2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #include <linux/cgroup.h>
30 #include <linux/ctype.h>
31 #include <linux/errno.h>
33 #include <linux/kernel.h>
34 #include <linux/list.h>
36 #include <linux/mutex.h>
37 #include <linux/mount.h>
38 #include <linux/pagemap.h>
39 #include <linux/proc_fs.h>
40 #include <linux/rcupdate.h>
41 #include <linux/sched.h>
42 #include <linux/backing-dev.h>
43 #include <linux/seq_file.h>
44 #include <linux/slab.h>
45 #include <linux/magic.h>
46 #include <linux/spinlock.h>
47 #include <linux/string.h>
48 #include <linux/sort.h>
49 #include <linux/kmod.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/cgroupstats.h>
53 #include <linux/hash.h>
54 #include <linux/namei.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/eventfd.h>
59 #include <linux/poll.h>
61 #include <asm/atomic.h>
63 static DEFINE_MUTEX(cgroup_mutex);
66 * Generate an array of cgroup subsystem pointers. At boot time, this is
67 * populated up to CGROUP_BUILTIN_SUBSYS_COUNT, and modular subsystems are
68 * registered after that. The mutable section of this array is protected by
71 #define SUBSYS(_x) &_x ## _subsys,
72 static struct cgroup_subsys *subsys[CGROUP_SUBSYS_COUNT] = {
73 #include <linux/cgroup_subsys.h>
76 #define MAX_CGROUP_ROOT_NAMELEN 64
79 * A cgroupfs_root represents the root of a cgroup hierarchy,
80 * and may be associated with a superblock to form an active
83 struct cgroupfs_root {
84 struct super_block *sb;
87 * The bitmask of subsystems intended to be attached to this
90 unsigned long subsys_bits;
92 /* Unique id for this hierarchy. */
95 /* The bitmask of subsystems currently attached to this hierarchy */
96 unsigned long actual_subsys_bits;
98 /* A list running through the attached subsystems */
99 struct list_head subsys_list;
101 /* The root cgroup for this hierarchy */
102 struct cgroup top_cgroup;
104 /* Tracks how many cgroups are currently defined in hierarchy.*/
105 int number_of_cgroups;
107 /* A list running through the active hierarchies */
108 struct list_head root_list;
110 /* Hierarchy-specific flags */
113 /* The path to use for release notifications. */
114 char release_agent_path[PATH_MAX];
116 /* The name for this hierarchy - may be empty */
117 char name[MAX_CGROUP_ROOT_NAMELEN];
121 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
122 * subsystems that are otherwise unattached - it never has more than a
123 * single cgroup, and all tasks are part of that cgroup.
125 static struct cgroupfs_root rootnode;
128 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
129 * cgroup_subsys->use_id != 0.
131 #define CSS_ID_MAX (65535)
134 * The css to which this ID points. This pointer is set to valid value
135 * after cgroup is populated. If cgroup is removed, this will be NULL.
136 * This pointer is expected to be RCU-safe because destroy()
137 * is called after synchronize_rcu(). But for safe use, css_is_removed()
138 * css_tryget() should be used for avoiding race.
140 struct cgroup_subsys_state __rcu *css;
146 * Depth in hierarchy which this ID belongs to.
148 unsigned short depth;
150 * ID is freed by RCU. (and lookup routine is RCU safe.)
152 struct rcu_head rcu_head;
154 * Hierarchy of CSS ID belongs to.
156 unsigned short stack[0]; /* Array of Length (depth+1) */
160 * cgroup_event represents events which userspace want to recieve.
162 struct cgroup_event {
164 * Cgroup which the event belongs to.
168 * Control file which the event associated.
172 * eventfd to signal userspace about the event.
174 struct eventfd_ctx *eventfd;
176 * Each of these stored in a list by the cgroup.
178 struct list_head list;
180 * All fields below needed to unregister event when
181 * userspace closes eventfd.
184 wait_queue_head_t *wqh;
186 struct work_struct remove;
189 /* The list of hierarchy roots */
191 static LIST_HEAD(roots);
192 static int root_count;
194 static DEFINE_IDA(hierarchy_ida);
195 static int next_hierarchy_id;
196 static DEFINE_SPINLOCK(hierarchy_id_lock);
198 /* dummytop is a shorthand for the dummy hierarchy's top cgroup */
199 #define dummytop (&rootnode.top_cgroup)
201 /* This flag indicates whether tasks in the fork and exit paths should
202 * check for fork/exit handlers to call. This avoids us having to do
203 * extra work in the fork/exit path if none of the subsystems need to
206 static int need_forkexit_callback __read_mostly;
208 #ifdef CONFIG_PROVE_LOCKING
209 int cgroup_lock_is_held(void)
211 return lockdep_is_held(&cgroup_mutex);
213 #else /* #ifdef CONFIG_PROVE_LOCKING */
214 int cgroup_lock_is_held(void)
216 return mutex_is_locked(&cgroup_mutex);
218 #endif /* #else #ifdef CONFIG_PROVE_LOCKING */
220 EXPORT_SYMBOL_GPL(cgroup_lock_is_held);
222 /* convenient tests for these bits */
223 inline int cgroup_is_removed(const struct cgroup *cgrp)
225 return test_bit(CGRP_REMOVED, &cgrp->flags);
228 /* bits in struct cgroupfs_root flags field */
230 ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
233 static int cgroup_is_releasable(const struct cgroup *cgrp)
236 (1 << CGRP_RELEASABLE) |
237 (1 << CGRP_NOTIFY_ON_RELEASE);
238 return (cgrp->flags & bits) == bits;
241 static int notify_on_release(const struct cgroup *cgrp)
243 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
246 static int clone_children(const struct cgroup *cgrp)
248 return test_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
252 * for_each_subsys() allows you to iterate on each subsystem attached to
253 * an active hierarchy
255 #define for_each_subsys(_root, _ss) \
256 list_for_each_entry(_ss, &_root->subsys_list, sibling)
258 /* for_each_active_root() allows you to iterate across the active hierarchies */
259 #define for_each_active_root(_root) \
260 list_for_each_entry(_root, &roots, root_list)
262 /* the list of cgroups eligible for automatic release. Protected by
263 * release_list_lock */
264 static LIST_HEAD(release_list);
265 static DEFINE_SPINLOCK(release_list_lock);
266 static void cgroup_release_agent(struct work_struct *work);
267 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
268 static void check_for_release(struct cgroup *cgrp);
270 /* Link structure for associating css_set objects with cgroups */
271 struct cg_cgroup_link {
273 * List running through cg_cgroup_links associated with a
274 * cgroup, anchored on cgroup->css_sets
276 struct list_head cgrp_link_list;
279 * List running through cg_cgroup_links pointing at a
280 * single css_set object, anchored on css_set->cg_links
282 struct list_head cg_link_list;
286 /* The default css_set - used by init and its children prior to any
287 * hierarchies being mounted. It contains a pointer to the root state
288 * for each subsystem. Also used to anchor the list of css_sets. Not
289 * reference-counted, to improve performance when child cgroups
290 * haven't been created.
293 static struct css_set init_css_set;
294 static struct cg_cgroup_link init_css_set_link;
296 static int cgroup_init_idr(struct cgroup_subsys *ss,
297 struct cgroup_subsys_state *css);
299 /* css_set_lock protects the list of css_set objects, and the
300 * chain of tasks off each css_set. Nests outside task->alloc_lock
301 * due to cgroup_iter_start() */
302 static DEFINE_RWLOCK(css_set_lock);
303 static int css_set_count;
306 * hash table for cgroup groups. This improves the performance to find
307 * an existing css_set. This hash doesn't (currently) take into
308 * account cgroups in empty hierarchies.
310 #define CSS_SET_HASH_BITS 7
311 #define CSS_SET_TABLE_SIZE (1 << CSS_SET_HASH_BITS)
312 static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
314 static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
318 unsigned long tmp = 0UL;
320 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
321 tmp += (unsigned long)css[i];
322 tmp = (tmp >> 16) ^ tmp;
324 index = hash_long(tmp, CSS_SET_HASH_BITS);
326 return &css_set_table[index];
329 static void free_css_set_rcu(struct rcu_head *obj)
331 struct css_set *cg = container_of(obj, struct css_set, rcu_head);
335 /* We don't maintain the lists running through each css_set to its
336 * task until after the first call to cgroup_iter_start(). This
337 * reduces the fork()/exit() overhead for people who have cgroups
338 * compiled into their kernel but not actually in use */
339 static int use_task_css_set_links __read_mostly;
341 static void __put_css_set(struct css_set *cg, int taskexit)
343 struct cg_cgroup_link *link;
344 struct cg_cgroup_link *saved_link;
346 * Ensure that the refcount doesn't hit zero while any readers
347 * can see it. Similar to atomic_dec_and_lock(), but for an
350 if (atomic_add_unless(&cg->refcount, -1, 1))
352 write_lock(&css_set_lock);
353 if (!atomic_dec_and_test(&cg->refcount)) {
354 write_unlock(&css_set_lock);
358 /* This css_set is dead. unlink it and release cgroup refcounts */
359 hlist_del(&cg->hlist);
362 list_for_each_entry_safe(link, saved_link, &cg->cg_links,
364 struct cgroup *cgrp = link->cgrp;
365 list_del(&link->cg_link_list);
366 list_del(&link->cgrp_link_list);
367 if (atomic_dec_and_test(&cgrp->count) &&
368 notify_on_release(cgrp)) {
370 set_bit(CGRP_RELEASABLE, &cgrp->flags);
371 check_for_release(cgrp);
377 write_unlock(&css_set_lock);
378 call_rcu(&cg->rcu_head, free_css_set_rcu);
382 * refcounted get/put for css_set objects
384 static inline void get_css_set(struct css_set *cg)
386 atomic_inc(&cg->refcount);
389 static inline void put_css_set(struct css_set *cg)
391 __put_css_set(cg, 0);
394 static inline void put_css_set_taskexit(struct css_set *cg)
396 __put_css_set(cg, 1);
400 * compare_css_sets - helper function for find_existing_css_set().
401 * @cg: candidate css_set being tested
402 * @old_cg: existing css_set for a task
403 * @new_cgrp: cgroup that's being entered by the task
404 * @template: desired set of css pointers in css_set (pre-calculated)
406 * Returns true if "cg" matches "old_cg" except for the hierarchy
407 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
409 static bool compare_css_sets(struct css_set *cg,
410 struct css_set *old_cg,
411 struct cgroup *new_cgrp,
412 struct cgroup_subsys_state *template[])
414 struct list_head *l1, *l2;
416 if (memcmp(template, cg->subsys, sizeof(cg->subsys))) {
417 /* Not all subsystems matched */
422 * Compare cgroup pointers in order to distinguish between
423 * different cgroups in heirarchies with no subsystems. We
424 * could get by with just this check alone (and skip the
425 * memcmp above) but on most setups the memcmp check will
426 * avoid the need for this more expensive check on almost all
431 l2 = &old_cg->cg_links;
433 struct cg_cgroup_link *cgl1, *cgl2;
434 struct cgroup *cg1, *cg2;
438 /* See if we reached the end - both lists are equal length. */
439 if (l1 == &cg->cg_links) {
440 BUG_ON(l2 != &old_cg->cg_links);
443 BUG_ON(l2 == &old_cg->cg_links);
445 /* Locate the cgroups associated with these links. */
446 cgl1 = list_entry(l1, struct cg_cgroup_link, cg_link_list);
447 cgl2 = list_entry(l2, struct cg_cgroup_link, cg_link_list);
450 /* Hierarchies should be linked in the same order. */
451 BUG_ON(cg1->root != cg2->root);
454 * If this hierarchy is the hierarchy of the cgroup
455 * that's changing, then we need to check that this
456 * css_set points to the new cgroup; if it's any other
457 * hierarchy, then this css_set should point to the
458 * same cgroup as the old css_set.
460 if (cg1->root == new_cgrp->root) {
472 * find_existing_css_set() is a helper for
473 * find_css_set(), and checks to see whether an existing
474 * css_set is suitable.
476 * oldcg: the cgroup group that we're using before the cgroup
479 * cgrp: the cgroup that we're moving into
481 * template: location in which to build the desired set of subsystem
482 * state objects for the new cgroup group
484 static struct css_set *find_existing_css_set(
485 struct css_set *oldcg,
487 struct cgroup_subsys_state *template[])
490 struct cgroupfs_root *root = cgrp->root;
491 struct hlist_head *hhead;
492 struct hlist_node *node;
496 * Build the set of subsystem state objects that we want to see in the
497 * new css_set. while subsystems can change globally, the entries here
498 * won't change, so no need for locking.
500 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
501 if (root->subsys_bits & (1UL << i)) {
502 /* Subsystem is in this hierarchy. So we want
503 * the subsystem state from the new
505 template[i] = cgrp->subsys[i];
507 /* Subsystem is not in this hierarchy, so we
508 * don't want to change the subsystem state */
509 template[i] = oldcg->subsys[i];
513 hhead = css_set_hash(template);
514 hlist_for_each_entry(cg, node, hhead, hlist) {
515 if (!compare_css_sets(cg, oldcg, cgrp, template))
518 /* This css_set matches what we need */
522 /* No existing cgroup group matched */
526 static void free_cg_links(struct list_head *tmp)
528 struct cg_cgroup_link *link;
529 struct cg_cgroup_link *saved_link;
531 list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
532 list_del(&link->cgrp_link_list);
538 * allocate_cg_links() allocates "count" cg_cgroup_link structures
539 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
540 * success or a negative error
542 static int allocate_cg_links(int count, struct list_head *tmp)
544 struct cg_cgroup_link *link;
547 for (i = 0; i < count; i++) {
548 link = kmalloc(sizeof(*link), GFP_KERNEL);
553 list_add(&link->cgrp_link_list, tmp);
559 * link_css_set - a helper function to link a css_set to a cgroup
560 * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
561 * @cg: the css_set to be linked
562 * @cgrp: the destination cgroup
564 static void link_css_set(struct list_head *tmp_cg_links,
565 struct css_set *cg, struct cgroup *cgrp)
567 struct cg_cgroup_link *link;
569 BUG_ON(list_empty(tmp_cg_links));
570 link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
574 atomic_inc(&cgrp->count);
575 list_move(&link->cgrp_link_list, &cgrp->css_sets);
577 * Always add links to the tail of the list so that the list
578 * is sorted by order of hierarchy creation
580 list_add_tail(&link->cg_link_list, &cg->cg_links);
584 * find_css_set() takes an existing cgroup group and a
585 * cgroup object, and returns a css_set object that's
586 * equivalent to the old group, but with the given cgroup
587 * substituted into the appropriate hierarchy. Must be called with
590 static struct css_set *find_css_set(
591 struct css_set *oldcg, struct cgroup *cgrp)
594 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
596 struct list_head tmp_cg_links;
598 struct hlist_head *hhead;
599 struct cg_cgroup_link *link;
601 /* First see if we already have a cgroup group that matches
603 read_lock(&css_set_lock);
604 res = find_existing_css_set(oldcg, cgrp, template);
607 read_unlock(&css_set_lock);
612 res = kmalloc(sizeof(*res), GFP_KERNEL);
616 /* Allocate all the cg_cgroup_link objects that we'll need */
617 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
622 atomic_set(&res->refcount, 1);
623 INIT_LIST_HEAD(&res->cg_links);
624 INIT_LIST_HEAD(&res->tasks);
625 INIT_HLIST_NODE(&res->hlist);
627 /* Copy the set of subsystem state objects generated in
628 * find_existing_css_set() */
629 memcpy(res->subsys, template, sizeof(res->subsys));
631 write_lock(&css_set_lock);
632 /* Add reference counts and links from the new css_set. */
633 list_for_each_entry(link, &oldcg->cg_links, cg_link_list) {
634 struct cgroup *c = link->cgrp;
635 if (c->root == cgrp->root)
637 link_css_set(&tmp_cg_links, res, c);
640 BUG_ON(!list_empty(&tmp_cg_links));
644 /* Add this cgroup group to the hash table */
645 hhead = css_set_hash(res->subsys);
646 hlist_add_head(&res->hlist, hhead);
648 write_unlock(&css_set_lock);
654 * Return the cgroup for "task" from the given hierarchy. Must be
655 * called with cgroup_mutex held.
657 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
658 struct cgroupfs_root *root)
661 struct cgroup *res = NULL;
663 BUG_ON(!mutex_is_locked(&cgroup_mutex));
664 read_lock(&css_set_lock);
666 * No need to lock the task - since we hold cgroup_mutex the
667 * task can't change groups, so the only thing that can happen
668 * is that it exits and its css is set back to init_css_set.
671 if (css == &init_css_set) {
672 res = &root->top_cgroup;
674 struct cg_cgroup_link *link;
675 list_for_each_entry(link, &css->cg_links, cg_link_list) {
676 struct cgroup *c = link->cgrp;
677 if (c->root == root) {
683 read_unlock(&css_set_lock);
689 * There is one global cgroup mutex. We also require taking
690 * task_lock() when dereferencing a task's cgroup subsys pointers.
691 * See "The task_lock() exception", at the end of this comment.
693 * A task must hold cgroup_mutex to modify cgroups.
695 * Any task can increment and decrement the count field without lock.
696 * So in general, code holding cgroup_mutex can't rely on the count
697 * field not changing. However, if the count goes to zero, then only
698 * cgroup_attach_task() can increment it again. Because a count of zero
699 * means that no tasks are currently attached, therefore there is no
700 * way a task attached to that cgroup can fork (the other way to
701 * increment the count). So code holding cgroup_mutex can safely
702 * assume that if the count is zero, it will stay zero. Similarly, if
703 * a task holds cgroup_mutex on a cgroup with zero count, it
704 * knows that the cgroup won't be removed, as cgroup_rmdir()
707 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
708 * (usually) take cgroup_mutex. These are the two most performance
709 * critical pieces of code here. The exception occurs on cgroup_exit(),
710 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
711 * is taken, and if the cgroup count is zero, a usermode call made
712 * to the release agent with the name of the cgroup (path relative to
713 * the root of cgroup file system) as the argument.
715 * A cgroup can only be deleted if both its 'count' of using tasks
716 * is zero, and its list of 'children' cgroups is empty. Since all
717 * tasks in the system use _some_ cgroup, and since there is always at
718 * least one task in the system (init, pid == 1), therefore, top_cgroup
719 * always has either children cgroups and/or using tasks. So we don't
720 * need a special hack to ensure that top_cgroup cannot be deleted.
722 * The task_lock() exception
724 * The need for this exception arises from the action of
725 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
726 * another. It does so using cgroup_mutex, however there are
727 * several performance critical places that need to reference
728 * task->cgroup without the expense of grabbing a system global
729 * mutex. Therefore except as noted below, when dereferencing or, as
730 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
731 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
732 * the task_struct routinely used for such matters.
734 * P.S. One more locking exception. RCU is used to guard the
735 * update of a tasks cgroup pointer by cgroup_attach_task()
739 * cgroup_lock - lock out any changes to cgroup structures
742 void cgroup_lock(void)
744 mutex_lock(&cgroup_mutex);
746 EXPORT_SYMBOL_GPL(cgroup_lock);
749 * cgroup_unlock - release lock on cgroup changes
751 * Undo the lock taken in a previous cgroup_lock() call.
753 void cgroup_unlock(void)
755 mutex_unlock(&cgroup_mutex);
757 EXPORT_SYMBOL_GPL(cgroup_unlock);
760 * A couple of forward declarations required, due to cyclic reference loop:
761 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
762 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
766 static struct dentry *cgroup_lookup(struct inode *dir,
767 struct dentry *dentry, struct nameidata *nd);
768 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
769 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
770 static int cgroup_populate_dir(struct cgroup *cgrp);
771 static const struct inode_operations cgroup_dir_inode_operations;
772 static const struct file_operations proc_cgroupstats_operations;
774 static struct backing_dev_info cgroup_backing_dev_info = {
776 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
779 static int alloc_css_id(struct cgroup_subsys *ss,
780 struct cgroup *parent, struct cgroup *child);
782 static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
784 struct inode *inode = new_inode(sb);
787 inode->i_ino = get_next_ino();
788 inode->i_mode = mode;
789 inode->i_uid = current_fsuid();
790 inode->i_gid = current_fsgid();
791 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
792 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
798 * Call subsys's pre_destroy handler.
799 * This is called before css refcnt check.
801 static int cgroup_call_pre_destroy(struct cgroup *cgrp)
803 struct cgroup_subsys *ss;
806 for_each_subsys(cgrp->root, ss)
807 if (ss->pre_destroy) {
808 ret = ss->pre_destroy(ss, cgrp);
816 static void free_cgroup_rcu(struct rcu_head *obj)
818 struct cgroup *cgrp = container_of(obj, struct cgroup, rcu_head);
823 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
825 /* is dentry a directory ? if so, kfree() associated cgroup */
826 if (S_ISDIR(inode->i_mode)) {
827 struct cgroup *cgrp = dentry->d_fsdata;
828 struct cgroup_subsys *ss;
829 BUG_ON(!(cgroup_is_removed(cgrp)));
830 /* It's possible for external users to be holding css
831 * reference counts on a cgroup; css_put() needs to
832 * be able to access the cgroup after decrementing
833 * the reference count in order to know if it needs to
834 * queue the cgroup to be handled by the release
838 mutex_lock(&cgroup_mutex);
840 * Release the subsystem state objects.
842 for_each_subsys(cgrp->root, ss)
843 ss->destroy(ss, cgrp);
845 cgrp->root->number_of_cgroups--;
846 mutex_unlock(&cgroup_mutex);
849 * Drop the active superblock reference that we took when we
852 deactivate_super(cgrp->root->sb);
855 * if we're getting rid of the cgroup, refcount should ensure
856 * that there are no pidlists left.
858 BUG_ON(!list_empty(&cgrp->pidlists));
860 call_rcu(&cgrp->rcu_head, free_cgroup_rcu);
865 static void remove_dir(struct dentry *d)
867 struct dentry *parent = dget(d->d_parent);
870 simple_rmdir(parent->d_inode, d);
874 static void cgroup_clear_directory(struct dentry *dentry)
876 struct list_head *node;
878 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
879 spin_lock(&dcache_lock);
880 spin_lock(&dentry->d_lock);
881 node = dentry->d_subdirs.next;
882 while (node != &dentry->d_subdirs) {
883 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
885 spin_lock_nested(&d->d_lock, DENTRY_D_LOCK_NESTED);
888 /* This should never be called on a cgroup
889 * directory with child cgroups */
890 BUG_ON(d->d_inode->i_mode & S_IFDIR);
891 dget_locked_dlock(d);
892 spin_unlock(&d->d_lock);
893 spin_unlock(&dentry->d_lock);
894 spin_unlock(&dcache_lock);
896 simple_unlink(dentry->d_inode, d);
898 spin_lock(&dcache_lock);
899 spin_lock(&dentry->d_lock);
901 spin_unlock(&d->d_lock);
902 node = dentry->d_subdirs.next;
904 spin_unlock(&dentry->d_lock);
905 spin_unlock(&dcache_lock);
909 * NOTE : the dentry must have been dget()'ed
911 static void cgroup_d_remove_dir(struct dentry *dentry)
913 struct dentry *parent;
915 cgroup_clear_directory(dentry);
917 spin_lock(&dcache_lock);
918 parent = dentry->d_parent;
919 spin_lock(&parent->d_lock);
920 spin_lock(&dentry->d_lock);
921 list_del_init(&dentry->d_u.d_child);
922 spin_unlock(&dentry->d_lock);
923 spin_unlock(&parent->d_lock);
924 spin_unlock(&dcache_lock);
929 * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
930 * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
931 * reference to css->refcnt. In general, this refcnt is expected to goes down
934 * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
936 DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
938 static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
940 if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
941 wake_up_all(&cgroup_rmdir_waitq);
944 void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
949 void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
951 cgroup_wakeup_rmdir_waiter(css->cgroup);
956 * Call with cgroup_mutex held. Drops reference counts on modules, including
957 * any duplicate ones that parse_cgroupfs_options took. If this function
958 * returns an error, no reference counts are touched.
960 static int rebind_subsystems(struct cgroupfs_root *root,
961 unsigned long final_bits)
963 unsigned long added_bits, removed_bits;
964 struct cgroup *cgrp = &root->top_cgroup;
967 BUG_ON(!mutex_is_locked(&cgroup_mutex));
969 removed_bits = root->actual_subsys_bits & ~final_bits;
970 added_bits = final_bits & ~root->actual_subsys_bits;
971 /* Check that any added subsystems are currently free */
972 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
973 unsigned long bit = 1UL << i;
974 struct cgroup_subsys *ss = subsys[i];
975 if (!(bit & added_bits))
978 * Nobody should tell us to do a subsys that doesn't exist:
979 * parse_cgroupfs_options should catch that case and refcounts
980 * ensure that subsystems won't disappear once selected.
983 if (ss->root != &rootnode) {
984 /* Subsystem isn't free */
989 /* Currently we don't handle adding/removing subsystems when
990 * any child cgroups exist. This is theoretically supportable
991 * but involves complex error handling, so it's being left until
993 if (root->number_of_cgroups > 1)
996 /* Process each subsystem */
997 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
998 struct cgroup_subsys *ss = subsys[i];
999 unsigned long bit = 1UL << i;
1000 if (bit & added_bits) {
1001 /* We're binding this subsystem to this hierarchy */
1003 BUG_ON(cgrp->subsys[i]);
1004 BUG_ON(!dummytop->subsys[i]);
1005 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
1006 mutex_lock(&ss->hierarchy_mutex);
1007 cgrp->subsys[i] = dummytop->subsys[i];
1008 cgrp->subsys[i]->cgroup = cgrp;
1009 list_move(&ss->sibling, &root->subsys_list);
1013 mutex_unlock(&ss->hierarchy_mutex);
1014 /* refcount was already taken, and we're keeping it */
1015 } else if (bit & removed_bits) {
1016 /* We're removing this subsystem */
1018 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
1019 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
1020 mutex_lock(&ss->hierarchy_mutex);
1022 ss->bind(ss, dummytop);
1023 dummytop->subsys[i]->cgroup = dummytop;
1024 cgrp->subsys[i] = NULL;
1025 subsys[i]->root = &rootnode;
1026 list_move(&ss->sibling, &rootnode.subsys_list);
1027 mutex_unlock(&ss->hierarchy_mutex);
1028 /* subsystem is now free - drop reference on module */
1029 module_put(ss->module);
1030 } else if (bit & final_bits) {
1031 /* Subsystem state should already exist */
1033 BUG_ON(!cgrp->subsys[i]);
1035 * a refcount was taken, but we already had one, so
1036 * drop the extra reference.
1038 module_put(ss->module);
1039 #ifdef CONFIG_MODULE_UNLOAD
1040 BUG_ON(ss->module && !module_refcount(ss->module));
1043 /* Subsystem state shouldn't exist */
1044 BUG_ON(cgrp->subsys[i]);
1047 root->subsys_bits = root->actual_subsys_bits = final_bits;
1053 static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
1055 struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
1056 struct cgroup_subsys *ss;
1058 mutex_lock(&cgroup_mutex);
1059 for_each_subsys(root, ss)
1060 seq_printf(seq, ",%s", ss->name);
1061 if (test_bit(ROOT_NOPREFIX, &root->flags))
1062 seq_puts(seq, ",noprefix");
1063 if (strlen(root->release_agent_path))
1064 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1065 if (clone_children(&root->top_cgroup))
1066 seq_puts(seq, ",clone_children");
1067 if (strlen(root->name))
1068 seq_printf(seq, ",name=%s", root->name);
1069 mutex_unlock(&cgroup_mutex);
1073 struct cgroup_sb_opts {
1074 unsigned long subsys_bits;
1075 unsigned long flags;
1076 char *release_agent;
1077 bool clone_children;
1079 /* User explicitly requested empty subsystem */
1082 struct cgroupfs_root *new_root;
1087 * Convert a hierarchy specifier into a bitmask of subsystems and flags. Call
1088 * with cgroup_mutex held to protect the subsys[] array. This function takes
1089 * refcounts on subsystems to be used, unless it returns error, in which case
1090 * no refcounts are taken.
1092 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1094 char *token, *o = data;
1095 bool all_ss = false, one_ss = false;
1096 unsigned long mask = (unsigned long)-1;
1098 bool module_pin_failed = false;
1100 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1102 #ifdef CONFIG_CPUSETS
1103 mask = ~(1UL << cpuset_subsys_id);
1106 memset(opts, 0, sizeof(*opts));
1108 while ((token = strsep(&o, ",")) != NULL) {
1111 if (!strcmp(token, "none")) {
1112 /* Explicitly have no subsystems */
1116 if (!strcmp(token, "all")) {
1117 /* Mutually exclusive option 'all' + subsystem name */
1123 if (!strcmp(token, "noprefix")) {
1124 set_bit(ROOT_NOPREFIX, &opts->flags);
1127 if (!strcmp(token, "clone_children")) {
1128 opts->clone_children = true;
1131 if (!strncmp(token, "release_agent=", 14)) {
1132 /* Specifying two release agents is forbidden */
1133 if (opts->release_agent)
1135 opts->release_agent =
1136 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1137 if (!opts->release_agent)
1141 if (!strncmp(token, "name=", 5)) {
1142 const char *name = token + 5;
1143 /* Can't specify an empty name */
1146 /* Must match [\w.-]+ */
1147 for (i = 0; i < strlen(name); i++) {
1151 if ((c == '.') || (c == '-') || (c == '_'))
1155 /* Specifying two names is forbidden */
1158 opts->name = kstrndup(name,
1159 MAX_CGROUP_ROOT_NAMELEN - 1,
1167 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1168 struct cgroup_subsys *ss = subsys[i];
1171 if (strcmp(token, ss->name))
1176 /* Mutually exclusive option 'all' + subsystem name */
1179 set_bit(i, &opts->subsys_bits);
1184 if (i == CGROUP_SUBSYS_COUNT)
1189 * If the 'all' option was specified select all the subsystems,
1190 * otherwise 'all, 'none' and a subsystem name options were not
1191 * specified, let's default to 'all'
1193 if (all_ss || (!all_ss && !one_ss && !opts->none)) {
1194 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1195 struct cgroup_subsys *ss = subsys[i];
1200 set_bit(i, &opts->subsys_bits);
1204 /* Consistency checks */
1207 * Option noprefix was introduced just for backward compatibility
1208 * with the old cpuset, so we allow noprefix only if mounting just
1209 * the cpuset subsystem.
1211 if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
1212 (opts->subsys_bits & mask))
1216 /* Can't specify "none" and some subsystems */
1217 if (opts->subsys_bits && opts->none)
1221 * We either have to specify by name or by subsystems. (So all
1222 * empty hierarchies must have a name).
1224 if (!opts->subsys_bits && !opts->name)
1228 * Grab references on all the modules we'll need, so the subsystems
1229 * don't dance around before rebind_subsystems attaches them. This may
1230 * take duplicate reference counts on a subsystem that's already used,
1231 * but rebind_subsystems handles this case.
1233 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1234 unsigned long bit = 1UL << i;
1236 if (!(bit & opts->subsys_bits))
1238 if (!try_module_get(subsys[i]->module)) {
1239 module_pin_failed = true;
1243 if (module_pin_failed) {
1245 * oops, one of the modules was going away. this means that we
1246 * raced with a module_delete call, and to the user this is
1247 * essentially a "subsystem doesn't exist" case.
1249 for (i--; i >= CGROUP_BUILTIN_SUBSYS_COUNT; i--) {
1250 /* drop refcounts only on the ones we took */
1251 unsigned long bit = 1UL << i;
1253 if (!(bit & opts->subsys_bits))
1255 module_put(subsys[i]->module);
1263 static void drop_parsed_module_refcounts(unsigned long subsys_bits)
1266 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1267 unsigned long bit = 1UL << i;
1269 if (!(bit & subsys_bits))
1271 module_put(subsys[i]->module);
1275 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1278 struct cgroupfs_root *root = sb->s_fs_info;
1279 struct cgroup *cgrp = &root->top_cgroup;
1280 struct cgroup_sb_opts opts;
1282 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1283 mutex_lock(&cgroup_mutex);
1285 /* See what subsystems are wanted */
1286 ret = parse_cgroupfs_options(data, &opts);
1290 /* Don't allow flags or name to change at remount */
1291 if (opts.flags != root->flags ||
1292 (opts.name && strcmp(opts.name, root->name))) {
1294 drop_parsed_module_refcounts(opts.subsys_bits);
1298 ret = rebind_subsystems(root, opts.subsys_bits);
1300 drop_parsed_module_refcounts(opts.subsys_bits);
1304 /* (re)populate subsystem files */
1305 cgroup_populate_dir(cgrp);
1307 if (opts.release_agent)
1308 strcpy(root->release_agent_path, opts.release_agent);
1310 kfree(opts.release_agent);
1312 mutex_unlock(&cgroup_mutex);
1313 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1317 static const struct super_operations cgroup_ops = {
1318 .statfs = simple_statfs,
1319 .drop_inode = generic_delete_inode,
1320 .show_options = cgroup_show_options,
1321 .remount_fs = cgroup_remount,
1324 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1326 INIT_LIST_HEAD(&cgrp->sibling);
1327 INIT_LIST_HEAD(&cgrp->children);
1328 INIT_LIST_HEAD(&cgrp->css_sets);
1329 INIT_LIST_HEAD(&cgrp->release_list);
1330 INIT_LIST_HEAD(&cgrp->pidlists);
1331 mutex_init(&cgrp->pidlist_mutex);
1332 INIT_LIST_HEAD(&cgrp->event_list);
1333 spin_lock_init(&cgrp->event_list_lock);
1336 static void init_cgroup_root(struct cgroupfs_root *root)
1338 struct cgroup *cgrp = &root->top_cgroup;
1339 INIT_LIST_HEAD(&root->subsys_list);
1340 INIT_LIST_HEAD(&root->root_list);
1341 root->number_of_cgroups = 1;
1343 cgrp->top_cgroup = cgrp;
1344 init_cgroup_housekeeping(cgrp);
1347 static bool init_root_id(struct cgroupfs_root *root)
1352 if (!ida_pre_get(&hierarchy_ida, GFP_KERNEL))
1354 spin_lock(&hierarchy_id_lock);
1355 /* Try to allocate the next unused ID */
1356 ret = ida_get_new_above(&hierarchy_ida, next_hierarchy_id,
1357 &root->hierarchy_id);
1359 /* Try again starting from 0 */
1360 ret = ida_get_new(&hierarchy_ida, &root->hierarchy_id);
1362 next_hierarchy_id = root->hierarchy_id + 1;
1363 } else if (ret != -EAGAIN) {
1364 /* Can only get here if the 31-bit IDR is full ... */
1367 spin_unlock(&hierarchy_id_lock);
1372 static int cgroup_test_super(struct super_block *sb, void *data)
1374 struct cgroup_sb_opts *opts = data;
1375 struct cgroupfs_root *root = sb->s_fs_info;
1377 /* If we asked for a name then it must match */
1378 if (opts->name && strcmp(opts->name, root->name))
1382 * If we asked for subsystems (or explicitly for no
1383 * subsystems) then they must match
1385 if ((opts->subsys_bits || opts->none)
1386 && (opts->subsys_bits != root->subsys_bits))
1392 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1394 struct cgroupfs_root *root;
1396 if (!opts->subsys_bits && !opts->none)
1399 root = kzalloc(sizeof(*root), GFP_KERNEL);
1401 return ERR_PTR(-ENOMEM);
1403 if (!init_root_id(root)) {
1405 return ERR_PTR(-ENOMEM);
1407 init_cgroup_root(root);
1409 root->subsys_bits = opts->subsys_bits;
1410 root->flags = opts->flags;
1411 if (opts->release_agent)
1412 strcpy(root->release_agent_path, opts->release_agent);
1414 strcpy(root->name, opts->name);
1415 if (opts->clone_children)
1416 set_bit(CGRP_CLONE_CHILDREN, &root->top_cgroup.flags);
1420 static void cgroup_drop_root(struct cgroupfs_root *root)
1425 BUG_ON(!root->hierarchy_id);
1426 spin_lock(&hierarchy_id_lock);
1427 ida_remove(&hierarchy_ida, root->hierarchy_id);
1428 spin_unlock(&hierarchy_id_lock);
1432 static int cgroup_set_super(struct super_block *sb, void *data)
1435 struct cgroup_sb_opts *opts = data;
1437 /* If we don't have a new root, we can't set up a new sb */
1438 if (!opts->new_root)
1441 BUG_ON(!opts->subsys_bits && !opts->none);
1443 ret = set_anon_super(sb, NULL);
1447 sb->s_fs_info = opts->new_root;
1448 opts->new_root->sb = sb;
1450 sb->s_blocksize = PAGE_CACHE_SIZE;
1451 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1452 sb->s_magic = CGROUP_SUPER_MAGIC;
1453 sb->s_op = &cgroup_ops;
1458 static int cgroup_get_rootdir(struct super_block *sb)
1460 struct inode *inode =
1461 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1462 struct dentry *dentry;
1467 inode->i_fop = &simple_dir_operations;
1468 inode->i_op = &cgroup_dir_inode_operations;
1469 /* directories start off with i_nlink == 2 (for "." entry) */
1471 dentry = d_alloc_root(inode);
1476 sb->s_root = dentry;
1480 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1481 int flags, const char *unused_dev_name,
1484 struct cgroup_sb_opts opts;
1485 struct cgroupfs_root *root;
1487 struct super_block *sb;
1488 struct cgroupfs_root *new_root;
1490 /* First find the desired set of subsystems */
1491 mutex_lock(&cgroup_mutex);
1492 ret = parse_cgroupfs_options(data, &opts);
1493 mutex_unlock(&cgroup_mutex);
1498 * Allocate a new cgroup root. We may not need it if we're
1499 * reusing an existing hierarchy.
1501 new_root = cgroup_root_from_opts(&opts);
1502 if (IS_ERR(new_root)) {
1503 ret = PTR_ERR(new_root);
1506 opts.new_root = new_root;
1508 /* Locate an existing or new sb for this hierarchy */
1509 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
1512 cgroup_drop_root(opts.new_root);
1516 root = sb->s_fs_info;
1518 if (root == opts.new_root) {
1519 /* We used the new root structure, so this is a new hierarchy */
1520 struct list_head tmp_cg_links;
1521 struct cgroup *root_cgrp = &root->top_cgroup;
1522 struct inode *inode;
1523 struct cgroupfs_root *existing_root;
1526 BUG_ON(sb->s_root != NULL);
1528 ret = cgroup_get_rootdir(sb);
1530 goto drop_new_super;
1531 inode = sb->s_root->d_inode;
1533 mutex_lock(&inode->i_mutex);
1534 mutex_lock(&cgroup_mutex);
1536 if (strlen(root->name)) {
1537 /* Check for name clashes with existing mounts */
1538 for_each_active_root(existing_root) {
1539 if (!strcmp(existing_root->name, root->name)) {
1541 mutex_unlock(&cgroup_mutex);
1542 mutex_unlock(&inode->i_mutex);
1543 goto drop_new_super;
1549 * We're accessing css_set_count without locking
1550 * css_set_lock here, but that's OK - it can only be
1551 * increased by someone holding cgroup_lock, and
1552 * that's us. The worst that can happen is that we
1553 * have some link structures left over
1555 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1557 mutex_unlock(&cgroup_mutex);
1558 mutex_unlock(&inode->i_mutex);
1559 goto drop_new_super;
1562 ret = rebind_subsystems(root, root->subsys_bits);
1563 if (ret == -EBUSY) {
1564 mutex_unlock(&cgroup_mutex);
1565 mutex_unlock(&inode->i_mutex);
1566 free_cg_links(&tmp_cg_links);
1567 goto drop_new_super;
1570 * There must be no failure case after here, since rebinding
1571 * takes care of subsystems' refcounts, which are explicitly
1572 * dropped in the failure exit path.
1575 /* EBUSY should be the only error here */
1578 list_add(&root->root_list, &roots);
1581 sb->s_root->d_fsdata = root_cgrp;
1582 root->top_cgroup.dentry = sb->s_root;
1584 /* Link the top cgroup in this hierarchy into all
1585 * the css_set objects */
1586 write_lock(&css_set_lock);
1587 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1588 struct hlist_head *hhead = &css_set_table[i];
1589 struct hlist_node *node;
1592 hlist_for_each_entry(cg, node, hhead, hlist)
1593 link_css_set(&tmp_cg_links, cg, root_cgrp);
1595 write_unlock(&css_set_lock);
1597 free_cg_links(&tmp_cg_links);
1599 BUG_ON(!list_empty(&root_cgrp->sibling));
1600 BUG_ON(!list_empty(&root_cgrp->children));
1601 BUG_ON(root->number_of_cgroups != 1);
1603 cgroup_populate_dir(root_cgrp);
1604 mutex_unlock(&cgroup_mutex);
1605 mutex_unlock(&inode->i_mutex);
1608 * We re-used an existing hierarchy - the new root (if
1609 * any) is not needed
1611 cgroup_drop_root(opts.new_root);
1612 /* no subsys rebinding, so refcounts don't change */
1613 drop_parsed_module_refcounts(opts.subsys_bits);
1616 kfree(opts.release_agent);
1618 return dget(sb->s_root);
1621 deactivate_locked_super(sb);
1623 drop_parsed_module_refcounts(opts.subsys_bits);
1625 kfree(opts.release_agent);
1627 return ERR_PTR(ret);
1630 static void cgroup_kill_sb(struct super_block *sb) {
1631 struct cgroupfs_root *root = sb->s_fs_info;
1632 struct cgroup *cgrp = &root->top_cgroup;
1634 struct cg_cgroup_link *link;
1635 struct cg_cgroup_link *saved_link;
1639 BUG_ON(root->number_of_cgroups != 1);
1640 BUG_ON(!list_empty(&cgrp->children));
1641 BUG_ON(!list_empty(&cgrp->sibling));
1643 mutex_lock(&cgroup_mutex);
1645 /* Rebind all subsystems back to the default hierarchy */
1646 ret = rebind_subsystems(root, 0);
1647 /* Shouldn't be able to fail ... */
1651 * Release all the links from css_sets to this hierarchy's
1654 write_lock(&css_set_lock);
1656 list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1658 list_del(&link->cg_link_list);
1659 list_del(&link->cgrp_link_list);
1662 write_unlock(&css_set_lock);
1664 if (!list_empty(&root->root_list)) {
1665 list_del(&root->root_list);
1669 mutex_unlock(&cgroup_mutex);
1671 kill_litter_super(sb);
1672 cgroup_drop_root(root);
1675 static struct file_system_type cgroup_fs_type = {
1677 .mount = cgroup_mount,
1678 .kill_sb = cgroup_kill_sb,
1681 static struct kobject *cgroup_kobj;
1683 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1685 return dentry->d_fsdata;
1688 static inline struct cftype *__d_cft(struct dentry *dentry)
1690 return dentry->d_fsdata;
1694 * cgroup_path - generate the path of a cgroup
1695 * @cgrp: the cgroup in question
1696 * @buf: the buffer to write the path into
1697 * @buflen: the length of the buffer
1699 * Called with cgroup_mutex held or else with an RCU-protected cgroup
1700 * reference. Writes path of cgroup into buf. Returns 0 on success,
1703 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1706 struct dentry *dentry = rcu_dereference_check(cgrp->dentry,
1707 rcu_read_lock_held() ||
1708 cgroup_lock_is_held());
1710 if (!dentry || cgrp == dummytop) {
1712 * Inactive subsystems have no dentry for their root
1719 start = buf + buflen;
1723 int len = dentry->d_name.len;
1725 if ((start -= len) < buf)
1726 return -ENAMETOOLONG;
1727 memcpy(start, dentry->d_name.name, len);
1728 cgrp = cgrp->parent;
1732 dentry = rcu_dereference_check(cgrp->dentry,
1733 rcu_read_lock_held() ||
1734 cgroup_lock_is_held());
1738 return -ENAMETOOLONG;
1741 memmove(buf, start, buf + buflen - start);
1744 EXPORT_SYMBOL_GPL(cgroup_path);
1747 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1748 * @cgrp: the cgroup the task is attaching to
1749 * @tsk: the task to be attached
1751 * Call holding cgroup_mutex. May take task_lock of
1752 * the task 'tsk' during call.
1754 int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1757 struct cgroup_subsys *ss, *failed_ss = NULL;
1758 struct cgroup *oldcgrp;
1760 struct css_set *newcg;
1761 struct cgroupfs_root *root = cgrp->root;
1763 /* Nothing to do if the task is already in that cgroup */
1764 oldcgrp = task_cgroup_from_root(tsk, root);
1765 if (cgrp == oldcgrp)
1768 for_each_subsys(root, ss) {
1769 if (ss->can_attach) {
1770 retval = ss->can_attach(ss, cgrp, tsk, false);
1773 * Remember on which subsystem the can_attach()
1774 * failed, so that we only call cancel_attach()
1775 * against the subsystems whose can_attach()
1776 * succeeded. (See below)
1789 * Locate or allocate a new css_set for this task,
1790 * based on its final set of cgroups
1792 newcg = find_css_set(cg, cgrp);
1800 if (tsk->flags & PF_EXITING) {
1806 rcu_assign_pointer(tsk->cgroups, newcg);
1809 /* Update the css_set linked lists if we're using them */
1810 write_lock(&css_set_lock);
1811 if (!list_empty(&tsk->cg_list)) {
1812 list_del(&tsk->cg_list);
1813 list_add(&tsk->cg_list, &newcg->tasks);
1815 write_unlock(&css_set_lock);
1817 for_each_subsys(root, ss) {
1819 ss->attach(ss, cgrp, oldcgrp, tsk, false);
1821 set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1826 * wake up rmdir() waiter. the rmdir should fail since the cgroup
1827 * is no longer empty.
1829 cgroup_wakeup_rmdir_waiter(cgrp);
1832 for_each_subsys(root, ss) {
1833 if (ss == failed_ss)
1835 * This subsystem was the one that failed the
1836 * can_attach() check earlier, so we don't need
1837 * to call cancel_attach() against it or any
1838 * remaining subsystems.
1841 if (ss->cancel_attach)
1842 ss->cancel_attach(ss, cgrp, tsk, false);
1849 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1850 * @from: attach to all cgroups of a given task
1851 * @tsk: the task to be attached
1853 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1855 struct cgroupfs_root *root;
1859 for_each_active_root(root) {
1860 struct cgroup *from_cg = task_cgroup_from_root(from, root);
1862 retval = cgroup_attach_task(from_cg, tsk);
1870 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
1873 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
1874 * held. May take task_lock of task
1876 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid)
1878 struct task_struct *tsk;
1879 const struct cred *cred = current_cred(), *tcred;
1884 tsk = find_task_by_vpid(pid);
1885 if (!tsk || tsk->flags & PF_EXITING) {
1890 tcred = __task_cred(tsk);
1892 cred->euid != tcred->uid &&
1893 cred->euid != tcred->suid) {
1897 get_task_struct(tsk);
1901 get_task_struct(tsk);
1904 ret = cgroup_attach_task(cgrp, tsk);
1905 put_task_struct(tsk);
1909 static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
1912 if (!cgroup_lock_live_group(cgrp))
1914 ret = attach_task_by_pid(cgrp, pid);
1920 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
1921 * @cgrp: the cgroup to be checked for liveness
1923 * On success, returns true; the lock should be later released with
1924 * cgroup_unlock(). On failure returns false with no lock held.
1926 bool cgroup_lock_live_group(struct cgroup *cgrp)
1928 mutex_lock(&cgroup_mutex);
1929 if (cgroup_is_removed(cgrp)) {
1930 mutex_unlock(&cgroup_mutex);
1935 EXPORT_SYMBOL_GPL(cgroup_lock_live_group);
1937 static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
1940 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1941 if (strlen(buffer) >= PATH_MAX)
1943 if (!cgroup_lock_live_group(cgrp))
1945 strcpy(cgrp->root->release_agent_path, buffer);
1950 static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
1951 struct seq_file *seq)
1953 if (!cgroup_lock_live_group(cgrp))
1955 seq_puts(seq, cgrp->root->release_agent_path);
1956 seq_putc(seq, '\n');
1961 /* A buffer size big enough for numbers or short strings */
1962 #define CGROUP_LOCAL_BUFFER_SIZE 64
1964 static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
1966 const char __user *userbuf,
1967 size_t nbytes, loff_t *unused_ppos)
1969 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
1975 if (nbytes >= sizeof(buffer))
1977 if (copy_from_user(buffer, userbuf, nbytes))
1980 buffer[nbytes] = 0; /* nul-terminate */
1981 if (cft->write_u64) {
1982 u64 val = simple_strtoull(strstrip(buffer), &end, 0);
1985 retval = cft->write_u64(cgrp, cft, val);
1987 s64 val = simple_strtoll(strstrip(buffer), &end, 0);
1990 retval = cft->write_s64(cgrp, cft, val);
1997 static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
1999 const char __user *userbuf,
2000 size_t nbytes, loff_t *unused_ppos)
2002 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
2004 size_t max_bytes = cft->max_write_len;
2005 char *buffer = local_buffer;
2008 max_bytes = sizeof(local_buffer) - 1;
2009 if (nbytes >= max_bytes)
2011 /* Allocate a dynamic buffer if we need one */
2012 if (nbytes >= sizeof(local_buffer)) {
2013 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
2017 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
2022 buffer[nbytes] = 0; /* nul-terminate */
2023 retval = cft->write_string(cgrp, cft, strstrip(buffer));
2027 if (buffer != local_buffer)
2032 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
2033 size_t nbytes, loff_t *ppos)
2035 struct cftype *cft = __d_cft(file->f_dentry);
2036 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2038 if (cgroup_is_removed(cgrp))
2041 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
2042 if (cft->write_u64 || cft->write_s64)
2043 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
2044 if (cft->write_string)
2045 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
2047 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
2048 return ret ? ret : nbytes;
2053 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
2055 char __user *buf, size_t nbytes,
2058 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2059 u64 val = cft->read_u64(cgrp, cft);
2060 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
2062 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2065 static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
2067 char __user *buf, size_t nbytes,
2070 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2071 s64 val = cft->read_s64(cgrp, cft);
2072 int len = sprintf(tmp, "%lld\n", (long long) val);
2074 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2077 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2078 size_t nbytes, loff_t *ppos)
2080 struct cftype *cft = __d_cft(file->f_dentry);
2081 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2083 if (cgroup_is_removed(cgrp))
2087 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2089 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2091 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2096 * seqfile ops/methods for returning structured data. Currently just
2097 * supports string->u64 maps, but can be extended in future.
2100 struct cgroup_seqfile_state {
2102 struct cgroup *cgroup;
2105 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2107 struct seq_file *sf = cb->state;
2108 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2111 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2113 struct cgroup_seqfile_state *state = m->private;
2114 struct cftype *cft = state->cft;
2115 if (cft->read_map) {
2116 struct cgroup_map_cb cb = {
2117 .fill = cgroup_map_add,
2120 return cft->read_map(state->cgroup, cft, &cb);
2122 return cft->read_seq_string(state->cgroup, cft, m);
2125 static int cgroup_seqfile_release(struct inode *inode, struct file *file)
2127 struct seq_file *seq = file->private_data;
2128 kfree(seq->private);
2129 return single_release(inode, file);
2132 static const struct file_operations cgroup_seqfile_operations = {
2134 .write = cgroup_file_write,
2135 .llseek = seq_lseek,
2136 .release = cgroup_seqfile_release,
2139 static int cgroup_file_open(struct inode *inode, struct file *file)
2144 err = generic_file_open(inode, file);
2147 cft = __d_cft(file->f_dentry);
2149 if (cft->read_map || cft->read_seq_string) {
2150 struct cgroup_seqfile_state *state =
2151 kzalloc(sizeof(*state), GFP_USER);
2155 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
2156 file->f_op = &cgroup_seqfile_operations;
2157 err = single_open(file, cgroup_seqfile_show, state);
2160 } else if (cft->open)
2161 err = cft->open(inode, file);
2168 static int cgroup_file_release(struct inode *inode, struct file *file)
2170 struct cftype *cft = __d_cft(file->f_dentry);
2172 return cft->release(inode, file);
2177 * cgroup_rename - Only allow simple rename of directories in place.
2179 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2180 struct inode *new_dir, struct dentry *new_dentry)
2182 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2184 if (new_dentry->d_inode)
2186 if (old_dir != new_dir)
2188 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2191 static const struct file_operations cgroup_file_operations = {
2192 .read = cgroup_file_read,
2193 .write = cgroup_file_write,
2194 .llseek = generic_file_llseek,
2195 .open = cgroup_file_open,
2196 .release = cgroup_file_release,
2199 static const struct inode_operations cgroup_dir_inode_operations = {
2200 .lookup = cgroup_lookup,
2201 .mkdir = cgroup_mkdir,
2202 .rmdir = cgroup_rmdir,
2203 .rename = cgroup_rename,
2207 * Check if a file is a control file
2209 static inline struct cftype *__file_cft(struct file *file)
2211 if (file->f_dentry->d_inode->i_fop != &cgroup_file_operations)
2212 return ERR_PTR(-EINVAL);
2213 return __d_cft(file->f_dentry);
2216 static int cgroup_delete_dentry(const struct dentry *dentry)
2221 static struct dentry *cgroup_lookup(struct inode *dir,
2222 struct dentry *dentry, struct nameidata *nd)
2224 static const struct dentry_operations cgroup_dentry_operations = {
2225 .d_delete = cgroup_delete_dentry,
2226 .d_iput = cgroup_diput,
2229 if (dentry->d_name.len > NAME_MAX)
2230 return ERR_PTR(-ENAMETOOLONG);
2231 dentry->d_op = &cgroup_dentry_operations;
2232 d_add(dentry, NULL);
2236 static int cgroup_create_file(struct dentry *dentry, mode_t mode,
2237 struct super_block *sb)
2239 struct inode *inode;
2243 if (dentry->d_inode)
2246 inode = cgroup_new_inode(mode, sb);
2250 if (S_ISDIR(mode)) {
2251 inode->i_op = &cgroup_dir_inode_operations;
2252 inode->i_fop = &simple_dir_operations;
2254 /* start off with i_nlink == 2 (for "." entry) */
2257 /* start with the directory inode held, so that we can
2258 * populate it without racing with another mkdir */
2259 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2260 } else if (S_ISREG(mode)) {
2262 inode->i_fop = &cgroup_file_operations;
2264 d_instantiate(dentry, inode);
2265 dget(dentry); /* Extra count - pin the dentry in core */
2270 * cgroup_create_dir - create a directory for an object.
2271 * @cgrp: the cgroup we create the directory for. It must have a valid
2272 * ->parent field. And we are going to fill its ->dentry field.
2273 * @dentry: dentry of the new cgroup
2274 * @mode: mode to set on new directory.
2276 static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
2279 struct dentry *parent;
2282 parent = cgrp->parent->dentry;
2283 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
2285 dentry->d_fsdata = cgrp;
2286 inc_nlink(parent->d_inode);
2287 rcu_assign_pointer(cgrp->dentry, dentry);
2296 * cgroup_file_mode - deduce file mode of a control file
2297 * @cft: the control file in question
2299 * returns cft->mode if ->mode is not 0
2300 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2301 * returns S_IRUGO if it has only a read handler
2302 * returns S_IWUSR if it has only a write hander
2304 static mode_t cgroup_file_mode(const struct cftype *cft)
2311 if (cft->read || cft->read_u64 || cft->read_s64 ||
2312 cft->read_map || cft->read_seq_string)
2315 if (cft->write || cft->write_u64 || cft->write_s64 ||
2316 cft->write_string || cft->trigger)
2322 int cgroup_add_file(struct cgroup *cgrp,
2323 struct cgroup_subsys *subsys,
2324 const struct cftype *cft)
2326 struct dentry *dir = cgrp->dentry;
2327 struct dentry *dentry;
2331 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2332 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
2333 strcpy(name, subsys->name);
2336 strcat(name, cft->name);
2337 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2338 dentry = lookup_one_len(name, dir, strlen(name));
2339 if (!IS_ERR(dentry)) {
2340 mode = cgroup_file_mode(cft);
2341 error = cgroup_create_file(dentry, mode | S_IFREG,
2344 dentry->d_fsdata = (void *)cft;
2347 error = PTR_ERR(dentry);
2350 EXPORT_SYMBOL_GPL(cgroup_add_file);
2352 int cgroup_add_files(struct cgroup *cgrp,
2353 struct cgroup_subsys *subsys,
2354 const struct cftype cft[],
2358 for (i = 0; i < count; i++) {
2359 err = cgroup_add_file(cgrp, subsys, &cft[i]);
2365 EXPORT_SYMBOL_GPL(cgroup_add_files);
2368 * cgroup_task_count - count the number of tasks in a cgroup.
2369 * @cgrp: the cgroup in question
2371 * Return the number of tasks in the cgroup.
2373 int cgroup_task_count(const struct cgroup *cgrp)
2376 struct cg_cgroup_link *link;
2378 read_lock(&css_set_lock);
2379 list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
2380 count += atomic_read(&link->cg->refcount);
2382 read_unlock(&css_set_lock);
2387 * Advance a list_head iterator. The iterator should be positioned at
2388 * the start of a css_set
2390 static void cgroup_advance_iter(struct cgroup *cgrp,
2391 struct cgroup_iter *it)
2393 struct list_head *l = it->cg_link;
2394 struct cg_cgroup_link *link;
2397 /* Advance to the next non-empty css_set */
2400 if (l == &cgrp->css_sets) {
2404 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
2406 } while (list_empty(&cg->tasks));
2408 it->task = cg->tasks.next;
2412 * To reduce the fork() overhead for systems that are not actually
2413 * using their cgroups capability, we don't maintain the lists running
2414 * through each css_set to its tasks until we see the list actually
2415 * used - in other words after the first call to cgroup_iter_start().
2417 * The tasklist_lock is not held here, as do_each_thread() and
2418 * while_each_thread() are protected by RCU.
2420 static void cgroup_enable_task_cg_lists(void)
2422 struct task_struct *p, *g;
2423 write_lock(&css_set_lock);
2424 use_task_css_set_links = 1;
2425 do_each_thread(g, p) {
2428 * We should check if the process is exiting, otherwise
2429 * it will race with cgroup_exit() in that the list
2430 * entry won't be deleted though the process has exited.
2432 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2433 list_add(&p->cg_list, &p->cgroups->tasks);
2435 } while_each_thread(g, p);
2436 write_unlock(&css_set_lock);
2439 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
2442 * The first time anyone tries to iterate across a cgroup,
2443 * we need to enable the list linking each css_set to its
2444 * tasks, and fix up all existing tasks.
2446 if (!use_task_css_set_links)
2447 cgroup_enable_task_cg_lists();
2449 read_lock(&css_set_lock);
2450 it->cg_link = &cgrp->css_sets;
2451 cgroup_advance_iter(cgrp, it);
2454 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
2455 struct cgroup_iter *it)
2457 struct task_struct *res;
2458 struct list_head *l = it->task;
2459 struct cg_cgroup_link *link;
2461 /* If the iterator cg is NULL, we have no tasks */
2464 res = list_entry(l, struct task_struct, cg_list);
2465 /* Advance iterator to find next entry */
2467 link = list_entry(it->cg_link, struct cg_cgroup_link, cgrp_link_list);
2468 if (l == &link->cg->tasks) {
2469 /* We reached the end of this task list - move on to
2470 * the next cg_cgroup_link */
2471 cgroup_advance_iter(cgrp, it);
2478 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
2480 read_unlock(&css_set_lock);
2483 static inline int started_after_time(struct task_struct *t1,
2484 struct timespec *time,
2485 struct task_struct *t2)
2487 int start_diff = timespec_compare(&t1->start_time, time);
2488 if (start_diff > 0) {
2490 } else if (start_diff < 0) {
2494 * Arbitrarily, if two processes started at the same
2495 * time, we'll say that the lower pointer value
2496 * started first. Note that t2 may have exited by now
2497 * so this may not be a valid pointer any longer, but
2498 * that's fine - it still serves to distinguish
2499 * between two tasks started (effectively) simultaneously.
2506 * This function is a callback from heap_insert() and is used to order
2508 * In this case we order the heap in descending task start time.
2510 static inline int started_after(void *p1, void *p2)
2512 struct task_struct *t1 = p1;
2513 struct task_struct *t2 = p2;
2514 return started_after_time(t1, &t2->start_time, t2);
2518 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
2519 * @scan: struct cgroup_scanner containing arguments for the scan
2521 * Arguments include pointers to callback functions test_task() and
2523 * Iterate through all the tasks in a cgroup, calling test_task() for each,
2524 * and if it returns true, call process_task() for it also.
2525 * The test_task pointer may be NULL, meaning always true (select all tasks).
2526 * Effectively duplicates cgroup_iter_{start,next,end}()
2527 * but does not lock css_set_lock for the call to process_task().
2528 * The struct cgroup_scanner may be embedded in any structure of the caller's
2530 * It is guaranteed that process_task() will act on every task that
2531 * is a member of the cgroup for the duration of this call. This
2532 * function may or may not call process_task() for tasks that exit
2533 * or move to a different cgroup during the call, or are forked or
2534 * move into the cgroup during the call.
2536 * Note that test_task() may be called with locks held, and may in some
2537 * situations be called multiple times for the same task, so it should
2539 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
2540 * pre-allocated and will be used for heap operations (and its "gt" member will
2541 * be overwritten), else a temporary heap will be used (allocation of which
2542 * may cause this function to fail).
2544 int cgroup_scan_tasks(struct cgroup_scanner *scan)
2547 struct cgroup_iter it;
2548 struct task_struct *p, *dropped;
2549 /* Never dereference latest_task, since it's not refcounted */
2550 struct task_struct *latest_task = NULL;
2551 struct ptr_heap tmp_heap;
2552 struct ptr_heap *heap;
2553 struct timespec latest_time = { 0, 0 };
2556 /* The caller supplied our heap and pre-allocated its memory */
2558 heap->gt = &started_after;
2560 /* We need to allocate our own heap memory */
2562 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2564 /* cannot allocate the heap */
2570 * Scan tasks in the cgroup, using the scanner's "test_task" callback
2571 * to determine which are of interest, and using the scanner's
2572 * "process_task" callback to process any of them that need an update.
2573 * Since we don't want to hold any locks during the task updates,
2574 * gather tasks to be processed in a heap structure.
2575 * The heap is sorted by descending task start time.
2576 * If the statically-sized heap fills up, we overflow tasks that
2577 * started later, and in future iterations only consider tasks that
2578 * started after the latest task in the previous pass. This
2579 * guarantees forward progress and that we don't miss any tasks.
2582 cgroup_iter_start(scan->cg, &it);
2583 while ((p = cgroup_iter_next(scan->cg, &it))) {
2585 * Only affect tasks that qualify per the caller's callback,
2586 * if he provided one
2588 if (scan->test_task && !scan->test_task(p, scan))
2591 * Only process tasks that started after the last task
2594 if (!started_after_time(p, &latest_time, latest_task))
2596 dropped = heap_insert(heap, p);
2597 if (dropped == NULL) {
2599 * The new task was inserted; the heap wasn't
2603 } else if (dropped != p) {
2605 * The new task was inserted, and pushed out a
2609 put_task_struct(dropped);
2612 * Else the new task was newer than anything already in
2613 * the heap and wasn't inserted
2616 cgroup_iter_end(scan->cg, &it);
2619 for (i = 0; i < heap->size; i++) {
2620 struct task_struct *q = heap->ptrs[i];
2622 latest_time = q->start_time;
2625 /* Process the task per the caller's callback */
2626 scan->process_task(q, scan);
2630 * If we had to process any tasks at all, scan again
2631 * in case some of them were in the middle of forking
2632 * children that didn't get processed.
2633 * Not the most efficient way to do it, but it avoids
2634 * having to take callback_mutex in the fork path
2638 if (heap == &tmp_heap)
2639 heap_free(&tmp_heap);
2644 * Stuff for reading the 'tasks'/'procs' files.
2646 * Reading this file can return large amounts of data if a cgroup has
2647 * *lots* of attached tasks. So it may need several calls to read(),
2648 * but we cannot guarantee that the information we produce is correct
2649 * unless we produce it entirely atomically.
2654 * The following two functions "fix" the issue where there are more pids
2655 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2656 * TODO: replace with a kernel-wide solution to this problem
2658 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2659 static void *pidlist_allocate(int count)
2661 if (PIDLIST_TOO_LARGE(count))
2662 return vmalloc(count * sizeof(pid_t));
2664 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2666 static void pidlist_free(void *p)
2668 if (is_vmalloc_addr(p))
2673 static void *pidlist_resize(void *p, int newcount)
2676 /* note: if new alloc fails, old p will still be valid either way */
2677 if (is_vmalloc_addr(p)) {
2678 newlist = vmalloc(newcount * sizeof(pid_t));
2681 memcpy(newlist, p, newcount * sizeof(pid_t));
2684 newlist = krealloc(p, newcount * sizeof(pid_t), GFP_KERNEL);
2690 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2691 * If the new stripped list is sufficiently smaller and there's enough memory
2692 * to allocate a new buffer, will let go of the unneeded memory. Returns the
2693 * number of unique elements.
2695 /* is the size difference enough that we should re-allocate the array? */
2696 #define PIDLIST_REALLOC_DIFFERENCE(old, new) ((old) - PAGE_SIZE >= (new))
2697 static int pidlist_uniq(pid_t **p, int length)
2704 * we presume the 0th element is unique, so i starts at 1. trivial
2705 * edge cases first; no work needs to be done for either
2707 if (length == 0 || length == 1)
2709 /* src and dest walk down the list; dest counts unique elements */
2710 for (src = 1; src < length; src++) {
2711 /* find next unique element */
2712 while (list[src] == list[src-1]) {
2717 /* dest always points to where the next unique element goes */
2718 list[dest] = list[src];
2723 * if the length difference is large enough, we want to allocate a
2724 * smaller buffer to save memory. if this fails due to out of memory,
2725 * we'll just stay with what we've got.
2727 if (PIDLIST_REALLOC_DIFFERENCE(length, dest)) {
2728 newlist = pidlist_resize(list, dest);
2735 static int cmppid(const void *a, const void *b)
2737 return *(pid_t *)a - *(pid_t *)b;
2741 * find the appropriate pidlist for our purpose (given procs vs tasks)
2742 * returns with the lock on that pidlist already held, and takes care
2743 * of the use count, or returns NULL with no locks held if we're out of
2746 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
2747 enum cgroup_filetype type)
2749 struct cgroup_pidlist *l;
2750 /* don't need task_nsproxy() if we're looking at ourself */
2751 struct pid_namespace *ns = current->nsproxy->pid_ns;
2754 * We can't drop the pidlist_mutex before taking the l->mutex in case
2755 * the last ref-holder is trying to remove l from the list at the same
2756 * time. Holding the pidlist_mutex precludes somebody taking whichever
2757 * list we find out from under us - compare release_pid_array().
2759 mutex_lock(&cgrp->pidlist_mutex);
2760 list_for_each_entry(l, &cgrp->pidlists, links) {
2761 if (l->key.type == type && l->key.ns == ns) {
2762 /* make sure l doesn't vanish out from under us */
2763 down_write(&l->mutex);
2764 mutex_unlock(&cgrp->pidlist_mutex);
2768 /* entry not found; create a new one */
2769 l = kmalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
2771 mutex_unlock(&cgrp->pidlist_mutex);
2774 init_rwsem(&l->mutex);
2775 down_write(&l->mutex);
2777 l->key.ns = get_pid_ns(ns);
2778 l->use_count = 0; /* don't increment here */
2781 list_add(&l->links, &cgrp->pidlists);
2782 mutex_unlock(&cgrp->pidlist_mutex);
2787 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
2789 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
2790 struct cgroup_pidlist **lp)
2794 int pid, n = 0; /* used for populating the array */
2795 struct cgroup_iter it;
2796 struct task_struct *tsk;
2797 struct cgroup_pidlist *l;
2800 * If cgroup gets more users after we read count, we won't have
2801 * enough space - tough. This race is indistinguishable to the
2802 * caller from the case that the additional cgroup users didn't
2803 * show up until sometime later on.
2805 length = cgroup_task_count(cgrp);
2806 array = pidlist_allocate(length);
2809 /* now, populate the array */
2810 cgroup_iter_start(cgrp, &it);
2811 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2812 if (unlikely(n == length))
2814 /* get tgid or pid for procs or tasks file respectively */
2815 if (type == CGROUP_FILE_PROCS)
2816 pid = task_tgid_vnr(tsk);
2818 pid = task_pid_vnr(tsk);
2819 if (pid > 0) /* make sure to only use valid results */
2822 cgroup_iter_end(cgrp, &it);
2824 /* now sort & (if procs) strip out duplicates */
2825 sort(array, length, sizeof(pid_t), cmppid, NULL);
2826 if (type == CGROUP_FILE_PROCS)
2827 length = pidlist_uniq(&array, length);
2828 l = cgroup_pidlist_find(cgrp, type);
2830 pidlist_free(array);
2833 /* store array, freeing old if necessary - lock already held */
2834 pidlist_free(l->list);
2838 up_write(&l->mutex);
2844 * cgroupstats_build - build and fill cgroupstats
2845 * @stats: cgroupstats to fill information into
2846 * @dentry: A dentry entry belonging to the cgroup for which stats have
2849 * Build and fill cgroupstats so that taskstats can export it to user
2852 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2855 struct cgroup *cgrp;
2856 struct cgroup_iter it;
2857 struct task_struct *tsk;
2860 * Validate dentry by checking the superblock operations,
2861 * and make sure it's a directory.
2863 if (dentry->d_sb->s_op != &cgroup_ops ||
2864 !S_ISDIR(dentry->d_inode->i_mode))
2868 cgrp = dentry->d_fsdata;
2870 cgroup_iter_start(cgrp, &it);
2871 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2872 switch (tsk->state) {
2874 stats->nr_running++;
2876 case TASK_INTERRUPTIBLE:
2877 stats->nr_sleeping++;
2879 case TASK_UNINTERRUPTIBLE:
2880 stats->nr_uninterruptible++;
2883 stats->nr_stopped++;
2886 if (delayacct_is_task_waiting_on_io(tsk))
2887 stats->nr_io_wait++;
2891 cgroup_iter_end(cgrp, &it);
2899 * seq_file methods for the tasks/procs files. The seq_file position is the
2900 * next pid to display; the seq_file iterator is a pointer to the pid
2901 * in the cgroup->l->list array.
2904 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
2907 * Initially we receive a position value that corresponds to
2908 * one more than the last pid shown (or 0 on the first call or
2909 * after a seek to the start). Use a binary-search to find the
2910 * next pid to display, if any
2912 struct cgroup_pidlist *l = s->private;
2913 int index = 0, pid = *pos;
2916 down_read(&l->mutex);
2918 int end = l->length;
2920 while (index < end) {
2921 int mid = (index + end) / 2;
2922 if (l->list[mid] == pid) {
2925 } else if (l->list[mid] <= pid)
2931 /* If we're off the end of the array, we're done */
2932 if (index >= l->length)
2934 /* Update the abstract position to be the actual pid that we found */
2935 iter = l->list + index;
2940 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
2942 struct cgroup_pidlist *l = s->private;
2946 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
2948 struct cgroup_pidlist *l = s->private;
2950 pid_t *end = l->list + l->length;
2952 * Advance to the next pid in the array. If this goes off the
2964 static int cgroup_pidlist_show(struct seq_file *s, void *v)
2966 return seq_printf(s, "%d\n", *(int *)v);
2970 * seq_operations functions for iterating on pidlists through seq_file -
2971 * independent of whether it's tasks or procs
2973 static const struct seq_operations cgroup_pidlist_seq_operations = {
2974 .start = cgroup_pidlist_start,
2975 .stop = cgroup_pidlist_stop,
2976 .next = cgroup_pidlist_next,
2977 .show = cgroup_pidlist_show,
2980 static void cgroup_release_pid_array(struct cgroup_pidlist *l)
2983 * the case where we're the last user of this particular pidlist will
2984 * have us remove it from the cgroup's list, which entails taking the
2985 * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
2986 * pidlist_mutex, we have to take pidlist_mutex first.
2988 mutex_lock(&l->owner->pidlist_mutex);
2989 down_write(&l->mutex);
2990 BUG_ON(!l->use_count);
2991 if (!--l->use_count) {
2992 /* we're the last user if refcount is 0; remove and free */
2993 list_del(&l->links);
2994 mutex_unlock(&l->owner->pidlist_mutex);
2995 pidlist_free(l->list);
2996 put_pid_ns(l->key.ns);
2997 up_write(&l->mutex);
3001 mutex_unlock(&l->owner->pidlist_mutex);
3002 up_write(&l->mutex);
3005 static int cgroup_pidlist_release(struct inode *inode, struct file *file)
3007 struct cgroup_pidlist *l;
3008 if (!(file->f_mode & FMODE_READ))
3011 * the seq_file will only be initialized if the file was opened for
3012 * reading; hence we check if it's not null only in that case.
3014 l = ((struct seq_file *)file->private_data)->private;
3015 cgroup_release_pid_array(l);
3016 return seq_release(inode, file);
3019 static const struct file_operations cgroup_pidlist_operations = {
3021 .llseek = seq_lseek,
3022 .write = cgroup_file_write,
3023 .release = cgroup_pidlist_release,
3027 * The following functions handle opens on a file that displays a pidlist
3028 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
3031 /* helper function for the two below it */
3032 static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
3034 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
3035 struct cgroup_pidlist *l;
3038 /* Nothing to do for write-only files */
3039 if (!(file->f_mode & FMODE_READ))
3042 /* have the array populated */
3043 retval = pidlist_array_load(cgrp, type, &l);
3046 /* configure file information */
3047 file->f_op = &cgroup_pidlist_operations;
3049 retval = seq_open(file, &cgroup_pidlist_seq_operations);
3051 cgroup_release_pid_array(l);
3054 ((struct seq_file *)file->private_data)->private = l;
3057 static int cgroup_tasks_open(struct inode *unused, struct file *file)
3059 return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
3061 static int cgroup_procs_open(struct inode *unused, struct file *file)
3063 return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
3066 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
3069 return notify_on_release(cgrp);
3072 static int cgroup_write_notify_on_release(struct cgroup *cgrp,
3076 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
3078 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3080 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3085 * Unregister event and free resources.
3087 * Gets called from workqueue.
3089 static void cgroup_event_remove(struct work_struct *work)
3091 struct cgroup_event *event = container_of(work, struct cgroup_event,
3093 struct cgroup *cgrp = event->cgrp;
3095 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3097 eventfd_ctx_put(event->eventfd);
3103 * Gets called on POLLHUP on eventfd when user closes it.
3105 * Called with wqh->lock held and interrupts disabled.
3107 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3108 int sync, void *key)
3110 struct cgroup_event *event = container_of(wait,
3111 struct cgroup_event, wait);
3112 struct cgroup *cgrp = event->cgrp;
3113 unsigned long flags = (unsigned long)key;
3115 if (flags & POLLHUP) {
3116 __remove_wait_queue(event->wqh, &event->wait);
3117 spin_lock(&cgrp->event_list_lock);
3118 list_del(&event->list);
3119 spin_unlock(&cgrp->event_list_lock);
3121 * We are in atomic context, but cgroup_event_remove() may
3122 * sleep, so we have to call it in workqueue.
3124 schedule_work(&event->remove);
3130 static void cgroup_event_ptable_queue_proc(struct file *file,
3131 wait_queue_head_t *wqh, poll_table *pt)
3133 struct cgroup_event *event = container_of(pt,
3134 struct cgroup_event, pt);
3137 add_wait_queue(wqh, &event->wait);
3141 * Parse input and register new cgroup event handler.
3143 * Input must be in format '<event_fd> <control_fd> <args>'.
3144 * Interpretation of args is defined by control file implementation.
3146 static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3149 struct cgroup_event *event = NULL;
3150 unsigned int efd, cfd;
3151 struct file *efile = NULL;
3152 struct file *cfile = NULL;
3156 efd = simple_strtoul(buffer, &endp, 10);
3161 cfd = simple_strtoul(buffer, &endp, 10);
3162 if ((*endp != ' ') && (*endp != '\0'))
3166 event = kzalloc(sizeof(*event), GFP_KERNEL);
3170 INIT_LIST_HEAD(&event->list);
3171 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
3172 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
3173 INIT_WORK(&event->remove, cgroup_event_remove);
3175 efile = eventfd_fget(efd);
3176 if (IS_ERR(efile)) {
3177 ret = PTR_ERR(efile);
3181 event->eventfd = eventfd_ctx_fileget(efile);
3182 if (IS_ERR(event->eventfd)) {
3183 ret = PTR_ERR(event->eventfd);
3193 /* the process need read permission on control file */
3194 ret = file_permission(cfile, MAY_READ);
3198 event->cft = __file_cft(cfile);
3199 if (IS_ERR(event->cft)) {
3200 ret = PTR_ERR(event->cft);
3204 if (!event->cft->register_event || !event->cft->unregister_event) {
3209 ret = event->cft->register_event(cgrp, event->cft,
3210 event->eventfd, buffer);
3214 if (efile->f_op->poll(efile, &event->pt) & POLLHUP) {
3215 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3221 * Events should be removed after rmdir of cgroup directory, but before
3222 * destroying subsystem state objects. Let's take reference to cgroup
3223 * directory dentry to do that.
3227 spin_lock(&cgrp->event_list_lock);
3228 list_add(&event->list, &cgrp->event_list);
3229 spin_unlock(&cgrp->event_list_lock);
3240 if (event && event->eventfd && !IS_ERR(event->eventfd))
3241 eventfd_ctx_put(event->eventfd);
3243 if (!IS_ERR_OR_NULL(efile))
3251 static u64 cgroup_clone_children_read(struct cgroup *cgrp,
3254 return clone_children(cgrp);
3257 static int cgroup_clone_children_write(struct cgroup *cgrp,
3262 set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3264 clear_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3269 * for the common functions, 'private' gives the type of file
3271 /* for hysterical raisins, we can't put this on the older files */
3272 #define CGROUP_FILE_GENERIC_PREFIX "cgroup."
3273 static struct cftype files[] = {
3276 .open = cgroup_tasks_open,
3277 .write_u64 = cgroup_tasks_write,
3278 .release = cgroup_pidlist_release,
3279 .mode = S_IRUGO | S_IWUSR,
3282 .name = CGROUP_FILE_GENERIC_PREFIX "procs",
3283 .open = cgroup_procs_open,
3284 /* .write_u64 = cgroup_procs_write, TODO */
3285 .release = cgroup_pidlist_release,
3289 .name = "notify_on_release",
3290 .read_u64 = cgroup_read_notify_on_release,
3291 .write_u64 = cgroup_write_notify_on_release,
3294 .name = CGROUP_FILE_GENERIC_PREFIX "event_control",
3295 .write_string = cgroup_write_event_control,
3299 .name = "cgroup.clone_children",
3300 .read_u64 = cgroup_clone_children_read,
3301 .write_u64 = cgroup_clone_children_write,
3305 static struct cftype cft_release_agent = {
3306 .name = "release_agent",
3307 .read_seq_string = cgroup_release_agent_show,
3308 .write_string = cgroup_release_agent_write,
3309 .max_write_len = PATH_MAX,
3312 static int cgroup_populate_dir(struct cgroup *cgrp)
3315 struct cgroup_subsys *ss;
3317 /* First clear out any existing files */
3318 cgroup_clear_directory(cgrp->dentry);
3320 err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
3324 if (cgrp == cgrp->top_cgroup) {
3325 if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
3329 for_each_subsys(cgrp->root, ss) {
3330 if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
3333 /* This cgroup is ready now */
3334 for_each_subsys(cgrp->root, ss) {
3335 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3337 * Update id->css pointer and make this css visible from
3338 * CSS ID functions. This pointer will be dereferened
3339 * from RCU-read-side without locks.
3342 rcu_assign_pointer(css->id->css, css);
3348 static void init_cgroup_css(struct cgroup_subsys_state *css,
3349 struct cgroup_subsys *ss,
3350 struct cgroup *cgrp)
3353 atomic_set(&css->refcnt, 1);
3356 if (cgrp == dummytop)
3357 set_bit(CSS_ROOT, &css->flags);
3358 BUG_ON(cgrp->subsys[ss->subsys_id]);
3359 cgrp->subsys[ss->subsys_id] = css;
3362 static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
3364 /* We need to take each hierarchy_mutex in a consistent order */
3368 * No worry about a race with rebind_subsystems that might mess up the
3369 * locking order, since both parties are under cgroup_mutex.
3371 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3372 struct cgroup_subsys *ss = subsys[i];
3375 if (ss->root == root)
3376 mutex_lock(&ss->hierarchy_mutex);
3380 static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
3384 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3385 struct cgroup_subsys *ss = subsys[i];
3388 if (ss->root == root)
3389 mutex_unlock(&ss->hierarchy_mutex);
3394 * cgroup_create - create a cgroup
3395 * @parent: cgroup that will be parent of the new cgroup
3396 * @dentry: dentry of the new cgroup
3397 * @mode: mode to set on new inode
3399 * Must be called with the mutex on the parent inode held
3401 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
3404 struct cgroup *cgrp;
3405 struct cgroupfs_root *root = parent->root;
3407 struct cgroup_subsys *ss;
3408 struct super_block *sb = root->sb;
3410 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3414 /* Grab a reference on the superblock so the hierarchy doesn't
3415 * get deleted on unmount if there are child cgroups. This
3416 * can be done outside cgroup_mutex, since the sb can't
3417 * disappear while someone has an open control file on the
3419 atomic_inc(&sb->s_active);
3421 mutex_lock(&cgroup_mutex);
3423 init_cgroup_housekeeping(cgrp);
3425 cgrp->parent = parent;
3426 cgrp->root = parent->root;
3427 cgrp->top_cgroup = parent->top_cgroup;
3429 if (notify_on_release(parent))
3430 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3432 if (clone_children(parent))
3433 set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3435 for_each_subsys(root, ss) {
3436 struct cgroup_subsys_state *css = ss->create(ss, cgrp);
3442 init_cgroup_css(css, ss, cgrp);
3444 err = alloc_css_id(ss, parent, cgrp);
3448 /* At error, ->destroy() callback has to free assigned ID. */
3449 if (clone_children(parent) && ss->post_clone)
3450 ss->post_clone(ss, cgrp);
3453 cgroup_lock_hierarchy(root);
3454 list_add(&cgrp->sibling, &cgrp->parent->children);
3455 cgroup_unlock_hierarchy(root);
3456 root->number_of_cgroups++;
3458 err = cgroup_create_dir(cgrp, dentry, mode);
3462 /* The cgroup directory was pre-locked for us */
3463 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
3465 err = cgroup_populate_dir(cgrp);
3466 /* If err < 0, we have a half-filled directory - oh well ;) */
3468 mutex_unlock(&cgroup_mutex);
3469 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
3475 cgroup_lock_hierarchy(root);
3476 list_del(&cgrp->sibling);
3477 cgroup_unlock_hierarchy(root);
3478 root->number_of_cgroups--;
3482 for_each_subsys(root, ss) {
3483 if (cgrp->subsys[ss->subsys_id])
3484 ss->destroy(ss, cgrp);
3487 mutex_unlock(&cgroup_mutex);
3489 /* Release the reference count that we took on the superblock */
3490 deactivate_super(sb);
3496 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
3498 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
3500 /* the vfs holds inode->i_mutex already */
3501 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
3504 static int cgroup_has_css_refs(struct cgroup *cgrp)
3506 /* Check the reference count on each subsystem. Since we
3507 * already established that there are no tasks in the
3508 * cgroup, if the css refcount is also 1, then there should
3509 * be no outstanding references, so the subsystem is safe to
3510 * destroy. We scan across all subsystems rather than using
3511 * the per-hierarchy linked list of mounted subsystems since
3512 * we can be called via check_for_release() with no
3513 * synchronization other than RCU, and the subsystem linked
3514 * list isn't RCU-safe */
3517 * We won't need to lock the subsys array, because the subsystems
3518 * we're concerned about aren't going anywhere since our cgroup root
3519 * has a reference on them.
3521 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3522 struct cgroup_subsys *ss = subsys[i];
3523 struct cgroup_subsys_state *css;
3524 /* Skip subsystems not present or not in this hierarchy */
3525 if (ss == NULL || ss->root != cgrp->root)
3527 css = cgrp->subsys[ss->subsys_id];
3528 /* When called from check_for_release() it's possible
3529 * that by this point the cgroup has been removed
3530 * and the css deleted. But a false-positive doesn't
3531 * matter, since it can only happen if the cgroup
3532 * has been deleted and hence no longer needs the
3533 * release agent to be called anyway. */
3534 if (css && (atomic_read(&css->refcnt) > 1))
3541 * Atomically mark all (or else none) of the cgroup's CSS objects as
3542 * CSS_REMOVED. Return true on success, or false if the cgroup has
3543 * busy subsystems. Call with cgroup_mutex held
3546 static int cgroup_clear_css_refs(struct cgroup *cgrp)
3548 struct cgroup_subsys *ss;
3549 unsigned long flags;
3550 bool failed = false;
3551 local_irq_save(flags);
3552 for_each_subsys(cgrp->root, ss) {
3553 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3556 /* We can only remove a CSS with a refcnt==1 */
3557 refcnt = atomic_read(&css->refcnt);
3564 * Drop the refcnt to 0 while we check other
3565 * subsystems. This will cause any racing
3566 * css_tryget() to spin until we set the
3567 * CSS_REMOVED bits or abort
3569 if (atomic_cmpxchg(&css->refcnt, refcnt, 0) == refcnt)
3575 for_each_subsys(cgrp->root, ss) {
3576 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3579 * Restore old refcnt if we previously managed
3580 * to clear it from 1 to 0
3582 if (!atomic_read(&css->refcnt))
3583 atomic_set(&css->refcnt, 1);
3585 /* Commit the fact that the CSS is removed */
3586 set_bit(CSS_REMOVED, &css->flags);
3589 local_irq_restore(flags);
3593 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
3595 struct cgroup *cgrp = dentry->d_fsdata;
3597 struct cgroup *parent;
3599 struct cgroup_event *event, *tmp;
3602 /* the vfs holds both inode->i_mutex already */
3604 mutex_lock(&cgroup_mutex);
3605 if (atomic_read(&cgrp->count) != 0) {
3606 mutex_unlock(&cgroup_mutex);
3609 if (!list_empty(&cgrp->children)) {
3610 mutex_unlock(&cgroup_mutex);
3613 mutex_unlock(&cgroup_mutex);
3616 * In general, subsystem has no css->refcnt after pre_destroy(). But
3617 * in racy cases, subsystem may have to get css->refcnt after
3618 * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
3619 * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
3620 * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
3621 * and subsystem's reference count handling. Please see css_get/put
3622 * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
3624 set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3627 * Call pre_destroy handlers of subsys. Notify subsystems
3628 * that rmdir() request comes.
3630 ret = cgroup_call_pre_destroy(cgrp);
3632 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3636 mutex_lock(&cgroup_mutex);
3637 parent = cgrp->parent;
3638 if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
3639 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3640 mutex_unlock(&cgroup_mutex);
3643 prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
3644 if (!cgroup_clear_css_refs(cgrp)) {
3645 mutex_unlock(&cgroup_mutex);
3647 * Because someone may call cgroup_wakeup_rmdir_waiter() before
3648 * prepare_to_wait(), we need to check this flag.
3650 if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
3652 finish_wait(&cgroup_rmdir_waitq, &wait);
3653 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3654 if (signal_pending(current))
3658 /* NO css_tryget() can success after here. */
3659 finish_wait(&cgroup_rmdir_waitq, &wait);
3660 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3662 spin_lock(&release_list_lock);
3663 set_bit(CGRP_REMOVED, &cgrp->flags);
3664 if (!list_empty(&cgrp->release_list))
3665 list_del(&cgrp->release_list);
3666 spin_unlock(&release_list_lock);
3668 cgroup_lock_hierarchy(cgrp->root);
3669 /* delete this cgroup from parent->children */
3670 list_del(&cgrp->sibling);
3671 cgroup_unlock_hierarchy(cgrp->root);
3673 d = dget(cgrp->dentry);
3675 cgroup_d_remove_dir(d);
3678 set_bit(CGRP_RELEASABLE, &parent->flags);
3679 check_for_release(parent);
3682 * Unregister events and notify userspace.
3683 * Notify userspace about cgroup removing only after rmdir of cgroup
3684 * directory to avoid race between userspace and kernelspace
3686 spin_lock(&cgrp->event_list_lock);
3687 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
3688 list_del(&event->list);
3689 remove_wait_queue(event->wqh, &event->wait);
3690 eventfd_signal(event->eventfd, 1);
3691 schedule_work(&event->remove);
3693 spin_unlock(&cgrp->event_list_lock);
3695 mutex_unlock(&cgroup_mutex);
3699 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
3701 struct cgroup_subsys_state *css;
3703 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
3705 /* Create the top cgroup state for this subsystem */
3706 list_add(&ss->sibling, &rootnode.subsys_list);
3707 ss->root = &rootnode;
3708 css = ss->create(ss, dummytop);
3709 /* We don't handle early failures gracefully */
3710 BUG_ON(IS_ERR(css));
3711 init_cgroup_css(css, ss, dummytop);
3713 /* Update the init_css_set to contain a subsys
3714 * pointer to this state - since the subsystem is
3715 * newly registered, all tasks and hence the
3716 * init_css_set is in the subsystem's top cgroup. */
3717 init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
3719 need_forkexit_callback |= ss->fork || ss->exit;
3721 /* At system boot, before all subsystems have been
3722 * registered, no tasks have been forked, so we don't
3723 * need to invoke fork callbacks here. */
3724 BUG_ON(!list_empty(&init_task.tasks));
3726 mutex_init(&ss->hierarchy_mutex);
3727 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3730 /* this function shouldn't be used with modular subsystems, since they
3731 * need to register a subsys_id, among other things */
3736 * cgroup_load_subsys: load and register a modular subsystem at runtime
3737 * @ss: the subsystem to load
3739 * This function should be called in a modular subsystem's initcall. If the
3740 * subsystem is built as a module, it will be assigned a new subsys_id and set
3741 * up for use. If the subsystem is built-in anyway, work is delegated to the
3742 * simpler cgroup_init_subsys.
3744 int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
3747 struct cgroup_subsys_state *css;
3749 /* check name and function validity */
3750 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
3751 ss->create == NULL || ss->destroy == NULL)
3755 * we don't support callbacks in modular subsystems. this check is
3756 * before the ss->module check for consistency; a subsystem that could
3757 * be a module should still have no callbacks even if the user isn't
3758 * compiling it as one.
3760 if (ss->fork || ss->exit)
3764 * an optionally modular subsystem is built-in: we want to do nothing,
3765 * since cgroup_init_subsys will have already taken care of it.
3767 if (ss->module == NULL) {
3768 /* a few sanity checks */
3769 BUG_ON(ss->subsys_id >= CGROUP_BUILTIN_SUBSYS_COUNT);
3770 BUG_ON(subsys[ss->subsys_id] != ss);
3775 * need to register a subsys id before anything else - for example,
3776 * init_cgroup_css needs it.
3778 mutex_lock(&cgroup_mutex);
3779 /* find the first empty slot in the array */
3780 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
3781 if (subsys[i] == NULL)
3784 if (i == CGROUP_SUBSYS_COUNT) {
3785 /* maximum number of subsystems already registered! */
3786 mutex_unlock(&cgroup_mutex);
3789 /* assign ourselves the subsys_id */
3794 * no ss->create seems to need anything important in the ss struct, so
3795 * this can happen first (i.e. before the rootnode attachment).
3797 css = ss->create(ss, dummytop);
3799 /* failure case - need to deassign the subsys[] slot. */
3801 mutex_unlock(&cgroup_mutex);
3802 return PTR_ERR(css);
3805 list_add(&ss->sibling, &rootnode.subsys_list);
3806 ss->root = &rootnode;
3808 /* our new subsystem will be attached to the dummy hierarchy. */
3809 init_cgroup_css(css, ss, dummytop);
3810 /* init_idr must be after init_cgroup_css because it sets css->id. */
3812 int ret = cgroup_init_idr(ss, css);
3814 dummytop->subsys[ss->subsys_id] = NULL;
3815 ss->destroy(ss, dummytop);
3817 mutex_unlock(&cgroup_mutex);
3823 * Now we need to entangle the css into the existing css_sets. unlike
3824 * in cgroup_init_subsys, there are now multiple css_sets, so each one
3825 * will need a new pointer to it; done by iterating the css_set_table.
3826 * furthermore, modifying the existing css_sets will corrupt the hash
3827 * table state, so each changed css_set will need its hash recomputed.
3828 * this is all done under the css_set_lock.
3830 write_lock(&css_set_lock);
3831 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
3833 struct hlist_node *node, *tmp;
3834 struct hlist_head *bucket = &css_set_table[i], *new_bucket;
3836 hlist_for_each_entry_safe(cg, node, tmp, bucket, hlist) {
3837 /* skip entries that we already rehashed */
3838 if (cg->subsys[ss->subsys_id])
3840 /* remove existing entry */
3841 hlist_del(&cg->hlist);
3843 cg->subsys[ss->subsys_id] = css;
3844 /* recompute hash and restore entry */
3845 new_bucket = css_set_hash(cg->subsys);
3846 hlist_add_head(&cg->hlist, new_bucket);
3849 write_unlock(&css_set_lock);
3851 mutex_init(&ss->hierarchy_mutex);
3852 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3856 mutex_unlock(&cgroup_mutex);
3859 EXPORT_SYMBOL_GPL(cgroup_load_subsys);
3862 * cgroup_unload_subsys: unload a modular subsystem
3863 * @ss: the subsystem to unload
3865 * This function should be called in a modular subsystem's exitcall. When this
3866 * function is invoked, the refcount on the subsystem's module will be 0, so
3867 * the subsystem will not be attached to any hierarchy.
3869 void cgroup_unload_subsys(struct cgroup_subsys *ss)
3871 struct cg_cgroup_link *link;
3872 struct hlist_head *hhead;
3874 BUG_ON(ss->module == NULL);
3877 * we shouldn't be called if the subsystem is in use, and the use of
3878 * try_module_get in parse_cgroupfs_options should ensure that it
3879 * doesn't start being used while we're killing it off.
3881 BUG_ON(ss->root != &rootnode);
3883 mutex_lock(&cgroup_mutex);
3884 /* deassign the subsys_id */
3885 BUG_ON(ss->subsys_id < CGROUP_BUILTIN_SUBSYS_COUNT);
3886 subsys[ss->subsys_id] = NULL;
3888 /* remove subsystem from rootnode's list of subsystems */
3889 list_del(&ss->sibling);
3892 * disentangle the css from all css_sets attached to the dummytop. as
3893 * in loading, we need to pay our respects to the hashtable gods.
3895 write_lock(&css_set_lock);
3896 list_for_each_entry(link, &dummytop->css_sets, cgrp_link_list) {
3897 struct css_set *cg = link->cg;
3899 hlist_del(&cg->hlist);
3900 BUG_ON(!cg->subsys[ss->subsys_id]);
3901 cg->subsys[ss->subsys_id] = NULL;
3902 hhead = css_set_hash(cg->subsys);
3903 hlist_add_head(&cg->hlist, hhead);
3905 write_unlock(&css_set_lock);
3908 * remove subsystem's css from the dummytop and free it - need to free
3909 * before marking as null because ss->destroy needs the cgrp->subsys
3910 * pointer to find their state. note that this also takes care of
3911 * freeing the css_id.
3913 ss->destroy(ss, dummytop);
3914 dummytop->subsys[ss->subsys_id] = NULL;
3916 mutex_unlock(&cgroup_mutex);
3918 EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
3921 * cgroup_init_early - cgroup initialization at system boot
3923 * Initialize cgroups at system boot, and initialize any
3924 * subsystems that request early init.
3926 int __init cgroup_init_early(void)
3929 atomic_set(&init_css_set.refcount, 1);
3930 INIT_LIST_HEAD(&init_css_set.cg_links);
3931 INIT_LIST_HEAD(&init_css_set.tasks);
3932 INIT_HLIST_NODE(&init_css_set.hlist);
3934 init_cgroup_root(&rootnode);
3936 init_task.cgroups = &init_css_set;
3938 init_css_set_link.cg = &init_css_set;
3939 init_css_set_link.cgrp = dummytop;
3940 list_add(&init_css_set_link.cgrp_link_list,
3941 &rootnode.top_cgroup.css_sets);
3942 list_add(&init_css_set_link.cg_link_list,
3943 &init_css_set.cg_links);
3945 for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
3946 INIT_HLIST_HEAD(&css_set_table[i]);
3948 /* at bootup time, we don't worry about modular subsystems */
3949 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3950 struct cgroup_subsys *ss = subsys[i];
3953 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
3954 BUG_ON(!ss->create);
3955 BUG_ON(!ss->destroy);
3956 if (ss->subsys_id != i) {
3957 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
3958 ss->name, ss->subsys_id);
3963 cgroup_init_subsys(ss);
3969 * cgroup_init - cgroup initialization
3971 * Register cgroup filesystem and /proc file, and initialize
3972 * any subsystems that didn't request early init.
3974 int __init cgroup_init(void)
3978 struct hlist_head *hhead;
3980 err = bdi_init(&cgroup_backing_dev_info);
3984 /* at bootup time, we don't worry about modular subsystems */
3985 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3986 struct cgroup_subsys *ss = subsys[i];
3987 if (!ss->early_init)
3988 cgroup_init_subsys(ss);
3990 cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
3993 /* Add init_css_set to the hash table */
3994 hhead = css_set_hash(init_css_set.subsys);
3995 hlist_add_head(&init_css_set.hlist, hhead);
3996 BUG_ON(!init_root_id(&rootnode));
3998 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4004 err = register_filesystem(&cgroup_fs_type);
4006 kobject_put(cgroup_kobj);
4010 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4014 bdi_destroy(&cgroup_backing_dev_info);
4020 * proc_cgroup_show()
4021 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4022 * - Used for /proc/<pid>/cgroup.
4023 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4024 * doesn't really matter if tsk->cgroup changes after we read it,
4025 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4026 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4027 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4028 * cgroup to top_cgroup.
4031 /* TODO: Use a proper seq_file iterator */
4032 static int proc_cgroup_show(struct seq_file *m, void *v)
4035 struct task_struct *tsk;
4038 struct cgroupfs_root *root;
4041 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4047 tsk = get_pid_task(pid, PIDTYPE_PID);
4053 mutex_lock(&cgroup_mutex);
4055 for_each_active_root(root) {
4056 struct cgroup_subsys *ss;
4057 struct cgroup *cgrp;
4060 seq_printf(m, "%d:", root->hierarchy_id);
4061 for_each_subsys(root, ss)
4062 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4063 if (strlen(root->name))
4064 seq_printf(m, "%sname=%s", count ? "," : "",
4067 cgrp = task_cgroup_from_root(tsk, root);
4068 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4076 mutex_unlock(&cgroup_mutex);
4077 put_task_struct(tsk);
4084 static int cgroup_open(struct inode *inode, struct file *file)
4086 struct pid *pid = PROC_I(inode)->pid;
4087 return single_open(file, proc_cgroup_show, pid);
4090 const struct file_operations proc_cgroup_operations = {
4091 .open = cgroup_open,
4093 .llseek = seq_lseek,
4094 .release = single_release,
4097 /* Display information about each subsystem and each hierarchy */
4098 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4102 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4104 * ideally we don't want subsystems moving around while we do this.
4105 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4106 * subsys/hierarchy state.
4108 mutex_lock(&cgroup_mutex);
4109 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
4110 struct cgroup_subsys *ss = subsys[i];
4113 seq_printf(m, "%s\t%d\t%d\t%d\n",
4114 ss->name, ss->root->hierarchy_id,
4115 ss->root->number_of_cgroups, !ss->disabled);
4117 mutex_unlock(&cgroup_mutex);
4121 static int cgroupstats_open(struct inode *inode, struct file *file)
4123 return single_open(file, proc_cgroupstats_show, NULL);
4126 static const struct file_operations proc_cgroupstats_operations = {
4127 .open = cgroupstats_open,
4129 .llseek = seq_lseek,
4130 .release = single_release,
4134 * cgroup_fork - attach newly forked task to its parents cgroup.
4135 * @child: pointer to task_struct of forking parent process.
4137 * Description: A task inherits its parent's cgroup at fork().
4139 * A pointer to the shared css_set was automatically copied in
4140 * fork.c by dup_task_struct(). However, we ignore that copy, since
4141 * it was not made under the protection of RCU or cgroup_mutex, so
4142 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4143 * have already changed current->cgroups, allowing the previously
4144 * referenced cgroup group to be removed and freed.
4146 * At the point that cgroup_fork() is called, 'current' is the parent
4147 * task, and the passed argument 'child' points to the child task.
4149 void cgroup_fork(struct task_struct *child)
4152 child->cgroups = current->cgroups;
4153 get_css_set(child->cgroups);
4154 task_unlock(current);
4155 INIT_LIST_HEAD(&child->cg_list);
4159 * cgroup_fork_callbacks - run fork callbacks
4160 * @child: the new task
4162 * Called on a new task very soon before adding it to the
4163 * tasklist. No need to take any locks since no-one can
4164 * be operating on this task.
4166 void cgroup_fork_callbacks(struct task_struct *child)
4168 if (need_forkexit_callback) {
4171 * forkexit callbacks are only supported for builtin
4172 * subsystems, and the builtin section of the subsys array is
4173 * immutable, so we don't need to lock the subsys array here.
4175 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4176 struct cgroup_subsys *ss = subsys[i];
4178 ss->fork(ss, child);
4184 * cgroup_post_fork - called on a new task after adding it to the task list
4185 * @child: the task in question
4187 * Adds the task to the list running through its css_set if necessary.
4188 * Has to be after the task is visible on the task list in case we race
4189 * with the first call to cgroup_iter_start() - to guarantee that the
4190 * new task ends up on its list.
4192 void cgroup_post_fork(struct task_struct *child)
4194 if (use_task_css_set_links) {
4195 write_lock(&css_set_lock);
4197 if (list_empty(&child->cg_list))
4198 list_add(&child->cg_list, &child->cgroups->tasks);
4200 write_unlock(&css_set_lock);
4204 * cgroup_exit - detach cgroup from exiting task
4205 * @tsk: pointer to task_struct of exiting process
4206 * @run_callback: run exit callbacks?
4208 * Description: Detach cgroup from @tsk and release it.
4210 * Note that cgroups marked notify_on_release force every task in
4211 * them to take the global cgroup_mutex mutex when exiting.
4212 * This could impact scaling on very large systems. Be reluctant to
4213 * use notify_on_release cgroups where very high task exit scaling
4214 * is required on large systems.
4216 * the_top_cgroup_hack:
4218 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4220 * We call cgroup_exit() while the task is still competent to
4221 * handle notify_on_release(), then leave the task attached to the
4222 * root cgroup in each hierarchy for the remainder of its exit.
4224 * To do this properly, we would increment the reference count on
4225 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4226 * code we would add a second cgroup function call, to drop that
4227 * reference. This would just create an unnecessary hot spot on
4228 * the top_cgroup reference count, to no avail.
4230 * Normally, holding a reference to a cgroup without bumping its
4231 * count is unsafe. The cgroup could go away, or someone could
4232 * attach us to a different cgroup, decrementing the count on
4233 * the first cgroup that we never incremented. But in this case,
4234 * top_cgroup isn't going away, and either task has PF_EXITING set,
4235 * which wards off any cgroup_attach_task() attempts, or task is a failed
4236 * fork, never visible to cgroup_attach_task.
4238 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4243 if (run_callbacks && need_forkexit_callback) {
4245 * modular subsystems can't use callbacks, so no need to lock
4248 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4249 struct cgroup_subsys *ss = subsys[i];
4256 * Unlink from the css_set task list if necessary.
4257 * Optimistically check cg_list before taking
4260 if (!list_empty(&tsk->cg_list)) {
4261 write_lock(&css_set_lock);
4262 if (!list_empty(&tsk->cg_list))
4263 list_del(&tsk->cg_list);
4264 write_unlock(&css_set_lock);
4267 /* Reassign the task to the init_css_set. */
4270 tsk->cgroups = &init_css_set;
4273 put_css_set_taskexit(cg);
4277 * cgroup_clone - clone the cgroup the given subsystem is attached to
4278 * @tsk: the task to be moved
4279 * @subsys: the given subsystem
4280 * @nodename: the name for the new cgroup
4282 * Duplicate the current cgroup in the hierarchy that the given
4283 * subsystem is attached to, and move this task into the new
4286 int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
4289 struct dentry *dentry;
4291 struct cgroup *parent, *child;
4292 struct inode *inode;
4294 struct cgroupfs_root *root;
4295 struct cgroup_subsys *ss;
4297 /* We shouldn't be called by an unregistered subsystem */
4298 BUG_ON(!subsys->active);
4300 /* First figure out what hierarchy and cgroup we're dealing
4301 * with, and pin them so we can drop cgroup_mutex */
4302 mutex_lock(&cgroup_mutex);
4304 root = subsys->root;
4305 if (root == &rootnode) {
4306 mutex_unlock(&cgroup_mutex);
4310 /* Pin the hierarchy */
4311 if (!atomic_inc_not_zero(&root->sb->s_active)) {
4312 /* We race with the final deactivate_super() */
4313 mutex_unlock(&cgroup_mutex);
4317 /* Keep the cgroup alive */
4319 parent = task_cgroup(tsk, subsys->subsys_id);
4324 mutex_unlock(&cgroup_mutex);
4326 /* Now do the VFS work to create a cgroup */
4327 inode = parent->dentry->d_inode;
4329 /* Hold the parent directory mutex across this operation to
4330 * stop anyone else deleting the new cgroup */
4331 mutex_lock(&inode->i_mutex);
4332 dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
4333 if (IS_ERR(dentry)) {
4335 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
4337 ret = PTR_ERR(dentry);
4341 /* Create the cgroup directory, which also creates the cgroup */
4342 ret = vfs_mkdir(inode, dentry, 0755);
4343 child = __d_cgrp(dentry);
4347 "Failed to create cgroup %s: %d\n", nodename,
4352 /* The cgroup now exists. Retake cgroup_mutex and check
4353 * that we're still in the same state that we thought we
4355 mutex_lock(&cgroup_mutex);
4356 if ((root != subsys->root) ||
4357 (parent != task_cgroup(tsk, subsys->subsys_id))) {
4358 /* Aargh, we raced ... */
4359 mutex_unlock(&inode->i_mutex);
4362 deactivate_super(root->sb);
4363 /* The cgroup is still accessible in the VFS, but
4364 * we're not going to try to rmdir() it at this
4367 "Race in cgroup_clone() - leaking cgroup %s\n",
4372 /* do any required auto-setup */
4373 for_each_subsys(root, ss) {
4375 ss->post_clone(ss, child);
4378 /* All seems fine. Finish by moving the task into the new cgroup */
4379 ret = cgroup_attach_task(child, tsk);
4380 mutex_unlock(&cgroup_mutex);
4383 mutex_unlock(&inode->i_mutex);
4385 mutex_lock(&cgroup_mutex);
4387 mutex_unlock(&cgroup_mutex);
4388 deactivate_super(root->sb);
4393 * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
4394 * @cgrp: the cgroup in question
4395 * @task: the task in question
4397 * See if @cgrp is a descendant of @task's cgroup in the appropriate
4400 * If we are sending in dummytop, then presumably we are creating
4401 * the top cgroup in the subsystem.
4403 * Called only by the ns (nsproxy) cgroup.
4405 int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
4408 struct cgroup *target;
4410 if (cgrp == dummytop)
4413 target = task_cgroup_from_root(task, cgrp->root);
4414 while (cgrp != target && cgrp!= cgrp->top_cgroup)
4415 cgrp = cgrp->parent;
4416 ret = (cgrp == target);
4420 static void check_for_release(struct cgroup *cgrp)
4422 /* All of these checks rely on RCU to keep the cgroup
4423 * structure alive */
4424 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
4425 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
4426 /* Control Group is currently removeable. If it's not
4427 * already queued for a userspace notification, queue
4429 int need_schedule_work = 0;
4430 spin_lock(&release_list_lock);
4431 if (!cgroup_is_removed(cgrp) &&
4432 list_empty(&cgrp->release_list)) {
4433 list_add(&cgrp->release_list, &release_list);
4434 need_schedule_work = 1;
4436 spin_unlock(&release_list_lock);
4437 if (need_schedule_work)
4438 schedule_work(&release_agent_work);
4442 /* Caller must verify that the css is not for root cgroup */
4443 void __css_put(struct cgroup_subsys_state *css, int count)
4445 struct cgroup *cgrp = css->cgroup;
4448 val = atomic_sub_return(count, &css->refcnt);
4450 if (notify_on_release(cgrp)) {
4451 set_bit(CGRP_RELEASABLE, &cgrp->flags);
4452 check_for_release(cgrp);
4454 cgroup_wakeup_rmdir_waiter(cgrp);
4457 WARN_ON_ONCE(val < 1);
4459 EXPORT_SYMBOL_GPL(__css_put);
4462 * Notify userspace when a cgroup is released, by running the
4463 * configured release agent with the name of the cgroup (path
4464 * relative to the root of cgroup file system) as the argument.
4466 * Most likely, this user command will try to rmdir this cgroup.
4468 * This races with the possibility that some other task will be
4469 * attached to this cgroup before it is removed, or that some other
4470 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4471 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4472 * unused, and this cgroup will be reprieved from its death sentence,
4473 * to continue to serve a useful existence. Next time it's released,
4474 * we will get notified again, if it still has 'notify_on_release' set.
4476 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4477 * means only wait until the task is successfully execve()'d. The
4478 * separate release agent task is forked by call_usermodehelper(),
4479 * then control in this thread returns here, without waiting for the
4480 * release agent task. We don't bother to wait because the caller of
4481 * this routine has no use for the exit status of the release agent
4482 * task, so no sense holding our caller up for that.
4484 static void cgroup_release_agent(struct work_struct *work)
4486 BUG_ON(work != &release_agent_work);
4487 mutex_lock(&cgroup_mutex);
4488 spin_lock(&release_list_lock);
4489 while (!list_empty(&release_list)) {
4490 char *argv[3], *envp[3];
4492 char *pathbuf = NULL, *agentbuf = NULL;
4493 struct cgroup *cgrp = list_entry(release_list.next,
4496 list_del_init(&cgrp->release_list);
4497 spin_unlock(&release_list_lock);
4498 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4501 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4503 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4508 argv[i++] = agentbuf;
4509 argv[i++] = pathbuf;
4513 /* minimal command environment */
4514 envp[i++] = "HOME=/";
4515 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4518 /* Drop the lock while we invoke the usermode helper,
4519 * since the exec could involve hitting disk and hence
4520 * be a slow process */
4521 mutex_unlock(&cgroup_mutex);
4522 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4523 mutex_lock(&cgroup_mutex);
4527 spin_lock(&release_list_lock);
4529 spin_unlock(&release_list_lock);
4530 mutex_unlock(&cgroup_mutex);
4533 static int __init cgroup_disable(char *str)
4538 while ((token = strsep(&str, ",")) != NULL) {
4542 * cgroup_disable, being at boot time, can't know about module
4543 * subsystems, so we don't worry about them.
4545 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4546 struct cgroup_subsys *ss = subsys[i];
4548 if (!strcmp(token, ss->name)) {
4550 printk(KERN_INFO "Disabling %s control group"
4551 " subsystem\n", ss->name);
4558 __setup("cgroup_disable=", cgroup_disable);
4561 * Functons for CSS ID.
4565 *To get ID other than 0, this should be called when !cgroup_is_removed().
4567 unsigned short css_id(struct cgroup_subsys_state *css)
4569 struct css_id *cssid;
4572 * This css_id() can return correct value when somone has refcnt
4573 * on this or this is under rcu_read_lock(). Once css->id is allocated,
4574 * it's unchanged until freed.
4576 cssid = rcu_dereference_check(css->id,
4577 rcu_read_lock_held() || atomic_read(&css->refcnt));
4583 EXPORT_SYMBOL_GPL(css_id);
4585 unsigned short css_depth(struct cgroup_subsys_state *css)
4587 struct css_id *cssid;
4589 cssid = rcu_dereference_check(css->id,
4590 rcu_read_lock_held() || atomic_read(&css->refcnt));
4593 return cssid->depth;
4596 EXPORT_SYMBOL_GPL(css_depth);
4599 * css_is_ancestor - test "root" css is an ancestor of "child"
4600 * @child: the css to be tested.
4601 * @root: the css supporsed to be an ancestor of the child.
4603 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
4604 * this function reads css->id, this use rcu_dereference() and rcu_read_lock().
4605 * But, considering usual usage, the csses should be valid objects after test.
4606 * Assuming that the caller will do some action to the child if this returns
4607 * returns true, the caller must take "child";s reference count.
4608 * If "child" is valid object and this returns true, "root" is valid, too.
4611 bool css_is_ancestor(struct cgroup_subsys_state *child,
4612 const struct cgroup_subsys_state *root)
4614 struct css_id *child_id;
4615 struct css_id *root_id;
4619 child_id = rcu_dereference(child->id);
4620 root_id = rcu_dereference(root->id);
4623 || (child_id->depth < root_id->depth)
4624 || (child_id->stack[root_id->depth] != root_id->id))
4630 static void __free_css_id_cb(struct rcu_head *head)
4634 id = container_of(head, struct css_id, rcu_head);
4638 void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
4640 struct css_id *id = css->id;
4641 /* When this is called before css_id initialization, id can be NULL */
4645 BUG_ON(!ss->use_id);
4647 rcu_assign_pointer(id->css, NULL);
4648 rcu_assign_pointer(css->id, NULL);
4649 spin_lock(&ss->id_lock);
4650 idr_remove(&ss->idr, id->id);
4651 spin_unlock(&ss->id_lock);
4652 call_rcu(&id->rcu_head, __free_css_id_cb);
4654 EXPORT_SYMBOL_GPL(free_css_id);
4657 * This is called by init or create(). Then, calls to this function are
4658 * always serialized (By cgroup_mutex() at create()).
4661 static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
4663 struct css_id *newid;
4664 int myid, error, size;
4666 BUG_ON(!ss->use_id);
4668 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
4669 newid = kzalloc(size, GFP_KERNEL);
4671 return ERR_PTR(-ENOMEM);
4673 if (unlikely(!idr_pre_get(&ss->idr, GFP_KERNEL))) {
4677 spin_lock(&ss->id_lock);
4678 /* Don't use 0. allocates an ID of 1-65535 */
4679 error = idr_get_new_above(&ss->idr, newid, 1, &myid);
4680 spin_unlock(&ss->id_lock);
4682 /* Returns error when there are no free spaces for new ID.*/
4687 if (myid > CSS_ID_MAX)
4691 newid->depth = depth;
4695 spin_lock(&ss->id_lock);
4696 idr_remove(&ss->idr, myid);
4697 spin_unlock(&ss->id_lock);
4700 return ERR_PTR(error);
4704 static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
4705 struct cgroup_subsys_state *rootcss)
4707 struct css_id *newid;
4709 spin_lock_init(&ss->id_lock);
4712 newid = get_new_cssid(ss, 0);
4714 return PTR_ERR(newid);
4716 newid->stack[0] = newid->id;
4717 newid->css = rootcss;
4718 rootcss->id = newid;
4722 static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
4723 struct cgroup *child)
4725 int subsys_id, i, depth = 0;
4726 struct cgroup_subsys_state *parent_css, *child_css;
4727 struct css_id *child_id, *parent_id;
4729 subsys_id = ss->subsys_id;
4730 parent_css = parent->subsys[subsys_id];
4731 child_css = child->subsys[subsys_id];
4732 parent_id = parent_css->id;
4733 depth = parent_id->depth + 1;
4735 child_id = get_new_cssid(ss, depth);
4736 if (IS_ERR(child_id))
4737 return PTR_ERR(child_id);
4739 for (i = 0; i < depth; i++)
4740 child_id->stack[i] = parent_id->stack[i];
4741 child_id->stack[depth] = child_id->id;
4743 * child_id->css pointer will be set after this cgroup is available
4744 * see cgroup_populate_dir()
4746 rcu_assign_pointer(child_css->id, child_id);
4752 * css_lookup - lookup css by id
4753 * @ss: cgroup subsys to be looked into.
4756 * Returns pointer to cgroup_subsys_state if there is valid one with id.
4757 * NULL if not. Should be called under rcu_read_lock()
4759 struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
4761 struct css_id *cssid = NULL;
4763 BUG_ON(!ss->use_id);
4764 cssid = idr_find(&ss->idr, id);
4766 if (unlikely(!cssid))
4769 return rcu_dereference(cssid->css);
4771 EXPORT_SYMBOL_GPL(css_lookup);
4774 * css_get_next - lookup next cgroup under specified hierarchy.
4775 * @ss: pointer to subsystem
4776 * @id: current position of iteration.
4777 * @root: pointer to css. search tree under this.
4778 * @foundid: position of found object.
4780 * Search next css under the specified hierarchy of rootid. Calling under
4781 * rcu_read_lock() is necessary. Returns NULL if it reaches the end.
4783 struct cgroup_subsys_state *
4784 css_get_next(struct cgroup_subsys *ss, int id,
4785 struct cgroup_subsys_state *root, int *foundid)
4787 struct cgroup_subsys_state *ret = NULL;
4790 int rootid = css_id(root);
4791 int depth = css_depth(root);
4796 BUG_ON(!ss->use_id);
4797 /* fill start point for scan */
4801 * scan next entry from bitmap(tree), tmpid is updated after
4804 spin_lock(&ss->id_lock);
4805 tmp = idr_get_next(&ss->idr, &tmpid);
4806 spin_unlock(&ss->id_lock);
4810 if (tmp->depth >= depth && tmp->stack[depth] == rootid) {
4811 ret = rcu_dereference(tmp->css);
4817 /* continue to scan from next id */
4823 #ifdef CONFIG_CGROUP_DEBUG
4824 static struct cgroup_subsys_state *debug_create(struct cgroup_subsys *ss,
4825 struct cgroup *cont)
4827 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4830 return ERR_PTR(-ENOMEM);
4835 static void debug_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
4837 kfree(cont->subsys[debug_subsys_id]);
4840 static u64 cgroup_refcount_read(struct cgroup *cont, struct cftype *cft)
4842 return atomic_read(&cont->count);
4845 static u64 debug_taskcount_read(struct cgroup *cont, struct cftype *cft)
4847 return cgroup_task_count(cont);
4850 static u64 current_css_set_read(struct cgroup *cont, struct cftype *cft)
4852 return (u64)(unsigned long)current->cgroups;
4855 static u64 current_css_set_refcount_read(struct cgroup *cont,
4861 count = atomic_read(¤t->cgroups->refcount);
4866 static int current_css_set_cg_links_read(struct cgroup *cont,
4868 struct seq_file *seq)
4870 struct cg_cgroup_link *link;
4873 read_lock(&css_set_lock);
4875 cg = rcu_dereference(current->cgroups);
4876 list_for_each_entry(link, &cg->cg_links, cg_link_list) {
4877 struct cgroup *c = link->cgrp;
4881 name = c->dentry->d_name.name;
4884 seq_printf(seq, "Root %d group %s\n",
4885 c->root->hierarchy_id, name);
4888 read_unlock(&css_set_lock);
4892 #define MAX_TASKS_SHOWN_PER_CSS 25
4893 static int cgroup_css_links_read(struct cgroup *cont,
4895 struct seq_file *seq)
4897 struct cg_cgroup_link *link;
4899 read_lock(&css_set_lock);
4900 list_for_each_entry(link, &cont->css_sets, cgrp_link_list) {
4901 struct css_set *cg = link->cg;
4902 struct task_struct *task;
4904 seq_printf(seq, "css_set %p\n", cg);
4905 list_for_each_entry(task, &cg->tasks, cg_list) {
4906 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4907 seq_puts(seq, " ...\n");
4910 seq_printf(seq, " task %d\n",
4911 task_pid_vnr(task));
4915 read_unlock(&css_set_lock);
4919 static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
4921 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
4924 static struct cftype debug_files[] = {
4926 .name = "cgroup_refcount",
4927 .read_u64 = cgroup_refcount_read,
4930 .name = "taskcount",
4931 .read_u64 = debug_taskcount_read,
4935 .name = "current_css_set",
4936 .read_u64 = current_css_set_read,
4940 .name = "current_css_set_refcount",
4941 .read_u64 = current_css_set_refcount_read,
4945 .name = "current_css_set_cg_links",
4946 .read_seq_string = current_css_set_cg_links_read,
4950 .name = "cgroup_css_links",
4951 .read_seq_string = cgroup_css_links_read,
4955 .name = "releasable",
4956 .read_u64 = releasable_read,
4960 static int debug_populate(struct cgroup_subsys *ss, struct cgroup *cont)
4962 return cgroup_add_files(cont, ss, debug_files,
4963 ARRAY_SIZE(debug_files));
4966 struct cgroup_subsys debug_subsys = {
4968 .create = debug_create,
4969 .destroy = debug_destroy,
4970 .populate = debug_populate,
4971 .subsys_id = debug_subsys_id,
4973 #endif /* CONFIG_CGROUP_DEBUG */