iwlagn: move no_sleep_autoadjust as part of iwlagn_mod_params
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / cgroup.c
blob909a35510af53656c684021b113977c10bfbbac8
1 /*
2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #include <linux/cgroup.h>
30 #include <linux/ctype.h>
31 #include <linux/errno.h>
32 #include <linux/fs.h>
33 #include <linux/kernel.h>
34 #include <linux/list.h>
35 #include <linux/mm.h>
36 #include <linux/mutex.h>
37 #include <linux/mount.h>
38 #include <linux/pagemap.h>
39 #include <linux/proc_fs.h>
40 #include <linux/rcupdate.h>
41 #include <linux/sched.h>
42 #include <linux/backing-dev.h>
43 #include <linux/seq_file.h>
44 #include <linux/slab.h>
45 #include <linux/magic.h>
46 #include <linux/spinlock.h>
47 #include <linux/string.h>
48 #include <linux/sort.h>
49 #include <linux/kmod.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/cgroupstats.h>
53 #include <linux/hash.h>
54 #include <linux/namei.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/eventfd.h>
59 #include <linux/poll.h>
61 #include <asm/atomic.h>
63 static DEFINE_MUTEX(cgroup_mutex);
66 * Generate an array of cgroup subsystem pointers. At boot time, this is
67 * populated up to CGROUP_BUILTIN_SUBSYS_COUNT, and modular subsystems are
68 * registered after that. The mutable section of this array is protected by
69 * cgroup_mutex.
71 #define SUBSYS(_x) &_x ## _subsys,
72 static struct cgroup_subsys *subsys[CGROUP_SUBSYS_COUNT] = {
73 #include <linux/cgroup_subsys.h>
76 #define MAX_CGROUP_ROOT_NAMELEN 64
79 * A cgroupfs_root represents the root of a cgroup hierarchy,
80 * and may be associated with a superblock to form an active
81 * hierarchy
83 struct cgroupfs_root {
84 struct super_block *sb;
87 * The bitmask of subsystems intended to be attached to this
88 * hierarchy
90 unsigned long subsys_bits;
92 /* Unique id for this hierarchy. */
93 int hierarchy_id;
95 /* The bitmask of subsystems currently attached to this hierarchy */
96 unsigned long actual_subsys_bits;
98 /* A list running through the attached subsystems */
99 struct list_head subsys_list;
101 /* The root cgroup for this hierarchy */
102 struct cgroup top_cgroup;
104 /* Tracks how many cgroups are currently defined in hierarchy.*/
105 int number_of_cgroups;
107 /* A list running through the active hierarchies */
108 struct list_head root_list;
110 /* Hierarchy-specific flags */
111 unsigned long flags;
113 /* The path to use for release notifications. */
114 char release_agent_path[PATH_MAX];
116 /* The name for this hierarchy - may be empty */
117 char name[MAX_CGROUP_ROOT_NAMELEN];
121 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
122 * subsystems that are otherwise unattached - it never has more than a
123 * single cgroup, and all tasks are part of that cgroup.
125 static struct cgroupfs_root rootnode;
128 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
129 * cgroup_subsys->use_id != 0.
131 #define CSS_ID_MAX (65535)
132 struct css_id {
134 * The css to which this ID points. This pointer is set to valid value
135 * after cgroup is populated. If cgroup is removed, this will be NULL.
136 * This pointer is expected to be RCU-safe because destroy()
137 * is called after synchronize_rcu(). But for safe use, css_is_removed()
138 * css_tryget() should be used for avoiding race.
140 struct cgroup_subsys_state __rcu *css;
142 * ID of this css.
144 unsigned short id;
146 * Depth in hierarchy which this ID belongs to.
148 unsigned short depth;
150 * ID is freed by RCU. (and lookup routine is RCU safe.)
152 struct rcu_head rcu_head;
154 * Hierarchy of CSS ID belongs to.
156 unsigned short stack[0]; /* Array of Length (depth+1) */
160 * cgroup_event represents events which userspace want to receive.
162 struct cgroup_event {
164 * Cgroup which the event belongs to.
166 struct cgroup *cgrp;
168 * Control file which the event associated.
170 struct cftype *cft;
172 * eventfd to signal userspace about the event.
174 struct eventfd_ctx *eventfd;
176 * Each of these stored in a list by the cgroup.
178 struct list_head list;
180 * All fields below needed to unregister event when
181 * userspace closes eventfd.
183 poll_table pt;
184 wait_queue_head_t *wqh;
185 wait_queue_t wait;
186 struct work_struct remove;
189 /* The list of hierarchy roots */
191 static LIST_HEAD(roots);
192 static int root_count;
194 static DEFINE_IDA(hierarchy_ida);
195 static int next_hierarchy_id;
196 static DEFINE_SPINLOCK(hierarchy_id_lock);
198 /* dummytop is a shorthand for the dummy hierarchy's top cgroup */
199 #define dummytop (&rootnode.top_cgroup)
201 /* This flag indicates whether tasks in the fork and exit paths should
202 * check for fork/exit handlers to call. This avoids us having to do
203 * extra work in the fork/exit path if none of the subsystems need to
204 * be called.
206 static int need_forkexit_callback __read_mostly;
208 #ifdef CONFIG_PROVE_LOCKING
209 int cgroup_lock_is_held(void)
211 return lockdep_is_held(&cgroup_mutex);
213 #else /* #ifdef CONFIG_PROVE_LOCKING */
214 int cgroup_lock_is_held(void)
216 return mutex_is_locked(&cgroup_mutex);
218 #endif /* #else #ifdef CONFIG_PROVE_LOCKING */
220 EXPORT_SYMBOL_GPL(cgroup_lock_is_held);
222 /* convenient tests for these bits */
223 inline int cgroup_is_removed(const struct cgroup *cgrp)
225 return test_bit(CGRP_REMOVED, &cgrp->flags);
228 /* bits in struct cgroupfs_root flags field */
229 enum {
230 ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
233 static int cgroup_is_releasable(const struct cgroup *cgrp)
235 const int bits =
236 (1 << CGRP_RELEASABLE) |
237 (1 << CGRP_NOTIFY_ON_RELEASE);
238 return (cgrp->flags & bits) == bits;
241 static int notify_on_release(const struct cgroup *cgrp)
243 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
246 static int clone_children(const struct cgroup *cgrp)
248 return test_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
252 * for_each_subsys() allows you to iterate on each subsystem attached to
253 * an active hierarchy
255 #define for_each_subsys(_root, _ss) \
256 list_for_each_entry(_ss, &_root->subsys_list, sibling)
258 /* for_each_active_root() allows you to iterate across the active hierarchies */
259 #define for_each_active_root(_root) \
260 list_for_each_entry(_root, &roots, root_list)
262 /* the list of cgroups eligible for automatic release. Protected by
263 * release_list_lock */
264 static LIST_HEAD(release_list);
265 static DEFINE_SPINLOCK(release_list_lock);
266 static void cgroup_release_agent(struct work_struct *work);
267 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
268 static void check_for_release(struct cgroup *cgrp);
270 /* Link structure for associating css_set objects with cgroups */
271 struct cg_cgroup_link {
273 * List running through cg_cgroup_links associated with a
274 * cgroup, anchored on cgroup->css_sets
276 struct list_head cgrp_link_list;
277 struct cgroup *cgrp;
279 * List running through cg_cgroup_links pointing at a
280 * single css_set object, anchored on css_set->cg_links
282 struct list_head cg_link_list;
283 struct css_set *cg;
286 /* The default css_set - used by init and its children prior to any
287 * hierarchies being mounted. It contains a pointer to the root state
288 * for each subsystem. Also used to anchor the list of css_sets. Not
289 * reference-counted, to improve performance when child cgroups
290 * haven't been created.
293 static struct css_set init_css_set;
294 static struct cg_cgroup_link init_css_set_link;
296 static int cgroup_init_idr(struct cgroup_subsys *ss,
297 struct cgroup_subsys_state *css);
299 /* css_set_lock protects the list of css_set objects, and the
300 * chain of tasks off each css_set. Nests outside task->alloc_lock
301 * due to cgroup_iter_start() */
302 static DEFINE_RWLOCK(css_set_lock);
303 static int css_set_count;
306 * hash table for cgroup groups. This improves the performance to find
307 * an existing css_set. This hash doesn't (currently) take into
308 * account cgroups in empty hierarchies.
310 #define CSS_SET_HASH_BITS 7
311 #define CSS_SET_TABLE_SIZE (1 << CSS_SET_HASH_BITS)
312 static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
314 static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
316 int i;
317 int index;
318 unsigned long tmp = 0UL;
320 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
321 tmp += (unsigned long)css[i];
322 tmp = (tmp >> 16) ^ tmp;
324 index = hash_long(tmp, CSS_SET_HASH_BITS);
326 return &css_set_table[index];
329 /* We don't maintain the lists running through each css_set to its
330 * task until after the first call to cgroup_iter_start(). This
331 * reduces the fork()/exit() overhead for people who have cgroups
332 * compiled into their kernel but not actually in use */
333 static int use_task_css_set_links __read_mostly;
335 static void __put_css_set(struct css_set *cg, int taskexit)
337 struct cg_cgroup_link *link;
338 struct cg_cgroup_link *saved_link;
340 * Ensure that the refcount doesn't hit zero while any readers
341 * can see it. Similar to atomic_dec_and_lock(), but for an
342 * rwlock
344 if (atomic_add_unless(&cg->refcount, -1, 1))
345 return;
346 write_lock(&css_set_lock);
347 if (!atomic_dec_and_test(&cg->refcount)) {
348 write_unlock(&css_set_lock);
349 return;
352 /* This css_set is dead. unlink it and release cgroup refcounts */
353 hlist_del(&cg->hlist);
354 css_set_count--;
356 list_for_each_entry_safe(link, saved_link, &cg->cg_links,
357 cg_link_list) {
358 struct cgroup *cgrp = link->cgrp;
359 list_del(&link->cg_link_list);
360 list_del(&link->cgrp_link_list);
361 if (atomic_dec_and_test(&cgrp->count) &&
362 notify_on_release(cgrp)) {
363 if (taskexit)
364 set_bit(CGRP_RELEASABLE, &cgrp->flags);
365 check_for_release(cgrp);
368 kfree(link);
371 write_unlock(&css_set_lock);
372 kfree_rcu(cg, rcu_head);
376 * refcounted get/put for css_set objects
378 static inline void get_css_set(struct css_set *cg)
380 atomic_inc(&cg->refcount);
383 static inline void put_css_set(struct css_set *cg)
385 __put_css_set(cg, 0);
388 static inline void put_css_set_taskexit(struct css_set *cg)
390 __put_css_set(cg, 1);
394 * compare_css_sets - helper function for find_existing_css_set().
395 * @cg: candidate css_set being tested
396 * @old_cg: existing css_set for a task
397 * @new_cgrp: cgroup that's being entered by the task
398 * @template: desired set of css pointers in css_set (pre-calculated)
400 * Returns true if "cg" matches "old_cg" except for the hierarchy
401 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
403 static bool compare_css_sets(struct css_set *cg,
404 struct css_set *old_cg,
405 struct cgroup *new_cgrp,
406 struct cgroup_subsys_state *template[])
408 struct list_head *l1, *l2;
410 if (memcmp(template, cg->subsys, sizeof(cg->subsys))) {
411 /* Not all subsystems matched */
412 return false;
416 * Compare cgroup pointers in order to distinguish between
417 * different cgroups in heirarchies with no subsystems. We
418 * could get by with just this check alone (and skip the
419 * memcmp above) but on most setups the memcmp check will
420 * avoid the need for this more expensive check on almost all
421 * candidates.
424 l1 = &cg->cg_links;
425 l2 = &old_cg->cg_links;
426 while (1) {
427 struct cg_cgroup_link *cgl1, *cgl2;
428 struct cgroup *cg1, *cg2;
430 l1 = l1->next;
431 l2 = l2->next;
432 /* See if we reached the end - both lists are equal length. */
433 if (l1 == &cg->cg_links) {
434 BUG_ON(l2 != &old_cg->cg_links);
435 break;
436 } else {
437 BUG_ON(l2 == &old_cg->cg_links);
439 /* Locate the cgroups associated with these links. */
440 cgl1 = list_entry(l1, struct cg_cgroup_link, cg_link_list);
441 cgl2 = list_entry(l2, struct cg_cgroup_link, cg_link_list);
442 cg1 = cgl1->cgrp;
443 cg2 = cgl2->cgrp;
444 /* Hierarchies should be linked in the same order. */
445 BUG_ON(cg1->root != cg2->root);
448 * If this hierarchy is the hierarchy of the cgroup
449 * that's changing, then we need to check that this
450 * css_set points to the new cgroup; if it's any other
451 * hierarchy, then this css_set should point to the
452 * same cgroup as the old css_set.
454 if (cg1->root == new_cgrp->root) {
455 if (cg1 != new_cgrp)
456 return false;
457 } else {
458 if (cg1 != cg2)
459 return false;
462 return true;
466 * find_existing_css_set() is a helper for
467 * find_css_set(), and checks to see whether an existing
468 * css_set is suitable.
470 * oldcg: the cgroup group that we're using before the cgroup
471 * transition
473 * cgrp: the cgroup that we're moving into
475 * template: location in which to build the desired set of subsystem
476 * state objects for the new cgroup group
478 static struct css_set *find_existing_css_set(
479 struct css_set *oldcg,
480 struct cgroup *cgrp,
481 struct cgroup_subsys_state *template[])
483 int i;
484 struct cgroupfs_root *root = cgrp->root;
485 struct hlist_head *hhead;
486 struct hlist_node *node;
487 struct css_set *cg;
490 * Build the set of subsystem state objects that we want to see in the
491 * new css_set. while subsystems can change globally, the entries here
492 * won't change, so no need for locking.
494 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
495 if (root->subsys_bits & (1UL << i)) {
496 /* Subsystem is in this hierarchy. So we want
497 * the subsystem state from the new
498 * cgroup */
499 template[i] = cgrp->subsys[i];
500 } else {
501 /* Subsystem is not in this hierarchy, so we
502 * don't want to change the subsystem state */
503 template[i] = oldcg->subsys[i];
507 hhead = css_set_hash(template);
508 hlist_for_each_entry(cg, node, hhead, hlist) {
509 if (!compare_css_sets(cg, oldcg, cgrp, template))
510 continue;
512 /* This css_set matches what we need */
513 return cg;
516 /* No existing cgroup group matched */
517 return NULL;
520 static void free_cg_links(struct list_head *tmp)
522 struct cg_cgroup_link *link;
523 struct cg_cgroup_link *saved_link;
525 list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
526 list_del(&link->cgrp_link_list);
527 kfree(link);
532 * allocate_cg_links() allocates "count" cg_cgroup_link structures
533 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
534 * success or a negative error
536 static int allocate_cg_links(int count, struct list_head *tmp)
538 struct cg_cgroup_link *link;
539 int i;
540 INIT_LIST_HEAD(tmp);
541 for (i = 0; i < count; i++) {
542 link = kmalloc(sizeof(*link), GFP_KERNEL);
543 if (!link) {
544 free_cg_links(tmp);
545 return -ENOMEM;
547 list_add(&link->cgrp_link_list, tmp);
549 return 0;
553 * link_css_set - a helper function to link a css_set to a cgroup
554 * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
555 * @cg: the css_set to be linked
556 * @cgrp: the destination cgroup
558 static void link_css_set(struct list_head *tmp_cg_links,
559 struct css_set *cg, struct cgroup *cgrp)
561 struct cg_cgroup_link *link;
563 BUG_ON(list_empty(tmp_cg_links));
564 link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
565 cgrp_link_list);
566 link->cg = cg;
567 link->cgrp = cgrp;
568 atomic_inc(&cgrp->count);
569 list_move(&link->cgrp_link_list, &cgrp->css_sets);
571 * Always add links to the tail of the list so that the list
572 * is sorted by order of hierarchy creation
574 list_add_tail(&link->cg_link_list, &cg->cg_links);
578 * find_css_set() takes an existing cgroup group and a
579 * cgroup object, and returns a css_set object that's
580 * equivalent to the old group, but with the given cgroup
581 * substituted into the appropriate hierarchy. Must be called with
582 * cgroup_mutex held
584 static struct css_set *find_css_set(
585 struct css_set *oldcg, struct cgroup *cgrp)
587 struct css_set *res;
588 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
590 struct list_head tmp_cg_links;
592 struct hlist_head *hhead;
593 struct cg_cgroup_link *link;
595 /* First see if we already have a cgroup group that matches
596 * the desired set */
597 read_lock(&css_set_lock);
598 res = find_existing_css_set(oldcg, cgrp, template);
599 if (res)
600 get_css_set(res);
601 read_unlock(&css_set_lock);
603 if (res)
604 return res;
606 res = kmalloc(sizeof(*res), GFP_KERNEL);
607 if (!res)
608 return NULL;
610 /* Allocate all the cg_cgroup_link objects that we'll need */
611 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
612 kfree(res);
613 return NULL;
616 atomic_set(&res->refcount, 1);
617 INIT_LIST_HEAD(&res->cg_links);
618 INIT_LIST_HEAD(&res->tasks);
619 INIT_HLIST_NODE(&res->hlist);
621 /* Copy the set of subsystem state objects generated in
622 * find_existing_css_set() */
623 memcpy(res->subsys, template, sizeof(res->subsys));
625 write_lock(&css_set_lock);
626 /* Add reference counts and links from the new css_set. */
627 list_for_each_entry(link, &oldcg->cg_links, cg_link_list) {
628 struct cgroup *c = link->cgrp;
629 if (c->root == cgrp->root)
630 c = cgrp;
631 link_css_set(&tmp_cg_links, res, c);
634 BUG_ON(!list_empty(&tmp_cg_links));
636 css_set_count++;
638 /* Add this cgroup group to the hash table */
639 hhead = css_set_hash(res->subsys);
640 hlist_add_head(&res->hlist, hhead);
642 write_unlock(&css_set_lock);
644 return res;
648 * Return the cgroup for "task" from the given hierarchy. Must be
649 * called with cgroup_mutex held.
651 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
652 struct cgroupfs_root *root)
654 struct css_set *css;
655 struct cgroup *res = NULL;
657 BUG_ON(!mutex_is_locked(&cgroup_mutex));
658 read_lock(&css_set_lock);
660 * No need to lock the task - since we hold cgroup_mutex the
661 * task can't change groups, so the only thing that can happen
662 * is that it exits and its css is set back to init_css_set.
664 css = task->cgroups;
665 if (css == &init_css_set) {
666 res = &root->top_cgroup;
667 } else {
668 struct cg_cgroup_link *link;
669 list_for_each_entry(link, &css->cg_links, cg_link_list) {
670 struct cgroup *c = link->cgrp;
671 if (c->root == root) {
672 res = c;
673 break;
677 read_unlock(&css_set_lock);
678 BUG_ON(!res);
679 return res;
683 * There is one global cgroup mutex. We also require taking
684 * task_lock() when dereferencing a task's cgroup subsys pointers.
685 * See "The task_lock() exception", at the end of this comment.
687 * A task must hold cgroup_mutex to modify cgroups.
689 * Any task can increment and decrement the count field without lock.
690 * So in general, code holding cgroup_mutex can't rely on the count
691 * field not changing. However, if the count goes to zero, then only
692 * cgroup_attach_task() can increment it again. Because a count of zero
693 * means that no tasks are currently attached, therefore there is no
694 * way a task attached to that cgroup can fork (the other way to
695 * increment the count). So code holding cgroup_mutex can safely
696 * assume that if the count is zero, it will stay zero. Similarly, if
697 * a task holds cgroup_mutex on a cgroup with zero count, it
698 * knows that the cgroup won't be removed, as cgroup_rmdir()
699 * needs that mutex.
701 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
702 * (usually) take cgroup_mutex. These are the two most performance
703 * critical pieces of code here. The exception occurs on cgroup_exit(),
704 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
705 * is taken, and if the cgroup count is zero, a usermode call made
706 * to the release agent with the name of the cgroup (path relative to
707 * the root of cgroup file system) as the argument.
709 * A cgroup can only be deleted if both its 'count' of using tasks
710 * is zero, and its list of 'children' cgroups is empty. Since all
711 * tasks in the system use _some_ cgroup, and since there is always at
712 * least one task in the system (init, pid == 1), therefore, top_cgroup
713 * always has either children cgroups and/or using tasks. So we don't
714 * need a special hack to ensure that top_cgroup cannot be deleted.
716 * The task_lock() exception
718 * The need for this exception arises from the action of
719 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
720 * another. It does so using cgroup_mutex, however there are
721 * several performance critical places that need to reference
722 * task->cgroup without the expense of grabbing a system global
723 * mutex. Therefore except as noted below, when dereferencing or, as
724 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
725 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
726 * the task_struct routinely used for such matters.
728 * P.S. One more locking exception. RCU is used to guard the
729 * update of a tasks cgroup pointer by cgroup_attach_task()
733 * cgroup_lock - lock out any changes to cgroup structures
736 void cgroup_lock(void)
738 mutex_lock(&cgroup_mutex);
740 EXPORT_SYMBOL_GPL(cgroup_lock);
743 * cgroup_unlock - release lock on cgroup changes
745 * Undo the lock taken in a previous cgroup_lock() call.
747 void cgroup_unlock(void)
749 mutex_unlock(&cgroup_mutex);
751 EXPORT_SYMBOL_GPL(cgroup_unlock);
754 * A couple of forward declarations required, due to cyclic reference loop:
755 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
756 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
757 * -> cgroup_mkdir.
760 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
761 static struct dentry *cgroup_lookup(struct inode *, struct dentry *, struct nameidata *);
762 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
763 static int cgroup_populate_dir(struct cgroup *cgrp);
764 static const struct inode_operations cgroup_dir_inode_operations;
765 static const struct file_operations proc_cgroupstats_operations;
767 static struct backing_dev_info cgroup_backing_dev_info = {
768 .name = "cgroup",
769 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
772 static int alloc_css_id(struct cgroup_subsys *ss,
773 struct cgroup *parent, struct cgroup *child);
775 static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
777 struct inode *inode = new_inode(sb);
779 if (inode) {
780 inode->i_ino = get_next_ino();
781 inode->i_mode = mode;
782 inode->i_uid = current_fsuid();
783 inode->i_gid = current_fsgid();
784 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
785 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
787 return inode;
791 * Call subsys's pre_destroy handler.
792 * This is called before css refcnt check.
794 static int cgroup_call_pre_destroy(struct cgroup *cgrp)
796 struct cgroup_subsys *ss;
797 int ret = 0;
799 for_each_subsys(cgrp->root, ss)
800 if (ss->pre_destroy) {
801 ret = ss->pre_destroy(ss, cgrp);
802 if (ret)
803 break;
806 return ret;
809 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
811 /* is dentry a directory ? if so, kfree() associated cgroup */
812 if (S_ISDIR(inode->i_mode)) {
813 struct cgroup *cgrp = dentry->d_fsdata;
814 struct cgroup_subsys *ss;
815 BUG_ON(!(cgroup_is_removed(cgrp)));
816 /* It's possible for external users to be holding css
817 * reference counts on a cgroup; css_put() needs to
818 * be able to access the cgroup after decrementing
819 * the reference count in order to know if it needs to
820 * queue the cgroup to be handled by the release
821 * agent */
822 synchronize_rcu();
824 mutex_lock(&cgroup_mutex);
826 * Release the subsystem state objects.
828 for_each_subsys(cgrp->root, ss)
829 ss->destroy(ss, cgrp);
831 cgrp->root->number_of_cgroups--;
832 mutex_unlock(&cgroup_mutex);
835 * Drop the active superblock reference that we took when we
836 * created the cgroup
838 deactivate_super(cgrp->root->sb);
841 * if we're getting rid of the cgroup, refcount should ensure
842 * that there are no pidlists left.
844 BUG_ON(!list_empty(&cgrp->pidlists));
846 kfree_rcu(cgrp, rcu_head);
848 iput(inode);
851 static int cgroup_delete(const struct dentry *d)
853 return 1;
856 static void remove_dir(struct dentry *d)
858 struct dentry *parent = dget(d->d_parent);
860 d_delete(d);
861 simple_rmdir(parent->d_inode, d);
862 dput(parent);
865 static void cgroup_clear_directory(struct dentry *dentry)
867 struct list_head *node;
869 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
870 spin_lock(&dentry->d_lock);
871 node = dentry->d_subdirs.next;
872 while (node != &dentry->d_subdirs) {
873 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
875 spin_lock_nested(&d->d_lock, DENTRY_D_LOCK_NESTED);
876 list_del_init(node);
877 if (d->d_inode) {
878 /* This should never be called on a cgroup
879 * directory with child cgroups */
880 BUG_ON(d->d_inode->i_mode & S_IFDIR);
881 dget_dlock(d);
882 spin_unlock(&d->d_lock);
883 spin_unlock(&dentry->d_lock);
884 d_delete(d);
885 simple_unlink(dentry->d_inode, d);
886 dput(d);
887 spin_lock(&dentry->d_lock);
888 } else
889 spin_unlock(&d->d_lock);
890 node = dentry->d_subdirs.next;
892 spin_unlock(&dentry->d_lock);
896 * NOTE : the dentry must have been dget()'ed
898 static void cgroup_d_remove_dir(struct dentry *dentry)
900 struct dentry *parent;
902 cgroup_clear_directory(dentry);
904 parent = dentry->d_parent;
905 spin_lock(&parent->d_lock);
906 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
907 list_del_init(&dentry->d_u.d_child);
908 spin_unlock(&dentry->d_lock);
909 spin_unlock(&parent->d_lock);
910 remove_dir(dentry);
914 * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
915 * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
916 * reference to css->refcnt. In general, this refcnt is expected to goes down
917 * to zero, soon.
919 * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
921 DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
923 static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
925 if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
926 wake_up_all(&cgroup_rmdir_waitq);
929 void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
931 css_get(css);
934 void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
936 cgroup_wakeup_rmdir_waiter(css->cgroup);
937 css_put(css);
941 * Call with cgroup_mutex held. Drops reference counts on modules, including
942 * any duplicate ones that parse_cgroupfs_options took. If this function
943 * returns an error, no reference counts are touched.
945 static int rebind_subsystems(struct cgroupfs_root *root,
946 unsigned long final_bits)
948 unsigned long added_bits, removed_bits;
949 struct cgroup *cgrp = &root->top_cgroup;
950 int i;
952 BUG_ON(!mutex_is_locked(&cgroup_mutex));
954 removed_bits = root->actual_subsys_bits & ~final_bits;
955 added_bits = final_bits & ~root->actual_subsys_bits;
956 /* Check that any added subsystems are currently free */
957 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
958 unsigned long bit = 1UL << i;
959 struct cgroup_subsys *ss = subsys[i];
960 if (!(bit & added_bits))
961 continue;
963 * Nobody should tell us to do a subsys that doesn't exist:
964 * parse_cgroupfs_options should catch that case and refcounts
965 * ensure that subsystems won't disappear once selected.
967 BUG_ON(ss == NULL);
968 if (ss->root != &rootnode) {
969 /* Subsystem isn't free */
970 return -EBUSY;
974 /* Currently we don't handle adding/removing subsystems when
975 * any child cgroups exist. This is theoretically supportable
976 * but involves complex error handling, so it's being left until
977 * later */
978 if (root->number_of_cgroups > 1)
979 return -EBUSY;
981 /* Process each subsystem */
982 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
983 struct cgroup_subsys *ss = subsys[i];
984 unsigned long bit = 1UL << i;
985 if (bit & added_bits) {
986 /* We're binding this subsystem to this hierarchy */
987 BUG_ON(ss == NULL);
988 BUG_ON(cgrp->subsys[i]);
989 BUG_ON(!dummytop->subsys[i]);
990 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
991 mutex_lock(&ss->hierarchy_mutex);
992 cgrp->subsys[i] = dummytop->subsys[i];
993 cgrp->subsys[i]->cgroup = cgrp;
994 list_move(&ss->sibling, &root->subsys_list);
995 ss->root = root;
996 if (ss->bind)
997 ss->bind(ss, cgrp);
998 mutex_unlock(&ss->hierarchy_mutex);
999 /* refcount was already taken, and we're keeping it */
1000 } else if (bit & removed_bits) {
1001 /* We're removing this subsystem */
1002 BUG_ON(ss == NULL);
1003 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
1004 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
1005 mutex_lock(&ss->hierarchy_mutex);
1006 if (ss->bind)
1007 ss->bind(ss, dummytop);
1008 dummytop->subsys[i]->cgroup = dummytop;
1009 cgrp->subsys[i] = NULL;
1010 subsys[i]->root = &rootnode;
1011 list_move(&ss->sibling, &rootnode.subsys_list);
1012 mutex_unlock(&ss->hierarchy_mutex);
1013 /* subsystem is now free - drop reference on module */
1014 module_put(ss->module);
1015 } else if (bit & final_bits) {
1016 /* Subsystem state should already exist */
1017 BUG_ON(ss == NULL);
1018 BUG_ON(!cgrp->subsys[i]);
1020 * a refcount was taken, but we already had one, so
1021 * drop the extra reference.
1023 module_put(ss->module);
1024 #ifdef CONFIG_MODULE_UNLOAD
1025 BUG_ON(ss->module && !module_refcount(ss->module));
1026 #endif
1027 } else {
1028 /* Subsystem state shouldn't exist */
1029 BUG_ON(cgrp->subsys[i]);
1032 root->subsys_bits = root->actual_subsys_bits = final_bits;
1033 synchronize_rcu();
1035 return 0;
1038 static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
1040 struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
1041 struct cgroup_subsys *ss;
1043 mutex_lock(&cgroup_mutex);
1044 for_each_subsys(root, ss)
1045 seq_printf(seq, ",%s", ss->name);
1046 if (test_bit(ROOT_NOPREFIX, &root->flags))
1047 seq_puts(seq, ",noprefix");
1048 if (strlen(root->release_agent_path))
1049 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1050 if (clone_children(&root->top_cgroup))
1051 seq_puts(seq, ",clone_children");
1052 if (strlen(root->name))
1053 seq_printf(seq, ",name=%s", root->name);
1054 mutex_unlock(&cgroup_mutex);
1055 return 0;
1058 struct cgroup_sb_opts {
1059 unsigned long subsys_bits;
1060 unsigned long flags;
1061 char *release_agent;
1062 bool clone_children;
1063 char *name;
1064 /* User explicitly requested empty subsystem */
1065 bool none;
1067 struct cgroupfs_root *new_root;
1072 * Convert a hierarchy specifier into a bitmask of subsystems and flags. Call
1073 * with cgroup_mutex held to protect the subsys[] array. This function takes
1074 * refcounts on subsystems to be used, unless it returns error, in which case
1075 * no refcounts are taken.
1077 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1079 char *token, *o = data;
1080 bool all_ss = false, one_ss = false;
1081 unsigned long mask = (unsigned long)-1;
1082 int i;
1083 bool module_pin_failed = false;
1085 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1087 #ifdef CONFIG_CPUSETS
1088 mask = ~(1UL << cpuset_subsys_id);
1089 #endif
1091 memset(opts, 0, sizeof(*opts));
1093 while ((token = strsep(&o, ",")) != NULL) {
1094 if (!*token)
1095 return -EINVAL;
1096 if (!strcmp(token, "none")) {
1097 /* Explicitly have no subsystems */
1098 opts->none = true;
1099 continue;
1101 if (!strcmp(token, "all")) {
1102 /* Mutually exclusive option 'all' + subsystem name */
1103 if (one_ss)
1104 return -EINVAL;
1105 all_ss = true;
1106 continue;
1108 if (!strcmp(token, "noprefix")) {
1109 set_bit(ROOT_NOPREFIX, &opts->flags);
1110 continue;
1112 if (!strcmp(token, "clone_children")) {
1113 opts->clone_children = true;
1114 continue;
1116 if (!strncmp(token, "release_agent=", 14)) {
1117 /* Specifying two release agents is forbidden */
1118 if (opts->release_agent)
1119 return -EINVAL;
1120 opts->release_agent =
1121 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1122 if (!opts->release_agent)
1123 return -ENOMEM;
1124 continue;
1126 if (!strncmp(token, "name=", 5)) {
1127 const char *name = token + 5;
1128 /* Can't specify an empty name */
1129 if (!strlen(name))
1130 return -EINVAL;
1131 /* Must match [\w.-]+ */
1132 for (i = 0; i < strlen(name); i++) {
1133 char c = name[i];
1134 if (isalnum(c))
1135 continue;
1136 if ((c == '.') || (c == '-') || (c == '_'))
1137 continue;
1138 return -EINVAL;
1140 /* Specifying two names is forbidden */
1141 if (opts->name)
1142 return -EINVAL;
1143 opts->name = kstrndup(name,
1144 MAX_CGROUP_ROOT_NAMELEN - 1,
1145 GFP_KERNEL);
1146 if (!opts->name)
1147 return -ENOMEM;
1149 continue;
1152 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1153 struct cgroup_subsys *ss = subsys[i];
1154 if (ss == NULL)
1155 continue;
1156 if (strcmp(token, ss->name))
1157 continue;
1158 if (ss->disabled)
1159 continue;
1161 /* Mutually exclusive option 'all' + subsystem name */
1162 if (all_ss)
1163 return -EINVAL;
1164 set_bit(i, &opts->subsys_bits);
1165 one_ss = true;
1167 break;
1169 if (i == CGROUP_SUBSYS_COUNT)
1170 return -ENOENT;
1174 * If the 'all' option was specified select all the subsystems,
1175 * otherwise 'all, 'none' and a subsystem name options were not
1176 * specified, let's default to 'all'
1178 if (all_ss || (!all_ss && !one_ss && !opts->none)) {
1179 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1180 struct cgroup_subsys *ss = subsys[i];
1181 if (ss == NULL)
1182 continue;
1183 if (ss->disabled)
1184 continue;
1185 set_bit(i, &opts->subsys_bits);
1189 /* Consistency checks */
1192 * Option noprefix was introduced just for backward compatibility
1193 * with the old cpuset, so we allow noprefix only if mounting just
1194 * the cpuset subsystem.
1196 if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
1197 (opts->subsys_bits & mask))
1198 return -EINVAL;
1201 /* Can't specify "none" and some subsystems */
1202 if (opts->subsys_bits && opts->none)
1203 return -EINVAL;
1206 * We either have to specify by name or by subsystems. (So all
1207 * empty hierarchies must have a name).
1209 if (!opts->subsys_bits && !opts->name)
1210 return -EINVAL;
1213 * Grab references on all the modules we'll need, so the subsystems
1214 * don't dance around before rebind_subsystems attaches them. This may
1215 * take duplicate reference counts on a subsystem that's already used,
1216 * but rebind_subsystems handles this case.
1218 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1219 unsigned long bit = 1UL << i;
1221 if (!(bit & opts->subsys_bits))
1222 continue;
1223 if (!try_module_get(subsys[i]->module)) {
1224 module_pin_failed = true;
1225 break;
1228 if (module_pin_failed) {
1230 * oops, one of the modules was going away. this means that we
1231 * raced with a module_delete call, and to the user this is
1232 * essentially a "subsystem doesn't exist" case.
1234 for (i--; i >= CGROUP_BUILTIN_SUBSYS_COUNT; i--) {
1235 /* drop refcounts only on the ones we took */
1236 unsigned long bit = 1UL << i;
1238 if (!(bit & opts->subsys_bits))
1239 continue;
1240 module_put(subsys[i]->module);
1242 return -ENOENT;
1245 return 0;
1248 static void drop_parsed_module_refcounts(unsigned long subsys_bits)
1250 int i;
1251 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1252 unsigned long bit = 1UL << i;
1254 if (!(bit & subsys_bits))
1255 continue;
1256 module_put(subsys[i]->module);
1260 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1262 int ret = 0;
1263 struct cgroupfs_root *root = sb->s_fs_info;
1264 struct cgroup *cgrp = &root->top_cgroup;
1265 struct cgroup_sb_opts opts;
1267 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1268 mutex_lock(&cgroup_mutex);
1270 /* See what subsystems are wanted */
1271 ret = parse_cgroupfs_options(data, &opts);
1272 if (ret)
1273 goto out_unlock;
1275 /* Don't allow flags or name to change at remount */
1276 if (opts.flags != root->flags ||
1277 (opts.name && strcmp(opts.name, root->name))) {
1278 ret = -EINVAL;
1279 drop_parsed_module_refcounts(opts.subsys_bits);
1280 goto out_unlock;
1283 ret = rebind_subsystems(root, opts.subsys_bits);
1284 if (ret) {
1285 drop_parsed_module_refcounts(opts.subsys_bits);
1286 goto out_unlock;
1289 /* (re)populate subsystem files */
1290 cgroup_populate_dir(cgrp);
1292 if (opts.release_agent)
1293 strcpy(root->release_agent_path, opts.release_agent);
1294 out_unlock:
1295 kfree(opts.release_agent);
1296 kfree(opts.name);
1297 mutex_unlock(&cgroup_mutex);
1298 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1299 return ret;
1302 static const struct super_operations cgroup_ops = {
1303 .statfs = simple_statfs,
1304 .drop_inode = generic_delete_inode,
1305 .show_options = cgroup_show_options,
1306 .remount_fs = cgroup_remount,
1309 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1311 INIT_LIST_HEAD(&cgrp->sibling);
1312 INIT_LIST_HEAD(&cgrp->children);
1313 INIT_LIST_HEAD(&cgrp->css_sets);
1314 INIT_LIST_HEAD(&cgrp->release_list);
1315 INIT_LIST_HEAD(&cgrp->pidlists);
1316 mutex_init(&cgrp->pidlist_mutex);
1317 INIT_LIST_HEAD(&cgrp->event_list);
1318 spin_lock_init(&cgrp->event_list_lock);
1321 static void init_cgroup_root(struct cgroupfs_root *root)
1323 struct cgroup *cgrp = &root->top_cgroup;
1324 INIT_LIST_HEAD(&root->subsys_list);
1325 INIT_LIST_HEAD(&root->root_list);
1326 root->number_of_cgroups = 1;
1327 cgrp->root = root;
1328 cgrp->top_cgroup = cgrp;
1329 init_cgroup_housekeeping(cgrp);
1332 static bool init_root_id(struct cgroupfs_root *root)
1334 int ret = 0;
1336 do {
1337 if (!ida_pre_get(&hierarchy_ida, GFP_KERNEL))
1338 return false;
1339 spin_lock(&hierarchy_id_lock);
1340 /* Try to allocate the next unused ID */
1341 ret = ida_get_new_above(&hierarchy_ida, next_hierarchy_id,
1342 &root->hierarchy_id);
1343 if (ret == -ENOSPC)
1344 /* Try again starting from 0 */
1345 ret = ida_get_new(&hierarchy_ida, &root->hierarchy_id);
1346 if (!ret) {
1347 next_hierarchy_id = root->hierarchy_id + 1;
1348 } else if (ret != -EAGAIN) {
1349 /* Can only get here if the 31-bit IDR is full ... */
1350 BUG_ON(ret);
1352 spin_unlock(&hierarchy_id_lock);
1353 } while (ret);
1354 return true;
1357 static int cgroup_test_super(struct super_block *sb, void *data)
1359 struct cgroup_sb_opts *opts = data;
1360 struct cgroupfs_root *root = sb->s_fs_info;
1362 /* If we asked for a name then it must match */
1363 if (opts->name && strcmp(opts->name, root->name))
1364 return 0;
1367 * If we asked for subsystems (or explicitly for no
1368 * subsystems) then they must match
1370 if ((opts->subsys_bits || opts->none)
1371 && (opts->subsys_bits != root->subsys_bits))
1372 return 0;
1374 return 1;
1377 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1379 struct cgroupfs_root *root;
1381 if (!opts->subsys_bits && !opts->none)
1382 return NULL;
1384 root = kzalloc(sizeof(*root), GFP_KERNEL);
1385 if (!root)
1386 return ERR_PTR(-ENOMEM);
1388 if (!init_root_id(root)) {
1389 kfree(root);
1390 return ERR_PTR(-ENOMEM);
1392 init_cgroup_root(root);
1394 root->subsys_bits = opts->subsys_bits;
1395 root->flags = opts->flags;
1396 if (opts->release_agent)
1397 strcpy(root->release_agent_path, opts->release_agent);
1398 if (opts->name)
1399 strcpy(root->name, opts->name);
1400 if (opts->clone_children)
1401 set_bit(CGRP_CLONE_CHILDREN, &root->top_cgroup.flags);
1402 return root;
1405 static void cgroup_drop_root(struct cgroupfs_root *root)
1407 if (!root)
1408 return;
1410 BUG_ON(!root->hierarchy_id);
1411 spin_lock(&hierarchy_id_lock);
1412 ida_remove(&hierarchy_ida, root->hierarchy_id);
1413 spin_unlock(&hierarchy_id_lock);
1414 kfree(root);
1417 static int cgroup_set_super(struct super_block *sb, void *data)
1419 int ret;
1420 struct cgroup_sb_opts *opts = data;
1422 /* If we don't have a new root, we can't set up a new sb */
1423 if (!opts->new_root)
1424 return -EINVAL;
1426 BUG_ON(!opts->subsys_bits && !opts->none);
1428 ret = set_anon_super(sb, NULL);
1429 if (ret)
1430 return ret;
1432 sb->s_fs_info = opts->new_root;
1433 opts->new_root->sb = sb;
1435 sb->s_blocksize = PAGE_CACHE_SIZE;
1436 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1437 sb->s_magic = CGROUP_SUPER_MAGIC;
1438 sb->s_op = &cgroup_ops;
1440 return 0;
1443 static int cgroup_get_rootdir(struct super_block *sb)
1445 static const struct dentry_operations cgroup_dops = {
1446 .d_iput = cgroup_diput,
1447 .d_delete = cgroup_delete,
1450 struct inode *inode =
1451 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1452 struct dentry *dentry;
1454 if (!inode)
1455 return -ENOMEM;
1457 inode->i_fop = &simple_dir_operations;
1458 inode->i_op = &cgroup_dir_inode_operations;
1459 /* directories start off with i_nlink == 2 (for "." entry) */
1460 inc_nlink(inode);
1461 dentry = d_alloc_root(inode);
1462 if (!dentry) {
1463 iput(inode);
1464 return -ENOMEM;
1466 sb->s_root = dentry;
1467 /* for everything else we want ->d_op set */
1468 sb->s_d_op = &cgroup_dops;
1469 return 0;
1472 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1473 int flags, const char *unused_dev_name,
1474 void *data)
1476 struct cgroup_sb_opts opts;
1477 struct cgroupfs_root *root;
1478 int ret = 0;
1479 struct super_block *sb;
1480 struct cgroupfs_root *new_root;
1482 /* First find the desired set of subsystems */
1483 mutex_lock(&cgroup_mutex);
1484 ret = parse_cgroupfs_options(data, &opts);
1485 mutex_unlock(&cgroup_mutex);
1486 if (ret)
1487 goto out_err;
1490 * Allocate a new cgroup root. We may not need it if we're
1491 * reusing an existing hierarchy.
1493 new_root = cgroup_root_from_opts(&opts);
1494 if (IS_ERR(new_root)) {
1495 ret = PTR_ERR(new_root);
1496 goto drop_modules;
1498 opts.new_root = new_root;
1500 /* Locate an existing or new sb for this hierarchy */
1501 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
1502 if (IS_ERR(sb)) {
1503 ret = PTR_ERR(sb);
1504 cgroup_drop_root(opts.new_root);
1505 goto drop_modules;
1508 root = sb->s_fs_info;
1509 BUG_ON(!root);
1510 if (root == opts.new_root) {
1511 /* We used the new root structure, so this is a new hierarchy */
1512 struct list_head tmp_cg_links;
1513 struct cgroup *root_cgrp = &root->top_cgroup;
1514 struct inode *inode;
1515 struct cgroupfs_root *existing_root;
1516 int i;
1518 BUG_ON(sb->s_root != NULL);
1520 ret = cgroup_get_rootdir(sb);
1521 if (ret)
1522 goto drop_new_super;
1523 inode = sb->s_root->d_inode;
1525 mutex_lock(&inode->i_mutex);
1526 mutex_lock(&cgroup_mutex);
1528 if (strlen(root->name)) {
1529 /* Check for name clashes with existing mounts */
1530 for_each_active_root(existing_root) {
1531 if (!strcmp(existing_root->name, root->name)) {
1532 ret = -EBUSY;
1533 mutex_unlock(&cgroup_mutex);
1534 mutex_unlock(&inode->i_mutex);
1535 goto drop_new_super;
1541 * We're accessing css_set_count without locking
1542 * css_set_lock here, but that's OK - it can only be
1543 * increased by someone holding cgroup_lock, and
1544 * that's us. The worst that can happen is that we
1545 * have some link structures left over
1547 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1548 if (ret) {
1549 mutex_unlock(&cgroup_mutex);
1550 mutex_unlock(&inode->i_mutex);
1551 goto drop_new_super;
1554 ret = rebind_subsystems(root, root->subsys_bits);
1555 if (ret == -EBUSY) {
1556 mutex_unlock(&cgroup_mutex);
1557 mutex_unlock(&inode->i_mutex);
1558 free_cg_links(&tmp_cg_links);
1559 goto drop_new_super;
1562 * There must be no failure case after here, since rebinding
1563 * takes care of subsystems' refcounts, which are explicitly
1564 * dropped in the failure exit path.
1567 /* EBUSY should be the only error here */
1568 BUG_ON(ret);
1570 list_add(&root->root_list, &roots);
1571 root_count++;
1573 sb->s_root->d_fsdata = root_cgrp;
1574 root->top_cgroup.dentry = sb->s_root;
1576 /* Link the top cgroup in this hierarchy into all
1577 * the css_set objects */
1578 write_lock(&css_set_lock);
1579 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1580 struct hlist_head *hhead = &css_set_table[i];
1581 struct hlist_node *node;
1582 struct css_set *cg;
1584 hlist_for_each_entry(cg, node, hhead, hlist)
1585 link_css_set(&tmp_cg_links, cg, root_cgrp);
1587 write_unlock(&css_set_lock);
1589 free_cg_links(&tmp_cg_links);
1591 BUG_ON(!list_empty(&root_cgrp->sibling));
1592 BUG_ON(!list_empty(&root_cgrp->children));
1593 BUG_ON(root->number_of_cgroups != 1);
1595 cgroup_populate_dir(root_cgrp);
1596 mutex_unlock(&cgroup_mutex);
1597 mutex_unlock(&inode->i_mutex);
1598 } else {
1600 * We re-used an existing hierarchy - the new root (if
1601 * any) is not needed
1603 cgroup_drop_root(opts.new_root);
1604 /* no subsys rebinding, so refcounts don't change */
1605 drop_parsed_module_refcounts(opts.subsys_bits);
1608 kfree(opts.release_agent);
1609 kfree(opts.name);
1610 return dget(sb->s_root);
1612 drop_new_super:
1613 deactivate_locked_super(sb);
1614 drop_modules:
1615 drop_parsed_module_refcounts(opts.subsys_bits);
1616 out_err:
1617 kfree(opts.release_agent);
1618 kfree(opts.name);
1619 return ERR_PTR(ret);
1622 static void cgroup_kill_sb(struct super_block *sb) {
1623 struct cgroupfs_root *root = sb->s_fs_info;
1624 struct cgroup *cgrp = &root->top_cgroup;
1625 int ret;
1626 struct cg_cgroup_link *link;
1627 struct cg_cgroup_link *saved_link;
1629 BUG_ON(!root);
1631 BUG_ON(root->number_of_cgroups != 1);
1632 BUG_ON(!list_empty(&cgrp->children));
1633 BUG_ON(!list_empty(&cgrp->sibling));
1635 mutex_lock(&cgroup_mutex);
1637 /* Rebind all subsystems back to the default hierarchy */
1638 ret = rebind_subsystems(root, 0);
1639 /* Shouldn't be able to fail ... */
1640 BUG_ON(ret);
1643 * Release all the links from css_sets to this hierarchy's
1644 * root cgroup
1646 write_lock(&css_set_lock);
1648 list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1649 cgrp_link_list) {
1650 list_del(&link->cg_link_list);
1651 list_del(&link->cgrp_link_list);
1652 kfree(link);
1654 write_unlock(&css_set_lock);
1656 if (!list_empty(&root->root_list)) {
1657 list_del(&root->root_list);
1658 root_count--;
1661 mutex_unlock(&cgroup_mutex);
1663 kill_litter_super(sb);
1664 cgroup_drop_root(root);
1667 static struct file_system_type cgroup_fs_type = {
1668 .name = "cgroup",
1669 .mount = cgroup_mount,
1670 .kill_sb = cgroup_kill_sb,
1673 static struct kobject *cgroup_kobj;
1675 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1677 return dentry->d_fsdata;
1680 static inline struct cftype *__d_cft(struct dentry *dentry)
1682 return dentry->d_fsdata;
1686 * cgroup_path - generate the path of a cgroup
1687 * @cgrp: the cgroup in question
1688 * @buf: the buffer to write the path into
1689 * @buflen: the length of the buffer
1691 * Called with cgroup_mutex held or else with an RCU-protected cgroup
1692 * reference. Writes path of cgroup into buf. Returns 0 on success,
1693 * -errno on error.
1695 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1697 char *start;
1698 struct dentry *dentry = rcu_dereference_check(cgrp->dentry,
1699 rcu_read_lock_held() ||
1700 cgroup_lock_is_held());
1702 if (!dentry || cgrp == dummytop) {
1704 * Inactive subsystems have no dentry for their root
1705 * cgroup
1707 strcpy(buf, "/");
1708 return 0;
1711 start = buf + buflen;
1713 *--start = '\0';
1714 for (;;) {
1715 int len = dentry->d_name.len;
1717 if ((start -= len) < buf)
1718 return -ENAMETOOLONG;
1719 memcpy(start, dentry->d_name.name, len);
1720 cgrp = cgrp->parent;
1721 if (!cgrp)
1722 break;
1724 dentry = rcu_dereference_check(cgrp->dentry,
1725 rcu_read_lock_held() ||
1726 cgroup_lock_is_held());
1727 if (!cgrp->parent)
1728 continue;
1729 if (--start < buf)
1730 return -ENAMETOOLONG;
1731 *start = '/';
1733 memmove(buf, start, buf + buflen - start);
1734 return 0;
1736 EXPORT_SYMBOL_GPL(cgroup_path);
1739 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1740 * @cgrp: the cgroup the task is attaching to
1741 * @tsk: the task to be attached
1743 * Call holding cgroup_mutex. May take task_lock of
1744 * the task 'tsk' during call.
1746 int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1748 int retval = 0;
1749 struct cgroup_subsys *ss, *failed_ss = NULL;
1750 struct cgroup *oldcgrp;
1751 struct css_set *cg;
1752 struct css_set *newcg;
1753 struct cgroupfs_root *root = cgrp->root;
1755 /* Nothing to do if the task is already in that cgroup */
1756 oldcgrp = task_cgroup_from_root(tsk, root);
1757 if (cgrp == oldcgrp)
1758 return 0;
1760 for_each_subsys(root, ss) {
1761 if (ss->can_attach) {
1762 retval = ss->can_attach(ss, cgrp, tsk, false);
1763 if (retval) {
1765 * Remember on which subsystem the can_attach()
1766 * failed, so that we only call cancel_attach()
1767 * against the subsystems whose can_attach()
1768 * succeeded. (See below)
1770 failed_ss = ss;
1771 goto out;
1776 task_lock(tsk);
1777 cg = tsk->cgroups;
1778 get_css_set(cg);
1779 task_unlock(tsk);
1781 * Locate or allocate a new css_set for this task,
1782 * based on its final set of cgroups
1784 newcg = find_css_set(cg, cgrp);
1785 put_css_set(cg);
1786 if (!newcg) {
1787 retval = -ENOMEM;
1788 goto out;
1791 task_lock(tsk);
1792 if (tsk->flags & PF_EXITING) {
1793 task_unlock(tsk);
1794 put_css_set(newcg);
1795 retval = -ESRCH;
1796 goto out;
1798 rcu_assign_pointer(tsk->cgroups, newcg);
1799 task_unlock(tsk);
1801 /* Update the css_set linked lists if we're using them */
1802 write_lock(&css_set_lock);
1803 if (!list_empty(&tsk->cg_list))
1804 list_move(&tsk->cg_list, &newcg->tasks);
1805 write_unlock(&css_set_lock);
1807 for_each_subsys(root, ss) {
1808 if (ss->attach)
1809 ss->attach(ss, cgrp, oldcgrp, tsk, false);
1811 set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1812 synchronize_rcu();
1813 put_css_set(cg);
1816 * wake up rmdir() waiter. the rmdir should fail since the cgroup
1817 * is no longer empty.
1819 cgroup_wakeup_rmdir_waiter(cgrp);
1820 out:
1821 if (retval) {
1822 for_each_subsys(root, ss) {
1823 if (ss == failed_ss)
1825 * This subsystem was the one that failed the
1826 * can_attach() check earlier, so we don't need
1827 * to call cancel_attach() against it or any
1828 * remaining subsystems.
1830 break;
1831 if (ss->cancel_attach)
1832 ss->cancel_attach(ss, cgrp, tsk, false);
1835 return retval;
1839 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1840 * @from: attach to all cgroups of a given task
1841 * @tsk: the task to be attached
1843 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1845 struct cgroupfs_root *root;
1846 int retval = 0;
1848 cgroup_lock();
1849 for_each_active_root(root) {
1850 struct cgroup *from_cg = task_cgroup_from_root(from, root);
1852 retval = cgroup_attach_task(from_cg, tsk);
1853 if (retval)
1854 break;
1856 cgroup_unlock();
1858 return retval;
1860 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
1863 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
1864 * held. May take task_lock of task
1866 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid)
1868 struct task_struct *tsk;
1869 const struct cred *cred = current_cred(), *tcred;
1870 int ret;
1872 if (pid) {
1873 rcu_read_lock();
1874 tsk = find_task_by_vpid(pid);
1875 if (!tsk || tsk->flags & PF_EXITING) {
1876 rcu_read_unlock();
1877 return -ESRCH;
1880 tcred = __task_cred(tsk);
1881 if (cred->euid &&
1882 cred->euid != tcred->uid &&
1883 cred->euid != tcred->suid) {
1884 rcu_read_unlock();
1885 return -EACCES;
1887 get_task_struct(tsk);
1888 rcu_read_unlock();
1889 } else {
1890 tsk = current;
1891 get_task_struct(tsk);
1894 ret = cgroup_attach_task(cgrp, tsk);
1895 put_task_struct(tsk);
1896 return ret;
1899 static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
1901 int ret;
1902 if (!cgroup_lock_live_group(cgrp))
1903 return -ENODEV;
1904 ret = attach_task_by_pid(cgrp, pid);
1905 cgroup_unlock();
1906 return ret;
1910 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
1911 * @cgrp: the cgroup to be checked for liveness
1913 * On success, returns true; the lock should be later released with
1914 * cgroup_unlock(). On failure returns false with no lock held.
1916 bool cgroup_lock_live_group(struct cgroup *cgrp)
1918 mutex_lock(&cgroup_mutex);
1919 if (cgroup_is_removed(cgrp)) {
1920 mutex_unlock(&cgroup_mutex);
1921 return false;
1923 return true;
1925 EXPORT_SYMBOL_GPL(cgroup_lock_live_group);
1927 static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
1928 const char *buffer)
1930 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1931 if (strlen(buffer) >= PATH_MAX)
1932 return -EINVAL;
1933 if (!cgroup_lock_live_group(cgrp))
1934 return -ENODEV;
1935 strcpy(cgrp->root->release_agent_path, buffer);
1936 cgroup_unlock();
1937 return 0;
1940 static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
1941 struct seq_file *seq)
1943 if (!cgroup_lock_live_group(cgrp))
1944 return -ENODEV;
1945 seq_puts(seq, cgrp->root->release_agent_path);
1946 seq_putc(seq, '\n');
1947 cgroup_unlock();
1948 return 0;
1951 /* A buffer size big enough for numbers or short strings */
1952 #define CGROUP_LOCAL_BUFFER_SIZE 64
1954 static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
1955 struct file *file,
1956 const char __user *userbuf,
1957 size_t nbytes, loff_t *unused_ppos)
1959 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
1960 int retval = 0;
1961 char *end;
1963 if (!nbytes)
1964 return -EINVAL;
1965 if (nbytes >= sizeof(buffer))
1966 return -E2BIG;
1967 if (copy_from_user(buffer, userbuf, nbytes))
1968 return -EFAULT;
1970 buffer[nbytes] = 0; /* nul-terminate */
1971 if (cft->write_u64) {
1972 u64 val = simple_strtoull(strstrip(buffer), &end, 0);
1973 if (*end)
1974 return -EINVAL;
1975 retval = cft->write_u64(cgrp, cft, val);
1976 } else {
1977 s64 val = simple_strtoll(strstrip(buffer), &end, 0);
1978 if (*end)
1979 return -EINVAL;
1980 retval = cft->write_s64(cgrp, cft, val);
1982 if (!retval)
1983 retval = nbytes;
1984 return retval;
1987 static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
1988 struct file *file,
1989 const char __user *userbuf,
1990 size_t nbytes, loff_t *unused_ppos)
1992 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
1993 int retval = 0;
1994 size_t max_bytes = cft->max_write_len;
1995 char *buffer = local_buffer;
1997 if (!max_bytes)
1998 max_bytes = sizeof(local_buffer) - 1;
1999 if (nbytes >= max_bytes)
2000 return -E2BIG;
2001 /* Allocate a dynamic buffer if we need one */
2002 if (nbytes >= sizeof(local_buffer)) {
2003 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
2004 if (buffer == NULL)
2005 return -ENOMEM;
2007 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
2008 retval = -EFAULT;
2009 goto out;
2012 buffer[nbytes] = 0; /* nul-terminate */
2013 retval = cft->write_string(cgrp, cft, strstrip(buffer));
2014 if (!retval)
2015 retval = nbytes;
2016 out:
2017 if (buffer != local_buffer)
2018 kfree(buffer);
2019 return retval;
2022 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
2023 size_t nbytes, loff_t *ppos)
2025 struct cftype *cft = __d_cft(file->f_dentry);
2026 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2028 if (cgroup_is_removed(cgrp))
2029 return -ENODEV;
2030 if (cft->write)
2031 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
2032 if (cft->write_u64 || cft->write_s64)
2033 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
2034 if (cft->write_string)
2035 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
2036 if (cft->trigger) {
2037 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
2038 return ret ? ret : nbytes;
2040 return -EINVAL;
2043 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
2044 struct file *file,
2045 char __user *buf, size_t nbytes,
2046 loff_t *ppos)
2048 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2049 u64 val = cft->read_u64(cgrp, cft);
2050 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
2052 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2055 static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
2056 struct file *file,
2057 char __user *buf, size_t nbytes,
2058 loff_t *ppos)
2060 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2061 s64 val = cft->read_s64(cgrp, cft);
2062 int len = sprintf(tmp, "%lld\n", (long long) val);
2064 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2067 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2068 size_t nbytes, loff_t *ppos)
2070 struct cftype *cft = __d_cft(file->f_dentry);
2071 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2073 if (cgroup_is_removed(cgrp))
2074 return -ENODEV;
2076 if (cft->read)
2077 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2078 if (cft->read_u64)
2079 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2080 if (cft->read_s64)
2081 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2082 return -EINVAL;
2086 * seqfile ops/methods for returning structured data. Currently just
2087 * supports string->u64 maps, but can be extended in future.
2090 struct cgroup_seqfile_state {
2091 struct cftype *cft;
2092 struct cgroup *cgroup;
2095 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2097 struct seq_file *sf = cb->state;
2098 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2101 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2103 struct cgroup_seqfile_state *state = m->private;
2104 struct cftype *cft = state->cft;
2105 if (cft->read_map) {
2106 struct cgroup_map_cb cb = {
2107 .fill = cgroup_map_add,
2108 .state = m,
2110 return cft->read_map(state->cgroup, cft, &cb);
2112 return cft->read_seq_string(state->cgroup, cft, m);
2115 static int cgroup_seqfile_release(struct inode *inode, struct file *file)
2117 struct seq_file *seq = file->private_data;
2118 kfree(seq->private);
2119 return single_release(inode, file);
2122 static const struct file_operations cgroup_seqfile_operations = {
2123 .read = seq_read,
2124 .write = cgroup_file_write,
2125 .llseek = seq_lseek,
2126 .release = cgroup_seqfile_release,
2129 static int cgroup_file_open(struct inode *inode, struct file *file)
2131 int err;
2132 struct cftype *cft;
2134 err = generic_file_open(inode, file);
2135 if (err)
2136 return err;
2137 cft = __d_cft(file->f_dentry);
2139 if (cft->read_map || cft->read_seq_string) {
2140 struct cgroup_seqfile_state *state =
2141 kzalloc(sizeof(*state), GFP_USER);
2142 if (!state)
2143 return -ENOMEM;
2144 state->cft = cft;
2145 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
2146 file->f_op = &cgroup_seqfile_operations;
2147 err = single_open(file, cgroup_seqfile_show, state);
2148 if (err < 0)
2149 kfree(state);
2150 } else if (cft->open)
2151 err = cft->open(inode, file);
2152 else
2153 err = 0;
2155 return err;
2158 static int cgroup_file_release(struct inode *inode, struct file *file)
2160 struct cftype *cft = __d_cft(file->f_dentry);
2161 if (cft->release)
2162 return cft->release(inode, file);
2163 return 0;
2167 * cgroup_rename - Only allow simple rename of directories in place.
2169 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2170 struct inode *new_dir, struct dentry *new_dentry)
2172 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2173 return -ENOTDIR;
2174 if (new_dentry->d_inode)
2175 return -EEXIST;
2176 if (old_dir != new_dir)
2177 return -EIO;
2178 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2181 static const struct file_operations cgroup_file_operations = {
2182 .read = cgroup_file_read,
2183 .write = cgroup_file_write,
2184 .llseek = generic_file_llseek,
2185 .open = cgroup_file_open,
2186 .release = cgroup_file_release,
2189 static const struct inode_operations cgroup_dir_inode_operations = {
2190 .lookup = cgroup_lookup,
2191 .mkdir = cgroup_mkdir,
2192 .rmdir = cgroup_rmdir,
2193 .rename = cgroup_rename,
2196 static struct dentry *cgroup_lookup(struct inode *dir, struct dentry *dentry, struct nameidata *nd)
2198 if (dentry->d_name.len > NAME_MAX)
2199 return ERR_PTR(-ENAMETOOLONG);
2200 d_add(dentry, NULL);
2201 return NULL;
2205 * Check if a file is a control file
2207 static inline struct cftype *__file_cft(struct file *file)
2209 if (file->f_dentry->d_inode->i_fop != &cgroup_file_operations)
2210 return ERR_PTR(-EINVAL);
2211 return __d_cft(file->f_dentry);
2214 static int cgroup_create_file(struct dentry *dentry, mode_t mode,
2215 struct super_block *sb)
2217 struct inode *inode;
2219 if (!dentry)
2220 return -ENOENT;
2221 if (dentry->d_inode)
2222 return -EEXIST;
2224 inode = cgroup_new_inode(mode, sb);
2225 if (!inode)
2226 return -ENOMEM;
2228 if (S_ISDIR(mode)) {
2229 inode->i_op = &cgroup_dir_inode_operations;
2230 inode->i_fop = &simple_dir_operations;
2232 /* start off with i_nlink == 2 (for "." entry) */
2233 inc_nlink(inode);
2235 /* start with the directory inode held, so that we can
2236 * populate it without racing with another mkdir */
2237 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2238 } else if (S_ISREG(mode)) {
2239 inode->i_size = 0;
2240 inode->i_fop = &cgroup_file_operations;
2242 d_instantiate(dentry, inode);
2243 dget(dentry); /* Extra count - pin the dentry in core */
2244 return 0;
2248 * cgroup_create_dir - create a directory for an object.
2249 * @cgrp: the cgroup we create the directory for. It must have a valid
2250 * ->parent field. And we are going to fill its ->dentry field.
2251 * @dentry: dentry of the new cgroup
2252 * @mode: mode to set on new directory.
2254 static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
2255 mode_t mode)
2257 struct dentry *parent;
2258 int error = 0;
2260 parent = cgrp->parent->dentry;
2261 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
2262 if (!error) {
2263 dentry->d_fsdata = cgrp;
2264 inc_nlink(parent->d_inode);
2265 rcu_assign_pointer(cgrp->dentry, dentry);
2266 dget(dentry);
2268 dput(dentry);
2270 return error;
2274 * cgroup_file_mode - deduce file mode of a control file
2275 * @cft: the control file in question
2277 * returns cft->mode if ->mode is not 0
2278 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2279 * returns S_IRUGO if it has only a read handler
2280 * returns S_IWUSR if it has only a write hander
2282 static mode_t cgroup_file_mode(const struct cftype *cft)
2284 mode_t mode = 0;
2286 if (cft->mode)
2287 return cft->mode;
2289 if (cft->read || cft->read_u64 || cft->read_s64 ||
2290 cft->read_map || cft->read_seq_string)
2291 mode |= S_IRUGO;
2293 if (cft->write || cft->write_u64 || cft->write_s64 ||
2294 cft->write_string || cft->trigger)
2295 mode |= S_IWUSR;
2297 return mode;
2300 int cgroup_add_file(struct cgroup *cgrp,
2301 struct cgroup_subsys *subsys,
2302 const struct cftype *cft)
2304 struct dentry *dir = cgrp->dentry;
2305 struct dentry *dentry;
2306 int error;
2307 mode_t mode;
2309 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2310 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
2311 strcpy(name, subsys->name);
2312 strcat(name, ".");
2314 strcat(name, cft->name);
2315 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2316 dentry = lookup_one_len(name, dir, strlen(name));
2317 if (!IS_ERR(dentry)) {
2318 mode = cgroup_file_mode(cft);
2319 error = cgroup_create_file(dentry, mode | S_IFREG,
2320 cgrp->root->sb);
2321 if (!error)
2322 dentry->d_fsdata = (void *)cft;
2323 dput(dentry);
2324 } else
2325 error = PTR_ERR(dentry);
2326 return error;
2328 EXPORT_SYMBOL_GPL(cgroup_add_file);
2330 int cgroup_add_files(struct cgroup *cgrp,
2331 struct cgroup_subsys *subsys,
2332 const struct cftype cft[],
2333 int count)
2335 int i, err;
2336 for (i = 0; i < count; i++) {
2337 err = cgroup_add_file(cgrp, subsys, &cft[i]);
2338 if (err)
2339 return err;
2341 return 0;
2343 EXPORT_SYMBOL_GPL(cgroup_add_files);
2346 * cgroup_task_count - count the number of tasks in a cgroup.
2347 * @cgrp: the cgroup in question
2349 * Return the number of tasks in the cgroup.
2351 int cgroup_task_count(const struct cgroup *cgrp)
2353 int count = 0;
2354 struct cg_cgroup_link *link;
2356 read_lock(&css_set_lock);
2357 list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
2358 count += atomic_read(&link->cg->refcount);
2360 read_unlock(&css_set_lock);
2361 return count;
2365 * Advance a list_head iterator. The iterator should be positioned at
2366 * the start of a css_set
2368 static void cgroup_advance_iter(struct cgroup *cgrp,
2369 struct cgroup_iter *it)
2371 struct list_head *l = it->cg_link;
2372 struct cg_cgroup_link *link;
2373 struct css_set *cg;
2375 /* Advance to the next non-empty css_set */
2376 do {
2377 l = l->next;
2378 if (l == &cgrp->css_sets) {
2379 it->cg_link = NULL;
2380 return;
2382 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
2383 cg = link->cg;
2384 } while (list_empty(&cg->tasks));
2385 it->cg_link = l;
2386 it->task = cg->tasks.next;
2390 * To reduce the fork() overhead for systems that are not actually
2391 * using their cgroups capability, we don't maintain the lists running
2392 * through each css_set to its tasks until we see the list actually
2393 * used - in other words after the first call to cgroup_iter_start().
2395 * The tasklist_lock is not held here, as do_each_thread() and
2396 * while_each_thread() are protected by RCU.
2398 static void cgroup_enable_task_cg_lists(void)
2400 struct task_struct *p, *g;
2401 write_lock(&css_set_lock);
2402 use_task_css_set_links = 1;
2403 do_each_thread(g, p) {
2404 task_lock(p);
2406 * We should check if the process is exiting, otherwise
2407 * it will race with cgroup_exit() in that the list
2408 * entry won't be deleted though the process has exited.
2410 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2411 list_add(&p->cg_list, &p->cgroups->tasks);
2412 task_unlock(p);
2413 } while_each_thread(g, p);
2414 write_unlock(&css_set_lock);
2417 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
2420 * The first time anyone tries to iterate across a cgroup,
2421 * we need to enable the list linking each css_set to its
2422 * tasks, and fix up all existing tasks.
2424 if (!use_task_css_set_links)
2425 cgroup_enable_task_cg_lists();
2427 read_lock(&css_set_lock);
2428 it->cg_link = &cgrp->css_sets;
2429 cgroup_advance_iter(cgrp, it);
2432 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
2433 struct cgroup_iter *it)
2435 struct task_struct *res;
2436 struct list_head *l = it->task;
2437 struct cg_cgroup_link *link;
2439 /* If the iterator cg is NULL, we have no tasks */
2440 if (!it->cg_link)
2441 return NULL;
2442 res = list_entry(l, struct task_struct, cg_list);
2443 /* Advance iterator to find next entry */
2444 l = l->next;
2445 link = list_entry(it->cg_link, struct cg_cgroup_link, cgrp_link_list);
2446 if (l == &link->cg->tasks) {
2447 /* We reached the end of this task list - move on to
2448 * the next cg_cgroup_link */
2449 cgroup_advance_iter(cgrp, it);
2450 } else {
2451 it->task = l;
2453 return res;
2456 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
2458 read_unlock(&css_set_lock);
2461 static inline int started_after_time(struct task_struct *t1,
2462 struct timespec *time,
2463 struct task_struct *t2)
2465 int start_diff = timespec_compare(&t1->start_time, time);
2466 if (start_diff > 0) {
2467 return 1;
2468 } else if (start_diff < 0) {
2469 return 0;
2470 } else {
2472 * Arbitrarily, if two processes started at the same
2473 * time, we'll say that the lower pointer value
2474 * started first. Note that t2 may have exited by now
2475 * so this may not be a valid pointer any longer, but
2476 * that's fine - it still serves to distinguish
2477 * between two tasks started (effectively) simultaneously.
2479 return t1 > t2;
2484 * This function is a callback from heap_insert() and is used to order
2485 * the heap.
2486 * In this case we order the heap in descending task start time.
2488 static inline int started_after(void *p1, void *p2)
2490 struct task_struct *t1 = p1;
2491 struct task_struct *t2 = p2;
2492 return started_after_time(t1, &t2->start_time, t2);
2496 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
2497 * @scan: struct cgroup_scanner containing arguments for the scan
2499 * Arguments include pointers to callback functions test_task() and
2500 * process_task().
2501 * Iterate through all the tasks in a cgroup, calling test_task() for each,
2502 * and if it returns true, call process_task() for it also.
2503 * The test_task pointer may be NULL, meaning always true (select all tasks).
2504 * Effectively duplicates cgroup_iter_{start,next,end}()
2505 * but does not lock css_set_lock for the call to process_task().
2506 * The struct cgroup_scanner may be embedded in any structure of the caller's
2507 * creation.
2508 * It is guaranteed that process_task() will act on every task that
2509 * is a member of the cgroup for the duration of this call. This
2510 * function may or may not call process_task() for tasks that exit
2511 * or move to a different cgroup during the call, or are forked or
2512 * move into the cgroup during the call.
2514 * Note that test_task() may be called with locks held, and may in some
2515 * situations be called multiple times for the same task, so it should
2516 * be cheap.
2517 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
2518 * pre-allocated and will be used for heap operations (and its "gt" member will
2519 * be overwritten), else a temporary heap will be used (allocation of which
2520 * may cause this function to fail).
2522 int cgroup_scan_tasks(struct cgroup_scanner *scan)
2524 int retval, i;
2525 struct cgroup_iter it;
2526 struct task_struct *p, *dropped;
2527 /* Never dereference latest_task, since it's not refcounted */
2528 struct task_struct *latest_task = NULL;
2529 struct ptr_heap tmp_heap;
2530 struct ptr_heap *heap;
2531 struct timespec latest_time = { 0, 0 };
2533 if (scan->heap) {
2534 /* The caller supplied our heap and pre-allocated its memory */
2535 heap = scan->heap;
2536 heap->gt = &started_after;
2537 } else {
2538 /* We need to allocate our own heap memory */
2539 heap = &tmp_heap;
2540 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2541 if (retval)
2542 /* cannot allocate the heap */
2543 return retval;
2546 again:
2548 * Scan tasks in the cgroup, using the scanner's "test_task" callback
2549 * to determine which are of interest, and using the scanner's
2550 * "process_task" callback to process any of them that need an update.
2551 * Since we don't want to hold any locks during the task updates,
2552 * gather tasks to be processed in a heap structure.
2553 * The heap is sorted by descending task start time.
2554 * If the statically-sized heap fills up, we overflow tasks that
2555 * started later, and in future iterations only consider tasks that
2556 * started after the latest task in the previous pass. This
2557 * guarantees forward progress and that we don't miss any tasks.
2559 heap->size = 0;
2560 cgroup_iter_start(scan->cg, &it);
2561 while ((p = cgroup_iter_next(scan->cg, &it))) {
2563 * Only affect tasks that qualify per the caller's callback,
2564 * if he provided one
2566 if (scan->test_task && !scan->test_task(p, scan))
2567 continue;
2569 * Only process tasks that started after the last task
2570 * we processed
2572 if (!started_after_time(p, &latest_time, latest_task))
2573 continue;
2574 dropped = heap_insert(heap, p);
2575 if (dropped == NULL) {
2577 * The new task was inserted; the heap wasn't
2578 * previously full
2580 get_task_struct(p);
2581 } else if (dropped != p) {
2583 * The new task was inserted, and pushed out a
2584 * different task
2586 get_task_struct(p);
2587 put_task_struct(dropped);
2590 * Else the new task was newer than anything already in
2591 * the heap and wasn't inserted
2594 cgroup_iter_end(scan->cg, &it);
2596 if (heap->size) {
2597 for (i = 0; i < heap->size; i++) {
2598 struct task_struct *q = heap->ptrs[i];
2599 if (i == 0) {
2600 latest_time = q->start_time;
2601 latest_task = q;
2603 /* Process the task per the caller's callback */
2604 scan->process_task(q, scan);
2605 put_task_struct(q);
2608 * If we had to process any tasks at all, scan again
2609 * in case some of them were in the middle of forking
2610 * children that didn't get processed.
2611 * Not the most efficient way to do it, but it avoids
2612 * having to take callback_mutex in the fork path
2614 goto again;
2616 if (heap == &tmp_heap)
2617 heap_free(&tmp_heap);
2618 return 0;
2622 * Stuff for reading the 'tasks'/'procs' files.
2624 * Reading this file can return large amounts of data if a cgroup has
2625 * *lots* of attached tasks. So it may need several calls to read(),
2626 * but we cannot guarantee that the information we produce is correct
2627 * unless we produce it entirely atomically.
2632 * The following two functions "fix" the issue where there are more pids
2633 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2634 * TODO: replace with a kernel-wide solution to this problem
2636 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2637 static void *pidlist_allocate(int count)
2639 if (PIDLIST_TOO_LARGE(count))
2640 return vmalloc(count * sizeof(pid_t));
2641 else
2642 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2644 static void pidlist_free(void *p)
2646 if (is_vmalloc_addr(p))
2647 vfree(p);
2648 else
2649 kfree(p);
2651 static void *pidlist_resize(void *p, int newcount)
2653 void *newlist;
2654 /* note: if new alloc fails, old p will still be valid either way */
2655 if (is_vmalloc_addr(p)) {
2656 newlist = vmalloc(newcount * sizeof(pid_t));
2657 if (!newlist)
2658 return NULL;
2659 memcpy(newlist, p, newcount * sizeof(pid_t));
2660 vfree(p);
2661 } else {
2662 newlist = krealloc(p, newcount * sizeof(pid_t), GFP_KERNEL);
2664 return newlist;
2668 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2669 * If the new stripped list is sufficiently smaller and there's enough memory
2670 * to allocate a new buffer, will let go of the unneeded memory. Returns the
2671 * number of unique elements.
2673 /* is the size difference enough that we should re-allocate the array? */
2674 #define PIDLIST_REALLOC_DIFFERENCE(old, new) ((old) - PAGE_SIZE >= (new))
2675 static int pidlist_uniq(pid_t **p, int length)
2677 int src, dest = 1;
2678 pid_t *list = *p;
2679 pid_t *newlist;
2682 * we presume the 0th element is unique, so i starts at 1. trivial
2683 * edge cases first; no work needs to be done for either
2685 if (length == 0 || length == 1)
2686 return length;
2687 /* src and dest walk down the list; dest counts unique elements */
2688 for (src = 1; src < length; src++) {
2689 /* find next unique element */
2690 while (list[src] == list[src-1]) {
2691 src++;
2692 if (src == length)
2693 goto after;
2695 /* dest always points to where the next unique element goes */
2696 list[dest] = list[src];
2697 dest++;
2699 after:
2701 * if the length difference is large enough, we want to allocate a
2702 * smaller buffer to save memory. if this fails due to out of memory,
2703 * we'll just stay with what we've got.
2705 if (PIDLIST_REALLOC_DIFFERENCE(length, dest)) {
2706 newlist = pidlist_resize(list, dest);
2707 if (newlist)
2708 *p = newlist;
2710 return dest;
2713 static int cmppid(const void *a, const void *b)
2715 return *(pid_t *)a - *(pid_t *)b;
2719 * find the appropriate pidlist for our purpose (given procs vs tasks)
2720 * returns with the lock on that pidlist already held, and takes care
2721 * of the use count, or returns NULL with no locks held if we're out of
2722 * memory.
2724 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
2725 enum cgroup_filetype type)
2727 struct cgroup_pidlist *l;
2728 /* don't need task_nsproxy() if we're looking at ourself */
2729 struct pid_namespace *ns = current->nsproxy->pid_ns;
2732 * We can't drop the pidlist_mutex before taking the l->mutex in case
2733 * the last ref-holder is trying to remove l from the list at the same
2734 * time. Holding the pidlist_mutex precludes somebody taking whichever
2735 * list we find out from under us - compare release_pid_array().
2737 mutex_lock(&cgrp->pidlist_mutex);
2738 list_for_each_entry(l, &cgrp->pidlists, links) {
2739 if (l->key.type == type && l->key.ns == ns) {
2740 /* make sure l doesn't vanish out from under us */
2741 down_write(&l->mutex);
2742 mutex_unlock(&cgrp->pidlist_mutex);
2743 return l;
2746 /* entry not found; create a new one */
2747 l = kmalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
2748 if (!l) {
2749 mutex_unlock(&cgrp->pidlist_mutex);
2750 return l;
2752 init_rwsem(&l->mutex);
2753 down_write(&l->mutex);
2754 l->key.type = type;
2755 l->key.ns = get_pid_ns(ns);
2756 l->use_count = 0; /* don't increment here */
2757 l->list = NULL;
2758 l->owner = cgrp;
2759 list_add(&l->links, &cgrp->pidlists);
2760 mutex_unlock(&cgrp->pidlist_mutex);
2761 return l;
2765 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
2767 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
2768 struct cgroup_pidlist **lp)
2770 pid_t *array;
2771 int length;
2772 int pid, n = 0; /* used for populating the array */
2773 struct cgroup_iter it;
2774 struct task_struct *tsk;
2775 struct cgroup_pidlist *l;
2778 * If cgroup gets more users after we read count, we won't have
2779 * enough space - tough. This race is indistinguishable to the
2780 * caller from the case that the additional cgroup users didn't
2781 * show up until sometime later on.
2783 length = cgroup_task_count(cgrp);
2784 array = pidlist_allocate(length);
2785 if (!array)
2786 return -ENOMEM;
2787 /* now, populate the array */
2788 cgroup_iter_start(cgrp, &it);
2789 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2790 if (unlikely(n == length))
2791 break;
2792 /* get tgid or pid for procs or tasks file respectively */
2793 if (type == CGROUP_FILE_PROCS)
2794 pid = task_tgid_vnr(tsk);
2795 else
2796 pid = task_pid_vnr(tsk);
2797 if (pid > 0) /* make sure to only use valid results */
2798 array[n++] = pid;
2800 cgroup_iter_end(cgrp, &it);
2801 length = n;
2802 /* now sort & (if procs) strip out duplicates */
2803 sort(array, length, sizeof(pid_t), cmppid, NULL);
2804 if (type == CGROUP_FILE_PROCS)
2805 length = pidlist_uniq(&array, length);
2806 l = cgroup_pidlist_find(cgrp, type);
2807 if (!l) {
2808 pidlist_free(array);
2809 return -ENOMEM;
2811 /* store array, freeing old if necessary - lock already held */
2812 pidlist_free(l->list);
2813 l->list = array;
2814 l->length = length;
2815 l->use_count++;
2816 up_write(&l->mutex);
2817 *lp = l;
2818 return 0;
2822 * cgroupstats_build - build and fill cgroupstats
2823 * @stats: cgroupstats to fill information into
2824 * @dentry: A dentry entry belonging to the cgroup for which stats have
2825 * been requested.
2827 * Build and fill cgroupstats so that taskstats can export it to user
2828 * space.
2830 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2832 int ret = -EINVAL;
2833 struct cgroup *cgrp;
2834 struct cgroup_iter it;
2835 struct task_struct *tsk;
2838 * Validate dentry by checking the superblock operations,
2839 * and make sure it's a directory.
2841 if (dentry->d_sb->s_op != &cgroup_ops ||
2842 !S_ISDIR(dentry->d_inode->i_mode))
2843 goto err;
2845 ret = 0;
2846 cgrp = dentry->d_fsdata;
2848 cgroup_iter_start(cgrp, &it);
2849 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2850 switch (tsk->state) {
2851 case TASK_RUNNING:
2852 stats->nr_running++;
2853 break;
2854 case TASK_INTERRUPTIBLE:
2855 stats->nr_sleeping++;
2856 break;
2857 case TASK_UNINTERRUPTIBLE:
2858 stats->nr_uninterruptible++;
2859 break;
2860 case TASK_STOPPED:
2861 stats->nr_stopped++;
2862 break;
2863 default:
2864 if (delayacct_is_task_waiting_on_io(tsk))
2865 stats->nr_io_wait++;
2866 break;
2869 cgroup_iter_end(cgrp, &it);
2871 err:
2872 return ret;
2877 * seq_file methods for the tasks/procs files. The seq_file position is the
2878 * next pid to display; the seq_file iterator is a pointer to the pid
2879 * in the cgroup->l->list array.
2882 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
2885 * Initially we receive a position value that corresponds to
2886 * one more than the last pid shown (or 0 on the first call or
2887 * after a seek to the start). Use a binary-search to find the
2888 * next pid to display, if any
2890 struct cgroup_pidlist *l = s->private;
2891 int index = 0, pid = *pos;
2892 int *iter;
2894 down_read(&l->mutex);
2895 if (pid) {
2896 int end = l->length;
2898 while (index < end) {
2899 int mid = (index + end) / 2;
2900 if (l->list[mid] == pid) {
2901 index = mid;
2902 break;
2903 } else if (l->list[mid] <= pid)
2904 index = mid + 1;
2905 else
2906 end = mid;
2909 /* If we're off the end of the array, we're done */
2910 if (index >= l->length)
2911 return NULL;
2912 /* Update the abstract position to be the actual pid that we found */
2913 iter = l->list + index;
2914 *pos = *iter;
2915 return iter;
2918 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
2920 struct cgroup_pidlist *l = s->private;
2921 up_read(&l->mutex);
2924 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
2926 struct cgroup_pidlist *l = s->private;
2927 pid_t *p = v;
2928 pid_t *end = l->list + l->length;
2930 * Advance to the next pid in the array. If this goes off the
2931 * end, we're done
2933 p++;
2934 if (p >= end) {
2935 return NULL;
2936 } else {
2937 *pos = *p;
2938 return p;
2942 static int cgroup_pidlist_show(struct seq_file *s, void *v)
2944 return seq_printf(s, "%d\n", *(int *)v);
2948 * seq_operations functions for iterating on pidlists through seq_file -
2949 * independent of whether it's tasks or procs
2951 static const struct seq_operations cgroup_pidlist_seq_operations = {
2952 .start = cgroup_pidlist_start,
2953 .stop = cgroup_pidlist_stop,
2954 .next = cgroup_pidlist_next,
2955 .show = cgroup_pidlist_show,
2958 static void cgroup_release_pid_array(struct cgroup_pidlist *l)
2961 * the case where we're the last user of this particular pidlist will
2962 * have us remove it from the cgroup's list, which entails taking the
2963 * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
2964 * pidlist_mutex, we have to take pidlist_mutex first.
2966 mutex_lock(&l->owner->pidlist_mutex);
2967 down_write(&l->mutex);
2968 BUG_ON(!l->use_count);
2969 if (!--l->use_count) {
2970 /* we're the last user if refcount is 0; remove and free */
2971 list_del(&l->links);
2972 mutex_unlock(&l->owner->pidlist_mutex);
2973 pidlist_free(l->list);
2974 put_pid_ns(l->key.ns);
2975 up_write(&l->mutex);
2976 kfree(l);
2977 return;
2979 mutex_unlock(&l->owner->pidlist_mutex);
2980 up_write(&l->mutex);
2983 static int cgroup_pidlist_release(struct inode *inode, struct file *file)
2985 struct cgroup_pidlist *l;
2986 if (!(file->f_mode & FMODE_READ))
2987 return 0;
2989 * the seq_file will only be initialized if the file was opened for
2990 * reading; hence we check if it's not null only in that case.
2992 l = ((struct seq_file *)file->private_data)->private;
2993 cgroup_release_pid_array(l);
2994 return seq_release(inode, file);
2997 static const struct file_operations cgroup_pidlist_operations = {
2998 .read = seq_read,
2999 .llseek = seq_lseek,
3000 .write = cgroup_file_write,
3001 .release = cgroup_pidlist_release,
3005 * The following functions handle opens on a file that displays a pidlist
3006 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
3007 * in the cgroup.
3009 /* helper function for the two below it */
3010 static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
3012 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
3013 struct cgroup_pidlist *l;
3014 int retval;
3016 /* Nothing to do for write-only files */
3017 if (!(file->f_mode & FMODE_READ))
3018 return 0;
3020 /* have the array populated */
3021 retval = pidlist_array_load(cgrp, type, &l);
3022 if (retval)
3023 return retval;
3024 /* configure file information */
3025 file->f_op = &cgroup_pidlist_operations;
3027 retval = seq_open(file, &cgroup_pidlist_seq_operations);
3028 if (retval) {
3029 cgroup_release_pid_array(l);
3030 return retval;
3032 ((struct seq_file *)file->private_data)->private = l;
3033 return 0;
3035 static int cgroup_tasks_open(struct inode *unused, struct file *file)
3037 return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
3039 static int cgroup_procs_open(struct inode *unused, struct file *file)
3041 return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
3044 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
3045 struct cftype *cft)
3047 return notify_on_release(cgrp);
3050 static int cgroup_write_notify_on_release(struct cgroup *cgrp,
3051 struct cftype *cft,
3052 u64 val)
3054 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
3055 if (val)
3056 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3057 else
3058 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3059 return 0;
3063 * Unregister event and free resources.
3065 * Gets called from workqueue.
3067 static void cgroup_event_remove(struct work_struct *work)
3069 struct cgroup_event *event = container_of(work, struct cgroup_event,
3070 remove);
3071 struct cgroup *cgrp = event->cgrp;
3073 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3075 eventfd_ctx_put(event->eventfd);
3076 kfree(event);
3077 dput(cgrp->dentry);
3081 * Gets called on POLLHUP on eventfd when user closes it.
3083 * Called with wqh->lock held and interrupts disabled.
3085 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3086 int sync, void *key)
3088 struct cgroup_event *event = container_of(wait,
3089 struct cgroup_event, wait);
3090 struct cgroup *cgrp = event->cgrp;
3091 unsigned long flags = (unsigned long)key;
3093 if (flags & POLLHUP) {
3094 __remove_wait_queue(event->wqh, &event->wait);
3095 spin_lock(&cgrp->event_list_lock);
3096 list_del(&event->list);
3097 spin_unlock(&cgrp->event_list_lock);
3099 * We are in atomic context, but cgroup_event_remove() may
3100 * sleep, so we have to call it in workqueue.
3102 schedule_work(&event->remove);
3105 return 0;
3108 static void cgroup_event_ptable_queue_proc(struct file *file,
3109 wait_queue_head_t *wqh, poll_table *pt)
3111 struct cgroup_event *event = container_of(pt,
3112 struct cgroup_event, pt);
3114 event->wqh = wqh;
3115 add_wait_queue(wqh, &event->wait);
3119 * Parse input and register new cgroup event handler.
3121 * Input must be in format '<event_fd> <control_fd> <args>'.
3122 * Interpretation of args is defined by control file implementation.
3124 static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3125 const char *buffer)
3127 struct cgroup_event *event = NULL;
3128 unsigned int efd, cfd;
3129 struct file *efile = NULL;
3130 struct file *cfile = NULL;
3131 char *endp;
3132 int ret;
3134 efd = simple_strtoul(buffer, &endp, 10);
3135 if (*endp != ' ')
3136 return -EINVAL;
3137 buffer = endp + 1;
3139 cfd = simple_strtoul(buffer, &endp, 10);
3140 if ((*endp != ' ') && (*endp != '\0'))
3141 return -EINVAL;
3142 buffer = endp + 1;
3144 event = kzalloc(sizeof(*event), GFP_KERNEL);
3145 if (!event)
3146 return -ENOMEM;
3147 event->cgrp = cgrp;
3148 INIT_LIST_HEAD(&event->list);
3149 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
3150 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
3151 INIT_WORK(&event->remove, cgroup_event_remove);
3153 efile = eventfd_fget(efd);
3154 if (IS_ERR(efile)) {
3155 ret = PTR_ERR(efile);
3156 goto fail;
3159 event->eventfd = eventfd_ctx_fileget(efile);
3160 if (IS_ERR(event->eventfd)) {
3161 ret = PTR_ERR(event->eventfd);
3162 goto fail;
3165 cfile = fget(cfd);
3166 if (!cfile) {
3167 ret = -EBADF;
3168 goto fail;
3171 /* the process need read permission on control file */
3172 ret = file_permission(cfile, MAY_READ);
3173 if (ret < 0)
3174 goto fail;
3176 event->cft = __file_cft(cfile);
3177 if (IS_ERR(event->cft)) {
3178 ret = PTR_ERR(event->cft);
3179 goto fail;
3182 if (!event->cft->register_event || !event->cft->unregister_event) {
3183 ret = -EINVAL;
3184 goto fail;
3187 ret = event->cft->register_event(cgrp, event->cft,
3188 event->eventfd, buffer);
3189 if (ret)
3190 goto fail;
3192 if (efile->f_op->poll(efile, &event->pt) & POLLHUP) {
3193 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3194 ret = 0;
3195 goto fail;
3199 * Events should be removed after rmdir of cgroup directory, but before
3200 * destroying subsystem state objects. Let's take reference to cgroup
3201 * directory dentry to do that.
3203 dget(cgrp->dentry);
3205 spin_lock(&cgrp->event_list_lock);
3206 list_add(&event->list, &cgrp->event_list);
3207 spin_unlock(&cgrp->event_list_lock);
3209 fput(cfile);
3210 fput(efile);
3212 return 0;
3214 fail:
3215 if (cfile)
3216 fput(cfile);
3218 if (event && event->eventfd && !IS_ERR(event->eventfd))
3219 eventfd_ctx_put(event->eventfd);
3221 if (!IS_ERR_OR_NULL(efile))
3222 fput(efile);
3224 kfree(event);
3226 return ret;
3229 static u64 cgroup_clone_children_read(struct cgroup *cgrp,
3230 struct cftype *cft)
3232 return clone_children(cgrp);
3235 static int cgroup_clone_children_write(struct cgroup *cgrp,
3236 struct cftype *cft,
3237 u64 val)
3239 if (val)
3240 set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3241 else
3242 clear_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3243 return 0;
3247 * for the common functions, 'private' gives the type of file
3249 /* for hysterical raisins, we can't put this on the older files */
3250 #define CGROUP_FILE_GENERIC_PREFIX "cgroup."
3251 static struct cftype files[] = {
3253 .name = "tasks",
3254 .open = cgroup_tasks_open,
3255 .write_u64 = cgroup_tasks_write,
3256 .release = cgroup_pidlist_release,
3257 .mode = S_IRUGO | S_IWUSR,
3260 .name = CGROUP_FILE_GENERIC_PREFIX "procs",
3261 .open = cgroup_procs_open,
3262 /* .write_u64 = cgroup_procs_write, TODO */
3263 .release = cgroup_pidlist_release,
3264 .mode = S_IRUGO,
3267 .name = "notify_on_release",
3268 .read_u64 = cgroup_read_notify_on_release,
3269 .write_u64 = cgroup_write_notify_on_release,
3272 .name = CGROUP_FILE_GENERIC_PREFIX "event_control",
3273 .write_string = cgroup_write_event_control,
3274 .mode = S_IWUGO,
3277 .name = "cgroup.clone_children",
3278 .read_u64 = cgroup_clone_children_read,
3279 .write_u64 = cgroup_clone_children_write,
3283 static struct cftype cft_release_agent = {
3284 .name = "release_agent",
3285 .read_seq_string = cgroup_release_agent_show,
3286 .write_string = cgroup_release_agent_write,
3287 .max_write_len = PATH_MAX,
3290 static int cgroup_populate_dir(struct cgroup *cgrp)
3292 int err;
3293 struct cgroup_subsys *ss;
3295 /* First clear out any existing files */
3296 cgroup_clear_directory(cgrp->dentry);
3298 err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
3299 if (err < 0)
3300 return err;
3302 if (cgrp == cgrp->top_cgroup) {
3303 if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
3304 return err;
3307 for_each_subsys(cgrp->root, ss) {
3308 if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
3309 return err;
3311 /* This cgroup is ready now */
3312 for_each_subsys(cgrp->root, ss) {
3313 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3315 * Update id->css pointer and make this css visible from
3316 * CSS ID functions. This pointer will be dereferened
3317 * from RCU-read-side without locks.
3319 if (css->id)
3320 rcu_assign_pointer(css->id->css, css);
3323 return 0;
3326 static void init_cgroup_css(struct cgroup_subsys_state *css,
3327 struct cgroup_subsys *ss,
3328 struct cgroup *cgrp)
3330 css->cgroup = cgrp;
3331 atomic_set(&css->refcnt, 1);
3332 css->flags = 0;
3333 css->id = NULL;
3334 if (cgrp == dummytop)
3335 set_bit(CSS_ROOT, &css->flags);
3336 BUG_ON(cgrp->subsys[ss->subsys_id]);
3337 cgrp->subsys[ss->subsys_id] = css;
3340 static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
3342 /* We need to take each hierarchy_mutex in a consistent order */
3343 int i;
3346 * No worry about a race with rebind_subsystems that might mess up the
3347 * locking order, since both parties are under cgroup_mutex.
3349 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3350 struct cgroup_subsys *ss = subsys[i];
3351 if (ss == NULL)
3352 continue;
3353 if (ss->root == root)
3354 mutex_lock(&ss->hierarchy_mutex);
3358 static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
3360 int i;
3362 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3363 struct cgroup_subsys *ss = subsys[i];
3364 if (ss == NULL)
3365 continue;
3366 if (ss->root == root)
3367 mutex_unlock(&ss->hierarchy_mutex);
3372 * cgroup_create - create a cgroup
3373 * @parent: cgroup that will be parent of the new cgroup
3374 * @dentry: dentry of the new cgroup
3375 * @mode: mode to set on new inode
3377 * Must be called with the mutex on the parent inode held
3379 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
3380 mode_t mode)
3382 struct cgroup *cgrp;
3383 struct cgroupfs_root *root = parent->root;
3384 int err = 0;
3385 struct cgroup_subsys *ss;
3386 struct super_block *sb = root->sb;
3388 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3389 if (!cgrp)
3390 return -ENOMEM;
3392 /* Grab a reference on the superblock so the hierarchy doesn't
3393 * get deleted on unmount if there are child cgroups. This
3394 * can be done outside cgroup_mutex, since the sb can't
3395 * disappear while someone has an open control file on the
3396 * fs */
3397 atomic_inc(&sb->s_active);
3399 mutex_lock(&cgroup_mutex);
3401 init_cgroup_housekeeping(cgrp);
3403 cgrp->parent = parent;
3404 cgrp->root = parent->root;
3405 cgrp->top_cgroup = parent->top_cgroup;
3407 if (notify_on_release(parent))
3408 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3410 if (clone_children(parent))
3411 set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3413 for_each_subsys(root, ss) {
3414 struct cgroup_subsys_state *css = ss->create(ss, cgrp);
3416 if (IS_ERR(css)) {
3417 err = PTR_ERR(css);
3418 goto err_destroy;
3420 init_cgroup_css(css, ss, cgrp);
3421 if (ss->use_id) {
3422 err = alloc_css_id(ss, parent, cgrp);
3423 if (err)
3424 goto err_destroy;
3426 /* At error, ->destroy() callback has to free assigned ID. */
3427 if (clone_children(parent) && ss->post_clone)
3428 ss->post_clone(ss, cgrp);
3431 cgroup_lock_hierarchy(root);
3432 list_add(&cgrp->sibling, &cgrp->parent->children);
3433 cgroup_unlock_hierarchy(root);
3434 root->number_of_cgroups++;
3436 err = cgroup_create_dir(cgrp, dentry, mode);
3437 if (err < 0)
3438 goto err_remove;
3440 /* The cgroup directory was pre-locked for us */
3441 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
3443 err = cgroup_populate_dir(cgrp);
3444 /* If err < 0, we have a half-filled directory - oh well ;) */
3446 mutex_unlock(&cgroup_mutex);
3447 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
3449 return 0;
3451 err_remove:
3453 cgroup_lock_hierarchy(root);
3454 list_del(&cgrp->sibling);
3455 cgroup_unlock_hierarchy(root);
3456 root->number_of_cgroups--;
3458 err_destroy:
3460 for_each_subsys(root, ss) {
3461 if (cgrp->subsys[ss->subsys_id])
3462 ss->destroy(ss, cgrp);
3465 mutex_unlock(&cgroup_mutex);
3467 /* Release the reference count that we took on the superblock */
3468 deactivate_super(sb);
3470 kfree(cgrp);
3471 return err;
3474 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
3476 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
3478 /* the vfs holds inode->i_mutex already */
3479 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
3482 static int cgroup_has_css_refs(struct cgroup *cgrp)
3484 /* Check the reference count on each subsystem. Since we
3485 * already established that there are no tasks in the
3486 * cgroup, if the css refcount is also 1, then there should
3487 * be no outstanding references, so the subsystem is safe to
3488 * destroy. We scan across all subsystems rather than using
3489 * the per-hierarchy linked list of mounted subsystems since
3490 * we can be called via check_for_release() with no
3491 * synchronization other than RCU, and the subsystem linked
3492 * list isn't RCU-safe */
3493 int i;
3495 * We won't need to lock the subsys array, because the subsystems
3496 * we're concerned about aren't going anywhere since our cgroup root
3497 * has a reference on them.
3499 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3500 struct cgroup_subsys *ss = subsys[i];
3501 struct cgroup_subsys_state *css;
3502 /* Skip subsystems not present or not in this hierarchy */
3503 if (ss == NULL || ss->root != cgrp->root)
3504 continue;
3505 css = cgrp->subsys[ss->subsys_id];
3506 /* When called from check_for_release() it's possible
3507 * that by this point the cgroup has been removed
3508 * and the css deleted. But a false-positive doesn't
3509 * matter, since it can only happen if the cgroup
3510 * has been deleted and hence no longer needs the
3511 * release agent to be called anyway. */
3512 if (css && (atomic_read(&css->refcnt) > 1))
3513 return 1;
3515 return 0;
3519 * Atomically mark all (or else none) of the cgroup's CSS objects as
3520 * CSS_REMOVED. Return true on success, or false if the cgroup has
3521 * busy subsystems. Call with cgroup_mutex held
3524 static int cgroup_clear_css_refs(struct cgroup *cgrp)
3526 struct cgroup_subsys *ss;
3527 unsigned long flags;
3528 bool failed = false;
3529 local_irq_save(flags);
3530 for_each_subsys(cgrp->root, ss) {
3531 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3532 int refcnt;
3533 while (1) {
3534 /* We can only remove a CSS with a refcnt==1 */
3535 refcnt = atomic_read(&css->refcnt);
3536 if (refcnt > 1) {
3537 failed = true;
3538 goto done;
3540 BUG_ON(!refcnt);
3542 * Drop the refcnt to 0 while we check other
3543 * subsystems. This will cause any racing
3544 * css_tryget() to spin until we set the
3545 * CSS_REMOVED bits or abort
3547 if (atomic_cmpxchg(&css->refcnt, refcnt, 0) == refcnt)
3548 break;
3549 cpu_relax();
3552 done:
3553 for_each_subsys(cgrp->root, ss) {
3554 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3555 if (failed) {
3557 * Restore old refcnt if we previously managed
3558 * to clear it from 1 to 0
3560 if (!atomic_read(&css->refcnt))
3561 atomic_set(&css->refcnt, 1);
3562 } else {
3563 /* Commit the fact that the CSS is removed */
3564 set_bit(CSS_REMOVED, &css->flags);
3567 local_irq_restore(flags);
3568 return !failed;
3571 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
3573 struct cgroup *cgrp = dentry->d_fsdata;
3574 struct dentry *d;
3575 struct cgroup *parent;
3576 DEFINE_WAIT(wait);
3577 struct cgroup_event *event, *tmp;
3578 int ret;
3580 /* the vfs holds both inode->i_mutex already */
3581 again:
3582 mutex_lock(&cgroup_mutex);
3583 if (atomic_read(&cgrp->count) != 0) {
3584 mutex_unlock(&cgroup_mutex);
3585 return -EBUSY;
3587 if (!list_empty(&cgrp->children)) {
3588 mutex_unlock(&cgroup_mutex);
3589 return -EBUSY;
3591 mutex_unlock(&cgroup_mutex);
3594 * In general, subsystem has no css->refcnt after pre_destroy(). But
3595 * in racy cases, subsystem may have to get css->refcnt after
3596 * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
3597 * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
3598 * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
3599 * and subsystem's reference count handling. Please see css_get/put
3600 * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
3602 set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3605 * Call pre_destroy handlers of subsys. Notify subsystems
3606 * that rmdir() request comes.
3608 ret = cgroup_call_pre_destroy(cgrp);
3609 if (ret) {
3610 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3611 return ret;
3614 mutex_lock(&cgroup_mutex);
3615 parent = cgrp->parent;
3616 if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
3617 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3618 mutex_unlock(&cgroup_mutex);
3619 return -EBUSY;
3621 prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
3622 if (!cgroup_clear_css_refs(cgrp)) {
3623 mutex_unlock(&cgroup_mutex);
3625 * Because someone may call cgroup_wakeup_rmdir_waiter() before
3626 * prepare_to_wait(), we need to check this flag.
3628 if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
3629 schedule();
3630 finish_wait(&cgroup_rmdir_waitq, &wait);
3631 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3632 if (signal_pending(current))
3633 return -EINTR;
3634 goto again;
3636 /* NO css_tryget() can success after here. */
3637 finish_wait(&cgroup_rmdir_waitq, &wait);
3638 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3640 spin_lock(&release_list_lock);
3641 set_bit(CGRP_REMOVED, &cgrp->flags);
3642 if (!list_empty(&cgrp->release_list))
3643 list_del_init(&cgrp->release_list);
3644 spin_unlock(&release_list_lock);
3646 cgroup_lock_hierarchy(cgrp->root);
3647 /* delete this cgroup from parent->children */
3648 list_del_init(&cgrp->sibling);
3649 cgroup_unlock_hierarchy(cgrp->root);
3651 d = dget(cgrp->dentry);
3653 cgroup_d_remove_dir(d);
3654 dput(d);
3656 set_bit(CGRP_RELEASABLE, &parent->flags);
3657 check_for_release(parent);
3660 * Unregister events and notify userspace.
3661 * Notify userspace about cgroup removing only after rmdir of cgroup
3662 * directory to avoid race between userspace and kernelspace
3664 spin_lock(&cgrp->event_list_lock);
3665 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
3666 list_del(&event->list);
3667 remove_wait_queue(event->wqh, &event->wait);
3668 eventfd_signal(event->eventfd, 1);
3669 schedule_work(&event->remove);
3671 spin_unlock(&cgrp->event_list_lock);
3673 mutex_unlock(&cgroup_mutex);
3674 return 0;
3677 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
3679 struct cgroup_subsys_state *css;
3681 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
3683 /* Create the top cgroup state for this subsystem */
3684 list_add(&ss->sibling, &rootnode.subsys_list);
3685 ss->root = &rootnode;
3686 css = ss->create(ss, dummytop);
3687 /* We don't handle early failures gracefully */
3688 BUG_ON(IS_ERR(css));
3689 init_cgroup_css(css, ss, dummytop);
3691 /* Update the init_css_set to contain a subsys
3692 * pointer to this state - since the subsystem is
3693 * newly registered, all tasks and hence the
3694 * init_css_set is in the subsystem's top cgroup. */
3695 init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
3697 need_forkexit_callback |= ss->fork || ss->exit;
3699 /* At system boot, before all subsystems have been
3700 * registered, no tasks have been forked, so we don't
3701 * need to invoke fork callbacks here. */
3702 BUG_ON(!list_empty(&init_task.tasks));
3704 mutex_init(&ss->hierarchy_mutex);
3705 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3706 ss->active = 1;
3708 /* this function shouldn't be used with modular subsystems, since they
3709 * need to register a subsys_id, among other things */
3710 BUG_ON(ss->module);
3714 * cgroup_load_subsys: load and register a modular subsystem at runtime
3715 * @ss: the subsystem to load
3717 * This function should be called in a modular subsystem's initcall. If the
3718 * subsystem is built as a module, it will be assigned a new subsys_id and set
3719 * up for use. If the subsystem is built-in anyway, work is delegated to the
3720 * simpler cgroup_init_subsys.
3722 int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
3724 int i;
3725 struct cgroup_subsys_state *css;
3727 /* check name and function validity */
3728 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
3729 ss->create == NULL || ss->destroy == NULL)
3730 return -EINVAL;
3733 * we don't support callbacks in modular subsystems. this check is
3734 * before the ss->module check for consistency; a subsystem that could
3735 * be a module should still have no callbacks even if the user isn't
3736 * compiling it as one.
3738 if (ss->fork || ss->exit)
3739 return -EINVAL;
3742 * an optionally modular subsystem is built-in: we want to do nothing,
3743 * since cgroup_init_subsys will have already taken care of it.
3745 if (ss->module == NULL) {
3746 /* a few sanity checks */
3747 BUG_ON(ss->subsys_id >= CGROUP_BUILTIN_SUBSYS_COUNT);
3748 BUG_ON(subsys[ss->subsys_id] != ss);
3749 return 0;
3753 * need to register a subsys id before anything else - for example,
3754 * init_cgroup_css needs it.
3756 mutex_lock(&cgroup_mutex);
3757 /* find the first empty slot in the array */
3758 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
3759 if (subsys[i] == NULL)
3760 break;
3762 if (i == CGROUP_SUBSYS_COUNT) {
3763 /* maximum number of subsystems already registered! */
3764 mutex_unlock(&cgroup_mutex);
3765 return -EBUSY;
3767 /* assign ourselves the subsys_id */
3768 ss->subsys_id = i;
3769 subsys[i] = ss;
3772 * no ss->create seems to need anything important in the ss struct, so
3773 * this can happen first (i.e. before the rootnode attachment).
3775 css = ss->create(ss, dummytop);
3776 if (IS_ERR(css)) {
3777 /* failure case - need to deassign the subsys[] slot. */
3778 subsys[i] = NULL;
3779 mutex_unlock(&cgroup_mutex);
3780 return PTR_ERR(css);
3783 list_add(&ss->sibling, &rootnode.subsys_list);
3784 ss->root = &rootnode;
3786 /* our new subsystem will be attached to the dummy hierarchy. */
3787 init_cgroup_css(css, ss, dummytop);
3788 /* init_idr must be after init_cgroup_css because it sets css->id. */
3789 if (ss->use_id) {
3790 int ret = cgroup_init_idr(ss, css);
3791 if (ret) {
3792 dummytop->subsys[ss->subsys_id] = NULL;
3793 ss->destroy(ss, dummytop);
3794 subsys[i] = NULL;
3795 mutex_unlock(&cgroup_mutex);
3796 return ret;
3801 * Now we need to entangle the css into the existing css_sets. unlike
3802 * in cgroup_init_subsys, there are now multiple css_sets, so each one
3803 * will need a new pointer to it; done by iterating the css_set_table.
3804 * furthermore, modifying the existing css_sets will corrupt the hash
3805 * table state, so each changed css_set will need its hash recomputed.
3806 * this is all done under the css_set_lock.
3808 write_lock(&css_set_lock);
3809 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
3810 struct css_set *cg;
3811 struct hlist_node *node, *tmp;
3812 struct hlist_head *bucket = &css_set_table[i], *new_bucket;
3814 hlist_for_each_entry_safe(cg, node, tmp, bucket, hlist) {
3815 /* skip entries that we already rehashed */
3816 if (cg->subsys[ss->subsys_id])
3817 continue;
3818 /* remove existing entry */
3819 hlist_del(&cg->hlist);
3820 /* set new value */
3821 cg->subsys[ss->subsys_id] = css;
3822 /* recompute hash and restore entry */
3823 new_bucket = css_set_hash(cg->subsys);
3824 hlist_add_head(&cg->hlist, new_bucket);
3827 write_unlock(&css_set_lock);
3829 mutex_init(&ss->hierarchy_mutex);
3830 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3831 ss->active = 1;
3833 /* success! */
3834 mutex_unlock(&cgroup_mutex);
3835 return 0;
3837 EXPORT_SYMBOL_GPL(cgroup_load_subsys);
3840 * cgroup_unload_subsys: unload a modular subsystem
3841 * @ss: the subsystem to unload
3843 * This function should be called in a modular subsystem's exitcall. When this
3844 * function is invoked, the refcount on the subsystem's module will be 0, so
3845 * the subsystem will not be attached to any hierarchy.
3847 void cgroup_unload_subsys(struct cgroup_subsys *ss)
3849 struct cg_cgroup_link *link;
3850 struct hlist_head *hhead;
3852 BUG_ON(ss->module == NULL);
3855 * we shouldn't be called if the subsystem is in use, and the use of
3856 * try_module_get in parse_cgroupfs_options should ensure that it
3857 * doesn't start being used while we're killing it off.
3859 BUG_ON(ss->root != &rootnode);
3861 mutex_lock(&cgroup_mutex);
3862 /* deassign the subsys_id */
3863 BUG_ON(ss->subsys_id < CGROUP_BUILTIN_SUBSYS_COUNT);
3864 subsys[ss->subsys_id] = NULL;
3866 /* remove subsystem from rootnode's list of subsystems */
3867 list_del_init(&ss->sibling);
3870 * disentangle the css from all css_sets attached to the dummytop. as
3871 * in loading, we need to pay our respects to the hashtable gods.
3873 write_lock(&css_set_lock);
3874 list_for_each_entry(link, &dummytop->css_sets, cgrp_link_list) {
3875 struct css_set *cg = link->cg;
3877 hlist_del(&cg->hlist);
3878 BUG_ON(!cg->subsys[ss->subsys_id]);
3879 cg->subsys[ss->subsys_id] = NULL;
3880 hhead = css_set_hash(cg->subsys);
3881 hlist_add_head(&cg->hlist, hhead);
3883 write_unlock(&css_set_lock);
3886 * remove subsystem's css from the dummytop and free it - need to free
3887 * before marking as null because ss->destroy needs the cgrp->subsys
3888 * pointer to find their state. note that this also takes care of
3889 * freeing the css_id.
3891 ss->destroy(ss, dummytop);
3892 dummytop->subsys[ss->subsys_id] = NULL;
3894 mutex_unlock(&cgroup_mutex);
3896 EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
3899 * cgroup_init_early - cgroup initialization at system boot
3901 * Initialize cgroups at system boot, and initialize any
3902 * subsystems that request early init.
3904 int __init cgroup_init_early(void)
3906 int i;
3907 atomic_set(&init_css_set.refcount, 1);
3908 INIT_LIST_HEAD(&init_css_set.cg_links);
3909 INIT_LIST_HEAD(&init_css_set.tasks);
3910 INIT_HLIST_NODE(&init_css_set.hlist);
3911 css_set_count = 1;
3912 init_cgroup_root(&rootnode);
3913 root_count = 1;
3914 init_task.cgroups = &init_css_set;
3916 init_css_set_link.cg = &init_css_set;
3917 init_css_set_link.cgrp = dummytop;
3918 list_add(&init_css_set_link.cgrp_link_list,
3919 &rootnode.top_cgroup.css_sets);
3920 list_add(&init_css_set_link.cg_link_list,
3921 &init_css_set.cg_links);
3923 for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
3924 INIT_HLIST_HEAD(&css_set_table[i]);
3926 /* at bootup time, we don't worry about modular subsystems */
3927 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3928 struct cgroup_subsys *ss = subsys[i];
3930 BUG_ON(!ss->name);
3931 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
3932 BUG_ON(!ss->create);
3933 BUG_ON(!ss->destroy);
3934 if (ss->subsys_id != i) {
3935 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
3936 ss->name, ss->subsys_id);
3937 BUG();
3940 if (ss->early_init)
3941 cgroup_init_subsys(ss);
3943 return 0;
3947 * cgroup_init - cgroup initialization
3949 * Register cgroup filesystem and /proc file, and initialize
3950 * any subsystems that didn't request early init.
3952 int __init cgroup_init(void)
3954 int err;
3955 int i;
3956 struct hlist_head *hhead;
3958 err = bdi_init(&cgroup_backing_dev_info);
3959 if (err)
3960 return err;
3962 /* at bootup time, we don't worry about modular subsystems */
3963 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3964 struct cgroup_subsys *ss = subsys[i];
3965 if (!ss->early_init)
3966 cgroup_init_subsys(ss);
3967 if (ss->use_id)
3968 cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
3971 /* Add init_css_set to the hash table */
3972 hhead = css_set_hash(init_css_set.subsys);
3973 hlist_add_head(&init_css_set.hlist, hhead);
3974 BUG_ON(!init_root_id(&rootnode));
3976 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
3977 if (!cgroup_kobj) {
3978 err = -ENOMEM;
3979 goto out;
3982 err = register_filesystem(&cgroup_fs_type);
3983 if (err < 0) {
3984 kobject_put(cgroup_kobj);
3985 goto out;
3988 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
3990 out:
3991 if (err)
3992 bdi_destroy(&cgroup_backing_dev_info);
3994 return err;
3998 * proc_cgroup_show()
3999 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4000 * - Used for /proc/<pid>/cgroup.
4001 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4002 * doesn't really matter if tsk->cgroup changes after we read it,
4003 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4004 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4005 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4006 * cgroup to top_cgroup.
4009 /* TODO: Use a proper seq_file iterator */
4010 static int proc_cgroup_show(struct seq_file *m, void *v)
4012 struct pid *pid;
4013 struct task_struct *tsk;
4014 char *buf;
4015 int retval;
4016 struct cgroupfs_root *root;
4018 retval = -ENOMEM;
4019 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4020 if (!buf)
4021 goto out;
4023 retval = -ESRCH;
4024 pid = m->private;
4025 tsk = get_pid_task(pid, PIDTYPE_PID);
4026 if (!tsk)
4027 goto out_free;
4029 retval = 0;
4031 mutex_lock(&cgroup_mutex);
4033 for_each_active_root(root) {
4034 struct cgroup_subsys *ss;
4035 struct cgroup *cgrp;
4036 int count = 0;
4038 seq_printf(m, "%d:", root->hierarchy_id);
4039 for_each_subsys(root, ss)
4040 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4041 if (strlen(root->name))
4042 seq_printf(m, "%sname=%s", count ? "," : "",
4043 root->name);
4044 seq_putc(m, ':');
4045 cgrp = task_cgroup_from_root(tsk, root);
4046 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4047 if (retval < 0)
4048 goto out_unlock;
4049 seq_puts(m, buf);
4050 seq_putc(m, '\n');
4053 out_unlock:
4054 mutex_unlock(&cgroup_mutex);
4055 put_task_struct(tsk);
4056 out_free:
4057 kfree(buf);
4058 out:
4059 return retval;
4062 static int cgroup_open(struct inode *inode, struct file *file)
4064 struct pid *pid = PROC_I(inode)->pid;
4065 return single_open(file, proc_cgroup_show, pid);
4068 const struct file_operations proc_cgroup_operations = {
4069 .open = cgroup_open,
4070 .read = seq_read,
4071 .llseek = seq_lseek,
4072 .release = single_release,
4075 /* Display information about each subsystem and each hierarchy */
4076 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4078 int i;
4080 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4082 * ideally we don't want subsystems moving around while we do this.
4083 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4084 * subsys/hierarchy state.
4086 mutex_lock(&cgroup_mutex);
4087 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
4088 struct cgroup_subsys *ss = subsys[i];
4089 if (ss == NULL)
4090 continue;
4091 seq_printf(m, "%s\t%d\t%d\t%d\n",
4092 ss->name, ss->root->hierarchy_id,
4093 ss->root->number_of_cgroups, !ss->disabled);
4095 mutex_unlock(&cgroup_mutex);
4096 return 0;
4099 static int cgroupstats_open(struct inode *inode, struct file *file)
4101 return single_open(file, proc_cgroupstats_show, NULL);
4104 static const struct file_operations proc_cgroupstats_operations = {
4105 .open = cgroupstats_open,
4106 .read = seq_read,
4107 .llseek = seq_lseek,
4108 .release = single_release,
4112 * cgroup_fork - attach newly forked task to its parents cgroup.
4113 * @child: pointer to task_struct of forking parent process.
4115 * Description: A task inherits its parent's cgroup at fork().
4117 * A pointer to the shared css_set was automatically copied in
4118 * fork.c by dup_task_struct(). However, we ignore that copy, since
4119 * it was not made under the protection of RCU or cgroup_mutex, so
4120 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4121 * have already changed current->cgroups, allowing the previously
4122 * referenced cgroup group to be removed and freed.
4124 * At the point that cgroup_fork() is called, 'current' is the parent
4125 * task, and the passed argument 'child' points to the child task.
4127 void cgroup_fork(struct task_struct *child)
4129 task_lock(current);
4130 child->cgroups = current->cgroups;
4131 get_css_set(child->cgroups);
4132 task_unlock(current);
4133 INIT_LIST_HEAD(&child->cg_list);
4137 * cgroup_fork_callbacks - run fork callbacks
4138 * @child: the new task
4140 * Called on a new task very soon before adding it to the
4141 * tasklist. No need to take any locks since no-one can
4142 * be operating on this task.
4144 void cgroup_fork_callbacks(struct task_struct *child)
4146 if (need_forkexit_callback) {
4147 int i;
4149 * forkexit callbacks are only supported for builtin
4150 * subsystems, and the builtin section of the subsys array is
4151 * immutable, so we don't need to lock the subsys array here.
4153 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4154 struct cgroup_subsys *ss = subsys[i];
4155 if (ss->fork)
4156 ss->fork(ss, child);
4162 * cgroup_post_fork - called on a new task after adding it to the task list
4163 * @child: the task in question
4165 * Adds the task to the list running through its css_set if necessary.
4166 * Has to be after the task is visible on the task list in case we race
4167 * with the first call to cgroup_iter_start() - to guarantee that the
4168 * new task ends up on its list.
4170 void cgroup_post_fork(struct task_struct *child)
4172 if (use_task_css_set_links) {
4173 write_lock(&css_set_lock);
4174 task_lock(child);
4175 if (list_empty(&child->cg_list))
4176 list_add(&child->cg_list, &child->cgroups->tasks);
4177 task_unlock(child);
4178 write_unlock(&css_set_lock);
4182 * cgroup_exit - detach cgroup from exiting task
4183 * @tsk: pointer to task_struct of exiting process
4184 * @run_callback: run exit callbacks?
4186 * Description: Detach cgroup from @tsk and release it.
4188 * Note that cgroups marked notify_on_release force every task in
4189 * them to take the global cgroup_mutex mutex when exiting.
4190 * This could impact scaling on very large systems. Be reluctant to
4191 * use notify_on_release cgroups where very high task exit scaling
4192 * is required on large systems.
4194 * the_top_cgroup_hack:
4196 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4198 * We call cgroup_exit() while the task is still competent to
4199 * handle notify_on_release(), then leave the task attached to the
4200 * root cgroup in each hierarchy for the remainder of its exit.
4202 * To do this properly, we would increment the reference count on
4203 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4204 * code we would add a second cgroup function call, to drop that
4205 * reference. This would just create an unnecessary hot spot on
4206 * the top_cgroup reference count, to no avail.
4208 * Normally, holding a reference to a cgroup without bumping its
4209 * count is unsafe. The cgroup could go away, or someone could
4210 * attach us to a different cgroup, decrementing the count on
4211 * the first cgroup that we never incremented. But in this case,
4212 * top_cgroup isn't going away, and either task has PF_EXITING set,
4213 * which wards off any cgroup_attach_task() attempts, or task is a failed
4214 * fork, never visible to cgroup_attach_task.
4216 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4218 struct css_set *cg;
4219 int i;
4222 * Unlink from the css_set task list if necessary.
4223 * Optimistically check cg_list before taking
4224 * css_set_lock
4226 if (!list_empty(&tsk->cg_list)) {
4227 write_lock(&css_set_lock);
4228 if (!list_empty(&tsk->cg_list))
4229 list_del_init(&tsk->cg_list);
4230 write_unlock(&css_set_lock);
4233 /* Reassign the task to the init_css_set. */
4234 task_lock(tsk);
4235 cg = tsk->cgroups;
4236 tsk->cgroups = &init_css_set;
4238 if (run_callbacks && need_forkexit_callback) {
4240 * modular subsystems can't use callbacks, so no need to lock
4241 * the subsys array
4243 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4244 struct cgroup_subsys *ss = subsys[i];
4245 if (ss->exit) {
4246 struct cgroup *old_cgrp =
4247 rcu_dereference_raw(cg->subsys[i])->cgroup;
4248 struct cgroup *cgrp = task_cgroup(tsk, i);
4249 ss->exit(ss, cgrp, old_cgrp, tsk);
4253 task_unlock(tsk);
4255 if (cg)
4256 put_css_set_taskexit(cg);
4260 * cgroup_clone - clone the cgroup the given subsystem is attached to
4261 * @tsk: the task to be moved
4262 * @subsys: the given subsystem
4263 * @nodename: the name for the new cgroup
4265 * Duplicate the current cgroup in the hierarchy that the given
4266 * subsystem is attached to, and move this task into the new
4267 * child.
4269 int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
4270 char *nodename)
4272 struct dentry *dentry;
4273 int ret = 0;
4274 struct cgroup *parent, *child;
4275 struct inode *inode;
4276 struct css_set *cg;
4277 struct cgroupfs_root *root;
4278 struct cgroup_subsys *ss;
4280 /* We shouldn't be called by an unregistered subsystem */
4281 BUG_ON(!subsys->active);
4283 /* First figure out what hierarchy and cgroup we're dealing
4284 * with, and pin them so we can drop cgroup_mutex */
4285 mutex_lock(&cgroup_mutex);
4286 again:
4287 root = subsys->root;
4288 if (root == &rootnode) {
4289 mutex_unlock(&cgroup_mutex);
4290 return 0;
4293 /* Pin the hierarchy */
4294 if (!atomic_inc_not_zero(&root->sb->s_active)) {
4295 /* We race with the final deactivate_super() */
4296 mutex_unlock(&cgroup_mutex);
4297 return 0;
4300 /* Keep the cgroup alive */
4301 task_lock(tsk);
4302 parent = task_cgroup(tsk, subsys->subsys_id);
4303 cg = tsk->cgroups;
4304 get_css_set(cg);
4305 task_unlock(tsk);
4307 mutex_unlock(&cgroup_mutex);
4309 /* Now do the VFS work to create a cgroup */
4310 inode = parent->dentry->d_inode;
4312 /* Hold the parent directory mutex across this operation to
4313 * stop anyone else deleting the new cgroup */
4314 mutex_lock(&inode->i_mutex);
4315 dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
4316 if (IS_ERR(dentry)) {
4317 printk(KERN_INFO
4318 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
4319 PTR_ERR(dentry));
4320 ret = PTR_ERR(dentry);
4321 goto out_release;
4324 /* Create the cgroup directory, which also creates the cgroup */
4325 ret = vfs_mkdir(inode, dentry, 0755);
4326 child = __d_cgrp(dentry);
4327 dput(dentry);
4328 if (ret) {
4329 printk(KERN_INFO
4330 "Failed to create cgroup %s: %d\n", nodename,
4331 ret);
4332 goto out_release;
4335 /* The cgroup now exists. Retake cgroup_mutex and check
4336 * that we're still in the same state that we thought we
4337 * were. */
4338 mutex_lock(&cgroup_mutex);
4339 if ((root != subsys->root) ||
4340 (parent != task_cgroup(tsk, subsys->subsys_id))) {
4341 /* Aargh, we raced ... */
4342 mutex_unlock(&inode->i_mutex);
4343 put_css_set(cg);
4345 deactivate_super(root->sb);
4346 /* The cgroup is still accessible in the VFS, but
4347 * we're not going to try to rmdir() it at this
4348 * point. */
4349 printk(KERN_INFO
4350 "Race in cgroup_clone() - leaking cgroup %s\n",
4351 nodename);
4352 goto again;
4355 /* do any required auto-setup */
4356 for_each_subsys(root, ss) {
4357 if (ss->post_clone)
4358 ss->post_clone(ss, child);
4361 /* All seems fine. Finish by moving the task into the new cgroup */
4362 ret = cgroup_attach_task(child, tsk);
4363 mutex_unlock(&cgroup_mutex);
4365 out_release:
4366 mutex_unlock(&inode->i_mutex);
4368 mutex_lock(&cgroup_mutex);
4369 put_css_set(cg);
4370 mutex_unlock(&cgroup_mutex);
4371 deactivate_super(root->sb);
4372 return ret;
4376 * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
4377 * @cgrp: the cgroup in question
4378 * @task: the task in question
4380 * See if @cgrp is a descendant of @task's cgroup in the appropriate
4381 * hierarchy.
4383 * If we are sending in dummytop, then presumably we are creating
4384 * the top cgroup in the subsystem.
4386 * Called only by the ns (nsproxy) cgroup.
4388 int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
4390 int ret;
4391 struct cgroup *target;
4393 if (cgrp == dummytop)
4394 return 1;
4396 target = task_cgroup_from_root(task, cgrp->root);
4397 while (cgrp != target && cgrp!= cgrp->top_cgroup)
4398 cgrp = cgrp->parent;
4399 ret = (cgrp == target);
4400 return ret;
4403 static void check_for_release(struct cgroup *cgrp)
4405 /* All of these checks rely on RCU to keep the cgroup
4406 * structure alive */
4407 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
4408 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
4409 /* Control Group is currently removeable. If it's not
4410 * already queued for a userspace notification, queue
4411 * it now */
4412 int need_schedule_work = 0;
4413 spin_lock(&release_list_lock);
4414 if (!cgroup_is_removed(cgrp) &&
4415 list_empty(&cgrp->release_list)) {
4416 list_add(&cgrp->release_list, &release_list);
4417 need_schedule_work = 1;
4419 spin_unlock(&release_list_lock);
4420 if (need_schedule_work)
4421 schedule_work(&release_agent_work);
4425 /* Caller must verify that the css is not for root cgroup */
4426 void __css_put(struct cgroup_subsys_state *css, int count)
4428 struct cgroup *cgrp = css->cgroup;
4429 int val;
4430 rcu_read_lock();
4431 val = atomic_sub_return(count, &css->refcnt);
4432 if (val == 1) {
4433 if (notify_on_release(cgrp)) {
4434 set_bit(CGRP_RELEASABLE, &cgrp->flags);
4435 check_for_release(cgrp);
4437 cgroup_wakeup_rmdir_waiter(cgrp);
4439 rcu_read_unlock();
4440 WARN_ON_ONCE(val < 1);
4442 EXPORT_SYMBOL_GPL(__css_put);
4445 * Notify userspace when a cgroup is released, by running the
4446 * configured release agent with the name of the cgroup (path
4447 * relative to the root of cgroup file system) as the argument.
4449 * Most likely, this user command will try to rmdir this cgroup.
4451 * This races with the possibility that some other task will be
4452 * attached to this cgroup before it is removed, or that some other
4453 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4454 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4455 * unused, and this cgroup will be reprieved from its death sentence,
4456 * to continue to serve a useful existence. Next time it's released,
4457 * we will get notified again, if it still has 'notify_on_release' set.
4459 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4460 * means only wait until the task is successfully execve()'d. The
4461 * separate release agent task is forked by call_usermodehelper(),
4462 * then control in this thread returns here, without waiting for the
4463 * release agent task. We don't bother to wait because the caller of
4464 * this routine has no use for the exit status of the release agent
4465 * task, so no sense holding our caller up for that.
4467 static void cgroup_release_agent(struct work_struct *work)
4469 BUG_ON(work != &release_agent_work);
4470 mutex_lock(&cgroup_mutex);
4471 spin_lock(&release_list_lock);
4472 while (!list_empty(&release_list)) {
4473 char *argv[3], *envp[3];
4474 int i;
4475 char *pathbuf = NULL, *agentbuf = NULL;
4476 struct cgroup *cgrp = list_entry(release_list.next,
4477 struct cgroup,
4478 release_list);
4479 list_del_init(&cgrp->release_list);
4480 spin_unlock(&release_list_lock);
4481 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4482 if (!pathbuf)
4483 goto continue_free;
4484 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4485 goto continue_free;
4486 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4487 if (!agentbuf)
4488 goto continue_free;
4490 i = 0;
4491 argv[i++] = agentbuf;
4492 argv[i++] = pathbuf;
4493 argv[i] = NULL;
4495 i = 0;
4496 /* minimal command environment */
4497 envp[i++] = "HOME=/";
4498 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4499 envp[i] = NULL;
4501 /* Drop the lock while we invoke the usermode helper,
4502 * since the exec could involve hitting disk and hence
4503 * be a slow process */
4504 mutex_unlock(&cgroup_mutex);
4505 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4506 mutex_lock(&cgroup_mutex);
4507 continue_free:
4508 kfree(pathbuf);
4509 kfree(agentbuf);
4510 spin_lock(&release_list_lock);
4512 spin_unlock(&release_list_lock);
4513 mutex_unlock(&cgroup_mutex);
4516 static int __init cgroup_disable(char *str)
4518 int i;
4519 char *token;
4521 while ((token = strsep(&str, ",")) != NULL) {
4522 if (!*token)
4523 continue;
4525 * cgroup_disable, being at boot time, can't know about module
4526 * subsystems, so we don't worry about them.
4528 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4529 struct cgroup_subsys *ss = subsys[i];
4531 if (!strcmp(token, ss->name)) {
4532 ss->disabled = 1;
4533 printk(KERN_INFO "Disabling %s control group"
4534 " subsystem\n", ss->name);
4535 break;
4539 return 1;
4541 __setup("cgroup_disable=", cgroup_disable);
4544 * Functons for CSS ID.
4548 *To get ID other than 0, this should be called when !cgroup_is_removed().
4550 unsigned short css_id(struct cgroup_subsys_state *css)
4552 struct css_id *cssid;
4555 * This css_id() can return correct value when somone has refcnt
4556 * on this or this is under rcu_read_lock(). Once css->id is allocated,
4557 * it's unchanged until freed.
4559 cssid = rcu_dereference_check(css->id,
4560 rcu_read_lock_held() || atomic_read(&css->refcnt));
4562 if (cssid)
4563 return cssid->id;
4564 return 0;
4566 EXPORT_SYMBOL_GPL(css_id);
4568 unsigned short css_depth(struct cgroup_subsys_state *css)
4570 struct css_id *cssid;
4572 cssid = rcu_dereference_check(css->id,
4573 rcu_read_lock_held() || atomic_read(&css->refcnt));
4575 if (cssid)
4576 return cssid->depth;
4577 return 0;
4579 EXPORT_SYMBOL_GPL(css_depth);
4582 * css_is_ancestor - test "root" css is an ancestor of "child"
4583 * @child: the css to be tested.
4584 * @root: the css supporsed to be an ancestor of the child.
4586 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
4587 * this function reads css->id, this use rcu_dereference() and rcu_read_lock().
4588 * But, considering usual usage, the csses should be valid objects after test.
4589 * Assuming that the caller will do some action to the child if this returns
4590 * returns true, the caller must take "child";s reference count.
4591 * If "child" is valid object and this returns true, "root" is valid, too.
4594 bool css_is_ancestor(struct cgroup_subsys_state *child,
4595 const struct cgroup_subsys_state *root)
4597 struct css_id *child_id;
4598 struct css_id *root_id;
4599 bool ret = true;
4601 rcu_read_lock();
4602 child_id = rcu_dereference(child->id);
4603 root_id = rcu_dereference(root->id);
4604 if (!child_id
4605 || !root_id
4606 || (child_id->depth < root_id->depth)
4607 || (child_id->stack[root_id->depth] != root_id->id))
4608 ret = false;
4609 rcu_read_unlock();
4610 return ret;
4613 void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
4615 struct css_id *id = css->id;
4616 /* When this is called before css_id initialization, id can be NULL */
4617 if (!id)
4618 return;
4620 BUG_ON(!ss->use_id);
4622 rcu_assign_pointer(id->css, NULL);
4623 rcu_assign_pointer(css->id, NULL);
4624 spin_lock(&ss->id_lock);
4625 idr_remove(&ss->idr, id->id);
4626 spin_unlock(&ss->id_lock);
4627 kfree_rcu(id, rcu_head);
4629 EXPORT_SYMBOL_GPL(free_css_id);
4632 * This is called by init or create(). Then, calls to this function are
4633 * always serialized (By cgroup_mutex() at create()).
4636 static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
4638 struct css_id *newid;
4639 int myid, error, size;
4641 BUG_ON(!ss->use_id);
4643 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
4644 newid = kzalloc(size, GFP_KERNEL);
4645 if (!newid)
4646 return ERR_PTR(-ENOMEM);
4647 /* get id */
4648 if (unlikely(!idr_pre_get(&ss->idr, GFP_KERNEL))) {
4649 error = -ENOMEM;
4650 goto err_out;
4652 spin_lock(&ss->id_lock);
4653 /* Don't use 0. allocates an ID of 1-65535 */
4654 error = idr_get_new_above(&ss->idr, newid, 1, &myid);
4655 spin_unlock(&ss->id_lock);
4657 /* Returns error when there are no free spaces for new ID.*/
4658 if (error) {
4659 error = -ENOSPC;
4660 goto err_out;
4662 if (myid > CSS_ID_MAX)
4663 goto remove_idr;
4665 newid->id = myid;
4666 newid->depth = depth;
4667 return newid;
4668 remove_idr:
4669 error = -ENOSPC;
4670 spin_lock(&ss->id_lock);
4671 idr_remove(&ss->idr, myid);
4672 spin_unlock(&ss->id_lock);
4673 err_out:
4674 kfree(newid);
4675 return ERR_PTR(error);
4679 static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
4680 struct cgroup_subsys_state *rootcss)
4682 struct css_id *newid;
4684 spin_lock_init(&ss->id_lock);
4685 idr_init(&ss->idr);
4687 newid = get_new_cssid(ss, 0);
4688 if (IS_ERR(newid))
4689 return PTR_ERR(newid);
4691 newid->stack[0] = newid->id;
4692 newid->css = rootcss;
4693 rootcss->id = newid;
4694 return 0;
4697 static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
4698 struct cgroup *child)
4700 int subsys_id, i, depth = 0;
4701 struct cgroup_subsys_state *parent_css, *child_css;
4702 struct css_id *child_id, *parent_id;
4704 subsys_id = ss->subsys_id;
4705 parent_css = parent->subsys[subsys_id];
4706 child_css = child->subsys[subsys_id];
4707 parent_id = parent_css->id;
4708 depth = parent_id->depth + 1;
4710 child_id = get_new_cssid(ss, depth);
4711 if (IS_ERR(child_id))
4712 return PTR_ERR(child_id);
4714 for (i = 0; i < depth; i++)
4715 child_id->stack[i] = parent_id->stack[i];
4716 child_id->stack[depth] = child_id->id;
4718 * child_id->css pointer will be set after this cgroup is available
4719 * see cgroup_populate_dir()
4721 rcu_assign_pointer(child_css->id, child_id);
4723 return 0;
4727 * css_lookup - lookup css by id
4728 * @ss: cgroup subsys to be looked into.
4729 * @id: the id
4731 * Returns pointer to cgroup_subsys_state if there is valid one with id.
4732 * NULL if not. Should be called under rcu_read_lock()
4734 struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
4736 struct css_id *cssid = NULL;
4738 BUG_ON(!ss->use_id);
4739 cssid = idr_find(&ss->idr, id);
4741 if (unlikely(!cssid))
4742 return NULL;
4744 return rcu_dereference(cssid->css);
4746 EXPORT_SYMBOL_GPL(css_lookup);
4749 * css_get_next - lookup next cgroup under specified hierarchy.
4750 * @ss: pointer to subsystem
4751 * @id: current position of iteration.
4752 * @root: pointer to css. search tree under this.
4753 * @foundid: position of found object.
4755 * Search next css under the specified hierarchy of rootid. Calling under
4756 * rcu_read_lock() is necessary. Returns NULL if it reaches the end.
4758 struct cgroup_subsys_state *
4759 css_get_next(struct cgroup_subsys *ss, int id,
4760 struct cgroup_subsys_state *root, int *foundid)
4762 struct cgroup_subsys_state *ret = NULL;
4763 struct css_id *tmp;
4764 int tmpid;
4765 int rootid = css_id(root);
4766 int depth = css_depth(root);
4768 if (!rootid)
4769 return NULL;
4771 BUG_ON(!ss->use_id);
4772 /* fill start point for scan */
4773 tmpid = id;
4774 while (1) {
4776 * scan next entry from bitmap(tree), tmpid is updated after
4777 * idr_get_next().
4779 spin_lock(&ss->id_lock);
4780 tmp = idr_get_next(&ss->idr, &tmpid);
4781 spin_unlock(&ss->id_lock);
4783 if (!tmp)
4784 break;
4785 if (tmp->depth >= depth && tmp->stack[depth] == rootid) {
4786 ret = rcu_dereference(tmp->css);
4787 if (ret) {
4788 *foundid = tmpid;
4789 break;
4792 /* continue to scan from next id */
4793 tmpid = tmpid + 1;
4795 return ret;
4799 * get corresponding css from file open on cgroupfs directory
4801 struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id)
4803 struct cgroup *cgrp;
4804 struct inode *inode;
4805 struct cgroup_subsys_state *css;
4807 inode = f->f_dentry->d_inode;
4808 /* check in cgroup filesystem dir */
4809 if (inode->i_op != &cgroup_dir_inode_operations)
4810 return ERR_PTR(-EBADF);
4812 if (id < 0 || id >= CGROUP_SUBSYS_COUNT)
4813 return ERR_PTR(-EINVAL);
4815 /* get cgroup */
4816 cgrp = __d_cgrp(f->f_dentry);
4817 css = cgrp->subsys[id];
4818 return css ? css : ERR_PTR(-ENOENT);
4821 #ifdef CONFIG_CGROUP_DEBUG
4822 static struct cgroup_subsys_state *debug_create(struct cgroup_subsys *ss,
4823 struct cgroup *cont)
4825 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4827 if (!css)
4828 return ERR_PTR(-ENOMEM);
4830 return css;
4833 static void debug_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
4835 kfree(cont->subsys[debug_subsys_id]);
4838 static u64 cgroup_refcount_read(struct cgroup *cont, struct cftype *cft)
4840 return atomic_read(&cont->count);
4843 static u64 debug_taskcount_read(struct cgroup *cont, struct cftype *cft)
4845 return cgroup_task_count(cont);
4848 static u64 current_css_set_read(struct cgroup *cont, struct cftype *cft)
4850 return (u64)(unsigned long)current->cgroups;
4853 static u64 current_css_set_refcount_read(struct cgroup *cont,
4854 struct cftype *cft)
4856 u64 count;
4858 rcu_read_lock();
4859 count = atomic_read(&current->cgroups->refcount);
4860 rcu_read_unlock();
4861 return count;
4864 static int current_css_set_cg_links_read(struct cgroup *cont,
4865 struct cftype *cft,
4866 struct seq_file *seq)
4868 struct cg_cgroup_link *link;
4869 struct css_set *cg;
4871 read_lock(&css_set_lock);
4872 rcu_read_lock();
4873 cg = rcu_dereference(current->cgroups);
4874 list_for_each_entry(link, &cg->cg_links, cg_link_list) {
4875 struct cgroup *c = link->cgrp;
4876 const char *name;
4878 if (c->dentry)
4879 name = c->dentry->d_name.name;
4880 else
4881 name = "?";
4882 seq_printf(seq, "Root %d group %s\n",
4883 c->root->hierarchy_id, name);
4885 rcu_read_unlock();
4886 read_unlock(&css_set_lock);
4887 return 0;
4890 #define MAX_TASKS_SHOWN_PER_CSS 25
4891 static int cgroup_css_links_read(struct cgroup *cont,
4892 struct cftype *cft,
4893 struct seq_file *seq)
4895 struct cg_cgroup_link *link;
4897 read_lock(&css_set_lock);
4898 list_for_each_entry(link, &cont->css_sets, cgrp_link_list) {
4899 struct css_set *cg = link->cg;
4900 struct task_struct *task;
4901 int count = 0;
4902 seq_printf(seq, "css_set %p\n", cg);
4903 list_for_each_entry(task, &cg->tasks, cg_list) {
4904 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4905 seq_puts(seq, " ...\n");
4906 break;
4907 } else {
4908 seq_printf(seq, " task %d\n",
4909 task_pid_vnr(task));
4913 read_unlock(&css_set_lock);
4914 return 0;
4917 static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
4919 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
4922 static struct cftype debug_files[] = {
4924 .name = "cgroup_refcount",
4925 .read_u64 = cgroup_refcount_read,
4928 .name = "taskcount",
4929 .read_u64 = debug_taskcount_read,
4933 .name = "current_css_set",
4934 .read_u64 = current_css_set_read,
4938 .name = "current_css_set_refcount",
4939 .read_u64 = current_css_set_refcount_read,
4943 .name = "current_css_set_cg_links",
4944 .read_seq_string = current_css_set_cg_links_read,
4948 .name = "cgroup_css_links",
4949 .read_seq_string = cgroup_css_links_read,
4953 .name = "releasable",
4954 .read_u64 = releasable_read,
4958 static int debug_populate(struct cgroup_subsys *ss, struct cgroup *cont)
4960 return cgroup_add_files(cont, ss, debug_files,
4961 ARRAY_SIZE(debug_files));
4964 struct cgroup_subsys debug_subsys = {
4965 .name = "debug",
4966 .create = debug_create,
4967 .destroy = debug_destroy,
4968 .populate = debug_populate,
4969 .subsys_id = debug_subsys_id,
4971 #endif /* CONFIG_CGROUP_DEBUG */