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/errno.h>
28 #include <linux/kernel.h>
29 #include <linux/list.h>
31 #include <linux/mutex.h>
32 #include <linux/mount.h>
33 #include <linux/pagemap.h>
34 #include <linux/proc_fs.h>
35 #include <linux/rcupdate.h>
36 #include <linux/sched.h>
37 #include <linux/backing-dev.h>
38 #include <linux/seq_file.h>
39 #include <linux/slab.h>
40 #include <linux/magic.h>
41 #include <linux/spinlock.h>
42 #include <linux/string.h>
43 #include <linux/sort.h>
44 #include <linux/kmod.h>
45 #include <linux/delayacct.h>
46 #include <linux/cgroupstats.h>
47 #include <linux/hash.h>
48 #include <linux/namei.h>
50 #include <asm/atomic.h>
52 static DEFINE_MUTEX(cgroup_mutex);
54 /* Generate an array of cgroup subsystem pointers */
55 #define SUBSYS(_x) &_x ## _subsys,
57 static struct cgroup_subsys *subsys[] = {
58 #include <linux/cgroup_subsys.h>
62 * A cgroupfs_root represents the root of a cgroup hierarchy,
63 * and may be associated with a superblock to form an active
66 struct cgroupfs_root {
67 struct super_block *sb;
70 * The bitmask of subsystems intended to be attached to this
73 unsigned long subsys_bits;
75 /* The bitmask of subsystems currently attached to this hierarchy */
76 unsigned long actual_subsys_bits;
78 /* A list running through the attached subsystems */
79 struct list_head subsys_list;
81 /* The root cgroup for this hierarchy */
82 struct cgroup top_cgroup;
84 /* Tracks how many cgroups are currently defined in hierarchy.*/
85 int number_of_cgroups;
87 /* A list running through the mounted hierarchies */
88 struct list_head root_list;
90 /* Hierarchy-specific flags */
93 /* The path to use for release notifications. */
94 char release_agent_path[PATH_MAX];
99 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
100 * subsystems that are otherwise unattached - it never has more than a
101 * single cgroup, and all tasks are part of that cgroup.
103 static struct cgroupfs_root rootnode;
105 /* The list of hierarchy roots */
107 static LIST_HEAD(roots);
108 static int root_count;
110 /* dummytop is a shorthand for the dummy hierarchy's top cgroup */
111 #define dummytop (&rootnode.top_cgroup)
113 /* This flag indicates whether tasks in the fork and exit paths should
114 * check for fork/exit handlers to call. This avoids us having to do
115 * extra work in the fork/exit path if none of the subsystems need to
118 static int need_forkexit_callback __read_mostly;
119 static int need_mm_owner_callback __read_mostly;
121 /* convenient tests for these bits */
122 inline int cgroup_is_removed(const struct cgroup *cgrp)
124 return test_bit(CGRP_REMOVED, &cgrp->flags);
127 /* bits in struct cgroupfs_root flags field */
129 ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
132 static int cgroup_is_releasable(const struct cgroup *cgrp)
135 (1 << CGRP_RELEASABLE) |
136 (1 << CGRP_NOTIFY_ON_RELEASE);
137 return (cgrp->flags & bits) == bits;
140 static int notify_on_release(const struct cgroup *cgrp)
142 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
146 * for_each_subsys() allows you to iterate on each subsystem attached to
147 * an active hierarchy
149 #define for_each_subsys(_root, _ss) \
150 list_for_each_entry(_ss, &_root->subsys_list, sibling)
152 /* for_each_root() allows you to iterate across the active hierarchies */
153 #define for_each_root(_root) \
154 list_for_each_entry(_root, &roots, root_list)
156 /* the list of cgroups eligible for automatic release. Protected by
157 * release_list_lock */
158 static LIST_HEAD(release_list);
159 static DEFINE_SPINLOCK(release_list_lock);
160 static void cgroup_release_agent(struct work_struct *work);
161 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
162 static void check_for_release(struct cgroup *cgrp);
164 /* Link structure for associating css_set objects with cgroups */
165 struct cg_cgroup_link {
167 * List running through cg_cgroup_links associated with a
168 * cgroup, anchored on cgroup->css_sets
170 struct list_head cgrp_link_list;
172 * List running through cg_cgroup_links pointing at a
173 * single css_set object, anchored on css_set->cg_links
175 struct list_head cg_link_list;
179 /* The default css_set - used by init and its children prior to any
180 * hierarchies being mounted. It contains a pointer to the root state
181 * for each subsystem. Also used to anchor the list of css_sets. Not
182 * reference-counted, to improve performance when child cgroups
183 * haven't been created.
186 static struct css_set init_css_set;
187 static struct cg_cgroup_link init_css_set_link;
189 /* css_set_lock protects the list of css_set objects, and the
190 * chain of tasks off each css_set. Nests outside task->alloc_lock
191 * due to cgroup_iter_start() */
192 static DEFINE_RWLOCK(css_set_lock);
193 static int css_set_count;
195 /* hash table for cgroup groups. This improves the performance to
196 * find an existing css_set */
197 #define CSS_SET_HASH_BITS 7
198 #define CSS_SET_TABLE_SIZE (1 << CSS_SET_HASH_BITS)
199 static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
201 static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
205 unsigned long tmp = 0UL;
207 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
208 tmp += (unsigned long)css[i];
209 tmp = (tmp >> 16) ^ tmp;
211 index = hash_long(tmp, CSS_SET_HASH_BITS);
213 return &css_set_table[index];
216 /* We don't maintain the lists running through each css_set to its
217 * task until after the first call to cgroup_iter_start(). This
218 * reduces the fork()/exit() overhead for people who have cgroups
219 * compiled into their kernel but not actually in use */
220 static int use_task_css_set_links __read_mostly;
222 /* When we create or destroy a css_set, the operation simply
223 * takes/releases a reference count on all the cgroups referenced
224 * by subsystems in this css_set. This can end up multiple-counting
225 * some cgroups, but that's OK - the ref-count is just a
226 * busy/not-busy indicator; ensuring that we only count each cgroup
227 * once would require taking a global lock to ensure that no
228 * subsystems moved between hierarchies while we were doing so.
230 * Possible TODO: decide at boot time based on the number of
231 * registered subsystems and the number of CPUs or NUMA nodes whether
232 * it's better for performance to ref-count every subsystem, or to
233 * take a global lock and only add one ref count to each hierarchy.
237 * unlink a css_set from the list and free it
239 static void unlink_css_set(struct css_set *cg)
241 struct cg_cgroup_link *link;
242 struct cg_cgroup_link *saved_link;
244 hlist_del(&cg->hlist);
247 list_for_each_entry_safe(link, saved_link, &cg->cg_links,
249 list_del(&link->cg_link_list);
250 list_del(&link->cgrp_link_list);
255 static void __put_css_set(struct css_set *cg, int taskexit)
259 * Ensure that the refcount doesn't hit zero while any readers
260 * can see it. Similar to atomic_dec_and_lock(), but for an
263 if (atomic_add_unless(&cg->refcount, -1, 1))
265 write_lock(&css_set_lock);
266 if (!atomic_dec_and_test(&cg->refcount)) {
267 write_unlock(&css_set_lock);
271 write_unlock(&css_set_lock);
274 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
275 struct cgroup *cgrp = cg->subsys[i]->cgroup;
276 if (atomic_dec_and_test(&cgrp->count) &&
277 notify_on_release(cgrp)) {
279 set_bit(CGRP_RELEASABLE, &cgrp->flags);
280 check_for_release(cgrp);
288 * refcounted get/put for css_set objects
290 static inline void get_css_set(struct css_set *cg)
292 atomic_inc(&cg->refcount);
295 static inline void put_css_set(struct css_set *cg)
297 __put_css_set(cg, 0);
300 static inline void put_css_set_taskexit(struct css_set *cg)
302 __put_css_set(cg, 1);
306 * find_existing_css_set() is a helper for
307 * find_css_set(), and checks to see whether an existing
308 * css_set is suitable.
310 * oldcg: the cgroup group that we're using before the cgroup
313 * cgrp: the cgroup that we're moving into
315 * template: location in which to build the desired set of subsystem
316 * state objects for the new cgroup group
318 static struct css_set *find_existing_css_set(
319 struct css_set *oldcg,
321 struct cgroup_subsys_state *template[])
324 struct cgroupfs_root *root = cgrp->root;
325 struct hlist_head *hhead;
326 struct hlist_node *node;
329 /* Built the set of subsystem state objects that we want to
330 * see in the new css_set */
331 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
332 if (root->subsys_bits & (1UL << i)) {
333 /* Subsystem is in this hierarchy. So we want
334 * the subsystem state from the new
336 template[i] = cgrp->subsys[i];
338 /* Subsystem is not in this hierarchy, so we
339 * don't want to change the subsystem state */
340 template[i] = oldcg->subsys[i];
344 hhead = css_set_hash(template);
345 hlist_for_each_entry(cg, node, hhead, hlist) {
346 if (!memcmp(template, cg->subsys, sizeof(cg->subsys))) {
347 /* All subsystems matched */
352 /* No existing cgroup group matched */
356 static void free_cg_links(struct list_head *tmp)
358 struct cg_cgroup_link *link;
359 struct cg_cgroup_link *saved_link;
361 list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
362 list_del(&link->cgrp_link_list);
368 * allocate_cg_links() allocates "count" cg_cgroup_link structures
369 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
370 * success or a negative error
372 static int allocate_cg_links(int count, struct list_head *tmp)
374 struct cg_cgroup_link *link;
377 for (i = 0; i < count; i++) {
378 link = kmalloc(sizeof(*link), GFP_KERNEL);
383 list_add(&link->cgrp_link_list, tmp);
389 * find_css_set() takes an existing cgroup group and a
390 * cgroup object, and returns a css_set object that's
391 * equivalent to the old group, but with the given cgroup
392 * substituted into the appropriate hierarchy. Must be called with
395 static struct css_set *find_css_set(
396 struct css_set *oldcg, struct cgroup *cgrp)
399 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
402 struct list_head tmp_cg_links;
403 struct cg_cgroup_link *link;
405 struct hlist_head *hhead;
407 /* First see if we already have a cgroup group that matches
409 read_lock(&css_set_lock);
410 res = find_existing_css_set(oldcg, cgrp, template);
413 read_unlock(&css_set_lock);
418 res = kmalloc(sizeof(*res), GFP_KERNEL);
422 /* Allocate all the cg_cgroup_link objects that we'll need */
423 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
428 atomic_set(&res->refcount, 1);
429 INIT_LIST_HEAD(&res->cg_links);
430 INIT_LIST_HEAD(&res->tasks);
431 INIT_HLIST_NODE(&res->hlist);
433 /* Copy the set of subsystem state objects generated in
434 * find_existing_css_set() */
435 memcpy(res->subsys, template, sizeof(res->subsys));
437 write_lock(&css_set_lock);
438 /* Add reference counts and links from the new css_set. */
439 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
440 struct cgroup *cgrp = res->subsys[i]->cgroup;
441 struct cgroup_subsys *ss = subsys[i];
442 atomic_inc(&cgrp->count);
444 * We want to add a link once per cgroup, so we
445 * only do it for the first subsystem in each
448 if (ss->root->subsys_list.next == &ss->sibling) {
449 BUG_ON(list_empty(&tmp_cg_links));
450 link = list_entry(tmp_cg_links.next,
451 struct cg_cgroup_link,
453 list_del(&link->cgrp_link_list);
454 list_add(&link->cgrp_link_list, &cgrp->css_sets);
456 list_add(&link->cg_link_list, &res->cg_links);
459 if (list_empty(&rootnode.subsys_list)) {
460 link = list_entry(tmp_cg_links.next,
461 struct cg_cgroup_link,
463 list_del(&link->cgrp_link_list);
464 list_add(&link->cgrp_link_list, &dummytop->css_sets);
466 list_add(&link->cg_link_list, &res->cg_links);
469 BUG_ON(!list_empty(&tmp_cg_links));
473 /* Add this cgroup group to the hash table */
474 hhead = css_set_hash(res->subsys);
475 hlist_add_head(&res->hlist, hhead);
477 write_unlock(&css_set_lock);
483 * There is one global cgroup mutex. We also require taking
484 * task_lock() when dereferencing a task's cgroup subsys pointers.
485 * See "The task_lock() exception", at the end of this comment.
487 * A task must hold cgroup_mutex to modify cgroups.
489 * Any task can increment and decrement the count field without lock.
490 * So in general, code holding cgroup_mutex can't rely on the count
491 * field not changing. However, if the count goes to zero, then only
492 * cgroup_attach_task() can increment it again. Because a count of zero
493 * means that no tasks are currently attached, therefore there is no
494 * way a task attached to that cgroup can fork (the other way to
495 * increment the count). So code holding cgroup_mutex can safely
496 * assume that if the count is zero, it will stay zero. Similarly, if
497 * a task holds cgroup_mutex on a cgroup with zero count, it
498 * knows that the cgroup won't be removed, as cgroup_rmdir()
501 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
502 * (usually) take cgroup_mutex. These are the two most performance
503 * critical pieces of code here. The exception occurs on cgroup_exit(),
504 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
505 * is taken, and if the cgroup count is zero, a usermode call made
506 * to the release agent with the name of the cgroup (path relative to
507 * the root of cgroup file system) as the argument.
509 * A cgroup can only be deleted if both its 'count' of using tasks
510 * is zero, and its list of 'children' cgroups is empty. Since all
511 * tasks in the system use _some_ cgroup, and since there is always at
512 * least one task in the system (init, pid == 1), therefore, top_cgroup
513 * always has either children cgroups and/or using tasks. So we don't
514 * need a special hack to ensure that top_cgroup cannot be deleted.
516 * The task_lock() exception
518 * The need for this exception arises from the action of
519 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
520 * another. It does so using cgroup_mutex, however there are
521 * several performance critical places that need to reference
522 * task->cgroup without the expense of grabbing a system global
523 * mutex. Therefore except as noted below, when dereferencing or, as
524 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
525 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
526 * the task_struct routinely used for such matters.
528 * P.S. One more locking exception. RCU is used to guard the
529 * update of a tasks cgroup pointer by cgroup_attach_task()
533 * cgroup_lock - lock out any changes to cgroup structures
536 void cgroup_lock(void)
538 mutex_lock(&cgroup_mutex);
542 * cgroup_unlock - release lock on cgroup changes
544 * Undo the lock taken in a previous cgroup_lock() call.
546 void cgroup_unlock(void)
548 mutex_unlock(&cgroup_mutex);
552 * A couple of forward declarations required, due to cyclic reference loop:
553 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
554 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
558 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
559 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
560 static int cgroup_populate_dir(struct cgroup *cgrp);
561 static struct inode_operations cgroup_dir_inode_operations;
562 static struct file_operations proc_cgroupstats_operations;
564 static struct backing_dev_info cgroup_backing_dev_info = {
565 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
568 static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
570 struct inode *inode = new_inode(sb);
573 inode->i_mode = mode;
574 inode->i_uid = current->fsuid;
575 inode->i_gid = current->fsgid;
577 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
578 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
584 * Call subsys's pre_destroy handler.
585 * This is called before css refcnt check.
587 static void cgroup_call_pre_destroy(struct cgroup *cgrp)
589 struct cgroup_subsys *ss;
590 for_each_subsys(cgrp->root, ss)
591 if (ss->pre_destroy && cgrp->subsys[ss->subsys_id])
592 ss->pre_destroy(ss, cgrp);
596 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
598 /* is dentry a directory ? if so, kfree() associated cgroup */
599 if (S_ISDIR(inode->i_mode)) {
600 struct cgroup *cgrp = dentry->d_fsdata;
601 struct cgroup_subsys *ss;
602 BUG_ON(!(cgroup_is_removed(cgrp)));
603 /* It's possible for external users to be holding css
604 * reference counts on a cgroup; css_put() needs to
605 * be able to access the cgroup after decrementing
606 * the reference count in order to know if it needs to
607 * queue the cgroup to be handled by the release
611 mutex_lock(&cgroup_mutex);
613 * Release the subsystem state objects.
615 for_each_subsys(cgrp->root, ss) {
616 if (cgrp->subsys[ss->subsys_id])
617 ss->destroy(ss, cgrp);
620 cgrp->root->number_of_cgroups--;
621 mutex_unlock(&cgroup_mutex);
623 /* Drop the active superblock reference that we took when we
624 * created the cgroup */
625 deactivate_super(cgrp->root->sb);
632 static void remove_dir(struct dentry *d)
634 struct dentry *parent = dget(d->d_parent);
637 simple_rmdir(parent->d_inode, d);
641 static void cgroup_clear_directory(struct dentry *dentry)
643 struct list_head *node;
645 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
646 spin_lock(&dcache_lock);
647 node = dentry->d_subdirs.next;
648 while (node != &dentry->d_subdirs) {
649 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
652 /* This should never be called on a cgroup
653 * directory with child cgroups */
654 BUG_ON(d->d_inode->i_mode & S_IFDIR);
656 spin_unlock(&dcache_lock);
658 simple_unlink(dentry->d_inode, d);
660 spin_lock(&dcache_lock);
662 node = dentry->d_subdirs.next;
664 spin_unlock(&dcache_lock);
668 * NOTE : the dentry must have been dget()'ed
670 static void cgroup_d_remove_dir(struct dentry *dentry)
672 cgroup_clear_directory(dentry);
674 spin_lock(&dcache_lock);
675 list_del_init(&dentry->d_u.d_child);
676 spin_unlock(&dcache_lock);
680 static int rebind_subsystems(struct cgroupfs_root *root,
681 unsigned long final_bits)
683 unsigned long added_bits, removed_bits;
684 struct cgroup *cgrp = &root->top_cgroup;
687 removed_bits = root->actual_subsys_bits & ~final_bits;
688 added_bits = final_bits & ~root->actual_subsys_bits;
689 /* Check that any added subsystems are currently free */
690 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
691 unsigned long bit = 1UL << i;
692 struct cgroup_subsys *ss = subsys[i];
693 if (!(bit & added_bits))
695 if (ss->root != &rootnode) {
696 /* Subsystem isn't free */
701 /* Currently we don't handle adding/removing subsystems when
702 * any child cgroups exist. This is theoretically supportable
703 * but involves complex error handling, so it's being left until
705 if (!list_empty(&cgrp->children))
708 /* Process each subsystem */
709 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
710 struct cgroup_subsys *ss = subsys[i];
711 unsigned long bit = 1UL << i;
712 if (bit & added_bits) {
713 /* We're binding this subsystem to this hierarchy */
714 BUG_ON(cgrp->subsys[i]);
715 BUG_ON(!dummytop->subsys[i]);
716 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
717 cgrp->subsys[i] = dummytop->subsys[i];
718 cgrp->subsys[i]->cgroup = cgrp;
719 list_add(&ss->sibling, &root->subsys_list);
720 rcu_assign_pointer(ss->root, root);
724 } else if (bit & removed_bits) {
725 /* We're removing this subsystem */
726 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
727 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
729 ss->bind(ss, dummytop);
730 dummytop->subsys[i]->cgroup = dummytop;
731 cgrp->subsys[i] = NULL;
732 rcu_assign_pointer(subsys[i]->root, &rootnode);
733 list_del(&ss->sibling);
734 } else if (bit & final_bits) {
735 /* Subsystem state should already exist */
736 BUG_ON(!cgrp->subsys[i]);
738 /* Subsystem state shouldn't exist */
739 BUG_ON(cgrp->subsys[i]);
742 root->subsys_bits = root->actual_subsys_bits = final_bits;
748 static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
750 struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
751 struct cgroup_subsys *ss;
753 mutex_lock(&cgroup_mutex);
754 for_each_subsys(root, ss)
755 seq_printf(seq, ",%s", ss->name);
756 if (test_bit(ROOT_NOPREFIX, &root->flags))
757 seq_puts(seq, ",noprefix");
758 if (strlen(root->release_agent_path))
759 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
760 mutex_unlock(&cgroup_mutex);
764 struct cgroup_sb_opts {
765 unsigned long subsys_bits;
770 /* Convert a hierarchy specifier into a bitmask of subsystems and
772 static int parse_cgroupfs_options(char *data,
773 struct cgroup_sb_opts *opts)
775 char *token, *o = data ?: "all";
777 opts->subsys_bits = 0;
779 opts->release_agent = NULL;
781 while ((token = strsep(&o, ",")) != NULL) {
784 if (!strcmp(token, "all")) {
785 /* Add all non-disabled subsystems */
787 opts->subsys_bits = 0;
788 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
789 struct cgroup_subsys *ss = subsys[i];
791 opts->subsys_bits |= 1ul << i;
793 } else if (!strcmp(token, "noprefix")) {
794 set_bit(ROOT_NOPREFIX, &opts->flags);
795 } else if (!strncmp(token, "release_agent=", 14)) {
796 /* Specifying two release agents is forbidden */
797 if (opts->release_agent)
799 opts->release_agent = kzalloc(PATH_MAX, GFP_KERNEL);
800 if (!opts->release_agent)
802 strncpy(opts->release_agent, token + 14, PATH_MAX - 1);
803 opts->release_agent[PATH_MAX - 1] = 0;
805 struct cgroup_subsys *ss;
807 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
809 if (!strcmp(token, ss->name)) {
811 set_bit(i, &opts->subsys_bits);
815 if (i == CGROUP_SUBSYS_COUNT)
820 /* We can't have an empty hierarchy */
821 if (!opts->subsys_bits)
827 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
830 struct cgroupfs_root *root = sb->s_fs_info;
831 struct cgroup *cgrp = &root->top_cgroup;
832 struct cgroup_sb_opts opts;
834 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
835 mutex_lock(&cgroup_mutex);
837 /* See what subsystems are wanted */
838 ret = parse_cgroupfs_options(data, &opts);
842 /* Don't allow flags to change at remount */
843 if (opts.flags != root->flags) {
848 ret = rebind_subsystems(root, opts.subsys_bits);
850 /* (re)populate subsystem files */
852 cgroup_populate_dir(cgrp);
854 if (opts.release_agent)
855 strcpy(root->release_agent_path, opts.release_agent);
857 if (opts.release_agent)
858 kfree(opts.release_agent);
859 mutex_unlock(&cgroup_mutex);
860 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
864 static struct super_operations cgroup_ops = {
865 .statfs = simple_statfs,
866 .drop_inode = generic_delete_inode,
867 .show_options = cgroup_show_options,
868 .remount_fs = cgroup_remount,
871 static void init_cgroup_housekeeping(struct cgroup *cgrp)
873 INIT_LIST_HEAD(&cgrp->sibling);
874 INIT_LIST_HEAD(&cgrp->children);
875 INIT_LIST_HEAD(&cgrp->css_sets);
876 INIT_LIST_HEAD(&cgrp->release_list);
877 init_rwsem(&cgrp->pids_mutex);
879 static void init_cgroup_root(struct cgroupfs_root *root)
881 struct cgroup *cgrp = &root->top_cgroup;
882 INIT_LIST_HEAD(&root->subsys_list);
883 INIT_LIST_HEAD(&root->root_list);
884 root->number_of_cgroups = 1;
886 cgrp->top_cgroup = cgrp;
887 init_cgroup_housekeeping(cgrp);
890 static int cgroup_test_super(struct super_block *sb, void *data)
892 struct cgroupfs_root *new = data;
893 struct cgroupfs_root *root = sb->s_fs_info;
895 /* First check subsystems */
896 if (new->subsys_bits != root->subsys_bits)
899 /* Next check flags */
900 if (new->flags != root->flags)
906 static int cgroup_set_super(struct super_block *sb, void *data)
909 struct cgroupfs_root *root = data;
911 ret = set_anon_super(sb, NULL);
915 sb->s_fs_info = root;
918 sb->s_blocksize = PAGE_CACHE_SIZE;
919 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
920 sb->s_magic = CGROUP_SUPER_MAGIC;
921 sb->s_op = &cgroup_ops;
926 static int cgroup_get_rootdir(struct super_block *sb)
928 struct inode *inode =
929 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
930 struct dentry *dentry;
935 inode->i_fop = &simple_dir_operations;
936 inode->i_op = &cgroup_dir_inode_operations;
937 /* directories start off with i_nlink == 2 (for "." entry) */
939 dentry = d_alloc_root(inode);
948 static int cgroup_get_sb(struct file_system_type *fs_type,
949 int flags, const char *unused_dev_name,
950 void *data, struct vfsmount *mnt)
952 struct cgroup_sb_opts opts;
954 struct super_block *sb;
955 struct cgroupfs_root *root;
956 struct list_head tmp_cg_links;
958 /* First find the desired set of subsystems */
959 ret = parse_cgroupfs_options(data, &opts);
961 if (opts.release_agent)
962 kfree(opts.release_agent);
966 root = kzalloc(sizeof(*root), GFP_KERNEL);
968 if (opts.release_agent)
969 kfree(opts.release_agent);
973 init_cgroup_root(root);
974 root->subsys_bits = opts.subsys_bits;
975 root->flags = opts.flags;
976 if (opts.release_agent) {
977 strcpy(root->release_agent_path, opts.release_agent);
978 kfree(opts.release_agent);
981 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, root);
988 if (sb->s_fs_info != root) {
989 /* Reusing an existing superblock */
990 BUG_ON(sb->s_root == NULL);
995 struct cgroup *cgrp = &root->top_cgroup;
999 BUG_ON(sb->s_root != NULL);
1001 ret = cgroup_get_rootdir(sb);
1003 goto drop_new_super;
1004 inode = sb->s_root->d_inode;
1006 mutex_lock(&inode->i_mutex);
1007 mutex_lock(&cgroup_mutex);
1010 * We're accessing css_set_count without locking
1011 * css_set_lock here, but that's OK - it can only be
1012 * increased by someone holding cgroup_lock, and
1013 * that's us. The worst that can happen is that we
1014 * have some link structures left over
1016 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1018 mutex_unlock(&cgroup_mutex);
1019 mutex_unlock(&inode->i_mutex);
1020 goto drop_new_super;
1023 ret = rebind_subsystems(root, root->subsys_bits);
1024 if (ret == -EBUSY) {
1025 mutex_unlock(&cgroup_mutex);
1026 mutex_unlock(&inode->i_mutex);
1027 goto drop_new_super;
1030 /* EBUSY should be the only error here */
1033 list_add(&root->root_list, &roots);
1036 sb->s_root->d_fsdata = &root->top_cgroup;
1037 root->top_cgroup.dentry = sb->s_root;
1039 /* Link the top cgroup in this hierarchy into all
1040 * the css_set objects */
1041 write_lock(&css_set_lock);
1042 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1043 struct hlist_head *hhead = &css_set_table[i];
1044 struct hlist_node *node;
1047 hlist_for_each_entry(cg, node, hhead, hlist) {
1048 struct cg_cgroup_link *link;
1050 BUG_ON(list_empty(&tmp_cg_links));
1051 link = list_entry(tmp_cg_links.next,
1052 struct cg_cgroup_link,
1054 list_del(&link->cgrp_link_list);
1056 list_add(&link->cgrp_link_list,
1057 &root->top_cgroup.css_sets);
1058 list_add(&link->cg_link_list, &cg->cg_links);
1061 write_unlock(&css_set_lock);
1063 free_cg_links(&tmp_cg_links);
1065 BUG_ON(!list_empty(&cgrp->sibling));
1066 BUG_ON(!list_empty(&cgrp->children));
1067 BUG_ON(root->number_of_cgroups != 1);
1069 cgroup_populate_dir(cgrp);
1070 mutex_unlock(&inode->i_mutex);
1071 mutex_unlock(&cgroup_mutex);
1074 return simple_set_mnt(mnt, sb);
1077 up_write(&sb->s_umount);
1078 deactivate_super(sb);
1079 free_cg_links(&tmp_cg_links);
1083 static void cgroup_kill_sb(struct super_block *sb) {
1084 struct cgroupfs_root *root = sb->s_fs_info;
1085 struct cgroup *cgrp = &root->top_cgroup;
1087 struct cg_cgroup_link *link;
1088 struct cg_cgroup_link *saved_link;
1092 BUG_ON(root->number_of_cgroups != 1);
1093 BUG_ON(!list_empty(&cgrp->children));
1094 BUG_ON(!list_empty(&cgrp->sibling));
1096 mutex_lock(&cgroup_mutex);
1098 /* Rebind all subsystems back to the default hierarchy */
1099 ret = rebind_subsystems(root, 0);
1100 /* Shouldn't be able to fail ... */
1104 * Release all the links from css_sets to this hierarchy's
1107 write_lock(&css_set_lock);
1109 list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1111 list_del(&link->cg_link_list);
1112 list_del(&link->cgrp_link_list);
1115 write_unlock(&css_set_lock);
1117 if (!list_empty(&root->root_list)) {
1118 list_del(&root->root_list);
1121 mutex_unlock(&cgroup_mutex);
1124 kill_litter_super(sb);
1127 static struct file_system_type cgroup_fs_type = {
1129 .get_sb = cgroup_get_sb,
1130 .kill_sb = cgroup_kill_sb,
1133 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1135 return dentry->d_fsdata;
1138 static inline struct cftype *__d_cft(struct dentry *dentry)
1140 return dentry->d_fsdata;
1144 * cgroup_path - generate the path of a cgroup
1145 * @cgrp: the cgroup in question
1146 * @buf: the buffer to write the path into
1147 * @buflen: the length of the buffer
1149 * Called with cgroup_mutex held. Writes path of cgroup into buf.
1150 * Returns 0 on success, -errno on error.
1152 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1156 if (cgrp == dummytop) {
1158 * Inactive subsystems have no dentry for their root
1165 start = buf + buflen;
1169 int len = cgrp->dentry->d_name.len;
1170 if ((start -= len) < buf)
1171 return -ENAMETOOLONG;
1172 memcpy(start, cgrp->dentry->d_name.name, len);
1173 cgrp = cgrp->parent;
1179 return -ENAMETOOLONG;
1182 memmove(buf, start, buf + buflen - start);
1187 * Return the first subsystem attached to a cgroup's hierarchy, and
1191 static void get_first_subsys(const struct cgroup *cgrp,
1192 struct cgroup_subsys_state **css, int *subsys_id)
1194 const struct cgroupfs_root *root = cgrp->root;
1195 const struct cgroup_subsys *test_ss;
1196 BUG_ON(list_empty(&root->subsys_list));
1197 test_ss = list_entry(root->subsys_list.next,
1198 struct cgroup_subsys, sibling);
1200 *css = cgrp->subsys[test_ss->subsys_id];
1204 *subsys_id = test_ss->subsys_id;
1208 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1209 * @cgrp: the cgroup the task is attaching to
1210 * @tsk: the task to be attached
1212 * Call holding cgroup_mutex. May take task_lock of
1213 * the task 'tsk' during call.
1215 int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1218 struct cgroup_subsys *ss;
1219 struct cgroup *oldcgrp;
1220 struct css_set *cg = tsk->cgroups;
1221 struct css_set *newcg;
1222 struct cgroupfs_root *root = cgrp->root;
1225 get_first_subsys(cgrp, NULL, &subsys_id);
1227 /* Nothing to do if the task is already in that cgroup */
1228 oldcgrp = task_cgroup(tsk, subsys_id);
1229 if (cgrp == oldcgrp)
1232 for_each_subsys(root, ss) {
1233 if (ss->can_attach) {
1234 retval = ss->can_attach(ss, cgrp, tsk);
1241 * Locate or allocate a new css_set for this task,
1242 * based on its final set of cgroups
1244 newcg = find_css_set(cg, cgrp);
1249 if (tsk->flags & PF_EXITING) {
1254 rcu_assign_pointer(tsk->cgroups, newcg);
1257 /* Update the css_set linked lists if we're using them */
1258 write_lock(&css_set_lock);
1259 if (!list_empty(&tsk->cg_list)) {
1260 list_del(&tsk->cg_list);
1261 list_add(&tsk->cg_list, &newcg->tasks);
1263 write_unlock(&css_set_lock);
1265 for_each_subsys(root, ss) {
1267 ss->attach(ss, cgrp, oldcgrp, tsk);
1269 set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1276 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
1277 * held. May take task_lock of task
1279 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid)
1281 struct task_struct *tsk;
1286 tsk = find_task_by_vpid(pid);
1287 if (!tsk || tsk->flags & PF_EXITING) {
1291 get_task_struct(tsk);
1294 if ((current->euid) && (current->euid != tsk->uid)
1295 && (current->euid != tsk->suid)) {
1296 put_task_struct(tsk);
1301 get_task_struct(tsk);
1304 ret = cgroup_attach_task(cgrp, tsk);
1305 put_task_struct(tsk);
1309 static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
1312 if (!cgroup_lock_live_group(cgrp))
1314 ret = attach_task_by_pid(cgrp, pid);
1319 /* The various types of files and directories in a cgroup file system */
1320 enum cgroup_filetype {
1324 FILE_NOTIFY_ON_RELEASE,
1329 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
1330 * @cgrp: the cgroup to be checked for liveness
1332 * On success, returns true; the lock should be later released with
1333 * cgroup_unlock(). On failure returns false with no lock held.
1335 bool cgroup_lock_live_group(struct cgroup *cgrp)
1337 mutex_lock(&cgroup_mutex);
1338 if (cgroup_is_removed(cgrp)) {
1339 mutex_unlock(&cgroup_mutex);
1345 static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
1348 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1349 if (!cgroup_lock_live_group(cgrp))
1351 strcpy(cgrp->root->release_agent_path, buffer);
1356 static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
1357 struct seq_file *seq)
1359 if (!cgroup_lock_live_group(cgrp))
1361 seq_puts(seq, cgrp->root->release_agent_path);
1362 seq_putc(seq, '\n');
1367 /* A buffer size big enough for numbers or short strings */
1368 #define CGROUP_LOCAL_BUFFER_SIZE 64
1370 static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
1372 const char __user *userbuf,
1373 size_t nbytes, loff_t *unused_ppos)
1375 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
1381 if (nbytes >= sizeof(buffer))
1383 if (copy_from_user(buffer, userbuf, nbytes))
1386 buffer[nbytes] = 0; /* nul-terminate */
1388 if (cft->write_u64) {
1389 u64 val = simple_strtoull(buffer, &end, 0);
1392 retval = cft->write_u64(cgrp, cft, val);
1394 s64 val = simple_strtoll(buffer, &end, 0);
1397 retval = cft->write_s64(cgrp, cft, val);
1404 static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
1406 const char __user *userbuf,
1407 size_t nbytes, loff_t *unused_ppos)
1409 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
1411 size_t max_bytes = cft->max_write_len;
1412 char *buffer = local_buffer;
1415 max_bytes = sizeof(local_buffer) - 1;
1416 if (nbytes >= max_bytes)
1418 /* Allocate a dynamic buffer if we need one */
1419 if (nbytes >= sizeof(local_buffer)) {
1420 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
1424 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
1429 buffer[nbytes] = 0; /* nul-terminate */
1431 retval = cft->write_string(cgrp, cft, buffer);
1435 if (buffer != local_buffer)
1440 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
1441 size_t nbytes, loff_t *ppos)
1443 struct cftype *cft = __d_cft(file->f_dentry);
1444 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1446 if (!cft || cgroup_is_removed(cgrp))
1449 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
1450 if (cft->write_u64 || cft->write_s64)
1451 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
1452 if (cft->write_string)
1453 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
1455 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
1456 return ret ? ret : nbytes;
1461 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
1463 char __user *buf, size_t nbytes,
1466 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
1467 u64 val = cft->read_u64(cgrp, cft);
1468 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
1470 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1473 static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
1475 char __user *buf, size_t nbytes,
1478 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
1479 s64 val = cft->read_s64(cgrp, cft);
1480 int len = sprintf(tmp, "%lld\n", (long long) val);
1482 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1485 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
1486 size_t nbytes, loff_t *ppos)
1488 struct cftype *cft = __d_cft(file->f_dentry);
1489 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1491 if (!cft || cgroup_is_removed(cgrp))
1495 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
1497 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
1499 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
1504 * seqfile ops/methods for returning structured data. Currently just
1505 * supports string->u64 maps, but can be extended in future.
1508 struct cgroup_seqfile_state {
1510 struct cgroup *cgroup;
1513 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
1515 struct seq_file *sf = cb->state;
1516 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
1519 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
1521 struct cgroup_seqfile_state *state = m->private;
1522 struct cftype *cft = state->cft;
1523 if (cft->read_map) {
1524 struct cgroup_map_cb cb = {
1525 .fill = cgroup_map_add,
1528 return cft->read_map(state->cgroup, cft, &cb);
1530 return cft->read_seq_string(state->cgroup, cft, m);
1533 static int cgroup_seqfile_release(struct inode *inode, struct file *file)
1535 struct seq_file *seq = file->private_data;
1536 kfree(seq->private);
1537 return single_release(inode, file);
1540 static struct file_operations cgroup_seqfile_operations = {
1542 .write = cgroup_file_write,
1543 .llseek = seq_lseek,
1544 .release = cgroup_seqfile_release,
1547 static int cgroup_file_open(struct inode *inode, struct file *file)
1552 err = generic_file_open(inode, file);
1556 cft = __d_cft(file->f_dentry);
1559 if (cft->read_map || cft->read_seq_string) {
1560 struct cgroup_seqfile_state *state =
1561 kzalloc(sizeof(*state), GFP_USER);
1565 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
1566 file->f_op = &cgroup_seqfile_operations;
1567 err = single_open(file, cgroup_seqfile_show, state);
1570 } else if (cft->open)
1571 err = cft->open(inode, file);
1578 static int cgroup_file_release(struct inode *inode, struct file *file)
1580 struct cftype *cft = __d_cft(file->f_dentry);
1582 return cft->release(inode, file);
1587 * cgroup_rename - Only allow simple rename of directories in place.
1589 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
1590 struct inode *new_dir, struct dentry *new_dentry)
1592 if (!S_ISDIR(old_dentry->d_inode->i_mode))
1594 if (new_dentry->d_inode)
1596 if (old_dir != new_dir)
1598 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
1601 static struct file_operations cgroup_file_operations = {
1602 .read = cgroup_file_read,
1603 .write = cgroup_file_write,
1604 .llseek = generic_file_llseek,
1605 .open = cgroup_file_open,
1606 .release = cgroup_file_release,
1609 static struct inode_operations cgroup_dir_inode_operations = {
1610 .lookup = simple_lookup,
1611 .mkdir = cgroup_mkdir,
1612 .rmdir = cgroup_rmdir,
1613 .rename = cgroup_rename,
1616 static int cgroup_create_file(struct dentry *dentry, int mode,
1617 struct super_block *sb)
1619 static struct dentry_operations cgroup_dops = {
1620 .d_iput = cgroup_diput,
1623 struct inode *inode;
1627 if (dentry->d_inode)
1630 inode = cgroup_new_inode(mode, sb);
1634 if (S_ISDIR(mode)) {
1635 inode->i_op = &cgroup_dir_inode_operations;
1636 inode->i_fop = &simple_dir_operations;
1638 /* start off with i_nlink == 2 (for "." entry) */
1641 /* start with the directory inode held, so that we can
1642 * populate it without racing with another mkdir */
1643 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1644 } else if (S_ISREG(mode)) {
1646 inode->i_fop = &cgroup_file_operations;
1648 dentry->d_op = &cgroup_dops;
1649 d_instantiate(dentry, inode);
1650 dget(dentry); /* Extra count - pin the dentry in core */
1655 * cgroup_create_dir - create a directory for an object.
1656 * @cgrp: the cgroup we create the directory for. It must have a valid
1657 * ->parent field. And we are going to fill its ->dentry field.
1658 * @dentry: dentry of the new cgroup
1659 * @mode: mode to set on new directory.
1661 static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
1664 struct dentry *parent;
1667 parent = cgrp->parent->dentry;
1668 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
1670 dentry->d_fsdata = cgrp;
1671 inc_nlink(parent->d_inode);
1672 cgrp->dentry = dentry;
1680 int cgroup_add_file(struct cgroup *cgrp,
1681 struct cgroup_subsys *subsys,
1682 const struct cftype *cft)
1684 struct dentry *dir = cgrp->dentry;
1685 struct dentry *dentry;
1688 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
1689 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
1690 strcpy(name, subsys->name);
1693 strcat(name, cft->name);
1694 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
1695 dentry = lookup_one_len(name, dir, strlen(name));
1696 if (!IS_ERR(dentry)) {
1697 error = cgroup_create_file(dentry, 0644 | S_IFREG,
1700 dentry->d_fsdata = (void *)cft;
1703 error = PTR_ERR(dentry);
1707 int cgroup_add_files(struct cgroup *cgrp,
1708 struct cgroup_subsys *subsys,
1709 const struct cftype cft[],
1713 for (i = 0; i < count; i++) {
1714 err = cgroup_add_file(cgrp, subsys, &cft[i]);
1722 * cgroup_task_count - count the number of tasks in a cgroup.
1723 * @cgrp: the cgroup in question
1725 * Return the number of tasks in the cgroup.
1727 int cgroup_task_count(const struct cgroup *cgrp)
1730 struct cg_cgroup_link *link;
1732 read_lock(&css_set_lock);
1733 list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
1734 count += atomic_read(&link->cg->refcount);
1736 read_unlock(&css_set_lock);
1741 * Advance a list_head iterator. The iterator should be positioned at
1742 * the start of a css_set
1744 static void cgroup_advance_iter(struct cgroup *cgrp,
1745 struct cgroup_iter *it)
1747 struct list_head *l = it->cg_link;
1748 struct cg_cgroup_link *link;
1751 /* Advance to the next non-empty css_set */
1754 if (l == &cgrp->css_sets) {
1758 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
1760 } while (list_empty(&cg->tasks));
1762 it->task = cg->tasks.next;
1766 * To reduce the fork() overhead for systems that are not actually
1767 * using their cgroups capability, we don't maintain the lists running
1768 * through each css_set to its tasks until we see the list actually
1769 * used - in other words after the first call to cgroup_iter_start().
1771 * The tasklist_lock is not held here, as do_each_thread() and
1772 * while_each_thread() are protected by RCU.
1774 static void cgroup_enable_task_cg_lists(void)
1776 struct task_struct *p, *g;
1777 write_lock(&css_set_lock);
1778 use_task_css_set_links = 1;
1779 do_each_thread(g, p) {
1782 * We should check if the process is exiting, otherwise
1783 * it will race with cgroup_exit() in that the list
1784 * entry won't be deleted though the process has exited.
1786 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
1787 list_add(&p->cg_list, &p->cgroups->tasks);
1789 } while_each_thread(g, p);
1790 write_unlock(&css_set_lock);
1793 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
1796 * The first time anyone tries to iterate across a cgroup,
1797 * we need to enable the list linking each css_set to its
1798 * tasks, and fix up all existing tasks.
1800 if (!use_task_css_set_links)
1801 cgroup_enable_task_cg_lists();
1803 read_lock(&css_set_lock);
1804 it->cg_link = &cgrp->css_sets;
1805 cgroup_advance_iter(cgrp, it);
1808 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
1809 struct cgroup_iter *it)
1811 struct task_struct *res;
1812 struct list_head *l = it->task;
1814 /* If the iterator cg is NULL, we have no tasks */
1817 res = list_entry(l, struct task_struct, cg_list);
1818 /* Advance iterator to find next entry */
1820 if (l == &res->cgroups->tasks) {
1821 /* We reached the end of this task list - move on to
1822 * the next cg_cgroup_link */
1823 cgroup_advance_iter(cgrp, it);
1830 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
1832 read_unlock(&css_set_lock);
1835 static inline int started_after_time(struct task_struct *t1,
1836 struct timespec *time,
1837 struct task_struct *t2)
1839 int start_diff = timespec_compare(&t1->start_time, time);
1840 if (start_diff > 0) {
1842 } else if (start_diff < 0) {
1846 * Arbitrarily, if two processes started at the same
1847 * time, we'll say that the lower pointer value
1848 * started first. Note that t2 may have exited by now
1849 * so this may not be a valid pointer any longer, but
1850 * that's fine - it still serves to distinguish
1851 * between two tasks started (effectively) simultaneously.
1858 * This function is a callback from heap_insert() and is used to order
1860 * In this case we order the heap in descending task start time.
1862 static inline int started_after(void *p1, void *p2)
1864 struct task_struct *t1 = p1;
1865 struct task_struct *t2 = p2;
1866 return started_after_time(t1, &t2->start_time, t2);
1870 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
1871 * @scan: struct cgroup_scanner containing arguments for the scan
1873 * Arguments include pointers to callback functions test_task() and
1875 * Iterate through all the tasks in a cgroup, calling test_task() for each,
1876 * and if it returns true, call process_task() for it also.
1877 * The test_task pointer may be NULL, meaning always true (select all tasks).
1878 * Effectively duplicates cgroup_iter_{start,next,end}()
1879 * but does not lock css_set_lock for the call to process_task().
1880 * The struct cgroup_scanner may be embedded in any structure of the caller's
1882 * It is guaranteed that process_task() will act on every task that
1883 * is a member of the cgroup for the duration of this call. This
1884 * function may or may not call process_task() for tasks that exit
1885 * or move to a different cgroup during the call, or are forked or
1886 * move into the cgroup during the call.
1888 * Note that test_task() may be called with locks held, and may in some
1889 * situations be called multiple times for the same task, so it should
1891 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
1892 * pre-allocated and will be used for heap operations (and its "gt" member will
1893 * be overwritten), else a temporary heap will be used (allocation of which
1894 * may cause this function to fail).
1896 int cgroup_scan_tasks(struct cgroup_scanner *scan)
1899 struct cgroup_iter it;
1900 struct task_struct *p, *dropped;
1901 /* Never dereference latest_task, since it's not refcounted */
1902 struct task_struct *latest_task = NULL;
1903 struct ptr_heap tmp_heap;
1904 struct ptr_heap *heap;
1905 struct timespec latest_time = { 0, 0 };
1908 /* The caller supplied our heap and pre-allocated its memory */
1910 heap->gt = &started_after;
1912 /* We need to allocate our own heap memory */
1914 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
1916 /* cannot allocate the heap */
1922 * Scan tasks in the cgroup, using the scanner's "test_task" callback
1923 * to determine which are of interest, and using the scanner's
1924 * "process_task" callback to process any of them that need an update.
1925 * Since we don't want to hold any locks during the task updates,
1926 * gather tasks to be processed in a heap structure.
1927 * The heap is sorted by descending task start time.
1928 * If the statically-sized heap fills up, we overflow tasks that
1929 * started later, and in future iterations only consider tasks that
1930 * started after the latest task in the previous pass. This
1931 * guarantees forward progress and that we don't miss any tasks.
1934 cgroup_iter_start(scan->cg, &it);
1935 while ((p = cgroup_iter_next(scan->cg, &it))) {
1937 * Only affect tasks that qualify per the caller's callback,
1938 * if he provided one
1940 if (scan->test_task && !scan->test_task(p, scan))
1943 * Only process tasks that started after the last task
1946 if (!started_after_time(p, &latest_time, latest_task))
1948 dropped = heap_insert(heap, p);
1949 if (dropped == NULL) {
1951 * The new task was inserted; the heap wasn't
1955 } else if (dropped != p) {
1957 * The new task was inserted, and pushed out a
1961 put_task_struct(dropped);
1964 * Else the new task was newer than anything already in
1965 * the heap and wasn't inserted
1968 cgroup_iter_end(scan->cg, &it);
1971 for (i = 0; i < heap->size; i++) {
1972 struct task_struct *q = heap->ptrs[i];
1974 latest_time = q->start_time;
1977 /* Process the task per the caller's callback */
1978 scan->process_task(q, scan);
1982 * If we had to process any tasks at all, scan again
1983 * in case some of them were in the middle of forking
1984 * children that didn't get processed.
1985 * Not the most efficient way to do it, but it avoids
1986 * having to take callback_mutex in the fork path
1990 if (heap == &tmp_heap)
1991 heap_free(&tmp_heap);
1996 * Stuff for reading the 'tasks' file.
1998 * Reading this file can return large amounts of data if a cgroup has
1999 * *lots* of attached tasks. So it may need several calls to read(),
2000 * but we cannot guarantee that the information we produce is correct
2001 * unless we produce it entirely atomically.
2006 * Load into 'pidarray' up to 'npids' of the tasks using cgroup
2007 * 'cgrp'. Return actual number of pids loaded. No need to
2008 * task_lock(p) when reading out p->cgroup, since we're in an RCU
2009 * read section, so the css_set can't go away, and is
2010 * immutable after creation.
2012 static int pid_array_load(pid_t *pidarray, int npids, struct cgroup *cgrp)
2015 struct cgroup_iter it;
2016 struct task_struct *tsk;
2017 cgroup_iter_start(cgrp, &it);
2018 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2019 if (unlikely(n == npids))
2021 pidarray[n++] = task_pid_vnr(tsk);
2023 cgroup_iter_end(cgrp, &it);
2028 * cgroupstats_build - build and fill cgroupstats
2029 * @stats: cgroupstats to fill information into
2030 * @dentry: A dentry entry belonging to the cgroup for which stats have
2033 * Build and fill cgroupstats so that taskstats can export it to user
2036 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2039 struct cgroup *cgrp;
2040 struct cgroup_iter it;
2041 struct task_struct *tsk;
2043 * Validate dentry by checking the superblock operations
2045 if (dentry->d_sb->s_op != &cgroup_ops)
2049 cgrp = dentry->d_fsdata;
2052 cgroup_iter_start(cgrp, &it);
2053 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2054 switch (tsk->state) {
2056 stats->nr_running++;
2058 case TASK_INTERRUPTIBLE:
2059 stats->nr_sleeping++;
2061 case TASK_UNINTERRUPTIBLE:
2062 stats->nr_uninterruptible++;
2065 stats->nr_stopped++;
2068 if (delayacct_is_task_waiting_on_io(tsk))
2069 stats->nr_io_wait++;
2073 cgroup_iter_end(cgrp, &it);
2080 static int cmppid(const void *a, const void *b)
2082 return *(pid_t *)a - *(pid_t *)b;
2087 * seq_file methods for the "tasks" file. The seq_file position is the
2088 * next pid to display; the seq_file iterator is a pointer to the pid
2089 * in the cgroup->tasks_pids array.
2092 static void *cgroup_tasks_start(struct seq_file *s, loff_t *pos)
2095 * Initially we receive a position value that corresponds to
2096 * one more than the last pid shown (or 0 on the first call or
2097 * after a seek to the start). Use a binary-search to find the
2098 * next pid to display, if any
2100 struct cgroup *cgrp = s->private;
2101 int index = 0, pid = *pos;
2104 down_read(&cgrp->pids_mutex);
2106 int end = cgrp->pids_length;
2108 while (index < end) {
2109 int mid = (index + end) / 2;
2110 if (cgrp->tasks_pids[mid] == pid) {
2113 } else if (cgrp->tasks_pids[mid] <= pid)
2119 /* If we're off the end of the array, we're done */
2120 if (index >= cgrp->pids_length)
2122 /* Update the abstract position to be the actual pid that we found */
2123 iter = cgrp->tasks_pids + index;
2128 static void cgroup_tasks_stop(struct seq_file *s, void *v)
2130 struct cgroup *cgrp = s->private;
2131 up_read(&cgrp->pids_mutex);
2134 static void *cgroup_tasks_next(struct seq_file *s, void *v, loff_t *pos)
2136 struct cgroup *cgrp = s->private;
2138 int *end = cgrp->tasks_pids + cgrp->pids_length;
2141 * Advance to the next pid in the array. If this goes off the
2153 static int cgroup_tasks_show(struct seq_file *s, void *v)
2155 return seq_printf(s, "%d\n", *(int *)v);
2158 static struct seq_operations cgroup_tasks_seq_operations = {
2159 .start = cgroup_tasks_start,
2160 .stop = cgroup_tasks_stop,
2161 .next = cgroup_tasks_next,
2162 .show = cgroup_tasks_show,
2165 static void release_cgroup_pid_array(struct cgroup *cgrp)
2167 down_write(&cgrp->pids_mutex);
2168 BUG_ON(!cgrp->pids_use_count);
2169 if (!--cgrp->pids_use_count) {
2170 kfree(cgrp->tasks_pids);
2171 cgrp->tasks_pids = NULL;
2172 cgrp->pids_length = 0;
2174 up_write(&cgrp->pids_mutex);
2177 static int cgroup_tasks_release(struct inode *inode, struct file *file)
2179 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2181 if (!(file->f_mode & FMODE_READ))
2184 release_cgroup_pid_array(cgrp);
2185 return seq_release(inode, file);
2188 static struct file_operations cgroup_tasks_operations = {
2190 .llseek = seq_lseek,
2191 .write = cgroup_file_write,
2192 .release = cgroup_tasks_release,
2196 * Handle an open on 'tasks' file. Prepare an array containing the
2197 * process id's of tasks currently attached to the cgroup being opened.
2200 static int cgroup_tasks_open(struct inode *unused, struct file *file)
2202 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2207 /* Nothing to do for write-only files */
2208 if (!(file->f_mode & FMODE_READ))
2212 * If cgroup gets more users after we read count, we won't have
2213 * enough space - tough. This race is indistinguishable to the
2214 * caller from the case that the additional cgroup users didn't
2215 * show up until sometime later on.
2217 npids = cgroup_task_count(cgrp);
2218 pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL);
2221 npids = pid_array_load(pidarray, npids, cgrp);
2222 sort(pidarray, npids, sizeof(pid_t), cmppid, NULL);
2225 * Store the array in the cgroup, freeing the old
2226 * array if necessary
2228 down_write(&cgrp->pids_mutex);
2229 kfree(cgrp->tasks_pids);
2230 cgrp->tasks_pids = pidarray;
2231 cgrp->pids_length = npids;
2232 cgrp->pids_use_count++;
2233 up_write(&cgrp->pids_mutex);
2235 file->f_op = &cgroup_tasks_operations;
2237 retval = seq_open(file, &cgroup_tasks_seq_operations);
2239 release_cgroup_pid_array(cgrp);
2242 ((struct seq_file *)file->private_data)->private = cgrp;
2246 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
2249 return notify_on_release(cgrp);
2252 static int cgroup_write_notify_on_release(struct cgroup *cgrp,
2256 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
2258 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2260 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2265 * for the common functions, 'private' gives the type of file
2267 static struct cftype files[] = {
2270 .open = cgroup_tasks_open,
2271 .write_u64 = cgroup_tasks_write,
2272 .release = cgroup_tasks_release,
2273 .private = FILE_TASKLIST,
2277 .name = "notify_on_release",
2278 .read_u64 = cgroup_read_notify_on_release,
2279 .write_u64 = cgroup_write_notify_on_release,
2280 .private = FILE_NOTIFY_ON_RELEASE,
2284 static struct cftype cft_release_agent = {
2285 .name = "release_agent",
2286 .read_seq_string = cgroup_release_agent_show,
2287 .write_string = cgroup_release_agent_write,
2288 .max_write_len = PATH_MAX,
2289 .private = FILE_RELEASE_AGENT,
2292 static int cgroup_populate_dir(struct cgroup *cgrp)
2295 struct cgroup_subsys *ss;
2297 /* First clear out any existing files */
2298 cgroup_clear_directory(cgrp->dentry);
2300 err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
2304 if (cgrp == cgrp->top_cgroup) {
2305 if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
2309 for_each_subsys(cgrp->root, ss) {
2310 if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
2317 static void init_cgroup_css(struct cgroup_subsys_state *css,
2318 struct cgroup_subsys *ss,
2319 struct cgroup *cgrp)
2322 atomic_set(&css->refcnt, 0);
2324 if (cgrp == dummytop)
2325 set_bit(CSS_ROOT, &css->flags);
2326 BUG_ON(cgrp->subsys[ss->subsys_id]);
2327 cgrp->subsys[ss->subsys_id] = css;
2331 * cgroup_create - create a cgroup
2332 * @parent: cgroup that will be parent of the new cgroup
2333 * @dentry: dentry of the new cgroup
2334 * @mode: mode to set on new inode
2336 * Must be called with the mutex on the parent inode held
2338 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
2341 struct cgroup *cgrp;
2342 struct cgroupfs_root *root = parent->root;
2344 struct cgroup_subsys *ss;
2345 struct super_block *sb = root->sb;
2347 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
2351 /* Grab a reference on the superblock so the hierarchy doesn't
2352 * get deleted on unmount if there are child cgroups. This
2353 * can be done outside cgroup_mutex, since the sb can't
2354 * disappear while someone has an open control file on the
2356 atomic_inc(&sb->s_active);
2358 mutex_lock(&cgroup_mutex);
2360 init_cgroup_housekeeping(cgrp);
2362 cgrp->parent = parent;
2363 cgrp->root = parent->root;
2364 cgrp->top_cgroup = parent->top_cgroup;
2366 if (notify_on_release(parent))
2367 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2369 for_each_subsys(root, ss) {
2370 struct cgroup_subsys_state *css = ss->create(ss, cgrp);
2375 init_cgroup_css(css, ss, cgrp);
2378 list_add(&cgrp->sibling, &cgrp->parent->children);
2379 root->number_of_cgroups++;
2381 err = cgroup_create_dir(cgrp, dentry, mode);
2385 /* The cgroup directory was pre-locked for us */
2386 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
2388 err = cgroup_populate_dir(cgrp);
2389 /* If err < 0, we have a half-filled directory - oh well ;) */
2391 mutex_unlock(&cgroup_mutex);
2392 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
2398 list_del(&cgrp->sibling);
2399 root->number_of_cgroups--;
2403 for_each_subsys(root, ss) {
2404 if (cgrp->subsys[ss->subsys_id])
2405 ss->destroy(ss, cgrp);
2408 mutex_unlock(&cgroup_mutex);
2410 /* Release the reference count that we took on the superblock */
2411 deactivate_super(sb);
2417 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
2419 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
2421 /* the vfs holds inode->i_mutex already */
2422 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
2425 static int cgroup_has_css_refs(struct cgroup *cgrp)
2427 /* Check the reference count on each subsystem. Since we
2428 * already established that there are no tasks in the
2429 * cgroup, if the css refcount is also 0, then there should
2430 * be no outstanding references, so the subsystem is safe to
2431 * destroy. We scan across all subsystems rather than using
2432 * the per-hierarchy linked list of mounted subsystems since
2433 * we can be called via check_for_release() with no
2434 * synchronization other than RCU, and the subsystem linked
2435 * list isn't RCU-safe */
2437 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2438 struct cgroup_subsys *ss = subsys[i];
2439 struct cgroup_subsys_state *css;
2440 /* Skip subsystems not in this hierarchy */
2441 if (ss->root != cgrp->root)
2443 css = cgrp->subsys[ss->subsys_id];
2444 /* When called from check_for_release() it's possible
2445 * that by this point the cgroup has been removed
2446 * and the css deleted. But a false-positive doesn't
2447 * matter, since it can only happen if the cgroup
2448 * has been deleted and hence no longer needs the
2449 * release agent to be called anyway. */
2450 if (css && atomic_read(&css->refcnt))
2456 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
2458 struct cgroup *cgrp = dentry->d_fsdata;
2460 struct cgroup *parent;
2461 struct super_block *sb;
2462 struct cgroupfs_root *root;
2464 /* the vfs holds both inode->i_mutex already */
2466 mutex_lock(&cgroup_mutex);
2467 if (atomic_read(&cgrp->count) != 0) {
2468 mutex_unlock(&cgroup_mutex);
2471 if (!list_empty(&cgrp->children)) {
2472 mutex_unlock(&cgroup_mutex);
2476 parent = cgrp->parent;
2481 * Call pre_destroy handlers of subsys. Notify subsystems
2482 * that rmdir() request comes.
2484 cgroup_call_pre_destroy(cgrp);
2486 if (cgroup_has_css_refs(cgrp)) {
2487 mutex_unlock(&cgroup_mutex);
2491 spin_lock(&release_list_lock);
2492 set_bit(CGRP_REMOVED, &cgrp->flags);
2493 if (!list_empty(&cgrp->release_list))
2494 list_del(&cgrp->release_list);
2495 spin_unlock(&release_list_lock);
2496 /* delete my sibling from parent->children */
2497 list_del(&cgrp->sibling);
2498 spin_lock(&cgrp->dentry->d_lock);
2499 d = dget(cgrp->dentry);
2500 spin_unlock(&d->d_lock);
2502 cgroup_d_remove_dir(d);
2505 set_bit(CGRP_RELEASABLE, &parent->flags);
2506 check_for_release(parent);
2508 mutex_unlock(&cgroup_mutex);
2512 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
2514 struct cgroup_subsys_state *css;
2516 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
2518 /* Create the top cgroup state for this subsystem */
2519 ss->root = &rootnode;
2520 css = ss->create(ss, dummytop);
2521 /* We don't handle early failures gracefully */
2522 BUG_ON(IS_ERR(css));
2523 init_cgroup_css(css, ss, dummytop);
2525 /* Update the init_css_set to contain a subsys
2526 * pointer to this state - since the subsystem is
2527 * newly registered, all tasks and hence the
2528 * init_css_set is in the subsystem's top cgroup. */
2529 init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
2531 need_forkexit_callback |= ss->fork || ss->exit;
2532 need_mm_owner_callback |= !!ss->mm_owner_changed;
2534 /* At system boot, before all subsystems have been
2535 * registered, no tasks have been forked, so we don't
2536 * need to invoke fork callbacks here. */
2537 BUG_ON(!list_empty(&init_task.tasks));
2543 * cgroup_init_early - cgroup initialization at system boot
2545 * Initialize cgroups at system boot, and initialize any
2546 * subsystems that request early init.
2548 int __init cgroup_init_early(void)
2551 atomic_set(&init_css_set.refcount, 1);
2552 INIT_LIST_HEAD(&init_css_set.cg_links);
2553 INIT_LIST_HEAD(&init_css_set.tasks);
2554 INIT_HLIST_NODE(&init_css_set.hlist);
2556 init_cgroup_root(&rootnode);
2557 list_add(&rootnode.root_list, &roots);
2559 init_task.cgroups = &init_css_set;
2561 init_css_set_link.cg = &init_css_set;
2562 list_add(&init_css_set_link.cgrp_link_list,
2563 &rootnode.top_cgroup.css_sets);
2564 list_add(&init_css_set_link.cg_link_list,
2565 &init_css_set.cg_links);
2567 for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
2568 INIT_HLIST_HEAD(&css_set_table[i]);
2570 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2571 struct cgroup_subsys *ss = subsys[i];
2574 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
2575 BUG_ON(!ss->create);
2576 BUG_ON(!ss->destroy);
2577 if (ss->subsys_id != i) {
2578 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
2579 ss->name, ss->subsys_id);
2584 cgroup_init_subsys(ss);
2590 * cgroup_init - cgroup initialization
2592 * Register cgroup filesystem and /proc file, and initialize
2593 * any subsystems that didn't request early init.
2595 int __init cgroup_init(void)
2599 struct hlist_head *hhead;
2601 err = bdi_init(&cgroup_backing_dev_info);
2605 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2606 struct cgroup_subsys *ss = subsys[i];
2607 if (!ss->early_init)
2608 cgroup_init_subsys(ss);
2611 /* Add init_css_set to the hash table */
2612 hhead = css_set_hash(init_css_set.subsys);
2613 hlist_add_head(&init_css_set.hlist, hhead);
2615 err = register_filesystem(&cgroup_fs_type);
2619 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
2623 bdi_destroy(&cgroup_backing_dev_info);
2629 * proc_cgroup_show()
2630 * - Print task's cgroup paths into seq_file, one line for each hierarchy
2631 * - Used for /proc/<pid>/cgroup.
2632 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
2633 * doesn't really matter if tsk->cgroup changes after we read it,
2634 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
2635 * anyway. No need to check that tsk->cgroup != NULL, thanks to
2636 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
2637 * cgroup to top_cgroup.
2640 /* TODO: Use a proper seq_file iterator */
2641 static int proc_cgroup_show(struct seq_file *m, void *v)
2644 struct task_struct *tsk;
2647 struct cgroupfs_root *root;
2650 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2656 tsk = get_pid_task(pid, PIDTYPE_PID);
2662 mutex_lock(&cgroup_mutex);
2664 for_each_root(root) {
2665 struct cgroup_subsys *ss;
2666 struct cgroup *cgrp;
2670 /* Skip this hierarchy if it has no active subsystems */
2671 if (!root->actual_subsys_bits)
2673 seq_printf(m, "%lu:", root->subsys_bits);
2674 for_each_subsys(root, ss)
2675 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
2677 get_first_subsys(&root->top_cgroup, NULL, &subsys_id);
2678 cgrp = task_cgroup(tsk, subsys_id);
2679 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
2687 mutex_unlock(&cgroup_mutex);
2688 put_task_struct(tsk);
2695 static int cgroup_open(struct inode *inode, struct file *file)
2697 struct pid *pid = PROC_I(inode)->pid;
2698 return single_open(file, proc_cgroup_show, pid);
2701 struct file_operations proc_cgroup_operations = {
2702 .open = cgroup_open,
2704 .llseek = seq_lseek,
2705 .release = single_release,
2708 /* Display information about each subsystem and each hierarchy */
2709 static int proc_cgroupstats_show(struct seq_file *m, void *v)
2713 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
2714 mutex_lock(&cgroup_mutex);
2715 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2716 struct cgroup_subsys *ss = subsys[i];
2717 seq_printf(m, "%s\t%lu\t%d\t%d\n",
2718 ss->name, ss->root->subsys_bits,
2719 ss->root->number_of_cgroups, !ss->disabled);
2721 mutex_unlock(&cgroup_mutex);
2725 static int cgroupstats_open(struct inode *inode, struct file *file)
2727 return single_open(file, proc_cgroupstats_show, NULL);
2730 static struct file_operations proc_cgroupstats_operations = {
2731 .open = cgroupstats_open,
2733 .llseek = seq_lseek,
2734 .release = single_release,
2738 * cgroup_fork - attach newly forked task to its parents cgroup.
2739 * @child: pointer to task_struct of forking parent process.
2741 * Description: A task inherits its parent's cgroup at fork().
2743 * A pointer to the shared css_set was automatically copied in
2744 * fork.c by dup_task_struct(). However, we ignore that copy, since
2745 * it was not made under the protection of RCU or cgroup_mutex, so
2746 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
2747 * have already changed current->cgroups, allowing the previously
2748 * referenced cgroup group to be removed and freed.
2750 * At the point that cgroup_fork() is called, 'current' is the parent
2751 * task, and the passed argument 'child' points to the child task.
2753 void cgroup_fork(struct task_struct *child)
2756 child->cgroups = current->cgroups;
2757 get_css_set(child->cgroups);
2758 task_unlock(current);
2759 INIT_LIST_HEAD(&child->cg_list);
2763 * cgroup_fork_callbacks - run fork callbacks
2764 * @child: the new task
2766 * Called on a new task very soon before adding it to the
2767 * tasklist. No need to take any locks since no-one can
2768 * be operating on this task.
2770 void cgroup_fork_callbacks(struct task_struct *child)
2772 if (need_forkexit_callback) {
2774 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2775 struct cgroup_subsys *ss = subsys[i];
2777 ss->fork(ss, child);
2782 #ifdef CONFIG_MM_OWNER
2784 * cgroup_mm_owner_callbacks - run callbacks when the mm->owner changes
2787 * Called on every change to mm->owner. mm_init_owner() does not
2788 * invoke this routine, since it assigns the mm->owner the first time
2789 * and does not change it.
2791 * The callbacks are invoked with mmap_sem held in read mode.
2793 void cgroup_mm_owner_callbacks(struct task_struct *old, struct task_struct *new)
2795 struct cgroup *oldcgrp, *newcgrp = NULL;
2797 if (need_mm_owner_callback) {
2799 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2800 struct cgroup_subsys *ss = subsys[i];
2801 oldcgrp = task_cgroup(old, ss->subsys_id);
2803 newcgrp = task_cgroup(new, ss->subsys_id);
2804 if (oldcgrp == newcgrp)
2806 if (ss->mm_owner_changed)
2807 ss->mm_owner_changed(ss, oldcgrp, newcgrp, new);
2811 #endif /* CONFIG_MM_OWNER */
2814 * cgroup_post_fork - called on a new task after adding it to the task list
2815 * @child: the task in question
2817 * Adds the task to the list running through its css_set if necessary.
2818 * Has to be after the task is visible on the task list in case we race
2819 * with the first call to cgroup_iter_start() - to guarantee that the
2820 * new task ends up on its list.
2822 void cgroup_post_fork(struct task_struct *child)
2824 if (use_task_css_set_links) {
2825 write_lock(&css_set_lock);
2826 if (list_empty(&child->cg_list))
2827 list_add(&child->cg_list, &child->cgroups->tasks);
2828 write_unlock(&css_set_lock);
2832 * cgroup_exit - detach cgroup from exiting task
2833 * @tsk: pointer to task_struct of exiting process
2834 * @run_callback: run exit callbacks?
2836 * Description: Detach cgroup from @tsk and release it.
2838 * Note that cgroups marked notify_on_release force every task in
2839 * them to take the global cgroup_mutex mutex when exiting.
2840 * This could impact scaling on very large systems. Be reluctant to
2841 * use notify_on_release cgroups where very high task exit scaling
2842 * is required on large systems.
2844 * the_top_cgroup_hack:
2846 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
2848 * We call cgroup_exit() while the task is still competent to
2849 * handle notify_on_release(), then leave the task attached to the
2850 * root cgroup in each hierarchy for the remainder of its exit.
2852 * To do this properly, we would increment the reference count on
2853 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
2854 * code we would add a second cgroup function call, to drop that
2855 * reference. This would just create an unnecessary hot spot on
2856 * the top_cgroup reference count, to no avail.
2858 * Normally, holding a reference to a cgroup without bumping its
2859 * count is unsafe. The cgroup could go away, or someone could
2860 * attach us to a different cgroup, decrementing the count on
2861 * the first cgroup that we never incremented. But in this case,
2862 * top_cgroup isn't going away, and either task has PF_EXITING set,
2863 * which wards off any cgroup_attach_task() attempts, or task is a failed
2864 * fork, never visible to cgroup_attach_task.
2866 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
2871 if (run_callbacks && need_forkexit_callback) {
2872 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2873 struct cgroup_subsys *ss = subsys[i];
2880 * Unlink from the css_set task list if necessary.
2881 * Optimistically check cg_list before taking
2884 if (!list_empty(&tsk->cg_list)) {
2885 write_lock(&css_set_lock);
2886 if (!list_empty(&tsk->cg_list))
2887 list_del(&tsk->cg_list);
2888 write_unlock(&css_set_lock);
2891 /* Reassign the task to the init_css_set. */
2894 tsk->cgroups = &init_css_set;
2897 put_css_set_taskexit(cg);
2901 * cgroup_clone - clone the cgroup the given subsystem is attached to
2902 * @tsk: the task to be moved
2903 * @subsys: the given subsystem
2904 * @nodename: the name for the new cgroup
2906 * Duplicate the current cgroup in the hierarchy that the given
2907 * subsystem is attached to, and move this task into the new
2910 int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
2913 struct dentry *dentry;
2915 struct cgroup *parent, *child;
2916 struct inode *inode;
2918 struct cgroupfs_root *root;
2919 struct cgroup_subsys *ss;
2921 /* We shouldn't be called by an unregistered subsystem */
2922 BUG_ON(!subsys->active);
2924 /* First figure out what hierarchy and cgroup we're dealing
2925 * with, and pin them so we can drop cgroup_mutex */
2926 mutex_lock(&cgroup_mutex);
2928 root = subsys->root;
2929 if (root == &rootnode) {
2931 "Not cloning cgroup for unused subsystem %s\n",
2933 mutex_unlock(&cgroup_mutex);
2937 parent = task_cgroup(tsk, subsys->subsys_id);
2939 /* Pin the hierarchy */
2940 atomic_inc(&parent->root->sb->s_active);
2942 /* Keep the cgroup alive */
2944 mutex_unlock(&cgroup_mutex);
2946 /* Now do the VFS work to create a cgroup */
2947 inode = parent->dentry->d_inode;
2949 /* Hold the parent directory mutex across this operation to
2950 * stop anyone else deleting the new cgroup */
2951 mutex_lock(&inode->i_mutex);
2952 dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
2953 if (IS_ERR(dentry)) {
2955 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
2957 ret = PTR_ERR(dentry);
2961 /* Create the cgroup directory, which also creates the cgroup */
2962 ret = vfs_mkdir(inode, dentry, S_IFDIR | 0755);
2963 child = __d_cgrp(dentry);
2967 "Failed to create cgroup %s: %d\n", nodename,
2974 "Couldn't find new cgroup %s\n", nodename);
2979 /* The cgroup now exists. Retake cgroup_mutex and check
2980 * that we're still in the same state that we thought we
2982 mutex_lock(&cgroup_mutex);
2983 if ((root != subsys->root) ||
2984 (parent != task_cgroup(tsk, subsys->subsys_id))) {
2985 /* Aargh, we raced ... */
2986 mutex_unlock(&inode->i_mutex);
2989 deactivate_super(parent->root->sb);
2990 /* The cgroup is still accessible in the VFS, but
2991 * we're not going to try to rmdir() it at this
2994 "Race in cgroup_clone() - leaking cgroup %s\n",
2999 /* do any required auto-setup */
3000 for_each_subsys(root, ss) {
3002 ss->post_clone(ss, child);
3005 /* All seems fine. Finish by moving the task into the new cgroup */
3006 ret = cgroup_attach_task(child, tsk);
3007 mutex_unlock(&cgroup_mutex);
3010 mutex_unlock(&inode->i_mutex);
3012 mutex_lock(&cgroup_mutex);
3014 mutex_unlock(&cgroup_mutex);
3015 deactivate_super(parent->root->sb);
3020 * cgroup_is_descendant - see if @cgrp is a descendant of current task's cgrp
3021 * @cgrp: the cgroup in question
3023 * See if @cgrp is a descendant of the current task's cgroup in
3024 * the appropriate hierarchy.
3026 * If we are sending in dummytop, then presumably we are creating
3027 * the top cgroup in the subsystem.
3029 * Called only by the ns (nsproxy) cgroup.
3031 int cgroup_is_descendant(const struct cgroup *cgrp)
3034 struct cgroup *target;
3037 if (cgrp == dummytop)
3040 get_first_subsys(cgrp, NULL, &subsys_id);
3041 target = task_cgroup(current, subsys_id);
3042 while (cgrp != target && cgrp!= cgrp->top_cgroup)
3043 cgrp = cgrp->parent;
3044 ret = (cgrp == target);
3048 static void check_for_release(struct cgroup *cgrp)
3050 /* All of these checks rely on RCU to keep the cgroup
3051 * structure alive */
3052 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
3053 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
3054 /* Control Group is currently removeable. If it's not
3055 * already queued for a userspace notification, queue
3057 int need_schedule_work = 0;
3058 spin_lock(&release_list_lock);
3059 if (!cgroup_is_removed(cgrp) &&
3060 list_empty(&cgrp->release_list)) {
3061 list_add(&cgrp->release_list, &release_list);
3062 need_schedule_work = 1;
3064 spin_unlock(&release_list_lock);
3065 if (need_schedule_work)
3066 schedule_work(&release_agent_work);
3070 void __css_put(struct cgroup_subsys_state *css)
3072 struct cgroup *cgrp = css->cgroup;
3074 if (atomic_dec_and_test(&css->refcnt) && notify_on_release(cgrp)) {
3075 set_bit(CGRP_RELEASABLE, &cgrp->flags);
3076 check_for_release(cgrp);
3082 * Notify userspace when a cgroup is released, by running the
3083 * configured release agent with the name of the cgroup (path
3084 * relative to the root of cgroup file system) as the argument.
3086 * Most likely, this user command will try to rmdir this cgroup.
3088 * This races with the possibility that some other task will be
3089 * attached to this cgroup before it is removed, or that some other
3090 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
3091 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
3092 * unused, and this cgroup will be reprieved from its death sentence,
3093 * to continue to serve a useful existence. Next time it's released,
3094 * we will get notified again, if it still has 'notify_on_release' set.
3096 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
3097 * means only wait until the task is successfully execve()'d. The
3098 * separate release agent task is forked by call_usermodehelper(),
3099 * then control in this thread returns here, without waiting for the
3100 * release agent task. We don't bother to wait because the caller of
3101 * this routine has no use for the exit status of the release agent
3102 * task, so no sense holding our caller up for that.
3104 static void cgroup_release_agent(struct work_struct *work)
3106 BUG_ON(work != &release_agent_work);
3107 mutex_lock(&cgroup_mutex);
3108 spin_lock(&release_list_lock);
3109 while (!list_empty(&release_list)) {
3110 char *argv[3], *envp[3];
3112 char *pathbuf = NULL, *agentbuf = NULL;
3113 struct cgroup *cgrp = list_entry(release_list.next,
3116 list_del_init(&cgrp->release_list);
3117 spin_unlock(&release_list_lock);
3118 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
3121 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
3123 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
3128 argv[i++] = agentbuf;
3129 argv[i++] = pathbuf;
3133 /* minimal command environment */
3134 envp[i++] = "HOME=/";
3135 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
3138 /* Drop the lock while we invoke the usermode helper,
3139 * since the exec could involve hitting disk and hence
3140 * be a slow process */
3141 mutex_unlock(&cgroup_mutex);
3142 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
3143 mutex_lock(&cgroup_mutex);
3147 spin_lock(&release_list_lock);
3149 spin_unlock(&release_list_lock);
3150 mutex_unlock(&cgroup_mutex);
3153 static int __init cgroup_disable(char *str)
3158 while ((token = strsep(&str, ",")) != NULL) {
3162 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3163 struct cgroup_subsys *ss = subsys[i];
3165 if (!strcmp(token, ss->name)) {
3167 printk(KERN_INFO "Disabling %s control group"
3168 " subsystem\n", ss->name);
3175 __setup("cgroup_disable=", cgroup_disable);