3943 init_css_set_link.cg = &init_css_set;
3944 init_css_set_link.cgrp = dummytop;
3945 list_add(&init_css_set_link.cgrp_link_list,
3946 &rootnode.top_cgroup.css_sets);
3947 list_add(&init_css_set_link.cg_link_list,
3948 &init_css_set.cg_links);
3950 for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
3951 INIT_HLIST_HEAD(&css_set_table[i]);
3953 /* at bootup time, we don't worry about modular subsystems */
3954 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3955 struct cgroup_subsys *ss = subsys[i];
3958 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
3959 BUG_ON(!ss->create);
3960 BUG_ON(!ss->destroy);
3961 if (ss->subsys_id != i) {
3962 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
3963 ss->name, ss->subsys_id);
3968 cgroup_init_subsys(ss);
3974 * cgroup_init - cgroup initialization
3976 * Register cgroup filesystem and /proc file, and initialize
3977 * any subsystems that didn't request early init.
3979 int __init cgroup_init(void)
3983 struct hlist_head *hhead;
3985 err = bdi_init(&cgroup_backing_dev_info);
3989 /* at bootup time, we don't worry about modular subsystems */
3990 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3991 struct cgroup_subsys *ss = subsys[i];
3992 if (!ss->early_init)
3993 cgroup_init_subsys(ss);
3995 cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
3998 /* Add init_css_set to the hash table */
3999 hhead = css_set_hash(init_css_set.subsys);
4000 hlist_add_head(&init_css_set.hlist, hhead);
4001 BUG_ON(!init_root_id(&rootnode));
4003 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4009 err = register_filesystem(&cgroup_fs_type);
4011 kobject_put(cgroup_kobj);
4015 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4019 bdi_destroy(&cgroup_backing_dev_info);
4025 * proc_cgroup_show()
4026 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4027 * - Used for /proc/<pid>/cgroup.
4028 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4029 * doesn't really matter if tsk->cgroup changes after we read it,
4030 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4031 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4032 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4033 * cgroup to top_cgroup.
4036 /* TODO: Use a proper seq_file iterator */
4037 static int proc_cgroup_show(struct seq_file *m, void *v)
4040 struct task_struct *tsk;
4043 struct cgroupfs_root *root;
4046 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4052 tsk = get_pid_task(pid, PIDTYPE_PID);
4058 mutex_lock(&cgroup_mutex);
4060 for_each_active_root(root) {
4061 struct cgroup_subsys *ss;
4062 struct cgroup *cgrp;
4065 seq_printf(m, "%d:", root->hierarchy_id);
4066 for_each_subsys(root, ss)
4067 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4068 if (strlen(root->name))
4069 seq_printf(m, "%sname=%s", count ? "," : "",
4072 cgrp = task_cgroup_from_root(tsk, root);
4073 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4081 mutex_unlock(&cgroup_mutex);
4082 put_task_struct(tsk);
4089 static int cgroup_open(struct inode *inode, struct file *file)
4091 struct pid *pid = PROC_I(inode)->pid;
4092 return single_open(file, proc_cgroup_show, pid);
4095 const struct file_operations proc_cgroup_operations = {
4096 .open = cgroup_open,
4098 .llseek = seq_lseek,
4099 .release = single_release,
4102 /* Display information about each subsystem and each hierarchy */
4103 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4107 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4109 * ideally we don't want subsystems moving around while we do this.
4110 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4111 * subsys/hierarchy state.
4113 mutex_lock(&cgroup_mutex);
4114 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
4115 struct cgroup_subsys *ss = subsys[i];
4118 seq_printf(m, "%s\t%d\t%d\t%d\n",
4119 ss->name, ss->root->hierarchy_id,
4120 ss->root->number_of_cgroups, !ss->disabled);
4122 mutex_unlock(&cgroup_mutex);
4126 static int cgroupstats_open(struct inode *inode, struct file *file)
4128 return single_open(file, proc_cgroupstats_show, NULL);
4131 static const struct file_operations proc_cgroupstats_operations = {
4132 .open = cgroupstats_open,
4134 .llseek = seq_lseek,
4135 .release = single_release,
4139 * cgroup_fork - attach newly forked task to its parents cgroup.
4140 * @child: pointer to task_struct of forking parent process.
4142 * Description: A task inherits its parent's cgroup at fork().
4144 * A pointer to the shared css_set was automatically copied in
4145 * fork.c by dup_task_struct(). However, we ignore that copy, since
4146 * it was not made under the protection of RCU or cgroup_mutex, so
4147 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4148 * have already changed current->cgroups, allowing the previously
4149 * referenced cgroup group to be removed and freed.
4151 * At the point that cgroup_fork() is called, 'current' is the parent
4152 * task, and the passed argument 'child' points to the child task.
4154 void cgroup_fork(struct task_struct *child)
4157 child->cgroups = current->cgroups;
4158 get_css_set(child->cgroups);
4159 task_unlock(current);
4160 INIT_LIST_HEAD(&child->cg_list);
4164 * cgroup_fork_callbacks - run fork callbacks
4165 * @child: the new task
4167 * Called on a new task very soon before adding it to the
4168 * tasklist. No need to take any locks since no-one can
4169 * be operating on this task.
4171 void cgroup_fork_callbacks(struct task_struct *child)
4173 if (need_forkexit_callback) {
4176 * forkexit callbacks are only supported for builtin
4177 * subsystems, and the builtin section of the subsys array is
4178 * immutable, so we don't need to lock the subsys array here.
4180 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4181 struct cgroup_subsys *ss = subsys[i];
4183 ss->fork(ss, child);
4189 * cgroup_post_fork - called on a new task after adding it to the task list
4190 * @child: the task in question
4192 * Adds the task to the list running through its css_set if necessary.
4193 * Has to be after the task is visible on the task list in case we race
4194 * with the first call to cgroup_iter_start() - to guarantee that the
4195 * new task ends up on its list.
4197 void cgroup_post_fork(struct task_struct *child)
4199 if (use_task_css_set_links) {
4200 write_lock(&css_set_lock);
4202 if (list_empty(&child->cg_list))
4203 list_add(&child->cg_list, &child->cgroups->tasks);
4205 write_unlock(&css_set_lock);
4209 * cgroup_exit - detach cgroup from exiting task
4210 * @tsk: pointer to task_struct of exiting process
4211 * @run_callback: run exit callbacks?
4213 * Description: Detach cgroup from @tsk and release it.
4215 * Note that cgroups marked notify_on_release force every task in
4216 * them to take the global cgroup_mutex mutex when exiting.
4217 * This could impact scaling on very large systems. Be reluctant to
4218 * use notify_on_release cgroups where very high task exit scaling
4219 * is required on large systems.
4221 * the_top_cgroup_hack:
4223 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4225 * We call cgroup_exit() while the task is still competent to
4226 * handle notify_on_release(), then leave the task attached to the
4227 * root cgroup in each hierarchy for the remainder of its exit.
4229 * To do this properly, we would increment the reference count on
4230 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4231 * code we would add a second cgroup function call, to drop that
4232 * reference. This would just create an unnecessary hot spot on
4233 * the top_cgroup reference count, to no avail.
4235 * Normally, holding a reference to a cgroup without bumping its
4236 * count is unsafe. The cgroup could go away, or someone could
4237 * attach us to a different cgroup, decrementing the count on
4238 * the first cgroup that we never incremented. But in this case,
4239 * top_cgroup isn't going away, and either task has PF_EXITING set,
4240 * which wards off any cgroup_attach_task() attempts, or task is a failed
4241 * fork, never visible to cgroup_attach_task.
4243 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4249 * Unlink from the css_set task list if necessary.
4250 * Optimistically check cg_list before taking
4253 if (!list_empty(&tsk->cg_list)) {
4254 write_lock(&css_set_lock);
4255 if (!list_empty(&tsk->cg_list))
4256 list_del_init(&tsk->cg_list);
4257 write_unlock(&css_set_lock);
4260 /* Reassign the task to the init_css_set. */
4263 tsk->cgroups = &init_css_set;
4265 if (run_callbacks && need_forkexit_callback) {
4267 * modular subsystems can't use callbacks, so no need to lock
4270 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4271 struct cgroup_subsys *ss = subsys[i];
4273 struct cgroup *old_cgrp =
4274 rcu_dereference_raw(cg->subsys[i])->cgroup;
4275 struct cgroup *cgrp = task_cgroup(tsk, i);
4276 ss->exit(ss, cgrp, old_cgrp, tsk);
4287 * cgroup_clone - clone the cgroup the given subsystem is attached to
4288 * @tsk: the task to be moved
4289 * @subsys: the given subsystem
4290 * @nodename: the name for the new cgroup
4292 * Duplicate the current cgroup in the hierarchy that the given
4293 * subsystem is attached to, and move this task into the new
4296 int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
4299 struct dentry *dentry;
4301 struct cgroup *parent, *child;
4302 struct inode *inode;
4304 struct cgroupfs_root *root;
4305 struct cgroup_subsys *ss;
4307 /* We shouldn't be called by an unregistered subsystem */
4308 BUG_ON(!subsys->active);
4310 /* First figure out what hierarchy and cgroup we're dealing
4311 * with, and pin them so we can drop cgroup_mutex */
4312 mutex_lock(&cgroup_mutex);
4314 root = subsys->root;
4315 if (root == &rootnode) {
4316 mutex_unlock(&cgroup_mutex);
4320 /* Pin the hierarchy */
4321 if (!atomic_inc_not_zero(&root->sb->s_active)) {
4322 /* We race with the final deactivate_super() */
4323 mutex_unlock(&cgroup_mutex);
4327 /* Keep the cgroup alive */
4329 parent = task_cgroup(tsk, subsys->subsys_id);
4334 mutex_unlock(&cgroup_mutex);
4336 /* Now do the VFS work to create a cgroup */
4337 inode = parent->dentry->d_inode;
4339 /* Hold the parent directory mutex across this operation to
4340 * stop anyone else deleting the new cgroup */
4341 mutex_lock(&inode->i_mutex);
4342 dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
4343 if (IS_ERR(dentry)) {
4345 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
4347 ret = PTR_ERR(dentry);
4351 /* Create the cgroup directory, which also creates the cgroup */
4352 ret = vfs_mkdir(inode, dentry, 0755);
4353 child = __d_cgrp(dentry);
4357 "Failed to create cgroup %s: %d\n", nodename,
4362 /* The cgroup now exists. Retake cgroup_mutex and check
4363 * that we're still in the same state that we thought we
4365 mutex_lock(&cgroup_mutex);
4366 if ((root != subsys->root) ||
4367 (parent != task_cgroup(tsk, subsys->subsys_id))) {
4368 /* Aargh, we raced ... */
4369 mutex_unlock(&inode->i_mutex);
4372 deactivate_super(root->sb);
4373 /* The cgroup is still accessible in the VFS, but
4374 * we're not going to try to rmdir() it at this
4377 "Race in cgroup_clone() - leaking cgroup %s\n",
4382 /* do any required auto-setup */
4383 for_each_subsys(root, ss) {
4385 ss->post_clone(ss, child);
4388 /* All seems fine. Finish by moving the task into the new cgroup */
4389 ret = cgroup_attach_task(child, tsk);
4390 mutex_unlock(&cgroup_mutex);
4393 mutex_unlock(&inode->i_mutex);
4395 mutex_lock(&cgroup_mutex);
4397 mutex_unlock(&cgroup_mutex);
4398 deactivate_super(root->sb);
4403 * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
4404 * @cgrp: the cgroup in question
4405 * @task: the task in question
4407 * See if @cgrp is a descendant of @task's cgroup in the appropriate
4410 * If we are sending in dummytop, then presumably we are creating
4411 * the top cgroup in the subsystem.
4413 * Called only by the ns (nsproxy) cgroup.
4415 int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
4418 struct cgroup *target;
4420 if (cgrp == dummytop)
4423 target = task_cgroup_from_root(task, cgrp->root);
4424 while (cgrp != target && cgrp!= cgrp->top_cgroup)
4425 cgrp = cgrp->parent;
4426 ret = (cgrp == target);
4430 static void check_for_release(struct cgroup *cgrp)
4432 /* All of these checks rely on RCU to keep the cgroup
4433 * structure alive */
4434 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
4435 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
4436 /* Control Group is currently removeable. If it's not
4437 * already queued for a userspace notification, queue
4439 int need_schedule_work = 0;
4440 spin_lock(&release_list_lock);
4441 if (!cgroup_is_removed(cgrp) &&
4442 list_empty(&cgrp->release_list)) {
4443 list_add(&cgrp->release_list, &release_list);
4444 need_schedule_work = 1;
4446 spin_unlock(&release_list_lock);
4447 if (need_schedule_work)
4448 schedule_work(&release_agent_work);
4452 /* Caller must verify that the css is not for root cgroup */
4453 void __css_get(struct cgroup_subsys_state *css, int count)
4455 atomic_add(count, &css->refcnt);
4456 set_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4458 EXPORT_SYMBOL_GPL(__css_get);
4460 /* Caller must verify that the css is not for root cgroup */
4461 void __css_put(struct cgroup_subsys_state *css, int count)
4463 struct cgroup *cgrp = css->cgroup;
4466 val = atomic_sub_return(count, &css->refcnt);
4468 check_for_release(cgrp);
4469 cgroup_wakeup_rmdir_waiter(cgrp);
4472 WARN_ON_ONCE(val < 1);
4474 EXPORT_SYMBOL_GPL(__css_put);
4477 * Notify userspace when a cgroup is released, by running the
4478 * configured release agent with the name of the cgroup (path
4479 * relative to the root of cgroup file system) as the argument.
4481 * Most likely, this user command will try to rmdir this cgroup.
4483 * This races with the possibility that some other task will be
4484 * attached to this cgroup before it is removed, or that some other
4485 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4486 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4487 * unused, and this cgroup will be reprieved from its death sentence,
4488 * to continue to serve a useful existence. Next time it's released,
4489 * we will get notified again, if it still has 'notify_on_release' set.
4491 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4492 * means only wait until the task is successfully execve()'d. The
4493 * separate release agent task is forked by call_usermodehelper(),
4494 * then control in this thread returns here, without waiting for the
4495 * release agent task. We don't bother to wait because the caller of
4496 * this routine has no use for the exit status of the release agent
4497 * task, so no sense holding our caller up for that.
4499 static void cgroup_release_agent(struct work_struct *work)
4501 BUG_ON(work != &release_agent_work);
4502 mutex_lock(&cgroup_mutex);
4503 spin_lock(&release_list_lock);
4504 while (!list_empty(&release_list)) {
4505 char *argv[3], *envp[3];
4507 char *pathbuf = NULL, *agentbuf = NULL;
4508 struct cgroup *cgrp = list_entry(release_list.next,
4511 list_del_init(&cgrp->release_list);
4512 spin_unlock(&release_list_lock);
4513 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4516 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4518 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4523 argv[i++] = agentbuf;
4524 argv[i++] = pathbuf;
4528 /* minimal command environment */
4529 envp[i++] = "HOME=/";
4530 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4533 /* Drop the lock while we invoke the usermode helper,
4534 * since the exec could involve hitting disk and hence
4535 * be a slow process */
4536 mutex_unlock(&cgroup_mutex);
4537 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4538 mutex_lock(&cgroup_mutex);
4542 spin_lock(&release_list_lock);
4544 spin_unlock(&release_list_lock);
4545 mutex_unlock(&cgroup_mutex);
4548 static int __init cgroup_disable(char *str)
4553 while ((token = strsep(&str, ",")) != NULL) {
4557 * cgroup_disable, being at boot time, can't know about module
4558 * subsystems, so we don't worry about them.
4560 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4561 struct cgroup_subsys *ss = subsys[i];
4563 if (!strcmp(token, ss->name)) {
4565 printk(KERN_INFO "Disabling %s control group"
4566 " subsystem\n", ss->name);
4573 __setup("cgroup_disable=", cgroup_disable);
4576 * Functons for CSS ID.
4580 *To get ID other than 0, this should be called when !cgroup_is_removed().
4582 unsigned short css_id(struct cgroup_subsys_state *css)
4584 struct css_id *cssid;
4587 * This css_id() can return correct value when somone has refcnt
4588 * on this or this is under rcu_read_lock(). Once css->id is allocated,
4589 * it's unchanged until freed.
4591 cssid = rcu_dereference_check(css->id,
4592 rcu_read_lock_held() || atomic_read(&css->refcnt));
4598 EXPORT_SYMBOL_GPL(css_id);
4600 unsigned short css_depth(struct cgroup_subsys_state *css)
4602 struct css_id *cssid;
4604 cssid = rcu_dereference_check(css->id,
4605 rcu_read_lock_held() || atomic_read(&css->refcnt));
4608 return cssid->depth;
4611 EXPORT_SYMBOL_GPL(css_depth);
4614 * css_is_ancestor - test "root" css is an ancestor of "child"
4615 * @child: the css to be tested.
4616 * @root: the css supporsed to be an ancestor of the child.
4618 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
4619 * this function reads css->id, this use rcu_dereference() and rcu_read_lock().
4620 * But, considering usual usage, the csses should be valid objects after test.
4621 * Assuming that the caller will do some action to the child if this returns
4622 * returns true, the caller must take "child";s reference count.
4623 * If "child" is valid object and this returns true, "root" is valid, too.
4626 bool css_is_ancestor(struct cgroup_subsys_state *child,
4627 const struct cgroup_subsys_state *root)
4629 struct css_id *child_id;
4630 struct css_id *root_id;
4634 child_id = rcu_dereference(child->id);
4635 root_id = rcu_dereference(root->id);
4638 || (child_id->depth < root_id->depth)
4639 || (child_id->stack[root_id->depth] != root_id->id))
4645 static void __free_css_id_cb(struct rcu_head *head)
4649 id = container_of(head, struct css_id, rcu_head);
4653 void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
4655 struct css_id *id = css->id;
4656 /* When this is called before css_id initialization, id can be NULL */
4660 BUG_ON(!ss->use_id);
4662 rcu_assign_pointer(id->css, NULL);
4663 rcu_assign_pointer(css->id, NULL);
4664 spin_lock(&ss->id_lock);
4665 idr_remove(&ss->idr, id->id);
4666 spin_unlock(&ss->id_lock);
4667 call_rcu(&id->rcu_head, __free_css_id_cb);
4669 EXPORT_SYMBOL_GPL(free_css_id);
4672 * This is called by init or create(). Then, calls to this function are
4673 * always serialized (By cgroup_mutex() at create()).
4676 static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
4678 struct css_id *newid;
4679 int myid, error, size;
4681 BUG_ON(!ss->use_id);
4683 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
4684 newid = kzalloc(size, GFP_KERNEL);
4686 return ERR_PTR(-ENOMEM);
4688 if (unlikely(!idr_pre_get(&ss->idr, GFP_KERNEL))) {
4692 spin_lock(&ss->id_lock);
4693 /* Don't use 0. allocates an ID of 1-65535 */
4694 error = idr_get_new_above(&ss->idr, newid, 1, &myid);
4695 spin_unlock(&ss->id_lock);
4697 /* Returns error when there are no free spaces for new ID.*/
4702 if (myid > CSS_ID_MAX)
4706 newid->depth = depth;
4710 spin_lock(&ss->id_lock);
4711 idr_remove(&ss->idr, myid);
4712 spin_unlock(&ss->id_lock);
4715 return ERR_PTR(error);
4719 static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
4720 struct cgroup_subsys_state *rootcss)
4722 struct css_id *newid;
4724 spin_lock_init(&ss->id_lock);
4727 newid = get_new_cssid(ss, 0);
4729 return PTR_ERR(newid);
4731 newid->stack[0] = newid->id;
4732 newid->css = rootcss;
4733 rootcss->id = newid;
4737 static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
4738 struct cgroup *child)
4740 int subsys_id, i, depth = 0;
4741 struct cgroup_subsys_state *parent_css, *child_css;
4742 struct css_id *child_id, *parent_id;
4744 subsys_id = ss->subsys_id;
4745 parent_css = parent->subsys[subsys_id];
4746 child_css = child->subsys[subsys_id];
4747 parent_id = parent_css->id;
4748 depth = parent_id->depth + 1;
4750 child_id = get_new_cssid(ss, depth);
4751 if (IS_ERR(child_id))
4752 return PTR_ERR(child_id);
4754 for (i = 0; i < depth; i++)
4755 child_id->stack[i] = parent_id->stack[i];
4756 child_id->stack[depth] = child_id->id;
4758 * child_id->css pointer will be set after this cgroup is available
4759 * see cgroup_populate_dir()
4761 rcu_assign_pointer(child_css->id, child_id);
4767 * css_lookup - lookup css by id
4768 * @ss: cgroup subsys to be looked into.
4771 * Returns pointer to cgroup_subsys_state if there is valid one with id.
4772 * NULL if not. Should be called under rcu_read_lock()
4774 struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
4776 struct css_id *cssid = NULL;
4778 BUG_ON(!ss->use_id);
4779 cssid = idr_find(&ss->idr, id);
4781 if (unlikely(!cssid))
4784 return rcu_dereference(cssid->css);
4786 EXPORT_SYMBOL_GPL(css_lookup);
4789 * css_get_next - lookup next cgroup under specified hierarchy.
4790 * @ss: pointer to subsystem
4791 * @id: current position of iteration.
4792 * @root: pointer to css. search tree under this.
4793 * @foundid: position of found object.
4795 * Search next css under the specified hierarchy of rootid. Calling under
4796 * rcu_read_lock() is necessary. Returns NULL if it reaches the end.
4798 struct cgroup_subsys_state *
4799 css_get_next(struct cgroup_subsys *ss, int id,
4800 struct cgroup_subsys_state *root, int *foundid)
4802 struct cgroup_subsys_state *ret = NULL;
4805 int rootid = css_id(root);
4806 int depth = css_depth(root);
4811 BUG_ON(!ss->use_id);
4812 /* fill start point for scan */
4816 * scan next entry from bitmap(tree), tmpid is updated after
4819 spin_lock(&ss->id_lock);
4820 tmp = idr_get_next(&ss->idr, &tmpid);
4821 spin_unlock(&ss->id_lock);
4825 if (tmp->depth >= depth && tmp->stack[depth] == rootid) {
4826 ret = rcu_dereference(tmp->css);
4832 /* continue to scan from next id */
4839 * get corresponding css from file open on cgroupfs directory
4841 struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id)
4843 struct cgroup *cgrp;
4844 struct inode *inode;
4845 struct cgroup_subsys_state *css;
4847 inode = f->f_dentry->d_inode;
4848 /* check in cgroup filesystem dir */
4849 if (inode->i_op != &cgroup_dir_inode_operations)
4850 return ERR_PTR(-EBADF);
4852 if (id < 0 || id >= CGROUP_SUBSYS_COUNT)
4853 return ERR_PTR(-EINVAL);
4856 cgrp = __d_cgrp(f->f_dentry);
4857 css = cgrp->subsys[id];
4858 return css ? css : ERR_PTR(-ENOENT);
4861 #ifdef CONFIG_CGROUP_DEBUG
4862 static struct cgroup_subsys_state *debug_create(struct cgroup_subsys *ss,
4863 struct cgroup *cont)
4865 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4868 return ERR_PTR(-ENOMEM);
4873 static void debug_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
4875 kfree(cont->subsys[debug_subsys_id]);
4878 static u64 cgroup_refcount_read(struct cgroup *cont, struct cftype *cft)
4880 return atomic_read(&cont->count);
4883 static u64 debug_taskcount_read(struct cgroup *cont, struct cftype *cft)
4885 return cgroup_task_count(cont);
4888 static u64 current_css_set_read(struct cgroup *cont, struct cftype *cft)
4890 return (u64)(unsigned long)current->cgroups;
4893 static u64 current_css_set_refcount_read(struct cgroup *cont,
4899 count = atomic_read(¤t->cgroups->refcount);
4904 static int current_css_set_cg_links_read(struct cgroup *cont,
4906 struct seq_file *seq)
4908 struct cg_cgroup_link *link;
4911 read_lock(&css_set_lock);
4913 cg = rcu_dereference(current->cgroups);
4914 list_for_each_entry(link, &cg->cg_links, cg_link_list) {
4915 struct cgroup *c = link->cgrp;
4919 name = c->dentry->d_name.name;
4922 seq_printf(seq, "Root %d group %s\n",
4923 c->root->hierarchy_id, name);
4926 read_unlock(&css_set_lock);
4930 #define MAX_TASKS_SHOWN_PER_CSS 25
4931 static int cgroup_css_links_read(struct cgroup *cont,
4933 struct seq_file *seq)
4935 struct cg_cgroup_link *link;
4937 read_lock(&css_set_lock);
4938 list_for_each_entry(link, &cont->css_sets, cgrp_link_list) {
4939 struct css_set *cg = link->cg;
4940 struct task_struct *task;
4942 seq_printf(seq, "css_set %p\n", cg);
4943 list_for_each_entry(task, &cg->tasks, cg_list) {
4944 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4945 seq_puts(seq, " ...\n");
4948 seq_printf(seq, " task %d\n",
4949 task_pid_vnr(task));
4953 read_unlock(&css_set_lock);
4957 static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
4959 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
4962 static struct cftype debug_files[] = {
4964 .name = "cgroup_refcount",
4965 .read_u64 = cgroup_refcount_read,
4968 .name = "taskcount",
4969 .read_u64 = debug_taskcount_read,
4973 .name = "current_css_set",
4974 .read_u64 = current_css_set_read,
4978 .name = "current_css_set_refcount",
4979 .read_u64 = current_css_set_refcount_read,
4983 .name = "current_css_set_cg_links",
4984 .read_seq_string = current_css_set_cg_links_read,
4988 .name = "cgroup_css_links",
4989 .read_seq_string = cgroup_css_links_read,
4993 .name = "releasable",
4994 .read_u64 = releasable_read,
4998 static int debug_populate(struct cgroup_subsys *ss, struct cgroup *cont)
5000 return cgroup_add_files(cont, ss, debug_files,
5001 ARRAY_SIZE(debug_files));
5004 struct cgroup_subsys debug_subsys = {
5006 .create = debug_create,
5007 .destroy = debug_destroy,
5008 .populate = debug_populate,
5009 .subsys_id = debug_subsys_id,
5011 #endif /* CONFIG_CGROUP_DEBUG */