2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Copyright notices from the original cpuset code:
8 * --------------------------------------------------
9 * Copyright (C) 2003 BULL SA.
10 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
12 * Portions derived from Patrick Mochel's sysfs code.
13 * sysfs is Copyright (c) 2001-3 Patrick Mochel
15 * 2003-10-10 Written by Simon Derr.
16 * 2003-10-22 Updates by Stephen Hemminger.
17 * 2004 May-July Rework by Paul Jackson.
18 * ---------------------------------------------------
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cgroup.h>
26 #include <linux/ctype.h>
27 #include <linux/errno.h>
29 #include <linux/kernel.h>
30 #include <linux/list.h>
32 #include <linux/mutex.h>
33 #include <linux/mount.h>
34 #include <linux/pagemap.h>
35 #include <linux/proc_fs.h>
36 #include <linux/rcupdate.h>
37 #include <linux/sched.h>
38 #include <linux/backing-dev.h>
39 #include <linux/seq_file.h>
40 #include <linux/slab.h>
41 #include <linux/magic.h>
42 #include <linux/spinlock.h>
43 #include <linux/string.h>
44 #include <linux/sort.h>
45 #include <linux/kmod.h>
46 #include <linux/delayacct.h>
47 #include <linux/cgroupstats.h>
48 #include <linux/hash.h>
49 #include <linux/namei.h>
50 #include <linux/smp_lock.h>
51 #include <linux/pid_namespace.h>
52 #include <linux/idr.h>
54 #include <asm/atomic.h>
56 static DEFINE_MUTEX(cgroup_mutex);
58 /* Generate an array of cgroup subsystem pointers */
59 #define SUBSYS(_x) &_x ## _subsys,
61 static struct cgroup_subsys *subsys[] = {
62 #include <linux/cgroup_subsys.h>
65 #define MAX_CGROUP_ROOT_NAMELEN 64
68 * A cgroupfs_root represents the root of a cgroup hierarchy,
69 * and may be associated with a superblock to form an active
72 struct cgroupfs_root {
73 struct super_block *sb;
76 * The bitmask of subsystems intended to be attached to this
79 unsigned long subsys_bits;
81 /* Unique id for this hierarchy. */
84 /* The bitmask of subsystems currently attached to this hierarchy */
85 unsigned long actual_subsys_bits;
87 /* A list running through the attached subsystems */
88 struct list_head subsys_list;
90 /* The root cgroup for this hierarchy */
91 struct cgroup top_cgroup;
93 /* Tracks how many cgroups are currently defined in hierarchy.*/
94 int number_of_cgroups;
96 /* A list running through the active hierarchies */
97 struct list_head root_list;
99 /* Hierarchy-specific flags */
102 /* The path to use for release notifications. */
103 char release_agent_path[PATH_MAX];
105 /* The name for this hierarchy - may be empty */
106 char name[MAX_CGROUP_ROOT_NAMELEN];
110 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
111 * subsystems that are otherwise unattached - it never has more than a
112 * single cgroup, and all tasks are part of that cgroup.
114 static struct cgroupfs_root rootnode;
117 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
118 * cgroup_subsys->use_id != 0.
120 #define CSS_ID_MAX (65535)
123 * The css to which this ID points. This pointer is set to valid value
124 * after cgroup is populated. If cgroup is removed, this will be NULL.
125 * This pointer is expected to be RCU-safe because destroy()
126 * is called after synchronize_rcu(). But for safe use, css_is_removed()
127 * css_tryget() should be used for avoiding race.
129 struct cgroup_subsys_state *css;
135 * Depth in hierarchy which this ID belongs to.
137 unsigned short depth;
139 * ID is freed by RCU. (and lookup routine is RCU safe.)
141 struct rcu_head rcu_head;
143 * Hierarchy of CSS ID belongs to.
145 unsigned short stack[0]; /* Array of Length (depth+1) */
149 /* The list of hierarchy roots */
151 static LIST_HEAD(roots);
152 static int root_count;
154 static DEFINE_IDA(hierarchy_ida);
155 static int next_hierarchy_id;
156 static DEFINE_SPINLOCK(hierarchy_id_lock);
158 /* dummytop is a shorthand for the dummy hierarchy's top cgroup */
159 #define dummytop (&rootnode.top_cgroup)
161 /* This flag indicates whether tasks in the fork and exit paths should
162 * check for fork/exit handlers to call. This avoids us having to do
163 * extra work in the fork/exit path if none of the subsystems need to
166 static int need_forkexit_callback __read_mostly;
168 /* convenient tests for these bits */
169 inline int cgroup_is_removed(const struct cgroup *cgrp)
171 return test_bit(CGRP_REMOVED, &cgrp->flags);
174 /* bits in struct cgroupfs_root flags field */
176 ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
179 static int cgroup_is_releasable(const struct cgroup *cgrp)
182 (1 << CGRP_RELEASABLE) |
183 (1 << CGRP_NOTIFY_ON_RELEASE);
184 return (cgrp->flags & bits) == bits;
187 static int notify_on_release(const struct cgroup *cgrp)
189 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
193 * for_each_subsys() allows you to iterate on each subsystem attached to
194 * an active hierarchy
196 #define for_each_subsys(_root, _ss) \
197 list_for_each_entry(_ss, &_root->subsys_list, sibling)
199 /* for_each_active_root() allows you to iterate across the active hierarchies */
200 #define for_each_active_root(_root) \
201 list_for_each_entry(_root, &roots, root_list)
203 /* the list of cgroups eligible for automatic release. Protected by
204 * release_list_lock */
205 static LIST_HEAD(release_list);
206 static DEFINE_SPINLOCK(release_list_lock);
207 static void cgroup_release_agent(struct work_struct *work);
208 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
209 static void check_for_release(struct cgroup *cgrp);
211 /* Link structure for associating css_set objects with cgroups */
212 struct cg_cgroup_link {
214 * List running through cg_cgroup_links associated with a
215 * cgroup, anchored on cgroup->css_sets
217 struct list_head cgrp_link_list;
220 * List running through cg_cgroup_links pointing at a
221 * single css_set object, anchored on css_set->cg_links
223 struct list_head cg_link_list;
227 /* The default css_set - used by init and its children prior to any
228 * hierarchies being mounted. It contains a pointer to the root state
229 * for each subsystem. Also used to anchor the list of css_sets. Not
230 * reference-counted, to improve performance when child cgroups
231 * haven't been created.
234 static struct css_set init_css_set;
235 static struct cg_cgroup_link init_css_set_link;
237 static int cgroup_subsys_init_idr(struct cgroup_subsys *ss);
239 /* css_set_lock protects the list of css_set objects, and the
240 * chain of tasks off each css_set. Nests outside task->alloc_lock
241 * due to cgroup_iter_start() */
242 static DEFINE_RWLOCK(css_set_lock);
243 static int css_set_count;
246 * hash table for cgroup groups. This improves the performance to find
247 * an existing css_set. This hash doesn't (currently) take into
248 * account cgroups in empty hierarchies.
250 #define CSS_SET_HASH_BITS 7
251 #define CSS_SET_TABLE_SIZE (1 << CSS_SET_HASH_BITS)
252 static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
254 static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
258 unsigned long tmp = 0UL;
260 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
261 tmp += (unsigned long)css[i];
262 tmp = (tmp >> 16) ^ tmp;
264 index = hash_long(tmp, CSS_SET_HASH_BITS);
266 return &css_set_table[index];
269 /* We don't maintain the lists running through each css_set to its
270 * task until after the first call to cgroup_iter_start(). This
271 * reduces the fork()/exit() overhead for people who have cgroups
272 * compiled into their kernel but not actually in use */
273 static int use_task_css_set_links __read_mostly;
275 static void __put_css_set(struct css_set *cg, int taskexit)
277 struct cg_cgroup_link *link;
278 struct cg_cgroup_link *saved_link;
280 * Ensure that the refcount doesn't hit zero while any readers
281 * can see it. Similar to atomic_dec_and_lock(), but for an
284 if (atomic_add_unless(&cg->refcount, -1, 1))
286 write_lock(&css_set_lock);
287 if (!atomic_dec_and_test(&cg->refcount)) {
288 write_unlock(&css_set_lock);
292 /* This css_set is dead. unlink it and release cgroup refcounts */
293 hlist_del(&cg->hlist);
296 list_for_each_entry_safe(link, saved_link, &cg->cg_links,
298 struct cgroup *cgrp = link->cgrp;
299 list_del(&link->cg_link_list);
300 list_del(&link->cgrp_link_list);
301 if (atomic_dec_and_test(&cgrp->count) &&
302 notify_on_release(cgrp)) {
304 set_bit(CGRP_RELEASABLE, &cgrp->flags);
305 check_for_release(cgrp);
311 write_unlock(&css_set_lock);
316 * refcounted get/put for css_set objects
318 static inline void get_css_set(struct css_set *cg)
320 atomic_inc(&cg->refcount);
323 static inline void put_css_set(struct css_set *cg)
325 __put_css_set(cg, 0);
328 static inline void put_css_set_taskexit(struct css_set *cg)
330 __put_css_set(cg, 1);
334 * compare_css_sets - helper function for find_existing_css_set().
335 * @cg: candidate css_set being tested
336 * @old_cg: existing css_set for a task
337 * @new_cgrp: cgroup that's being entered by the task
338 * @template: desired set of css pointers in css_set (pre-calculated)
340 * Returns true if "cg" matches "old_cg" except for the hierarchy
341 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
343 static bool compare_css_sets(struct css_set *cg,
344 struct css_set *old_cg,
345 struct cgroup *new_cgrp,
346 struct cgroup_subsys_state *template[])
348 struct list_head *l1, *l2;
350 if (memcmp(template, cg->subsys, sizeof(cg->subsys))) {
351 /* Not all subsystems matched */
356 * Compare cgroup pointers in order to distinguish between
357 * different cgroups in heirarchies with no subsystems. We
358 * could get by with just this check alone (and skip the
359 * memcmp above) but on most setups the memcmp check will
360 * avoid the need for this more expensive check on almost all
365 l2 = &old_cg->cg_links;
367 struct cg_cgroup_link *cgl1, *cgl2;
368 struct cgroup *cg1, *cg2;
372 /* See if we reached the end - both lists are equal length. */
373 if (l1 == &cg->cg_links) {
374 BUG_ON(l2 != &old_cg->cg_links);
377 BUG_ON(l2 == &old_cg->cg_links);
379 /* Locate the cgroups associated with these links. */
380 cgl1 = list_entry(l1, struct cg_cgroup_link, cg_link_list);
381 cgl2 = list_entry(l2, struct cg_cgroup_link, cg_link_list);
384 /* Hierarchies should be linked in the same order. */
385 BUG_ON(cg1->root != cg2->root);
388 * If this hierarchy is the hierarchy of the cgroup
389 * that's changing, then we need to check that this
390 * css_set points to the new cgroup; if it's any other
391 * hierarchy, then this css_set should point to the
392 * same cgroup as the old css_set.
394 if (cg1->root == new_cgrp->root) {
406 * find_existing_css_set() is a helper for
407 * find_css_set(), and checks to see whether an existing
408 * css_set is suitable.
410 * oldcg: the cgroup group that we're using before the cgroup
413 * cgrp: the cgroup that we're moving into
415 * template: location in which to build the desired set of subsystem
416 * state objects for the new cgroup group
418 static struct css_set *find_existing_css_set(
419 struct css_set *oldcg,
421 struct cgroup_subsys_state *template[])
424 struct cgroupfs_root *root = cgrp->root;
425 struct hlist_head *hhead;
426 struct hlist_node *node;
429 /* Built the set of subsystem state objects that we want to
430 * see in the new css_set */
431 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
432 if (root->subsys_bits & (1UL << i)) {
433 /* Subsystem is in this hierarchy. So we want
434 * the subsystem state from the new
436 template[i] = cgrp->subsys[i];
438 /* Subsystem is not in this hierarchy, so we
439 * don't want to change the subsystem state */
440 template[i] = oldcg->subsys[i];
444 hhead = css_set_hash(template);
445 hlist_for_each_entry(cg, node, hhead, hlist) {
446 if (!compare_css_sets(cg, oldcg, cgrp, template))
449 /* This css_set matches what we need */
453 /* No existing cgroup group matched */
457 static void free_cg_links(struct list_head *tmp)
459 struct cg_cgroup_link *link;
460 struct cg_cgroup_link *saved_link;
462 list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
463 list_del(&link->cgrp_link_list);
469 * allocate_cg_links() allocates "count" cg_cgroup_link structures
470 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
471 * success or a negative error
473 static int allocate_cg_links(int count, struct list_head *tmp)
475 struct cg_cgroup_link *link;
478 for (i = 0; i < count; i++) {
479 link = kmalloc(sizeof(*link), GFP_KERNEL);
484 list_add(&link->cgrp_link_list, tmp);
490 * link_css_set - a helper function to link a css_set to a cgroup
491 * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
492 * @cg: the css_set to be linked
493 * @cgrp: the destination cgroup
495 static void link_css_set(struct list_head *tmp_cg_links,
496 struct css_set *cg, struct cgroup *cgrp)
498 struct cg_cgroup_link *link;
500 BUG_ON(list_empty(tmp_cg_links));
501 link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
505 atomic_inc(&cgrp->count);
506 list_move(&link->cgrp_link_list, &cgrp->css_sets);
508 * Always add links to the tail of the list so that the list
509 * is sorted by order of hierarchy creation
511 list_add_tail(&link->cg_link_list, &cg->cg_links);
515 * find_css_set() takes an existing cgroup group and a
516 * cgroup object, and returns a css_set object that's
517 * equivalent to the old group, but with the given cgroup
518 * substituted into the appropriate hierarchy. Must be called with
521 static struct css_set *find_css_set(
522 struct css_set *oldcg, struct cgroup *cgrp)
525 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
527 struct list_head tmp_cg_links;
529 struct hlist_head *hhead;
530 struct cg_cgroup_link *link;
532 /* First see if we already have a cgroup group that matches
534 read_lock(&css_set_lock);
535 res = find_existing_css_set(oldcg, cgrp, template);
538 read_unlock(&css_set_lock);
543 res = kmalloc(sizeof(*res), GFP_KERNEL);
547 /* Allocate all the cg_cgroup_link objects that we'll need */
548 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
553 atomic_set(&res->refcount, 1);
554 INIT_LIST_HEAD(&res->cg_links);
555 INIT_LIST_HEAD(&res->tasks);
556 INIT_HLIST_NODE(&res->hlist);
558 /* Copy the set of subsystem state objects generated in
559 * find_existing_css_set() */
560 memcpy(res->subsys, template, sizeof(res->subsys));
562 write_lock(&css_set_lock);
563 /* Add reference counts and links from the new css_set. */
564 list_for_each_entry(link, &oldcg->cg_links, cg_link_list) {
565 struct cgroup *c = link->cgrp;
566 if (c->root == cgrp->root)
568 link_css_set(&tmp_cg_links, res, c);
571 BUG_ON(!list_empty(&tmp_cg_links));
575 /* Add this cgroup group to the hash table */
576 hhead = css_set_hash(res->subsys);
577 hlist_add_head(&res->hlist, hhead);
579 write_unlock(&css_set_lock);
585 * Return the cgroup for "task" from the given hierarchy. Must be
586 * called with cgroup_mutex held.
588 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
589 struct cgroupfs_root *root)
592 struct cgroup *res = NULL;
594 BUG_ON(!mutex_is_locked(&cgroup_mutex));
595 read_lock(&css_set_lock);
597 * No need to lock the task - since we hold cgroup_mutex the
598 * task can't change groups, so the only thing that can happen
599 * is that it exits and its css is set back to init_css_set.
602 if (css == &init_css_set) {
603 res = &root->top_cgroup;
605 struct cg_cgroup_link *link;
606 list_for_each_entry(link, &css->cg_links, cg_link_list) {
607 struct cgroup *c = link->cgrp;
608 if (c->root == root) {
614 read_unlock(&css_set_lock);
620 * There is one global cgroup mutex. We also require taking
621 * task_lock() when dereferencing a task's cgroup subsys pointers.
622 * See "The task_lock() exception", at the end of this comment.
624 * A task must hold cgroup_mutex to modify cgroups.
626 * Any task can increment and decrement the count field without lock.
627 * So in general, code holding cgroup_mutex can't rely on the count
628 * field not changing. However, if the count goes to zero, then only
629 * cgroup_attach_task() can increment it again. Because a count of zero
630 * means that no tasks are currently attached, therefore there is no
631 * way a task attached to that cgroup can fork (the other way to
632 * increment the count). So code holding cgroup_mutex can safely
633 * assume that if the count is zero, it will stay zero. Similarly, if
634 * a task holds cgroup_mutex on a cgroup with zero count, it
635 * knows that the cgroup won't be removed, as cgroup_rmdir()
638 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
639 * (usually) take cgroup_mutex. These are the two most performance
640 * critical pieces of code here. The exception occurs on cgroup_exit(),
641 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
642 * is taken, and if the cgroup count is zero, a usermode call made
643 * to the release agent with the name of the cgroup (path relative to
644 * the root of cgroup file system) as the argument.
646 * A cgroup can only be deleted if both its 'count' of using tasks
647 * is zero, and its list of 'children' cgroups is empty. Since all
648 * tasks in the system use _some_ cgroup, and since there is always at
649 * least one task in the system (init, pid == 1), therefore, top_cgroup
650 * always has either children cgroups and/or using tasks. So we don't
651 * need a special hack to ensure that top_cgroup cannot be deleted.
653 * The task_lock() exception
655 * The need for this exception arises from the action of
656 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
657 * another. It does so using cgroup_mutex, however there are
658 * several performance critical places that need to reference
659 * task->cgroup without the expense of grabbing a system global
660 * mutex. Therefore except as noted below, when dereferencing or, as
661 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
662 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
663 * the task_struct routinely used for such matters.
665 * P.S. One more locking exception. RCU is used to guard the
666 * update of a tasks cgroup pointer by cgroup_attach_task()
670 * cgroup_lock - lock out any changes to cgroup structures
673 void cgroup_lock(void)
675 mutex_lock(&cgroup_mutex);
679 * cgroup_unlock - release lock on cgroup changes
681 * Undo the lock taken in a previous cgroup_lock() call.
683 void cgroup_unlock(void)
685 mutex_unlock(&cgroup_mutex);
689 * A couple of forward declarations required, due to cyclic reference loop:
690 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
691 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
695 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
696 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
697 static int cgroup_populate_dir(struct cgroup *cgrp);
698 static const struct inode_operations cgroup_dir_inode_operations;
699 static struct file_operations proc_cgroupstats_operations;
701 static struct backing_dev_info cgroup_backing_dev_info = {
703 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
706 static int alloc_css_id(struct cgroup_subsys *ss,
707 struct cgroup *parent, struct cgroup *child);
709 static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
711 struct inode *inode = new_inode(sb);
714 inode->i_mode = mode;
715 inode->i_uid = current_fsuid();
716 inode->i_gid = current_fsgid();
717 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
718 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
724 * Call subsys's pre_destroy handler.
725 * This is called before css refcnt check.
727 static int cgroup_call_pre_destroy(struct cgroup *cgrp)
729 struct cgroup_subsys *ss;
732 for_each_subsys(cgrp->root, ss)
733 if (ss->pre_destroy) {
734 ret = ss->pre_destroy(ss, cgrp);
741 static void free_cgroup_rcu(struct rcu_head *obj)
743 struct cgroup *cgrp = container_of(obj, struct cgroup, rcu_head);
748 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
750 /* is dentry a directory ? if so, kfree() associated cgroup */
751 if (S_ISDIR(inode->i_mode)) {
752 struct cgroup *cgrp = dentry->d_fsdata;
753 struct cgroup_subsys *ss;
754 BUG_ON(!(cgroup_is_removed(cgrp)));
755 /* It's possible for external users to be holding css
756 * reference counts on a cgroup; css_put() needs to
757 * be able to access the cgroup after decrementing
758 * the reference count in order to know if it needs to
759 * queue the cgroup to be handled by the release
763 mutex_lock(&cgroup_mutex);
765 * Release the subsystem state objects.
767 for_each_subsys(cgrp->root, ss)
768 ss->destroy(ss, cgrp);
770 cgrp->root->number_of_cgroups--;
771 mutex_unlock(&cgroup_mutex);
774 * Drop the active superblock reference that we took when we
777 deactivate_super(cgrp->root->sb);
779 call_rcu(&cgrp->rcu_head, free_cgroup_rcu);
784 static void remove_dir(struct dentry *d)
786 struct dentry *parent = dget(d->d_parent);
789 simple_rmdir(parent->d_inode, d);
793 static void cgroup_clear_directory(struct dentry *dentry)
795 struct list_head *node;
797 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
798 spin_lock(&dcache_lock);
799 node = dentry->d_subdirs.next;
800 while (node != &dentry->d_subdirs) {
801 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
804 /* This should never be called on a cgroup
805 * directory with child cgroups */
806 BUG_ON(d->d_inode->i_mode & S_IFDIR);
808 spin_unlock(&dcache_lock);
810 simple_unlink(dentry->d_inode, d);
812 spin_lock(&dcache_lock);
814 node = dentry->d_subdirs.next;
816 spin_unlock(&dcache_lock);
820 * NOTE : the dentry must have been dget()'ed
822 static void cgroup_d_remove_dir(struct dentry *dentry)
824 cgroup_clear_directory(dentry);
826 spin_lock(&dcache_lock);
827 list_del_init(&dentry->d_u.d_child);
828 spin_unlock(&dcache_lock);
833 * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
834 * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
835 * reference to css->refcnt. In general, this refcnt is expected to goes down
838 * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
840 DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
842 static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
844 if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
845 wake_up_all(&cgroup_rmdir_waitq);
848 void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
853 void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
855 cgroup_wakeup_rmdir_waiter(css->cgroup);
860 static int rebind_subsystems(struct cgroupfs_root *root,
861 unsigned long final_bits)
863 unsigned long added_bits, removed_bits;
864 struct cgroup *cgrp = &root->top_cgroup;
867 removed_bits = root->actual_subsys_bits & ~final_bits;
868 added_bits = final_bits & ~root->actual_subsys_bits;
869 /* Check that any added subsystems are currently free */
870 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
871 unsigned long bit = 1UL << i;
872 struct cgroup_subsys *ss = subsys[i];
873 if (!(bit & added_bits))
875 if (ss->root != &rootnode) {
876 /* Subsystem isn't free */
881 /* Currently we don't handle adding/removing subsystems when
882 * any child cgroups exist. This is theoretically supportable
883 * but involves complex error handling, so it's being left until
885 if (root->number_of_cgroups > 1)
888 /* Process each subsystem */
889 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
890 struct cgroup_subsys *ss = subsys[i];
891 unsigned long bit = 1UL << i;
892 if (bit & added_bits) {
893 /* We're binding this subsystem to this hierarchy */
894 BUG_ON(cgrp->subsys[i]);
895 BUG_ON(!dummytop->subsys[i]);
896 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
897 mutex_lock(&ss->hierarchy_mutex);
898 cgrp->subsys[i] = dummytop->subsys[i];
899 cgrp->subsys[i]->cgroup = cgrp;
900 list_move(&ss->sibling, &root->subsys_list);
904 mutex_unlock(&ss->hierarchy_mutex);
905 } else if (bit & removed_bits) {
906 /* We're removing this subsystem */
907 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
908 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
909 mutex_lock(&ss->hierarchy_mutex);
911 ss->bind(ss, dummytop);
912 dummytop->subsys[i]->cgroup = dummytop;
913 cgrp->subsys[i] = NULL;
914 subsys[i]->root = &rootnode;
915 list_move(&ss->sibling, &rootnode.subsys_list);
916 mutex_unlock(&ss->hierarchy_mutex);
917 } else if (bit & final_bits) {
918 /* Subsystem state should already exist */
919 BUG_ON(!cgrp->subsys[i]);
921 /* Subsystem state shouldn't exist */
922 BUG_ON(cgrp->subsys[i]);
925 root->subsys_bits = root->actual_subsys_bits = final_bits;
931 static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
933 struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
934 struct cgroup_subsys *ss;
936 mutex_lock(&cgroup_mutex);
937 for_each_subsys(root, ss)
938 seq_printf(seq, ",%s", ss->name);
939 if (test_bit(ROOT_NOPREFIX, &root->flags))
940 seq_puts(seq, ",noprefix");
941 if (strlen(root->release_agent_path))
942 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
943 if (strlen(root->name))
944 seq_printf(seq, ",name=%s", root->name);
945 mutex_unlock(&cgroup_mutex);
949 struct cgroup_sb_opts {
950 unsigned long subsys_bits;
954 /* User explicitly requested empty subsystem */
957 struct cgroupfs_root *new_root;
961 /* Convert a hierarchy specifier into a bitmask of subsystems and
963 static int parse_cgroupfs_options(char *data,
964 struct cgroup_sb_opts *opts)
966 char *token, *o = data ?: "all";
967 unsigned long mask = (unsigned long)-1;
969 #ifdef CONFIG_CPUSETS
970 mask = ~(1UL << cpuset_subsys_id);
973 memset(opts, 0, sizeof(*opts));
975 while ((token = strsep(&o, ",")) != NULL) {
978 if (!strcmp(token, "all")) {
979 /* Add all non-disabled subsystems */
981 opts->subsys_bits = 0;
982 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
983 struct cgroup_subsys *ss = subsys[i];
985 opts->subsys_bits |= 1ul << i;
987 } else if (!strcmp(token, "none")) {
988 /* Explicitly have no subsystems */
990 } else if (!strcmp(token, "noprefix")) {
991 set_bit(ROOT_NOPREFIX, &opts->flags);
992 } else if (!strncmp(token, "release_agent=", 14)) {
993 /* Specifying two release agents is forbidden */
994 if (opts->release_agent)
996 opts->release_agent =
997 kstrndup(token + 14, PATH_MAX, GFP_KERNEL);
998 if (!opts->release_agent)
1000 } else if (!strncmp(token, "name=", 5)) {
1002 const char *name = token + 5;
1003 /* Can't specify an empty name */
1006 /* Must match [\w.-]+ */
1007 for (i = 0; i < strlen(name); i++) {
1011 if ((c == '.') || (c == '-') || (c == '_'))
1015 /* Specifying two names is forbidden */
1018 opts->name = kstrndup(name,
1019 MAX_CGROUP_ROOT_NAMELEN,
1024 struct cgroup_subsys *ss;
1026 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1028 if (!strcmp(token, ss->name)) {
1030 set_bit(i, &opts->subsys_bits);
1034 if (i == CGROUP_SUBSYS_COUNT)
1039 /* Consistency checks */
1042 * Option noprefix was introduced just for backward compatibility
1043 * with the old cpuset, so we allow noprefix only if mounting just
1044 * the cpuset subsystem.
1046 if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
1047 (opts->subsys_bits & mask))
1051 /* Can't specify "none" and some subsystems */
1052 if (opts->subsys_bits && opts->none)
1056 * We either have to specify by name or by subsystems. (So all
1057 * empty hierarchies must have a name).
1059 if (!opts->subsys_bits && !opts->name)
1065 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1068 struct cgroupfs_root *root = sb->s_fs_info;
1069 struct cgroup *cgrp = &root->top_cgroup;
1070 struct cgroup_sb_opts opts;
1073 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1074 mutex_lock(&cgroup_mutex);
1076 /* See what subsystems are wanted */
1077 ret = parse_cgroupfs_options(data, &opts);
1081 /* Don't allow flags to change at remount */
1082 if (opts.flags != root->flags) {
1087 /* Don't allow name to change at remount */
1088 if (opts.name && strcmp(opts.name, root->name)) {
1093 ret = rebind_subsystems(root, opts.subsys_bits);
1097 /* (re)populate subsystem files */
1098 cgroup_populate_dir(cgrp);
1100 if (opts.release_agent)
1101 strcpy(root->release_agent_path, opts.release_agent);
1103 kfree(opts.release_agent);
1105 mutex_unlock(&cgroup_mutex);
1106 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1111 static const struct super_operations cgroup_ops = {
1112 .statfs = simple_statfs,
1113 .drop_inode = generic_delete_inode,
1114 .show_options = cgroup_show_options,
1115 .remount_fs = cgroup_remount,
1118 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1120 INIT_LIST_HEAD(&cgrp->sibling);
1121 INIT_LIST_HEAD(&cgrp->children);
1122 INIT_LIST_HEAD(&cgrp->css_sets);
1123 INIT_LIST_HEAD(&cgrp->release_list);
1124 INIT_LIST_HEAD(&cgrp->pids_list);
1125 init_rwsem(&cgrp->pids_mutex);
1128 static void init_cgroup_root(struct cgroupfs_root *root)
1130 struct cgroup *cgrp = &root->top_cgroup;
1131 INIT_LIST_HEAD(&root->subsys_list);
1132 INIT_LIST_HEAD(&root->root_list);
1133 root->number_of_cgroups = 1;
1135 cgrp->top_cgroup = cgrp;
1136 init_cgroup_housekeeping(cgrp);
1139 static bool init_root_id(struct cgroupfs_root *root)
1144 if (!ida_pre_get(&hierarchy_ida, GFP_KERNEL))
1146 spin_lock(&hierarchy_id_lock);
1147 /* Try to allocate the next unused ID */
1148 ret = ida_get_new_above(&hierarchy_ida, next_hierarchy_id,
1149 &root->hierarchy_id);
1151 /* Try again starting from 0 */
1152 ret = ida_get_new(&hierarchy_ida, &root->hierarchy_id);
1154 next_hierarchy_id = root->hierarchy_id + 1;
1155 } else if (ret != -EAGAIN) {
1156 /* Can only get here if the 31-bit IDR is full ... */
1159 spin_unlock(&hierarchy_id_lock);
1164 static int cgroup_test_super(struct super_block *sb, void *data)
1166 struct cgroup_sb_opts *opts = data;
1167 struct cgroupfs_root *root = sb->s_fs_info;
1169 /* If we asked for a name then it must match */
1170 if (opts->name && strcmp(opts->name, root->name))
1174 * If we asked for subsystems (or explicitly for no
1175 * subsystems) then they must match
1177 if ((opts->subsys_bits || opts->none)
1178 && (opts->subsys_bits != root->subsys_bits))
1184 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1186 struct cgroupfs_root *root;
1188 if (!opts->subsys_bits && !opts->none)
1191 root = kzalloc(sizeof(*root), GFP_KERNEL);
1193 return ERR_PTR(-ENOMEM);
1195 if (!init_root_id(root)) {
1197 return ERR_PTR(-ENOMEM);
1199 init_cgroup_root(root);
1201 root->subsys_bits = opts->subsys_bits;
1202 root->flags = opts->flags;
1203 if (opts->release_agent)
1204 strcpy(root->release_agent_path, opts->release_agent);
1206 strcpy(root->name, opts->name);
1210 static void cgroup_drop_root(struct cgroupfs_root *root)
1215 BUG_ON(!root->hierarchy_id);
1216 spin_lock(&hierarchy_id_lock);
1217 ida_remove(&hierarchy_ida, root->hierarchy_id);
1218 spin_unlock(&hierarchy_id_lock);
1222 static int cgroup_set_super(struct super_block *sb, void *data)
1225 struct cgroup_sb_opts *opts = data;
1227 /* If we don't have a new root, we can't set up a new sb */
1228 if (!opts->new_root)
1231 BUG_ON(!opts->subsys_bits && !opts->none);
1233 ret = set_anon_super(sb, NULL);
1237 sb->s_fs_info = opts->new_root;
1238 opts->new_root->sb = sb;
1240 sb->s_blocksize = PAGE_CACHE_SIZE;
1241 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1242 sb->s_magic = CGROUP_SUPER_MAGIC;
1243 sb->s_op = &cgroup_ops;
1248 static int cgroup_get_rootdir(struct super_block *sb)
1250 struct inode *inode =
1251 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1252 struct dentry *dentry;
1257 inode->i_fop = &simple_dir_operations;
1258 inode->i_op = &cgroup_dir_inode_operations;
1259 /* directories start off with i_nlink == 2 (for "." entry) */
1261 dentry = d_alloc_root(inode);
1266 sb->s_root = dentry;
1270 static int cgroup_get_sb(struct file_system_type *fs_type,
1271 int flags, const char *unused_dev_name,
1272 void *data, struct vfsmount *mnt)
1274 struct cgroup_sb_opts opts;
1275 struct cgroupfs_root *root;
1277 struct super_block *sb;
1278 struct cgroupfs_root *new_root;
1280 /* First find the desired set of subsystems */
1281 ret = parse_cgroupfs_options(data, &opts);
1286 * Allocate a new cgroup root. We may not need it if we're
1287 * reusing an existing hierarchy.
1289 new_root = cgroup_root_from_opts(&opts);
1290 if (IS_ERR(new_root)) {
1291 ret = PTR_ERR(new_root);
1294 opts.new_root = new_root;
1296 /* Locate an existing or new sb for this hierarchy */
1297 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
1300 cgroup_drop_root(opts.new_root);
1304 root = sb->s_fs_info;
1306 if (root == opts.new_root) {
1307 /* We used the new root structure, so this is a new hierarchy */
1308 struct list_head tmp_cg_links;
1309 struct cgroup *root_cgrp = &root->top_cgroup;
1310 struct inode *inode;
1311 struct cgroupfs_root *existing_root;
1314 BUG_ON(sb->s_root != NULL);
1316 ret = cgroup_get_rootdir(sb);
1318 goto drop_new_super;
1319 inode = sb->s_root->d_inode;
1321 mutex_lock(&inode->i_mutex);
1322 mutex_lock(&cgroup_mutex);
1324 if (strlen(root->name)) {
1325 /* Check for name clashes with existing mounts */
1326 for_each_active_root(existing_root) {
1327 if (!strcmp(existing_root->name, root->name)) {
1329 mutex_unlock(&cgroup_mutex);
1330 mutex_unlock(&inode->i_mutex);
1331 goto drop_new_super;
1337 * We're accessing css_set_count without locking
1338 * css_set_lock here, but that's OK - it can only be
1339 * increased by someone holding cgroup_lock, and
1340 * that's us. The worst that can happen is that we
1341 * have some link structures left over
1343 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1345 mutex_unlock(&cgroup_mutex);
1346 mutex_unlock(&inode->i_mutex);
1347 goto drop_new_super;
1350 ret = rebind_subsystems(root, root->subsys_bits);
1351 if (ret == -EBUSY) {
1352 mutex_unlock(&cgroup_mutex);
1353 mutex_unlock(&inode->i_mutex);
1354 free_cg_links(&tmp_cg_links);
1355 goto drop_new_super;
1358 /* EBUSY should be the only error here */
1361 list_add(&root->root_list, &roots);
1364 sb->s_root->d_fsdata = root_cgrp;
1365 root->top_cgroup.dentry = sb->s_root;
1367 /* Link the top cgroup in this hierarchy into all
1368 * the css_set objects */
1369 write_lock(&css_set_lock);
1370 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1371 struct hlist_head *hhead = &css_set_table[i];
1372 struct hlist_node *node;
1375 hlist_for_each_entry(cg, node, hhead, hlist)
1376 link_css_set(&tmp_cg_links, cg, root_cgrp);
1378 write_unlock(&css_set_lock);
1380 free_cg_links(&tmp_cg_links);
1382 BUG_ON(!list_empty(&root_cgrp->sibling));
1383 BUG_ON(!list_empty(&root_cgrp->children));
1384 BUG_ON(root->number_of_cgroups != 1);
1386 cgroup_populate_dir(root_cgrp);
1387 mutex_unlock(&cgroup_mutex);
1388 mutex_unlock(&inode->i_mutex);
1391 * We re-used an existing hierarchy - the new root (if
1392 * any) is not needed
1394 cgroup_drop_root(opts.new_root);
1397 simple_set_mnt(mnt, sb);
1398 kfree(opts.release_agent);
1403 deactivate_locked_super(sb);
1405 kfree(opts.release_agent);
1411 static void cgroup_kill_sb(struct super_block *sb) {
1412 struct cgroupfs_root *root = sb->s_fs_info;
1413 struct cgroup *cgrp = &root->top_cgroup;
1415 struct cg_cgroup_link *link;
1416 struct cg_cgroup_link *saved_link;
1420 BUG_ON(root->number_of_cgroups != 1);
1421 BUG_ON(!list_empty(&cgrp->children));
1422 BUG_ON(!list_empty(&cgrp->sibling));
1424 mutex_lock(&cgroup_mutex);
1426 /* Rebind all subsystems back to the default hierarchy */
1427 ret = rebind_subsystems(root, 0);
1428 /* Shouldn't be able to fail ... */
1432 * Release all the links from css_sets to this hierarchy's
1435 write_lock(&css_set_lock);
1437 list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1439 list_del(&link->cg_link_list);
1440 list_del(&link->cgrp_link_list);
1443 write_unlock(&css_set_lock);
1445 if (!list_empty(&root->root_list)) {
1446 list_del(&root->root_list);
1450 mutex_unlock(&cgroup_mutex);
1452 kill_litter_super(sb);
1453 cgroup_drop_root(root);
1456 static struct file_system_type cgroup_fs_type = {
1458 .get_sb = cgroup_get_sb,
1459 .kill_sb = cgroup_kill_sb,
1462 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1464 return dentry->d_fsdata;
1467 static inline struct cftype *__d_cft(struct dentry *dentry)
1469 return dentry->d_fsdata;
1473 * cgroup_path - generate the path of a cgroup
1474 * @cgrp: the cgroup in question
1475 * @buf: the buffer to write the path into
1476 * @buflen: the length of the buffer
1478 * Called with cgroup_mutex held or else with an RCU-protected cgroup
1479 * reference. Writes path of cgroup into buf. Returns 0 on success,
1482 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1485 struct dentry *dentry = rcu_dereference(cgrp->dentry);
1487 if (!dentry || cgrp == dummytop) {
1489 * Inactive subsystems have no dentry for their root
1496 start = buf + buflen;
1500 int len = dentry->d_name.len;
1501 if ((start -= len) < buf)
1502 return -ENAMETOOLONG;
1503 memcpy(start, cgrp->dentry->d_name.name, len);
1504 cgrp = cgrp->parent;
1507 dentry = rcu_dereference(cgrp->dentry);
1511 return -ENAMETOOLONG;
1514 memmove(buf, start, buf + buflen - start);
1519 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1520 * @cgrp: the cgroup the task is attaching to
1521 * @tsk: the task to be attached
1523 * Call holding cgroup_mutex. May take task_lock of
1524 * the task 'tsk' during call.
1526 int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1529 struct cgroup_subsys *ss;
1530 struct cgroup *oldcgrp;
1532 struct css_set *newcg;
1533 struct cgroupfs_root *root = cgrp->root;
1535 /* Nothing to do if the task is already in that cgroup */
1536 oldcgrp = task_cgroup_from_root(tsk, root);
1537 if (cgrp == oldcgrp)
1540 for_each_subsys(root, ss) {
1541 if (ss->can_attach) {
1542 retval = ss->can_attach(ss, cgrp, tsk);
1553 * Locate or allocate a new css_set for this task,
1554 * based on its final set of cgroups
1556 newcg = find_css_set(cg, cgrp);
1562 if (tsk->flags & PF_EXITING) {
1567 rcu_assign_pointer(tsk->cgroups, newcg);
1570 /* Update the css_set linked lists if we're using them */
1571 write_lock(&css_set_lock);
1572 if (!list_empty(&tsk->cg_list)) {
1573 list_del(&tsk->cg_list);
1574 list_add(&tsk->cg_list, &newcg->tasks);
1576 write_unlock(&css_set_lock);
1578 for_each_subsys(root, ss) {
1580 ss->attach(ss, cgrp, oldcgrp, tsk);
1582 set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1587 * wake up rmdir() waiter. the rmdir should fail since the cgroup
1588 * is no longer empty.
1590 cgroup_wakeup_rmdir_waiter(cgrp);
1595 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
1596 * held. May take task_lock of task
1598 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid)
1600 struct task_struct *tsk;
1601 const struct cred *cred = current_cred(), *tcred;
1606 tsk = find_task_by_vpid(pid);
1607 if (!tsk || tsk->flags & PF_EXITING) {
1612 tcred = __task_cred(tsk);
1614 cred->euid != tcred->uid &&
1615 cred->euid != tcred->suid) {
1619 get_task_struct(tsk);
1623 get_task_struct(tsk);
1626 ret = cgroup_attach_task(cgrp, tsk);
1627 put_task_struct(tsk);
1631 static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
1634 if (!cgroup_lock_live_group(cgrp))
1636 ret = attach_task_by_pid(cgrp, pid);
1641 /* The various types of files and directories in a cgroup file system */
1642 enum cgroup_filetype {
1646 FILE_NOTIFY_ON_RELEASE,
1651 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
1652 * @cgrp: the cgroup to be checked for liveness
1654 * On success, returns true; the lock should be later released with
1655 * cgroup_unlock(). On failure returns false with no lock held.
1657 bool cgroup_lock_live_group(struct cgroup *cgrp)
1659 mutex_lock(&cgroup_mutex);
1660 if (cgroup_is_removed(cgrp)) {
1661 mutex_unlock(&cgroup_mutex);
1667 static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
1670 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1671 if (!cgroup_lock_live_group(cgrp))
1673 strcpy(cgrp->root->release_agent_path, buffer);
1678 static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
1679 struct seq_file *seq)
1681 if (!cgroup_lock_live_group(cgrp))
1683 seq_puts(seq, cgrp->root->release_agent_path);
1684 seq_putc(seq, '\n');
1689 /* A buffer size big enough for numbers or short strings */
1690 #define CGROUP_LOCAL_BUFFER_SIZE 64
1692 static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
1694 const char __user *userbuf,
1695 size_t nbytes, loff_t *unused_ppos)
1697 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
1703 if (nbytes >= sizeof(buffer))
1705 if (copy_from_user(buffer, userbuf, nbytes))
1708 buffer[nbytes] = 0; /* nul-terminate */
1710 if (cft->write_u64) {
1711 u64 val = simple_strtoull(buffer, &end, 0);
1714 retval = cft->write_u64(cgrp, cft, val);
1716 s64 val = simple_strtoll(buffer, &end, 0);
1719 retval = cft->write_s64(cgrp, cft, val);
1726 static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
1728 const char __user *userbuf,
1729 size_t nbytes, loff_t *unused_ppos)
1731 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
1733 size_t max_bytes = cft->max_write_len;
1734 char *buffer = local_buffer;
1737 max_bytes = sizeof(local_buffer) - 1;
1738 if (nbytes >= max_bytes)
1740 /* Allocate a dynamic buffer if we need one */
1741 if (nbytes >= sizeof(local_buffer)) {
1742 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
1746 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
1751 buffer[nbytes] = 0; /* nul-terminate */
1753 retval = cft->write_string(cgrp, cft, buffer);
1757 if (buffer != local_buffer)
1762 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
1763 size_t nbytes, loff_t *ppos)
1765 struct cftype *cft = __d_cft(file->f_dentry);
1766 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1768 if (cgroup_is_removed(cgrp))
1771 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
1772 if (cft->write_u64 || cft->write_s64)
1773 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
1774 if (cft->write_string)
1775 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
1777 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
1778 return ret ? ret : nbytes;
1783 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
1785 char __user *buf, size_t nbytes,
1788 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
1789 u64 val = cft->read_u64(cgrp, cft);
1790 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
1792 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1795 static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
1797 char __user *buf, size_t nbytes,
1800 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
1801 s64 val = cft->read_s64(cgrp, cft);
1802 int len = sprintf(tmp, "%lld\n", (long long) val);
1804 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1807 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
1808 size_t nbytes, loff_t *ppos)
1810 struct cftype *cft = __d_cft(file->f_dentry);
1811 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1813 if (cgroup_is_removed(cgrp))
1817 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
1819 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
1821 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
1826 * seqfile ops/methods for returning structured data. Currently just
1827 * supports string->u64 maps, but can be extended in future.
1830 struct cgroup_seqfile_state {
1832 struct cgroup *cgroup;
1835 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
1837 struct seq_file *sf = cb->state;
1838 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
1841 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
1843 struct cgroup_seqfile_state *state = m->private;
1844 struct cftype *cft = state->cft;
1845 if (cft->read_map) {
1846 struct cgroup_map_cb cb = {
1847 .fill = cgroup_map_add,
1850 return cft->read_map(state->cgroup, cft, &cb);
1852 return cft->read_seq_string(state->cgroup, cft, m);
1855 static int cgroup_seqfile_release(struct inode *inode, struct file *file)
1857 struct seq_file *seq = file->private_data;
1858 kfree(seq->private);
1859 return single_release(inode, file);
1862 static struct file_operations cgroup_seqfile_operations = {
1864 .write = cgroup_file_write,
1865 .llseek = seq_lseek,
1866 .release = cgroup_seqfile_release,
1869 static int cgroup_file_open(struct inode *inode, struct file *file)
1874 err = generic_file_open(inode, file);
1877 cft = __d_cft(file->f_dentry);
1879 if (cft->read_map || cft->read_seq_string) {
1880 struct cgroup_seqfile_state *state =
1881 kzalloc(sizeof(*state), GFP_USER);
1885 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
1886 file->f_op = &cgroup_seqfile_operations;
1887 err = single_open(file, cgroup_seqfile_show, state);
1890 } else if (cft->open)
1891 err = cft->open(inode, file);
1898 static int cgroup_file_release(struct inode *inode, struct file *file)
1900 struct cftype *cft = __d_cft(file->f_dentry);
1902 return cft->release(inode, file);
1907 * cgroup_rename - Only allow simple rename of directories in place.
1909 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
1910 struct inode *new_dir, struct dentry *new_dentry)
1912 if (!S_ISDIR(old_dentry->d_inode->i_mode))
1914 if (new_dentry->d_inode)
1916 if (old_dir != new_dir)
1918 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
1921 static struct file_operations cgroup_file_operations = {
1922 .read = cgroup_file_read,
1923 .write = cgroup_file_write,
1924 .llseek = generic_file_llseek,
1925 .open = cgroup_file_open,
1926 .release = cgroup_file_release,
1929 static const struct inode_operations cgroup_dir_inode_operations = {
1930 .lookup = simple_lookup,
1931 .mkdir = cgroup_mkdir,
1932 .rmdir = cgroup_rmdir,
1933 .rename = cgroup_rename,
1936 static int cgroup_create_file(struct dentry *dentry, mode_t mode,
1937 struct super_block *sb)
1939 static const struct dentry_operations cgroup_dops = {
1940 .d_iput = cgroup_diput,
1943 struct inode *inode;
1947 if (dentry->d_inode)
1950 inode = cgroup_new_inode(mode, sb);
1954 if (S_ISDIR(mode)) {
1955 inode->i_op = &cgroup_dir_inode_operations;
1956 inode->i_fop = &simple_dir_operations;
1958 /* start off with i_nlink == 2 (for "." entry) */
1961 /* start with the directory inode held, so that we can
1962 * populate it without racing with another mkdir */
1963 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1964 } else if (S_ISREG(mode)) {
1966 inode->i_fop = &cgroup_file_operations;
1968 dentry->d_op = &cgroup_dops;
1969 d_instantiate(dentry, inode);
1970 dget(dentry); /* Extra count - pin the dentry in core */
1975 * cgroup_create_dir - create a directory for an object.
1976 * @cgrp: the cgroup we create the directory for. It must have a valid
1977 * ->parent field. And we are going to fill its ->dentry field.
1978 * @dentry: dentry of the new cgroup
1979 * @mode: mode to set on new directory.
1981 static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
1984 struct dentry *parent;
1987 parent = cgrp->parent->dentry;
1988 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
1990 dentry->d_fsdata = cgrp;
1991 inc_nlink(parent->d_inode);
1992 rcu_assign_pointer(cgrp->dentry, dentry);
2001 * cgroup_file_mode - deduce file mode of a control file
2002 * @cft: the control file in question
2004 * returns cft->mode if ->mode is not 0
2005 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2006 * returns S_IRUGO if it has only a read handler
2007 * returns S_IWUSR if it has only a write hander
2009 static mode_t cgroup_file_mode(const struct cftype *cft)
2016 if (cft->read || cft->read_u64 || cft->read_s64 ||
2017 cft->read_map || cft->read_seq_string)
2020 if (cft->write || cft->write_u64 || cft->write_s64 ||
2021 cft->write_string || cft->trigger)
2027 int cgroup_add_file(struct cgroup *cgrp,
2028 struct cgroup_subsys *subsys,
2029 const struct cftype *cft)
2031 struct dentry *dir = cgrp->dentry;
2032 struct dentry *dentry;
2036 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2037 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
2038 strcpy(name, subsys->name);
2041 strcat(name, cft->name);
2042 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2043 dentry = lookup_one_len(name, dir, strlen(name));
2044 if (!IS_ERR(dentry)) {
2045 mode = cgroup_file_mode(cft);
2046 error = cgroup_create_file(dentry, mode | S_IFREG,
2049 dentry->d_fsdata = (void *)cft;
2052 error = PTR_ERR(dentry);
2056 int cgroup_add_files(struct cgroup *cgrp,
2057 struct cgroup_subsys *subsys,
2058 const struct cftype cft[],
2062 for (i = 0; i < count; i++) {
2063 err = cgroup_add_file(cgrp, subsys, &cft[i]);
2071 * cgroup_task_count - count the number of tasks in a cgroup.
2072 * @cgrp: the cgroup in question
2074 * Return the number of tasks in the cgroup.
2076 int cgroup_task_count(const struct cgroup *cgrp)
2079 struct cg_cgroup_link *link;
2081 read_lock(&css_set_lock);
2082 list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
2083 count += atomic_read(&link->cg->refcount);
2085 read_unlock(&css_set_lock);
2090 * Advance a list_head iterator. The iterator should be positioned at
2091 * the start of a css_set
2093 static void cgroup_advance_iter(struct cgroup *cgrp,
2094 struct cgroup_iter *it)
2096 struct list_head *l = it->cg_link;
2097 struct cg_cgroup_link *link;
2100 /* Advance to the next non-empty css_set */
2103 if (l == &cgrp->css_sets) {
2107 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
2109 } while (list_empty(&cg->tasks));
2111 it->task = cg->tasks.next;
2115 * To reduce the fork() overhead for systems that are not actually
2116 * using their cgroups capability, we don't maintain the lists running
2117 * through each css_set to its tasks until we see the list actually
2118 * used - in other words after the first call to cgroup_iter_start().
2120 * The tasklist_lock is not held here, as do_each_thread() and
2121 * while_each_thread() are protected by RCU.
2123 static void cgroup_enable_task_cg_lists(void)
2125 struct task_struct *p, *g;
2126 write_lock(&css_set_lock);
2127 use_task_css_set_links = 1;
2128 do_each_thread(g, p) {
2131 * We should check if the process is exiting, otherwise
2132 * it will race with cgroup_exit() in that the list
2133 * entry won't be deleted though the process has exited.
2135 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2136 list_add(&p->cg_list, &p->cgroups->tasks);
2138 } while_each_thread(g, p);
2139 write_unlock(&css_set_lock);
2142 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
2145 * The first time anyone tries to iterate across a cgroup,
2146 * we need to enable the list linking each css_set to its
2147 * tasks, and fix up all existing tasks.
2149 if (!use_task_css_set_links)
2150 cgroup_enable_task_cg_lists();
2152 read_lock(&css_set_lock);
2153 it->cg_link = &cgrp->css_sets;
2154 cgroup_advance_iter(cgrp, it);
2157 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
2158 struct cgroup_iter *it)
2160 struct task_struct *res;
2161 struct list_head *l = it->task;
2162 struct cg_cgroup_link *link;
2164 /* If the iterator cg is NULL, we have no tasks */
2167 res = list_entry(l, struct task_struct, cg_list);
2168 /* Advance iterator to find next entry */
2170 link = list_entry(it->cg_link, struct cg_cgroup_link, cgrp_link_list);
2171 if (l == &link->cg->tasks) {
2172 /* We reached the end of this task list - move on to
2173 * the next cg_cgroup_link */
2174 cgroup_advance_iter(cgrp, it);
2181 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
2183 read_unlock(&css_set_lock);
2186 static inline int started_after_time(struct task_struct *t1,
2187 struct timespec *time,
2188 struct task_struct *t2)
2190 int start_diff = timespec_compare(&t1->start_time, time);
2191 if (start_diff > 0) {
2193 } else if (start_diff < 0) {
2197 * Arbitrarily, if two processes started at the same
2198 * time, we'll say that the lower pointer value
2199 * started first. Note that t2 may have exited by now
2200 * so this may not be a valid pointer any longer, but
2201 * that's fine - it still serves to distinguish
2202 * between two tasks started (effectively) simultaneously.
2209 * This function is a callback from heap_insert() and is used to order
2211 * In this case we order the heap in descending task start time.
2213 static inline int started_after(void *p1, void *p2)
2215 struct task_struct *t1 = p1;
2216 struct task_struct *t2 = p2;
2217 return started_after_time(t1, &t2->start_time, t2);
2221 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
2222 * @scan: struct cgroup_scanner containing arguments for the scan
2224 * Arguments include pointers to callback functions test_task() and
2226 * Iterate through all the tasks in a cgroup, calling test_task() for each,
2227 * and if it returns true, call process_task() for it also.
2228 * The test_task pointer may be NULL, meaning always true (select all tasks).
2229 * Effectively duplicates cgroup_iter_{start,next,end}()
2230 * but does not lock css_set_lock for the call to process_task().
2231 * The struct cgroup_scanner may be embedded in any structure of the caller's
2233 * It is guaranteed that process_task() will act on every task that
2234 * is a member of the cgroup for the duration of this call. This
2235 * function may or may not call process_task() for tasks that exit
2236 * or move to a different cgroup during the call, or are forked or
2237 * move into the cgroup during the call.
2239 * Note that test_task() may be called with locks held, and may in some
2240 * situations be called multiple times for the same task, so it should
2242 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
2243 * pre-allocated and will be used for heap operations (and its "gt" member will
2244 * be overwritten), else a temporary heap will be used (allocation of which
2245 * may cause this function to fail).
2247 int cgroup_scan_tasks(struct cgroup_scanner *scan)
2250 struct cgroup_iter it;
2251 struct task_struct *p, *dropped;
2252 /* Never dereference latest_task, since it's not refcounted */
2253 struct task_struct *latest_task = NULL;
2254 struct ptr_heap tmp_heap;
2255 struct ptr_heap *heap;
2256 struct timespec latest_time = { 0, 0 };
2259 /* The caller supplied our heap and pre-allocated its memory */
2261 heap->gt = &started_after;
2263 /* We need to allocate our own heap memory */
2265 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2267 /* cannot allocate the heap */
2273 * Scan tasks in the cgroup, using the scanner's "test_task" callback
2274 * to determine which are of interest, and using the scanner's
2275 * "process_task" callback to process any of them that need an update.
2276 * Since we don't want to hold any locks during the task updates,
2277 * gather tasks to be processed in a heap structure.
2278 * The heap is sorted by descending task start time.
2279 * If the statically-sized heap fills up, we overflow tasks that
2280 * started later, and in future iterations only consider tasks that
2281 * started after the latest task in the previous pass. This
2282 * guarantees forward progress and that we don't miss any tasks.
2285 cgroup_iter_start(scan->cg, &it);
2286 while ((p = cgroup_iter_next(scan->cg, &it))) {
2288 * Only affect tasks that qualify per the caller's callback,
2289 * if he provided one
2291 if (scan->test_task && !scan->test_task(p, scan))
2294 * Only process tasks that started after the last task
2297 if (!started_after_time(p, &latest_time, latest_task))
2299 dropped = heap_insert(heap, p);
2300 if (dropped == NULL) {
2302 * The new task was inserted; the heap wasn't
2306 } else if (dropped != p) {
2308 * The new task was inserted, and pushed out a
2312 put_task_struct(dropped);
2315 * Else the new task was newer than anything already in
2316 * the heap and wasn't inserted
2319 cgroup_iter_end(scan->cg, &it);
2322 for (i = 0; i < heap->size; i++) {
2323 struct task_struct *q = heap->ptrs[i];
2325 latest_time = q->start_time;
2328 /* Process the task per the caller's callback */
2329 scan->process_task(q, scan);
2333 * If we had to process any tasks at all, scan again
2334 * in case some of them were in the middle of forking
2335 * children that didn't get processed.
2336 * Not the most efficient way to do it, but it avoids
2337 * having to take callback_mutex in the fork path
2341 if (heap == &tmp_heap)
2342 heap_free(&tmp_heap);
2347 * Stuff for reading the 'tasks' file.
2349 * Reading this file can return large amounts of data if a cgroup has
2350 * *lots* of attached tasks. So it may need several calls to read(),
2351 * but we cannot guarantee that the information we produce is correct
2352 * unless we produce it entirely atomically.
2357 * Load into 'pidarray' up to 'npids' of the tasks using cgroup
2358 * 'cgrp'. Return actual number of pids loaded. No need to
2359 * task_lock(p) when reading out p->cgroup, since we're in an RCU
2360 * read section, so the css_set can't go away, and is
2361 * immutable after creation.
2363 static int pid_array_load(pid_t *pidarray, int npids, struct cgroup *cgrp)
2366 struct cgroup_iter it;
2367 struct task_struct *tsk;
2368 cgroup_iter_start(cgrp, &it);
2369 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2370 if (unlikely(n == npids))
2372 pid = task_pid_vnr(tsk);
2374 pidarray[n++] = pid;
2376 cgroup_iter_end(cgrp, &it);
2381 * cgroupstats_build - build and fill cgroupstats
2382 * @stats: cgroupstats to fill information into
2383 * @dentry: A dentry entry belonging to the cgroup for which stats have
2386 * Build and fill cgroupstats so that taskstats can export it to user
2389 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2392 struct cgroup *cgrp;
2393 struct cgroup_iter it;
2394 struct task_struct *tsk;
2397 * Validate dentry by checking the superblock operations,
2398 * and make sure it's a directory.
2400 if (dentry->d_sb->s_op != &cgroup_ops ||
2401 !S_ISDIR(dentry->d_inode->i_mode))
2405 cgrp = dentry->d_fsdata;
2407 cgroup_iter_start(cgrp, &it);
2408 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2409 switch (tsk->state) {
2411 stats->nr_running++;
2413 case TASK_INTERRUPTIBLE:
2414 stats->nr_sleeping++;
2416 case TASK_UNINTERRUPTIBLE:
2417 stats->nr_uninterruptible++;
2420 stats->nr_stopped++;
2423 if (delayacct_is_task_waiting_on_io(tsk))
2424 stats->nr_io_wait++;
2428 cgroup_iter_end(cgrp, &it);
2435 * Cache pids for all threads in the same pid namespace that are
2436 * opening the same "tasks" file.
2438 struct cgroup_pids {
2439 /* The node in cgrp->pids_list */
2440 struct list_head list;
2441 /* The cgroup those pids belong to */
2442 struct cgroup *cgrp;
2443 /* The namepsace those pids belong to */
2444 struct pid_namespace *ns;
2445 /* Array of process ids in the cgroup */
2447 /* How many files are using the this tasks_pids array */
2449 /* Length of the current tasks_pids array */
2453 static int cmppid(const void *a, const void *b)
2455 return *(pid_t *)a - *(pid_t *)b;
2459 * seq_file methods for the "tasks" file. The seq_file position is the
2460 * next pid to display; the seq_file iterator is a pointer to the pid
2461 * in the cgroup->tasks_pids array.
2464 static void *cgroup_tasks_start(struct seq_file *s, loff_t *pos)
2467 * Initially we receive a position value that corresponds to
2468 * one more than the last pid shown (or 0 on the first call or
2469 * after a seek to the start). Use a binary-search to find the
2470 * next pid to display, if any
2472 struct cgroup_pids *cp = s->private;
2473 struct cgroup *cgrp = cp->cgrp;
2474 int index = 0, pid = *pos;
2477 down_read(&cgrp->pids_mutex);
2479 int end = cp->length;
2481 while (index < end) {
2482 int mid = (index + end) / 2;
2483 if (cp->tasks_pids[mid] == pid) {
2486 } else if (cp->tasks_pids[mid] <= pid)
2492 /* If we're off the end of the array, we're done */
2493 if (index >= cp->length)
2495 /* Update the abstract position to be the actual pid that we found */
2496 iter = cp->tasks_pids + index;
2501 static void cgroup_tasks_stop(struct seq_file *s, void *v)
2503 struct cgroup_pids *cp = s->private;
2504 struct cgroup *cgrp = cp->cgrp;
2505 up_read(&cgrp->pids_mutex);
2508 static void *cgroup_tasks_next(struct seq_file *s, void *v, loff_t *pos)
2510 struct cgroup_pids *cp = s->private;
2512 int *end = cp->tasks_pids + cp->length;
2515 * Advance to the next pid in the array. If this goes off the
2527 static int cgroup_tasks_show(struct seq_file *s, void *v)
2529 return seq_printf(s, "%d\n", *(int *)v);
2532 static const struct seq_operations cgroup_tasks_seq_operations = {
2533 .start = cgroup_tasks_start,
2534 .stop = cgroup_tasks_stop,
2535 .next = cgroup_tasks_next,
2536 .show = cgroup_tasks_show,
2539 static void release_cgroup_pid_array(struct cgroup_pids *cp)
2541 struct cgroup *cgrp = cp->cgrp;
2543 down_write(&cgrp->pids_mutex);
2544 BUG_ON(!cp->use_count);
2545 if (!--cp->use_count) {
2546 list_del(&cp->list);
2548 kfree(cp->tasks_pids);
2551 up_write(&cgrp->pids_mutex);
2554 static int cgroup_tasks_release(struct inode *inode, struct file *file)
2556 struct seq_file *seq;
2557 struct cgroup_pids *cp;
2559 if (!(file->f_mode & FMODE_READ))
2562 seq = file->private_data;
2565 release_cgroup_pid_array(cp);
2566 return seq_release(inode, file);
2569 static struct file_operations cgroup_tasks_operations = {
2571 .llseek = seq_lseek,
2572 .write = cgroup_file_write,
2573 .release = cgroup_tasks_release,
2577 * Handle an open on 'tasks' file. Prepare an array containing the
2578 * process id's of tasks currently attached to the cgroup being opened.
2581 static int cgroup_tasks_open(struct inode *unused, struct file *file)
2583 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2584 struct pid_namespace *ns = current->nsproxy->pid_ns;
2585 struct cgroup_pids *cp;
2590 /* Nothing to do for write-only files */
2591 if (!(file->f_mode & FMODE_READ))
2595 * If cgroup gets more users after we read count, we won't have
2596 * enough space - tough. This race is indistinguishable to the
2597 * caller from the case that the additional cgroup users didn't
2598 * show up until sometime later on.
2600 npids = cgroup_task_count(cgrp);
2601 pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL);
2604 npids = pid_array_load(pidarray, npids, cgrp);
2605 sort(pidarray, npids, sizeof(pid_t), cmppid, NULL);
2608 * Store the array in the cgroup, freeing the old
2609 * array if necessary
2611 down_write(&cgrp->pids_mutex);
2613 list_for_each_entry(cp, &cgrp->pids_list, list) {
2618 cp = kzalloc(sizeof(*cp), GFP_KERNEL);
2620 up_write(&cgrp->pids_mutex);
2627 list_add(&cp->list, &cgrp->pids_list);
2629 kfree(cp->tasks_pids);
2630 cp->tasks_pids = pidarray;
2633 up_write(&cgrp->pids_mutex);
2635 file->f_op = &cgroup_tasks_operations;
2637 retval = seq_open(file, &cgroup_tasks_seq_operations);
2639 release_cgroup_pid_array(cp);
2642 ((struct seq_file *)file->private_data)->private = cp;
2646 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
2649 return notify_on_release(cgrp);
2652 static int cgroup_write_notify_on_release(struct cgroup *cgrp,
2656 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
2658 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2660 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2665 * for the common functions, 'private' gives the type of file
2667 static struct cftype files[] = {
2670 .open = cgroup_tasks_open,
2671 .write_u64 = cgroup_tasks_write,
2672 .release = cgroup_tasks_release,
2673 .private = FILE_TASKLIST,
2674 .mode = S_IRUGO | S_IWUSR,
2678 .name = "notify_on_release",
2679 .read_u64 = cgroup_read_notify_on_release,
2680 .write_u64 = cgroup_write_notify_on_release,
2681 .private = FILE_NOTIFY_ON_RELEASE,
2685 static struct cftype cft_release_agent = {
2686 .name = "release_agent",
2687 .read_seq_string = cgroup_release_agent_show,
2688 .write_string = cgroup_release_agent_write,
2689 .max_write_len = PATH_MAX,
2690 .private = FILE_RELEASE_AGENT,
2693 static int cgroup_populate_dir(struct cgroup *cgrp)
2696 struct cgroup_subsys *ss;
2698 /* First clear out any existing files */
2699 cgroup_clear_directory(cgrp->dentry);
2701 err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
2705 if (cgrp == cgrp->top_cgroup) {
2706 if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
2710 for_each_subsys(cgrp->root, ss) {
2711 if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
2714 /* This cgroup is ready now */
2715 for_each_subsys(cgrp->root, ss) {
2716 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
2718 * Update id->css pointer and make this css visible from
2719 * CSS ID functions. This pointer will be dereferened
2720 * from RCU-read-side without locks.
2723 rcu_assign_pointer(css->id->css, css);
2729 static void init_cgroup_css(struct cgroup_subsys_state *css,
2730 struct cgroup_subsys *ss,
2731 struct cgroup *cgrp)
2734 atomic_set(&css->refcnt, 1);
2737 if (cgrp == dummytop)
2738 set_bit(CSS_ROOT, &css->flags);
2739 BUG_ON(cgrp->subsys[ss->subsys_id]);
2740 cgrp->subsys[ss->subsys_id] = css;
2743 static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
2745 /* We need to take each hierarchy_mutex in a consistent order */
2748 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2749 struct cgroup_subsys *ss = subsys[i];
2750 if (ss->root == root)
2751 mutex_lock(&ss->hierarchy_mutex);
2755 static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
2759 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2760 struct cgroup_subsys *ss = subsys[i];
2761 if (ss->root == root)
2762 mutex_unlock(&ss->hierarchy_mutex);
2767 * cgroup_create - create a cgroup
2768 * @parent: cgroup that will be parent of the new cgroup
2769 * @dentry: dentry of the new cgroup
2770 * @mode: mode to set on new inode
2772 * Must be called with the mutex on the parent inode held
2774 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
2777 struct cgroup *cgrp;
2778 struct cgroupfs_root *root = parent->root;
2780 struct cgroup_subsys *ss;
2781 struct super_block *sb = root->sb;
2783 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
2787 /* Grab a reference on the superblock so the hierarchy doesn't
2788 * get deleted on unmount if there are child cgroups. This
2789 * can be done outside cgroup_mutex, since the sb can't
2790 * disappear while someone has an open control file on the
2792 atomic_inc(&sb->s_active);
2794 mutex_lock(&cgroup_mutex);
2796 init_cgroup_housekeeping(cgrp);
2798 cgrp->parent = parent;
2799 cgrp->root = parent->root;
2800 cgrp->top_cgroup = parent->top_cgroup;
2802 if (notify_on_release(parent))
2803 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2805 for_each_subsys(root, ss) {
2806 struct cgroup_subsys_state *css = ss->create(ss, cgrp);
2811 init_cgroup_css(css, ss, cgrp);
2813 if (alloc_css_id(ss, parent, cgrp))
2815 /* At error, ->destroy() callback has to free assigned ID. */
2818 cgroup_lock_hierarchy(root);
2819 list_add(&cgrp->sibling, &cgrp->parent->children);
2820 cgroup_unlock_hierarchy(root);
2821 root->number_of_cgroups++;
2823 err = cgroup_create_dir(cgrp, dentry, mode);
2827 /* The cgroup directory was pre-locked for us */
2828 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
2830 err = cgroup_populate_dir(cgrp);
2831 /* If err < 0, we have a half-filled directory - oh well ;) */
2833 mutex_unlock(&cgroup_mutex);
2834 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
2840 cgroup_lock_hierarchy(root);
2841 list_del(&cgrp->sibling);
2842 cgroup_unlock_hierarchy(root);
2843 root->number_of_cgroups--;
2847 for_each_subsys(root, ss) {
2848 if (cgrp->subsys[ss->subsys_id])
2849 ss->destroy(ss, cgrp);
2852 mutex_unlock(&cgroup_mutex);
2854 /* Release the reference count that we took on the superblock */
2855 deactivate_super(sb);
2861 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
2863 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
2865 /* the vfs holds inode->i_mutex already */
2866 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
2869 static int cgroup_has_css_refs(struct cgroup *cgrp)
2871 /* Check the reference count on each subsystem. Since we
2872 * already established that there are no tasks in the
2873 * cgroup, if the css refcount is also 1, then there should
2874 * be no outstanding references, so the subsystem is safe to
2875 * destroy. We scan across all subsystems rather than using
2876 * the per-hierarchy linked list of mounted subsystems since
2877 * we can be called via check_for_release() with no
2878 * synchronization other than RCU, and the subsystem linked
2879 * list isn't RCU-safe */
2881 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2882 struct cgroup_subsys *ss = subsys[i];
2883 struct cgroup_subsys_state *css;
2884 /* Skip subsystems not in this hierarchy */
2885 if (ss->root != cgrp->root)
2887 css = cgrp->subsys[ss->subsys_id];
2888 /* When called from check_for_release() it's possible
2889 * that by this point the cgroup has been removed
2890 * and the css deleted. But a false-positive doesn't
2891 * matter, since it can only happen if the cgroup
2892 * has been deleted and hence no longer needs the
2893 * release agent to be called anyway. */
2894 if (css && (atomic_read(&css->refcnt) > 1))
2901 * Atomically mark all (or else none) of the cgroup's CSS objects as
2902 * CSS_REMOVED. Return true on success, or false if the cgroup has
2903 * busy subsystems. Call with cgroup_mutex held
2906 static int cgroup_clear_css_refs(struct cgroup *cgrp)
2908 struct cgroup_subsys *ss;
2909 unsigned long flags;
2910 bool failed = false;
2911 local_irq_save(flags);
2912 for_each_subsys(cgrp->root, ss) {
2913 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
2916 /* We can only remove a CSS with a refcnt==1 */
2917 refcnt = atomic_read(&css->refcnt);
2924 * Drop the refcnt to 0 while we check other
2925 * subsystems. This will cause any racing
2926 * css_tryget() to spin until we set the
2927 * CSS_REMOVED bits or abort
2929 if (atomic_cmpxchg(&css->refcnt, refcnt, 0) == refcnt)
2935 for_each_subsys(cgrp->root, ss) {
2936 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
2939 * Restore old refcnt if we previously managed
2940 * to clear it from 1 to 0
2942 if (!atomic_read(&css->refcnt))
2943 atomic_set(&css->refcnt, 1);
2945 /* Commit the fact that the CSS is removed */
2946 set_bit(CSS_REMOVED, &css->flags);
2949 local_irq_restore(flags);
2953 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
2955 struct cgroup *cgrp = dentry->d_fsdata;
2957 struct cgroup *parent;
2961 /* the vfs holds both inode->i_mutex already */
2963 mutex_lock(&cgroup_mutex);
2964 if (atomic_read(&cgrp->count) != 0) {
2965 mutex_unlock(&cgroup_mutex);
2968 if (!list_empty(&cgrp->children)) {
2969 mutex_unlock(&cgroup_mutex);
2972 mutex_unlock(&cgroup_mutex);
2975 * In general, subsystem has no css->refcnt after pre_destroy(). But
2976 * in racy cases, subsystem may have to get css->refcnt after
2977 * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
2978 * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
2979 * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
2980 * and subsystem's reference count handling. Please see css_get/put
2981 * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
2983 set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
2986 * Call pre_destroy handlers of subsys. Notify subsystems
2987 * that rmdir() request comes.
2989 ret = cgroup_call_pre_destroy(cgrp);
2991 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
2995 mutex_lock(&cgroup_mutex);
2996 parent = cgrp->parent;
2997 if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
2998 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
2999 mutex_unlock(&cgroup_mutex);
3002 prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
3003 if (!cgroup_clear_css_refs(cgrp)) {
3004 mutex_unlock(&cgroup_mutex);
3006 * Because someone may call cgroup_wakeup_rmdir_waiter() before
3007 * prepare_to_wait(), we need to check this flag.
3009 if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
3011 finish_wait(&cgroup_rmdir_waitq, &wait);
3012 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3013 if (signal_pending(current))
3017 /* NO css_tryget() can success after here. */
3018 finish_wait(&cgroup_rmdir_waitq, &wait);
3019 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3021 spin_lock(&release_list_lock);
3022 set_bit(CGRP_REMOVED, &cgrp->flags);
3023 if (!list_empty(&cgrp->release_list))
3024 list_del(&cgrp->release_list);
3025 spin_unlock(&release_list_lock);
3027 cgroup_lock_hierarchy(cgrp->root);
3028 /* delete this cgroup from parent->children */
3029 list_del(&cgrp->sibling);
3030 cgroup_unlock_hierarchy(cgrp->root);
3032 spin_lock(&cgrp->dentry->d_lock);
3033 d = dget(cgrp->dentry);
3034 spin_unlock(&d->d_lock);
3036 cgroup_d_remove_dir(d);
3039 set_bit(CGRP_RELEASABLE, &parent->flags);
3040 check_for_release(parent);
3042 mutex_unlock(&cgroup_mutex);
3046 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
3048 struct cgroup_subsys_state *css;
3050 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
3052 /* Create the top cgroup state for this subsystem */
3053 list_add(&ss->sibling, &rootnode.subsys_list);
3054 ss->root = &rootnode;
3055 css = ss->create(ss, dummytop);
3056 /* We don't handle early failures gracefully */
3057 BUG_ON(IS_ERR(css));
3058 init_cgroup_css(css, ss, dummytop);
3060 /* Update the init_css_set to contain a subsys
3061 * pointer to this state - since the subsystem is
3062 * newly registered, all tasks and hence the
3063 * init_css_set is in the subsystem's top cgroup. */
3064 init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
3066 need_forkexit_callback |= ss->fork || ss->exit;
3068 /* At system boot, before all subsystems have been
3069 * registered, no tasks have been forked, so we don't
3070 * need to invoke fork callbacks here. */
3071 BUG_ON(!list_empty(&init_task.tasks));
3073 mutex_init(&ss->hierarchy_mutex);
3074 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3079 * cgroup_init_early - cgroup initialization at system boot
3081 * Initialize cgroups at system boot, and initialize any
3082 * subsystems that request early init.
3084 int __init cgroup_init_early(void)
3087 atomic_set(&init_css_set.refcount, 1);
3088 INIT_LIST_HEAD(&init_css_set.cg_links);
3089 INIT_LIST_HEAD(&init_css_set.tasks);
3090 INIT_HLIST_NODE(&init_css_set.hlist);
3092 init_cgroup_root(&rootnode);
3094 init_task.cgroups = &init_css_set;
3096 init_css_set_link.cg = &init_css_set;
3097 init_css_set_link.cgrp = dummytop;
3098 list_add(&init_css_set_link.cgrp_link_list,
3099 &rootnode.top_cgroup.css_sets);
3100 list_add(&init_css_set_link.cg_link_list,
3101 &init_css_set.cg_links);
3103 for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
3104 INIT_HLIST_HEAD(&css_set_table[i]);
3106 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3107 struct cgroup_subsys *ss = subsys[i];
3110 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
3111 BUG_ON(!ss->create);
3112 BUG_ON(!ss->destroy);
3113 if (ss->subsys_id != i) {
3114 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
3115 ss->name, ss->subsys_id);
3120 cgroup_init_subsys(ss);
3126 * cgroup_init - cgroup initialization
3128 * Register cgroup filesystem and /proc file, and initialize
3129 * any subsystems that didn't request early init.
3131 int __init cgroup_init(void)
3135 struct hlist_head *hhead;
3137 err = bdi_init(&cgroup_backing_dev_info);
3141 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3142 struct cgroup_subsys *ss = subsys[i];
3143 if (!ss->early_init)
3144 cgroup_init_subsys(ss);
3146 cgroup_subsys_init_idr(ss);
3149 /* Add init_css_set to the hash table */
3150 hhead = css_set_hash(init_css_set.subsys);
3151 hlist_add_head(&init_css_set.hlist, hhead);
3152 BUG_ON(!init_root_id(&rootnode));
3153 err = register_filesystem(&cgroup_fs_type);
3157 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
3161 bdi_destroy(&cgroup_backing_dev_info);
3167 * proc_cgroup_show()
3168 * - Print task's cgroup paths into seq_file, one line for each hierarchy
3169 * - Used for /proc/<pid>/cgroup.
3170 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
3171 * doesn't really matter if tsk->cgroup changes after we read it,
3172 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
3173 * anyway. No need to check that tsk->cgroup != NULL, thanks to
3174 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
3175 * cgroup to top_cgroup.
3178 /* TODO: Use a proper seq_file iterator */
3179 static int proc_cgroup_show(struct seq_file *m, void *v)
3182 struct task_struct *tsk;
3185 struct cgroupfs_root *root;
3188 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
3194 tsk = get_pid_task(pid, PIDTYPE_PID);
3200 mutex_lock(&cgroup_mutex);
3202 for_each_active_root(root) {
3203 struct cgroup_subsys *ss;
3204 struct cgroup *cgrp;
3207 seq_printf(m, "%d:", root->hierarchy_id);
3208 for_each_subsys(root, ss)
3209 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
3210 if (strlen(root->name))
3211 seq_printf(m, "%sname=%s", count ? "," : "",
3214 cgrp = task_cgroup_from_root(tsk, root);
3215 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
3223 mutex_unlock(&cgroup_mutex);
3224 put_task_struct(tsk);
3231 static int cgroup_open(struct inode *inode, struct file *file)
3233 struct pid *pid = PROC_I(inode)->pid;
3234 return single_open(file, proc_cgroup_show, pid);
3237 struct file_operations proc_cgroup_operations = {
3238 .open = cgroup_open,
3240 .llseek = seq_lseek,
3241 .release = single_release,
3244 /* Display information about each subsystem and each hierarchy */
3245 static int proc_cgroupstats_show(struct seq_file *m, void *v)
3249 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
3250 mutex_lock(&cgroup_mutex);
3251 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3252 struct cgroup_subsys *ss = subsys[i];
3253 seq_printf(m, "%s\t%d\t%d\t%d\n",
3254 ss->name, ss->root->hierarchy_id,
3255 ss->root->number_of_cgroups, !ss->disabled);
3257 mutex_unlock(&cgroup_mutex);
3261 static int cgroupstats_open(struct inode *inode, struct file *file)
3263 return single_open(file, proc_cgroupstats_show, NULL);
3266 static struct file_operations proc_cgroupstats_operations = {
3267 .open = cgroupstats_open,
3269 .llseek = seq_lseek,
3270 .release = single_release,
3274 * cgroup_fork - attach newly forked task to its parents cgroup.
3275 * @child: pointer to task_struct of forking parent process.
3277 * Description: A task inherits its parent's cgroup at fork().
3279 * A pointer to the shared css_set was automatically copied in
3280 * fork.c by dup_task_struct(). However, we ignore that copy, since
3281 * it was not made under the protection of RCU or cgroup_mutex, so
3282 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
3283 * have already changed current->cgroups, allowing the previously
3284 * referenced cgroup group to be removed and freed.
3286 * At the point that cgroup_fork() is called, 'current' is the parent
3287 * task, and the passed argument 'child' points to the child task.
3289 void cgroup_fork(struct task_struct *child)
3292 child->cgroups = current->cgroups;
3293 get_css_set(child->cgroups);
3294 task_unlock(current);
3295 INIT_LIST_HEAD(&child->cg_list);
3299 * cgroup_fork_callbacks - run fork callbacks
3300 * @child: the new task
3302 * Called on a new task very soon before adding it to the
3303 * tasklist. No need to take any locks since no-one can
3304 * be operating on this task.
3306 void cgroup_fork_callbacks(struct task_struct *child)
3308 if (need_forkexit_callback) {
3310 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3311 struct cgroup_subsys *ss = subsys[i];
3313 ss->fork(ss, child);
3319 * cgroup_post_fork - called on a new task after adding it to the task list
3320 * @child: the task in question
3322 * Adds the task to the list running through its css_set if necessary.
3323 * Has to be after the task is visible on the task list in case we race
3324 * with the first call to cgroup_iter_start() - to guarantee that the
3325 * new task ends up on its list.
3327 void cgroup_post_fork(struct task_struct *child)
3329 if (use_task_css_set_links) {
3330 write_lock(&css_set_lock);
3332 if (list_empty(&child->cg_list))
3333 list_add(&child->cg_list, &child->cgroups->tasks);
3335 write_unlock(&css_set_lock);
3339 * cgroup_exit - detach cgroup from exiting task
3340 * @tsk: pointer to task_struct of exiting process
3341 * @run_callback: run exit callbacks?
3343 * Description: Detach cgroup from @tsk and release it.
3345 * Note that cgroups marked notify_on_release force every task in
3346 * them to take the global cgroup_mutex mutex when exiting.
3347 * This could impact scaling on very large systems. Be reluctant to
3348 * use notify_on_release cgroups where very high task exit scaling
3349 * is required on large systems.
3351 * the_top_cgroup_hack:
3353 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
3355 * We call cgroup_exit() while the task is still competent to
3356 * handle notify_on_release(), then leave the task attached to the
3357 * root cgroup in each hierarchy for the remainder of its exit.
3359 * To do this properly, we would increment the reference count on
3360 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
3361 * code we would add a second cgroup function call, to drop that
3362 * reference. This would just create an unnecessary hot spot on
3363 * the top_cgroup reference count, to no avail.
3365 * Normally, holding a reference to a cgroup without bumping its
3366 * count is unsafe. The cgroup could go away, or someone could
3367 * attach us to a different cgroup, decrementing the count on
3368 * the first cgroup that we never incremented. But in this case,
3369 * top_cgroup isn't going away, and either task has PF_EXITING set,
3370 * which wards off any cgroup_attach_task() attempts, or task is a failed
3371 * fork, never visible to cgroup_attach_task.
3373 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
3378 if (run_callbacks && need_forkexit_callback) {
3379 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3380 struct cgroup_subsys *ss = subsys[i];
3387 * Unlink from the css_set task list if necessary.
3388 * Optimistically check cg_list before taking
3391 if (!list_empty(&tsk->cg_list)) {
3392 write_lock(&css_set_lock);
3393 if (!list_empty(&tsk->cg_list))
3394 list_del(&tsk->cg_list);
3395 write_unlock(&css_set_lock);
3398 /* Reassign the task to the init_css_set. */
3401 tsk->cgroups = &init_css_set;
3404 put_css_set_taskexit(cg);
3408 * cgroup_clone - clone the cgroup the given subsystem is attached to
3409 * @tsk: the task to be moved
3410 * @subsys: the given subsystem
3411 * @nodename: the name for the new cgroup
3413 * Duplicate the current cgroup in the hierarchy that the given
3414 * subsystem is attached to, and move this task into the new
3417 int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
3420 struct dentry *dentry;
3422 struct cgroup *parent, *child;
3423 struct inode *inode;
3425 struct cgroupfs_root *root;
3426 struct cgroup_subsys *ss;
3428 /* We shouldn't be called by an unregistered subsystem */
3429 BUG_ON(!subsys->active);
3431 /* First figure out what hierarchy and cgroup we're dealing
3432 * with, and pin them so we can drop cgroup_mutex */
3433 mutex_lock(&cgroup_mutex);
3435 root = subsys->root;
3436 if (root == &rootnode) {
3437 mutex_unlock(&cgroup_mutex);
3441 /* Pin the hierarchy */
3442 if (!atomic_inc_not_zero(&root->sb->s_active)) {
3443 /* We race with the final deactivate_super() */
3444 mutex_unlock(&cgroup_mutex);
3448 /* Keep the cgroup alive */
3450 parent = task_cgroup(tsk, subsys->subsys_id);
3455 mutex_unlock(&cgroup_mutex);
3457 /* Now do the VFS work to create a cgroup */
3458 inode = parent->dentry->d_inode;
3460 /* Hold the parent directory mutex across this operation to
3461 * stop anyone else deleting the new cgroup */
3462 mutex_lock(&inode->i_mutex);
3463 dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
3464 if (IS_ERR(dentry)) {
3466 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
3468 ret = PTR_ERR(dentry);
3472 /* Create the cgroup directory, which also creates the cgroup */
3473 ret = vfs_mkdir(inode, dentry, 0755);
3474 child = __d_cgrp(dentry);
3478 "Failed to create cgroup %s: %d\n", nodename,
3483 /* The cgroup now exists. Retake cgroup_mutex and check
3484 * that we're still in the same state that we thought we
3486 mutex_lock(&cgroup_mutex);
3487 if ((root != subsys->root) ||
3488 (parent != task_cgroup(tsk, subsys->subsys_id))) {
3489 /* Aargh, we raced ... */
3490 mutex_unlock(&inode->i_mutex);
3493 deactivate_super(root->sb);
3494 /* The cgroup is still accessible in the VFS, but
3495 * we're not going to try to rmdir() it at this
3498 "Race in cgroup_clone() - leaking cgroup %s\n",
3503 /* do any required auto-setup */
3504 for_each_subsys(root, ss) {
3506 ss->post_clone(ss, child);
3509 /* All seems fine. Finish by moving the task into the new cgroup */
3510 ret = cgroup_attach_task(child, tsk);
3511 mutex_unlock(&cgroup_mutex);
3514 mutex_unlock(&inode->i_mutex);
3516 mutex_lock(&cgroup_mutex);
3518 mutex_unlock(&cgroup_mutex);
3519 deactivate_super(root->sb);
3524 * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
3525 * @cgrp: the cgroup in question
3526 * @task: the task in question
3528 * See if @cgrp is a descendant of @task's cgroup in the appropriate
3531 * If we are sending in dummytop, then presumably we are creating
3532 * the top cgroup in the subsystem.
3534 * Called only by the ns (nsproxy) cgroup.
3536 int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
3539 struct cgroup *target;
3541 if (cgrp == dummytop)
3544 target = task_cgroup_from_root(task, cgrp->root);
3545 while (cgrp != target && cgrp!= cgrp->top_cgroup)
3546 cgrp = cgrp->parent;
3547 ret = (cgrp == target);
3551 static void check_for_release(struct cgroup *cgrp)
3553 /* All of these checks rely on RCU to keep the cgroup
3554 * structure alive */
3555 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
3556 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
3557 /* Control Group is currently removeable. If it's not
3558 * already queued for a userspace notification, queue
3560 int need_schedule_work = 0;
3561 spin_lock(&release_list_lock);
3562 if (!cgroup_is_removed(cgrp) &&
3563 list_empty(&cgrp->release_list)) {
3564 list_add(&cgrp->release_list, &release_list);
3565 need_schedule_work = 1;
3567 spin_unlock(&release_list_lock);
3568 if (need_schedule_work)
3569 schedule_work(&release_agent_work);
3573 void __css_put(struct cgroup_subsys_state *css)
3575 struct cgroup *cgrp = css->cgroup;
3577 if (atomic_dec_return(&css->refcnt) == 1) {
3578 if (notify_on_release(cgrp)) {
3579 set_bit(CGRP_RELEASABLE, &cgrp->flags);
3580 check_for_release(cgrp);
3582 cgroup_wakeup_rmdir_waiter(cgrp);
3588 * Notify userspace when a cgroup is released, by running the
3589 * configured release agent with the name of the cgroup (path
3590 * relative to the root of cgroup file system) as the argument.
3592 * Most likely, this user command will try to rmdir this cgroup.
3594 * This races with the possibility that some other task will be
3595 * attached to this cgroup before it is removed, or that some other
3596 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
3597 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
3598 * unused, and this cgroup will be reprieved from its death sentence,
3599 * to continue to serve a useful existence. Next time it's released,
3600 * we will get notified again, if it still has 'notify_on_release' set.
3602 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
3603 * means only wait until the task is successfully execve()'d. The
3604 * separate release agent task is forked by call_usermodehelper(),
3605 * then control in this thread returns here, without waiting for the
3606 * release agent task. We don't bother to wait because the caller of
3607 * this routine has no use for the exit status of the release agent
3608 * task, so no sense holding our caller up for that.
3610 static void cgroup_release_agent(struct work_struct *work)
3612 BUG_ON(work != &release_agent_work);
3613 mutex_lock(&cgroup_mutex);
3614 spin_lock(&release_list_lock);
3615 while (!list_empty(&release_list)) {
3616 char *argv[3], *envp[3];
3618 char *pathbuf = NULL, *agentbuf = NULL;
3619 struct cgroup *cgrp = list_entry(release_list.next,
3622 list_del_init(&cgrp->release_list);
3623 spin_unlock(&release_list_lock);
3624 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
3627 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
3629 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
3634 argv[i++] = agentbuf;
3635 argv[i++] = pathbuf;
3639 /* minimal command environment */
3640 envp[i++] = "HOME=/";
3641 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
3644 /* Drop the lock while we invoke the usermode helper,
3645 * since the exec could involve hitting disk and hence
3646 * be a slow process */
3647 mutex_unlock(&cgroup_mutex);
3648 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
3649 mutex_lock(&cgroup_mutex);
3653 spin_lock(&release_list_lock);
3655 spin_unlock(&release_list_lock);
3656 mutex_unlock(&cgroup_mutex);
3659 static int __init cgroup_disable(char *str)
3664 while ((token = strsep(&str, ",")) != NULL) {
3668 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3669 struct cgroup_subsys *ss = subsys[i];
3671 if (!strcmp(token, ss->name)) {
3673 printk(KERN_INFO "Disabling %s control group"
3674 " subsystem\n", ss->name);
3681 __setup("cgroup_disable=", cgroup_disable);
3684 * Functons for CSS ID.
3688 *To get ID other than 0, this should be called when !cgroup_is_removed().
3690 unsigned short css_id(struct cgroup_subsys_state *css)
3692 struct css_id *cssid = rcu_dereference(css->id);
3699 unsigned short css_depth(struct cgroup_subsys_state *css)
3701 struct css_id *cssid = rcu_dereference(css->id);
3704 return cssid->depth;
3708 bool css_is_ancestor(struct cgroup_subsys_state *child,
3709 const struct cgroup_subsys_state *root)
3711 struct css_id *child_id = rcu_dereference(child->id);
3712 struct css_id *root_id = rcu_dereference(root->id);
3714 if (!child_id || !root_id || (child_id->depth < root_id->depth))
3716 return child_id->stack[root_id->depth] == root_id->id;
3719 static void __free_css_id_cb(struct rcu_head *head)
3723 id = container_of(head, struct css_id, rcu_head);
3727 void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
3729 struct css_id *id = css->id;
3730 /* When this is called before css_id initialization, id can be NULL */
3734 BUG_ON(!ss->use_id);
3736 rcu_assign_pointer(id->css, NULL);
3737 rcu_assign_pointer(css->id, NULL);
3738 spin_lock(&ss->id_lock);
3739 idr_remove(&ss->idr, id->id);
3740 spin_unlock(&ss->id_lock);
3741 call_rcu(&id->rcu_head, __free_css_id_cb);
3745 * This is called by init or create(). Then, calls to this function are
3746 * always serialized (By cgroup_mutex() at create()).
3749 static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
3751 struct css_id *newid;
3752 int myid, error, size;
3754 BUG_ON(!ss->use_id);
3756 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
3757 newid = kzalloc(size, GFP_KERNEL);
3759 return ERR_PTR(-ENOMEM);
3761 if (unlikely(!idr_pre_get(&ss->idr, GFP_KERNEL))) {
3765 spin_lock(&ss->id_lock);
3766 /* Don't use 0. allocates an ID of 1-65535 */
3767 error = idr_get_new_above(&ss->idr, newid, 1, &myid);
3768 spin_unlock(&ss->id_lock);
3770 /* Returns error when there are no free spaces for new ID.*/
3775 if (myid > CSS_ID_MAX)
3779 newid->depth = depth;
3783 spin_lock(&ss->id_lock);
3784 idr_remove(&ss->idr, myid);
3785 spin_unlock(&ss->id_lock);
3788 return ERR_PTR(error);
3792 static int __init cgroup_subsys_init_idr(struct cgroup_subsys *ss)
3794 struct css_id *newid;
3795 struct cgroup_subsys_state *rootcss;
3797 spin_lock_init(&ss->id_lock);
3800 rootcss = init_css_set.subsys[ss->subsys_id];
3801 newid = get_new_cssid(ss, 0);
3803 return PTR_ERR(newid);
3805 newid->stack[0] = newid->id;
3806 newid->css = rootcss;
3807 rootcss->id = newid;
3811 static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
3812 struct cgroup *child)
3814 int subsys_id, i, depth = 0;
3815 struct cgroup_subsys_state *parent_css, *child_css;
3816 struct css_id *child_id, *parent_id = NULL;
3818 subsys_id = ss->subsys_id;
3819 parent_css = parent->subsys[subsys_id];
3820 child_css = child->subsys[subsys_id];
3821 depth = css_depth(parent_css) + 1;
3822 parent_id = parent_css->id;
3824 child_id = get_new_cssid(ss, depth);
3825 if (IS_ERR(child_id))
3826 return PTR_ERR(child_id);
3828 for (i = 0; i < depth; i++)
3829 child_id->stack[i] = parent_id->stack[i];
3830 child_id->stack[depth] = child_id->id;
3832 * child_id->css pointer will be set after this cgroup is available
3833 * see cgroup_populate_dir()
3835 rcu_assign_pointer(child_css->id, child_id);
3841 * css_lookup - lookup css by id
3842 * @ss: cgroup subsys to be looked into.
3845 * Returns pointer to cgroup_subsys_state if there is valid one with id.
3846 * NULL if not. Should be called under rcu_read_lock()
3848 struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
3850 struct css_id *cssid = NULL;
3852 BUG_ON(!ss->use_id);
3853 cssid = idr_find(&ss->idr, id);
3855 if (unlikely(!cssid))
3858 return rcu_dereference(cssid->css);
3862 * css_get_next - lookup next cgroup under specified hierarchy.
3863 * @ss: pointer to subsystem
3864 * @id: current position of iteration.
3865 * @root: pointer to css. search tree under this.
3866 * @foundid: position of found object.
3868 * Search next css under the specified hierarchy of rootid. Calling under
3869 * rcu_read_lock() is necessary. Returns NULL if it reaches the end.
3871 struct cgroup_subsys_state *
3872 css_get_next(struct cgroup_subsys *ss, int id,
3873 struct cgroup_subsys_state *root, int *foundid)
3875 struct cgroup_subsys_state *ret = NULL;
3878 int rootid = css_id(root);
3879 int depth = css_depth(root);
3884 BUG_ON(!ss->use_id);
3885 /* fill start point for scan */
3889 * scan next entry from bitmap(tree), tmpid is updated after
3892 spin_lock(&ss->id_lock);
3893 tmp = idr_get_next(&ss->idr, &tmpid);
3894 spin_unlock(&ss->id_lock);
3898 if (tmp->depth >= depth && tmp->stack[depth] == rootid) {
3899 ret = rcu_dereference(tmp->css);
3905 /* continue to scan from next id */
3911 #ifdef CONFIG_CGROUP_DEBUG
3912 static struct cgroup_subsys_state *debug_create(struct cgroup_subsys *ss,
3913 struct cgroup *cont)
3915 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
3918 return ERR_PTR(-ENOMEM);
3923 static void debug_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
3925 kfree(cont->subsys[debug_subsys_id]);
3928 static u64 cgroup_refcount_read(struct cgroup *cont, struct cftype *cft)
3930 return atomic_read(&cont->count);
3933 static u64 debug_taskcount_read(struct cgroup *cont, struct cftype *cft)
3935 return cgroup_task_count(cont);
3938 static u64 current_css_set_read(struct cgroup *cont, struct cftype *cft)
3940 return (u64)(unsigned long)current->cgroups;
3943 static u64 current_css_set_refcount_read(struct cgroup *cont,
3949 count = atomic_read(¤t->cgroups->refcount);
3954 static int current_css_set_cg_links_read(struct cgroup *cont,
3956 struct seq_file *seq)
3958 struct cg_cgroup_link *link;
3961 read_lock(&css_set_lock);
3963 cg = rcu_dereference(current->cgroups);
3964 list_for_each_entry(link, &cg->cg_links, cg_link_list) {
3965 struct cgroup *c = link->cgrp;
3969 name = c->dentry->d_name.name;
3972 seq_printf(seq, "Root %d group %s\n",
3973 c->root->hierarchy_id, name);
3976 read_unlock(&css_set_lock);
3980 #define MAX_TASKS_SHOWN_PER_CSS 25
3981 static int cgroup_css_links_read(struct cgroup *cont,
3983 struct seq_file *seq)
3985 struct cg_cgroup_link *link;
3987 read_lock(&css_set_lock);
3988 list_for_each_entry(link, &cont->css_sets, cgrp_link_list) {
3989 struct css_set *cg = link->cg;
3990 struct task_struct *task;
3992 seq_printf(seq, "css_set %p\n", cg);
3993 list_for_each_entry(task, &cg->tasks, cg_list) {
3994 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
3995 seq_puts(seq, " ...\n");
3998 seq_printf(seq, " task %d\n",
3999 task_pid_vnr(task));
4003 read_unlock(&css_set_lock);
4007 static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
4009 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
4012 static struct cftype debug_files[] = {
4014 .name = "cgroup_refcount",
4015 .read_u64 = cgroup_refcount_read,
4018 .name = "taskcount",
4019 .read_u64 = debug_taskcount_read,
4023 .name = "current_css_set",
4024 .read_u64 = current_css_set_read,
4028 .name = "current_css_set_refcount",
4029 .read_u64 = current_css_set_refcount_read,
4033 .name = "current_css_set_cg_links",
4034 .read_seq_string = current_css_set_cg_links_read,
4038 .name = "cgroup_css_links",
4039 .read_seq_string = cgroup_css_links_read,
4043 .name = "releasable",
4044 .read_u64 = releasable_read,
4048 static int debug_populate(struct cgroup_subsys *ss, struct cgroup *cont)
4050 return cgroup_add_files(cont, ss, debug_files,
4051 ARRAY_SIZE(debug_files));
4054 struct cgroup_subsys debug_subsys = {
4056 .create = debug_create,
4057 .destroy = debug_destroy,
4058 .populate = debug_populate,
4059 .subsys_id = debug_subsys_id,
4061 #endif /* CONFIG_CGROUP_DEBUG */