2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #include <linux/cgroup.h>
30 #include <linux/ctype.h>
31 #include <linux/errno.h>
33 #include <linux/kernel.h>
34 #include <linux/list.h>
36 #include <linux/mutex.h>
37 #include <linux/mount.h>
38 #include <linux/pagemap.h>
39 #include <linux/proc_fs.h>
40 #include <linux/rcupdate.h>
41 #include <linux/sched.h>
42 #include <linux/backing-dev.h>
43 #include <linux/seq_file.h>
44 #include <linux/slab.h>
45 #include <linux/magic.h>
46 #include <linux/spinlock.h>
47 #include <linux/string.h>
48 #include <linux/sort.h>
49 #include <linux/kmod.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/cgroupstats.h>
53 #include <linux/hash.h>
54 #include <linux/namei.h>
55 #include <linux/smp_lock.h>
56 #include <linux/pid_namespace.h>
57 #include <linux/idr.h>
58 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
59 #include <linux/eventfd.h>
60 #include <linux/poll.h>
61 #include <linux/capability.h>
63 #include <asm/atomic.h>
65 static DEFINE_MUTEX(cgroup_mutex);
68 * Generate an array of cgroup subsystem pointers. At boot time, this is
69 * populated up to CGROUP_BUILTIN_SUBSYS_COUNT, and modular subsystems are
70 * registered after that. The mutable section of this array is protected by
73 #define SUBSYS(_x) &_x ## _subsys,
74 static struct cgroup_subsys *subsys[CGROUP_SUBSYS_COUNT] = {
75 #include <linux/cgroup_subsys.h>
78 #define MAX_CGROUP_ROOT_NAMELEN 64
81 * A cgroupfs_root represents the root of a cgroup hierarchy,
82 * and may be associated with a superblock to form an active
85 struct cgroupfs_root {
86 struct super_block *sb;
89 * The bitmask of subsystems intended to be attached to this
92 unsigned long subsys_bits;
94 /* Unique id for this hierarchy. */
97 /* The bitmask of subsystems currently attached to this hierarchy */
98 unsigned long actual_subsys_bits;
100 /* A list running through the attached subsystems */
101 struct list_head subsys_list;
103 /* The root cgroup for this hierarchy */
104 struct cgroup top_cgroup;
106 /* Tracks how many cgroups are currently defined in hierarchy.*/
107 int number_of_cgroups;
109 /* A list running through the active hierarchies */
110 struct list_head root_list;
112 /* Hierarchy-specific flags */
115 /* The path to use for release notifications. */
116 char release_agent_path[PATH_MAX];
118 /* The name for this hierarchy - may be empty */
119 char name[MAX_CGROUP_ROOT_NAMELEN];
123 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
124 * subsystems that are otherwise unattached - it never has more than a
125 * single cgroup, and all tasks are part of that cgroup.
127 static struct cgroupfs_root rootnode;
130 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
131 * cgroup_subsys->use_id != 0.
133 #define CSS_ID_MAX (65535)
136 * The css to which this ID points. This pointer is set to valid value
137 * after cgroup is populated. If cgroup is removed, this will be NULL.
138 * This pointer is expected to be RCU-safe because destroy()
139 * is called after synchronize_rcu(). But for safe use, css_is_removed()
140 * css_tryget() should be used for avoiding race.
142 struct cgroup_subsys_state *css;
148 * Depth in hierarchy which this ID belongs to.
150 unsigned short depth;
152 * ID is freed by RCU. (and lookup routine is RCU safe.)
154 struct rcu_head rcu_head;
156 * Hierarchy of CSS ID belongs to.
158 unsigned short stack[0]; /* Array of Length (depth+1) */
162 * cgroup_event represents events which userspace want to recieve.
164 struct cgroup_event {
166 * Cgroup which the event belongs to.
170 * Control file which the event associated.
174 * eventfd to signal userspace about the event.
176 struct eventfd_ctx *eventfd;
178 * Each of these stored in a list by the cgroup.
180 struct list_head list;
182 * All fields below needed to unregister event when
183 * userspace closes eventfd.
186 wait_queue_head_t *wqh;
188 struct work_struct remove;
191 /* The list of hierarchy roots */
193 static LIST_HEAD(roots);
194 static int root_count;
196 static DEFINE_IDA(hierarchy_ida);
197 static int next_hierarchy_id;
198 static DEFINE_SPINLOCK(hierarchy_id_lock);
200 /* dummytop is a shorthand for the dummy hierarchy's top cgroup */
201 #define dummytop (&rootnode.top_cgroup)
203 /* This flag indicates whether tasks in the fork and exit paths should
204 * check for fork/exit handlers to call. This avoids us having to do
205 * extra work in the fork/exit path if none of the subsystems need to
208 static int need_forkexit_callback __read_mostly;
210 #ifdef CONFIG_PROVE_LOCKING
211 int cgroup_lock_is_held(void)
213 return lockdep_is_held(&cgroup_mutex);
215 #else /* #ifdef CONFIG_PROVE_LOCKING */
216 int cgroup_lock_is_held(void)
218 return mutex_is_locked(&cgroup_mutex);
220 #endif /* #else #ifdef CONFIG_PROVE_LOCKING */
222 EXPORT_SYMBOL_GPL(cgroup_lock_is_held);
224 /* convenient tests for these bits */
225 inline int cgroup_is_removed(const struct cgroup *cgrp)
227 return test_bit(CGRP_REMOVED, &cgrp->flags);
230 /* bits in struct cgroupfs_root flags field */
232 ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
235 static int cgroup_is_releasable(const struct cgroup *cgrp)
238 (1 << CGRP_RELEASABLE) |
239 (1 << CGRP_NOTIFY_ON_RELEASE);
240 return (cgrp->flags & bits) == bits;
243 static int notify_on_release(const struct cgroup *cgrp)
245 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
249 * for_each_subsys() allows you to iterate on each subsystem attached to
250 * an active hierarchy
252 #define for_each_subsys(_root, _ss) \
253 list_for_each_entry(_ss, &_root->subsys_list, sibling)
255 /* for_each_active_root() allows you to iterate across the active hierarchies */
256 #define for_each_active_root(_root) \
257 list_for_each_entry(_root, &roots, root_list)
259 /* the list of cgroups eligible for automatic release. Protected by
260 * release_list_lock */
261 static LIST_HEAD(release_list);
262 static DEFINE_SPINLOCK(release_list_lock);
263 static void cgroup_release_agent(struct work_struct *work);
264 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
265 static void check_for_release(struct cgroup *cgrp);
267 /* Link structure for associating css_set objects with cgroups */
268 struct cg_cgroup_link {
270 * List running through cg_cgroup_links associated with a
271 * cgroup, anchored on cgroup->css_sets
273 struct list_head cgrp_link_list;
276 * List running through cg_cgroup_links pointing at a
277 * single css_set object, anchored on css_set->cg_links
279 struct list_head cg_link_list;
283 /* The default css_set - used by init and its children prior to any
284 * hierarchies being mounted. It contains a pointer to the root state
285 * for each subsystem. Also used to anchor the list of css_sets. Not
286 * reference-counted, to improve performance when child cgroups
287 * haven't been created.
290 static struct css_set init_css_set;
291 static struct cg_cgroup_link init_css_set_link;
293 static int cgroup_init_idr(struct cgroup_subsys *ss,
294 struct cgroup_subsys_state *css);
296 /* css_set_lock protects the list of css_set objects, and the
297 * chain of tasks off each css_set. Nests outside task->alloc_lock
298 * due to cgroup_iter_start() */
299 static DEFINE_RWLOCK(css_set_lock);
300 static int css_set_count;
303 * hash table for cgroup groups. This improves the performance to find
304 * an existing css_set. This hash doesn't (currently) take into
305 * account cgroups in empty hierarchies.
307 #define CSS_SET_HASH_BITS 7
308 #define CSS_SET_TABLE_SIZE (1 << CSS_SET_HASH_BITS)
309 static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
311 static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
315 unsigned long tmp = 0UL;
317 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
318 tmp += (unsigned long)css[i];
319 tmp = (tmp >> 16) ^ tmp;
321 index = hash_long(tmp, CSS_SET_HASH_BITS);
323 return &css_set_table[index];
326 static void free_css_set_rcu(struct rcu_head *obj)
328 struct css_set *cg = container_of(obj, struct css_set, rcu_head);
332 /* We don't maintain the lists running through each css_set to its
333 * task until after the first call to cgroup_iter_start(). This
334 * reduces the fork()/exit() overhead for people who have cgroups
335 * compiled into their kernel but not actually in use */
336 static int use_task_css_set_links __read_mostly;
339 * refcounted get/put for css_set objects
341 static inline void get_css_set(struct css_set *cg)
343 atomic_inc(&cg->refcount);
346 static void put_css_set(struct css_set *cg)
348 struct cg_cgroup_link *link;
349 struct cg_cgroup_link *saved_link;
351 * Ensure that the refcount doesn't hit zero while any readers
352 * can see it. Similar to atomic_dec_and_lock(), but for an
355 if (atomic_add_unless(&cg->refcount, -1, 1))
357 write_lock(&css_set_lock);
358 if (!atomic_dec_and_test(&cg->refcount)) {
359 write_unlock(&css_set_lock);
363 /* This css_set is dead. unlink it and release cgroup refcounts */
364 hlist_del(&cg->hlist);
367 list_for_each_entry_safe(link, saved_link, &cg->cg_links,
369 struct cgroup *cgrp = link->cgrp;
370 list_del(&link->cg_link_list);
371 list_del(&link->cgrp_link_list);
372 if (atomic_dec_and_test(&cgrp->count))
373 check_for_release(cgrp);
378 write_unlock(&css_set_lock);
379 call_rcu(&cg->rcu_head, free_css_set_rcu);
383 * compare_css_sets - helper function for find_existing_css_set().
384 * @cg: candidate css_set being tested
385 * @old_cg: existing css_set for a task
386 * @new_cgrp: cgroup that's being entered by the task
387 * @template: desired set of css pointers in css_set (pre-calculated)
389 * Returns true if "cg" matches "old_cg" except for the hierarchy
390 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
392 static bool compare_css_sets(struct css_set *cg,
393 struct css_set *old_cg,
394 struct cgroup *new_cgrp,
395 struct cgroup_subsys_state *template[])
397 struct list_head *l1, *l2;
399 if (memcmp(template, cg->subsys, sizeof(cg->subsys))) {
400 /* Not all subsystems matched */
405 * Compare cgroup pointers in order to distinguish between
406 * different cgroups in heirarchies with no subsystems. We
407 * could get by with just this check alone (and skip the
408 * memcmp above) but on most setups the memcmp check will
409 * avoid the need for this more expensive check on almost all
414 l2 = &old_cg->cg_links;
416 struct cg_cgroup_link *cgl1, *cgl2;
417 struct cgroup *cg1, *cg2;
421 /* See if we reached the end - both lists are equal length. */
422 if (l1 == &cg->cg_links) {
423 BUG_ON(l2 != &old_cg->cg_links);
426 BUG_ON(l2 == &old_cg->cg_links);
428 /* Locate the cgroups associated with these links. */
429 cgl1 = list_entry(l1, struct cg_cgroup_link, cg_link_list);
430 cgl2 = list_entry(l2, struct cg_cgroup_link, cg_link_list);
433 /* Hierarchies should be linked in the same order. */
434 BUG_ON(cg1->root != cg2->root);
437 * If this hierarchy is the hierarchy of the cgroup
438 * that's changing, then we need to check that this
439 * css_set points to the new cgroup; if it's any other
440 * hierarchy, then this css_set should point to the
441 * same cgroup as the old css_set.
443 if (cg1->root == new_cgrp->root) {
455 * find_existing_css_set() is a helper for
456 * find_css_set(), and checks to see whether an existing
457 * css_set is suitable.
459 * oldcg: the cgroup group that we're using before the cgroup
462 * cgrp: the cgroup that we're moving into
464 * template: location in which to build the desired set of subsystem
465 * state objects for the new cgroup group
467 static struct css_set *find_existing_css_set(
468 struct css_set *oldcg,
470 struct cgroup_subsys_state *template[])
473 struct cgroupfs_root *root = cgrp->root;
474 struct hlist_head *hhead;
475 struct hlist_node *node;
479 * Build the set of subsystem state objects that we want to see in the
480 * new css_set. while subsystems can change globally, the entries here
481 * won't change, so no need for locking.
483 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
484 if (root->subsys_bits & (1UL << i)) {
485 /* Subsystem is in this hierarchy. So we want
486 * the subsystem state from the new
488 template[i] = cgrp->subsys[i];
490 /* Subsystem is not in this hierarchy, so we
491 * don't want to change the subsystem state */
492 template[i] = oldcg->subsys[i];
496 hhead = css_set_hash(template);
497 hlist_for_each_entry(cg, node, hhead, hlist) {
498 if (!compare_css_sets(cg, oldcg, cgrp, template))
501 /* This css_set matches what we need */
505 /* No existing cgroup group matched */
509 static void free_cg_links(struct list_head *tmp)
511 struct cg_cgroup_link *link;
512 struct cg_cgroup_link *saved_link;
514 list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
515 list_del(&link->cgrp_link_list);
521 * allocate_cg_links() allocates "count" cg_cgroup_link structures
522 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
523 * success or a negative error
525 static int allocate_cg_links(int count, struct list_head *tmp)
527 struct cg_cgroup_link *link;
530 for (i = 0; i < count; i++) {
531 link = kmalloc(sizeof(*link), GFP_KERNEL);
536 list_add(&link->cgrp_link_list, tmp);
542 * link_css_set - a helper function to link a css_set to a cgroup
543 * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
544 * @cg: the css_set to be linked
545 * @cgrp: the destination cgroup
547 static void link_css_set(struct list_head *tmp_cg_links,
548 struct css_set *cg, struct cgroup *cgrp)
550 struct cg_cgroup_link *link;
552 BUG_ON(list_empty(tmp_cg_links));
553 link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
557 atomic_inc(&cgrp->count);
558 list_move(&link->cgrp_link_list, &cgrp->css_sets);
560 * Always add links to the tail of the list so that the list
561 * is sorted by order of hierarchy creation
563 list_add_tail(&link->cg_link_list, &cg->cg_links);
567 * find_css_set() takes an existing cgroup group and a
568 * cgroup object, and returns a css_set object that's
569 * equivalent to the old group, but with the given cgroup
570 * substituted into the appropriate hierarchy. Must be called with
573 static struct css_set *find_css_set(
574 struct css_set *oldcg, struct cgroup *cgrp)
577 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
579 struct list_head tmp_cg_links;
581 struct hlist_head *hhead;
582 struct cg_cgroup_link *link;
584 /* First see if we already have a cgroup group that matches
586 read_lock(&css_set_lock);
587 res = find_existing_css_set(oldcg, cgrp, template);
590 read_unlock(&css_set_lock);
595 res = kmalloc(sizeof(*res), GFP_KERNEL);
599 /* Allocate all the cg_cgroup_link objects that we'll need */
600 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
605 atomic_set(&res->refcount, 1);
606 INIT_LIST_HEAD(&res->cg_links);
607 INIT_LIST_HEAD(&res->tasks);
608 INIT_HLIST_NODE(&res->hlist);
610 /* Copy the set of subsystem state objects generated in
611 * find_existing_css_set() */
612 memcpy(res->subsys, template, sizeof(res->subsys));
614 write_lock(&css_set_lock);
615 /* Add reference counts and links from the new css_set. */
616 list_for_each_entry(link, &oldcg->cg_links, cg_link_list) {
617 struct cgroup *c = link->cgrp;
618 if (c->root == cgrp->root)
620 link_css_set(&tmp_cg_links, res, c);
623 BUG_ON(!list_empty(&tmp_cg_links));
627 /* Add this cgroup group to the hash table */
628 hhead = css_set_hash(res->subsys);
629 hlist_add_head(&res->hlist, hhead);
631 write_unlock(&css_set_lock);
637 * Return the cgroup for "task" from the given hierarchy. Must be
638 * called with cgroup_mutex held.
640 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
641 struct cgroupfs_root *root)
644 struct cgroup *res = NULL;
646 BUG_ON(!mutex_is_locked(&cgroup_mutex));
647 read_lock(&css_set_lock);
649 * No need to lock the task - since we hold cgroup_mutex the
650 * task can't change groups, so the only thing that can happen
651 * is that it exits and its css is set back to init_css_set.
654 if (css == &init_css_set) {
655 res = &root->top_cgroup;
657 struct cg_cgroup_link *link;
658 list_for_each_entry(link, &css->cg_links, cg_link_list) {
659 struct cgroup *c = link->cgrp;
660 if (c->root == root) {
666 read_unlock(&css_set_lock);
672 * There is one global cgroup mutex. We also require taking
673 * task_lock() when dereferencing a task's cgroup subsys pointers.
674 * See "The task_lock() exception", at the end of this comment.
676 * A task must hold cgroup_mutex to modify cgroups.
678 * Any task can increment and decrement the count field without lock.
679 * So in general, code holding cgroup_mutex can't rely on the count
680 * field not changing. However, if the count goes to zero, then only
681 * cgroup_attach_task() can increment it again. Because a count of zero
682 * means that no tasks are currently attached, therefore there is no
683 * way a task attached to that cgroup can fork (the other way to
684 * increment the count). So code holding cgroup_mutex can safely
685 * assume that if the count is zero, it will stay zero. Similarly, if
686 * a task holds cgroup_mutex on a cgroup with zero count, it
687 * knows that the cgroup won't be removed, as cgroup_rmdir()
690 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
691 * (usually) take cgroup_mutex. These are the two most performance
692 * critical pieces of code here. The exception occurs on cgroup_exit(),
693 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
694 * is taken, and if the cgroup count is zero, a usermode call made
695 * to the release agent with the name of the cgroup (path relative to
696 * the root of cgroup file system) as the argument.
698 * A cgroup can only be deleted if both its 'count' of using tasks
699 * is zero, and its list of 'children' cgroups is empty. Since all
700 * tasks in the system use _some_ cgroup, and since there is always at
701 * least one task in the system (init, pid == 1), therefore, top_cgroup
702 * always has either children cgroups and/or using tasks. So we don't
703 * need a special hack to ensure that top_cgroup cannot be deleted.
705 * The task_lock() exception
707 * The need for this exception arises from the action of
708 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
709 * another. It does so using cgroup_mutex, however there are
710 * several performance critical places that need to reference
711 * task->cgroup without the expense of grabbing a system global
712 * mutex. Therefore except as noted below, when dereferencing or, as
713 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
714 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
715 * the task_struct routinely used for such matters.
717 * P.S. One more locking exception. RCU is used to guard the
718 * update of a tasks cgroup pointer by cgroup_attach_task()
722 * cgroup_lock - lock out any changes to cgroup structures
725 void cgroup_lock(void)
727 mutex_lock(&cgroup_mutex);
729 EXPORT_SYMBOL_GPL(cgroup_lock);
732 * cgroup_unlock - release lock on cgroup changes
734 * Undo the lock taken in a previous cgroup_lock() call.
736 void cgroup_unlock(void)
738 mutex_unlock(&cgroup_mutex);
740 EXPORT_SYMBOL_GPL(cgroup_unlock);
743 * A couple of forward declarations required, due to cyclic reference loop:
744 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
745 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
749 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
750 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
751 static int cgroup_populate_dir(struct cgroup *cgrp);
752 static const struct inode_operations cgroup_dir_inode_operations;
753 static const struct file_operations proc_cgroupstats_operations;
755 static struct backing_dev_info cgroup_backing_dev_info = {
757 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
760 static int alloc_css_id(struct cgroup_subsys *ss,
761 struct cgroup *parent, struct cgroup *child);
763 static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
765 struct inode *inode = new_inode(sb);
768 inode->i_mode = mode;
769 inode->i_uid = current_fsuid();
770 inode->i_gid = current_fsgid();
771 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
772 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
778 * Call subsys's pre_destroy handler.
779 * This is called before css refcnt check.
781 static int cgroup_call_pre_destroy(struct cgroup *cgrp)
783 struct cgroup_subsys *ss;
786 for_each_subsys(cgrp->root, ss)
787 if (ss->pre_destroy) {
788 ret = ss->pre_destroy(ss, cgrp);
796 static void free_cgroup_rcu(struct rcu_head *obj)
798 struct cgroup *cgrp = container_of(obj, struct cgroup, rcu_head);
803 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
805 /* is dentry a directory ? if so, kfree() associated cgroup */
806 if (S_ISDIR(inode->i_mode)) {
807 struct cgroup *cgrp = dentry->d_fsdata;
808 struct cgroup_subsys *ss;
809 BUG_ON(!(cgroup_is_removed(cgrp)));
810 /* It's possible for external users to be holding css
811 * reference counts on a cgroup; css_put() needs to
812 * be able to access the cgroup after decrementing
813 * the reference count in order to know if it needs to
814 * queue the cgroup to be handled by the release
818 mutex_lock(&cgroup_mutex);
820 * Release the subsystem state objects.
822 for_each_subsys(cgrp->root, ss)
823 ss->destroy(ss, cgrp);
825 cgrp->root->number_of_cgroups--;
826 mutex_unlock(&cgroup_mutex);
829 * Drop the active superblock reference that we took when we
832 deactivate_super(cgrp->root->sb);
835 * if we're getting rid of the cgroup, refcount should ensure
836 * that there are no pidlists left.
838 BUG_ON(!list_empty(&cgrp->pidlists));
840 call_rcu(&cgrp->rcu_head, free_cgroup_rcu);
845 static void remove_dir(struct dentry *d)
847 struct dentry *parent = dget(d->d_parent);
850 simple_rmdir(parent->d_inode, d);
854 static void cgroup_clear_directory(struct dentry *dentry)
856 struct list_head *node;
858 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
859 spin_lock(&dcache_lock);
860 node = dentry->d_subdirs.next;
861 while (node != &dentry->d_subdirs) {
862 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
865 /* This should never be called on a cgroup
866 * directory with child cgroups */
867 BUG_ON(d->d_inode->i_mode & S_IFDIR);
869 spin_unlock(&dcache_lock);
871 simple_unlink(dentry->d_inode, d);
873 spin_lock(&dcache_lock);
875 node = dentry->d_subdirs.next;
877 spin_unlock(&dcache_lock);
881 * NOTE : the dentry must have been dget()'ed
883 static void cgroup_d_remove_dir(struct dentry *dentry)
885 cgroup_clear_directory(dentry);
887 spin_lock(&dcache_lock);
888 list_del_init(&dentry->d_u.d_child);
889 spin_unlock(&dcache_lock);
894 * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
895 * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
896 * reference to css->refcnt. In general, this refcnt is expected to goes down
899 * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
901 DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
903 static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
905 if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
906 wake_up_all(&cgroup_rmdir_waitq);
909 void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
914 void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
916 cgroup_wakeup_rmdir_waiter(css->cgroup);
921 * Call with cgroup_mutex held. Drops reference counts on modules, including
922 * any duplicate ones that parse_cgroupfs_options took. If this function
923 * returns an error, no reference counts are touched.
925 static int rebind_subsystems(struct cgroupfs_root *root,
926 unsigned long final_bits)
928 unsigned long added_bits, removed_bits;
929 struct cgroup *cgrp = &root->top_cgroup;
932 BUG_ON(!mutex_is_locked(&cgroup_mutex));
934 removed_bits = root->actual_subsys_bits & ~final_bits;
935 added_bits = final_bits & ~root->actual_subsys_bits;
936 /* Check that any added subsystems are currently free */
937 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
938 unsigned long bit = 1UL << i;
939 struct cgroup_subsys *ss = subsys[i];
940 if (!(bit & added_bits))
943 * Nobody should tell us to do a subsys that doesn't exist:
944 * parse_cgroupfs_options should catch that case and refcounts
945 * ensure that subsystems won't disappear once selected.
948 if (ss->root != &rootnode) {
949 /* Subsystem isn't free */
954 /* Currently we don't handle adding/removing subsystems when
955 * any child cgroups exist. This is theoretically supportable
956 * but involves complex error handling, so it's being left until
958 if (root->number_of_cgroups > 1)
961 /* Process each subsystem */
962 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
963 struct cgroup_subsys *ss = subsys[i];
964 unsigned long bit = 1UL << i;
965 if (bit & added_bits) {
966 /* We're binding this subsystem to this hierarchy */
968 BUG_ON(cgrp->subsys[i]);
969 BUG_ON(!dummytop->subsys[i]);
970 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
971 mutex_lock(&ss->hierarchy_mutex);
972 cgrp->subsys[i] = dummytop->subsys[i];
973 cgrp->subsys[i]->cgroup = cgrp;
974 list_move(&ss->sibling, &root->subsys_list);
978 mutex_unlock(&ss->hierarchy_mutex);
979 /* refcount was already taken, and we're keeping it */
980 } else if (bit & removed_bits) {
981 /* We're removing this subsystem */
983 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
984 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
985 mutex_lock(&ss->hierarchy_mutex);
987 ss->bind(ss, dummytop);
988 dummytop->subsys[i]->cgroup = dummytop;
989 cgrp->subsys[i] = NULL;
990 subsys[i]->root = &rootnode;
991 list_move(&ss->sibling, &rootnode.subsys_list);
992 mutex_unlock(&ss->hierarchy_mutex);
993 /* subsystem is now free - drop reference on module */
994 module_put(ss->module);
995 } else if (bit & final_bits) {
996 /* Subsystem state should already exist */
998 BUG_ON(!cgrp->subsys[i]);
1000 * a refcount was taken, but we already had one, so
1001 * drop the extra reference.
1003 module_put(ss->module);
1004 #ifdef CONFIG_MODULE_UNLOAD
1005 BUG_ON(ss->module && !module_refcount(ss->module));
1008 /* Subsystem state shouldn't exist */
1009 BUG_ON(cgrp->subsys[i]);
1012 root->subsys_bits = root->actual_subsys_bits = final_bits;
1018 static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
1020 struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
1021 struct cgroup_subsys *ss;
1023 mutex_lock(&cgroup_mutex);
1024 for_each_subsys(root, ss)
1025 seq_printf(seq, ",%s", ss->name);
1026 if (test_bit(ROOT_NOPREFIX, &root->flags))
1027 seq_puts(seq, ",noprefix");
1028 if (strlen(root->release_agent_path))
1029 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1030 if (strlen(root->name))
1031 seq_printf(seq, ",name=%s", root->name);
1032 mutex_unlock(&cgroup_mutex);
1036 struct cgroup_sb_opts {
1037 unsigned long subsys_bits;
1038 unsigned long flags;
1039 char *release_agent;
1041 /* User explicitly requested empty subsystem */
1044 struct cgroupfs_root *new_root;
1049 * Convert a hierarchy specifier into a bitmask of subsystems and flags. Call
1050 * with cgroup_mutex held to protect the subsys[] array. This function takes
1051 * refcounts on subsystems to be used, unless it returns error, in which case
1052 * no refcounts are taken.
1054 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1056 char *token, *o = data ?: "all";
1057 unsigned long mask = (unsigned long)-1;
1059 bool module_pin_failed = false;
1061 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1063 #ifdef CONFIG_CPUSETS
1064 mask = ~(1UL << cpuset_subsys_id);
1067 memset(opts, 0, sizeof(*opts));
1069 while ((token = strsep(&o, ",")) != NULL) {
1072 if (!strcmp(token, "all")) {
1073 /* Add all non-disabled subsystems */
1074 opts->subsys_bits = 0;
1075 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1076 struct cgroup_subsys *ss = subsys[i];
1080 opts->subsys_bits |= 1ul << i;
1082 } else if (!strcmp(token, "none")) {
1083 /* Explicitly have no subsystems */
1085 } else if (!strcmp(token, "noprefix")) {
1086 set_bit(ROOT_NOPREFIX, &opts->flags);
1087 } else if (!strncmp(token, "release_agent=", 14)) {
1088 /* Specifying two release agents is forbidden */
1089 if (opts->release_agent)
1091 opts->release_agent =
1092 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1093 if (!opts->release_agent)
1095 } else if (!strncmp(token, "name=", 5)) {
1096 const char *name = token + 5;
1097 /* Can't specify an empty name */
1100 /* Must match [\w.-]+ */
1101 for (i = 0; i < strlen(name); i++) {
1105 if ((c == '.') || (c == '-') || (c == '_'))
1109 /* Specifying two names is forbidden */
1112 opts->name = kstrndup(name,
1113 MAX_CGROUP_ROOT_NAMELEN - 1,
1118 struct cgroup_subsys *ss;
1119 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1123 if (!strcmp(token, ss->name)) {
1125 set_bit(i, &opts->subsys_bits);
1129 if (i == CGROUP_SUBSYS_COUNT)
1134 /* Consistency checks */
1137 * Option noprefix was introduced just for backward compatibility
1138 * with the old cpuset, so we allow noprefix only if mounting just
1139 * the cpuset subsystem.
1141 if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
1142 (opts->subsys_bits & mask))
1146 /* Can't specify "none" and some subsystems */
1147 if (opts->subsys_bits && opts->none)
1151 * We either have to specify by name or by subsystems. (So all
1152 * empty hierarchies must have a name).
1154 if (!opts->subsys_bits && !opts->name)
1158 * Grab references on all the modules we'll need, so the subsystems
1159 * don't dance around before rebind_subsystems attaches them. This may
1160 * take duplicate reference counts on a subsystem that's already used,
1161 * but rebind_subsystems handles this case.
1163 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1164 unsigned long bit = 1UL << i;
1166 if (!(bit & opts->subsys_bits))
1168 if (!try_module_get(subsys[i]->module)) {
1169 module_pin_failed = true;
1173 if (module_pin_failed) {
1175 * oops, one of the modules was going away. this means that we
1176 * raced with a module_delete call, and to the user this is
1177 * essentially a "subsystem doesn't exist" case.
1179 for (i--; i >= CGROUP_BUILTIN_SUBSYS_COUNT; i--) {
1180 /* drop refcounts only on the ones we took */
1181 unsigned long bit = 1UL << i;
1183 if (!(bit & opts->subsys_bits))
1185 module_put(subsys[i]->module);
1193 static void drop_parsed_module_refcounts(unsigned long subsys_bits)
1196 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1197 unsigned long bit = 1UL << i;
1199 if (!(bit & subsys_bits))
1201 module_put(subsys[i]->module);
1205 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1208 struct cgroupfs_root *root = sb->s_fs_info;
1209 struct cgroup *cgrp = &root->top_cgroup;
1210 struct cgroup_sb_opts opts;
1213 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1214 mutex_lock(&cgroup_mutex);
1216 /* See what subsystems are wanted */
1217 ret = parse_cgroupfs_options(data, &opts);
1221 /* Don't allow flags or name to change at remount */
1222 if (opts.flags != root->flags ||
1223 (opts.name && strcmp(opts.name, root->name))) {
1225 drop_parsed_module_refcounts(opts.subsys_bits);
1229 ret = rebind_subsystems(root, opts.subsys_bits);
1231 drop_parsed_module_refcounts(opts.subsys_bits);
1235 /* (re)populate subsystem files */
1236 cgroup_populate_dir(cgrp);
1238 if (opts.release_agent)
1239 strcpy(root->release_agent_path, opts.release_agent);
1241 kfree(opts.release_agent);
1243 mutex_unlock(&cgroup_mutex);
1244 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1249 static const struct super_operations cgroup_ops = {
1250 .statfs = simple_statfs,
1251 .drop_inode = generic_delete_inode,
1252 .show_options = cgroup_show_options,
1253 .remount_fs = cgroup_remount,
1256 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1258 INIT_LIST_HEAD(&cgrp->sibling);
1259 INIT_LIST_HEAD(&cgrp->children);
1260 INIT_LIST_HEAD(&cgrp->css_sets);
1261 INIT_LIST_HEAD(&cgrp->release_list);
1262 INIT_LIST_HEAD(&cgrp->pidlists);
1263 mutex_init(&cgrp->pidlist_mutex);
1264 INIT_LIST_HEAD(&cgrp->event_list);
1265 spin_lock_init(&cgrp->event_list_lock);
1268 static void init_cgroup_root(struct cgroupfs_root *root)
1270 struct cgroup *cgrp = &root->top_cgroup;
1271 INIT_LIST_HEAD(&root->subsys_list);
1272 INIT_LIST_HEAD(&root->root_list);
1273 root->number_of_cgroups = 1;
1275 cgrp->top_cgroup = cgrp;
1276 init_cgroup_housekeeping(cgrp);
1279 static bool init_root_id(struct cgroupfs_root *root)
1284 if (!ida_pre_get(&hierarchy_ida, GFP_KERNEL))
1286 spin_lock(&hierarchy_id_lock);
1287 /* Try to allocate the next unused ID */
1288 ret = ida_get_new_above(&hierarchy_ida, next_hierarchy_id,
1289 &root->hierarchy_id);
1291 /* Try again starting from 0 */
1292 ret = ida_get_new(&hierarchy_ida, &root->hierarchy_id);
1294 next_hierarchy_id = root->hierarchy_id + 1;
1295 } else if (ret != -EAGAIN) {
1296 /* Can only get here if the 31-bit IDR is full ... */
1299 spin_unlock(&hierarchy_id_lock);
1304 static int cgroup_test_super(struct super_block *sb, void *data)
1306 struct cgroup_sb_opts *opts = data;
1307 struct cgroupfs_root *root = sb->s_fs_info;
1309 /* If we asked for a name then it must match */
1310 if (opts->name && strcmp(opts->name, root->name))
1314 * If we asked for subsystems (or explicitly for no
1315 * subsystems) then they must match
1317 if ((opts->subsys_bits || opts->none)
1318 && (opts->subsys_bits != root->subsys_bits))
1324 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1326 struct cgroupfs_root *root;
1328 if (!opts->subsys_bits && !opts->none)
1331 root = kzalloc(sizeof(*root), GFP_KERNEL);
1333 return ERR_PTR(-ENOMEM);
1335 if (!init_root_id(root)) {
1337 return ERR_PTR(-ENOMEM);
1339 init_cgroup_root(root);
1341 root->subsys_bits = opts->subsys_bits;
1342 root->flags = opts->flags;
1343 if (opts->release_agent)
1344 strcpy(root->release_agent_path, opts->release_agent);
1346 strcpy(root->name, opts->name);
1350 static void cgroup_drop_root(struct cgroupfs_root *root)
1355 BUG_ON(!root->hierarchy_id);
1356 spin_lock(&hierarchy_id_lock);
1357 ida_remove(&hierarchy_ida, root->hierarchy_id);
1358 spin_unlock(&hierarchy_id_lock);
1362 static int cgroup_set_super(struct super_block *sb, void *data)
1365 struct cgroup_sb_opts *opts = data;
1367 /* If we don't have a new root, we can't set up a new sb */
1368 if (!opts->new_root)
1371 BUG_ON(!opts->subsys_bits && !opts->none);
1373 ret = set_anon_super(sb, NULL);
1377 sb->s_fs_info = opts->new_root;
1378 opts->new_root->sb = sb;
1380 sb->s_blocksize = PAGE_CACHE_SIZE;
1381 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1382 sb->s_magic = CGROUP_SUPER_MAGIC;
1383 sb->s_op = &cgroup_ops;
1388 static int cgroup_get_rootdir(struct super_block *sb)
1390 struct inode *inode =
1391 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1392 struct dentry *dentry;
1397 inode->i_fop = &simple_dir_operations;
1398 inode->i_op = &cgroup_dir_inode_operations;
1399 /* directories start off with i_nlink == 2 (for "." entry) */
1401 dentry = d_alloc_root(inode);
1406 sb->s_root = dentry;
1410 static int cgroup_get_sb(struct file_system_type *fs_type,
1411 int flags, const char *unused_dev_name,
1412 void *data, struct vfsmount *mnt)
1414 struct cgroup_sb_opts opts;
1415 struct cgroupfs_root *root;
1417 struct super_block *sb;
1418 struct cgroupfs_root *new_root;
1420 /* First find the desired set of subsystems */
1421 mutex_lock(&cgroup_mutex);
1422 ret = parse_cgroupfs_options(data, &opts);
1423 mutex_unlock(&cgroup_mutex);
1428 * Allocate a new cgroup root. We may not need it if we're
1429 * reusing an existing hierarchy.
1431 new_root = cgroup_root_from_opts(&opts);
1432 if (IS_ERR(new_root)) {
1433 ret = PTR_ERR(new_root);
1436 opts.new_root = new_root;
1438 /* Locate an existing or new sb for this hierarchy */
1439 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
1442 cgroup_drop_root(opts.new_root);
1446 root = sb->s_fs_info;
1448 if (root == opts.new_root) {
1449 /* We used the new root structure, so this is a new hierarchy */
1450 struct list_head tmp_cg_links;
1451 struct cgroup *root_cgrp = &root->top_cgroup;
1452 struct inode *inode;
1453 struct cgroupfs_root *existing_root;
1456 BUG_ON(sb->s_root != NULL);
1458 ret = cgroup_get_rootdir(sb);
1460 goto drop_new_super;
1461 inode = sb->s_root->d_inode;
1463 mutex_lock(&inode->i_mutex);
1464 mutex_lock(&cgroup_mutex);
1466 if (strlen(root->name)) {
1467 /* Check for name clashes with existing mounts */
1468 for_each_active_root(existing_root) {
1469 if (!strcmp(existing_root->name, root->name)) {
1471 mutex_unlock(&cgroup_mutex);
1472 mutex_unlock(&inode->i_mutex);
1473 goto drop_new_super;
1479 * We're accessing css_set_count without locking
1480 * css_set_lock here, but that's OK - it can only be
1481 * increased by someone holding cgroup_lock, and
1482 * that's us. The worst that can happen is that we
1483 * have some link structures left over
1485 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1487 mutex_unlock(&cgroup_mutex);
1488 mutex_unlock(&inode->i_mutex);
1489 goto drop_new_super;
1492 ret = rebind_subsystems(root, root->subsys_bits);
1493 if (ret == -EBUSY) {
1494 mutex_unlock(&cgroup_mutex);
1495 mutex_unlock(&inode->i_mutex);
1496 free_cg_links(&tmp_cg_links);
1497 goto drop_new_super;
1500 * There must be no failure case after here, since rebinding
1501 * takes care of subsystems' refcounts, which are explicitly
1502 * dropped in the failure exit path.
1505 /* EBUSY should be the only error here */
1508 list_add(&root->root_list, &roots);
1511 sb->s_root->d_fsdata = root_cgrp;
1512 root->top_cgroup.dentry = sb->s_root;
1514 /* Link the top cgroup in this hierarchy into all
1515 * the css_set objects */
1516 write_lock(&css_set_lock);
1517 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1518 struct hlist_head *hhead = &css_set_table[i];
1519 struct hlist_node *node;
1522 hlist_for_each_entry(cg, node, hhead, hlist)
1523 link_css_set(&tmp_cg_links, cg, root_cgrp);
1525 write_unlock(&css_set_lock);
1527 free_cg_links(&tmp_cg_links);
1529 BUG_ON(!list_empty(&root_cgrp->sibling));
1530 BUG_ON(!list_empty(&root_cgrp->children));
1531 BUG_ON(root->number_of_cgroups != 1);
1533 cgroup_populate_dir(root_cgrp);
1534 mutex_unlock(&cgroup_mutex);
1535 mutex_unlock(&inode->i_mutex);
1538 * We re-used an existing hierarchy - the new root (if
1539 * any) is not needed
1541 cgroup_drop_root(opts.new_root);
1542 /* no subsys rebinding, so refcounts don't change */
1543 drop_parsed_module_refcounts(opts.subsys_bits);
1546 simple_set_mnt(mnt, sb);
1547 kfree(opts.release_agent);
1552 deactivate_locked_super(sb);
1554 drop_parsed_module_refcounts(opts.subsys_bits);
1556 kfree(opts.release_agent);
1562 static void cgroup_kill_sb(struct super_block *sb) {
1563 struct cgroupfs_root *root = sb->s_fs_info;
1564 struct cgroup *cgrp = &root->top_cgroup;
1566 struct cg_cgroup_link *link;
1567 struct cg_cgroup_link *saved_link;
1571 BUG_ON(root->number_of_cgroups != 1);
1572 BUG_ON(!list_empty(&cgrp->children));
1573 BUG_ON(!list_empty(&cgrp->sibling));
1575 mutex_lock(&cgroup_mutex);
1577 /* Rebind all subsystems back to the default hierarchy */
1578 ret = rebind_subsystems(root, 0);
1579 /* Shouldn't be able to fail ... */
1583 * Release all the links from css_sets to this hierarchy's
1586 write_lock(&css_set_lock);
1588 list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1590 list_del(&link->cg_link_list);
1591 list_del(&link->cgrp_link_list);
1594 write_unlock(&css_set_lock);
1596 if (!list_empty(&root->root_list)) {
1597 list_del(&root->root_list);
1601 mutex_unlock(&cgroup_mutex);
1603 kill_litter_super(sb);
1604 cgroup_drop_root(root);
1607 static struct file_system_type cgroup_fs_type = {
1609 .get_sb = cgroup_get_sb,
1610 .kill_sb = cgroup_kill_sb,
1613 static struct kobject *cgroup_kobj;
1615 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1617 return dentry->d_fsdata;
1620 static inline struct cftype *__d_cft(struct dentry *dentry)
1622 return dentry->d_fsdata;
1626 * cgroup_path - generate the path of a cgroup
1627 * @cgrp: the cgroup in question
1628 * @buf: the buffer to write the path into
1629 * @buflen: the length of the buffer
1631 * Called with cgroup_mutex held or else with an RCU-protected cgroup
1632 * reference. Writes path of cgroup into buf. Returns 0 on success,
1635 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1638 struct dentry *dentry = rcu_dereference_check(cgrp->dentry,
1639 rcu_read_lock_held() ||
1640 cgroup_lock_is_held());
1642 if (!dentry || cgrp == dummytop) {
1644 * Inactive subsystems have no dentry for their root
1651 start = buf + buflen;
1655 int len = dentry->d_name.len;
1657 if ((start -= len) < buf)
1658 return -ENAMETOOLONG;
1659 memcpy(start, dentry->d_name.name, len);
1660 cgrp = cgrp->parent;
1664 dentry = rcu_dereference_check(cgrp->dentry,
1665 rcu_read_lock_held() ||
1666 cgroup_lock_is_held());
1670 return -ENAMETOOLONG;
1673 memmove(buf, start, buf + buflen - start);
1676 EXPORT_SYMBOL_GPL(cgroup_path);
1679 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1680 * @cgrp: the cgroup the task is attaching to
1681 * @tsk: the task to be attached
1683 * Call holding cgroup_mutex. May take task_lock of
1684 * the task 'tsk' during call.
1686 int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1689 struct cgroup_subsys *ss, *failed_ss = NULL;
1690 struct cgroup *oldcgrp;
1692 struct css_set *newcg;
1693 struct cgroupfs_root *root = cgrp->root;
1695 /* Nothing to do if the task is already in that cgroup */
1696 oldcgrp = task_cgroup_from_root(tsk, root);
1697 if (cgrp == oldcgrp)
1700 for_each_subsys(root, ss) {
1701 if (ss->can_attach) {
1702 retval = ss->can_attach(ss, cgrp, tsk, false);
1705 * Remember on which subsystem the can_attach()
1706 * failed, so that we only call cancel_attach()
1707 * against the subsystems whose can_attach()
1708 * succeeded. (See below)
1713 } else if (!capable(CAP_SYS_ADMIN)) {
1714 const struct cred *cred = current_cred(), *tcred;
1716 /* No can_attach() - check perms generically */
1717 tcred = __task_cred(tsk);
1718 if (cred->euid != tcred->uid &&
1719 cred->euid != tcred->suid) {
1730 * Locate or allocate a new css_set for this task,
1731 * based on its final set of cgroups
1733 newcg = find_css_set(cg, cgrp);
1741 if (tsk->flags & PF_EXITING) {
1747 rcu_assign_pointer(tsk->cgroups, newcg);
1750 /* Update the css_set linked lists if we're using them */
1751 write_lock(&css_set_lock);
1752 if (!list_empty(&tsk->cg_list)) {
1753 list_del(&tsk->cg_list);
1754 list_add(&tsk->cg_list, &newcg->tasks);
1756 write_unlock(&css_set_lock);
1758 for_each_subsys(root, ss) {
1760 ss->attach(ss, cgrp, oldcgrp, tsk, false);
1762 set_bit(CGRP_RELEASABLE, &cgrp->flags);
1767 * wake up rmdir() waiter. the rmdir should fail since the cgroup
1768 * is no longer empty.
1770 cgroup_wakeup_rmdir_waiter(cgrp);
1773 for_each_subsys(root, ss) {
1774 if (ss == failed_ss)
1776 * This subsystem was the one that failed the
1777 * can_attach() check earlier, so we don't need
1778 * to call cancel_attach() against it or any
1779 * remaining subsystems.
1782 if (ss->cancel_attach)
1783 ss->cancel_attach(ss, cgrp, tsk, false);
1790 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1791 * @from: attach to all cgroups of a given task
1792 * @tsk: the task to be attached
1794 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1796 struct cgroupfs_root *root;
1800 for_each_active_root(root) {
1801 struct cgroup *from_cg = task_cgroup_from_root(from, root);
1803 retval = cgroup_attach_task(from_cg, tsk);
1811 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
1814 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
1815 * held. May take task_lock of task
1817 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid)
1819 struct task_struct *tsk;
1824 tsk = find_task_by_vpid(pid);
1825 if (!tsk || tsk->flags & PF_EXITING) {
1829 get_task_struct(tsk);
1833 get_task_struct(tsk);
1836 ret = cgroup_attach_task(cgrp, tsk);
1837 put_task_struct(tsk);
1841 static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
1844 if (!cgroup_lock_live_group(cgrp))
1846 ret = attach_task_by_pid(cgrp, pid);
1852 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
1853 * @cgrp: the cgroup to be checked for liveness
1855 * On success, returns true; the lock should be later released with
1856 * cgroup_unlock(). On failure returns false with no lock held.
1858 bool cgroup_lock_live_group(struct cgroup *cgrp)
1860 mutex_lock(&cgroup_mutex);
1861 if (cgroup_is_removed(cgrp)) {
1862 mutex_unlock(&cgroup_mutex);
1867 EXPORT_SYMBOL_GPL(cgroup_lock_live_group);
1869 static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
1872 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1873 if (!cgroup_lock_live_group(cgrp))
1875 strcpy(cgrp->root->release_agent_path, buffer);
1880 static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
1881 struct seq_file *seq)
1883 if (!cgroup_lock_live_group(cgrp))
1885 seq_puts(seq, cgrp->root->release_agent_path);
1886 seq_putc(seq, '\n');
1891 /* A buffer size big enough for numbers or short strings */
1892 #define CGROUP_LOCAL_BUFFER_SIZE 64
1894 static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
1896 const char __user *userbuf,
1897 size_t nbytes, loff_t *unused_ppos)
1899 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
1905 if (nbytes >= sizeof(buffer))
1907 if (copy_from_user(buffer, userbuf, nbytes))
1910 buffer[nbytes] = 0; /* nul-terminate */
1911 if (cft->write_u64) {
1912 u64 val = simple_strtoull(strstrip(buffer), &end, 0);
1915 retval = cft->write_u64(cgrp, cft, val);
1917 s64 val = simple_strtoll(strstrip(buffer), &end, 0);
1920 retval = cft->write_s64(cgrp, cft, val);
1927 static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
1929 const char __user *userbuf,
1930 size_t nbytes, loff_t *unused_ppos)
1932 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
1934 size_t max_bytes = cft->max_write_len;
1935 char *buffer = local_buffer;
1938 max_bytes = sizeof(local_buffer) - 1;
1939 if (nbytes >= max_bytes)
1941 /* Allocate a dynamic buffer if we need one */
1942 if (nbytes >= sizeof(local_buffer)) {
1943 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
1947 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
1952 buffer[nbytes] = 0; /* nul-terminate */
1953 retval = cft->write_string(cgrp, cft, strstrip(buffer));
1957 if (buffer != local_buffer)
1962 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
1963 size_t nbytes, loff_t *ppos)
1965 struct cftype *cft = __d_cft(file->f_dentry);
1966 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1968 if (cgroup_is_removed(cgrp))
1971 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
1972 if (cft->write_u64 || cft->write_s64)
1973 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
1974 if (cft->write_string)
1975 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
1977 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
1978 return ret ? ret : nbytes;
1983 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
1985 char __user *buf, size_t nbytes,
1988 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
1989 u64 val = cft->read_u64(cgrp, cft);
1990 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
1992 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1995 static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
1997 char __user *buf, size_t nbytes,
2000 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2001 s64 val = cft->read_s64(cgrp, cft);
2002 int len = sprintf(tmp, "%lld\n", (long long) val);
2004 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2007 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2008 size_t nbytes, loff_t *ppos)
2010 struct cftype *cft = __d_cft(file->f_dentry);
2011 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2013 if (cgroup_is_removed(cgrp))
2017 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2019 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2021 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2026 * seqfile ops/methods for returning structured data. Currently just
2027 * supports string->u64 maps, but can be extended in future.
2030 struct cgroup_seqfile_state {
2032 struct cgroup *cgroup;
2035 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2037 struct seq_file *sf = cb->state;
2038 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2041 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2043 struct cgroup_seqfile_state *state = m->private;
2044 struct cftype *cft = state->cft;
2045 if (cft->read_map) {
2046 struct cgroup_map_cb cb = {
2047 .fill = cgroup_map_add,
2050 return cft->read_map(state->cgroup, cft, &cb);
2052 return cft->read_seq_string(state->cgroup, cft, m);
2055 static int cgroup_seqfile_release(struct inode *inode, struct file *file)
2057 struct seq_file *seq = file->private_data;
2058 kfree(seq->private);
2059 return single_release(inode, file);
2062 static const struct file_operations cgroup_seqfile_operations = {
2064 .write = cgroup_file_write,
2065 .llseek = seq_lseek,
2066 .release = cgroup_seqfile_release,
2069 static int cgroup_file_open(struct inode *inode, struct file *file)
2074 err = generic_file_open(inode, file);
2077 cft = __d_cft(file->f_dentry);
2079 if (cft->read_map || cft->read_seq_string) {
2080 struct cgroup_seqfile_state *state =
2081 kzalloc(sizeof(*state), GFP_USER);
2085 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
2086 file->f_op = &cgroup_seqfile_operations;
2087 err = single_open(file, cgroup_seqfile_show, state);
2090 } else if (cft->open)
2091 err = cft->open(inode, file);
2098 static int cgroup_file_release(struct inode *inode, struct file *file)
2100 struct cftype *cft = __d_cft(file->f_dentry);
2102 return cft->release(inode, file);
2107 * cgroup_rename - Only allow simple rename of directories in place.
2109 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2110 struct inode *new_dir, struct dentry *new_dentry)
2112 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2114 if (new_dentry->d_inode)
2116 if (old_dir != new_dir)
2118 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2121 static const struct file_operations cgroup_file_operations = {
2122 .read = cgroup_file_read,
2123 .write = cgroup_file_write,
2124 .llseek = generic_file_llseek,
2125 .open = cgroup_file_open,
2126 .release = cgroup_file_release,
2129 static const struct inode_operations cgroup_dir_inode_operations = {
2130 .lookup = simple_lookup,
2131 .mkdir = cgroup_mkdir,
2132 .rmdir = cgroup_rmdir,
2133 .rename = cgroup_rename,
2137 * Check if a file is a control file
2139 static inline struct cftype *__file_cft(struct file *file)
2141 if (file->f_dentry->d_inode->i_fop != &cgroup_file_operations)
2142 return ERR_PTR(-EINVAL);
2143 return __d_cft(file->f_dentry);
2146 static int cgroup_create_file(struct dentry *dentry, mode_t mode,
2147 struct super_block *sb)
2149 static const struct dentry_operations cgroup_dops = {
2150 .d_iput = cgroup_diput,
2153 struct inode *inode;
2157 if (dentry->d_inode)
2160 inode = cgroup_new_inode(mode, sb);
2164 if (S_ISDIR(mode)) {
2165 inode->i_op = &cgroup_dir_inode_operations;
2166 inode->i_fop = &simple_dir_operations;
2168 /* start off with i_nlink == 2 (for "." entry) */
2171 /* start with the directory inode held, so that we can
2172 * populate it without racing with another mkdir */
2173 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2174 } else if (S_ISREG(mode)) {
2176 inode->i_fop = &cgroup_file_operations;
2178 dentry->d_op = &cgroup_dops;
2179 d_instantiate(dentry, inode);
2180 dget(dentry); /* Extra count - pin the dentry in core */
2185 * cgroup_create_dir - create a directory for an object.
2186 * @cgrp: the cgroup we create the directory for. It must have a valid
2187 * ->parent field. And we are going to fill its ->dentry field.
2188 * @dentry: dentry of the new cgroup
2189 * @mode: mode to set on new directory.
2191 static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
2194 struct dentry *parent;
2197 parent = cgrp->parent->dentry;
2198 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
2200 dentry->d_fsdata = cgrp;
2201 inc_nlink(parent->d_inode);
2202 rcu_assign_pointer(cgrp->dentry, dentry);
2211 * cgroup_file_mode - deduce file mode of a control file
2212 * @cft: the control file in question
2214 * returns cft->mode if ->mode is not 0
2215 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2216 * returns S_IRUGO if it has only a read handler
2217 * returns S_IWUSR if it has only a write hander
2219 static mode_t cgroup_file_mode(const struct cftype *cft)
2226 if (cft->read || cft->read_u64 || cft->read_s64 ||
2227 cft->read_map || cft->read_seq_string)
2230 if (cft->write || cft->write_u64 || cft->write_s64 ||
2231 cft->write_string || cft->trigger)
2237 int cgroup_add_file(struct cgroup *cgrp,
2238 struct cgroup_subsys *subsys,
2239 const struct cftype *cft)
2241 struct dentry *dir = cgrp->dentry;
2242 struct dentry *dentry;
2246 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2247 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
2248 strcpy(name, subsys->name);
2251 strcat(name, cft->name);
2252 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2253 dentry = lookup_one_len(name, dir, strlen(name));
2254 if (!IS_ERR(dentry)) {
2255 mode = cgroup_file_mode(cft);
2256 error = cgroup_create_file(dentry, mode | S_IFREG,
2259 dentry->d_fsdata = (void *)cft;
2262 error = PTR_ERR(dentry);
2265 EXPORT_SYMBOL_GPL(cgroup_add_file);
2267 int cgroup_add_files(struct cgroup *cgrp,
2268 struct cgroup_subsys *subsys,
2269 const struct cftype cft[],
2273 for (i = 0; i < count; i++) {
2274 err = cgroup_add_file(cgrp, subsys, &cft[i]);
2280 EXPORT_SYMBOL_GPL(cgroup_add_files);
2283 * cgroup_task_count - count the number of tasks in a cgroup.
2284 * @cgrp: the cgroup in question
2286 * Return the number of tasks in the cgroup.
2288 int cgroup_task_count(const struct cgroup *cgrp)
2291 struct cg_cgroup_link *link;
2293 read_lock(&css_set_lock);
2294 list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
2295 count += atomic_read(&link->cg->refcount);
2297 read_unlock(&css_set_lock);
2302 * Advance a list_head iterator. The iterator should be positioned at
2303 * the start of a css_set
2305 static void cgroup_advance_iter(struct cgroup *cgrp,
2306 struct cgroup_iter *it)
2308 struct list_head *l = it->cg_link;
2309 struct cg_cgroup_link *link;
2312 /* Advance to the next non-empty css_set */
2315 if (l == &cgrp->css_sets) {
2319 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
2321 } while (list_empty(&cg->tasks));
2323 it->task = cg->tasks.next;
2327 * To reduce the fork() overhead for systems that are not actually
2328 * using their cgroups capability, we don't maintain the lists running
2329 * through each css_set to its tasks until we see the list actually
2330 * used - in other words after the first call to cgroup_iter_start().
2332 * The tasklist_lock is not held here, as do_each_thread() and
2333 * while_each_thread() are protected by RCU.
2335 static void cgroup_enable_task_cg_lists(void)
2337 struct task_struct *p, *g;
2338 write_lock(&css_set_lock);
2339 use_task_css_set_links = 1;
2340 do_each_thread(g, p) {
2343 * We should check if the process is exiting, otherwise
2344 * it will race with cgroup_exit() in that the list
2345 * entry won't be deleted though the process has exited.
2347 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2348 list_add(&p->cg_list, &p->cgroups->tasks);
2350 } while_each_thread(g, p);
2351 write_unlock(&css_set_lock);
2354 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
2357 * The first time anyone tries to iterate across a cgroup,
2358 * we need to enable the list linking each css_set to its
2359 * tasks, and fix up all existing tasks.
2361 if (!use_task_css_set_links)
2362 cgroup_enable_task_cg_lists();
2364 read_lock(&css_set_lock);
2365 it->cg_link = &cgrp->css_sets;
2366 cgroup_advance_iter(cgrp, it);
2369 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
2370 struct cgroup_iter *it)
2372 struct task_struct *res;
2373 struct list_head *l = it->task;
2374 struct cg_cgroup_link *link;
2376 /* If the iterator cg is NULL, we have no tasks */
2379 res = list_entry(l, struct task_struct, cg_list);
2380 /* Advance iterator to find next entry */
2382 link = list_entry(it->cg_link, struct cg_cgroup_link, cgrp_link_list);
2383 if (l == &link->cg->tasks) {
2384 /* We reached the end of this task list - move on to
2385 * the next cg_cgroup_link */
2386 cgroup_advance_iter(cgrp, it);
2393 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
2395 read_unlock(&css_set_lock);
2398 static inline int started_after_time(struct task_struct *t1,
2399 struct timespec *time,
2400 struct task_struct *t2)
2402 int start_diff = timespec_compare(&t1->start_time, time);
2403 if (start_diff > 0) {
2405 } else if (start_diff < 0) {
2409 * Arbitrarily, if two processes started at the same
2410 * time, we'll say that the lower pointer value
2411 * started first. Note that t2 may have exited by now
2412 * so this may not be a valid pointer any longer, but
2413 * that's fine - it still serves to distinguish
2414 * between two tasks started (effectively) simultaneously.
2421 * This function is a callback from heap_insert() and is used to order
2423 * In this case we order the heap in descending task start time.
2425 static inline int started_after(void *p1, void *p2)
2427 struct task_struct *t1 = p1;
2428 struct task_struct *t2 = p2;
2429 return started_after_time(t1, &t2->start_time, t2);
2433 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
2434 * @scan: struct cgroup_scanner containing arguments for the scan
2436 * Arguments include pointers to callback functions test_task() and
2438 * Iterate through all the tasks in a cgroup, calling test_task() for each,
2439 * and if it returns true, call process_task() for it also.
2440 * The test_task pointer may be NULL, meaning always true (select all tasks).
2441 * Effectively duplicates cgroup_iter_{start,next,end}()
2442 * but does not lock css_set_lock for the call to process_task().
2443 * The struct cgroup_scanner may be embedded in any structure of the caller's
2445 * It is guaranteed that process_task() will act on every task that
2446 * is a member of the cgroup for the duration of this call. This
2447 * function may or may not call process_task() for tasks that exit
2448 * or move to a different cgroup during the call, or are forked or
2449 * move into the cgroup during the call.
2451 * Note that test_task() may be called with locks held, and may in some
2452 * situations be called multiple times for the same task, so it should
2454 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
2455 * pre-allocated and will be used for heap operations (and its "gt" member will
2456 * be overwritten), else a temporary heap will be used (allocation of which
2457 * may cause this function to fail).
2459 int cgroup_scan_tasks(struct cgroup_scanner *scan)
2462 struct cgroup_iter it;
2463 struct task_struct *p, *dropped;
2464 /* Never dereference latest_task, since it's not refcounted */
2465 struct task_struct *latest_task = NULL;
2466 struct ptr_heap tmp_heap;
2467 struct ptr_heap *heap;
2468 struct timespec latest_time = { 0, 0 };
2471 /* The caller supplied our heap and pre-allocated its memory */
2473 heap->gt = &started_after;
2475 /* We need to allocate our own heap memory */
2477 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2479 /* cannot allocate the heap */
2485 * Scan tasks in the cgroup, using the scanner's "test_task" callback
2486 * to determine which are of interest, and using the scanner's
2487 * "process_task" callback to process any of them that need an update.
2488 * Since we don't want to hold any locks during the task updates,
2489 * gather tasks to be processed in a heap structure.
2490 * The heap is sorted by descending task start time.
2491 * If the statically-sized heap fills up, we overflow tasks that
2492 * started later, and in future iterations only consider tasks that
2493 * started after the latest task in the previous pass. This
2494 * guarantees forward progress and that we don't miss any tasks.
2497 cgroup_iter_start(scan->cg, &it);
2498 while ((p = cgroup_iter_next(scan->cg, &it))) {
2500 * Only affect tasks that qualify per the caller's callback,
2501 * if he provided one
2503 if (scan->test_task && !scan->test_task(p, scan))
2506 * Only process tasks that started after the last task
2509 if (!started_after_time(p, &latest_time, latest_task))
2511 dropped = heap_insert(heap, p);
2512 if (dropped == NULL) {
2514 * The new task was inserted; the heap wasn't
2518 } else if (dropped != p) {
2520 * The new task was inserted, and pushed out a
2524 put_task_struct(dropped);
2527 * Else the new task was newer than anything already in
2528 * the heap and wasn't inserted
2531 cgroup_iter_end(scan->cg, &it);
2534 for (i = 0; i < heap->size; i++) {
2535 struct task_struct *q = heap->ptrs[i];
2537 latest_time = q->start_time;
2540 /* Process the task per the caller's callback */
2541 scan->process_task(q, scan);
2545 * If we had to process any tasks at all, scan again
2546 * in case some of them were in the middle of forking
2547 * children that didn't get processed.
2548 * Not the most efficient way to do it, but it avoids
2549 * having to take callback_mutex in the fork path
2553 if (heap == &tmp_heap)
2554 heap_free(&tmp_heap);
2559 * Stuff for reading the 'tasks'/'procs' files.
2561 * Reading this file can return large amounts of data if a cgroup has
2562 * *lots* of attached tasks. So it may need several calls to read(),
2563 * but we cannot guarantee that the information we produce is correct
2564 * unless we produce it entirely atomically.
2569 * The following two functions "fix" the issue where there are more pids
2570 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2571 * TODO: replace with a kernel-wide solution to this problem
2573 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2574 static void *pidlist_allocate(int count)
2576 if (PIDLIST_TOO_LARGE(count))
2577 return vmalloc(count * sizeof(pid_t));
2579 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2581 static void pidlist_free(void *p)
2583 if (is_vmalloc_addr(p))
2588 static void *pidlist_resize(void *p, int newcount)
2591 /* note: if new alloc fails, old p will still be valid either way */
2592 if (is_vmalloc_addr(p)) {
2593 newlist = vmalloc(newcount * sizeof(pid_t));
2596 memcpy(newlist, p, newcount * sizeof(pid_t));
2599 newlist = krealloc(p, newcount * sizeof(pid_t), GFP_KERNEL);
2605 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2606 * If the new stripped list is sufficiently smaller and there's enough memory
2607 * to allocate a new buffer, will let go of the unneeded memory. Returns the
2608 * number of unique elements.
2610 /* is the size difference enough that we should re-allocate the array? */
2611 #define PIDLIST_REALLOC_DIFFERENCE(old, new) ((old) - PAGE_SIZE >= (new))
2612 static int pidlist_uniq(pid_t **p, int length)
2619 * we presume the 0th element is unique, so i starts at 1. trivial
2620 * edge cases first; no work needs to be done for either
2622 if (length == 0 || length == 1)
2624 /* src and dest walk down the list; dest counts unique elements */
2625 for (src = 1; src < length; src++) {
2626 /* find next unique element */
2627 while (list[src] == list[src-1]) {
2632 /* dest always points to where the next unique element goes */
2633 list[dest] = list[src];
2638 * if the length difference is large enough, we want to allocate a
2639 * smaller buffer to save memory. if this fails due to out of memory,
2640 * we'll just stay with what we've got.
2642 if (PIDLIST_REALLOC_DIFFERENCE(length, dest)) {
2643 newlist = pidlist_resize(list, dest);
2650 static int cmppid(const void *a, const void *b)
2652 return *(pid_t *)a - *(pid_t *)b;
2656 * find the appropriate pidlist for our purpose (given procs vs tasks)
2657 * returns with the lock on that pidlist already held, and takes care
2658 * of the use count, or returns NULL with no locks held if we're out of
2661 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
2662 enum cgroup_filetype type)
2664 struct cgroup_pidlist *l;
2665 /* don't need task_nsproxy() if we're looking at ourself */
2666 struct pid_namespace *ns = current->nsproxy->pid_ns;
2669 * We can't drop the pidlist_mutex before taking the l->mutex in case
2670 * the last ref-holder is trying to remove l from the list at the same
2671 * time. Holding the pidlist_mutex precludes somebody taking whichever
2672 * list we find out from under us - compare release_pid_array().
2674 mutex_lock(&cgrp->pidlist_mutex);
2675 list_for_each_entry(l, &cgrp->pidlists, links) {
2676 if (l->key.type == type && l->key.ns == ns) {
2677 /* make sure l doesn't vanish out from under us */
2678 down_write(&l->mutex);
2679 mutex_unlock(&cgrp->pidlist_mutex);
2683 /* entry not found; create a new one */
2684 l = kmalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
2686 mutex_unlock(&cgrp->pidlist_mutex);
2689 init_rwsem(&l->mutex);
2690 down_write(&l->mutex);
2692 l->key.ns = get_pid_ns(ns);
2693 l->use_count = 0; /* don't increment here */
2696 list_add(&l->links, &cgrp->pidlists);
2697 mutex_unlock(&cgrp->pidlist_mutex);
2702 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
2704 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
2705 struct cgroup_pidlist **lp)
2709 int pid, n = 0; /* used for populating the array */
2710 struct cgroup_iter it;
2711 struct task_struct *tsk;
2712 struct cgroup_pidlist *l;
2715 * If cgroup gets more users after we read count, we won't have
2716 * enough space - tough. This race is indistinguishable to the
2717 * caller from the case that the additional cgroup users didn't
2718 * show up until sometime later on.
2720 length = cgroup_task_count(cgrp);
2721 array = pidlist_allocate(length);
2724 /* now, populate the array */
2725 cgroup_iter_start(cgrp, &it);
2726 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2727 if (unlikely(n == length))
2729 /* get tgid or pid for procs or tasks file respectively */
2730 if (type == CGROUP_FILE_PROCS)
2731 pid = task_tgid_vnr(tsk);
2733 pid = task_pid_vnr(tsk);
2734 if (pid > 0) /* make sure to only use valid results */
2737 cgroup_iter_end(cgrp, &it);
2739 /* now sort & (if procs) strip out duplicates */
2740 sort(array, length, sizeof(pid_t), cmppid, NULL);
2741 if (type == CGROUP_FILE_PROCS)
2742 length = pidlist_uniq(&array, length);
2743 l = cgroup_pidlist_find(cgrp, type);
2745 pidlist_free(array);
2748 /* store array, freeing old if necessary - lock already held */
2749 pidlist_free(l->list);
2753 up_write(&l->mutex);
2759 * cgroupstats_build - build and fill cgroupstats
2760 * @stats: cgroupstats to fill information into
2761 * @dentry: A dentry entry belonging to the cgroup for which stats have
2764 * Build and fill cgroupstats so that taskstats can export it to user
2767 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2770 struct cgroup *cgrp;
2771 struct cgroup_iter it;
2772 struct task_struct *tsk;
2775 * Validate dentry by checking the superblock operations,
2776 * and make sure it's a directory.
2778 if (dentry->d_sb->s_op != &cgroup_ops ||
2779 !S_ISDIR(dentry->d_inode->i_mode))
2783 cgrp = dentry->d_fsdata;
2785 cgroup_iter_start(cgrp, &it);
2786 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2787 switch (tsk->state) {
2789 stats->nr_running++;
2791 case TASK_INTERRUPTIBLE:
2792 stats->nr_sleeping++;
2794 case TASK_UNINTERRUPTIBLE:
2795 stats->nr_uninterruptible++;
2798 stats->nr_stopped++;
2801 if (delayacct_is_task_waiting_on_io(tsk))
2802 stats->nr_io_wait++;
2806 cgroup_iter_end(cgrp, &it);
2814 * seq_file methods for the tasks/procs files. The seq_file position is the
2815 * next pid to display; the seq_file iterator is a pointer to the pid
2816 * in the cgroup->l->list array.
2819 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
2822 * Initially we receive a position value that corresponds to
2823 * one more than the last pid shown (or 0 on the first call or
2824 * after a seek to the start). Use a binary-search to find the
2825 * next pid to display, if any
2827 struct cgroup_pidlist *l = s->private;
2828 int index = 0, pid = *pos;
2831 down_read(&l->mutex);
2833 int end = l->length;
2835 while (index < end) {
2836 int mid = (index + end) / 2;
2837 if (l->list[mid] == pid) {
2840 } else if (l->list[mid] <= pid)
2846 /* If we're off the end of the array, we're done */
2847 if (index >= l->length)
2849 /* Update the abstract position to be the actual pid that we found */
2850 iter = l->list + index;
2855 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
2857 struct cgroup_pidlist *l = s->private;
2861 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
2863 struct cgroup_pidlist *l = s->private;
2865 pid_t *end = l->list + l->length;
2867 * Advance to the next pid in the array. If this goes off the
2879 static int cgroup_pidlist_show(struct seq_file *s, void *v)
2881 return seq_printf(s, "%d\n", *(int *)v);
2885 * seq_operations functions for iterating on pidlists through seq_file -
2886 * independent of whether it's tasks or procs
2888 static const struct seq_operations cgroup_pidlist_seq_operations = {
2889 .start = cgroup_pidlist_start,
2890 .stop = cgroup_pidlist_stop,
2891 .next = cgroup_pidlist_next,
2892 .show = cgroup_pidlist_show,
2895 static void cgroup_release_pid_array(struct cgroup_pidlist *l)
2898 * the case where we're the last user of this particular pidlist will
2899 * have us remove it from the cgroup's list, which entails taking the
2900 * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
2901 * pidlist_mutex, we have to take pidlist_mutex first.
2903 mutex_lock(&l->owner->pidlist_mutex);
2904 down_write(&l->mutex);
2905 BUG_ON(!l->use_count);
2906 if (!--l->use_count) {
2907 /* we're the last user if refcount is 0; remove and free */
2908 list_del(&l->links);
2909 mutex_unlock(&l->owner->pidlist_mutex);
2910 pidlist_free(l->list);
2911 put_pid_ns(l->key.ns);
2912 up_write(&l->mutex);
2916 mutex_unlock(&l->owner->pidlist_mutex);
2917 up_write(&l->mutex);
2920 static int cgroup_pidlist_release(struct inode *inode, struct file *file)
2922 struct cgroup_pidlist *l;
2923 if (!(file->f_mode & FMODE_READ))
2926 * the seq_file will only be initialized if the file was opened for
2927 * reading; hence we check if it's not null only in that case.
2929 l = ((struct seq_file *)file->private_data)->private;
2930 cgroup_release_pid_array(l);
2931 return seq_release(inode, file);
2934 static const struct file_operations cgroup_pidlist_operations = {
2936 .llseek = seq_lseek,
2937 .write = cgroup_file_write,
2938 .release = cgroup_pidlist_release,
2942 * The following functions handle opens on a file that displays a pidlist
2943 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
2946 /* helper function for the two below it */
2947 static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
2949 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2950 struct cgroup_pidlist *l;
2953 /* Nothing to do for write-only files */
2954 if (!(file->f_mode & FMODE_READ))
2957 /* have the array populated */
2958 retval = pidlist_array_load(cgrp, type, &l);
2961 /* configure file information */
2962 file->f_op = &cgroup_pidlist_operations;
2964 retval = seq_open(file, &cgroup_pidlist_seq_operations);
2966 cgroup_release_pid_array(l);
2969 ((struct seq_file *)file->private_data)->private = l;
2972 static int cgroup_tasks_open(struct inode *unused, struct file *file)
2974 return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
2976 static int cgroup_procs_open(struct inode *unused, struct file *file)
2978 return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
2981 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
2984 return notify_on_release(cgrp);
2987 static int cgroup_write_notify_on_release(struct cgroup *cgrp,
2991 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
2993 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2995 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3000 * Unregister event and free resources.
3002 * Gets called from workqueue.
3004 static void cgroup_event_remove(struct work_struct *work)
3006 struct cgroup_event *event = container_of(work, struct cgroup_event,
3008 struct cgroup *cgrp = event->cgrp;
3010 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3012 eventfd_ctx_put(event->eventfd);
3018 * Gets called on POLLHUP on eventfd when user closes it.
3020 * Called with wqh->lock held and interrupts disabled.
3022 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3023 int sync, void *key)
3025 struct cgroup_event *event = container_of(wait,
3026 struct cgroup_event, wait);
3027 struct cgroup *cgrp = event->cgrp;
3028 unsigned long flags = (unsigned long)key;
3030 if (flags & POLLHUP) {
3031 __remove_wait_queue(event->wqh, &event->wait);
3032 spin_lock(&cgrp->event_list_lock);
3033 list_del(&event->list);
3034 spin_unlock(&cgrp->event_list_lock);
3036 * We are in atomic context, but cgroup_event_remove() may
3037 * sleep, so we have to call it in workqueue.
3039 schedule_work(&event->remove);
3045 static void cgroup_event_ptable_queue_proc(struct file *file,
3046 wait_queue_head_t *wqh, poll_table *pt)
3048 struct cgroup_event *event = container_of(pt,
3049 struct cgroup_event, pt);
3052 add_wait_queue(wqh, &event->wait);
3056 * Parse input and register new cgroup event handler.
3058 * Input must be in format '<event_fd> <control_fd> <args>'.
3059 * Interpretation of args is defined by control file implementation.
3061 static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3064 struct cgroup_event *event = NULL;
3065 unsigned int efd, cfd;
3066 struct file *efile = NULL;
3067 struct file *cfile = NULL;
3071 efd = simple_strtoul(buffer, &endp, 10);
3076 cfd = simple_strtoul(buffer, &endp, 10);
3077 if ((*endp != ' ') && (*endp != '\0'))
3081 event = kzalloc(sizeof(*event), GFP_KERNEL);
3085 INIT_LIST_HEAD(&event->list);
3086 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
3087 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
3088 INIT_WORK(&event->remove, cgroup_event_remove);
3090 efile = eventfd_fget(efd);
3091 if (IS_ERR(efile)) {
3092 ret = PTR_ERR(efile);
3096 event->eventfd = eventfd_ctx_fileget(efile);
3097 if (IS_ERR(event->eventfd)) {
3098 ret = PTR_ERR(event->eventfd);
3108 /* the process need read permission on control file */
3109 ret = file_permission(cfile, MAY_READ);
3113 event->cft = __file_cft(cfile);
3114 if (IS_ERR(event->cft)) {
3115 ret = PTR_ERR(event->cft);
3119 if (!event->cft->register_event || !event->cft->unregister_event) {
3124 ret = event->cft->register_event(cgrp, event->cft,
3125 event->eventfd, buffer);
3129 if (efile->f_op->poll(efile, &event->pt) & POLLHUP) {
3130 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3136 * Events should be removed after rmdir of cgroup directory, but before
3137 * destroying subsystem state objects. Let's take reference to cgroup
3138 * directory dentry to do that.
3142 spin_lock(&cgrp->event_list_lock);
3143 list_add(&event->list, &cgrp->event_list);
3144 spin_unlock(&cgrp->event_list_lock);
3155 if (event && event->eventfd && !IS_ERR(event->eventfd))
3156 eventfd_ctx_put(event->eventfd);
3158 if (!IS_ERR_OR_NULL(efile))
3167 * for the common functions, 'private' gives the type of file
3169 /* for hysterical raisins, we can't put this on the older files */
3170 #define CGROUP_FILE_GENERIC_PREFIX "cgroup."
3171 static struct cftype files[] = {
3174 .open = cgroup_tasks_open,
3175 .write_u64 = cgroup_tasks_write,
3176 .release = cgroup_pidlist_release,
3177 .mode = S_IRUGO | S_IWUSR,
3180 .name = CGROUP_FILE_GENERIC_PREFIX "procs",
3181 .open = cgroup_procs_open,
3182 /* .write_u64 = cgroup_procs_write, TODO */
3183 .release = cgroup_pidlist_release,
3187 .name = "notify_on_release",
3188 .read_u64 = cgroup_read_notify_on_release,
3189 .write_u64 = cgroup_write_notify_on_release,
3192 .name = CGROUP_FILE_GENERIC_PREFIX "event_control",
3193 .write_string = cgroup_write_event_control,
3198 static struct cftype cft_release_agent = {
3199 .name = "release_agent",
3200 .read_seq_string = cgroup_release_agent_show,
3201 .write_string = cgroup_release_agent_write,
3202 .max_write_len = PATH_MAX,
3205 static int cgroup_populate_dir(struct cgroup *cgrp)
3208 struct cgroup_subsys *ss;
3210 /* First clear out any existing files */
3211 cgroup_clear_directory(cgrp->dentry);
3213 err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
3217 if (cgrp == cgrp->top_cgroup) {
3218 if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
3222 for_each_subsys(cgrp->root, ss) {
3223 if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
3226 /* This cgroup is ready now */
3227 for_each_subsys(cgrp->root, ss) {
3228 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3230 * Update id->css pointer and make this css visible from
3231 * CSS ID functions. This pointer will be dereferened
3232 * from RCU-read-side without locks.
3235 rcu_assign_pointer(css->id->css, css);
3241 static void init_cgroup_css(struct cgroup_subsys_state *css,
3242 struct cgroup_subsys *ss,
3243 struct cgroup *cgrp)
3246 atomic_set(&css->refcnt, 1);
3249 if (cgrp == dummytop)
3250 set_bit(CSS_ROOT, &css->flags);
3251 BUG_ON(cgrp->subsys[ss->subsys_id]);
3252 cgrp->subsys[ss->subsys_id] = css;
3255 static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
3257 /* We need to take each hierarchy_mutex in a consistent order */
3261 * No worry about a race with rebind_subsystems that might mess up the
3262 * locking order, since both parties are under cgroup_mutex.
3264 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3265 struct cgroup_subsys *ss = subsys[i];
3268 if (ss->root == root)
3269 mutex_lock(&ss->hierarchy_mutex);
3273 static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
3277 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3278 struct cgroup_subsys *ss = subsys[i];
3281 if (ss->root == root)
3282 mutex_unlock(&ss->hierarchy_mutex);
3287 * cgroup_create - create a cgroup
3288 * @parent: cgroup that will be parent of the new cgroup
3289 * @dentry: dentry of the new cgroup
3290 * @mode: mode to set on new inode
3292 * Must be called with the mutex on the parent inode held
3294 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
3297 struct cgroup *cgrp;
3298 struct cgroupfs_root *root = parent->root;
3300 struct cgroup_subsys *ss;
3301 struct super_block *sb = root->sb;
3303 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3307 /* Grab a reference on the superblock so the hierarchy doesn't
3308 * get deleted on unmount if there are child cgroups. This
3309 * can be done outside cgroup_mutex, since the sb can't
3310 * disappear while someone has an open control file on the
3312 atomic_inc(&sb->s_active);
3314 mutex_lock(&cgroup_mutex);
3316 init_cgroup_housekeeping(cgrp);
3318 cgrp->parent = parent;
3319 cgrp->root = parent->root;
3320 cgrp->top_cgroup = parent->top_cgroup;
3322 if (notify_on_release(parent))
3323 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3325 for_each_subsys(root, ss) {
3326 struct cgroup_subsys_state *css = ss->create(ss, cgrp);
3332 init_cgroup_css(css, ss, cgrp);
3334 err = alloc_css_id(ss, parent, cgrp);
3338 /* At error, ->destroy() callback has to free assigned ID. */
3341 cgroup_lock_hierarchy(root);
3342 list_add(&cgrp->sibling, &cgrp->parent->children);
3343 cgroup_unlock_hierarchy(root);
3344 root->number_of_cgroups++;
3346 err = cgroup_create_dir(cgrp, dentry, mode);
3350 set_bit(CGRP_RELEASABLE, &parent->flags);
3352 /* The cgroup directory was pre-locked for us */
3353 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
3355 err = cgroup_populate_dir(cgrp);
3356 /* If err < 0, we have a half-filled directory - oh well ;) */
3358 mutex_unlock(&cgroup_mutex);
3359 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
3365 cgroup_lock_hierarchy(root);
3366 list_del(&cgrp->sibling);
3367 cgroup_unlock_hierarchy(root);
3368 root->number_of_cgroups--;
3372 for_each_subsys(root, ss) {
3373 if (cgrp->subsys[ss->subsys_id])
3374 ss->destroy(ss, cgrp);
3377 mutex_unlock(&cgroup_mutex);
3379 /* Release the reference count that we took on the superblock */
3380 deactivate_super(sb);
3386 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
3388 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
3390 /* the vfs holds inode->i_mutex already */
3391 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
3394 static int cgroup_has_css_refs(struct cgroup *cgrp)
3396 /* Check the reference count on each subsystem. Since we
3397 * already established that there are no tasks in the
3398 * cgroup, if the css refcount is also 1, then there should
3399 * be no outstanding references, so the subsystem is safe to
3400 * destroy. We scan across all subsystems rather than using
3401 * the per-hierarchy linked list of mounted subsystems since
3402 * we can be called via check_for_release() with no
3403 * synchronization other than RCU, and the subsystem linked
3404 * list isn't RCU-safe */
3407 * We won't need to lock the subsys array, because the subsystems
3408 * we're concerned about aren't going anywhere since our cgroup root
3409 * has a reference on them.
3411 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3412 struct cgroup_subsys *ss = subsys[i];
3413 struct cgroup_subsys_state *css;
3414 /* Skip subsystems not present or not in this hierarchy */
3415 if (ss == NULL || ss->root != cgrp->root)
3417 css = cgrp->subsys[ss->subsys_id];
3418 /* When called from check_for_release() it's possible
3419 * that by this point the cgroup has been removed
3420 * and the css deleted. But a false-positive doesn't
3421 * matter, since it can only happen if the cgroup
3422 * has been deleted and hence no longer needs the
3423 * release agent to be called anyway. */
3424 if (css && (atomic_read(&css->refcnt) > 1))
3431 * Atomically mark all (or else none) of the cgroup's CSS objects as
3432 * CSS_REMOVED. Return true on success, or false if the cgroup has
3433 * busy subsystems. Call with cgroup_mutex held
3436 static int cgroup_clear_css_refs(struct cgroup *cgrp)
3438 struct cgroup_subsys *ss;
3439 unsigned long flags;
3440 bool failed = false;
3441 local_irq_save(flags);
3442 for_each_subsys(cgrp->root, ss) {
3443 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3446 /* We can only remove a CSS with a refcnt==1 */
3447 refcnt = atomic_read(&css->refcnt);
3454 * Drop the refcnt to 0 while we check other
3455 * subsystems. This will cause any racing
3456 * css_tryget() to spin until we set the
3457 * CSS_REMOVED bits or abort
3459 if (atomic_cmpxchg(&css->refcnt, refcnt, 0) == refcnt)
3465 for_each_subsys(cgrp->root, ss) {
3466 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3469 * Restore old refcnt if we previously managed
3470 * to clear it from 1 to 0
3472 if (!atomic_read(&css->refcnt))
3473 atomic_set(&css->refcnt, 1);
3475 /* Commit the fact that the CSS is removed */
3476 set_bit(CSS_REMOVED, &css->flags);
3479 local_irq_restore(flags);
3483 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
3485 struct cgroup *cgrp = dentry->d_fsdata;
3487 struct cgroup *parent;
3489 struct cgroup_event *event, *tmp;
3492 /* the vfs holds both inode->i_mutex already */
3494 mutex_lock(&cgroup_mutex);
3495 if (atomic_read(&cgrp->count) != 0) {
3496 mutex_unlock(&cgroup_mutex);
3499 if (!list_empty(&cgrp->children)) {
3500 mutex_unlock(&cgroup_mutex);
3503 mutex_unlock(&cgroup_mutex);
3506 * In general, subsystem has no css->refcnt after pre_destroy(). But
3507 * in racy cases, subsystem may have to get css->refcnt after
3508 * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
3509 * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
3510 * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
3511 * and subsystem's reference count handling. Please see css_get/put
3512 * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
3514 set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3517 * Call pre_destroy handlers of subsys. Notify subsystems
3518 * that rmdir() request comes.
3520 ret = cgroup_call_pre_destroy(cgrp);
3522 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3526 mutex_lock(&cgroup_mutex);
3527 parent = cgrp->parent;
3528 if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
3529 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3530 mutex_unlock(&cgroup_mutex);
3533 prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
3534 if (!cgroup_clear_css_refs(cgrp)) {
3535 mutex_unlock(&cgroup_mutex);
3537 * Because someone may call cgroup_wakeup_rmdir_waiter() before
3538 * prepare_to_wait(), we need to check this flag.
3540 if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
3542 finish_wait(&cgroup_rmdir_waitq, &wait);
3543 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3544 if (signal_pending(current))
3548 /* NO css_tryget() can success after here. */
3549 finish_wait(&cgroup_rmdir_waitq, &wait);
3550 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3552 spin_lock(&release_list_lock);
3553 set_bit(CGRP_REMOVED, &cgrp->flags);
3554 if (!list_empty(&cgrp->release_list))
3555 list_del(&cgrp->release_list);
3556 spin_unlock(&release_list_lock);
3558 cgroup_lock_hierarchy(cgrp->root);
3559 /* delete this cgroup from parent->children */
3560 list_del(&cgrp->sibling);
3561 cgroup_unlock_hierarchy(cgrp->root);
3563 spin_lock(&cgrp->dentry->d_lock);
3564 d = dget(cgrp->dentry);
3565 spin_unlock(&d->d_lock);
3567 cgroup_d_remove_dir(d);
3570 check_for_release(parent);
3573 * Unregister events and notify userspace.
3574 * Notify userspace about cgroup removing only after rmdir of cgroup
3575 * directory to avoid race between userspace and kernelspace
3577 spin_lock(&cgrp->event_list_lock);
3578 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
3579 list_del(&event->list);
3580 remove_wait_queue(event->wqh, &event->wait);
3581 eventfd_signal(event->eventfd, 1);
3582 schedule_work(&event->remove);
3584 spin_unlock(&cgrp->event_list_lock);
3586 mutex_unlock(&cgroup_mutex);
3590 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
3592 struct cgroup_subsys_state *css;
3594 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
3596 /* Create the top cgroup state for this subsystem */
3597 list_add(&ss->sibling, &rootnode.subsys_list);
3598 ss->root = &rootnode;
3599 css = ss->create(ss, dummytop);
3600 /* We don't handle early failures gracefully */
3601 BUG_ON(IS_ERR(css));
3602 init_cgroup_css(css, ss, dummytop);
3604 /* Update the init_css_set to contain a subsys
3605 * pointer to this state - since the subsystem is
3606 * newly registered, all tasks and hence the
3607 * init_css_set is in the subsystem's top cgroup. */
3608 init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
3610 need_forkexit_callback |= ss->fork || ss->exit;
3612 /* At system boot, before all subsystems have been
3613 * registered, no tasks have been forked, so we don't
3614 * need to invoke fork callbacks here. */
3615 BUG_ON(!list_empty(&init_task.tasks));
3617 mutex_init(&ss->hierarchy_mutex);
3618 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3621 /* this function shouldn't be used with modular subsystems, since they
3622 * need to register a subsys_id, among other things */
3627 * cgroup_load_subsys: load and register a modular subsystem at runtime
3628 * @ss: the subsystem to load
3630 * This function should be called in a modular subsystem's initcall. If the
3631 * subsystem is built as a module, it will be assigned a new subsys_id and set
3632 * up for use. If the subsystem is built-in anyway, work is delegated to the
3633 * simpler cgroup_init_subsys.
3635 int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
3638 struct cgroup_subsys_state *css;
3640 /* check name and function validity */
3641 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
3642 ss->create == NULL || ss->destroy == NULL)
3646 * we don't support callbacks in modular subsystems. this check is
3647 * before the ss->module check for consistency; a subsystem that could
3648 * be a module should still have no callbacks even if the user isn't
3649 * compiling it as one.
3651 if (ss->fork || ss->exit)
3655 * an optionally modular subsystem is built-in: we want to do nothing,
3656 * since cgroup_init_subsys will have already taken care of it.
3658 if (ss->module == NULL) {
3659 /* a few sanity checks */
3660 BUG_ON(ss->subsys_id >= CGROUP_BUILTIN_SUBSYS_COUNT);
3661 BUG_ON(subsys[ss->subsys_id] != ss);
3666 * need to register a subsys id before anything else - for example,
3667 * init_cgroup_css needs it.
3669 mutex_lock(&cgroup_mutex);
3670 /* find the first empty slot in the array */
3671 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
3672 if (subsys[i] == NULL)
3675 if (i == CGROUP_SUBSYS_COUNT) {
3676 /* maximum number of subsystems already registered! */
3677 mutex_unlock(&cgroup_mutex);
3680 /* assign ourselves the subsys_id */
3685 * no ss->create seems to need anything important in the ss struct, so
3686 * this can happen first (i.e. before the rootnode attachment).
3688 css = ss->create(ss, dummytop);
3690 /* failure case - need to deassign the subsys[] slot. */
3692 mutex_unlock(&cgroup_mutex);
3693 return PTR_ERR(css);
3696 list_add(&ss->sibling, &rootnode.subsys_list);
3697 ss->root = &rootnode;
3699 /* our new subsystem will be attached to the dummy hierarchy. */
3700 init_cgroup_css(css, ss, dummytop);
3701 /* init_idr must be after init_cgroup_css because it sets css->id. */
3703 int ret = cgroup_init_idr(ss, css);
3705 dummytop->subsys[ss->subsys_id] = NULL;
3706 ss->destroy(ss, dummytop);
3708 mutex_unlock(&cgroup_mutex);
3714 * Now we need to entangle the css into the existing css_sets. unlike
3715 * in cgroup_init_subsys, there are now multiple css_sets, so each one
3716 * will need a new pointer to it; done by iterating the css_set_table.
3717 * furthermore, modifying the existing css_sets will corrupt the hash
3718 * table state, so each changed css_set will need its hash recomputed.
3719 * this is all done under the css_set_lock.
3721 write_lock(&css_set_lock);
3722 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
3724 struct hlist_node *node, *tmp;
3725 struct hlist_head *bucket = &css_set_table[i], *new_bucket;
3727 hlist_for_each_entry_safe(cg, node, tmp, bucket, hlist) {
3728 /* skip entries that we already rehashed */
3729 if (cg->subsys[ss->subsys_id])
3731 /* remove existing entry */
3732 hlist_del(&cg->hlist);
3734 cg->subsys[ss->subsys_id] = css;
3735 /* recompute hash and restore entry */
3736 new_bucket = css_set_hash(cg->subsys);
3737 hlist_add_head(&cg->hlist, new_bucket);
3740 write_unlock(&css_set_lock);
3742 mutex_init(&ss->hierarchy_mutex);
3743 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3747 mutex_unlock(&cgroup_mutex);
3750 EXPORT_SYMBOL_GPL(cgroup_load_subsys);
3753 * cgroup_unload_subsys: unload a modular subsystem
3754 * @ss: the subsystem to unload
3756 * This function should be called in a modular subsystem's exitcall. When this
3757 * function is invoked, the refcount on the subsystem's module will be 0, so
3758 * the subsystem will not be attached to any hierarchy.
3760 void cgroup_unload_subsys(struct cgroup_subsys *ss)
3762 struct cg_cgroup_link *link;
3763 struct hlist_head *hhead;
3765 BUG_ON(ss->module == NULL);
3768 * we shouldn't be called if the subsystem is in use, and the use of
3769 * try_module_get in parse_cgroupfs_options should ensure that it
3770 * doesn't start being used while we're killing it off.
3772 BUG_ON(ss->root != &rootnode);
3774 mutex_lock(&cgroup_mutex);
3775 /* deassign the subsys_id */
3776 BUG_ON(ss->subsys_id < CGROUP_BUILTIN_SUBSYS_COUNT);
3777 subsys[ss->subsys_id] = NULL;
3779 /* remove subsystem from rootnode's list of subsystems */
3780 list_del(&ss->sibling);
3783 * disentangle the css from all css_sets attached to the dummytop. as
3784 * in loading, we need to pay our respects to the hashtable gods.
3786 write_lock(&css_set_lock);
3787 list_for_each_entry(link, &dummytop->css_sets, cgrp_link_list) {
3788 struct css_set *cg = link->cg;
3790 hlist_del(&cg->hlist);
3791 BUG_ON(!cg->subsys[ss->subsys_id]);
3792 cg->subsys[ss->subsys_id] = NULL;
3793 hhead = css_set_hash(cg->subsys);
3794 hlist_add_head(&cg->hlist, hhead);
3796 write_unlock(&css_set_lock);
3799 * remove subsystem's css from the dummytop and free it - need to free
3800 * before marking as null because ss->destroy needs the cgrp->subsys
3801 * pointer to find their state. note that this also takes care of
3802 * freeing the css_id.
3804 ss->destroy(ss, dummytop);
3805 dummytop->subsys[ss->subsys_id] = NULL;
3807 mutex_unlock(&cgroup_mutex);
3809 EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
3812 * cgroup_init_early - cgroup initialization at system boot
3814 * Initialize cgroups at system boot, and initialize any
3815 * subsystems that request early init.
3817 int __init cgroup_init_early(void)
3820 atomic_set(&init_css_set.refcount, 1);
3821 INIT_LIST_HEAD(&init_css_set.cg_links);
3822 INIT_LIST_HEAD(&init_css_set.tasks);
3823 INIT_HLIST_NODE(&init_css_set.hlist);
3825 init_cgroup_root(&rootnode);
3827 init_task.cgroups = &init_css_set;
3829 init_css_set_link.cg = &init_css_set;
3830 init_css_set_link.cgrp = dummytop;
3831 list_add(&init_css_set_link.cgrp_link_list,
3832 &rootnode.top_cgroup.css_sets);
3833 list_add(&init_css_set_link.cg_link_list,
3834 &init_css_set.cg_links);
3836 for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
3837 INIT_HLIST_HEAD(&css_set_table[i]);
3839 /* at bootup time, we don't worry about modular subsystems */
3840 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3841 struct cgroup_subsys *ss = subsys[i];
3844 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
3845 BUG_ON(!ss->create);
3846 BUG_ON(!ss->destroy);
3847 if (ss->subsys_id != i) {
3848 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
3849 ss->name, ss->subsys_id);
3854 cgroup_init_subsys(ss);
3860 * cgroup_init - cgroup initialization
3862 * Register cgroup filesystem and /proc file, and initialize
3863 * any subsystems that didn't request early init.
3865 int __init cgroup_init(void)
3869 struct hlist_head *hhead;
3871 err = bdi_init(&cgroup_backing_dev_info);
3875 /* at bootup time, we don't worry about modular subsystems */
3876 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3877 struct cgroup_subsys *ss = subsys[i];
3878 if (!ss->early_init)
3879 cgroup_init_subsys(ss);
3881 cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
3884 /* Add init_css_set to the hash table */
3885 hhead = css_set_hash(init_css_set.subsys);
3886 hlist_add_head(&init_css_set.hlist, hhead);
3887 BUG_ON(!init_root_id(&rootnode));
3889 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
3895 err = register_filesystem(&cgroup_fs_type);
3897 kobject_put(cgroup_kobj);
3901 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
3905 bdi_destroy(&cgroup_backing_dev_info);
3911 * proc_cgroup_show()
3912 * - Print task's cgroup paths into seq_file, one line for each hierarchy
3913 * - Used for /proc/<pid>/cgroup.
3914 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
3915 * doesn't really matter if tsk->cgroup changes after we read it,
3916 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
3917 * anyway. No need to check that tsk->cgroup != NULL, thanks to
3918 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
3919 * cgroup to top_cgroup.
3922 /* TODO: Use a proper seq_file iterator */
3923 static int proc_cgroup_show(struct seq_file *m, void *v)
3926 struct task_struct *tsk;
3929 struct cgroupfs_root *root;
3932 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
3938 tsk = get_pid_task(pid, PIDTYPE_PID);
3944 mutex_lock(&cgroup_mutex);
3946 for_each_active_root(root) {
3947 struct cgroup_subsys *ss;
3948 struct cgroup *cgrp;
3951 seq_printf(m, "%d:", root->hierarchy_id);
3952 for_each_subsys(root, ss)
3953 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
3954 if (strlen(root->name))
3955 seq_printf(m, "%sname=%s", count ? "," : "",
3958 cgrp = task_cgroup_from_root(tsk, root);
3959 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
3967 mutex_unlock(&cgroup_mutex);
3968 put_task_struct(tsk);
3975 static int cgroup_open(struct inode *inode, struct file *file)
3977 struct pid *pid = PROC_I(inode)->pid;
3978 return single_open(file, proc_cgroup_show, pid);
3981 const struct file_operations proc_cgroup_operations = {
3982 .open = cgroup_open,
3984 .llseek = seq_lseek,
3985 .release = single_release,
3988 /* Display information about each subsystem and each hierarchy */
3989 static int proc_cgroupstats_show(struct seq_file *m, void *v)
3993 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
3995 * ideally we don't want subsystems moving around while we do this.
3996 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
3997 * subsys/hierarchy state.
3999 mutex_lock(&cgroup_mutex);
4000 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
4001 struct cgroup_subsys *ss = subsys[i];
4004 seq_printf(m, "%s\t%d\t%d\t%d\n",
4005 ss->name, ss->root->hierarchy_id,
4006 ss->root->number_of_cgroups, !ss->disabled);
4008 mutex_unlock(&cgroup_mutex);
4012 static int cgroupstats_open(struct inode *inode, struct file *file)
4014 return single_open(file, proc_cgroupstats_show, NULL);
4017 static const struct file_operations proc_cgroupstats_operations = {
4018 .open = cgroupstats_open,
4020 .llseek = seq_lseek,
4021 .release = single_release,
4025 * cgroup_fork - attach newly forked task to its parents cgroup.
4026 * @child: pointer to task_struct of forking parent process.
4028 * Description: A task inherits its parent's cgroup at fork().
4030 * A pointer to the shared css_set was automatically copied in
4031 * fork.c by dup_task_struct(). However, we ignore that copy, since
4032 * it was not made under the protection of RCU or cgroup_mutex, so
4033 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4034 * have already changed current->cgroups, allowing the previously
4035 * referenced cgroup group to be removed and freed.
4037 * At the point that cgroup_fork() is called, 'current' is the parent
4038 * task, and the passed argument 'child' points to the child task.
4040 void cgroup_fork(struct task_struct *child)
4043 child->cgroups = current->cgroups;
4044 get_css_set(child->cgroups);
4045 task_unlock(current);
4046 INIT_LIST_HEAD(&child->cg_list);
4050 * cgroup_fork_callbacks - run fork callbacks
4051 * @child: the new task
4053 * Called on a new task very soon before adding it to the
4054 * tasklist. No need to take any locks since no-one can
4055 * be operating on this task.
4057 void cgroup_fork_callbacks(struct task_struct *child)
4059 if (need_forkexit_callback) {
4062 * forkexit callbacks are only supported for builtin
4063 * subsystems, and the builtin section of the subsys array is
4064 * immutable, so we don't need to lock the subsys array here.
4066 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4067 struct cgroup_subsys *ss = subsys[i];
4069 ss->fork(ss, child);
4075 * cgroup_post_fork - called on a new task after adding it to the task list
4076 * @child: the task in question
4078 * Adds the task to the list running through its css_set if necessary.
4079 * Has to be after the task is visible on the task list in case we race
4080 * with the first call to cgroup_iter_start() - to guarantee that the
4081 * new task ends up on its list.
4083 void cgroup_post_fork(struct task_struct *child)
4085 if (use_task_css_set_links) {
4086 write_lock(&css_set_lock);
4088 if (list_empty(&child->cg_list))
4089 list_add(&child->cg_list, &child->cgroups->tasks);
4091 write_unlock(&css_set_lock);
4095 * cgroup_exit - detach cgroup from exiting task
4096 * @tsk: pointer to task_struct of exiting process
4097 * @run_callback: run exit callbacks?
4099 * Description: Detach cgroup from @tsk and release it.
4101 * Note that cgroups marked notify_on_release force every task in
4102 * them to take the global cgroup_mutex mutex when exiting.
4103 * This could impact scaling on very large systems. Be reluctant to
4104 * use notify_on_release cgroups where very high task exit scaling
4105 * is required on large systems.
4107 * the_top_cgroup_hack:
4109 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4111 * We call cgroup_exit() while the task is still competent to
4112 * handle notify_on_release(), then leave the task attached to the
4113 * root cgroup in each hierarchy for the remainder of its exit.
4115 * To do this properly, we would increment the reference count on
4116 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4117 * code we would add a second cgroup function call, to drop that
4118 * reference. This would just create an unnecessary hot spot on
4119 * the top_cgroup reference count, to no avail.
4121 * Normally, holding a reference to a cgroup without bumping its
4122 * count is unsafe. The cgroup could go away, or someone could
4123 * attach us to a different cgroup, decrementing the count on
4124 * the first cgroup that we never incremented. But in this case,
4125 * top_cgroup isn't going away, and either task has PF_EXITING set,
4126 * which wards off any cgroup_attach_task() attempts, or task is a failed
4127 * fork, never visible to cgroup_attach_task.
4129 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4134 if (run_callbacks && need_forkexit_callback) {
4136 * modular subsystems can't use callbacks, so no need to lock
4139 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4140 struct cgroup_subsys *ss = subsys[i];
4147 * Unlink from the css_set task list if necessary.
4148 * Optimistically check cg_list before taking
4151 if (!list_empty(&tsk->cg_list)) {
4152 write_lock(&css_set_lock);
4153 if (!list_empty(&tsk->cg_list))
4154 list_del(&tsk->cg_list);
4155 write_unlock(&css_set_lock);
4158 /* Reassign the task to the init_css_set. */
4161 tsk->cgroups = &init_css_set;
4168 * cgroup_clone - clone the cgroup the given subsystem is attached to
4169 * @tsk: the task to be moved
4170 * @subsys: the given subsystem
4171 * @nodename: the name for the new cgroup
4173 * Duplicate the current cgroup in the hierarchy that the given
4174 * subsystem is attached to, and move this task into the new
4177 int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
4180 struct dentry *dentry;
4182 struct cgroup *parent, *child;
4183 struct inode *inode;
4185 struct cgroupfs_root *root;
4186 struct cgroup_subsys *ss;
4188 /* We shouldn't be called by an unregistered subsystem */
4189 BUG_ON(!subsys->active);
4191 /* First figure out what hierarchy and cgroup we're dealing
4192 * with, and pin them so we can drop cgroup_mutex */
4193 mutex_lock(&cgroup_mutex);
4195 root = subsys->root;
4196 if (root == &rootnode) {
4197 mutex_unlock(&cgroup_mutex);
4201 /* Pin the hierarchy */
4202 if (!atomic_inc_not_zero(&root->sb->s_active)) {
4203 /* We race with the final deactivate_super() */
4204 mutex_unlock(&cgroup_mutex);
4208 /* Keep the cgroup alive */
4210 parent = task_cgroup(tsk, subsys->subsys_id);
4215 mutex_unlock(&cgroup_mutex);
4217 /* Now do the VFS work to create a cgroup */
4218 inode = parent->dentry->d_inode;
4220 /* Hold the parent directory mutex across this operation to
4221 * stop anyone else deleting the new cgroup */
4222 mutex_lock(&inode->i_mutex);
4223 dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
4224 if (IS_ERR(dentry)) {
4226 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
4228 ret = PTR_ERR(dentry);
4232 /* Create the cgroup directory, which also creates the cgroup */
4233 ret = vfs_mkdir(inode, dentry, 0755);
4234 child = __d_cgrp(dentry);
4238 "Failed to create cgroup %s: %d\n", nodename,
4243 /* The cgroup now exists. Retake cgroup_mutex and check
4244 * that we're still in the same state that we thought we
4246 mutex_lock(&cgroup_mutex);
4247 if ((root != subsys->root) ||
4248 (parent != task_cgroup(tsk, subsys->subsys_id))) {
4249 /* Aargh, we raced ... */
4250 mutex_unlock(&inode->i_mutex);
4253 deactivate_super(root->sb);
4254 /* The cgroup is still accessible in the VFS, but
4255 * we're not going to try to rmdir() it at this
4258 "Race in cgroup_clone() - leaking cgroup %s\n",
4263 /* do any required auto-setup */
4264 for_each_subsys(root, ss) {
4266 ss->post_clone(ss, child);
4269 /* All seems fine. Finish by moving the task into the new cgroup */
4270 ret = cgroup_attach_task(child, tsk);
4271 mutex_unlock(&cgroup_mutex);
4274 mutex_unlock(&inode->i_mutex);
4276 mutex_lock(&cgroup_mutex);
4278 mutex_unlock(&cgroup_mutex);
4279 deactivate_super(root->sb);
4284 * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
4285 * @cgrp: the cgroup in question
4286 * @task: the task in question
4288 * See if @cgrp is a descendant of @task's cgroup in the appropriate
4291 * If we are sending in dummytop, then presumably we are creating
4292 * the top cgroup in the subsystem.
4294 * Called only by the ns (nsproxy) cgroup.
4296 int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
4299 struct cgroup *target;
4301 if (cgrp == dummytop)
4304 target = task_cgroup_from_root(task, cgrp->root);
4305 while (cgrp != target && cgrp!= cgrp->top_cgroup)
4306 cgrp = cgrp->parent;
4307 ret = (cgrp == target);
4311 static void check_for_release(struct cgroup *cgrp)
4313 /* All of these checks rely on RCU to keep the cgroup
4314 * structure alive */
4315 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
4316 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
4317 /* Control Group is currently removeable. If it's not
4318 * already queued for a userspace notification, queue
4320 int need_schedule_work = 0;
4321 spin_lock(&release_list_lock);
4322 if (!cgroup_is_removed(cgrp) &&
4323 list_empty(&cgrp->release_list)) {
4324 list_add(&cgrp->release_list, &release_list);
4325 need_schedule_work = 1;
4327 spin_unlock(&release_list_lock);
4328 if (need_schedule_work)
4329 schedule_work(&release_agent_work);
4333 /* Caller must verify that the css is not for root cgroup */
4334 void __css_get(struct cgroup_subsys_state *css, int count)
4336 atomic_add(count, &css->refcnt);
4337 set_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4339 EXPORT_SYMBOL_GPL(__css_get);
4341 /* Caller must verify that the css is not for root cgroup */
4342 void __css_put(struct cgroup_subsys_state *css, int count)
4344 struct cgroup *cgrp = css->cgroup;
4347 val = atomic_sub_return(count, &css->refcnt);
4349 check_for_release(cgrp);
4350 cgroup_wakeup_rmdir_waiter(cgrp);
4353 WARN_ON_ONCE(val < 1);
4355 EXPORT_SYMBOL_GPL(__css_put);
4358 * Notify userspace when a cgroup is released, by running the
4359 * configured release agent with the name of the cgroup (path
4360 * relative to the root of cgroup file system) as the argument.
4362 * Most likely, this user command will try to rmdir this cgroup.
4364 * This races with the possibility that some other task will be
4365 * attached to this cgroup before it is removed, or that some other
4366 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4367 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4368 * unused, and this cgroup will be reprieved from its death sentence,
4369 * to continue to serve a useful existence. Next time it's released,
4370 * we will get notified again, if it still has 'notify_on_release' set.
4372 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4373 * means only wait until the task is successfully execve()'d. The
4374 * separate release agent task is forked by call_usermodehelper(),
4375 * then control in this thread returns here, without waiting for the
4376 * release agent task. We don't bother to wait because the caller of
4377 * this routine has no use for the exit status of the release agent
4378 * task, so no sense holding our caller up for that.
4380 static void cgroup_release_agent(struct work_struct *work)
4382 BUG_ON(work != &release_agent_work);
4383 mutex_lock(&cgroup_mutex);
4384 spin_lock(&release_list_lock);
4385 while (!list_empty(&release_list)) {
4386 char *argv[3], *envp[3];
4388 char *pathbuf = NULL, *agentbuf = NULL;
4389 struct cgroup *cgrp = list_entry(release_list.next,
4392 list_del_init(&cgrp->release_list);
4393 spin_unlock(&release_list_lock);
4394 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4397 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4399 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4404 argv[i++] = agentbuf;
4405 argv[i++] = pathbuf;
4409 /* minimal command environment */
4410 envp[i++] = "HOME=/";
4411 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4414 /* Drop the lock while we invoke the usermode helper,
4415 * since the exec could involve hitting disk and hence
4416 * be a slow process */
4417 mutex_unlock(&cgroup_mutex);
4418 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4419 mutex_lock(&cgroup_mutex);
4423 spin_lock(&release_list_lock);
4425 spin_unlock(&release_list_lock);
4426 mutex_unlock(&cgroup_mutex);
4429 static int __init cgroup_disable(char *str)
4434 while ((token = strsep(&str, ",")) != NULL) {
4438 * cgroup_disable, being at boot time, can't know about module
4439 * subsystems, so we don't worry about them.
4441 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4442 struct cgroup_subsys *ss = subsys[i];
4444 if (!strcmp(token, ss->name)) {
4446 printk(KERN_INFO "Disabling %s control group"
4447 " subsystem\n", ss->name);
4454 __setup("cgroup_disable=", cgroup_disable);
4457 * Functons for CSS ID.
4461 *To get ID other than 0, this should be called when !cgroup_is_removed().
4463 unsigned short css_id(struct cgroup_subsys_state *css)
4465 struct css_id *cssid;
4468 * This css_id() can return correct value when somone has refcnt
4469 * on this or this is under rcu_read_lock(). Once css->id is allocated,
4470 * it's unchanged until freed.
4472 cssid = rcu_dereference_check(css->id,
4473 rcu_read_lock_held() || atomic_read(&css->refcnt));
4479 EXPORT_SYMBOL_GPL(css_id);
4481 unsigned short css_depth(struct cgroup_subsys_state *css)
4483 struct css_id *cssid;
4485 cssid = rcu_dereference_check(css->id,
4486 rcu_read_lock_held() || atomic_read(&css->refcnt));
4489 return cssid->depth;
4492 EXPORT_SYMBOL_GPL(css_depth);
4495 * css_is_ancestor - test "root" css is an ancestor of "child"
4496 * @child: the css to be tested.
4497 * @root: the css supporsed to be an ancestor of the child.
4499 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
4500 * this function reads css->id, this use rcu_dereference() and rcu_read_lock().
4501 * But, considering usual usage, the csses should be valid objects after test.
4502 * Assuming that the caller will do some action to the child if this returns
4503 * returns true, the caller must take "child";s reference count.
4504 * If "child" is valid object and this returns true, "root" is valid, too.
4507 bool css_is_ancestor(struct cgroup_subsys_state *child,
4508 const struct cgroup_subsys_state *root)
4510 struct css_id *child_id;
4511 struct css_id *root_id;
4515 child_id = rcu_dereference(child->id);
4516 root_id = rcu_dereference(root->id);
4519 || (child_id->depth < root_id->depth)
4520 || (child_id->stack[root_id->depth] != root_id->id))
4526 static void __free_css_id_cb(struct rcu_head *head)
4530 id = container_of(head, struct css_id, rcu_head);
4534 void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
4536 struct css_id *id = css->id;
4537 /* When this is called before css_id initialization, id can be NULL */
4541 BUG_ON(!ss->use_id);
4543 rcu_assign_pointer(id->css, NULL);
4544 rcu_assign_pointer(css->id, NULL);
4545 spin_lock(&ss->id_lock);
4546 idr_remove(&ss->idr, id->id);
4547 spin_unlock(&ss->id_lock);
4548 call_rcu(&id->rcu_head, __free_css_id_cb);
4550 EXPORT_SYMBOL_GPL(free_css_id);
4553 * This is called by init or create(). Then, calls to this function are
4554 * always serialized (By cgroup_mutex() at create()).
4557 static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
4559 struct css_id *newid;
4560 int myid, error, size;
4562 BUG_ON(!ss->use_id);
4564 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
4565 newid = kzalloc(size, GFP_KERNEL);
4567 return ERR_PTR(-ENOMEM);
4569 if (unlikely(!idr_pre_get(&ss->idr, GFP_KERNEL))) {
4573 spin_lock(&ss->id_lock);
4574 /* Don't use 0. allocates an ID of 1-65535 */
4575 error = idr_get_new_above(&ss->idr, newid, 1, &myid);
4576 spin_unlock(&ss->id_lock);
4578 /* Returns error when there are no free spaces for new ID.*/
4583 if (myid > CSS_ID_MAX)
4587 newid->depth = depth;
4591 spin_lock(&ss->id_lock);
4592 idr_remove(&ss->idr, myid);
4593 spin_unlock(&ss->id_lock);
4596 return ERR_PTR(error);
4600 static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
4601 struct cgroup_subsys_state *rootcss)
4603 struct css_id *newid;
4605 spin_lock_init(&ss->id_lock);
4608 newid = get_new_cssid(ss, 0);
4610 return PTR_ERR(newid);
4612 newid->stack[0] = newid->id;
4613 newid->css = rootcss;
4614 rootcss->id = newid;
4618 static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
4619 struct cgroup *child)
4621 int subsys_id, i, depth = 0;
4622 struct cgroup_subsys_state *parent_css, *child_css;
4623 struct css_id *child_id, *parent_id;
4625 subsys_id = ss->subsys_id;
4626 parent_css = parent->subsys[subsys_id];
4627 child_css = child->subsys[subsys_id];
4628 parent_id = parent_css->id;
4629 depth = parent_id->depth + 1;
4631 child_id = get_new_cssid(ss, depth);
4632 if (IS_ERR(child_id))
4633 return PTR_ERR(child_id);
4635 for (i = 0; i < depth; i++)
4636 child_id->stack[i] = parent_id->stack[i];
4637 child_id->stack[depth] = child_id->id;
4639 * child_id->css pointer will be set after this cgroup is available
4640 * see cgroup_populate_dir()
4642 rcu_assign_pointer(child_css->id, child_id);
4648 * css_lookup - lookup css by id
4649 * @ss: cgroup subsys to be looked into.
4652 * Returns pointer to cgroup_subsys_state if there is valid one with id.
4653 * NULL if not. Should be called under rcu_read_lock()
4655 struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
4657 struct css_id *cssid = NULL;
4659 BUG_ON(!ss->use_id);
4660 cssid = idr_find(&ss->idr, id);
4662 if (unlikely(!cssid))
4665 return rcu_dereference(cssid->css);
4667 EXPORT_SYMBOL_GPL(css_lookup);
4670 * css_get_next - lookup next cgroup under specified hierarchy.
4671 * @ss: pointer to subsystem
4672 * @id: current position of iteration.
4673 * @root: pointer to css. search tree under this.
4674 * @foundid: position of found object.
4676 * Search next css under the specified hierarchy of rootid. Calling under
4677 * rcu_read_lock() is necessary. Returns NULL if it reaches the end.
4679 struct cgroup_subsys_state *
4680 css_get_next(struct cgroup_subsys *ss, int id,
4681 struct cgroup_subsys_state *root, int *foundid)
4683 struct cgroup_subsys_state *ret = NULL;
4686 int rootid = css_id(root);
4687 int depth = css_depth(root);
4692 BUG_ON(!ss->use_id);
4693 /* fill start point for scan */
4697 * scan next entry from bitmap(tree), tmpid is updated after
4700 spin_lock(&ss->id_lock);
4701 tmp = idr_get_next(&ss->idr, &tmpid);
4702 spin_unlock(&ss->id_lock);
4706 if (tmp->depth >= depth && tmp->stack[depth] == rootid) {
4707 ret = rcu_dereference(tmp->css);
4713 /* continue to scan from next id */
4719 #ifdef CONFIG_CGROUP_DEBUG
4720 static struct cgroup_subsys_state *debug_create(struct cgroup_subsys *ss,
4721 struct cgroup *cont)
4723 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4726 return ERR_PTR(-ENOMEM);
4731 static void debug_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
4733 kfree(cont->subsys[debug_subsys_id]);
4736 static u64 cgroup_refcount_read(struct cgroup *cont, struct cftype *cft)
4738 return atomic_read(&cont->count);
4741 static u64 debug_taskcount_read(struct cgroup *cont, struct cftype *cft)
4743 return cgroup_task_count(cont);
4746 static u64 current_css_set_read(struct cgroup *cont, struct cftype *cft)
4748 return (u64)(unsigned long)current->cgroups;
4751 static u64 current_css_set_refcount_read(struct cgroup *cont,
4757 count = atomic_read(¤t->cgroups->refcount);
4762 static int current_css_set_cg_links_read(struct cgroup *cont,
4764 struct seq_file *seq)
4766 struct cg_cgroup_link *link;
4769 read_lock(&css_set_lock);
4771 cg = rcu_dereference(current->cgroups);
4772 list_for_each_entry(link, &cg->cg_links, cg_link_list) {
4773 struct cgroup *c = link->cgrp;
4777 name = c->dentry->d_name.name;
4780 seq_printf(seq, "Root %d group %s\n",
4781 c->root->hierarchy_id, name);
4784 read_unlock(&css_set_lock);
4788 #define MAX_TASKS_SHOWN_PER_CSS 25
4789 static int cgroup_css_links_read(struct cgroup *cont,
4791 struct seq_file *seq)
4793 struct cg_cgroup_link *link;
4795 read_lock(&css_set_lock);
4796 list_for_each_entry(link, &cont->css_sets, cgrp_link_list) {
4797 struct css_set *cg = link->cg;
4798 struct task_struct *task;
4800 seq_printf(seq, "css_set %p\n", cg);
4801 list_for_each_entry(task, &cg->tasks, cg_list) {
4802 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4803 seq_puts(seq, " ...\n");
4806 seq_printf(seq, " task %d\n",
4807 task_pid_vnr(task));
4811 read_unlock(&css_set_lock);
4815 static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
4817 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
4820 static struct cftype debug_files[] = {
4822 .name = "cgroup_refcount",
4823 .read_u64 = cgroup_refcount_read,
4826 .name = "taskcount",
4827 .read_u64 = debug_taskcount_read,
4831 .name = "current_css_set",
4832 .read_u64 = current_css_set_read,
4836 .name = "current_css_set_refcount",
4837 .read_u64 = current_css_set_refcount_read,
4841 .name = "current_css_set_cg_links",
4842 .read_seq_string = current_css_set_cg_links_read,
4846 .name = "cgroup_css_links",
4847 .read_seq_string = cgroup_css_links_read,
4851 .name = "releasable",
4852 .read_u64 = releasable_read,
4856 static int debug_populate(struct cgroup_subsys *ss, struct cgroup *cont)
4858 return cgroup_add_files(cont, ss, debug_files,
4859 ARRAY_SIZE(debug_files));
4862 struct cgroup_subsys debug_subsys = {
4864 .create = debug_create,
4865 .destroy = debug_destroy,
4866 .populate = debug_populate,
4867 .subsys_id = debug_subsys_id,
4869 #endif /* CONFIG_CGROUP_DEBUG */