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/smp_lock.h>
56 #include <linux/pid_namespace.h>
57 #include <linux/idr.h>
58 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
59 #include <linux/eventfd.h>
60 #include <linux/poll.h>
61 #include <linux/capability.h>
63 #include <asm/atomic.h>
65 static DEFINE_MUTEX(cgroup_mutex);
68 * Generate an array of cgroup subsystem pointers. At boot time, this is
69 * populated up to CGROUP_BUILTIN_SUBSYS_COUNT, and modular subsystems are
70 * registered after that. The mutable section of this array is protected by
73 #define SUBSYS(_x) &_x ## _subsys,
74 static struct cgroup_subsys *subsys[CGROUP_SUBSYS_COUNT] = {
75 #include <linux/cgroup_subsys.h>
78 #define MAX_CGROUP_ROOT_NAMELEN 64
81 * A cgroupfs_root represents the root of a cgroup hierarchy,
82 * and may be associated with a superblock to form an active
85 struct cgroupfs_root {
86 struct super_block *sb;
89 * The bitmask of subsystems intended to be attached to this
92 unsigned long subsys_bits;
94 /* Unique id for this hierarchy. */
97 /* The bitmask of subsystems currently attached to this hierarchy */
98 unsigned long actual_subsys_bits;
100 /* A list running through the attached subsystems */
101 struct list_head subsys_list;
103 /* The root cgroup for this hierarchy */
104 struct cgroup top_cgroup;
106 /* Tracks how many cgroups are currently defined in hierarchy.*/
107 int number_of_cgroups;
109 /* A list running through the active hierarchies */
110 struct list_head root_list;
112 /* Hierarchy-specific flags */
115 /* The path to use for release notifications. */
116 char release_agent_path[PATH_MAX];
118 /* The name for this hierarchy - may be empty */
119 char name[MAX_CGROUP_ROOT_NAMELEN];
123 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
124 * subsystems that are otherwise unattached - it never has more than a
125 * single cgroup, and all tasks are part of that cgroup.
127 static struct cgroupfs_root rootnode;
130 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
131 * cgroup_subsys->use_id != 0.
133 #define CSS_ID_MAX (65535)
136 * The css to which this ID points. This pointer is set to valid value
137 * after cgroup is populated. If cgroup is removed, this will be NULL.
138 * This pointer is expected to be RCU-safe because destroy()
139 * is called after synchronize_rcu(). But for safe use, css_is_removed()
140 * css_tryget() should be used for avoiding race.
142 struct cgroup_subsys_state *css;
148 * Depth in hierarchy which this ID belongs to.
150 unsigned short depth;
152 * ID is freed by RCU. (and lookup routine is RCU safe.)
154 struct rcu_head rcu_head;
156 * Hierarchy of CSS ID belongs to.
158 unsigned short stack[0]; /* Array of Length (depth+1) */
162 * cgroup_event represents events which userspace want to recieve.
164 struct cgroup_event {
166 * Cgroup which the event belongs to.
170 * Control file which the event associated.
174 * eventfd to signal userspace about the event.
176 struct eventfd_ctx *eventfd;
178 * Each of these stored in a list by the cgroup.
180 struct list_head list;
182 * All fields below needed to unregister event when
183 * userspace closes eventfd.
186 wait_queue_head_t *wqh;
188 struct work_struct remove;
191 /* The list of hierarchy roots */
193 static LIST_HEAD(roots);
194 static int root_count;
196 static DEFINE_IDA(hierarchy_ida);
197 static int next_hierarchy_id;
198 static DEFINE_SPINLOCK(hierarchy_id_lock);
200 /* dummytop is a shorthand for the dummy hierarchy's top cgroup */
201 #define dummytop (&rootnode.top_cgroup)
203 /* This flag indicates whether tasks in the fork and exit paths should
204 * check for fork/exit handlers to call. This avoids us having to do
205 * extra work in the fork/exit path if none of the subsystems need to
208 static int need_forkexit_callback __read_mostly;
210 #ifdef CONFIG_PROVE_LOCKING
211 int cgroup_lock_is_held(void)
213 return lockdep_is_held(&cgroup_mutex);
215 #else /* #ifdef CONFIG_PROVE_LOCKING */
216 int cgroup_lock_is_held(void)
218 return mutex_is_locked(&cgroup_mutex);
220 #endif /* #else #ifdef CONFIG_PROVE_LOCKING */
222 EXPORT_SYMBOL_GPL(cgroup_lock_is_held);
224 /* convenient tests for these bits */
225 inline int cgroup_is_removed(const struct cgroup *cgrp)
227 return test_bit(CGRP_REMOVED, &cgrp->flags);
230 /* bits in struct cgroupfs_root flags field */
232 ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
235 static int cgroup_is_releasable(const struct cgroup *cgrp)
238 (1 << CGRP_RELEASABLE) |
239 (1 << CGRP_NOTIFY_ON_RELEASE);
240 return (cgrp->flags & bits) == bits;
243 static int notify_on_release(const struct cgroup *cgrp)
245 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
249 * for_each_subsys() allows you to iterate on each subsystem attached to
250 * an active hierarchy
252 #define for_each_subsys(_root, _ss) \
253 list_for_each_entry(_ss, &_root->subsys_list, sibling)
255 /* for_each_active_root() allows you to iterate across the active hierarchies */
256 #define for_each_active_root(_root) \
257 list_for_each_entry(_root, &roots, root_list)
259 /* the list of cgroups eligible for automatic release. Protected by
260 * release_list_lock */
261 static LIST_HEAD(release_list);
262 static DEFINE_SPINLOCK(release_list_lock);
263 static void cgroup_release_agent(struct work_struct *work);
264 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
265 static void check_for_release(struct cgroup *cgrp);
267 /* Link structure for associating css_set objects with cgroups */
268 struct cg_cgroup_link {
270 * List running through cg_cgroup_links associated with a
271 * cgroup, anchored on cgroup->css_sets
273 struct list_head cgrp_link_list;
276 * List running through cg_cgroup_links pointing at a
277 * single css_set object, anchored on css_set->cg_links
279 struct list_head cg_link_list;
283 /* The default css_set - used by init and its children prior to any
284 * hierarchies being mounted. It contains a pointer to the root state
285 * for each subsystem. Also used to anchor the list of css_sets. Not
286 * reference-counted, to improve performance when child cgroups
287 * haven't been created.
290 static struct css_set init_css_set;
291 static struct cg_cgroup_link init_css_set_link;
293 static int cgroup_init_idr(struct cgroup_subsys *ss,
294 struct cgroup_subsys_state *css);
296 /* css_set_lock protects the list of css_set objects, and the
297 * chain of tasks off each css_set. Nests outside task->alloc_lock
298 * due to cgroup_iter_start() */
299 static DEFINE_RWLOCK(css_set_lock);
300 static int css_set_count;
303 * hash table for cgroup groups. This improves the performance to find
304 * an existing css_set. This hash doesn't (currently) take into
305 * account cgroups in empty hierarchies.
307 #define CSS_SET_HASH_BITS 7
308 #define CSS_SET_TABLE_SIZE (1 << CSS_SET_HASH_BITS)
309 static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
311 static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
315 unsigned long tmp = 0UL;
317 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
318 tmp += (unsigned long)css[i];
319 tmp = (tmp >> 16) ^ tmp;
321 index = hash_long(tmp, CSS_SET_HASH_BITS);
323 return &css_set_table[index];
326 static void free_css_set_rcu(struct rcu_head *obj)
328 struct css_set *cg = container_of(obj, struct css_set, rcu_head);
332 /* We don't maintain the lists running through each css_set to its
333 * task until after the first call to cgroup_iter_start(). This
334 * reduces the fork()/exit() overhead for people who have cgroups
335 * compiled into their kernel but not actually in use */
336 static int use_task_css_set_links __read_mostly;
338 static void __put_css_set(struct css_set *cg, int taskexit)
340 struct cg_cgroup_link *link;
341 struct cg_cgroup_link *saved_link;
343 * Ensure that the refcount doesn't hit zero while any readers
344 * can see it. Similar to atomic_dec_and_lock(), but for an
347 if (atomic_add_unless(&cg->refcount, -1, 1))
349 write_lock(&css_set_lock);
350 if (!atomic_dec_and_test(&cg->refcount)) {
351 write_unlock(&css_set_lock);
355 /* This css_set is dead. unlink it and release cgroup refcounts */
356 hlist_del(&cg->hlist);
359 list_for_each_entry_safe(link, saved_link, &cg->cg_links,
361 struct cgroup *cgrp = link->cgrp;
362 list_del(&link->cg_link_list);
363 list_del(&link->cgrp_link_list);
364 if (atomic_dec_and_test(&cgrp->count) &&
365 notify_on_release(cgrp)) {
367 set_bit(CGRP_RELEASABLE, &cgrp->flags);
368 check_for_release(cgrp);
374 write_unlock(&css_set_lock);
375 call_rcu(&cg->rcu_head, free_css_set_rcu);
379 * refcounted get/put for css_set objects
381 static inline void get_css_set(struct css_set *cg)
383 atomic_inc(&cg->refcount);
386 static inline void put_css_set(struct css_set *cg)
388 __put_css_set(cg, 0);
391 static inline void put_css_set_taskexit(struct css_set *cg)
393 __put_css_set(cg, 1);
397 * compare_css_sets - helper function for find_existing_css_set().
398 * @cg: candidate css_set being tested
399 * @old_cg: existing css_set for a task
400 * @new_cgrp: cgroup that's being entered by the task
401 * @template: desired set of css pointers in css_set (pre-calculated)
403 * Returns true if "cg" matches "old_cg" except for the hierarchy
404 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
406 static bool compare_css_sets(struct css_set *cg,
407 struct css_set *old_cg,
408 struct cgroup *new_cgrp,
409 struct cgroup_subsys_state *template[])
411 struct list_head *l1, *l2;
413 if (memcmp(template, cg->subsys, sizeof(cg->subsys))) {
414 /* Not all subsystems matched */
419 * Compare cgroup pointers in order to distinguish between
420 * different cgroups in heirarchies with no subsystems. We
421 * could get by with just this check alone (and skip the
422 * memcmp above) but on most setups the memcmp check will
423 * avoid the need for this more expensive check on almost all
428 l2 = &old_cg->cg_links;
430 struct cg_cgroup_link *cgl1, *cgl2;
431 struct cgroup *cg1, *cg2;
435 /* See if we reached the end - both lists are equal length. */
436 if (l1 == &cg->cg_links) {
437 BUG_ON(l2 != &old_cg->cg_links);
440 BUG_ON(l2 == &old_cg->cg_links);
442 /* Locate the cgroups associated with these links. */
443 cgl1 = list_entry(l1, struct cg_cgroup_link, cg_link_list);
444 cgl2 = list_entry(l2, struct cg_cgroup_link, cg_link_list);
447 /* Hierarchies should be linked in the same order. */
448 BUG_ON(cg1->root != cg2->root);
451 * If this hierarchy is the hierarchy of the cgroup
452 * that's changing, then we need to check that this
453 * css_set points to the new cgroup; if it's any other
454 * hierarchy, then this css_set should point to the
455 * same cgroup as the old css_set.
457 if (cg1->root == new_cgrp->root) {
469 * find_existing_css_set() is a helper for
470 * find_css_set(), and checks to see whether an existing
471 * css_set is suitable.
473 * oldcg: the cgroup group that we're using before the cgroup
476 * cgrp: the cgroup that we're moving into
478 * template: location in which to build the desired set of subsystem
479 * state objects for the new cgroup group
481 static struct css_set *find_existing_css_set(
482 struct css_set *oldcg,
484 struct cgroup_subsys_state *template[])
487 struct cgroupfs_root *root = cgrp->root;
488 struct hlist_head *hhead;
489 struct hlist_node *node;
493 * Build the set of subsystem state objects that we want to see in the
494 * new css_set. while subsystems can change globally, the entries here
495 * won't change, so no need for locking.
497 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
498 if (root->subsys_bits & (1UL << i)) {
499 /* Subsystem is in this hierarchy. So we want
500 * the subsystem state from the new
502 template[i] = cgrp->subsys[i];
504 /* Subsystem is not in this hierarchy, so we
505 * don't want to change the subsystem state */
506 template[i] = oldcg->subsys[i];
510 hhead = css_set_hash(template);
511 hlist_for_each_entry(cg, node, hhead, hlist) {
512 if (!compare_css_sets(cg, oldcg, cgrp, template))
515 /* This css_set matches what we need */
519 /* No existing cgroup group matched */
523 static void free_cg_links(struct list_head *tmp)
525 struct cg_cgroup_link *link;
526 struct cg_cgroup_link *saved_link;
528 list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
529 list_del(&link->cgrp_link_list);
535 * allocate_cg_links() allocates "count" cg_cgroup_link structures
536 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
537 * success or a negative error
539 static int allocate_cg_links(int count, struct list_head *tmp)
541 struct cg_cgroup_link *link;
544 for (i = 0; i < count; i++) {
545 link = kmalloc(sizeof(*link), GFP_KERNEL);
550 list_add(&link->cgrp_link_list, tmp);
556 * link_css_set - a helper function to link a css_set to a cgroup
557 * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
558 * @cg: the css_set to be linked
559 * @cgrp: the destination cgroup
561 static void link_css_set(struct list_head *tmp_cg_links,
562 struct css_set *cg, struct cgroup *cgrp)
564 struct cg_cgroup_link *link;
566 BUG_ON(list_empty(tmp_cg_links));
567 link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
571 atomic_inc(&cgrp->count);
572 list_move(&link->cgrp_link_list, &cgrp->css_sets);
574 * Always add links to the tail of the list so that the list
575 * is sorted by order of hierarchy creation
577 list_add_tail(&link->cg_link_list, &cg->cg_links);
581 * find_css_set() takes an existing cgroup group and a
582 * cgroup object, and returns a css_set object that's
583 * equivalent to the old group, but with the given cgroup
584 * substituted into the appropriate hierarchy. Must be called with
587 static struct css_set *find_css_set(
588 struct css_set *oldcg, struct cgroup *cgrp)
591 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
593 struct list_head tmp_cg_links;
595 struct hlist_head *hhead;
596 struct cg_cgroup_link *link;
598 /* First see if we already have a cgroup group that matches
600 read_lock(&css_set_lock);
601 res = find_existing_css_set(oldcg, cgrp, template);
604 read_unlock(&css_set_lock);
609 res = kmalloc(sizeof(*res), GFP_KERNEL);
613 /* Allocate all the cg_cgroup_link objects that we'll need */
614 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
619 atomic_set(&res->refcount, 1);
620 INIT_LIST_HEAD(&res->cg_links);
621 INIT_LIST_HEAD(&res->tasks);
622 INIT_HLIST_NODE(&res->hlist);
624 /* Copy the set of subsystem state objects generated in
625 * find_existing_css_set() */
626 memcpy(res->subsys, template, sizeof(res->subsys));
628 write_lock(&css_set_lock);
629 /* Add reference counts and links from the new css_set. */
630 list_for_each_entry(link, &oldcg->cg_links, cg_link_list) {
631 struct cgroup *c = link->cgrp;
632 if (c->root == cgrp->root)
634 link_css_set(&tmp_cg_links, res, c);
637 BUG_ON(!list_empty(&tmp_cg_links));
641 /* Add this cgroup group to the hash table */
642 hhead = css_set_hash(res->subsys);
643 hlist_add_head(&res->hlist, hhead);
645 write_unlock(&css_set_lock);
651 * Return the cgroup for "task" from the given hierarchy. Must be
652 * called with cgroup_mutex held.
654 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
655 struct cgroupfs_root *root)
658 struct cgroup *res = NULL;
660 BUG_ON(!mutex_is_locked(&cgroup_mutex));
661 read_lock(&css_set_lock);
663 * No need to lock the task - since we hold cgroup_mutex the
664 * task can't change groups, so the only thing that can happen
665 * is that it exits and its css is set back to init_css_set.
668 if (css == &init_css_set) {
669 res = &root->top_cgroup;
671 struct cg_cgroup_link *link;
672 list_for_each_entry(link, &css->cg_links, cg_link_list) {
673 struct cgroup *c = link->cgrp;
674 if (c->root == root) {
680 read_unlock(&css_set_lock);
686 * There is one global cgroup mutex. We also require taking
687 * task_lock() when dereferencing a task's cgroup subsys pointers.
688 * See "The task_lock() exception", at the end of this comment.
690 * A task must hold cgroup_mutex to modify cgroups.
692 * Any task can increment and decrement the count field without lock.
693 * So in general, code holding cgroup_mutex can't rely on the count
694 * field not changing. However, if the count goes to zero, then only
695 * cgroup_attach_task() can increment it again. Because a count of zero
696 * means that no tasks are currently attached, therefore there is no
697 * way a task attached to that cgroup can fork (the other way to
698 * increment the count). So code holding cgroup_mutex can safely
699 * assume that if the count is zero, it will stay zero. Similarly, if
700 * a task holds cgroup_mutex on a cgroup with zero count, it
701 * knows that the cgroup won't be removed, as cgroup_rmdir()
704 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
705 * (usually) take cgroup_mutex. These are the two most performance
706 * critical pieces of code here. The exception occurs on cgroup_exit(),
707 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
708 * is taken, and if the cgroup count is zero, a usermode call made
709 * to the release agent with the name of the cgroup (path relative to
710 * the root of cgroup file system) as the argument.
712 * A cgroup can only be deleted if both its 'count' of using tasks
713 * is zero, and its list of 'children' cgroups is empty. Since all
714 * tasks in the system use _some_ cgroup, and since there is always at
715 * least one task in the system (init, pid == 1), therefore, top_cgroup
716 * always has either children cgroups and/or using tasks. So we don't
717 * need a special hack to ensure that top_cgroup cannot be deleted.
719 * The task_lock() exception
721 * The need for this exception arises from the action of
722 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
723 * another. It does so using cgroup_mutex, however there are
724 * several performance critical places that need to reference
725 * task->cgroup without the expense of grabbing a system global
726 * mutex. Therefore except as noted below, when dereferencing or, as
727 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
728 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
729 * the task_struct routinely used for such matters.
731 * P.S. One more locking exception. RCU is used to guard the
732 * update of a tasks cgroup pointer by cgroup_attach_task()
736 * cgroup_lock - lock out any changes to cgroup structures
739 void cgroup_lock(void)
741 mutex_lock(&cgroup_mutex);
743 EXPORT_SYMBOL_GPL(cgroup_lock);
746 * cgroup_unlock - release lock on cgroup changes
748 * Undo the lock taken in a previous cgroup_lock() call.
750 void cgroup_unlock(void)
752 mutex_unlock(&cgroup_mutex);
754 EXPORT_SYMBOL_GPL(cgroup_unlock);
757 * A couple of forward declarations required, due to cyclic reference loop:
758 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
759 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
763 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
764 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
765 static int cgroup_populate_dir(struct cgroup *cgrp);
766 static const struct inode_operations cgroup_dir_inode_operations;
767 static const struct file_operations proc_cgroupstats_operations;
769 static struct backing_dev_info cgroup_backing_dev_info = {
771 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
774 static int alloc_css_id(struct cgroup_subsys *ss,
775 struct cgroup *parent, struct cgroup *child);
777 static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
779 struct inode *inode = new_inode(sb);
782 inode->i_mode = mode;
783 inode->i_uid = current_fsuid();
784 inode->i_gid = current_fsgid();
785 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
786 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
792 * Call subsys's pre_destroy handler.
793 * This is called before css refcnt check.
795 static int cgroup_call_pre_destroy(struct cgroup *cgrp)
797 struct cgroup_subsys *ss;
800 for_each_subsys(cgrp->root, ss)
801 if (ss->pre_destroy) {
802 ret = ss->pre_destroy(ss, cgrp);
810 static void free_cgroup_rcu(struct rcu_head *obj)
812 struct cgroup *cgrp = container_of(obj, struct cgroup, rcu_head);
817 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
819 /* is dentry a directory ? if so, kfree() associated cgroup */
820 if (S_ISDIR(inode->i_mode)) {
821 struct cgroup *cgrp = dentry->d_fsdata;
822 struct cgroup_subsys *ss;
823 BUG_ON(!(cgroup_is_removed(cgrp)));
824 /* It's possible for external users to be holding css
825 * reference counts on a cgroup; css_put() needs to
826 * be able to access the cgroup after decrementing
827 * the reference count in order to know if it needs to
828 * queue the cgroup to be handled by the release
832 mutex_lock(&cgroup_mutex);
834 * Release the subsystem state objects.
836 for_each_subsys(cgrp->root, ss)
837 ss->destroy(ss, cgrp);
839 cgrp->root->number_of_cgroups--;
840 mutex_unlock(&cgroup_mutex);
843 * Drop the active superblock reference that we took when we
846 deactivate_super(cgrp->root->sb);
849 * if we're getting rid of the cgroup, refcount should ensure
850 * that there are no pidlists left.
852 BUG_ON(!list_empty(&cgrp->pidlists));
854 call_rcu(&cgrp->rcu_head, free_cgroup_rcu);
859 static void remove_dir(struct dentry *d)
861 struct dentry *parent = dget(d->d_parent);
864 simple_rmdir(parent->d_inode, d);
868 static void cgroup_clear_directory(struct dentry *dentry)
870 struct list_head *node;
872 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
873 spin_lock(&dcache_lock);
874 node = dentry->d_subdirs.next;
875 while (node != &dentry->d_subdirs) {
876 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
879 /* This should never be called on a cgroup
880 * directory with child cgroups */
881 BUG_ON(d->d_inode->i_mode & S_IFDIR);
883 spin_unlock(&dcache_lock);
885 simple_unlink(dentry->d_inode, d);
887 spin_lock(&dcache_lock);
889 node = dentry->d_subdirs.next;
891 spin_unlock(&dcache_lock);
895 * NOTE : the dentry must have been dget()'ed
897 static void cgroup_d_remove_dir(struct dentry *dentry)
899 cgroup_clear_directory(dentry);
901 spin_lock(&dcache_lock);
902 list_del_init(&dentry->d_u.d_child);
903 spin_unlock(&dcache_lock);
908 * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
909 * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
910 * reference to css->refcnt. In general, this refcnt is expected to goes down
913 * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
915 DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
917 static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
919 if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
920 wake_up_all(&cgroup_rmdir_waitq);
923 void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
928 void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
930 cgroup_wakeup_rmdir_waiter(css->cgroup);
935 * Call with cgroup_mutex held. Drops reference counts on modules, including
936 * any duplicate ones that parse_cgroupfs_options took. If this function
937 * returns an error, no reference counts are touched.
939 static int rebind_subsystems(struct cgroupfs_root *root,
940 unsigned long final_bits)
942 unsigned long added_bits, removed_bits;
943 struct cgroup *cgrp = &root->top_cgroup;
946 BUG_ON(!mutex_is_locked(&cgroup_mutex));
948 removed_bits = root->actual_subsys_bits & ~final_bits;
949 added_bits = final_bits & ~root->actual_subsys_bits;
950 /* Check that any added subsystems are currently free */
951 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
952 unsigned long bit = 1UL << i;
953 struct cgroup_subsys *ss = subsys[i];
954 if (!(bit & added_bits))
957 * Nobody should tell us to do a subsys that doesn't exist:
958 * parse_cgroupfs_options should catch that case and refcounts
959 * ensure that subsystems won't disappear once selected.
962 if (ss->root != &rootnode) {
963 /* Subsystem isn't free */
968 /* Currently we don't handle adding/removing subsystems when
969 * any child cgroups exist. This is theoretically supportable
970 * but involves complex error handling, so it's being left until
972 if (root->number_of_cgroups > 1)
975 /* Process each subsystem */
976 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
977 struct cgroup_subsys *ss = subsys[i];
978 unsigned long bit = 1UL << i;
979 if (bit & added_bits) {
980 /* We're binding this subsystem to this hierarchy */
982 BUG_ON(cgrp->subsys[i]);
983 BUG_ON(!dummytop->subsys[i]);
984 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
985 mutex_lock(&ss->hierarchy_mutex);
986 cgrp->subsys[i] = dummytop->subsys[i];
987 cgrp->subsys[i]->cgroup = cgrp;
988 list_move(&ss->sibling, &root->subsys_list);
992 mutex_unlock(&ss->hierarchy_mutex);
993 /* refcount was already taken, and we're keeping it */
994 } else if (bit & removed_bits) {
995 /* We're removing this subsystem */
997 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
998 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
999 mutex_lock(&ss->hierarchy_mutex);
1001 ss->bind(ss, dummytop);
1002 dummytop->subsys[i]->cgroup = dummytop;
1003 cgrp->subsys[i] = NULL;
1004 subsys[i]->root = &rootnode;
1005 list_move(&ss->sibling, &rootnode.subsys_list);
1006 mutex_unlock(&ss->hierarchy_mutex);
1007 /* subsystem is now free - drop reference on module */
1008 module_put(ss->module);
1009 } else if (bit & final_bits) {
1010 /* Subsystem state should already exist */
1012 BUG_ON(!cgrp->subsys[i]);
1014 * a refcount was taken, but we already had one, so
1015 * drop the extra reference.
1017 module_put(ss->module);
1018 #ifdef CONFIG_MODULE_UNLOAD
1019 BUG_ON(ss->module && !module_refcount(ss->module));
1022 /* Subsystem state shouldn't exist */
1023 BUG_ON(cgrp->subsys[i]);
1026 root->subsys_bits = root->actual_subsys_bits = final_bits;
1032 static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
1034 struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
1035 struct cgroup_subsys *ss;
1037 mutex_lock(&cgroup_mutex);
1038 for_each_subsys(root, ss)
1039 seq_printf(seq, ",%s", ss->name);
1040 if (test_bit(ROOT_NOPREFIX, &root->flags))
1041 seq_puts(seq, ",noprefix");
1042 if (strlen(root->release_agent_path))
1043 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1044 if (strlen(root->name))
1045 seq_printf(seq, ",name=%s", root->name);
1046 mutex_unlock(&cgroup_mutex);
1050 struct cgroup_sb_opts {
1051 unsigned long subsys_bits;
1052 unsigned long flags;
1053 char *release_agent;
1055 /* User explicitly requested empty subsystem */
1058 struct cgroupfs_root *new_root;
1063 * Convert a hierarchy specifier into a bitmask of subsystems and flags. Call
1064 * with cgroup_mutex held to protect the subsys[] array. This function takes
1065 * refcounts on subsystems to be used, unless it returns error, in which case
1066 * no refcounts are taken.
1068 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1070 char *token, *o = data ?: "all";
1071 unsigned long mask = (unsigned long)-1;
1073 bool module_pin_failed = false;
1075 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1077 #ifdef CONFIG_CPUSETS
1078 mask = ~(1UL << cpuset_subsys_id);
1081 memset(opts, 0, sizeof(*opts));
1083 while ((token = strsep(&o, ",")) != NULL) {
1086 if (!strcmp(token, "all")) {
1087 /* Add all non-disabled subsystems */
1088 opts->subsys_bits = 0;
1089 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1090 struct cgroup_subsys *ss = subsys[i];
1094 opts->subsys_bits |= 1ul << i;
1096 } else if (!strcmp(token, "none")) {
1097 /* Explicitly have no subsystems */
1099 } else if (!strcmp(token, "noprefix")) {
1100 set_bit(ROOT_NOPREFIX, &opts->flags);
1101 } else if (!strncmp(token, "release_agent=", 14)) {
1102 /* Specifying two release agents is forbidden */
1103 if (opts->release_agent)
1105 opts->release_agent =
1106 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1107 if (!opts->release_agent)
1109 } else if (!strncmp(token, "name=", 5)) {
1110 const char *name = token + 5;
1111 /* Can't specify an empty name */
1114 /* Must match [\w.-]+ */
1115 for (i = 0; i < strlen(name); i++) {
1119 if ((c == '.') || (c == '-') || (c == '_'))
1123 /* Specifying two names is forbidden */
1126 opts->name = kstrndup(name,
1127 MAX_CGROUP_ROOT_NAMELEN - 1,
1132 struct cgroup_subsys *ss;
1133 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1137 if (!strcmp(token, ss->name)) {
1139 set_bit(i, &opts->subsys_bits);
1143 if (i == CGROUP_SUBSYS_COUNT)
1148 /* Consistency checks */
1151 * Option noprefix was introduced just for backward compatibility
1152 * with the old cpuset, so we allow noprefix only if mounting just
1153 * the cpuset subsystem.
1155 if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
1156 (opts->subsys_bits & mask))
1160 /* Can't specify "none" and some subsystems */
1161 if (opts->subsys_bits && opts->none)
1165 * We either have to specify by name or by subsystems. (So all
1166 * empty hierarchies must have a name).
1168 if (!opts->subsys_bits && !opts->name)
1172 * Grab references on all the modules we'll need, so the subsystems
1173 * don't dance around before rebind_subsystems attaches them. This may
1174 * take duplicate reference counts on a subsystem that's already used,
1175 * but rebind_subsystems handles this case.
1177 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1178 unsigned long bit = 1UL << i;
1180 if (!(bit & opts->subsys_bits))
1182 if (!try_module_get(subsys[i]->module)) {
1183 module_pin_failed = true;
1187 if (module_pin_failed) {
1189 * oops, one of the modules was going away. this means that we
1190 * raced with a module_delete call, and to the user this is
1191 * essentially a "subsystem doesn't exist" case.
1193 for (i--; i >= CGROUP_BUILTIN_SUBSYS_COUNT; i--) {
1194 /* drop refcounts only on the ones we took */
1195 unsigned long bit = 1UL << i;
1197 if (!(bit & opts->subsys_bits))
1199 module_put(subsys[i]->module);
1207 static void drop_parsed_module_refcounts(unsigned long subsys_bits)
1210 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1211 unsigned long bit = 1UL << i;
1213 if (!(bit & subsys_bits))
1215 module_put(subsys[i]->module);
1219 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1222 struct cgroupfs_root *root = sb->s_fs_info;
1223 struct cgroup *cgrp = &root->top_cgroup;
1224 struct cgroup_sb_opts opts;
1227 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1228 mutex_lock(&cgroup_mutex);
1230 /* See what subsystems are wanted */
1231 ret = parse_cgroupfs_options(data, &opts);
1235 /* Don't allow flags or name to change at remount */
1236 if (opts.flags != root->flags ||
1237 (opts.name && strcmp(opts.name, root->name))) {
1239 drop_parsed_module_refcounts(opts.subsys_bits);
1243 ret = rebind_subsystems(root, opts.subsys_bits);
1245 drop_parsed_module_refcounts(opts.subsys_bits);
1249 /* (re)populate subsystem files */
1250 cgroup_populate_dir(cgrp);
1252 if (opts.release_agent)
1253 strcpy(root->release_agent_path, opts.release_agent);
1255 kfree(opts.release_agent);
1257 mutex_unlock(&cgroup_mutex);
1258 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1263 static const struct super_operations cgroup_ops = {
1264 .statfs = simple_statfs,
1265 .drop_inode = generic_delete_inode,
1266 .show_options = cgroup_show_options,
1267 .remount_fs = cgroup_remount,
1270 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1272 INIT_LIST_HEAD(&cgrp->sibling);
1273 INIT_LIST_HEAD(&cgrp->children);
1274 INIT_LIST_HEAD(&cgrp->css_sets);
1275 INIT_LIST_HEAD(&cgrp->release_list);
1276 INIT_LIST_HEAD(&cgrp->pidlists);
1277 mutex_init(&cgrp->pidlist_mutex);
1278 INIT_LIST_HEAD(&cgrp->event_list);
1279 spin_lock_init(&cgrp->event_list_lock);
1282 static void init_cgroup_root(struct cgroupfs_root *root)
1284 struct cgroup *cgrp = &root->top_cgroup;
1285 INIT_LIST_HEAD(&root->subsys_list);
1286 INIT_LIST_HEAD(&root->root_list);
1287 root->number_of_cgroups = 1;
1289 cgrp->top_cgroup = cgrp;
1290 init_cgroup_housekeeping(cgrp);
1293 static bool init_root_id(struct cgroupfs_root *root)
1298 if (!ida_pre_get(&hierarchy_ida, GFP_KERNEL))
1300 spin_lock(&hierarchy_id_lock);
1301 /* Try to allocate the next unused ID */
1302 ret = ida_get_new_above(&hierarchy_ida, next_hierarchy_id,
1303 &root->hierarchy_id);
1305 /* Try again starting from 0 */
1306 ret = ida_get_new(&hierarchy_ida, &root->hierarchy_id);
1308 next_hierarchy_id = root->hierarchy_id + 1;
1309 } else if (ret != -EAGAIN) {
1310 /* Can only get here if the 31-bit IDR is full ... */
1313 spin_unlock(&hierarchy_id_lock);
1318 static int cgroup_test_super(struct super_block *sb, void *data)
1320 struct cgroup_sb_opts *opts = data;
1321 struct cgroupfs_root *root = sb->s_fs_info;
1323 /* If we asked for a name then it must match */
1324 if (opts->name && strcmp(opts->name, root->name))
1328 * If we asked for subsystems (or explicitly for no
1329 * subsystems) then they must match
1331 if ((opts->subsys_bits || opts->none)
1332 && (opts->subsys_bits != root->subsys_bits))
1338 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1340 struct cgroupfs_root *root;
1342 if (!opts->subsys_bits && !opts->none)
1345 root = kzalloc(sizeof(*root), GFP_KERNEL);
1347 return ERR_PTR(-ENOMEM);
1349 if (!init_root_id(root)) {
1351 return ERR_PTR(-ENOMEM);
1353 init_cgroup_root(root);
1355 root->subsys_bits = opts->subsys_bits;
1356 root->flags = opts->flags;
1357 if (opts->release_agent)
1358 strcpy(root->release_agent_path, opts->release_agent);
1360 strcpy(root->name, opts->name);
1364 static void cgroup_drop_root(struct cgroupfs_root *root)
1369 BUG_ON(!root->hierarchy_id);
1370 spin_lock(&hierarchy_id_lock);
1371 ida_remove(&hierarchy_ida, root->hierarchy_id);
1372 spin_unlock(&hierarchy_id_lock);
1376 static int cgroup_set_super(struct super_block *sb, void *data)
1379 struct cgroup_sb_opts *opts = data;
1381 /* If we don't have a new root, we can't set up a new sb */
1382 if (!opts->new_root)
1385 BUG_ON(!opts->subsys_bits && !opts->none);
1387 ret = set_anon_super(sb, NULL);
1391 sb->s_fs_info = opts->new_root;
1392 opts->new_root->sb = sb;
1394 sb->s_blocksize = PAGE_CACHE_SIZE;
1395 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1396 sb->s_magic = CGROUP_SUPER_MAGIC;
1397 sb->s_op = &cgroup_ops;
1402 static int cgroup_get_rootdir(struct super_block *sb)
1404 struct inode *inode =
1405 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1406 struct dentry *dentry;
1411 inode->i_fop = &simple_dir_operations;
1412 inode->i_op = &cgroup_dir_inode_operations;
1413 /* directories start off with i_nlink == 2 (for "." entry) */
1415 dentry = d_alloc_root(inode);
1420 sb->s_root = dentry;
1424 static int cgroup_get_sb(struct file_system_type *fs_type,
1425 int flags, const char *unused_dev_name,
1426 void *data, struct vfsmount *mnt)
1428 struct cgroup_sb_opts opts;
1429 struct cgroupfs_root *root;
1431 struct super_block *sb;
1432 struct cgroupfs_root *new_root;
1434 /* First find the desired set of subsystems */
1435 mutex_lock(&cgroup_mutex);
1436 ret = parse_cgroupfs_options(data, &opts);
1437 mutex_unlock(&cgroup_mutex);
1442 * Allocate a new cgroup root. We may not need it if we're
1443 * reusing an existing hierarchy.
1445 new_root = cgroup_root_from_opts(&opts);
1446 if (IS_ERR(new_root)) {
1447 ret = PTR_ERR(new_root);
1450 opts.new_root = new_root;
1452 /* Locate an existing or new sb for this hierarchy */
1453 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
1456 cgroup_drop_root(opts.new_root);
1460 root = sb->s_fs_info;
1462 if (root == opts.new_root) {
1463 /* We used the new root structure, so this is a new hierarchy */
1464 struct list_head tmp_cg_links;
1465 struct cgroup *root_cgrp = &root->top_cgroup;
1466 struct inode *inode;
1467 struct cgroupfs_root *existing_root;
1470 BUG_ON(sb->s_root != NULL);
1472 ret = cgroup_get_rootdir(sb);
1474 goto drop_new_super;
1475 inode = sb->s_root->d_inode;
1477 mutex_lock(&inode->i_mutex);
1478 mutex_lock(&cgroup_mutex);
1480 if (strlen(root->name)) {
1481 /* Check for name clashes with existing mounts */
1482 for_each_active_root(existing_root) {
1483 if (!strcmp(existing_root->name, root->name)) {
1485 mutex_unlock(&cgroup_mutex);
1486 mutex_unlock(&inode->i_mutex);
1487 goto drop_new_super;
1493 * We're accessing css_set_count without locking
1494 * css_set_lock here, but that's OK - it can only be
1495 * increased by someone holding cgroup_lock, and
1496 * that's us. The worst that can happen is that we
1497 * have some link structures left over
1499 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1501 mutex_unlock(&cgroup_mutex);
1502 mutex_unlock(&inode->i_mutex);
1503 goto drop_new_super;
1506 ret = rebind_subsystems(root, root->subsys_bits);
1507 if (ret == -EBUSY) {
1508 mutex_unlock(&cgroup_mutex);
1509 mutex_unlock(&inode->i_mutex);
1510 free_cg_links(&tmp_cg_links);
1511 goto drop_new_super;
1514 * There must be no failure case after here, since rebinding
1515 * takes care of subsystems' refcounts, which are explicitly
1516 * dropped in the failure exit path.
1519 /* EBUSY should be the only error here */
1522 list_add(&root->root_list, &roots);
1525 sb->s_root->d_fsdata = root_cgrp;
1526 root->top_cgroup.dentry = sb->s_root;
1528 /* Link the top cgroup in this hierarchy into all
1529 * the css_set objects */
1530 write_lock(&css_set_lock);
1531 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1532 struct hlist_head *hhead = &css_set_table[i];
1533 struct hlist_node *node;
1536 hlist_for_each_entry(cg, node, hhead, hlist)
1537 link_css_set(&tmp_cg_links, cg, root_cgrp);
1539 write_unlock(&css_set_lock);
1541 free_cg_links(&tmp_cg_links);
1543 BUG_ON(!list_empty(&root_cgrp->sibling));
1544 BUG_ON(!list_empty(&root_cgrp->children));
1545 BUG_ON(root->number_of_cgroups != 1);
1547 cgroup_populate_dir(root_cgrp);
1548 mutex_unlock(&cgroup_mutex);
1549 mutex_unlock(&inode->i_mutex);
1552 * We re-used an existing hierarchy - the new root (if
1553 * any) is not needed
1555 cgroup_drop_root(opts.new_root);
1556 /* no subsys rebinding, so refcounts don't change */
1557 drop_parsed_module_refcounts(opts.subsys_bits);
1560 simple_set_mnt(mnt, sb);
1561 kfree(opts.release_agent);
1566 deactivate_locked_super(sb);
1568 drop_parsed_module_refcounts(opts.subsys_bits);
1570 kfree(opts.release_agent);
1576 static void cgroup_kill_sb(struct super_block *sb) {
1577 struct cgroupfs_root *root = sb->s_fs_info;
1578 struct cgroup *cgrp = &root->top_cgroup;
1580 struct cg_cgroup_link *link;
1581 struct cg_cgroup_link *saved_link;
1585 BUG_ON(root->number_of_cgroups != 1);
1586 BUG_ON(!list_empty(&cgrp->children));
1587 BUG_ON(!list_empty(&cgrp->sibling));
1589 mutex_lock(&cgroup_mutex);
1591 /* Rebind all subsystems back to the default hierarchy */
1592 ret = rebind_subsystems(root, 0);
1593 /* Shouldn't be able to fail ... */
1597 * Release all the links from css_sets to this hierarchy's
1600 write_lock(&css_set_lock);
1602 list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1604 list_del(&link->cg_link_list);
1605 list_del(&link->cgrp_link_list);
1608 write_unlock(&css_set_lock);
1610 if (!list_empty(&root->root_list)) {
1611 list_del(&root->root_list);
1615 mutex_unlock(&cgroup_mutex);
1617 kill_litter_super(sb);
1618 cgroup_drop_root(root);
1621 static struct file_system_type cgroup_fs_type = {
1623 .get_sb = cgroup_get_sb,
1624 .kill_sb = cgroup_kill_sb,
1627 static struct kobject *cgroup_kobj;
1629 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1631 return dentry->d_fsdata;
1634 static inline struct cftype *__d_cft(struct dentry *dentry)
1636 return dentry->d_fsdata;
1640 * cgroup_path - generate the path of a cgroup
1641 * @cgrp: the cgroup in question
1642 * @buf: the buffer to write the path into
1643 * @buflen: the length of the buffer
1645 * Called with cgroup_mutex held or else with an RCU-protected cgroup
1646 * reference. Writes path of cgroup into buf. Returns 0 on success,
1649 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1652 struct dentry *dentry = rcu_dereference_check(cgrp->dentry,
1653 rcu_read_lock_held() ||
1654 cgroup_lock_is_held());
1656 if (!dentry || cgrp == dummytop) {
1658 * Inactive subsystems have no dentry for their root
1665 start = buf + buflen;
1669 int len = dentry->d_name.len;
1671 if ((start -= len) < buf)
1672 return -ENAMETOOLONG;
1673 memcpy(start, dentry->d_name.name, len);
1674 cgrp = cgrp->parent;
1678 dentry = rcu_dereference_check(cgrp->dentry,
1679 rcu_read_lock_held() ||
1680 cgroup_lock_is_held());
1684 return -ENAMETOOLONG;
1687 memmove(buf, start, buf + buflen - start);
1690 EXPORT_SYMBOL_GPL(cgroup_path);
1693 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1694 * @cgrp: the cgroup the task is attaching to
1695 * @tsk: the task to be attached
1697 * Call holding cgroup_mutex. May take task_lock of
1698 * the task 'tsk' during call.
1700 int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1703 struct cgroup_subsys *ss, *failed_ss = NULL;
1704 struct cgroup *oldcgrp;
1706 struct css_set *newcg;
1707 struct cgroupfs_root *root = cgrp->root;
1709 /* Nothing to do if the task is already in that cgroup */
1710 oldcgrp = task_cgroup_from_root(tsk, root);
1711 if (cgrp == oldcgrp)
1714 for_each_subsys(root, ss) {
1715 if (ss->can_attach) {
1716 retval = ss->can_attach(ss, cgrp, tsk, false);
1719 * Remember on which subsystem the can_attach()
1720 * failed, so that we only call cancel_attach()
1721 * against the subsystems whose can_attach()
1722 * succeeded. (See below)
1727 } else if (!capable(CAP_SYS_ADMIN)) {
1728 const struct cred *cred = current_cred(), *tcred;
1730 /* No can_attach() - check perms generically */
1731 tcred = __task_cred(tsk);
1732 if (cred->euid != tcred->uid &&
1733 cred->euid != tcred->suid) {
1744 * Locate or allocate a new css_set for this task,
1745 * based on its final set of cgroups
1747 newcg = find_css_set(cg, cgrp);
1755 if (tsk->flags & PF_EXITING) {
1761 rcu_assign_pointer(tsk->cgroups, newcg);
1764 /* Update the css_set linked lists if we're using them */
1765 write_lock(&css_set_lock);
1766 if (!list_empty(&tsk->cg_list)) {
1767 list_del(&tsk->cg_list);
1768 list_add(&tsk->cg_list, &newcg->tasks);
1770 write_unlock(&css_set_lock);
1772 for_each_subsys(root, ss) {
1774 ss->attach(ss, cgrp, oldcgrp, tsk, false);
1776 set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1781 * wake up rmdir() waiter. the rmdir should fail since the cgroup
1782 * is no longer empty.
1784 cgroup_wakeup_rmdir_waiter(cgrp);
1787 for_each_subsys(root, ss) {
1788 if (ss == failed_ss)
1790 * This subsystem was the one that failed the
1791 * can_attach() check earlier, so we don't need
1792 * to call cancel_attach() against it or any
1793 * remaining subsystems.
1796 if (ss->cancel_attach)
1797 ss->cancel_attach(ss, cgrp, tsk, false);
1804 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1805 * @from: attach to all cgroups of a given task
1806 * @tsk: the task to be attached
1808 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1810 struct cgroupfs_root *root;
1814 for_each_active_root(root) {
1815 struct cgroup *from_cg = task_cgroup_from_root(from, root);
1817 retval = cgroup_attach_task(from_cg, tsk);
1825 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
1828 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
1829 * held. May take task_lock of task
1831 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid)
1833 struct task_struct *tsk;
1838 tsk = find_task_by_vpid(pid);
1839 if (!tsk || tsk->flags & PF_EXITING) {
1843 get_task_struct(tsk);
1847 get_task_struct(tsk);
1850 ret = cgroup_attach_task(cgrp, tsk);
1851 put_task_struct(tsk);
1855 static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
1858 if (!cgroup_lock_live_group(cgrp))
1860 ret = attach_task_by_pid(cgrp, pid);
1866 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
1867 * @cgrp: the cgroup to be checked for liveness
1869 * On success, returns true; the lock should be later released with
1870 * cgroup_unlock(). On failure returns false with no lock held.
1872 bool cgroup_lock_live_group(struct cgroup *cgrp)
1874 mutex_lock(&cgroup_mutex);
1875 if (cgroup_is_removed(cgrp)) {
1876 mutex_unlock(&cgroup_mutex);
1881 EXPORT_SYMBOL_GPL(cgroup_lock_live_group);
1883 static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
1886 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1887 if (!cgroup_lock_live_group(cgrp))
1889 strcpy(cgrp->root->release_agent_path, buffer);
1894 static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
1895 struct seq_file *seq)
1897 if (!cgroup_lock_live_group(cgrp))
1899 seq_puts(seq, cgrp->root->release_agent_path);
1900 seq_putc(seq, '\n');
1905 /* A buffer size big enough for numbers or short strings */
1906 #define CGROUP_LOCAL_BUFFER_SIZE 64
1908 static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
1910 const char __user *userbuf,
1911 size_t nbytes, loff_t *unused_ppos)
1913 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
1919 if (nbytes >= sizeof(buffer))
1921 if (copy_from_user(buffer, userbuf, nbytes))
1924 buffer[nbytes] = 0; /* nul-terminate */
1925 if (cft->write_u64) {
1926 u64 val = simple_strtoull(strstrip(buffer), &end, 0);
1929 retval = cft->write_u64(cgrp, cft, val);
1931 s64 val = simple_strtoll(strstrip(buffer), &end, 0);
1934 retval = cft->write_s64(cgrp, cft, val);
1941 static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
1943 const char __user *userbuf,
1944 size_t nbytes, loff_t *unused_ppos)
1946 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
1948 size_t max_bytes = cft->max_write_len;
1949 char *buffer = local_buffer;
1952 max_bytes = sizeof(local_buffer) - 1;
1953 if (nbytes >= max_bytes)
1955 /* Allocate a dynamic buffer if we need one */
1956 if (nbytes >= sizeof(local_buffer)) {
1957 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
1961 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
1966 buffer[nbytes] = 0; /* nul-terminate */
1967 retval = cft->write_string(cgrp, cft, strstrip(buffer));
1971 if (buffer != local_buffer)
1976 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
1977 size_t nbytes, loff_t *ppos)
1979 struct cftype *cft = __d_cft(file->f_dentry);
1980 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1982 if (cgroup_is_removed(cgrp))
1985 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
1986 if (cft->write_u64 || cft->write_s64)
1987 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
1988 if (cft->write_string)
1989 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
1991 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
1992 return ret ? ret : nbytes;
1997 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
1999 char __user *buf, size_t nbytes,
2002 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2003 u64 val = cft->read_u64(cgrp, cft);
2004 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
2006 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2009 static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
2011 char __user *buf, size_t nbytes,
2014 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2015 s64 val = cft->read_s64(cgrp, cft);
2016 int len = sprintf(tmp, "%lld\n", (long long) val);
2018 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2021 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2022 size_t nbytes, loff_t *ppos)
2024 struct cftype *cft = __d_cft(file->f_dentry);
2025 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2027 if (cgroup_is_removed(cgrp))
2031 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2033 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2035 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2040 * seqfile ops/methods for returning structured data. Currently just
2041 * supports string->u64 maps, but can be extended in future.
2044 struct cgroup_seqfile_state {
2046 struct cgroup *cgroup;
2049 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2051 struct seq_file *sf = cb->state;
2052 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2055 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2057 struct cgroup_seqfile_state *state = m->private;
2058 struct cftype *cft = state->cft;
2059 if (cft->read_map) {
2060 struct cgroup_map_cb cb = {
2061 .fill = cgroup_map_add,
2064 return cft->read_map(state->cgroup, cft, &cb);
2066 return cft->read_seq_string(state->cgroup, cft, m);
2069 static int cgroup_seqfile_release(struct inode *inode, struct file *file)
2071 struct seq_file *seq = file->private_data;
2072 kfree(seq->private);
2073 return single_release(inode, file);
2076 static const struct file_operations cgroup_seqfile_operations = {
2078 .write = cgroup_file_write,
2079 .llseek = seq_lseek,
2080 .release = cgroup_seqfile_release,
2083 static int cgroup_file_open(struct inode *inode, struct file *file)
2088 err = generic_file_open(inode, file);
2091 cft = __d_cft(file->f_dentry);
2093 if (cft->read_map || cft->read_seq_string) {
2094 struct cgroup_seqfile_state *state =
2095 kzalloc(sizeof(*state), GFP_USER);
2099 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
2100 file->f_op = &cgroup_seqfile_operations;
2101 err = single_open(file, cgroup_seqfile_show, state);
2104 } else if (cft->open)
2105 err = cft->open(inode, file);
2112 static int cgroup_file_release(struct inode *inode, struct file *file)
2114 struct cftype *cft = __d_cft(file->f_dentry);
2116 return cft->release(inode, file);
2121 * cgroup_rename - Only allow simple rename of directories in place.
2123 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2124 struct inode *new_dir, struct dentry *new_dentry)
2126 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2128 if (new_dentry->d_inode)
2130 if (old_dir != new_dir)
2132 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2135 static const struct file_operations cgroup_file_operations = {
2136 .read = cgroup_file_read,
2137 .write = cgroup_file_write,
2138 .llseek = generic_file_llseek,
2139 .open = cgroup_file_open,
2140 .release = cgroup_file_release,
2143 static const struct inode_operations cgroup_dir_inode_operations = {
2144 .lookup = simple_lookup,
2145 .mkdir = cgroup_mkdir,
2146 .rmdir = cgroup_rmdir,
2147 .rename = cgroup_rename,
2151 * Check if a file is a control file
2153 static inline struct cftype *__file_cft(struct file *file)
2155 if (file->f_dentry->d_inode->i_fop != &cgroup_file_operations)
2156 return ERR_PTR(-EINVAL);
2157 return __d_cft(file->f_dentry);
2160 static int cgroup_create_file(struct dentry *dentry, mode_t mode,
2161 struct super_block *sb)
2163 static const struct dentry_operations cgroup_dops = {
2164 .d_iput = cgroup_diput,
2167 struct inode *inode;
2171 if (dentry->d_inode)
2174 inode = cgroup_new_inode(mode, sb);
2178 if (S_ISDIR(mode)) {
2179 inode->i_op = &cgroup_dir_inode_operations;
2180 inode->i_fop = &simple_dir_operations;
2182 /* start off with i_nlink == 2 (for "." entry) */
2185 /* start with the directory inode held, so that we can
2186 * populate it without racing with another mkdir */
2187 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2188 } else if (S_ISREG(mode)) {
2190 inode->i_fop = &cgroup_file_operations;
2192 dentry->d_op = &cgroup_dops;
2193 d_instantiate(dentry, inode);
2194 dget(dentry); /* Extra count - pin the dentry in core */
2199 * cgroup_create_dir - create a directory for an object.
2200 * @cgrp: the cgroup we create the directory for. It must have a valid
2201 * ->parent field. And we are going to fill its ->dentry field.
2202 * @dentry: dentry of the new cgroup
2203 * @mode: mode to set on new directory.
2205 static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
2208 struct dentry *parent;
2211 parent = cgrp->parent->dentry;
2212 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
2214 dentry->d_fsdata = cgrp;
2215 inc_nlink(parent->d_inode);
2216 rcu_assign_pointer(cgrp->dentry, dentry);
2225 * cgroup_file_mode - deduce file mode of a control file
2226 * @cft: the control file in question
2228 * returns cft->mode if ->mode is not 0
2229 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2230 * returns S_IRUGO if it has only a read handler
2231 * returns S_IWUSR if it has only a write hander
2233 static mode_t cgroup_file_mode(const struct cftype *cft)
2240 if (cft->read || cft->read_u64 || cft->read_s64 ||
2241 cft->read_map || cft->read_seq_string)
2244 if (cft->write || cft->write_u64 || cft->write_s64 ||
2245 cft->write_string || cft->trigger)
2251 int cgroup_add_file(struct cgroup *cgrp,
2252 struct cgroup_subsys *subsys,
2253 const struct cftype *cft)
2255 struct dentry *dir = cgrp->dentry;
2256 struct dentry *dentry;
2260 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2261 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
2262 strcpy(name, subsys->name);
2265 strcat(name, cft->name);
2266 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2267 dentry = lookup_one_len(name, dir, strlen(name));
2268 if (!IS_ERR(dentry)) {
2269 mode = cgroup_file_mode(cft);
2270 error = cgroup_create_file(dentry, mode | S_IFREG,
2273 dentry->d_fsdata = (void *)cft;
2276 error = PTR_ERR(dentry);
2279 EXPORT_SYMBOL_GPL(cgroup_add_file);
2281 int cgroup_add_files(struct cgroup *cgrp,
2282 struct cgroup_subsys *subsys,
2283 const struct cftype cft[],
2287 for (i = 0; i < count; i++) {
2288 err = cgroup_add_file(cgrp, subsys, &cft[i]);
2294 EXPORT_SYMBOL_GPL(cgroup_add_files);
2297 * cgroup_task_count - count the number of tasks in a cgroup.
2298 * @cgrp: the cgroup in question
2300 * Return the number of tasks in the cgroup.
2302 int cgroup_task_count(const struct cgroup *cgrp)
2305 struct cg_cgroup_link *link;
2307 read_lock(&css_set_lock);
2308 list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
2309 count += atomic_read(&link->cg->refcount);
2311 read_unlock(&css_set_lock);
2316 * Advance a list_head iterator. The iterator should be positioned at
2317 * the start of a css_set
2319 static void cgroup_advance_iter(struct cgroup *cgrp,
2320 struct cgroup_iter *it)
2322 struct list_head *l = it->cg_link;
2323 struct cg_cgroup_link *link;
2326 /* Advance to the next non-empty css_set */
2329 if (l == &cgrp->css_sets) {
2333 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
2335 } while (list_empty(&cg->tasks));
2337 it->task = cg->tasks.next;
2341 * To reduce the fork() overhead for systems that are not actually
2342 * using their cgroups capability, we don't maintain the lists running
2343 * through each css_set to its tasks until we see the list actually
2344 * used - in other words after the first call to cgroup_iter_start().
2346 * The tasklist_lock is not held here, as do_each_thread() and
2347 * while_each_thread() are protected by RCU.
2349 static void cgroup_enable_task_cg_lists(void)
2351 struct task_struct *p, *g;
2352 write_lock(&css_set_lock);
2353 use_task_css_set_links = 1;
2354 do_each_thread(g, p) {
2357 * We should check if the process is exiting, otherwise
2358 * it will race with cgroup_exit() in that the list
2359 * entry won't be deleted though the process has exited.
2361 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2362 list_add(&p->cg_list, &p->cgroups->tasks);
2364 } while_each_thread(g, p);
2365 write_unlock(&css_set_lock);
2368 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
2371 * The first time anyone tries to iterate across a cgroup,
2372 * we need to enable the list linking each css_set to its
2373 * tasks, and fix up all existing tasks.
2375 if (!use_task_css_set_links)
2376 cgroup_enable_task_cg_lists();
2378 read_lock(&css_set_lock);
2379 it->cg_link = &cgrp->css_sets;
2380 cgroup_advance_iter(cgrp, it);
2383 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
2384 struct cgroup_iter *it)
2386 struct task_struct *res;
2387 struct list_head *l = it->task;
2388 struct cg_cgroup_link *link;
2390 /* If the iterator cg is NULL, we have no tasks */
2393 res = list_entry(l, struct task_struct, cg_list);
2394 /* Advance iterator to find next entry */
2396 link = list_entry(it->cg_link, struct cg_cgroup_link, cgrp_link_list);
2397 if (l == &link->cg->tasks) {
2398 /* We reached the end of this task list - move on to
2399 * the next cg_cgroup_link */
2400 cgroup_advance_iter(cgrp, it);
2407 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
2409 read_unlock(&css_set_lock);
2412 static inline int started_after_time(struct task_struct *t1,
2413 struct timespec *time,
2414 struct task_struct *t2)
2416 int start_diff = timespec_compare(&t1->start_time, time);
2417 if (start_diff > 0) {
2419 } else if (start_diff < 0) {
2423 * Arbitrarily, if two processes started at the same
2424 * time, we'll say that the lower pointer value
2425 * started first. Note that t2 may have exited by now
2426 * so this may not be a valid pointer any longer, but
2427 * that's fine - it still serves to distinguish
2428 * between two tasks started (effectively) simultaneously.
2435 * This function is a callback from heap_insert() and is used to order
2437 * In this case we order the heap in descending task start time.
2439 static inline int started_after(void *p1, void *p2)
2441 struct task_struct *t1 = p1;
2442 struct task_struct *t2 = p2;
2443 return started_after_time(t1, &t2->start_time, t2);
2447 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
2448 * @scan: struct cgroup_scanner containing arguments for the scan
2450 * Arguments include pointers to callback functions test_task() and
2452 * Iterate through all the tasks in a cgroup, calling test_task() for each,
2453 * and if it returns true, call process_task() for it also.
2454 * The test_task pointer may be NULL, meaning always true (select all tasks).
2455 * Effectively duplicates cgroup_iter_{start,next,end}()
2456 * but does not lock css_set_lock for the call to process_task().
2457 * The struct cgroup_scanner may be embedded in any structure of the caller's
2459 * It is guaranteed that process_task() will act on every task that
2460 * is a member of the cgroup for the duration of this call. This
2461 * function may or may not call process_task() for tasks that exit
2462 * or move to a different cgroup during the call, or are forked or
2463 * move into the cgroup during the call.
2465 * Note that test_task() may be called with locks held, and may in some
2466 * situations be called multiple times for the same task, so it should
2468 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
2469 * pre-allocated and will be used for heap operations (and its "gt" member will
2470 * be overwritten), else a temporary heap will be used (allocation of which
2471 * may cause this function to fail).
2473 int cgroup_scan_tasks(struct cgroup_scanner *scan)
2476 struct cgroup_iter it;
2477 struct task_struct *p, *dropped;
2478 /* Never dereference latest_task, since it's not refcounted */
2479 struct task_struct *latest_task = NULL;
2480 struct ptr_heap tmp_heap;
2481 struct ptr_heap *heap;
2482 struct timespec latest_time = { 0, 0 };
2485 /* The caller supplied our heap and pre-allocated its memory */
2487 heap->gt = &started_after;
2489 /* We need to allocate our own heap memory */
2491 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2493 /* cannot allocate the heap */
2499 * Scan tasks in the cgroup, using the scanner's "test_task" callback
2500 * to determine which are of interest, and using the scanner's
2501 * "process_task" callback to process any of them that need an update.
2502 * Since we don't want to hold any locks during the task updates,
2503 * gather tasks to be processed in a heap structure.
2504 * The heap is sorted by descending task start time.
2505 * If the statically-sized heap fills up, we overflow tasks that
2506 * started later, and in future iterations only consider tasks that
2507 * started after the latest task in the previous pass. This
2508 * guarantees forward progress and that we don't miss any tasks.
2511 cgroup_iter_start(scan->cg, &it);
2512 while ((p = cgroup_iter_next(scan->cg, &it))) {
2514 * Only affect tasks that qualify per the caller's callback,
2515 * if he provided one
2517 if (scan->test_task && !scan->test_task(p, scan))
2520 * Only process tasks that started after the last task
2523 if (!started_after_time(p, &latest_time, latest_task))
2525 dropped = heap_insert(heap, p);
2526 if (dropped == NULL) {
2528 * The new task was inserted; the heap wasn't
2532 } else if (dropped != p) {
2534 * The new task was inserted, and pushed out a
2538 put_task_struct(dropped);
2541 * Else the new task was newer than anything already in
2542 * the heap and wasn't inserted
2545 cgroup_iter_end(scan->cg, &it);
2548 for (i = 0; i < heap->size; i++) {
2549 struct task_struct *q = heap->ptrs[i];
2551 latest_time = q->start_time;
2554 /* Process the task per the caller's callback */
2555 scan->process_task(q, scan);
2559 * If we had to process any tasks at all, scan again
2560 * in case some of them were in the middle of forking
2561 * children that didn't get processed.
2562 * Not the most efficient way to do it, but it avoids
2563 * having to take callback_mutex in the fork path
2567 if (heap == &tmp_heap)
2568 heap_free(&tmp_heap);
2573 * Stuff for reading the 'tasks'/'procs' files.
2575 * Reading this file can return large amounts of data if a cgroup has
2576 * *lots* of attached tasks. So it may need several calls to read(),
2577 * but we cannot guarantee that the information we produce is correct
2578 * unless we produce it entirely atomically.
2583 * The following two functions "fix" the issue where there are more pids
2584 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2585 * TODO: replace with a kernel-wide solution to this problem
2587 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2588 static void *pidlist_allocate(int count)
2590 if (PIDLIST_TOO_LARGE(count))
2591 return vmalloc(count * sizeof(pid_t));
2593 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2595 static void pidlist_free(void *p)
2597 if (is_vmalloc_addr(p))
2602 static void *pidlist_resize(void *p, int newcount)
2605 /* note: if new alloc fails, old p will still be valid either way */
2606 if (is_vmalloc_addr(p)) {
2607 newlist = vmalloc(newcount * sizeof(pid_t));
2610 memcpy(newlist, p, newcount * sizeof(pid_t));
2613 newlist = krealloc(p, newcount * sizeof(pid_t), GFP_KERNEL);
2619 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2620 * If the new stripped list is sufficiently smaller and there's enough memory
2621 * to allocate a new buffer, will let go of the unneeded memory. Returns the
2622 * number of unique elements.
2624 /* is the size difference enough that we should re-allocate the array? */
2625 #define PIDLIST_REALLOC_DIFFERENCE(old, new) ((old) - PAGE_SIZE >= (new))
2626 static int pidlist_uniq(pid_t **p, int length)
2633 * we presume the 0th element is unique, so i starts at 1. trivial
2634 * edge cases first; no work needs to be done for either
2636 if (length == 0 || length == 1)
2638 /* src and dest walk down the list; dest counts unique elements */
2639 for (src = 1; src < length; src++) {
2640 /* find next unique element */
2641 while (list[src] == list[src-1]) {
2646 /* dest always points to where the next unique element goes */
2647 list[dest] = list[src];
2652 * if the length difference is large enough, we want to allocate a
2653 * smaller buffer to save memory. if this fails due to out of memory,
2654 * we'll just stay with what we've got.
2656 if (PIDLIST_REALLOC_DIFFERENCE(length, dest)) {
2657 newlist = pidlist_resize(list, dest);
2664 static int cmppid(const void *a, const void *b)
2666 return *(pid_t *)a - *(pid_t *)b;
2670 * find the appropriate pidlist for our purpose (given procs vs tasks)
2671 * returns with the lock on that pidlist already held, and takes care
2672 * of the use count, or returns NULL with no locks held if we're out of
2675 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
2676 enum cgroup_filetype type)
2678 struct cgroup_pidlist *l;
2679 /* don't need task_nsproxy() if we're looking at ourself */
2680 struct pid_namespace *ns = current->nsproxy->pid_ns;
2683 * We can't drop the pidlist_mutex before taking the l->mutex in case
2684 * the last ref-holder is trying to remove l from the list at the same
2685 * time. Holding the pidlist_mutex precludes somebody taking whichever
2686 * list we find out from under us - compare release_pid_array().
2688 mutex_lock(&cgrp->pidlist_mutex);
2689 list_for_each_entry(l, &cgrp->pidlists, links) {
2690 if (l->key.type == type && l->key.ns == ns) {
2691 /* make sure l doesn't vanish out from under us */
2692 down_write(&l->mutex);
2693 mutex_unlock(&cgrp->pidlist_mutex);
2697 /* entry not found; create a new one */
2698 l = kmalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
2700 mutex_unlock(&cgrp->pidlist_mutex);
2703 init_rwsem(&l->mutex);
2704 down_write(&l->mutex);
2706 l->key.ns = get_pid_ns(ns);
2707 l->use_count = 0; /* don't increment here */
2710 list_add(&l->links, &cgrp->pidlists);
2711 mutex_unlock(&cgrp->pidlist_mutex);
2716 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
2718 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
2719 struct cgroup_pidlist **lp)
2723 int pid, n = 0; /* used for populating the array */
2724 struct cgroup_iter it;
2725 struct task_struct *tsk;
2726 struct cgroup_pidlist *l;
2729 * If cgroup gets more users after we read count, we won't have
2730 * enough space - tough. This race is indistinguishable to the
2731 * caller from the case that the additional cgroup users didn't
2732 * show up until sometime later on.
2734 length = cgroup_task_count(cgrp);
2735 array = pidlist_allocate(length);
2738 /* now, populate the array */
2739 cgroup_iter_start(cgrp, &it);
2740 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2741 if (unlikely(n == length))
2743 /* get tgid or pid for procs or tasks file respectively */
2744 if (type == CGROUP_FILE_PROCS)
2745 pid = task_tgid_vnr(tsk);
2747 pid = task_pid_vnr(tsk);
2748 if (pid > 0) /* make sure to only use valid results */
2751 cgroup_iter_end(cgrp, &it);
2753 /* now sort & (if procs) strip out duplicates */
2754 sort(array, length, sizeof(pid_t), cmppid, NULL);
2755 if (type == CGROUP_FILE_PROCS)
2756 length = pidlist_uniq(&array, length);
2757 l = cgroup_pidlist_find(cgrp, type);
2759 pidlist_free(array);
2762 /* store array, freeing old if necessary - lock already held */
2763 pidlist_free(l->list);
2767 up_write(&l->mutex);
2773 * cgroupstats_build - build and fill cgroupstats
2774 * @stats: cgroupstats to fill information into
2775 * @dentry: A dentry entry belonging to the cgroup for which stats have
2778 * Build and fill cgroupstats so that taskstats can export it to user
2781 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2784 struct cgroup *cgrp;
2785 struct cgroup_iter it;
2786 struct task_struct *tsk;
2789 * Validate dentry by checking the superblock operations,
2790 * and make sure it's a directory.
2792 if (dentry->d_sb->s_op != &cgroup_ops ||
2793 !S_ISDIR(dentry->d_inode->i_mode))
2797 cgrp = dentry->d_fsdata;
2799 cgroup_iter_start(cgrp, &it);
2800 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2801 switch (tsk->state) {
2803 stats->nr_running++;
2805 case TASK_INTERRUPTIBLE:
2806 stats->nr_sleeping++;
2808 case TASK_UNINTERRUPTIBLE:
2809 stats->nr_uninterruptible++;
2812 stats->nr_stopped++;
2815 if (delayacct_is_task_waiting_on_io(tsk))
2816 stats->nr_io_wait++;
2820 cgroup_iter_end(cgrp, &it);
2828 * seq_file methods for the tasks/procs files. The seq_file position is the
2829 * next pid to display; the seq_file iterator is a pointer to the pid
2830 * in the cgroup->l->list array.
2833 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
2836 * Initially we receive a position value that corresponds to
2837 * one more than the last pid shown (or 0 on the first call or
2838 * after a seek to the start). Use a binary-search to find the
2839 * next pid to display, if any
2841 struct cgroup_pidlist *l = s->private;
2842 int index = 0, pid = *pos;
2845 down_read(&l->mutex);
2847 int end = l->length;
2849 while (index < end) {
2850 int mid = (index + end) / 2;
2851 if (l->list[mid] == pid) {
2854 } else if (l->list[mid] <= pid)
2860 /* If we're off the end of the array, we're done */
2861 if (index >= l->length)
2863 /* Update the abstract position to be the actual pid that we found */
2864 iter = l->list + index;
2869 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
2871 struct cgroup_pidlist *l = s->private;
2875 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
2877 struct cgroup_pidlist *l = s->private;
2879 pid_t *end = l->list + l->length;
2881 * Advance to the next pid in the array. If this goes off the
2893 static int cgroup_pidlist_show(struct seq_file *s, void *v)
2895 return seq_printf(s, "%d\n", *(int *)v);
2899 * seq_operations functions for iterating on pidlists through seq_file -
2900 * independent of whether it's tasks or procs
2902 static const struct seq_operations cgroup_pidlist_seq_operations = {
2903 .start = cgroup_pidlist_start,
2904 .stop = cgroup_pidlist_stop,
2905 .next = cgroup_pidlist_next,
2906 .show = cgroup_pidlist_show,
2909 static void cgroup_release_pid_array(struct cgroup_pidlist *l)
2912 * the case where we're the last user of this particular pidlist will
2913 * have us remove it from the cgroup's list, which entails taking the
2914 * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
2915 * pidlist_mutex, we have to take pidlist_mutex first.
2917 mutex_lock(&l->owner->pidlist_mutex);
2918 down_write(&l->mutex);
2919 BUG_ON(!l->use_count);
2920 if (!--l->use_count) {
2921 /* we're the last user if refcount is 0; remove and free */
2922 list_del(&l->links);
2923 mutex_unlock(&l->owner->pidlist_mutex);
2924 pidlist_free(l->list);
2925 put_pid_ns(l->key.ns);
2926 up_write(&l->mutex);
2930 mutex_unlock(&l->owner->pidlist_mutex);
2931 up_write(&l->mutex);
2934 static int cgroup_pidlist_release(struct inode *inode, struct file *file)
2936 struct cgroup_pidlist *l;
2937 if (!(file->f_mode & FMODE_READ))
2940 * the seq_file will only be initialized if the file was opened for
2941 * reading; hence we check if it's not null only in that case.
2943 l = ((struct seq_file *)file->private_data)->private;
2944 cgroup_release_pid_array(l);
2945 return seq_release(inode, file);
2948 static const struct file_operations cgroup_pidlist_operations = {
2950 .llseek = seq_lseek,
2951 .write = cgroup_file_write,
2952 .release = cgroup_pidlist_release,
2956 * The following functions handle opens on a file that displays a pidlist
2957 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
2960 /* helper function for the two below it */
2961 static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
2963 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2964 struct cgroup_pidlist *l;
2967 /* Nothing to do for write-only files */
2968 if (!(file->f_mode & FMODE_READ))
2971 /* have the array populated */
2972 retval = pidlist_array_load(cgrp, type, &l);
2975 /* configure file information */
2976 file->f_op = &cgroup_pidlist_operations;
2978 retval = seq_open(file, &cgroup_pidlist_seq_operations);
2980 cgroup_release_pid_array(l);
2983 ((struct seq_file *)file->private_data)->private = l;
2986 static int cgroup_tasks_open(struct inode *unused, struct file *file)
2988 return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
2990 static int cgroup_procs_open(struct inode *unused, struct file *file)
2992 return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
2995 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
2998 return notify_on_release(cgrp);
3001 static int cgroup_write_notify_on_release(struct cgroup *cgrp,
3005 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
3007 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3009 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3014 * Unregister event and free resources.
3016 * Gets called from workqueue.
3018 static void cgroup_event_remove(struct work_struct *work)
3020 struct cgroup_event *event = container_of(work, struct cgroup_event,
3022 struct cgroup *cgrp = event->cgrp;
3024 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3026 eventfd_ctx_put(event->eventfd);
3032 * Gets called on POLLHUP on eventfd when user closes it.
3034 * Called with wqh->lock held and interrupts disabled.
3036 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3037 int sync, void *key)
3039 struct cgroup_event *event = container_of(wait,
3040 struct cgroup_event, wait);
3041 struct cgroup *cgrp = event->cgrp;
3042 unsigned long flags = (unsigned long)key;
3044 if (flags & POLLHUP) {
3045 __remove_wait_queue(event->wqh, &event->wait);
3046 spin_lock(&cgrp->event_list_lock);
3047 list_del(&event->list);
3048 spin_unlock(&cgrp->event_list_lock);
3050 * We are in atomic context, but cgroup_event_remove() may
3051 * sleep, so we have to call it in workqueue.
3053 schedule_work(&event->remove);
3059 static void cgroup_event_ptable_queue_proc(struct file *file,
3060 wait_queue_head_t *wqh, poll_table *pt)
3062 struct cgroup_event *event = container_of(pt,
3063 struct cgroup_event, pt);
3066 add_wait_queue(wqh, &event->wait);
3070 * Parse input and register new cgroup event handler.
3072 * Input must be in format '<event_fd> <control_fd> <args>'.
3073 * Interpretation of args is defined by control file implementation.
3075 static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3078 struct cgroup_event *event = NULL;
3079 unsigned int efd, cfd;
3080 struct file *efile = NULL;
3081 struct file *cfile = NULL;
3085 efd = simple_strtoul(buffer, &endp, 10);
3090 cfd = simple_strtoul(buffer, &endp, 10);
3091 if ((*endp != ' ') && (*endp != '\0'))
3095 event = kzalloc(sizeof(*event), GFP_KERNEL);
3099 INIT_LIST_HEAD(&event->list);
3100 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
3101 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
3102 INIT_WORK(&event->remove, cgroup_event_remove);
3104 efile = eventfd_fget(efd);
3105 if (IS_ERR(efile)) {
3106 ret = PTR_ERR(efile);
3110 event->eventfd = eventfd_ctx_fileget(efile);
3111 if (IS_ERR(event->eventfd)) {
3112 ret = PTR_ERR(event->eventfd);
3122 /* the process need read permission on control file */
3123 ret = file_permission(cfile, MAY_READ);
3127 event->cft = __file_cft(cfile);
3128 if (IS_ERR(event->cft)) {
3129 ret = PTR_ERR(event->cft);
3133 if (!event->cft->register_event || !event->cft->unregister_event) {
3138 ret = event->cft->register_event(cgrp, event->cft,
3139 event->eventfd, buffer);
3143 if (efile->f_op->poll(efile, &event->pt) & POLLHUP) {
3144 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3150 * Events should be removed after rmdir of cgroup directory, but before
3151 * destroying subsystem state objects. Let's take reference to cgroup
3152 * directory dentry to do that.
3156 spin_lock(&cgrp->event_list_lock);
3157 list_add(&event->list, &cgrp->event_list);
3158 spin_unlock(&cgrp->event_list_lock);
3169 if (event && event->eventfd && !IS_ERR(event->eventfd))
3170 eventfd_ctx_put(event->eventfd);
3172 if (!IS_ERR_OR_NULL(efile))
3181 * for the common functions, 'private' gives the type of file
3183 /* for hysterical raisins, we can't put this on the older files */
3184 #define CGROUP_FILE_GENERIC_PREFIX "cgroup."
3185 static struct cftype files[] = {
3188 .open = cgroup_tasks_open,
3189 .write_u64 = cgroup_tasks_write,
3190 .release = cgroup_pidlist_release,
3191 .mode = S_IRUGO | S_IWUSR,
3194 .name = CGROUP_FILE_GENERIC_PREFIX "procs",
3195 .open = cgroup_procs_open,
3196 /* .write_u64 = cgroup_procs_write, TODO */
3197 .release = cgroup_pidlist_release,
3201 .name = "notify_on_release",
3202 .read_u64 = cgroup_read_notify_on_release,
3203 .write_u64 = cgroup_write_notify_on_release,
3206 .name = CGROUP_FILE_GENERIC_PREFIX "event_control",
3207 .write_string = cgroup_write_event_control,
3212 static struct cftype cft_release_agent = {
3213 .name = "release_agent",
3214 .read_seq_string = cgroup_release_agent_show,
3215 .write_string = cgroup_release_agent_write,
3216 .max_write_len = PATH_MAX,
3219 static int cgroup_populate_dir(struct cgroup *cgrp)
3222 struct cgroup_subsys *ss;
3224 /* First clear out any existing files */
3225 cgroup_clear_directory(cgrp->dentry);
3227 err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
3231 if (cgrp == cgrp->top_cgroup) {
3232 if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
3236 for_each_subsys(cgrp->root, ss) {
3237 if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
3240 /* This cgroup is ready now */
3241 for_each_subsys(cgrp->root, ss) {
3242 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3244 * Update id->css pointer and make this css visible from
3245 * CSS ID functions. This pointer will be dereferened
3246 * from RCU-read-side without locks.
3249 rcu_assign_pointer(css->id->css, css);
3255 static void init_cgroup_css(struct cgroup_subsys_state *css,
3256 struct cgroup_subsys *ss,
3257 struct cgroup *cgrp)
3260 atomic_set(&css->refcnt, 1);
3263 if (cgrp == dummytop)
3264 set_bit(CSS_ROOT, &css->flags);
3265 BUG_ON(cgrp->subsys[ss->subsys_id]);
3266 cgrp->subsys[ss->subsys_id] = css;
3269 static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
3271 /* We need to take each hierarchy_mutex in a consistent order */
3275 * No worry about a race with rebind_subsystems that might mess up the
3276 * locking order, since both parties are under cgroup_mutex.
3278 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3279 struct cgroup_subsys *ss = subsys[i];
3282 if (ss->root == root)
3283 mutex_lock(&ss->hierarchy_mutex);
3287 static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
3291 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3292 struct cgroup_subsys *ss = subsys[i];
3295 if (ss->root == root)
3296 mutex_unlock(&ss->hierarchy_mutex);
3301 * cgroup_create - create a cgroup
3302 * @parent: cgroup that will be parent of the new cgroup
3303 * @dentry: dentry of the new cgroup
3304 * @mode: mode to set on new inode
3306 * Must be called with the mutex on the parent inode held
3308 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
3311 struct cgroup *cgrp;
3312 struct cgroupfs_root *root = parent->root;
3314 struct cgroup_subsys *ss;
3315 struct super_block *sb = root->sb;
3317 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3321 /* Grab a reference on the superblock so the hierarchy doesn't
3322 * get deleted on unmount if there are child cgroups. This
3323 * can be done outside cgroup_mutex, since the sb can't
3324 * disappear while someone has an open control file on the
3326 atomic_inc(&sb->s_active);
3328 mutex_lock(&cgroup_mutex);
3330 init_cgroup_housekeeping(cgrp);
3332 cgrp->parent = parent;
3333 cgrp->root = parent->root;
3334 cgrp->top_cgroup = parent->top_cgroup;
3336 if (notify_on_release(parent))
3337 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3339 for_each_subsys(root, ss) {
3340 struct cgroup_subsys_state *css = ss->create(ss, cgrp);
3346 init_cgroup_css(css, ss, cgrp);
3348 err = alloc_css_id(ss, parent, cgrp);
3352 /* At error, ->destroy() callback has to free assigned ID. */
3355 cgroup_lock_hierarchy(root);
3356 list_add(&cgrp->sibling, &cgrp->parent->children);
3357 cgroup_unlock_hierarchy(root);
3358 root->number_of_cgroups++;
3360 err = cgroup_create_dir(cgrp, dentry, mode);
3364 /* The cgroup directory was pre-locked for us */
3365 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
3367 err = cgroup_populate_dir(cgrp);
3368 /* If err < 0, we have a half-filled directory - oh well ;) */
3370 mutex_unlock(&cgroup_mutex);
3371 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
3377 cgroup_lock_hierarchy(root);
3378 list_del(&cgrp->sibling);
3379 cgroup_unlock_hierarchy(root);
3380 root->number_of_cgroups--;
3384 for_each_subsys(root, ss) {
3385 if (cgrp->subsys[ss->subsys_id])
3386 ss->destroy(ss, cgrp);
3389 mutex_unlock(&cgroup_mutex);
3391 /* Release the reference count that we took on the superblock */
3392 deactivate_super(sb);
3398 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
3400 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
3402 /* the vfs holds inode->i_mutex already */
3403 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
3406 static int cgroup_has_css_refs(struct cgroup *cgrp)
3408 /* Check the reference count on each subsystem. Since we
3409 * already established that there are no tasks in the
3410 * cgroup, if the css refcount is also 1, then there should
3411 * be no outstanding references, so the subsystem is safe to
3412 * destroy. We scan across all subsystems rather than using
3413 * the per-hierarchy linked list of mounted subsystems since
3414 * we can be called via check_for_release() with no
3415 * synchronization other than RCU, and the subsystem linked
3416 * list isn't RCU-safe */
3419 * We won't need to lock the subsys array, because the subsystems
3420 * we're concerned about aren't going anywhere since our cgroup root
3421 * has a reference on them.
3423 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3424 struct cgroup_subsys *ss = subsys[i];
3425 struct cgroup_subsys_state *css;
3426 /* Skip subsystems not present or not in this hierarchy */
3427 if (ss == NULL || ss->root != cgrp->root)
3429 css = cgrp->subsys[ss->subsys_id];
3430 /* When called from check_for_release() it's possible
3431 * that by this point the cgroup has been removed
3432 * and the css deleted. But a false-positive doesn't
3433 * matter, since it can only happen if the cgroup
3434 * has been deleted and hence no longer needs the
3435 * release agent to be called anyway. */
3436 if (css && (atomic_read(&css->refcnt) > 1))
3443 * Atomically mark all (or else none) of the cgroup's CSS objects as
3444 * CSS_REMOVED. Return true on success, or false if the cgroup has
3445 * busy subsystems. Call with cgroup_mutex held
3448 static int cgroup_clear_css_refs(struct cgroup *cgrp)
3450 struct cgroup_subsys *ss;
3451 unsigned long flags;
3452 bool failed = false;
3453 local_irq_save(flags);
3454 for_each_subsys(cgrp->root, ss) {
3455 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3458 /* We can only remove a CSS with a refcnt==1 */
3459 refcnt = atomic_read(&css->refcnt);
3466 * Drop the refcnt to 0 while we check other
3467 * subsystems. This will cause any racing
3468 * css_tryget() to spin until we set the
3469 * CSS_REMOVED bits or abort
3471 if (atomic_cmpxchg(&css->refcnt, refcnt, 0) == refcnt)
3477 for_each_subsys(cgrp->root, ss) {
3478 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3481 * Restore old refcnt if we previously managed
3482 * to clear it from 1 to 0
3484 if (!atomic_read(&css->refcnt))
3485 atomic_set(&css->refcnt, 1);
3487 /* Commit the fact that the CSS is removed */
3488 set_bit(CSS_REMOVED, &css->flags);
3491 local_irq_restore(flags);
3495 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
3497 struct cgroup *cgrp = dentry->d_fsdata;
3499 struct cgroup *parent;
3501 struct cgroup_event *event, *tmp;
3504 /* the vfs holds both inode->i_mutex already */
3506 mutex_lock(&cgroup_mutex);
3507 if (atomic_read(&cgrp->count) != 0) {
3508 mutex_unlock(&cgroup_mutex);
3511 if (!list_empty(&cgrp->children)) {
3512 mutex_unlock(&cgroup_mutex);
3515 mutex_unlock(&cgroup_mutex);
3518 * In general, subsystem has no css->refcnt after pre_destroy(). But
3519 * in racy cases, subsystem may have to get css->refcnt after
3520 * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
3521 * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
3522 * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
3523 * and subsystem's reference count handling. Please see css_get/put
3524 * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
3526 set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3529 * Call pre_destroy handlers of subsys. Notify subsystems
3530 * that rmdir() request comes.
3532 ret = cgroup_call_pre_destroy(cgrp);
3534 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3538 mutex_lock(&cgroup_mutex);
3539 parent = cgrp->parent;
3540 if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
3541 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3542 mutex_unlock(&cgroup_mutex);
3545 prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
3546 if (!cgroup_clear_css_refs(cgrp)) {
3547 mutex_unlock(&cgroup_mutex);
3549 * Because someone may call cgroup_wakeup_rmdir_waiter() before
3550 * prepare_to_wait(), we need to check this flag.
3552 if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
3554 finish_wait(&cgroup_rmdir_waitq, &wait);
3555 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3556 if (signal_pending(current))
3560 /* NO css_tryget() can success after here. */
3561 finish_wait(&cgroup_rmdir_waitq, &wait);
3562 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3564 spin_lock(&release_list_lock);
3565 set_bit(CGRP_REMOVED, &cgrp->flags);
3566 if (!list_empty(&cgrp->release_list))
3567 list_del(&cgrp->release_list);
3568 spin_unlock(&release_list_lock);
3570 cgroup_lock_hierarchy(cgrp->root);
3571 /* delete this cgroup from parent->children */
3572 list_del(&cgrp->sibling);
3573 cgroup_unlock_hierarchy(cgrp->root);
3575 spin_lock(&cgrp->dentry->d_lock);
3576 d = dget(cgrp->dentry);
3577 spin_unlock(&d->d_lock);
3579 cgroup_d_remove_dir(d);
3582 set_bit(CGRP_RELEASABLE, &parent->flags);
3583 check_for_release(parent);
3586 * Unregister events and notify userspace.
3587 * Notify userspace about cgroup removing only after rmdir of cgroup
3588 * directory to avoid race between userspace and kernelspace
3590 spin_lock(&cgrp->event_list_lock);
3591 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
3592 list_del(&event->list);
3593 remove_wait_queue(event->wqh, &event->wait);
3594 eventfd_signal(event->eventfd, 1);
3595 schedule_work(&event->remove);
3597 spin_unlock(&cgrp->event_list_lock);
3599 mutex_unlock(&cgroup_mutex);
3603 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
3605 struct cgroup_subsys_state *css;
3607 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
3609 /* Create the top cgroup state for this subsystem */
3610 list_add(&ss->sibling, &rootnode.subsys_list);
3611 ss->root = &rootnode;
3612 css = ss->create(ss, dummytop);
3613 /* We don't handle early failures gracefully */
3614 BUG_ON(IS_ERR(css));
3615 init_cgroup_css(css, ss, dummytop);
3617 /* Update the init_css_set to contain a subsys
3618 * pointer to this state - since the subsystem is
3619 * newly registered, all tasks and hence the
3620 * init_css_set is in the subsystem's top cgroup. */
3621 init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
3623 need_forkexit_callback |= ss->fork || ss->exit;
3625 /* At system boot, before all subsystems have been
3626 * registered, no tasks have been forked, so we don't
3627 * need to invoke fork callbacks here. */
3628 BUG_ON(!list_empty(&init_task.tasks));
3630 mutex_init(&ss->hierarchy_mutex);
3631 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3634 /* this function shouldn't be used with modular subsystems, since they
3635 * need to register a subsys_id, among other things */
3640 * cgroup_load_subsys: load and register a modular subsystem at runtime
3641 * @ss: the subsystem to load
3643 * This function should be called in a modular subsystem's initcall. If the
3644 * subsystem is built as a module, it will be assigned a new subsys_id and set
3645 * up for use. If the subsystem is built-in anyway, work is delegated to the
3646 * simpler cgroup_init_subsys.
3648 int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
3651 struct cgroup_subsys_state *css;
3653 /* check name and function validity */
3654 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
3655 ss->create == NULL || ss->destroy == NULL)
3659 * we don't support callbacks in modular subsystems. this check is
3660 * before the ss->module check for consistency; a subsystem that could
3661 * be a module should still have no callbacks even if the user isn't
3662 * compiling it as one.
3664 if (ss->fork || ss->exit)
3668 * an optionally modular subsystem is built-in: we want to do nothing,
3669 * since cgroup_init_subsys will have already taken care of it.
3671 if (ss->module == NULL) {
3672 /* a few sanity checks */
3673 BUG_ON(ss->subsys_id >= CGROUP_BUILTIN_SUBSYS_COUNT);
3674 BUG_ON(subsys[ss->subsys_id] != ss);
3679 * need to register a subsys id before anything else - for example,
3680 * init_cgroup_css needs it.
3682 mutex_lock(&cgroup_mutex);
3683 /* find the first empty slot in the array */
3684 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
3685 if (subsys[i] == NULL)
3688 if (i == CGROUP_SUBSYS_COUNT) {
3689 /* maximum number of subsystems already registered! */
3690 mutex_unlock(&cgroup_mutex);
3693 /* assign ourselves the subsys_id */
3698 * no ss->create seems to need anything important in the ss struct, so
3699 * this can happen first (i.e. before the rootnode attachment).
3701 css = ss->create(ss, dummytop);
3703 /* failure case - need to deassign the subsys[] slot. */
3705 mutex_unlock(&cgroup_mutex);
3706 return PTR_ERR(css);
3709 list_add(&ss->sibling, &rootnode.subsys_list);
3710 ss->root = &rootnode;
3712 /* our new subsystem will be attached to the dummy hierarchy. */
3713 init_cgroup_css(css, ss, dummytop);
3714 /* init_idr must be after init_cgroup_css because it sets css->id. */
3716 int ret = cgroup_init_idr(ss, css);
3718 dummytop->subsys[ss->subsys_id] = NULL;
3719 ss->destroy(ss, dummytop);
3721 mutex_unlock(&cgroup_mutex);
3727 * Now we need to entangle the css into the existing css_sets. unlike
3728 * in cgroup_init_subsys, there are now multiple css_sets, so each one
3729 * will need a new pointer to it; done by iterating the css_set_table.
3730 * furthermore, modifying the existing css_sets will corrupt the hash
3731 * table state, so each changed css_set will need its hash recomputed.
3732 * this is all done under the css_set_lock.
3734 write_lock(&css_set_lock);
3735 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
3737 struct hlist_node *node, *tmp;
3738 struct hlist_head *bucket = &css_set_table[i], *new_bucket;
3740 hlist_for_each_entry_safe(cg, node, tmp, bucket, hlist) {
3741 /* skip entries that we already rehashed */
3742 if (cg->subsys[ss->subsys_id])
3744 /* remove existing entry */
3745 hlist_del(&cg->hlist);
3747 cg->subsys[ss->subsys_id] = css;
3748 /* recompute hash and restore entry */
3749 new_bucket = css_set_hash(cg->subsys);
3750 hlist_add_head(&cg->hlist, new_bucket);
3753 write_unlock(&css_set_lock);
3755 mutex_init(&ss->hierarchy_mutex);
3756 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3760 mutex_unlock(&cgroup_mutex);
3763 EXPORT_SYMBOL_GPL(cgroup_load_subsys);
3766 * cgroup_unload_subsys: unload a modular subsystem
3767 * @ss: the subsystem to unload
3769 * This function should be called in a modular subsystem's exitcall. When this
3770 * function is invoked, the refcount on the subsystem's module will be 0, so
3771 * the subsystem will not be attached to any hierarchy.
3773 void cgroup_unload_subsys(struct cgroup_subsys *ss)
3775 struct cg_cgroup_link *link;
3776 struct hlist_head *hhead;
3778 BUG_ON(ss->module == NULL);
3781 * we shouldn't be called if the subsystem is in use, and the use of
3782 * try_module_get in parse_cgroupfs_options should ensure that it
3783 * doesn't start being used while we're killing it off.
3785 BUG_ON(ss->root != &rootnode);
3787 mutex_lock(&cgroup_mutex);
3788 /* deassign the subsys_id */
3789 BUG_ON(ss->subsys_id < CGROUP_BUILTIN_SUBSYS_COUNT);
3790 subsys[ss->subsys_id] = NULL;
3792 /* remove subsystem from rootnode's list of subsystems */
3793 list_del(&ss->sibling);
3796 * disentangle the css from all css_sets attached to the dummytop. as
3797 * in loading, we need to pay our respects to the hashtable gods.
3799 write_lock(&css_set_lock);
3800 list_for_each_entry(link, &dummytop->css_sets, cgrp_link_list) {
3801 struct css_set *cg = link->cg;
3803 hlist_del(&cg->hlist);
3804 BUG_ON(!cg->subsys[ss->subsys_id]);
3805 cg->subsys[ss->subsys_id] = NULL;
3806 hhead = css_set_hash(cg->subsys);
3807 hlist_add_head(&cg->hlist, hhead);
3809 write_unlock(&css_set_lock);
3812 * remove subsystem's css from the dummytop and free it - need to free
3813 * before marking as null because ss->destroy needs the cgrp->subsys
3814 * pointer to find their state. note that this also takes care of
3815 * freeing the css_id.
3817 ss->destroy(ss, dummytop);
3818 dummytop->subsys[ss->subsys_id] = NULL;
3820 mutex_unlock(&cgroup_mutex);
3822 EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
3825 * cgroup_init_early - cgroup initialization at system boot
3827 * Initialize cgroups at system boot, and initialize any
3828 * subsystems that request early init.
3830 int __init cgroup_init_early(void)
3833 atomic_set(&init_css_set.refcount, 1);
3834 INIT_LIST_HEAD(&init_css_set.cg_links);
3835 INIT_LIST_HEAD(&init_css_set.tasks);
3836 INIT_HLIST_NODE(&init_css_set.hlist);
3838 init_cgroup_root(&rootnode);
3840 init_task.cgroups = &init_css_set;
3842 init_css_set_link.cg = &init_css_set;
3843 init_css_set_link.cgrp = dummytop;
3844 list_add(&init_css_set_link.cgrp_link_list,
3845 &rootnode.top_cgroup.css_sets);
3846 list_add(&init_css_set_link.cg_link_list,
3847 &init_css_set.cg_links);
3849 for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
3850 INIT_HLIST_HEAD(&css_set_table[i]);
3852 /* at bootup time, we don't worry about modular subsystems */
3853 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3854 struct cgroup_subsys *ss = subsys[i];
3857 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
3858 BUG_ON(!ss->create);
3859 BUG_ON(!ss->destroy);
3860 if (ss->subsys_id != i) {
3861 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
3862 ss->name, ss->subsys_id);
3867 cgroup_init_subsys(ss);
3873 * cgroup_init - cgroup initialization
3875 * Register cgroup filesystem and /proc file, and initialize
3876 * any subsystems that didn't request early init.
3878 int __init cgroup_init(void)
3882 struct hlist_head *hhead;
3884 err = bdi_init(&cgroup_backing_dev_info);
3888 /* at bootup time, we don't worry about modular subsystems */
3889 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3890 struct cgroup_subsys *ss = subsys[i];
3891 if (!ss->early_init)
3892 cgroup_init_subsys(ss);
3894 cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
3897 /* Add init_css_set to the hash table */
3898 hhead = css_set_hash(init_css_set.subsys);
3899 hlist_add_head(&init_css_set.hlist, hhead);
3900 BUG_ON(!init_root_id(&rootnode));
3902 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
3908 err = register_filesystem(&cgroup_fs_type);
3910 kobject_put(cgroup_kobj);
3914 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
3918 bdi_destroy(&cgroup_backing_dev_info);
3924 * proc_cgroup_show()
3925 * - Print task's cgroup paths into seq_file, one line for each hierarchy
3926 * - Used for /proc/<pid>/cgroup.
3927 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
3928 * doesn't really matter if tsk->cgroup changes after we read it,
3929 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
3930 * anyway. No need to check that tsk->cgroup != NULL, thanks to
3931 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
3932 * cgroup to top_cgroup.
3935 /* TODO: Use a proper seq_file iterator */
3936 static int proc_cgroup_show(struct seq_file *m, void *v)
3939 struct task_struct *tsk;
3942 struct cgroupfs_root *root;
3945 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
3951 tsk = get_pid_task(pid, PIDTYPE_PID);
3957 mutex_lock(&cgroup_mutex);
3959 for_each_active_root(root) {
3960 struct cgroup_subsys *ss;
3961 struct cgroup *cgrp;
3964 seq_printf(m, "%d:", root->hierarchy_id);
3965 for_each_subsys(root, ss)
3966 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
3967 if (strlen(root->name))
3968 seq_printf(m, "%sname=%s", count ? "," : "",
3971 cgrp = task_cgroup_from_root(tsk, root);
3972 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
3980 mutex_unlock(&cgroup_mutex);
3981 put_task_struct(tsk);
3988 static int cgroup_open(struct inode *inode, struct file *file)
3990 struct pid *pid = PROC_I(inode)->pid;
3991 return single_open(file, proc_cgroup_show, pid);
3994 const struct file_operations proc_cgroup_operations = {
3995 .open = cgroup_open,
3997 .llseek = seq_lseek,
3998 .release = single_release,
4001 /* Display information about each subsystem and each hierarchy */
4002 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4006 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4008 * ideally we don't want subsystems moving around while we do this.
4009 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4010 * subsys/hierarchy state.
4012 mutex_lock(&cgroup_mutex);
4013 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
4014 struct cgroup_subsys *ss = subsys[i];
4017 seq_printf(m, "%s\t%d\t%d\t%d\n",
4018 ss->name, ss->root->hierarchy_id,
4019 ss->root->number_of_cgroups, !ss->disabled);
4021 mutex_unlock(&cgroup_mutex);
4025 static int cgroupstats_open(struct inode *inode, struct file *file)
4027 return single_open(file, proc_cgroupstats_show, NULL);
4030 static const struct file_operations proc_cgroupstats_operations = {
4031 .open = cgroupstats_open,
4033 .llseek = seq_lseek,
4034 .release = single_release,
4038 * cgroup_fork - attach newly forked task to its parents cgroup.
4039 * @child: pointer to task_struct of forking parent process.
4041 * Description: A task inherits its parent's cgroup at fork().
4043 * A pointer to the shared css_set was automatically copied in
4044 * fork.c by dup_task_struct(). However, we ignore that copy, since
4045 * it was not made under the protection of RCU or cgroup_mutex, so
4046 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4047 * have already changed current->cgroups, allowing the previously
4048 * referenced cgroup group to be removed and freed.
4050 * At the point that cgroup_fork() is called, 'current' is the parent
4051 * task, and the passed argument 'child' points to the child task.
4053 void cgroup_fork(struct task_struct *child)
4056 child->cgroups = current->cgroups;
4057 get_css_set(child->cgroups);
4058 task_unlock(current);
4059 INIT_LIST_HEAD(&child->cg_list);
4063 * cgroup_fork_callbacks - run fork callbacks
4064 * @child: the new task
4066 * Called on a new task very soon before adding it to the
4067 * tasklist. No need to take any locks since no-one can
4068 * be operating on this task.
4070 void cgroup_fork_callbacks(struct task_struct *child)
4072 if (need_forkexit_callback) {
4075 * forkexit callbacks are only supported for builtin
4076 * subsystems, and the builtin section of the subsys array is
4077 * immutable, so we don't need to lock the subsys array here.
4079 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4080 struct cgroup_subsys *ss = subsys[i];
4082 ss->fork(ss, child);
4088 * cgroup_post_fork - called on a new task after adding it to the task list
4089 * @child: the task in question
4091 * Adds the task to the list running through its css_set if necessary.
4092 * Has to be after the task is visible on the task list in case we race
4093 * with the first call to cgroup_iter_start() - to guarantee that the
4094 * new task ends up on its list.
4096 void cgroup_post_fork(struct task_struct *child)
4098 if (use_task_css_set_links) {
4099 write_lock(&css_set_lock);
4101 if (list_empty(&child->cg_list))
4102 list_add(&child->cg_list, &child->cgroups->tasks);
4104 write_unlock(&css_set_lock);
4108 * cgroup_exit - detach cgroup from exiting task
4109 * @tsk: pointer to task_struct of exiting process
4110 * @run_callback: run exit callbacks?
4112 * Description: Detach cgroup from @tsk and release it.
4114 * Note that cgroups marked notify_on_release force every task in
4115 * them to take the global cgroup_mutex mutex when exiting.
4116 * This could impact scaling on very large systems. Be reluctant to
4117 * use notify_on_release cgroups where very high task exit scaling
4118 * is required on large systems.
4120 * the_top_cgroup_hack:
4122 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4124 * We call cgroup_exit() while the task is still competent to
4125 * handle notify_on_release(), then leave the task attached to the
4126 * root cgroup in each hierarchy for the remainder of its exit.
4128 * To do this properly, we would increment the reference count on
4129 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4130 * code we would add a second cgroup function call, to drop that
4131 * reference. This would just create an unnecessary hot spot on
4132 * the top_cgroup reference count, to no avail.
4134 * Normally, holding a reference to a cgroup without bumping its
4135 * count is unsafe. The cgroup could go away, or someone could
4136 * attach us to a different cgroup, decrementing the count on
4137 * the first cgroup that we never incremented. But in this case,
4138 * top_cgroup isn't going away, and either task has PF_EXITING set,
4139 * which wards off any cgroup_attach_task() attempts, or task is a failed
4140 * fork, never visible to cgroup_attach_task.
4142 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4147 if (run_callbacks && need_forkexit_callback) {
4149 * modular subsystems can't use callbacks, so no need to lock
4152 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4153 struct cgroup_subsys *ss = subsys[i];
4160 * Unlink from the css_set task list if necessary.
4161 * Optimistically check cg_list before taking
4164 if (!list_empty(&tsk->cg_list)) {
4165 write_lock(&css_set_lock);
4166 if (!list_empty(&tsk->cg_list))
4167 list_del(&tsk->cg_list);
4168 write_unlock(&css_set_lock);
4171 /* Reassign the task to the init_css_set. */
4174 tsk->cgroups = &init_css_set;
4177 put_css_set_taskexit(cg);
4181 * cgroup_clone - clone the cgroup the given subsystem is attached to
4182 * @tsk: the task to be moved
4183 * @subsys: the given subsystem
4184 * @nodename: the name for the new cgroup
4186 * Duplicate the current cgroup in the hierarchy that the given
4187 * subsystem is attached to, and move this task into the new
4190 int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
4193 struct dentry *dentry;
4195 struct cgroup *parent, *child;
4196 struct inode *inode;
4198 struct cgroupfs_root *root;
4199 struct cgroup_subsys *ss;
4201 /* We shouldn't be called by an unregistered subsystem */
4202 BUG_ON(!subsys->active);
4204 /* First figure out what hierarchy and cgroup we're dealing
4205 * with, and pin them so we can drop cgroup_mutex */
4206 mutex_lock(&cgroup_mutex);
4208 root = subsys->root;
4209 if (root == &rootnode) {
4210 mutex_unlock(&cgroup_mutex);
4214 /* Pin the hierarchy */
4215 if (!atomic_inc_not_zero(&root->sb->s_active)) {
4216 /* We race with the final deactivate_super() */
4217 mutex_unlock(&cgroup_mutex);
4221 /* Keep the cgroup alive */
4223 parent = task_cgroup(tsk, subsys->subsys_id);
4228 mutex_unlock(&cgroup_mutex);
4230 /* Now do the VFS work to create a cgroup */
4231 inode = parent->dentry->d_inode;
4233 /* Hold the parent directory mutex across this operation to
4234 * stop anyone else deleting the new cgroup */
4235 mutex_lock(&inode->i_mutex);
4236 dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
4237 if (IS_ERR(dentry)) {
4239 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
4241 ret = PTR_ERR(dentry);
4245 /* Create the cgroup directory, which also creates the cgroup */
4246 ret = vfs_mkdir(inode, dentry, 0755);
4247 child = __d_cgrp(dentry);
4251 "Failed to create cgroup %s: %d\n", nodename,
4256 /* The cgroup now exists. Retake cgroup_mutex and check
4257 * that we're still in the same state that we thought we
4259 mutex_lock(&cgroup_mutex);
4260 if ((root != subsys->root) ||
4261 (parent != task_cgroup(tsk, subsys->subsys_id))) {
4262 /* Aargh, we raced ... */
4263 mutex_unlock(&inode->i_mutex);
4266 deactivate_super(root->sb);
4267 /* The cgroup is still accessible in the VFS, but
4268 * we're not going to try to rmdir() it at this
4271 "Race in cgroup_clone() - leaking cgroup %s\n",
4276 /* do any required auto-setup */
4277 for_each_subsys(root, ss) {
4279 ss->post_clone(ss, child);
4282 /* All seems fine. Finish by moving the task into the new cgroup */
4283 ret = cgroup_attach_task(child, tsk);
4284 mutex_unlock(&cgroup_mutex);
4287 mutex_unlock(&inode->i_mutex);
4289 mutex_lock(&cgroup_mutex);
4291 mutex_unlock(&cgroup_mutex);
4292 deactivate_super(root->sb);
4297 * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
4298 * @cgrp: the cgroup in question
4299 * @task: the task in question
4301 * See if @cgrp is a descendant of @task's cgroup in the appropriate
4304 * If we are sending in dummytop, then presumably we are creating
4305 * the top cgroup in the subsystem.
4307 * Called only by the ns (nsproxy) cgroup.
4309 int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
4312 struct cgroup *target;
4314 if (cgrp == dummytop)
4317 target = task_cgroup_from_root(task, cgrp->root);
4318 while (cgrp != target && cgrp!= cgrp->top_cgroup)
4319 cgrp = cgrp->parent;
4320 ret = (cgrp == target);
4324 static void check_for_release(struct cgroup *cgrp)
4326 /* All of these checks rely on RCU to keep the cgroup
4327 * structure alive */
4328 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
4329 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
4330 /* Control Group is currently removeable. If it's not
4331 * already queued for a userspace notification, queue
4333 int need_schedule_work = 0;
4334 spin_lock(&release_list_lock);
4335 if (!cgroup_is_removed(cgrp) &&
4336 list_empty(&cgrp->release_list)) {
4337 list_add(&cgrp->release_list, &release_list);
4338 need_schedule_work = 1;
4340 spin_unlock(&release_list_lock);
4341 if (need_schedule_work)
4342 schedule_work(&release_agent_work);
4346 /* Caller must verify that the css is not for root cgroup */
4347 void __css_put(struct cgroup_subsys_state *css, int count)
4349 struct cgroup *cgrp = css->cgroup;
4352 val = atomic_sub_return(count, &css->refcnt);
4354 if (notify_on_release(cgrp)) {
4355 set_bit(CGRP_RELEASABLE, &cgrp->flags);
4356 check_for_release(cgrp);
4358 cgroup_wakeup_rmdir_waiter(cgrp);
4361 WARN_ON_ONCE(val < 1);
4363 EXPORT_SYMBOL_GPL(__css_put);
4366 * Notify userspace when a cgroup is released, by running the
4367 * configured release agent with the name of the cgroup (path
4368 * relative to the root of cgroup file system) as the argument.
4370 * Most likely, this user command will try to rmdir this cgroup.
4372 * This races with the possibility that some other task will be
4373 * attached to this cgroup before it is removed, or that some other
4374 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4375 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4376 * unused, and this cgroup will be reprieved from its death sentence,
4377 * to continue to serve a useful existence. Next time it's released,
4378 * we will get notified again, if it still has 'notify_on_release' set.
4380 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4381 * means only wait until the task is successfully execve()'d. The
4382 * separate release agent task is forked by call_usermodehelper(),
4383 * then control in this thread returns here, without waiting for the
4384 * release agent task. We don't bother to wait because the caller of
4385 * this routine has no use for the exit status of the release agent
4386 * task, so no sense holding our caller up for that.
4388 static void cgroup_release_agent(struct work_struct *work)
4390 BUG_ON(work != &release_agent_work);
4391 mutex_lock(&cgroup_mutex);
4392 spin_lock(&release_list_lock);
4393 while (!list_empty(&release_list)) {
4394 char *argv[3], *envp[3];
4396 char *pathbuf = NULL, *agentbuf = NULL;
4397 struct cgroup *cgrp = list_entry(release_list.next,
4400 list_del_init(&cgrp->release_list);
4401 spin_unlock(&release_list_lock);
4402 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4405 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4407 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4412 argv[i++] = agentbuf;
4413 argv[i++] = pathbuf;
4417 /* minimal command environment */
4418 envp[i++] = "HOME=/";
4419 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4422 /* Drop the lock while we invoke the usermode helper,
4423 * since the exec could involve hitting disk and hence
4424 * be a slow process */
4425 mutex_unlock(&cgroup_mutex);
4426 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4427 mutex_lock(&cgroup_mutex);
4431 spin_lock(&release_list_lock);
4433 spin_unlock(&release_list_lock);
4434 mutex_unlock(&cgroup_mutex);
4437 static int __init cgroup_disable(char *str)
4442 while ((token = strsep(&str, ",")) != NULL) {
4446 * cgroup_disable, being at boot time, can't know about module
4447 * subsystems, so we don't worry about them.
4449 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4450 struct cgroup_subsys *ss = subsys[i];
4452 if (!strcmp(token, ss->name)) {
4454 printk(KERN_INFO "Disabling %s control group"
4455 " subsystem\n", ss->name);
4462 __setup("cgroup_disable=", cgroup_disable);
4465 * Functons for CSS ID.
4469 *To get ID other than 0, this should be called when !cgroup_is_removed().
4471 unsigned short css_id(struct cgroup_subsys_state *css)
4473 struct css_id *cssid;
4476 * This css_id() can return correct value when somone has refcnt
4477 * on this or this is under rcu_read_lock(). Once css->id is allocated,
4478 * it's unchanged until freed.
4480 cssid = rcu_dereference_check(css->id,
4481 rcu_read_lock_held() || atomic_read(&css->refcnt));
4487 EXPORT_SYMBOL_GPL(css_id);
4489 unsigned short css_depth(struct cgroup_subsys_state *css)
4491 struct css_id *cssid;
4493 cssid = rcu_dereference_check(css->id,
4494 rcu_read_lock_held() || atomic_read(&css->refcnt));
4497 return cssid->depth;
4500 EXPORT_SYMBOL_GPL(css_depth);
4503 * css_is_ancestor - test "root" css is an ancestor of "child"
4504 * @child: the css to be tested.
4505 * @root: the css supporsed to be an ancestor of the child.
4507 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
4508 * this function reads css->id, this use rcu_dereference() and rcu_read_lock().
4509 * But, considering usual usage, the csses should be valid objects after test.
4510 * Assuming that the caller will do some action to the child if this returns
4511 * returns true, the caller must take "child";s reference count.
4512 * If "child" is valid object and this returns true, "root" is valid, too.
4515 bool css_is_ancestor(struct cgroup_subsys_state *child,
4516 const struct cgroup_subsys_state *root)
4518 struct css_id *child_id;
4519 struct css_id *root_id;
4523 child_id = rcu_dereference(child->id);
4524 root_id = rcu_dereference(root->id);
4527 || (child_id->depth < root_id->depth)
4528 || (child_id->stack[root_id->depth] != root_id->id))
4534 static void __free_css_id_cb(struct rcu_head *head)
4538 id = container_of(head, struct css_id, rcu_head);
4542 void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
4544 struct css_id *id = css->id;
4545 /* When this is called before css_id initialization, id can be NULL */
4549 BUG_ON(!ss->use_id);
4551 rcu_assign_pointer(id->css, NULL);
4552 rcu_assign_pointer(css->id, NULL);
4553 spin_lock(&ss->id_lock);
4554 idr_remove(&ss->idr, id->id);
4555 spin_unlock(&ss->id_lock);
4556 call_rcu(&id->rcu_head, __free_css_id_cb);
4558 EXPORT_SYMBOL_GPL(free_css_id);
4561 * This is called by init or create(). Then, calls to this function are
4562 * always serialized (By cgroup_mutex() at create()).
4565 static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
4567 struct css_id *newid;
4568 int myid, error, size;
4570 BUG_ON(!ss->use_id);
4572 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
4573 newid = kzalloc(size, GFP_KERNEL);
4575 return ERR_PTR(-ENOMEM);
4577 if (unlikely(!idr_pre_get(&ss->idr, GFP_KERNEL))) {
4581 spin_lock(&ss->id_lock);
4582 /* Don't use 0. allocates an ID of 1-65535 */
4583 error = idr_get_new_above(&ss->idr, newid, 1, &myid);
4584 spin_unlock(&ss->id_lock);
4586 /* Returns error when there are no free spaces for new ID.*/
4591 if (myid > CSS_ID_MAX)
4595 newid->depth = depth;
4599 spin_lock(&ss->id_lock);
4600 idr_remove(&ss->idr, myid);
4601 spin_unlock(&ss->id_lock);
4604 return ERR_PTR(error);
4608 static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
4609 struct cgroup_subsys_state *rootcss)
4611 struct css_id *newid;
4613 spin_lock_init(&ss->id_lock);
4616 newid = get_new_cssid(ss, 0);
4618 return PTR_ERR(newid);
4620 newid->stack[0] = newid->id;
4621 newid->css = rootcss;
4622 rootcss->id = newid;
4626 static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
4627 struct cgroup *child)
4629 int subsys_id, i, depth = 0;
4630 struct cgroup_subsys_state *parent_css, *child_css;
4631 struct css_id *child_id, *parent_id;
4633 subsys_id = ss->subsys_id;
4634 parent_css = parent->subsys[subsys_id];
4635 child_css = child->subsys[subsys_id];
4636 parent_id = parent_css->id;
4637 depth = parent_id->depth + 1;
4639 child_id = get_new_cssid(ss, depth);
4640 if (IS_ERR(child_id))
4641 return PTR_ERR(child_id);
4643 for (i = 0; i < depth; i++)
4644 child_id->stack[i] = parent_id->stack[i];
4645 child_id->stack[depth] = child_id->id;
4647 * child_id->css pointer will be set after this cgroup is available
4648 * see cgroup_populate_dir()
4650 rcu_assign_pointer(child_css->id, child_id);
4656 * css_lookup - lookup css by id
4657 * @ss: cgroup subsys to be looked into.
4660 * Returns pointer to cgroup_subsys_state if there is valid one with id.
4661 * NULL if not. Should be called under rcu_read_lock()
4663 struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
4665 struct css_id *cssid = NULL;
4667 BUG_ON(!ss->use_id);
4668 cssid = idr_find(&ss->idr, id);
4670 if (unlikely(!cssid))
4673 return rcu_dereference(cssid->css);
4675 EXPORT_SYMBOL_GPL(css_lookup);
4678 * css_get_next - lookup next cgroup under specified hierarchy.
4679 * @ss: pointer to subsystem
4680 * @id: current position of iteration.
4681 * @root: pointer to css. search tree under this.
4682 * @foundid: position of found object.
4684 * Search next css under the specified hierarchy of rootid. Calling under
4685 * rcu_read_lock() is necessary. Returns NULL if it reaches the end.
4687 struct cgroup_subsys_state *
4688 css_get_next(struct cgroup_subsys *ss, int id,
4689 struct cgroup_subsys_state *root, int *foundid)
4691 struct cgroup_subsys_state *ret = NULL;
4694 int rootid = css_id(root);
4695 int depth = css_depth(root);
4700 BUG_ON(!ss->use_id);
4701 /* fill start point for scan */
4705 * scan next entry from bitmap(tree), tmpid is updated after
4708 spin_lock(&ss->id_lock);
4709 tmp = idr_get_next(&ss->idr, &tmpid);
4710 spin_unlock(&ss->id_lock);
4714 if (tmp->depth >= depth && tmp->stack[depth] == rootid) {
4715 ret = rcu_dereference(tmp->css);
4721 /* continue to scan from next id */
4727 #ifdef CONFIG_CGROUP_DEBUG
4728 static struct cgroup_subsys_state *debug_create(struct cgroup_subsys *ss,
4729 struct cgroup *cont)
4731 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4734 return ERR_PTR(-ENOMEM);
4739 static void debug_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
4741 kfree(cont->subsys[debug_subsys_id]);
4744 static u64 cgroup_refcount_read(struct cgroup *cont, struct cftype *cft)
4746 return atomic_read(&cont->count);
4749 static u64 debug_taskcount_read(struct cgroup *cont, struct cftype *cft)
4751 return cgroup_task_count(cont);
4754 static u64 current_css_set_read(struct cgroup *cont, struct cftype *cft)
4756 return (u64)(unsigned long)current->cgroups;
4759 static u64 current_css_set_refcount_read(struct cgroup *cont,
4765 count = atomic_read(¤t->cgroups->refcount);
4770 static int current_css_set_cg_links_read(struct cgroup *cont,
4772 struct seq_file *seq)
4774 struct cg_cgroup_link *link;
4777 read_lock(&css_set_lock);
4779 cg = rcu_dereference(current->cgroups);
4780 list_for_each_entry(link, &cg->cg_links, cg_link_list) {
4781 struct cgroup *c = link->cgrp;
4785 name = c->dentry->d_name.name;
4788 seq_printf(seq, "Root %d group %s\n",
4789 c->root->hierarchy_id, name);
4792 read_unlock(&css_set_lock);
4796 #define MAX_TASKS_SHOWN_PER_CSS 25
4797 static int cgroup_css_links_read(struct cgroup *cont,
4799 struct seq_file *seq)
4801 struct cg_cgroup_link *link;
4803 read_lock(&css_set_lock);
4804 list_for_each_entry(link, &cont->css_sets, cgrp_link_list) {
4805 struct css_set *cg = link->cg;
4806 struct task_struct *task;
4808 seq_printf(seq, "css_set %p\n", cg);
4809 list_for_each_entry(task, &cg->tasks, cg_list) {
4810 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4811 seq_puts(seq, " ...\n");
4814 seq_printf(seq, " task %d\n",
4815 task_pid_vnr(task));
4819 read_unlock(&css_set_lock);
4823 static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
4825 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
4828 static struct cftype debug_files[] = {
4830 .name = "cgroup_refcount",
4831 .read_u64 = cgroup_refcount_read,
4834 .name = "taskcount",
4835 .read_u64 = debug_taskcount_read,
4839 .name = "current_css_set",
4840 .read_u64 = current_css_set_read,
4844 .name = "current_css_set_refcount",
4845 .read_u64 = current_css_set_refcount_read,
4849 .name = "current_css_set_cg_links",
4850 .read_seq_string = current_css_set_cg_links_read,
4854 .name = "cgroup_css_links",
4855 .read_seq_string = cgroup_css_links_read,
4859 .name = "releasable",
4860 .read_u64 = releasable_read,
4864 static int debug_populate(struct cgroup_subsys *ss, struct cgroup *cont)
4866 return cgroup_add_files(cont, ss, debug_files,
4867 ARRAY_SIZE(debug_files));
4870 struct cgroup_subsys debug_subsys = {
4872 .create = debug_create,
4873 .destroy = debug_destroy,
4874 .populate = debug_populate,
4875 .subsys_id = debug_subsys_id,
4877 #endif /* CONFIG_CGROUP_DEBUG */