projects/linux·blob v2.6.34 · refs

4842 lines · 126.3 KiB

1/*
2 * Generic process-grouping system.
3 *
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
6 *
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
10 *
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
15 *
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
18 *
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 * ---------------------------------------------------
23 *
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.
27 */
28
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/smp_lock.h>
56#include <linux/pid_namespace.h>
57#include <linux/idr.h>
58#include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
59#include <linux/eventfd.h>
60#include <linux/poll.h>
61
62#include <asm/atomic.h>
63
64static DEFINE_MUTEX(cgroup_mutex);
65
66/*
67 * Generate an array of cgroup subsystem pointers. At boot time, this is
68 * populated up to CGROUP_BUILTIN_SUBSYS_COUNT, and modular subsystems are
69 * registered after that. The mutable section of this array is protected by
70 * cgroup_mutex.
71 */
72#define SUBSYS(_x) &_x ## _subsys,
73static struct cgroup_subsys *subsys[CGROUP_SUBSYS_COUNT] = {
74#include <linux/cgroup_subsys.h>
75};
76
77#define MAX_CGROUP_ROOT_NAMELEN 64
78
79/*
80 * A cgroupfs_root represents the root of a cgroup hierarchy,
81 * and may be associated with a superblock to form an active
82 * hierarchy
83 */
84struct cgroupfs_root {
85 struct super_block *sb;
86
87 /*
88 * The bitmask of subsystems intended to be attached to this
89 * hierarchy
90 */
91 unsigned long subsys_bits;
92
93 /* Unique id for this hierarchy. */
94 int hierarchy_id;
95
96 /* The bitmask of subsystems currently attached to this hierarchy */
97 unsigned long actual_subsys_bits;
98
99 /* A list running through the attached subsystems */
100 struct list_head subsys_list;
101
102 /* The root cgroup for this hierarchy */
103 struct cgroup top_cgroup;
104
105 /* Tracks how many cgroups are currently defined in hierarchy.*/
106 int number_of_cgroups;
107
108 /* A list running through the active hierarchies */
109 struct list_head root_list;
110
111 /* Hierarchy-specific flags */
112 unsigned long flags;
113
114 /* The path to use for release notifications. */
115 char release_agent_path[PATH_MAX];
116
117 /* The name for this hierarchy - may be empty */
118 char name[MAX_CGROUP_ROOT_NAMELEN];
119};
120
121/*
122 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
123 * subsystems that are otherwise unattached - it never has more than a
124 * single cgroup, and all tasks are part of that cgroup.
125 */
126static struct cgroupfs_root rootnode;
127
128/*
129 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
130 * cgroup_subsys->use_id != 0.
131 */
132#define CSS_ID_MAX (65535)
133struct css_id {
134 /*
135 * The css to which this ID points. This pointer is set to valid value
136 * after cgroup is populated. If cgroup is removed, this will be NULL.
137 * This pointer is expected to be RCU-safe because destroy()
138 * is called after synchronize_rcu(). But for safe use, css_is_removed()
139 * css_tryget() should be used for avoiding race.
140 */
141 struct cgroup_subsys_state *css;
142 /*
143 * ID of this css.
144 */
145 unsigned short id;
146 /*
147 * Depth in hierarchy which this ID belongs to.
148 */
149 unsigned short depth;
150 /*
151 * ID is freed by RCU. (and lookup routine is RCU safe.)
152 */
153 struct rcu_head rcu_head;
154 /*
155 * Hierarchy of CSS ID belongs to.
156 */
157 unsigned short stack[0]; /* Array of Length (depth+1) */
158};
159
160/*
161 * cgroup_event represents events which userspace want to recieve.
162 */
163struct cgroup_event {
164 /*
165 * Cgroup which the event belongs to.
166 */
167 struct cgroup *cgrp;
168 /*
169 * Control file which the event associated.
170 */
171 struct cftype *cft;
172 /*
173 * eventfd to signal userspace about the event.
174 */
175 struct eventfd_ctx *eventfd;
176 /*
177 * Each of these stored in a list by the cgroup.
178 */
179 struct list_head list;
180 /*
181 * All fields below needed to unregister event when
182 * userspace closes eventfd.
183 */
184 poll_table pt;
185 wait_queue_head_t *wqh;
186 wait_queue_t wait;
187 struct work_struct remove;
188};
189
190/* The list of hierarchy roots */
191
192static LIST_HEAD(roots);
193static int root_count;
194
195static DEFINE_IDA(hierarchy_ida);
196static int next_hierarchy_id;
197static DEFINE_SPINLOCK(hierarchy_id_lock);
198
199/* dummytop is a shorthand for the dummy hierarchy's top cgroup */
200#define dummytop (&rootnode.top_cgroup)
201
202/* This flag indicates whether tasks in the fork and exit paths should
203 * check for fork/exit handlers to call. This avoids us having to do
204 * extra work in the fork/exit path if none of the subsystems need to
205 * be called.
206 */
207static int need_forkexit_callback __read_mostly;
208
209#ifdef CONFIG_PROVE_LOCKING
210int cgroup_lock_is_held(void)
211{
212 return lockdep_is_held(&cgroup_mutex);
213}
214#else /* #ifdef CONFIG_PROVE_LOCKING */
215int cgroup_lock_is_held(void)
216{
217 return mutex_is_locked(&cgroup_mutex);
218}
219#endif /* #else #ifdef CONFIG_PROVE_LOCKING */
220
221EXPORT_SYMBOL_GPL(cgroup_lock_is_held);
222
223/* convenient tests for these bits */
224inline int cgroup_is_removed(const struct cgroup *cgrp)
225{
226 return test_bit(CGRP_REMOVED, &cgrp->flags);
227}
228
229/* bits in struct cgroupfs_root flags field */
230enum {
231 ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
232};
233
234static int cgroup_is_releasable(const struct cgroup *cgrp)
235{
236 const int bits =
237 (1 << CGRP_RELEASABLE) |
238 (1 << CGRP_NOTIFY_ON_RELEASE);
239 return (cgrp->flags & bits) == bits;
240}
241
242static int notify_on_release(const struct cgroup *cgrp)
243{
244 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
245}
246
247/*
248 * for_each_subsys() allows you to iterate on each subsystem attached to
249 * an active hierarchy
250 */
251#define for_each_subsys(_root, _ss) \
252list_for_each_entry(_ss, &_root->subsys_list, sibling)
253
254/* for_each_active_root() allows you to iterate across the active hierarchies */
255#define for_each_active_root(_root) \
256list_for_each_entry(_root, &roots, root_list)
257
258/* the list of cgroups eligible for automatic release. Protected by
259 * release_list_lock */
260static LIST_HEAD(release_list);
261static DEFINE_SPINLOCK(release_list_lock);
262static void cgroup_release_agent(struct work_struct *work);
263static DECLARE_WORK(release_agent_work, cgroup_release_agent);
264static void check_for_release(struct cgroup *cgrp);
265
266/* Link structure for associating css_set objects with cgroups */
267struct cg_cgroup_link {
268 /*
269 * List running through cg_cgroup_links associated with a
270 * cgroup, anchored on cgroup->css_sets
271 */
272 struct list_head cgrp_link_list;
273 struct cgroup *cgrp;
274 /*
275 * List running through cg_cgroup_links pointing at a
276 * single css_set object, anchored on css_set->cg_links
277 */
278 struct list_head cg_link_list;
279 struct css_set *cg;
280};
281
282/* The default css_set - used by init and its children prior to any
283 * hierarchies being mounted. It contains a pointer to the root state
284 * for each subsystem. Also used to anchor the list of css_sets. Not
285 * reference-counted, to improve performance when child cgroups
286 * haven't been created.
287 */
288
289static struct css_set init_css_set;
290static struct cg_cgroup_link init_css_set_link;
291
292static int cgroup_init_idr(struct cgroup_subsys *ss,
293 struct cgroup_subsys_state *css);
294
295/* css_set_lock protects the list of css_set objects, and the
296 * chain of tasks off each css_set. Nests outside task->alloc_lock
297 * due to cgroup_iter_start() */
298static DEFINE_RWLOCK(css_set_lock);
299static int css_set_count;
300
301/*
302 * hash table for cgroup groups. This improves the performance to find
303 * an existing css_set. This hash doesn't (currently) take into
304 * account cgroups in empty hierarchies.
305 */
306#define CSS_SET_HASH_BITS 7
307#define CSS_SET_TABLE_SIZE (1 << CSS_SET_HASH_BITS)
308static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
309
310static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
311{
312 int i;
313 int index;
314 unsigned long tmp = 0UL;
315
316 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
317 tmp += (unsigned long)css[i];
318 tmp = (tmp >> 16) ^ tmp;
319
320 index = hash_long(tmp, CSS_SET_HASH_BITS);
321
322 return &css_set_table[index];
323}
324
325static void free_css_set_rcu(struct rcu_head *obj)
326{
327 struct css_set *cg = container_of(obj, struct css_set, rcu_head);
328 kfree(cg);
329}
330
331/* We don't maintain the lists running through each css_set to its
332 * task until after the first call to cgroup_iter_start(). This
333 * reduces the fork()/exit() overhead for people who have cgroups
334 * compiled into their kernel but not actually in use */
335static int use_task_css_set_links __read_mostly;
336
337static void __put_css_set(struct css_set *cg, int taskexit)
338{
339 struct cg_cgroup_link *link;
340 struct cg_cgroup_link *saved_link;
341 /*
342 * Ensure that the refcount doesn't hit zero while any readers
343 * can see it. Similar to atomic_dec_and_lock(), but for an
344 * rwlock
345 */
346 if (atomic_add_unless(&cg->refcount, -1, 1))
347 return;
348 write_lock(&css_set_lock);
349 if (!atomic_dec_and_test(&cg->refcount)) {
350 write_unlock(&css_set_lock);
351 return;
352 }
353
354 /* This css_set is dead. unlink it and release cgroup refcounts */
355 hlist_del(&cg->hlist);
356 css_set_count--;
357
358 list_for_each_entry_safe(link, saved_link, &cg->cg_links,
359 cg_link_list) {
360 struct cgroup *cgrp = link->cgrp;
361 list_del(&link->cg_link_list);
362 list_del(&link->cgrp_link_list);
363 if (atomic_dec_and_test(&cgrp->count) &&
364 notify_on_release(cgrp)) {
365 if (taskexit)
366 set_bit(CGRP_RELEASABLE, &cgrp->flags);
367 check_for_release(cgrp);
368 }
369
370 kfree(link);
371 }
372
373 write_unlock(&css_set_lock);
374 call_rcu(&cg->rcu_head, free_css_set_rcu);
375}
376
377/*
378 * refcounted get/put for css_set objects
379 */
380static inline void get_css_set(struct css_set *cg)
381{
382 atomic_inc(&cg->refcount);
383}
384
385static inline void put_css_set(struct css_set *cg)
386{
387 __put_css_set(cg, 0);
388}
389
390static inline void put_css_set_taskexit(struct css_set *cg)
391{
392 __put_css_set(cg, 1);
393}
394
395/*
396 * compare_css_sets - helper function for find_existing_css_set().
397 * @cg: candidate css_set being tested
398 * @old_cg: existing css_set for a task
399 * @new_cgrp: cgroup that's being entered by the task
400 * @template: desired set of css pointers in css_set (pre-calculated)
401 *
402 * Returns true if "cg" matches "old_cg" except for the hierarchy
403 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
404 */
405static bool compare_css_sets(struct css_set *cg,
406 struct css_set *old_cg,
407 struct cgroup *new_cgrp,
408 struct cgroup_subsys_state *template[])
409{
410 struct list_head *l1, *l2;
411
412 if (memcmp(template, cg->subsys, sizeof(cg->subsys))) {
413 /* Not all subsystems matched */
414 return false;
415 }
416
417 /*
418 * Compare cgroup pointers in order to distinguish between
419 * different cgroups in heirarchies with no subsystems. We
420 * could get by with just this check alone (and skip the
421 * memcmp above) but on most setups the memcmp check will
422 * avoid the need for this more expensive check on almost all
423 * candidates.
424 */
425
426 l1 = &cg->cg_links;
427 l2 = &old_cg->cg_links;
428 while (1) {
429 struct cg_cgroup_link *cgl1, *cgl2;
430 struct cgroup *cg1, *cg2;
431
432 l1 = l1->next;
433 l2 = l2->next;
434 /* See if we reached the end - both lists are equal length. */
435 if (l1 == &cg->cg_links) {
436 BUG_ON(l2 != &old_cg->cg_links);
437 break;
438 } else {
439 BUG_ON(l2 == &old_cg->cg_links);
440 }
441 /* Locate the cgroups associated with these links. */
442 cgl1 = list_entry(l1, struct cg_cgroup_link, cg_link_list);
443 cgl2 = list_entry(l2, struct cg_cgroup_link, cg_link_list);
444 cg1 = cgl1->cgrp;
445 cg2 = cgl2->cgrp;
446 /* Hierarchies should be linked in the same order. */
447 BUG_ON(cg1->root != cg2->root);
448
449 /*
450 * If this hierarchy is the hierarchy of the cgroup
451 * that's changing, then we need to check that this
452 * css_set points to the new cgroup; if it's any other
453 * hierarchy, then this css_set should point to the
454 * same cgroup as the old css_set.
455 */
456 if (cg1->root == new_cgrp->root) {
457 if (cg1 != new_cgrp)
458 return false;
459 } else {
460 if (cg1 != cg2)
461 return false;
462 }
463 }
464 return true;
465}
466
467/*
468 * find_existing_css_set() is a helper for
469 * find_css_set(), and checks to see whether an existing
470 * css_set is suitable.
471 *
472 * oldcg: the cgroup group that we're using before the cgroup
473 * transition
474 *
475 * cgrp: the cgroup that we're moving into
476 *
477 * template: location in which to build the desired set of subsystem
478 * state objects for the new cgroup group
479 */
480static struct css_set *find_existing_css_set(
481 struct css_set *oldcg,
482 struct cgroup *cgrp,
483 struct cgroup_subsys_state *template[])
484{
485 int i;
486 struct cgroupfs_root *root = cgrp->root;
487 struct hlist_head *hhead;
488 struct hlist_node *node;
489 struct css_set *cg;
490
491 /*
492 * Build the set of subsystem state objects that we want to see in the
493 * new css_set. while subsystems can change globally, the entries here
494 * won't change, so no need for locking.
495 */
496 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
497 if (root->subsys_bits & (1UL << i)) {
498 /* Subsystem is in this hierarchy. So we want
499 * the subsystem state from the new
500 * cgroup */
501 template[i] = cgrp->subsys[i];
502 } else {
503 /* Subsystem is not in this hierarchy, so we
504 * don't want to change the subsystem state */
505 template[i] = oldcg->subsys[i];
506 }
507 }
508
509 hhead = css_set_hash(template);
510 hlist_for_each_entry(cg, node, hhead, hlist) {
511 if (!compare_css_sets(cg, oldcg, cgrp, template))
512 continue;
513
514 /* This css_set matches what we need */
515 return cg;
516 }
517
518 /* No existing cgroup group matched */
519 return NULL;
520}
521
522static void free_cg_links(struct list_head *tmp)
523{
524 struct cg_cgroup_link *link;
525 struct cg_cgroup_link *saved_link;
526
527 list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
528 list_del(&link->cgrp_link_list);
529 kfree(link);
530 }
531}
532
533/*
534 * allocate_cg_links() allocates "count" cg_cgroup_link structures
535 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
536 * success or a negative error
537 */
538static int allocate_cg_links(int count, struct list_head *tmp)
539{
540 struct cg_cgroup_link *link;
541 int i;
542 INIT_LIST_HEAD(tmp);
543 for (i = 0; i < count; i++) {
544 link = kmalloc(sizeof(*link), GFP_KERNEL);
545 if (!link) {
546 free_cg_links(tmp);
547 return -ENOMEM;
548 }
549 list_add(&link->cgrp_link_list, tmp);
550 }
551 return 0;
552}
553
554/**
555 * link_css_set - a helper function to link a css_set to a cgroup
556 * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
557 * @cg: the css_set to be linked
558 * @cgrp: the destination cgroup
559 */
560static void link_css_set(struct list_head *tmp_cg_links,
561 struct css_set *cg, struct cgroup *cgrp)
562{
563 struct cg_cgroup_link *link;
564
565 BUG_ON(list_empty(tmp_cg_links));
566 link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
567 cgrp_link_list);
568 link->cg = cg;
569 link->cgrp = cgrp;
570 atomic_inc(&cgrp->count);
571 list_move(&link->cgrp_link_list, &cgrp->css_sets);
572 /*
573 * Always add links to the tail of the list so that the list
574 * is sorted by order of hierarchy creation
575 */
576 list_add_tail(&link->cg_link_list, &cg->cg_links);
577}
578
579/*
580 * find_css_set() takes an existing cgroup group and a
581 * cgroup object, and returns a css_set object that's
582 * equivalent to the old group, but with the given cgroup
583 * substituted into the appropriate hierarchy. Must be called with
584 * cgroup_mutex held
585 */
586static struct css_set *find_css_set(
587 struct css_set *oldcg, struct cgroup *cgrp)
588{
589 struct css_set *res;
590 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
591
592 struct list_head tmp_cg_links;
593
594 struct hlist_head *hhead;
595 struct cg_cgroup_link *link;
596
597 /* First see if we already have a cgroup group that matches
598 * the desired set */
599 read_lock(&css_set_lock);
600 res = find_existing_css_set(oldcg, cgrp, template);
601 if (res)
602 get_css_set(res);
603 read_unlock(&css_set_lock);
604
605 if (res)
606 return res;
607
608 res = kmalloc(sizeof(*res), GFP_KERNEL);
609 if (!res)
610 return NULL;
611
612 /* Allocate all the cg_cgroup_link objects that we'll need */
613 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
614 kfree(res);
615 return NULL;
616 }
617
618 atomic_set(&res->refcount, 1);
619 INIT_LIST_HEAD(&res->cg_links);
620 INIT_LIST_HEAD(&res->tasks);
621 INIT_HLIST_NODE(&res->hlist);
622
623 /* Copy the set of subsystem state objects generated in
624 * find_existing_css_set() */
625 memcpy(res->subsys, template, sizeof(res->subsys));
626
627 write_lock(&css_set_lock);
628 /* Add reference counts and links from the new css_set. */
629 list_for_each_entry(link, &oldcg->cg_links, cg_link_list) {
630 struct cgroup *c = link->cgrp;
631 if (c->root == cgrp->root)
632 c = cgrp;
633 link_css_set(&tmp_cg_links, res, c);
634 }
635
636 BUG_ON(!list_empty(&tmp_cg_links));
637
638 css_set_count++;
639
640 /* Add this cgroup group to the hash table */
641 hhead = css_set_hash(res->subsys);
642 hlist_add_head(&res->hlist, hhead);
643
644 write_unlock(&css_set_lock);
645
646 return res;
647}
648
649/*
650 * Return the cgroup for "task" from the given hierarchy. Must be
651 * called with cgroup_mutex held.
652 */
653static struct cgroup *task_cgroup_from_root(struct task_struct *task,
654 struct cgroupfs_root *root)
655{
656 struct css_set *css;
657 struct cgroup *res = NULL;
658
659 BUG_ON(!mutex_is_locked(&cgroup_mutex));
660 read_lock(&css_set_lock);
661 /*
662 * No need to lock the task - since we hold cgroup_mutex the
663 * task can't change groups, so the only thing that can happen
664 * is that it exits and its css is set back to init_css_set.
665 */
666 css = task->cgroups;
667 if (css == &init_css_set) {
668 res = &root->top_cgroup;
669 } else {
670 struct cg_cgroup_link *link;
671 list_for_each_entry(link, &css->cg_links, cg_link_list) {
672 struct cgroup *c = link->cgrp;
673 if (c->root == root) {
674 res = c;
675 break;
676 }
677 }
678 }
679 read_unlock(&css_set_lock);
680 BUG_ON(!res);
681 return res;
682}
683
684/*
685 * There is one global cgroup mutex. We also require taking
686 * task_lock() when dereferencing a task's cgroup subsys pointers.
687 * See "The task_lock() exception", at the end of this comment.
688 *
689 * A task must hold cgroup_mutex to modify cgroups.
690 *
691 * Any task can increment and decrement the count field without lock.
692 * So in general, code holding cgroup_mutex can't rely on the count
693 * field not changing. However, if the count goes to zero, then only
694 * cgroup_attach_task() can increment it again. Because a count of zero
695 * means that no tasks are currently attached, therefore there is no
696 * way a task attached to that cgroup can fork (the other way to
697 * increment the count). So code holding cgroup_mutex can safely
698 * assume that if the count is zero, it will stay zero. Similarly, if
699 * a task holds cgroup_mutex on a cgroup with zero count, it
700 * knows that the cgroup won't be removed, as cgroup_rmdir()
701 * needs that mutex.
702 *
703 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
704 * (usually) take cgroup_mutex. These are the two most performance
705 * critical pieces of code here. The exception occurs on cgroup_exit(),
706 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
707 * is taken, and if the cgroup count is zero, a usermode call made
708 * to the release agent with the name of the cgroup (path relative to
709 * the root of cgroup file system) as the argument.
710 *
711 * A cgroup can only be deleted if both its 'count' of using tasks
712 * is zero, and its list of 'children' cgroups is empty. Since all
713 * tasks in the system use _some_ cgroup, and since there is always at
714 * least one task in the system (init, pid == 1), therefore, top_cgroup
715 * always has either children cgroups and/or using tasks. So we don't
716 * need a special hack to ensure that top_cgroup cannot be deleted.
717 *
718 * The task_lock() exception
719 *
720 * The need for this exception arises from the action of
721 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
722 * another. It does so using cgroup_mutex, however there are
723 * several performance critical places that need to reference
724 * task->cgroup without the expense of grabbing a system global
725 * mutex. Therefore except as noted below, when dereferencing or, as
726 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
727 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
728 * the task_struct routinely used for such matters.
729 *
730 * P.S. One more locking exception. RCU is used to guard the
731 * update of a tasks cgroup pointer by cgroup_attach_task()
732 */
733
734/**
735 * cgroup_lock - lock out any changes to cgroup structures
736 *
737 */
738void cgroup_lock(void)
739{
740 mutex_lock(&cgroup_mutex);
741}
742EXPORT_SYMBOL_GPL(cgroup_lock);
743
744/**
745 * cgroup_unlock - release lock on cgroup changes
746 *
747 * Undo the lock taken in a previous cgroup_lock() call.
748 */
749void cgroup_unlock(void)
750{
751 mutex_unlock(&cgroup_mutex);
752}
753EXPORT_SYMBOL_GPL(cgroup_unlock);
754
755/*
756 * A couple of forward declarations required, due to cyclic reference loop:
757 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
758 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
759 * -> cgroup_mkdir.
760 */
761
762static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
763static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
764static int cgroup_populate_dir(struct cgroup *cgrp);
765static const struct inode_operations cgroup_dir_inode_operations;
766static const struct file_operations proc_cgroupstats_operations;
767
768static struct backing_dev_info cgroup_backing_dev_info = {
769 .name = "cgroup",
770 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
771};
772
773static int alloc_css_id(struct cgroup_subsys *ss,
774 struct cgroup *parent, struct cgroup *child);
775
776static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
777{
778 struct inode *inode = new_inode(sb);
779
780 if (inode) {
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;
786 }
787 return inode;
788}
789
790/*
791 * Call subsys's pre_destroy handler.
792 * This is called before css refcnt check.
793 */
794static int cgroup_call_pre_destroy(struct cgroup *cgrp)
795{
796 struct cgroup_subsys *ss;
797 int ret = 0;
798
799 for_each_subsys(cgrp->root, ss)
800 if (ss->pre_destroy) {
801 ret = ss->pre_destroy(ss, cgrp);
802 if (ret)
803 break;
804 }
805
806 return ret;
807}
808
809static void free_cgroup_rcu(struct rcu_head *obj)
810{
811 struct cgroup *cgrp = container_of(obj, struct cgroup, rcu_head);
812
813 kfree(cgrp);
814}
815
816static void cgroup_diput(struct dentry *dentry, struct inode *inode)
817{
818 /* is dentry a directory ? if so, kfree() associated cgroup */
819 if (S_ISDIR(inode->i_mode)) {
820 struct cgroup *cgrp = dentry->d_fsdata;
821 struct cgroup_subsys *ss;
822 BUG_ON(!(cgroup_is_removed(cgrp)));
823 /* It's possible for external users to be holding css
824 * reference counts on a cgroup; css_put() needs to
825 * be able to access the cgroup after decrementing
826 * the reference count in order to know if it needs to
827 * queue the cgroup to be handled by the release
828 * agent */
829 synchronize_rcu();
830
831 mutex_lock(&cgroup_mutex);
832 /*
833 * Release the subsystem state objects.
834 */
835 for_each_subsys(cgrp->root, ss)
836 ss->destroy(ss, cgrp);
837
838 cgrp->root->number_of_cgroups--;
839 mutex_unlock(&cgroup_mutex);
840
841 /*
842 * Drop the active superblock reference that we took when we
843 * created the cgroup
844 */
845 deactivate_super(cgrp->root->sb);
846
847 /*
848 * if we're getting rid of the cgroup, refcount should ensure
849 * that there are no pidlists left.
850 */
851 BUG_ON(!list_empty(&cgrp->pidlists));
852
853 call_rcu(&cgrp->rcu_head, free_cgroup_rcu);
854 }
855 iput(inode);
856}
857
858static void remove_dir(struct dentry *d)
859{
860 struct dentry *parent = dget(d->d_parent);
861
862 d_delete(d);
863 simple_rmdir(parent->d_inode, d);
864 dput(parent);
865}
866
867static void cgroup_clear_directory(struct dentry *dentry)
868{
869 struct list_head *node;
870
871 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
872 spin_lock(&dcache_lock);
873 node = dentry->d_subdirs.next;
874 while (node != &dentry->d_subdirs) {
875 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
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 d = dget_locked(d);
882 spin_unlock(&dcache_lock);
883 d_delete(d);
884 simple_unlink(dentry->d_inode, d);
885 dput(d);
886 spin_lock(&dcache_lock);
887 }
888 node = dentry->d_subdirs.next;
889 }
890 spin_unlock(&dcache_lock);
891}
892
893/*
894 * NOTE : the dentry must have been dget()'ed
895 */
896static void cgroup_d_remove_dir(struct dentry *dentry)
897{
898 cgroup_clear_directory(dentry);
899
900 spin_lock(&dcache_lock);
901 list_del_init(&dentry->d_u.d_child);
902 spin_unlock(&dcache_lock);
903 remove_dir(dentry);
904}
905
906/*
907 * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
908 * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
909 * reference to css->refcnt. In general, this refcnt is expected to goes down
910 * to zero, soon.
911 *
912 * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
913 */
914DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
915
916static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
917{
918 if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
919 wake_up_all(&cgroup_rmdir_waitq);
920}
921
922void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
923{
924 css_get(css);
925}
926
927void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
928{
929 cgroup_wakeup_rmdir_waiter(css->cgroup);
930 css_put(css);
931}
932
933/*
934 * Call with cgroup_mutex held. Drops reference counts on modules, including
935 * any duplicate ones that parse_cgroupfs_options took. If this function
936 * returns an error, no reference counts are touched.
937 */
938static int rebind_subsystems(struct cgroupfs_root *root,
939 unsigned long final_bits)
940{
941 unsigned long added_bits, removed_bits;
942 struct cgroup *cgrp = &root->top_cgroup;
943 int i;
944
945 BUG_ON(!mutex_is_locked(&cgroup_mutex));
946
947 removed_bits = root->actual_subsys_bits & ~final_bits;
948 added_bits = final_bits & ~root->actual_subsys_bits;
949 /* Check that any added subsystems are currently free */
950 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
951 unsigned long bit = 1UL << i;
952 struct cgroup_subsys *ss = subsys[i];
953 if (!(bit & added_bits))
954 continue;
955 /*
956 * Nobody should tell us to do a subsys that doesn't exist:
957 * parse_cgroupfs_options should catch that case and refcounts
958 * ensure that subsystems won't disappear once selected.
959 */
960 BUG_ON(ss == NULL);
961 if (ss->root != &rootnode) {
962 /* Subsystem isn't free */
963 return -EBUSY;
964 }
965 }
966
967 /* Currently we don't handle adding/removing subsystems when
968 * any child cgroups exist. This is theoretically supportable
969 * but involves complex error handling, so it's being left until
970 * later */
971 if (root->number_of_cgroups > 1)
972 return -EBUSY;
973
974 /* Process each subsystem */
975 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
976 struct cgroup_subsys *ss = subsys[i];
977 unsigned long bit = 1UL << i;
978 if (bit & added_bits) {
979 /* We're binding this subsystem to this hierarchy */
980 BUG_ON(ss == NULL);
981 BUG_ON(cgrp->subsys[i]);
982 BUG_ON(!dummytop->subsys[i]);
983 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
984 mutex_lock(&ss->hierarchy_mutex);
985 cgrp->subsys[i] = dummytop->subsys[i];
986 cgrp->subsys[i]->cgroup = cgrp;
987 list_move(&ss->sibling, &root->subsys_list);
988 ss->root = root;
989 if (ss->bind)
990 ss->bind(ss, cgrp);
991 mutex_unlock(&ss->hierarchy_mutex);
992 /* refcount was already taken, and we're keeping it */
993 } else if (bit & removed_bits) {
994 /* We're removing this subsystem */
995 BUG_ON(ss == NULL);
996 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
997 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
998 mutex_lock(&ss->hierarchy_mutex);
999 if (ss->bind)
1000 ss->bind(ss, dummytop);
1001 dummytop->subsys[i]->cgroup = dummytop;
1002 cgrp->subsys[i] = NULL;
1003 subsys[i]->root = &rootnode;
1004 list_move(&ss->sibling, &rootnode.subsys_list);
1005 mutex_unlock(&ss->hierarchy_mutex);
1006 /* subsystem is now free - drop reference on module */
1007 module_put(ss->module);
1008 } else if (bit & final_bits) {
1009 /* Subsystem state should already exist */
1010 BUG_ON(ss == NULL);
1011 BUG_ON(!cgrp->subsys[i]);
1012 /*
1013 * a refcount was taken, but we already had one, so
1014 * drop the extra reference.
1015 */
1016 module_put(ss->module);
1017#ifdef CONFIG_MODULE_UNLOAD
1018 BUG_ON(ss->module && !module_refcount(ss->module));
1019#endif
1020 } else {
1021 /* Subsystem state shouldn't exist */
1022 BUG_ON(cgrp->subsys[i]);
1023 }
1024 }
1025 root->subsys_bits = root->actual_subsys_bits = final_bits;
1026 synchronize_rcu();
1027
1028 return 0;
1029}
1030
1031static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
1032{
1033 struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
1034 struct cgroup_subsys *ss;
1035
1036 mutex_lock(&cgroup_mutex);
1037 for_each_subsys(root, ss)
1038 seq_printf(seq, ",%s", ss->name);
1039 if (test_bit(ROOT_NOPREFIX, &root->flags))
1040 seq_puts(seq, ",noprefix");
1041 if (strlen(root->release_agent_path))
1042 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1043 if (strlen(root->name))
1044 seq_printf(seq, ",name=%s", root->name);
1045 mutex_unlock(&cgroup_mutex);
1046 return 0;
1047}
1048
1049struct cgroup_sb_opts {
1050 unsigned long subsys_bits;
1051 unsigned long flags;
1052 char *release_agent;
1053 char *name;
1054 /* User explicitly requested empty subsystem */
1055 bool none;
1056
1057 struct cgroupfs_root *new_root;
1058
1059};
1060
1061/*
1062 * Convert a hierarchy specifier into a bitmask of subsystems and flags. Call
1063 * with cgroup_mutex held to protect the subsys[] array. This function takes
1064 * refcounts on subsystems to be used, unless it returns error, in which case
1065 * no refcounts are taken.
1066 */
1067static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1068{
1069 char *token, *o = data ?: "all";
1070 unsigned long mask = (unsigned long)-1;
1071 int i;
1072 bool module_pin_failed = false;
1073
1074 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1075
1076#ifdef CONFIG_CPUSETS
1077 mask = ~(1UL << cpuset_subsys_id);
1078#endif
1079
1080 memset(opts, 0, sizeof(*opts));
1081
1082 while ((token = strsep(&o, ",")) != NULL) {
1083 if (!*token)
1084 return -EINVAL;
1085 if (!strcmp(token, "all")) {
1086 /* Add all non-disabled subsystems */
1087 opts->subsys_bits = 0;
1088 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1089 struct cgroup_subsys *ss = subsys[i];
1090 if (ss == NULL)
1091 continue;
1092 if (!ss->disabled)
1093 opts->subsys_bits |= 1ul << i;
1094 }
1095 } else if (!strcmp(token, "none")) {
1096 /* Explicitly have no subsystems */
1097 opts->none = true;
1098 } else if (!strcmp(token, "noprefix")) {
1099 set_bit(ROOT_NOPREFIX, &opts->flags);
1100 } else if (!strncmp(token, "release_agent=", 14)) {
1101 /* Specifying two release agents is forbidden */
1102 if (opts->release_agent)
1103 return -EINVAL;
1104 opts->release_agent =
1105 kstrndup(token + 14, PATH_MAX, GFP_KERNEL);
1106 if (!opts->release_agent)
1107 return -ENOMEM;
1108 } else if (!strncmp(token, "name=", 5)) {
1109 const char *name = token + 5;
1110 /* Can't specify an empty name */
1111 if (!strlen(name))
1112 return -EINVAL;
1113 /* Must match [\w.-]+ */
1114 for (i = 0; i < strlen(name); i++) {
1115 char c = name[i];
1116 if (isalnum(c))
1117 continue;
1118 if ((c == '.') || (c == '-') || (c == '_'))
1119 continue;
1120 return -EINVAL;
1121 }
1122 /* Specifying two names is forbidden */
1123 if (opts->name)
1124 return -EINVAL;
1125 opts->name = kstrndup(name,
1126 MAX_CGROUP_ROOT_NAMELEN,
1127 GFP_KERNEL);
1128 if (!opts->name)
1129 return -ENOMEM;
1130 } else {
1131 struct cgroup_subsys *ss;
1132 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1133 ss = subsys[i];
1134 if (ss == NULL)
1135 continue;
1136 if (!strcmp(token, ss->name)) {
1137 if (!ss->disabled)
1138 set_bit(i, &opts->subsys_bits);
1139 break;
1140 }
1141 }
1142 if (i == CGROUP_SUBSYS_COUNT)
1143 return -ENOENT;
1144 }
1145 }
1146
1147 /* Consistency checks */
1148
1149 /*
1150 * Option noprefix was introduced just for backward compatibility
1151 * with the old cpuset, so we allow noprefix only if mounting just
1152 * the cpuset subsystem.
1153 */
1154 if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
1155 (opts->subsys_bits & mask))
1156 return -EINVAL;
1157
1158
1159 /* Can't specify "none" and some subsystems */
1160 if (opts->subsys_bits && opts->none)
1161 return -EINVAL;
1162
1163 /*
1164 * We either have to specify by name or by subsystems. (So all
1165 * empty hierarchies must have a name).
1166 */
1167 if (!opts->subsys_bits && !opts->name)
1168 return -EINVAL;
1169
1170 /*
1171 * Grab references on all the modules we'll need, so the subsystems
1172 * don't dance around before rebind_subsystems attaches them. This may
1173 * take duplicate reference counts on a subsystem that's already used,
1174 * but rebind_subsystems handles this case.
1175 */
1176 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1177 unsigned long bit = 1UL << i;
1178
1179 if (!(bit & opts->subsys_bits))
1180 continue;
1181 if (!try_module_get(subsys[i]->module)) {
1182 module_pin_failed = true;
1183 break;
1184 }
1185 }
1186 if (module_pin_failed) {
1187 /*
1188 * oops, one of the modules was going away. this means that we
1189 * raced with a module_delete call, and to the user this is
1190 * essentially a "subsystem doesn't exist" case.
1191 */
1192 for (i--; i >= CGROUP_BUILTIN_SUBSYS_COUNT; i--) {
1193 /* drop refcounts only on the ones we took */
1194 unsigned long bit = 1UL << i;
1195
1196 if (!(bit & opts->subsys_bits))
1197 continue;
1198 module_put(subsys[i]->module);
1199 }
1200 return -ENOENT;
1201 }
1202
1203 return 0;
1204}
1205
1206static void drop_parsed_module_refcounts(unsigned long subsys_bits)
1207{
1208 int i;
1209 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1210 unsigned long bit = 1UL << i;
1211
1212 if (!(bit & subsys_bits))
1213 continue;
1214 module_put(subsys[i]->module);
1215 }
1216}
1217
1218static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1219{
1220 int ret = 0;
1221 struct cgroupfs_root *root = sb->s_fs_info;
1222 struct cgroup *cgrp = &root->top_cgroup;
1223 struct cgroup_sb_opts opts;
1224
1225 lock_kernel();
1226 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1227 mutex_lock(&cgroup_mutex);
1228
1229 /* See what subsystems are wanted */
1230 ret = parse_cgroupfs_options(data, &opts);
1231 if (ret)
1232 goto out_unlock;
1233
1234 /* Don't allow flags or name to change at remount */
1235 if (opts.flags != root->flags ||
1236 (opts.name && strcmp(opts.name, root->name))) {
1237 ret = -EINVAL;
1238 drop_parsed_module_refcounts(opts.subsys_bits);
1239 goto out_unlock;
1240 }
1241
1242 ret = rebind_subsystems(root, opts.subsys_bits);
1243 if (ret) {
1244 drop_parsed_module_refcounts(opts.subsys_bits);
1245 goto out_unlock;
1246 }
1247
1248 /* (re)populate subsystem files */
1249 cgroup_populate_dir(cgrp);
1250
1251 if (opts.release_agent)
1252 strcpy(root->release_agent_path, opts.release_agent);
1253 out_unlock:
1254 kfree(opts.release_agent);
1255 kfree(opts.name);
1256 mutex_unlock(&cgroup_mutex);
1257 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1258 unlock_kernel();
1259 return ret;
1260}
1261
1262static const struct super_operations cgroup_ops = {
1263 .statfs = simple_statfs,
1264 .drop_inode = generic_delete_inode,
1265 .show_options = cgroup_show_options,
1266 .remount_fs = cgroup_remount,
1267};
1268
1269static void init_cgroup_housekeeping(struct cgroup *cgrp)
1270{
1271 INIT_LIST_HEAD(&cgrp->sibling);
1272 INIT_LIST_HEAD(&cgrp->children);
1273 INIT_LIST_HEAD(&cgrp->css_sets);
1274 INIT_LIST_HEAD(&cgrp->release_list);
1275 INIT_LIST_HEAD(&cgrp->pidlists);
1276 mutex_init(&cgrp->pidlist_mutex);
1277 INIT_LIST_HEAD(&cgrp->event_list);
1278 spin_lock_init(&cgrp->event_list_lock);
1279}
1280
1281static void init_cgroup_root(struct cgroupfs_root *root)
1282{
1283 struct cgroup *cgrp = &root->top_cgroup;
1284 INIT_LIST_HEAD(&root->subsys_list);
1285 INIT_LIST_HEAD(&root->root_list);
1286 root->number_of_cgroups = 1;
1287 cgrp->root = root;
1288 cgrp->top_cgroup = cgrp;
1289 init_cgroup_housekeeping(cgrp);
1290}
1291
1292static bool init_root_id(struct cgroupfs_root *root)
1293{
1294 int ret = 0;
1295
1296 do {
1297 if (!ida_pre_get(&hierarchy_ida, GFP_KERNEL))
1298 return false;
1299 spin_lock(&hierarchy_id_lock);
1300 /* Try to allocate the next unused ID */
1301 ret = ida_get_new_above(&hierarchy_ida, next_hierarchy_id,
1302 &root->hierarchy_id);
1303 if (ret == -ENOSPC)
1304 /* Try again starting from 0 */
1305 ret = ida_get_new(&hierarchy_ida, &root->hierarchy_id);
1306 if (!ret) {
1307 next_hierarchy_id = root->hierarchy_id + 1;
1308 } else if (ret != -EAGAIN) {
1309 /* Can only get here if the 31-bit IDR is full ... */
1310 BUG_ON(ret);
1311 }
1312 spin_unlock(&hierarchy_id_lock);
1313 } while (ret);
1314 return true;
1315}
1316
1317static int cgroup_test_super(struct super_block *sb, void *data)
1318{
1319 struct cgroup_sb_opts *opts = data;
1320 struct cgroupfs_root *root = sb->s_fs_info;
1321
1322 /* If we asked for a name then it must match */
1323 if (opts->name && strcmp(opts->name, root->name))
1324 return 0;
1325
1326 /*
1327 * If we asked for subsystems (or explicitly for no
1328 * subsystems) then they must match
1329 */
1330 if ((opts->subsys_bits || opts->none)
1331 && (opts->subsys_bits != root->subsys_bits))
1332 return 0;
1333
1334 return 1;
1335}
1336
1337static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1338{
1339 struct cgroupfs_root *root;
1340
1341 if (!opts->subsys_bits && !opts->none)
1342 return NULL;
1343
1344 root = kzalloc(sizeof(*root), GFP_KERNEL);
1345 if (!root)
1346 return ERR_PTR(-ENOMEM);
1347
1348 if (!init_root_id(root)) {
1349 kfree(root);
1350 return ERR_PTR(-ENOMEM);
1351 }
1352 init_cgroup_root(root);
1353
1354 root->subsys_bits = opts->subsys_bits;
1355 root->flags = opts->flags;
1356 if (opts->release_agent)
1357 strcpy(root->release_agent_path, opts->release_agent);
1358 if (opts->name)
1359 strcpy(root->name, opts->name);
1360 return root;
1361}
1362
1363static void cgroup_drop_root(struct cgroupfs_root *root)
1364{
1365 if (!root)
1366 return;
1367
1368 BUG_ON(!root->hierarchy_id);
1369 spin_lock(&hierarchy_id_lock);
1370 ida_remove(&hierarchy_ida, root->hierarchy_id);
1371 spin_unlock(&hierarchy_id_lock);
1372 kfree(root);
1373}
1374
1375static int cgroup_set_super(struct super_block *sb, void *data)
1376{
1377 int ret;
1378 struct cgroup_sb_opts *opts = data;
1379
1380 /* If we don't have a new root, we can't set up a new sb */
1381 if (!opts->new_root)
1382 return -EINVAL;
1383
1384 BUG_ON(!opts->subsys_bits && !opts->none);
1385
1386 ret = set_anon_super(sb, NULL);
1387 if (ret)
1388 return ret;
1389
1390 sb->s_fs_info = opts->new_root;
1391 opts->new_root->sb = sb;
1392
1393 sb->s_blocksize = PAGE_CACHE_SIZE;
1394 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1395 sb->s_magic = CGROUP_SUPER_MAGIC;
1396 sb->s_op = &cgroup_ops;
1397
1398 return 0;
1399}
1400
1401static int cgroup_get_rootdir(struct super_block *sb)
1402{
1403 struct inode *inode =
1404 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1405 struct dentry *dentry;
1406
1407 if (!inode)
1408 return -ENOMEM;
1409
1410 inode->i_fop = &simple_dir_operations;
1411 inode->i_op = &cgroup_dir_inode_operations;
1412 /* directories start off with i_nlink == 2 (for "." entry) */
1413 inc_nlink(inode);
1414 dentry = d_alloc_root(inode);
1415 if (!dentry) {
1416 iput(inode);
1417 return -ENOMEM;
1418 }
1419 sb->s_root = dentry;
1420 return 0;
1421}
1422
1423static int cgroup_get_sb(struct file_system_type *fs_type,
1424 int flags, const char *unused_dev_name,
1425 void *data, struct vfsmount *mnt)
1426{
1427 struct cgroup_sb_opts opts;
1428 struct cgroupfs_root *root;
1429 int ret = 0;
1430 struct super_block *sb;
1431 struct cgroupfs_root *new_root;
1432
1433 /* First find the desired set of subsystems */
1434 mutex_lock(&cgroup_mutex);
1435 ret = parse_cgroupfs_options(data, &opts);
1436 mutex_unlock(&cgroup_mutex);
1437 if (ret)
1438 goto out_err;
1439
1440 /*
1441 * Allocate a new cgroup root. We may not need it if we're
1442 * reusing an existing hierarchy.
1443 */
1444 new_root = cgroup_root_from_opts(&opts);
1445 if (IS_ERR(new_root)) {
1446 ret = PTR_ERR(new_root);
1447 goto drop_modules;
1448 }
1449 opts.new_root = new_root;
1450
1451 /* Locate an existing or new sb for this hierarchy */
1452 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
1453 if (IS_ERR(sb)) {
1454 ret = PTR_ERR(sb);
1455 cgroup_drop_root(opts.new_root);
1456 goto drop_modules;
1457 }
1458
1459 root = sb->s_fs_info;
1460 BUG_ON(!root);
1461 if (root == opts.new_root) {
1462 /* We used the new root structure, so this is a new hierarchy */
1463 struct list_head tmp_cg_links;
1464 struct cgroup *root_cgrp = &root->top_cgroup;
1465 struct inode *inode;
1466 struct cgroupfs_root *existing_root;
1467 int i;
1468
1469 BUG_ON(sb->s_root != NULL);
1470
1471 ret = cgroup_get_rootdir(sb);
1472 if (ret)
1473 goto drop_new_super;
1474 inode = sb->s_root->d_inode;
1475
1476 mutex_lock(&inode->i_mutex);
1477 mutex_lock(&cgroup_mutex);
1478
1479 if (strlen(root->name)) {
1480 /* Check for name clashes with existing mounts */
1481 for_each_active_root(existing_root) {
1482 if (!strcmp(existing_root->name, root->name)) {
1483 ret = -EBUSY;
1484 mutex_unlock(&cgroup_mutex);
1485 mutex_unlock(&inode->i_mutex);
1486 goto drop_new_super;
1487 }
1488 }
1489 }
1490
1491 /*
1492 * We're accessing css_set_count without locking
1493 * css_set_lock here, but that's OK - it can only be
1494 * increased by someone holding cgroup_lock, and
1495 * that's us. The worst that can happen is that we
1496 * have some link structures left over
1497 */
1498 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1499 if (ret) {
1500 mutex_unlock(&cgroup_mutex);
1501 mutex_unlock(&inode->i_mutex);
1502 goto drop_new_super;
1503 }
1504
1505 ret = rebind_subsystems(root, root->subsys_bits);
1506 if (ret == -EBUSY) {
1507 mutex_unlock(&cgroup_mutex);
1508 mutex_unlock(&inode->i_mutex);
1509 free_cg_links(&tmp_cg_links);
1510 goto drop_new_super;
1511 }
1512 /*
1513 * There must be no failure case after here, since rebinding
1514 * takes care of subsystems' refcounts, which are explicitly
1515 * dropped in the failure exit path.
1516 */
1517
1518 /* EBUSY should be the only error here */
1519 BUG_ON(ret);
1520
1521 list_add(&root->root_list, &roots);
1522 root_count++;
1523
1524 sb->s_root->d_fsdata = root_cgrp;
1525 root->top_cgroup.dentry = sb->s_root;
1526
1527 /* Link the top cgroup in this hierarchy into all
1528 * the css_set objects */
1529 write_lock(&css_set_lock);
1530 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1531 struct hlist_head *hhead = &css_set_table[i];
1532 struct hlist_node *node;
1533 struct css_set *cg;
1534
1535 hlist_for_each_entry(cg, node, hhead, hlist)
1536 link_css_set(&tmp_cg_links, cg, root_cgrp);
1537 }
1538 write_unlock(&css_set_lock);
1539
1540 free_cg_links(&tmp_cg_links);
1541
1542 BUG_ON(!list_empty(&root_cgrp->sibling));
1543 BUG_ON(!list_empty(&root_cgrp->children));
1544 BUG_ON(root->number_of_cgroups != 1);
1545
1546 cgroup_populate_dir(root_cgrp);
1547 mutex_unlock(&cgroup_mutex);
1548 mutex_unlock(&inode->i_mutex);
1549 } else {
1550 /*
1551 * We re-used an existing hierarchy - the new root (if
1552 * any) is not needed
1553 */
1554 cgroup_drop_root(opts.new_root);
1555 /* no subsys rebinding, so refcounts don't change */
1556 drop_parsed_module_refcounts(opts.subsys_bits);
1557 }
1558
1559 simple_set_mnt(mnt, sb);
1560 kfree(opts.release_agent);
1561 kfree(opts.name);
1562 return 0;
1563
1564 drop_new_super:
1565 deactivate_locked_super(sb);
1566 drop_modules:
1567 drop_parsed_module_refcounts(opts.subsys_bits);
1568 out_err:
1569 kfree(opts.release_agent);
1570 kfree(opts.name);
1571
1572 return ret;
1573}
1574
1575static void cgroup_kill_sb(struct super_block *sb) {
1576 struct cgroupfs_root *root = sb->s_fs_info;
1577 struct cgroup *cgrp = &root->top_cgroup;
1578 int ret;
1579 struct cg_cgroup_link *link;
1580 struct cg_cgroup_link *saved_link;
1581
1582 BUG_ON(!root);
1583
1584 BUG_ON(root->number_of_cgroups != 1);
1585 BUG_ON(!list_empty(&cgrp->children));
1586 BUG_ON(!list_empty(&cgrp->sibling));
1587
1588 mutex_lock(&cgroup_mutex);
1589
1590 /* Rebind all subsystems back to the default hierarchy */
1591 ret = rebind_subsystems(root, 0);
1592 /* Shouldn't be able to fail ... */
1593 BUG_ON(ret);
1594
1595 /*
1596 * Release all the links from css_sets to this hierarchy's
1597 * root cgroup
1598 */
1599 write_lock(&css_set_lock);
1600
1601 list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1602 cgrp_link_list) {
1603 list_del(&link->cg_link_list);
1604 list_del(&link->cgrp_link_list);
1605 kfree(link);
1606 }
1607 write_unlock(&css_set_lock);
1608
1609 if (!list_empty(&root->root_list)) {
1610 list_del(&root->root_list);
1611 root_count--;
1612 }
1613
1614 mutex_unlock(&cgroup_mutex);
1615
1616 kill_litter_super(sb);
1617 cgroup_drop_root(root);
1618}
1619
1620static struct file_system_type cgroup_fs_type = {
1621 .name = "cgroup",
1622 .get_sb = cgroup_get_sb,
1623 .kill_sb = cgroup_kill_sb,
1624};
1625
1626static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1627{
1628 return dentry->d_fsdata;
1629}
1630
1631static inline struct cftype *__d_cft(struct dentry *dentry)
1632{
1633 return dentry->d_fsdata;
1634}
1635
1636/**
1637 * cgroup_path - generate the path of a cgroup
1638 * @cgrp: the cgroup in question
1639 * @buf: the buffer to write the path into
1640 * @buflen: the length of the buffer
1641 *
1642 * Called with cgroup_mutex held or else with an RCU-protected cgroup
1643 * reference. Writes path of cgroup into buf. Returns 0 on success,
1644 * -errno on error.
1645 */
1646int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1647{
1648 char *start;
1649 struct dentry *dentry = rcu_dereference_check(cgrp->dentry,
1650 rcu_read_lock_held() ||
1651 cgroup_lock_is_held());
1652
1653 if (!dentry || cgrp == dummytop) {
1654 /*
1655 * Inactive subsystems have no dentry for their root
1656 * cgroup
1657 */
1658 strcpy(buf, "/");
1659 return 0;
1660 }
1661
1662 start = buf + buflen;
1663
1664 *--start = '\0';
1665 for (;;) {
1666 int len = dentry->d_name.len;
1667
1668 if ((start -= len) < buf)
1669 return -ENAMETOOLONG;
1670 memcpy(start, dentry->d_name.name, len);
1671 cgrp = cgrp->parent;
1672 if (!cgrp)
1673 break;
1674
1675 dentry = rcu_dereference_check(cgrp->dentry,
1676 rcu_read_lock_held() ||
1677 cgroup_lock_is_held());
1678 if (!cgrp->parent)
1679 continue;
1680 if (--start < buf)
1681 return -ENAMETOOLONG;
1682 *start = '/';
1683 }
1684 memmove(buf, start, buf + buflen - start);
1685 return 0;
1686}
1687EXPORT_SYMBOL_GPL(cgroup_path);
1688
1689/**
1690 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1691 * @cgrp: the cgroup the task is attaching to
1692 * @tsk: the task to be attached
1693 *
1694 * Call holding cgroup_mutex. May take task_lock of
1695 * the task 'tsk' during call.
1696 */
1697int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1698{
1699 int retval = 0;
1700 struct cgroup_subsys *ss, *failed_ss = NULL;
1701 struct cgroup *oldcgrp;
1702 struct css_set *cg;
1703 struct css_set *newcg;
1704 struct cgroupfs_root *root = cgrp->root;
1705
1706 /* Nothing to do if the task is already in that cgroup */
1707 oldcgrp = task_cgroup_from_root(tsk, root);
1708 if (cgrp == oldcgrp)
1709 return 0;
1710
1711 for_each_subsys(root, ss) {
1712 if (ss->can_attach) {
1713 retval = ss->can_attach(ss, cgrp, tsk, false);
1714 if (retval) {
1715 /*
1716 * Remember on which subsystem the can_attach()
1717 * failed, so that we only call cancel_attach()
1718 * against the subsystems whose can_attach()
1719 * succeeded. (See below)
1720 */
1721 failed_ss = ss;
1722 goto out;
1723 }
1724 }
1725 }
1726
1727 task_lock(tsk);
1728 cg = tsk->cgroups;
1729 get_css_set(cg);
1730 task_unlock(tsk);
1731 /*
1732 * Locate or allocate a new css_set for this task,
1733 * based on its final set of cgroups
1734 */
1735 newcg = find_css_set(cg, cgrp);
1736 put_css_set(cg);
1737 if (!newcg) {
1738 retval = -ENOMEM;
1739 goto out;
1740 }
1741
1742 task_lock(tsk);
1743 if (tsk->flags & PF_EXITING) {
1744 task_unlock(tsk);
1745 put_css_set(newcg);
1746 retval = -ESRCH;
1747 goto out;
1748 }
1749 rcu_assign_pointer(tsk->cgroups, newcg);
1750 task_unlock(tsk);
1751
1752 /* Update the css_set linked lists if we're using them */
1753 write_lock(&css_set_lock);
1754 if (!list_empty(&tsk->cg_list)) {
1755 list_del(&tsk->cg_list);
1756 list_add(&tsk->cg_list, &newcg->tasks);
1757 }
1758 write_unlock(&css_set_lock);
1759
1760 for_each_subsys(root, ss) {
1761 if (ss->attach)
1762 ss->attach(ss, cgrp, oldcgrp, tsk, false);
1763 }
1764 set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1765 synchronize_rcu();
1766 put_css_set(cg);
1767
1768 /*
1769 * wake up rmdir() waiter. the rmdir should fail since the cgroup
1770 * is no longer empty.
1771 */
1772 cgroup_wakeup_rmdir_waiter(cgrp);
1773out:
1774 if (retval) {
1775 for_each_subsys(root, ss) {
1776 if (ss == failed_ss)
1777 /*
1778 * This subsystem was the one that failed the
1779 * can_attach() check earlier, so we don't need
1780 * to call cancel_attach() against it or any
1781 * remaining subsystems.
1782 */
1783 break;
1784 if (ss->cancel_attach)
1785 ss->cancel_attach(ss, cgrp, tsk, false);
1786 }
1787 }
1788 return retval;
1789}
1790
1791/*
1792 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
1793 * held. May take task_lock of task
1794 */
1795static int attach_task_by_pid(struct cgroup *cgrp, u64 pid)
1796{
1797 struct task_struct *tsk;
1798 const struct cred *cred = current_cred(), *tcred;
1799 int ret;
1800
1801 if (pid) {
1802 rcu_read_lock();
1803 tsk = find_task_by_vpid(pid);
1804 if (!tsk || tsk->flags & PF_EXITING) {
1805 rcu_read_unlock();
1806 return -ESRCH;
1807 }
1808
1809 tcred = __task_cred(tsk);
1810 if (cred->euid &&
1811 cred->euid != tcred->uid &&
1812 cred->euid != tcred->suid) {
1813 rcu_read_unlock();
1814 return -EACCES;
1815 }
1816 get_task_struct(tsk);
1817 rcu_read_unlock();
1818 } else {
1819 tsk = current;
1820 get_task_struct(tsk);
1821 }
1822
1823 ret = cgroup_attach_task(cgrp, tsk);
1824 put_task_struct(tsk);
1825 return ret;
1826}
1827
1828static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
1829{
1830 int ret;
1831 if (!cgroup_lock_live_group(cgrp))
1832 return -ENODEV;
1833 ret = attach_task_by_pid(cgrp, pid);
1834 cgroup_unlock();
1835 return ret;
1836}
1837
1838/**
1839 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
1840 * @cgrp: the cgroup to be checked for liveness
1841 *
1842 * On success, returns true; the lock should be later released with
1843 * cgroup_unlock(). On failure returns false with no lock held.
1844 */
1845bool cgroup_lock_live_group(struct cgroup *cgrp)
1846{
1847 mutex_lock(&cgroup_mutex);
1848 if (cgroup_is_removed(cgrp)) {
1849 mutex_unlock(&cgroup_mutex);
1850 return false;
1851 }
1852 return true;
1853}
1854EXPORT_SYMBOL_GPL(cgroup_lock_live_group);
1855
1856static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
1857 const char *buffer)
1858{
1859 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1860 if (!cgroup_lock_live_group(cgrp))
1861 return -ENODEV;
1862 strcpy(cgrp->root->release_agent_path, buffer);
1863 cgroup_unlock();
1864 return 0;
1865}
1866
1867static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
1868 struct seq_file *seq)
1869{
1870 if (!cgroup_lock_live_group(cgrp))
1871 return -ENODEV;
1872 seq_puts(seq, cgrp->root->release_agent_path);
1873 seq_putc(seq, '\n');
1874 cgroup_unlock();
1875 return 0;
1876}
1877
1878/* A buffer size big enough for numbers or short strings */
1879#define CGROUP_LOCAL_BUFFER_SIZE 64
1880
1881static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
1882 struct file *file,
1883 const char __user *userbuf,
1884 size_t nbytes, loff_t *unused_ppos)
1885{
1886 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
1887 int retval = 0;
1888 char *end;
1889
1890 if (!nbytes)
1891 return -EINVAL;
1892 if (nbytes >= sizeof(buffer))
1893 return -E2BIG;
1894 if (copy_from_user(buffer, userbuf, nbytes))
1895 return -EFAULT;
1896
1897 buffer[nbytes] = 0; /* nul-terminate */
1898 if (cft->write_u64) {
1899 u64 val = simple_strtoull(strstrip(buffer), &end, 0);
1900 if (*end)
1901 return -EINVAL;
1902 retval = cft->write_u64(cgrp, cft, val);
1903 } else {
1904 s64 val = simple_strtoll(strstrip(buffer), &end, 0);
1905 if (*end)
1906 return -EINVAL;
1907 retval = cft->write_s64(cgrp, cft, val);
1908 }
1909 if (!retval)
1910 retval = nbytes;
1911 return retval;
1912}
1913
1914static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
1915 struct file *file,
1916 const char __user *userbuf,
1917 size_t nbytes, loff_t *unused_ppos)
1918{
1919 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
1920 int retval = 0;
1921 size_t max_bytes = cft->max_write_len;
1922 char *buffer = local_buffer;
1923
1924 if (!max_bytes)
1925 max_bytes = sizeof(local_buffer) - 1;
1926 if (nbytes >= max_bytes)
1927 return -E2BIG;
1928 /* Allocate a dynamic buffer if we need one */
1929 if (nbytes >= sizeof(local_buffer)) {
1930 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
1931 if (buffer == NULL)
1932 return -ENOMEM;
1933 }
1934 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
1935 retval = -EFAULT;
1936 goto out;
1937 }
1938
1939 buffer[nbytes] = 0; /* nul-terminate */
1940 retval = cft->write_string(cgrp, cft, strstrip(buffer));
1941 if (!retval)
1942 retval = nbytes;
1943out:
1944 if (buffer != local_buffer)
1945 kfree(buffer);
1946 return retval;
1947}
1948
1949static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
1950 size_t nbytes, loff_t *ppos)
1951{
1952 struct cftype *cft = __d_cft(file->f_dentry);
1953 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1954
1955 if (cgroup_is_removed(cgrp))
1956 return -ENODEV;
1957 if (cft->write)
1958 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
1959 if (cft->write_u64 || cft->write_s64)
1960 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
1961 if (cft->write_string)
1962 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
1963 if (cft->trigger) {
1964 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
1965 return ret ? ret : nbytes;
1966 }
1967 return -EINVAL;
1968}
1969
1970static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
1971 struct file *file,
1972 char __user *buf, size_t nbytes,
1973 loff_t *ppos)
1974{
1975 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
1976 u64 val = cft->read_u64(cgrp, cft);
1977 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
1978
1979 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1980}
1981
1982static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
1983 struct file *file,
1984 char __user *buf, size_t nbytes,
1985 loff_t *ppos)
1986{
1987 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
1988 s64 val = cft->read_s64(cgrp, cft);
1989 int len = sprintf(tmp, "%lld\n", (long long) val);
1990
1991 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1992}
1993
1994static ssize_t cgroup_file_read(struct file *file, char __user *buf,
1995 size_t nbytes, loff_t *ppos)
1996{
1997 struct cftype *cft = __d_cft(file->f_dentry);
1998 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1999
2000 if (cgroup_is_removed(cgrp))
2001 return -ENODEV;
2002
2003 if (cft->read)
2004 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2005 if (cft->read_u64)
2006 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2007 if (cft->read_s64)
2008 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2009 return -EINVAL;
2010}
2011
2012/*
2013 * seqfile ops/methods for returning structured data. Currently just
2014 * supports string->u64 maps, but can be extended in future.
2015 */
2016
2017struct cgroup_seqfile_state {
2018 struct cftype *cft;
2019 struct cgroup *cgroup;
2020};
2021
2022static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2023{
2024 struct seq_file *sf = cb->state;
2025 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2026}
2027
2028static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2029{
2030 struct cgroup_seqfile_state *state = m->private;
2031 struct cftype *cft = state->cft;
2032 if (cft->read_map) {
2033 struct cgroup_map_cb cb = {
2034 .fill = cgroup_map_add,
2035 .state = m,
2036 };
2037 return cft->read_map(state->cgroup, cft, &cb);
2038 }
2039 return cft->read_seq_string(state->cgroup, cft, m);
2040}
2041
2042static int cgroup_seqfile_release(struct inode *inode, struct file *file)
2043{
2044 struct seq_file *seq = file->private_data;
2045 kfree(seq->private);
2046 return single_release(inode, file);
2047}
2048
2049static const struct file_operations cgroup_seqfile_operations = {
2050 .read = seq_read,
2051 .write = cgroup_file_write,
2052 .llseek = seq_lseek,
2053 .release = cgroup_seqfile_release,
2054};
2055
2056static int cgroup_file_open(struct inode *inode, struct file *file)
2057{
2058 int err;
2059 struct cftype *cft;
2060
2061 err = generic_file_open(inode, file);
2062 if (err)
2063 return err;
2064 cft = __d_cft(file->f_dentry);
2065
2066 if (cft->read_map || cft->read_seq_string) {
2067 struct cgroup_seqfile_state *state =
2068 kzalloc(sizeof(*state), GFP_USER);
2069 if (!state)
2070 return -ENOMEM;
2071 state->cft = cft;
2072 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
2073 file->f_op = &cgroup_seqfile_operations;
2074 err = single_open(file, cgroup_seqfile_show, state);
2075 if (err < 0)
2076 kfree(state);
2077 } else if (cft->open)
2078 err = cft->open(inode, file);
2079 else
2080 err = 0;
2081
2082 return err;
2083}
2084
2085static int cgroup_file_release(struct inode *inode, struct file *file)
2086{
2087 struct cftype *cft = __d_cft(file->f_dentry);
2088 if (cft->release)
2089 return cft->release(inode, file);
2090 return 0;
2091}
2092
2093/*
2094 * cgroup_rename - Only allow simple rename of directories in place.
2095 */
2096static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2097 struct inode *new_dir, struct dentry *new_dentry)
2098{
2099 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2100 return -ENOTDIR;
2101 if (new_dentry->d_inode)
2102 return -EEXIST;
2103 if (old_dir != new_dir)
2104 return -EIO;
2105 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2106}
2107
2108static const struct file_operations cgroup_file_operations = {
2109 .read = cgroup_file_read,
2110 .write = cgroup_file_write,
2111 .llseek = generic_file_llseek,
2112 .open = cgroup_file_open,
2113 .release = cgroup_file_release,
2114};
2115
2116static const struct inode_operations cgroup_dir_inode_operations = {
2117 .lookup = simple_lookup,
2118 .mkdir = cgroup_mkdir,
2119 .rmdir = cgroup_rmdir,
2120 .rename = cgroup_rename,
2121};
2122
2123/*
2124 * Check if a file is a control file
2125 */
2126static inline struct cftype *__file_cft(struct file *file)
2127{
2128 if (file->f_dentry->d_inode->i_fop != &cgroup_file_operations)
2129 return ERR_PTR(-EINVAL);
2130 return __d_cft(file->f_dentry);
2131}
2132
2133static int cgroup_create_file(struct dentry *dentry, mode_t mode,
2134 struct super_block *sb)
2135{
2136 static const struct dentry_operations cgroup_dops = {
2137 .d_iput = cgroup_diput,
2138 };
2139
2140 struct inode *inode;
2141
2142 if (!dentry)
2143 return -ENOENT;
2144 if (dentry->d_inode)
2145 return -EEXIST;
2146
2147 inode = cgroup_new_inode(mode, sb);
2148 if (!inode)
2149 return -ENOMEM;
2150
2151 if (S_ISDIR(mode)) {
2152 inode->i_op = &cgroup_dir_inode_operations;
2153 inode->i_fop = &simple_dir_operations;
2154
2155 /* start off with i_nlink == 2 (for "." entry) */
2156 inc_nlink(inode);
2157
2158 /* start with the directory inode held, so that we can
2159 * populate it without racing with another mkdir */
2160 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2161 } else if (S_ISREG(mode)) {
2162 inode->i_size = 0;
2163 inode->i_fop = &cgroup_file_operations;
2164 }
2165 dentry->d_op = &cgroup_dops;
2166 d_instantiate(dentry, inode);
2167 dget(dentry); /* Extra count - pin the dentry in core */
2168 return 0;
2169}
2170
2171/*
2172 * cgroup_create_dir - create a directory for an object.
2173 * @cgrp: the cgroup we create the directory for. It must have a valid
2174 * ->parent field. And we are going to fill its ->dentry field.
2175 * @dentry: dentry of the new cgroup
2176 * @mode: mode to set on new directory.
2177 */
2178static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
2179 mode_t mode)
2180{
2181 struct dentry *parent;
2182 int error = 0;
2183
2184 parent = cgrp->parent->dentry;
2185 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
2186 if (!error) {
2187 dentry->d_fsdata = cgrp;
2188 inc_nlink(parent->d_inode);
2189 rcu_assign_pointer(cgrp->dentry, dentry);
2190 dget(dentry);
2191 }
2192 dput(dentry);
2193
2194 return error;
2195}
2196
2197/**
2198 * cgroup_file_mode - deduce file mode of a control file
2199 * @cft: the control file in question
2200 *
2201 * returns cft->mode if ->mode is not 0
2202 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2203 * returns S_IRUGO if it has only a read handler
2204 * returns S_IWUSR if it has only a write hander
2205 */
2206static mode_t cgroup_file_mode(const struct cftype *cft)
2207{
2208 mode_t mode = 0;
2209
2210 if (cft->mode)
2211 return cft->mode;
2212
2213 if (cft->read || cft->read_u64 || cft->read_s64 ||
2214 cft->read_map || cft->read_seq_string)
2215 mode |= S_IRUGO;
2216
2217 if (cft->write || cft->write_u64 || cft->write_s64 ||
2218 cft->write_string || cft->trigger)
2219 mode |= S_IWUSR;
2220
2221 return mode;
2222}
2223
2224int cgroup_add_file(struct cgroup *cgrp,
2225 struct cgroup_subsys *subsys,
2226 const struct cftype *cft)
2227{
2228 struct dentry *dir = cgrp->dentry;
2229 struct dentry *dentry;
2230 int error;
2231 mode_t mode;
2232
2233 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2234 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
2235 strcpy(name, subsys->name);
2236 strcat(name, ".");
2237 }
2238 strcat(name, cft->name);
2239 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2240 dentry = lookup_one_len(name, dir, strlen(name));
2241 if (!IS_ERR(dentry)) {
2242 mode = cgroup_file_mode(cft);
2243 error = cgroup_create_file(dentry, mode | S_IFREG,
2244 cgrp->root->sb);
2245 if (!error)
2246 dentry->d_fsdata = (void *)cft;
2247 dput(dentry);
2248 } else
2249 error = PTR_ERR(dentry);
2250 return error;
2251}
2252EXPORT_SYMBOL_GPL(cgroup_add_file);
2253
2254int cgroup_add_files(struct cgroup *cgrp,
2255 struct cgroup_subsys *subsys,
2256 const struct cftype cft[],
2257 int count)
2258{
2259 int i, err;
2260 for (i = 0; i < count; i++) {
2261 err = cgroup_add_file(cgrp, subsys, &cft[i]);
2262 if (err)
2263 return err;
2264 }
2265 return 0;
2266}
2267EXPORT_SYMBOL_GPL(cgroup_add_files);
2268
2269/**
2270 * cgroup_task_count - count the number of tasks in a cgroup.
2271 * @cgrp: the cgroup in question
2272 *
2273 * Return the number of tasks in the cgroup.
2274 */
2275int cgroup_task_count(const struct cgroup *cgrp)
2276{
2277 int count = 0;
2278 struct cg_cgroup_link *link;
2279
2280 read_lock(&css_set_lock);
2281 list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
2282 count += atomic_read(&link->cg->refcount);
2283 }
2284 read_unlock(&css_set_lock);
2285 return count;
2286}
2287
2288/*
2289 * Advance a list_head iterator. The iterator should be positioned at
2290 * the start of a css_set
2291 */
2292static void cgroup_advance_iter(struct cgroup *cgrp,
2293 struct cgroup_iter *it)
2294{
2295 struct list_head *l = it->cg_link;
2296 struct cg_cgroup_link *link;
2297 struct css_set *cg;
2298
2299 /* Advance to the next non-empty css_set */
2300 do {
2301 l = l->next;
2302 if (l == &cgrp->css_sets) {
2303 it->cg_link = NULL;
2304 return;
2305 }
2306 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
2307 cg = link->cg;
2308 } while (list_empty(&cg->tasks));
2309 it->cg_link = l;
2310 it->task = cg->tasks.next;
2311}
2312
2313/*
2314 * To reduce the fork() overhead for systems that are not actually
2315 * using their cgroups capability, we don't maintain the lists running
2316 * through each css_set to its tasks until we see the list actually
2317 * used - in other words after the first call to cgroup_iter_start().
2318 *
2319 * The tasklist_lock is not held here, as do_each_thread() and
2320 * while_each_thread() are protected by RCU.
2321 */
2322static void cgroup_enable_task_cg_lists(void)
2323{
2324 struct task_struct *p, *g;
2325 write_lock(&css_set_lock);
2326 use_task_css_set_links = 1;
2327 do_each_thread(g, p) {
2328 task_lock(p);
2329 /*
2330 * We should check if the process is exiting, otherwise
2331 * it will race with cgroup_exit() in that the list
2332 * entry won't be deleted though the process has exited.
2333 */
2334 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2335 list_add(&p->cg_list, &p->cgroups->tasks);
2336 task_unlock(p);
2337 } while_each_thread(g, p);
2338 write_unlock(&css_set_lock);
2339}
2340
2341void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
2342{
2343 /*
2344 * The first time anyone tries to iterate across a cgroup,
2345 * we need to enable the list linking each css_set to its
2346 * tasks, and fix up all existing tasks.
2347 */
2348 if (!use_task_css_set_links)
2349 cgroup_enable_task_cg_lists();
2350
2351 read_lock(&css_set_lock);
2352 it->cg_link = &cgrp->css_sets;
2353 cgroup_advance_iter(cgrp, it);
2354}
2355
2356struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
2357 struct cgroup_iter *it)
2358{
2359 struct task_struct *res;
2360 struct list_head *l = it->task;
2361 struct cg_cgroup_link *link;
2362
2363 /* If the iterator cg is NULL, we have no tasks */
2364 if (!it->cg_link)
2365 return NULL;
2366 res = list_entry(l, struct task_struct, cg_list);
2367 /* Advance iterator to find next entry */
2368 l = l->next;
2369 link = list_entry(it->cg_link, struct cg_cgroup_link, cgrp_link_list);
2370 if (l == &link->cg->tasks) {
2371 /* We reached the end of this task list - move on to
2372 * the next cg_cgroup_link */
2373 cgroup_advance_iter(cgrp, it);
2374 } else {
2375 it->task = l;
2376 }
2377 return res;
2378}
2379
2380void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
2381{
2382 read_unlock(&css_set_lock);
2383}
2384
2385static inline int started_after_time(struct task_struct *t1,
2386 struct timespec *time,
2387 struct task_struct *t2)
2388{
2389 int start_diff = timespec_compare(&t1->start_time, time);
2390 if (start_diff > 0) {
2391 return 1;
2392 } else if (start_diff < 0) {
2393 return 0;
2394 } else {
2395 /*
2396 * Arbitrarily, if two processes started at the same
2397 * time, we'll say that the lower pointer value
2398 * started first. Note that t2 may have exited by now
2399 * so this may not be a valid pointer any longer, but
2400 * that's fine - it still serves to distinguish
2401 * between two tasks started (effectively) simultaneously.
2402 */
2403 return t1 > t2;
2404 }
2405}
2406
2407/*
2408 * This function is a callback from heap_insert() and is used to order
2409 * the heap.
2410 * In this case we order the heap in descending task start time.
2411 */
2412static inline int started_after(void *p1, void *p2)
2413{
2414 struct task_struct *t1 = p1;
2415 struct task_struct *t2 = p2;
2416 return started_after_time(t1, &t2->start_time, t2);
2417}
2418
2419/**
2420 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
2421 * @scan: struct cgroup_scanner containing arguments for the scan
2422 *
2423 * Arguments include pointers to callback functions test_task() and
2424 * process_task().
2425 * Iterate through all the tasks in a cgroup, calling test_task() for each,
2426 * and if it returns true, call process_task() for it also.
2427 * The test_task pointer may be NULL, meaning always true (select all tasks).
2428 * Effectively duplicates cgroup_iter_{start,next,end}()
2429 * but does not lock css_set_lock for the call to process_task().
2430 * The struct cgroup_scanner may be embedded in any structure of the caller's
2431 * creation.
2432 * It is guaranteed that process_task() will act on every task that
2433 * is a member of the cgroup for the duration of this call. This
2434 * function may or may not call process_task() for tasks that exit
2435 * or move to a different cgroup during the call, or are forked or
2436 * move into the cgroup during the call.
2437 *
2438 * Note that test_task() may be called with locks held, and may in some
2439 * situations be called multiple times for the same task, so it should
2440 * be cheap.
2441 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
2442 * pre-allocated and will be used for heap operations (and its "gt" member will
2443 * be overwritten), else a temporary heap will be used (allocation of which
2444 * may cause this function to fail).
2445 */
2446int cgroup_scan_tasks(struct cgroup_scanner *scan)
2447{
2448 int retval, i;
2449 struct cgroup_iter it;
2450 struct task_struct *p, *dropped;
2451 /* Never dereference latest_task, since it's not refcounted */
2452 struct task_struct *latest_task = NULL;
2453 struct ptr_heap tmp_heap;
2454 struct ptr_heap *heap;
2455 struct timespec latest_time = { 0, 0 };
2456
2457 if (scan->heap) {
2458 /* The caller supplied our heap and pre-allocated its memory */
2459 heap = scan->heap;
2460 heap->gt = &started_after;
2461 } else {
2462 /* We need to allocate our own heap memory */
2463 heap = &tmp_heap;
2464 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2465 if (retval)
2466 /* cannot allocate the heap */
2467 return retval;
2468 }
2469
2470 again:
2471 /*
2472 * Scan tasks in the cgroup, using the scanner's "test_task" callback
2473 * to determine which are of interest, and using the scanner's
2474 * "process_task" callback to process any of them that need an update.
2475 * Since we don't want to hold any locks during the task updates,
2476 * gather tasks to be processed in a heap structure.
2477 * The heap is sorted by descending task start time.
2478 * If the statically-sized heap fills up, we overflow tasks that
2479 * started later, and in future iterations only consider tasks that
2480 * started after the latest task in the previous pass. This
2481 * guarantees forward progress and that we don't miss any tasks.
2482 */
2483 heap->size = 0;
2484 cgroup_iter_start(scan->cg, &it);
2485 while ((p = cgroup_iter_next(scan->cg, &it))) {
2486 /*
2487 * Only affect tasks that qualify per the caller's callback,
2488 * if he provided one
2489 */
2490 if (scan->test_task && !scan->test_task(p, scan))
2491 continue;
2492 /*
2493 * Only process tasks that started after the last task
2494 * we processed
2495 */
2496 if (!started_after_time(p, &latest_time, latest_task))
2497 continue;
2498 dropped = heap_insert(heap, p);
2499 if (dropped == NULL) {
2500 /*
2501 * The new task was inserted; the heap wasn't
2502 * previously full
2503 */
2504 get_task_struct(p);
2505 } else if (dropped != p) {
2506 /*
2507 * The new task was inserted, and pushed out a
2508 * different task
2509 */
2510 get_task_struct(p);
2511 put_task_struct(dropped);
2512 }
2513 /*
2514 * Else the new task was newer than anything already in
2515 * the heap and wasn't inserted
2516 */
2517 }
2518 cgroup_iter_end(scan->cg, &it);
2519
2520 if (heap->size) {
2521 for (i = 0; i < heap->size; i++) {
2522 struct task_struct *q = heap->ptrs[i];
2523 if (i == 0) {
2524 latest_time = q->start_time;
2525 latest_task = q;
2526 }
2527 /* Process the task per the caller's callback */
2528 scan->process_task(q, scan);
2529 put_task_struct(q);
2530 }
2531 /*
2532 * If we had to process any tasks at all, scan again
2533 * in case some of them were in the middle of forking
2534 * children that didn't get processed.
2535 * Not the most efficient way to do it, but it avoids
2536 * having to take callback_mutex in the fork path
2537 */
2538 goto again;
2539 }
2540 if (heap == &tmp_heap)
2541 heap_free(&tmp_heap);
2542 return 0;
2543}
2544
2545/*
2546 * Stuff for reading the 'tasks'/'procs' files.
2547 *
2548 * Reading this file can return large amounts of data if a cgroup has
2549 * *lots* of attached tasks. So it may need several calls to read(),
2550 * but we cannot guarantee that the information we produce is correct
2551 * unless we produce it entirely atomically.
2552 *
2553 */
2554
2555/*
2556 * The following two functions "fix" the issue where there are more pids
2557 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2558 * TODO: replace with a kernel-wide solution to this problem
2559 */
2560#define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2561static void *pidlist_allocate(int count)
2562{
2563 if (PIDLIST_TOO_LARGE(count))
2564 return vmalloc(count * sizeof(pid_t));
2565 else
2566 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2567}
2568static void pidlist_free(void *p)
2569{
2570 if (is_vmalloc_addr(p))
2571 vfree(p);
2572 else
2573 kfree(p);
2574}
2575static void *pidlist_resize(void *p, int newcount)
2576{
2577 void *newlist;
2578 /* note: if new alloc fails, old p will still be valid either way */
2579 if (is_vmalloc_addr(p)) {
2580 newlist = vmalloc(newcount * sizeof(pid_t));
2581 if (!newlist)
2582 return NULL;
2583 memcpy(newlist, p, newcount * sizeof(pid_t));
2584 vfree(p);
2585 } else {
2586 newlist = krealloc(p, newcount * sizeof(pid_t), GFP_KERNEL);
2587 }
2588 return newlist;
2589}
2590
2591/*
2592 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2593 * If the new stripped list is sufficiently smaller and there's enough memory
2594 * to allocate a new buffer, will let go of the unneeded memory. Returns the
2595 * number of unique elements.
2596 */
2597/* is the size difference enough that we should re-allocate the array? */
2598#define PIDLIST_REALLOC_DIFFERENCE(old, new) ((old) - PAGE_SIZE >= (new))
2599static int pidlist_uniq(pid_t **p, int length)
2600{
2601 int src, dest = 1;
2602 pid_t *list = *p;
2603 pid_t *newlist;
2604
2605 /*
2606 * we presume the 0th element is unique, so i starts at 1. trivial
2607 * edge cases first; no work needs to be done for either
2608 */
2609 if (length == 0 || length == 1)
2610 return length;
2611 /* src and dest walk down the list; dest counts unique elements */
2612 for (src = 1; src < length; src++) {
2613 /* find next unique element */
2614 while (list[src] == list[src-1]) {
2615 src++;
2616 if (src == length)
2617 goto after;
2618 }
2619 /* dest always points to where the next unique element goes */
2620 list[dest] = list[src];
2621 dest++;
2622 }
2623after:
2624 /*
2625 * if the length difference is large enough, we want to allocate a
2626 * smaller buffer to save memory. if this fails due to out of memory,
2627 * we'll just stay with what we've got.
2628 */
2629 if (PIDLIST_REALLOC_DIFFERENCE(length, dest)) {
2630 newlist = pidlist_resize(list, dest);
2631 if (newlist)
2632 *p = newlist;
2633 }
2634 return dest;
2635}
2636
2637static int cmppid(const void *a, const void *b)
2638{
2639 return *(pid_t *)a - *(pid_t *)b;
2640}
2641
2642/*
2643 * find the appropriate pidlist for our purpose (given procs vs tasks)
2644 * returns with the lock on that pidlist already held, and takes care
2645 * of the use count, or returns NULL with no locks held if we're out of
2646 * memory.
2647 */
2648static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
2649 enum cgroup_filetype type)
2650{
2651 struct cgroup_pidlist *l;
2652 /* don't need task_nsproxy() if we're looking at ourself */
2653 struct pid_namespace *ns = current->nsproxy->pid_ns;
2654
2655 /*
2656 * We can't drop the pidlist_mutex before taking the l->mutex in case
2657 * the last ref-holder is trying to remove l from the list at the same
2658 * time. Holding the pidlist_mutex precludes somebody taking whichever
2659 * list we find out from under us - compare release_pid_array().
2660 */
2661 mutex_lock(&cgrp->pidlist_mutex);
2662 list_for_each_entry(l, &cgrp->pidlists, links) {
2663 if (l->key.type == type && l->key.ns == ns) {
2664 /* make sure l doesn't vanish out from under us */
2665 down_write(&l->mutex);
2666 mutex_unlock(&cgrp->pidlist_mutex);
2667 return l;
2668 }
2669 }
2670 /* entry not found; create a new one */
2671 l = kmalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
2672 if (!l) {
2673 mutex_unlock(&cgrp->pidlist_mutex);
2674 return l;
2675 }
2676 init_rwsem(&l->mutex);
2677 down_write(&l->mutex);
2678 l->key.type = type;
2679 l->key.ns = get_pid_ns(ns);
2680 l->use_count = 0; /* don't increment here */
2681 l->list = NULL;
2682 l->owner = cgrp;
2683 list_add(&l->links, &cgrp->pidlists);
2684 mutex_unlock(&cgrp->pidlist_mutex);
2685 return l;
2686}
2687
2688/*
2689 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
2690 */
2691static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
2692 struct cgroup_pidlist **lp)
2693{
2694 pid_t *array;
2695 int length;
2696 int pid, n = 0; /* used for populating the array */
2697 struct cgroup_iter it;
2698 struct task_struct *tsk;
2699 struct cgroup_pidlist *l;
2700
2701 /*
2702 * If cgroup gets more users after we read count, we won't have
2703 * enough space - tough. This race is indistinguishable to the
2704 * caller from the case that the additional cgroup users didn't
2705 * show up until sometime later on.
2706 */
2707 length = cgroup_task_count(cgrp);
2708 array = pidlist_allocate(length);
2709 if (!array)
2710 return -ENOMEM;
2711 /* now, populate the array */
2712 cgroup_iter_start(cgrp, &it);
2713 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2714 if (unlikely(n == length))
2715 break;
2716 /* get tgid or pid for procs or tasks file respectively */
2717 if (type == CGROUP_FILE_PROCS)
2718 pid = task_tgid_vnr(tsk);
2719 else
2720 pid = task_pid_vnr(tsk);
2721 if (pid > 0) /* make sure to only use valid results */
2722 array[n++] = pid;
2723 }
2724 cgroup_iter_end(cgrp, &it);
2725 length = n;
2726 /* now sort & (if procs) strip out duplicates */
2727 sort(array, length, sizeof(pid_t), cmppid, NULL);
2728 if (type == CGROUP_FILE_PROCS)
2729 length = pidlist_uniq(&array, length);
2730 l = cgroup_pidlist_find(cgrp, type);
2731 if (!l) {
2732 pidlist_free(array);
2733 return -ENOMEM;
2734 }
2735 /* store array, freeing old if necessary - lock already held */
2736 pidlist_free(l->list);
2737 l->list = array;
2738 l->length = length;
2739 l->use_count++;
2740 up_write(&l->mutex);
2741 *lp = l;
2742 return 0;
2743}
2744
2745/**
2746 * cgroupstats_build - build and fill cgroupstats
2747 * @stats: cgroupstats to fill information into
2748 * @dentry: A dentry entry belonging to the cgroup for which stats have
2749 * been requested.
2750 *
2751 * Build and fill cgroupstats so that taskstats can export it to user
2752 * space.
2753 */
2754int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2755{
2756 int ret = -EINVAL;
2757 struct cgroup *cgrp;
2758 struct cgroup_iter it;
2759 struct task_struct *tsk;
2760
2761 /*
2762 * Validate dentry by checking the superblock operations,
2763 * and make sure it's a directory.
2764 */
2765 if (dentry->d_sb->s_op != &cgroup_ops ||
2766 !S_ISDIR(dentry->d_inode->i_mode))
2767 goto err;
2768
2769 ret = 0;
2770 cgrp = dentry->d_fsdata;
2771
2772 cgroup_iter_start(cgrp, &it);
2773 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2774 switch (tsk->state) {
2775 case TASK_RUNNING:
2776 stats->nr_running++;
2777 break;
2778 case TASK_INTERRUPTIBLE:
2779 stats->nr_sleeping++;
2780 break;
2781 case TASK_UNINTERRUPTIBLE:
2782 stats->nr_uninterruptible++;
2783 break;
2784 case TASK_STOPPED:
2785 stats->nr_stopped++;
2786 break;
2787 default:
2788 if (delayacct_is_task_waiting_on_io(tsk))
2789 stats->nr_io_wait++;
2790 break;
2791 }
2792 }
2793 cgroup_iter_end(cgrp, &it);
2794
2795err:
2796 return ret;
2797}
2798
2799
2800/*
2801 * seq_file methods for the tasks/procs files. The seq_file position is the
2802 * next pid to display; the seq_file iterator is a pointer to the pid
2803 * in the cgroup->l->list array.
2804 */
2805
2806static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
2807{
2808 /*
2809 * Initially we receive a position value that corresponds to
2810 * one more than the last pid shown (or 0 on the first call or
2811 * after a seek to the start). Use a binary-search to find the
2812 * next pid to display, if any
2813 */
2814 struct cgroup_pidlist *l = s->private;
2815 int index = 0, pid = *pos;
2816 int *iter;
2817
2818 down_read(&l->mutex);
2819 if (pid) {
2820 int end = l->length;
2821
2822 while (index < end) {
2823 int mid = (index + end) / 2;
2824 if (l->list[mid] == pid) {
2825 index = mid;
2826 break;
2827 } else if (l->list[mid] <= pid)
2828 index = mid + 1;
2829 else
2830 end = mid;
2831 }
2832 }
2833 /* If we're off the end of the array, we're done */
2834 if (index >= l->length)
2835 return NULL;
2836 /* Update the abstract position to be the actual pid that we found */
2837 iter = l->list + index;
2838 *pos = *iter;
2839 return iter;
2840}
2841
2842static void cgroup_pidlist_stop(struct seq_file *s, void *v)
2843{
2844 struct cgroup_pidlist *l = s->private;
2845 up_read(&l->mutex);
2846}
2847
2848static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
2849{
2850 struct cgroup_pidlist *l = s->private;
2851 pid_t *p = v;
2852 pid_t *end = l->list + l->length;
2853 /*
2854 * Advance to the next pid in the array. If this goes off the
2855 * end, we're done
2856 */
2857 p++;
2858 if (p >= end) {
2859 return NULL;
2860 } else {
2861 *pos = *p;
2862 return p;
2863 }
2864}
2865
2866static int cgroup_pidlist_show(struct seq_file *s, void *v)
2867{
2868 return seq_printf(s, "%d\n", *(int *)v);
2869}
2870
2871/*
2872 * seq_operations functions for iterating on pidlists through seq_file -
2873 * independent of whether it's tasks or procs
2874 */
2875static const struct seq_operations cgroup_pidlist_seq_operations = {
2876 .start = cgroup_pidlist_start,
2877 .stop = cgroup_pidlist_stop,
2878 .next = cgroup_pidlist_next,
2879 .show = cgroup_pidlist_show,
2880};
2881
2882static void cgroup_release_pid_array(struct cgroup_pidlist *l)
2883{
2884 /*
2885 * the case where we're the last user of this particular pidlist will
2886 * have us remove it from the cgroup's list, which entails taking the
2887 * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
2888 * pidlist_mutex, we have to take pidlist_mutex first.
2889 */
2890 mutex_lock(&l->owner->pidlist_mutex);
2891 down_write(&l->mutex);
2892 BUG_ON(!l->use_count);
2893 if (!--l->use_count) {
2894 /* we're the last user if refcount is 0; remove and free */
2895 list_del(&l->links);
2896 mutex_unlock(&l->owner->pidlist_mutex);
2897 pidlist_free(l->list);
2898 put_pid_ns(l->key.ns);
2899 up_write(&l->mutex);
2900 kfree(l);
2901 return;
2902 }
2903 mutex_unlock(&l->owner->pidlist_mutex);
2904 up_write(&l->mutex);
2905}
2906
2907static int cgroup_pidlist_release(struct inode *inode, struct file *file)
2908{
2909 struct cgroup_pidlist *l;
2910 if (!(file->f_mode & FMODE_READ))
2911 return 0;
2912 /*
2913 * the seq_file will only be initialized if the file was opened for
2914 * reading; hence we check if it's not null only in that case.
2915 */
2916 l = ((struct seq_file *)file->private_data)->private;
2917 cgroup_release_pid_array(l);
2918 return seq_release(inode, file);
2919}
2920
2921static const struct file_operations cgroup_pidlist_operations = {
2922 .read = seq_read,
2923 .llseek = seq_lseek,
2924 .write = cgroup_file_write,
2925 .release = cgroup_pidlist_release,
2926};
2927
2928/*
2929 * The following functions handle opens on a file that displays a pidlist
2930 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
2931 * in the cgroup.
2932 */
2933/* helper function for the two below it */
2934static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
2935{
2936 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2937 struct cgroup_pidlist *l;
2938 int retval;
2939
2940 /* Nothing to do for write-only files */
2941 if (!(file->f_mode & FMODE_READ))
2942 return 0;
2943
2944 /* have the array populated */
2945 retval = pidlist_array_load(cgrp, type, &l);
2946 if (retval)
2947 return retval;
2948 /* configure file information */
2949 file->f_op = &cgroup_pidlist_operations;
2950
2951 retval = seq_open(file, &cgroup_pidlist_seq_operations);
2952 if (retval) {
2953 cgroup_release_pid_array(l);
2954 return retval;
2955 }
2956 ((struct seq_file *)file->private_data)->private = l;
2957 return 0;
2958}
2959static int cgroup_tasks_open(struct inode *unused, struct file *file)
2960{
2961 return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
2962}
2963static int cgroup_procs_open(struct inode *unused, struct file *file)
2964{
2965 return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
2966}
2967
2968static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
2969 struct cftype *cft)
2970{
2971 return notify_on_release(cgrp);
2972}
2973
2974static int cgroup_write_notify_on_release(struct cgroup *cgrp,
2975 struct cftype *cft,
2976 u64 val)
2977{
2978 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
2979 if (val)
2980 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2981 else
2982 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2983 return 0;
2984}
2985
2986/*
2987 * Unregister event and free resources.
2988 *
2989 * Gets called from workqueue.
2990 */
2991static void cgroup_event_remove(struct work_struct *work)
2992{
2993 struct cgroup_event *event = container_of(work, struct cgroup_event,
2994 remove);
2995 struct cgroup *cgrp = event->cgrp;
2996
2997 /* TODO: check return code */
2998 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
2999
3000 eventfd_ctx_put(event->eventfd);
3001 kfree(event);
3002 dput(cgrp->dentry);
3003}
3004
3005/*
3006 * Gets called on POLLHUP on eventfd when user closes it.
3007 *
3008 * Called with wqh->lock held and interrupts disabled.
3009 */
3010static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3011 int sync, void *key)
3012{
3013 struct cgroup_event *event = container_of(wait,
3014 struct cgroup_event, wait);
3015 struct cgroup *cgrp = event->cgrp;
3016 unsigned long flags = (unsigned long)key;
3017
3018 if (flags & POLLHUP) {
3019 remove_wait_queue_locked(event->wqh, &event->wait);
3020 spin_lock(&cgrp->event_list_lock);
3021 list_del(&event->list);
3022 spin_unlock(&cgrp->event_list_lock);
3023 /*
3024 * We are in atomic context, but cgroup_event_remove() may
3025 * sleep, so we have to call it in workqueue.
3026 */
3027 schedule_work(&event->remove);
3028 }
3029
3030 return 0;
3031}
3032
3033static void cgroup_event_ptable_queue_proc(struct file *file,
3034 wait_queue_head_t *wqh, poll_table *pt)
3035{
3036 struct cgroup_event *event = container_of(pt,
3037 struct cgroup_event, pt);
3038
3039 event->wqh = wqh;
3040 add_wait_queue(wqh, &event->wait);
3041}
3042
3043/*
3044 * Parse input and register new cgroup event handler.
3045 *
3046 * Input must be in format '<event_fd> <control_fd> <args>'.
3047 * Interpretation of args is defined by control file implementation.
3048 */
3049static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3050 const char *buffer)
3051{
3052 struct cgroup_event *event = NULL;
3053 unsigned int efd, cfd;
3054 struct file *efile = NULL;
3055 struct file *cfile = NULL;
3056 char *endp;
3057 int ret;
3058
3059 efd = simple_strtoul(buffer, &endp, 10);
3060 if (*endp != ' ')
3061 return -EINVAL;
3062 buffer = endp + 1;
3063
3064 cfd = simple_strtoul(buffer, &endp, 10);
3065 if ((*endp != ' ') && (*endp != '\0'))
3066 return -EINVAL;
3067 buffer = endp + 1;
3068
3069 event = kzalloc(sizeof(*event), GFP_KERNEL);
3070 if (!event)
3071 return -ENOMEM;
3072 event->cgrp = cgrp;
3073 INIT_LIST_HEAD(&event->list);
3074 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
3075 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
3076 INIT_WORK(&event->remove, cgroup_event_remove);
3077
3078 efile = eventfd_fget(efd);
3079 if (IS_ERR(efile)) {
3080 ret = PTR_ERR(efile);
3081 goto fail;
3082 }
3083
3084 event->eventfd = eventfd_ctx_fileget(efile);
3085 if (IS_ERR(event->eventfd)) {
3086 ret = PTR_ERR(event->eventfd);
3087 goto fail;
3088 }
3089
3090 cfile = fget(cfd);
3091 if (!cfile) {
3092 ret = -EBADF;
3093 goto fail;
3094 }
3095
3096 /* the process need read permission on control file */
3097 ret = file_permission(cfile, MAY_READ);
3098 if (ret < 0)
3099 goto fail;
3100
3101 event->cft = __file_cft(cfile);
3102 if (IS_ERR(event->cft)) {
3103 ret = PTR_ERR(event->cft);
3104 goto fail;
3105 }
3106
3107 if (!event->cft->register_event || !event->cft->unregister_event) {
3108 ret = -EINVAL;
3109 goto fail;
3110 }
3111
3112 ret = event->cft->register_event(cgrp, event->cft,
3113 event->eventfd, buffer);
3114 if (ret)
3115 goto fail;
3116
3117 if (efile->f_op->poll(efile, &event->pt) & POLLHUP) {
3118 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3119 ret = 0;
3120 goto fail;
3121 }
3122
3123 /*
3124 * Events should be removed after rmdir of cgroup directory, but before
3125 * destroying subsystem state objects. Let's take reference to cgroup
3126 * directory dentry to do that.
3127 */
3128 dget(cgrp->dentry);
3129
3130 spin_lock(&cgrp->event_list_lock);
3131 list_add(&event->list, &cgrp->event_list);
3132 spin_unlock(&cgrp->event_list_lock);
3133
3134 fput(cfile);
3135 fput(efile);
3136
3137 return 0;
3138
3139fail:
3140 if (cfile)
3141 fput(cfile);
3142
3143 if (event && event->eventfd && !IS_ERR(event->eventfd))
3144 eventfd_ctx_put(event->eventfd);
3145
3146 if (!IS_ERR_OR_NULL(efile))
3147 fput(efile);
3148
3149 kfree(event);
3150
3151 return ret;
3152}
3153
3154/*
3155 * for the common functions, 'private' gives the type of file
3156 */
3157/* for hysterical raisins, we can't put this on the older files */
3158#define CGROUP_FILE_GENERIC_PREFIX "cgroup."
3159static struct cftype files[] = {
3160 {
3161 .name = "tasks",
3162 .open = cgroup_tasks_open,
3163 .write_u64 = cgroup_tasks_write,
3164 .release = cgroup_pidlist_release,
3165 .mode = S_IRUGO | S_IWUSR,
3166 },
3167 {
3168 .name = CGROUP_FILE_GENERIC_PREFIX "procs",
3169 .open = cgroup_procs_open,
3170 /* .write_u64 = cgroup_procs_write, TODO */
3171 .release = cgroup_pidlist_release,
3172 .mode = S_IRUGO,
3173 },
3174 {
3175 .name = "notify_on_release",
3176 .read_u64 = cgroup_read_notify_on_release,
3177 .write_u64 = cgroup_write_notify_on_release,
3178 },
3179 {
3180 .name = CGROUP_FILE_GENERIC_PREFIX "event_control",
3181 .write_string = cgroup_write_event_control,
3182 .mode = S_IWUGO,
3183 },
3184};
3185
3186static struct cftype cft_release_agent = {
3187 .name = "release_agent",
3188 .read_seq_string = cgroup_release_agent_show,
3189 .write_string = cgroup_release_agent_write,
3190 .max_write_len = PATH_MAX,
3191};
3192
3193static int cgroup_populate_dir(struct cgroup *cgrp)
3194{
3195 int err;
3196 struct cgroup_subsys *ss;
3197
3198 /* First clear out any existing files */
3199 cgroup_clear_directory(cgrp->dentry);
3200
3201 err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
3202 if (err < 0)
3203 return err;
3204
3205 if (cgrp == cgrp->top_cgroup) {
3206 if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
3207 return err;
3208 }
3209
3210 for_each_subsys(cgrp->root, ss) {
3211 if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
3212 return err;
3213 }
3214 /* This cgroup is ready now */
3215 for_each_subsys(cgrp->root, ss) {
3216 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3217 /*
3218 * Update id->css pointer and make this css visible from
3219 * CSS ID functions. This pointer will be dereferened
3220 * from RCU-read-side without locks.
3221 */
3222 if (css->id)
3223 rcu_assign_pointer(css->id->css, css);
3224 }
3225
3226 return 0;
3227}
3228
3229static void init_cgroup_css(struct cgroup_subsys_state *css,
3230 struct cgroup_subsys *ss,
3231 struct cgroup *cgrp)
3232{
3233 css->cgroup = cgrp;
3234 atomic_set(&css->refcnt, 1);
3235 css->flags = 0;
3236 css->id = NULL;
3237 if (cgrp == dummytop)
3238 set_bit(CSS_ROOT, &css->flags);
3239 BUG_ON(cgrp->subsys[ss->subsys_id]);
3240 cgrp->subsys[ss->subsys_id] = css;
3241}
3242
3243static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
3244{
3245 /* We need to take each hierarchy_mutex in a consistent order */
3246 int i;
3247
3248 /*
3249 * No worry about a race with rebind_subsystems that might mess up the
3250 * locking order, since both parties are under cgroup_mutex.
3251 */
3252 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3253 struct cgroup_subsys *ss = subsys[i];
3254 if (ss == NULL)
3255 continue;
3256 if (ss->root == root)
3257 mutex_lock(&ss->hierarchy_mutex);
3258 }
3259}
3260
3261static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
3262{
3263 int i;
3264
3265 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3266 struct cgroup_subsys *ss = subsys[i];
3267 if (ss == NULL)
3268 continue;
3269 if (ss->root == root)
3270 mutex_unlock(&ss->hierarchy_mutex);
3271 }
3272}
3273
3274/*
3275 * cgroup_create - create a cgroup
3276 * @parent: cgroup that will be parent of the new cgroup
3277 * @dentry: dentry of the new cgroup
3278 * @mode: mode to set on new inode
3279 *
3280 * Must be called with the mutex on the parent inode held
3281 */
3282static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
3283 mode_t mode)
3284{
3285 struct cgroup *cgrp;
3286 struct cgroupfs_root *root = parent->root;
3287 int err = 0;
3288 struct cgroup_subsys *ss;
3289 struct super_block *sb = root->sb;
3290
3291 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3292 if (!cgrp)
3293 return -ENOMEM;
3294
3295 /* Grab a reference on the superblock so the hierarchy doesn't
3296 * get deleted on unmount if there are child cgroups. This
3297 * can be done outside cgroup_mutex, since the sb can't
3298 * disappear while someone has an open control file on the
3299 * fs */
3300 atomic_inc(&sb->s_active);
3301
3302 mutex_lock(&cgroup_mutex);
3303
3304 init_cgroup_housekeeping(cgrp);
3305
3306 cgrp->parent = parent;
3307 cgrp->root = parent->root;
3308 cgrp->top_cgroup = parent->top_cgroup;
3309
3310 if (notify_on_release(parent))
3311 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3312
3313 for_each_subsys(root, ss) {
3314 struct cgroup_subsys_state *css = ss->create(ss, cgrp);
3315
3316 if (IS_ERR(css)) {
3317 err = PTR_ERR(css);
3318 goto err_destroy;
3319 }
3320 init_cgroup_css(css, ss, cgrp);
3321 if (ss->use_id) {
3322 err = alloc_css_id(ss, parent, cgrp);
3323 if (err)
3324 goto err_destroy;
3325 }
3326 /* At error, ->destroy() callback has to free assigned ID. */
3327 }
3328
3329 cgroup_lock_hierarchy(root);
3330 list_add(&cgrp->sibling, &cgrp->parent->children);
3331 cgroup_unlock_hierarchy(root);
3332 root->number_of_cgroups++;
3333
3334 err = cgroup_create_dir(cgrp, dentry, mode);
3335 if (err < 0)
3336 goto err_remove;
3337
3338 /* The cgroup directory was pre-locked for us */
3339 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
3340
3341 err = cgroup_populate_dir(cgrp);
3342 /* If err < 0, we have a half-filled directory - oh well ;) */
3343
3344 mutex_unlock(&cgroup_mutex);
3345 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
3346
3347 return 0;
3348
3349 err_remove:
3350
3351 cgroup_lock_hierarchy(root);
3352 list_del(&cgrp->sibling);
3353 cgroup_unlock_hierarchy(root);
3354 root->number_of_cgroups--;
3355
3356 err_destroy:
3357
3358 for_each_subsys(root, ss) {
3359 if (cgrp->subsys[ss->subsys_id])
3360 ss->destroy(ss, cgrp);
3361 }
3362
3363 mutex_unlock(&cgroup_mutex);
3364
3365 /* Release the reference count that we took on the superblock */
3366 deactivate_super(sb);
3367
3368 kfree(cgrp);
3369 return err;
3370}
3371
3372static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
3373{
3374 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
3375
3376 /* the vfs holds inode->i_mutex already */
3377 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
3378}
3379
3380static int cgroup_has_css_refs(struct cgroup *cgrp)
3381{
3382 /* Check the reference count on each subsystem. Since we
3383 * already established that there are no tasks in the
3384 * cgroup, if the css refcount is also 1, then there should
3385 * be no outstanding references, so the subsystem is safe to
3386 * destroy. We scan across all subsystems rather than using
3387 * the per-hierarchy linked list of mounted subsystems since
3388 * we can be called via check_for_release() with no
3389 * synchronization other than RCU, and the subsystem linked
3390 * list isn't RCU-safe */
3391 int i;
3392 /*
3393 * We won't need to lock the subsys array, because the subsystems
3394 * we're concerned about aren't going anywhere since our cgroup root
3395 * has a reference on them.
3396 */
3397 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3398 struct cgroup_subsys *ss = subsys[i];
3399 struct cgroup_subsys_state *css;
3400 /* Skip subsystems not present or not in this hierarchy */
3401 if (ss == NULL || ss->root != cgrp->root)
3402 continue;
3403 css = cgrp->subsys[ss->subsys_id];
3404 /* When called from check_for_release() it's possible
3405 * that by this point the cgroup has been removed
3406 * and the css deleted. But a false-positive doesn't
3407 * matter, since it can only happen if the cgroup
3408 * has been deleted and hence no longer needs the
3409 * release agent to be called anyway. */
3410 if (css && (atomic_read(&css->refcnt) > 1))
3411 return 1;
3412 }
3413 return 0;
3414}
3415
3416/*
3417 * Atomically mark all (or else none) of the cgroup's CSS objects as
3418 * CSS_REMOVED. Return true on success, or false if the cgroup has
3419 * busy subsystems. Call with cgroup_mutex held
3420 */
3421
3422static int cgroup_clear_css_refs(struct cgroup *cgrp)
3423{
3424 struct cgroup_subsys *ss;
3425 unsigned long flags;
3426 bool failed = false;
3427 local_irq_save(flags);
3428 for_each_subsys(cgrp->root, ss) {
3429 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3430 int refcnt;
3431 while (1) {
3432 /* We can only remove a CSS with a refcnt==1 */
3433 refcnt = atomic_read(&css->refcnt);
3434 if (refcnt > 1) {
3435 failed = true;
3436 goto done;
3437 }
3438 BUG_ON(!refcnt);
3439 /*
3440 * Drop the refcnt to 0 while we check other
3441 * subsystems. This will cause any racing
3442 * css_tryget() to spin until we set the
3443 * CSS_REMOVED bits or abort
3444 */
3445 if (atomic_cmpxchg(&css->refcnt, refcnt, 0) == refcnt)
3446 break;
3447 cpu_relax();
3448 }
3449 }
3450 done:
3451 for_each_subsys(cgrp->root, ss) {
3452 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3453 if (failed) {
3454 /*
3455 * Restore old refcnt if we previously managed
3456 * to clear it from 1 to 0
3457 */
3458 if (!atomic_read(&css->refcnt))
3459 atomic_set(&css->refcnt, 1);
3460 } else {
3461 /* Commit the fact that the CSS is removed */
3462 set_bit(CSS_REMOVED, &css->flags);
3463 }
3464 }
3465 local_irq_restore(flags);
3466 return !failed;
3467}
3468
3469static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
3470{
3471 struct cgroup *cgrp = dentry->d_fsdata;
3472 struct dentry *d;
3473 struct cgroup *parent;
3474 DEFINE_WAIT(wait);
3475 struct cgroup_event *event, *tmp;
3476 int ret;
3477
3478 /* the vfs holds both inode->i_mutex already */
3479again:
3480 mutex_lock(&cgroup_mutex);
3481 if (atomic_read(&cgrp->count) != 0) {
3482 mutex_unlock(&cgroup_mutex);
3483 return -EBUSY;
3484 }
3485 if (!list_empty(&cgrp->children)) {
3486 mutex_unlock(&cgroup_mutex);
3487 return -EBUSY;
3488 }
3489 mutex_unlock(&cgroup_mutex);
3490
3491 /*
3492 * In general, subsystem has no css->refcnt after pre_destroy(). But
3493 * in racy cases, subsystem may have to get css->refcnt after
3494 * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
3495 * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
3496 * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
3497 * and subsystem's reference count handling. Please see css_get/put
3498 * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
3499 */
3500 set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3501
3502 /*
3503 * Call pre_destroy handlers of subsys. Notify subsystems
3504 * that rmdir() request comes.
3505 */
3506 ret = cgroup_call_pre_destroy(cgrp);
3507 if (ret) {
3508 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3509 return ret;
3510 }
3511
3512 mutex_lock(&cgroup_mutex);
3513 parent = cgrp->parent;
3514 if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
3515 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3516 mutex_unlock(&cgroup_mutex);
3517 return -EBUSY;
3518 }
3519 prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
3520 if (!cgroup_clear_css_refs(cgrp)) {
3521 mutex_unlock(&cgroup_mutex);
3522 /*
3523 * Because someone may call cgroup_wakeup_rmdir_waiter() before
3524 * prepare_to_wait(), we need to check this flag.
3525 */
3526 if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
3527 schedule();
3528 finish_wait(&cgroup_rmdir_waitq, &wait);
3529 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3530 if (signal_pending(current))
3531 return -EINTR;
3532 goto again;
3533 }
3534 /* NO css_tryget() can success after here. */
3535 finish_wait(&cgroup_rmdir_waitq, &wait);
3536 clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3537
3538 spin_lock(&release_list_lock);
3539 set_bit(CGRP_REMOVED, &cgrp->flags);
3540 if (!list_empty(&cgrp->release_list))
3541 list_del(&cgrp->release_list);
3542 spin_unlock(&release_list_lock);
3543
3544 cgroup_lock_hierarchy(cgrp->root);
3545 /* delete this cgroup from parent->children */
3546 list_del(&cgrp->sibling);
3547 cgroup_unlock_hierarchy(cgrp->root);
3548
3549 spin_lock(&cgrp->dentry->d_lock);
3550 d = dget(cgrp->dentry);
3551 spin_unlock(&d->d_lock);
3552
3553 cgroup_d_remove_dir(d);
3554 dput(d);
3555
3556 set_bit(CGRP_RELEASABLE, &parent->flags);
3557 check_for_release(parent);
3558
3559 /*
3560 * Unregister events and notify userspace.
3561 * Notify userspace about cgroup removing only after rmdir of cgroup
3562 * directory to avoid race between userspace and kernelspace
3563 */
3564 spin_lock(&cgrp->event_list_lock);
3565 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
3566 list_del(&event->list);
3567 remove_wait_queue(event->wqh, &event->wait);
3568 eventfd_signal(event->eventfd, 1);
3569 schedule_work(&event->remove);
3570 }
3571 spin_unlock(&cgrp->event_list_lock);
3572
3573 mutex_unlock(&cgroup_mutex);
3574 return 0;
3575}
3576
3577static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
3578{
3579 struct cgroup_subsys_state *css;
3580
3581 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
3582
3583 /* Create the top cgroup state for this subsystem */
3584 list_add(&ss->sibling, &rootnode.subsys_list);
3585 ss->root = &rootnode;
3586 css = ss->create(ss, dummytop);
3587 /* We don't handle early failures gracefully */
3588 BUG_ON(IS_ERR(css));
3589 init_cgroup_css(css, ss, dummytop);
3590
3591 /* Update the init_css_set to contain a subsys
3592 * pointer to this state - since the subsystem is
3593 * newly registered, all tasks and hence the
3594 * init_css_set is in the subsystem's top cgroup. */
3595 init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
3596
3597 need_forkexit_callback |= ss->fork || ss->exit;
3598
3599 /* At system boot, before all subsystems have been
3600 * registered, no tasks have been forked, so we don't
3601 * need to invoke fork callbacks here. */
3602 BUG_ON(!list_empty(&init_task.tasks));
3603
3604 mutex_init(&ss->hierarchy_mutex);
3605 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3606 ss->active = 1;
3607
3608 /* this function shouldn't be used with modular subsystems, since they
3609 * need to register a subsys_id, among other things */
3610 BUG_ON(ss->module);
3611}
3612
3613/**
3614 * cgroup_load_subsys: load and register a modular subsystem at runtime
3615 * @ss: the subsystem to load
3616 *
3617 * This function should be called in a modular subsystem's initcall. If the
3618 * subsytem is built as a module, it will be assigned a new subsys_id and set
3619 * up for use. If the subsystem is built-in anyway, work is delegated to the
3620 * simpler cgroup_init_subsys.
3621 */
3622int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
3623{
3624 int i;
3625 struct cgroup_subsys_state *css;
3626
3627 /* check name and function validity */
3628 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
3629 ss->create == NULL || ss->destroy == NULL)
3630 return -EINVAL;
3631
3632 /*
3633 * we don't support callbacks in modular subsystems. this check is
3634 * before the ss->module check for consistency; a subsystem that could
3635 * be a module should still have no callbacks even if the user isn't
3636 * compiling it as one.
3637 */
3638 if (ss->fork || ss->exit)
3639 return -EINVAL;
3640
3641 /*
3642 * an optionally modular subsystem is built-in: we want to do nothing,
3643 * since cgroup_init_subsys will have already taken care of it.
3644 */
3645 if (ss->module == NULL) {
3646 /* a few sanity checks */
3647 BUG_ON(ss->subsys_id >= CGROUP_BUILTIN_SUBSYS_COUNT);
3648 BUG_ON(subsys[ss->subsys_id] != ss);
3649 return 0;
3650 }
3651
3652 /*
3653 * need to register a subsys id before anything else - for example,
3654 * init_cgroup_css needs it.
3655 */
3656 mutex_lock(&cgroup_mutex);
3657 /* find the first empty slot in the array */
3658 for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
3659 if (subsys[i] == NULL)
3660 break;
3661 }
3662 if (i == CGROUP_SUBSYS_COUNT) {
3663 /* maximum number of subsystems already registered! */
3664 mutex_unlock(&cgroup_mutex);
3665 return -EBUSY;
3666 }
3667 /* assign ourselves the subsys_id */
3668 ss->subsys_id = i;
3669 subsys[i] = ss;
3670
3671 /*
3672 * no ss->create seems to need anything important in the ss struct, so
3673 * this can happen first (i.e. before the rootnode attachment).
3674 */
3675 css = ss->create(ss, dummytop);
3676 if (IS_ERR(css)) {
3677 /* failure case - need to deassign the subsys[] slot. */
3678 subsys[i] = NULL;
3679 mutex_unlock(&cgroup_mutex);
3680 return PTR_ERR(css);
3681 }
3682
3683 list_add(&ss->sibling, &rootnode.subsys_list);
3684 ss->root = &rootnode;
3685
3686 /* our new subsystem will be attached to the dummy hierarchy. */
3687 init_cgroup_css(css, ss, dummytop);
3688 /* init_idr must be after init_cgroup_css because it sets css->id. */
3689 if (ss->use_id) {
3690 int ret = cgroup_init_idr(ss, css);
3691 if (ret) {
3692 dummytop->subsys[ss->subsys_id] = NULL;
3693 ss->destroy(ss, dummytop);
3694 subsys[i] = NULL;
3695 mutex_unlock(&cgroup_mutex);
3696 return ret;
3697 }
3698 }
3699
3700 /*
3701 * Now we need to entangle the css into the existing css_sets. unlike
3702 * in cgroup_init_subsys, there are now multiple css_sets, so each one
3703 * will need a new pointer to it; done by iterating the css_set_table.
3704 * furthermore, modifying the existing css_sets will corrupt the hash
3705 * table state, so each changed css_set will need its hash recomputed.
3706 * this is all done under the css_set_lock.
3707 */
3708 write_lock(&css_set_lock);
3709 for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
3710 struct css_set *cg;
3711 struct hlist_node *node, *tmp;
3712 struct hlist_head *bucket = &css_set_table[i], *new_bucket;
3713
3714 hlist_for_each_entry_safe(cg, node, tmp, bucket, hlist) {
3715 /* skip entries that we already rehashed */
3716 if (cg->subsys[ss->subsys_id])
3717 continue;
3718 /* remove existing entry */
3719 hlist_del(&cg->hlist);
3720 /* set new value */
3721 cg->subsys[ss->subsys_id] = css;
3722 /* recompute hash and restore entry */
3723 new_bucket = css_set_hash(cg->subsys);
3724 hlist_add_head(&cg->hlist, new_bucket);
3725 }
3726 }
3727 write_unlock(&css_set_lock);
3728
3729 mutex_init(&ss->hierarchy_mutex);
3730 lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
3731 ss->active = 1;
3732
3733 /* success! */
3734 mutex_unlock(&cgroup_mutex);
3735 return 0;
3736}
3737EXPORT_SYMBOL_GPL(cgroup_load_subsys);
3738
3739/**
3740 * cgroup_unload_subsys: unload a modular subsystem
3741 * @ss: the subsystem to unload
3742 *
3743 * This function should be called in a modular subsystem's exitcall. When this
3744 * function is invoked, the refcount on the subsystem's module will be 0, so
3745 * the subsystem will not be attached to any hierarchy.
3746 */
3747void cgroup_unload_subsys(struct cgroup_subsys *ss)
3748{
3749 struct cg_cgroup_link *link;
3750 struct hlist_head *hhead;
3751
3752 BUG_ON(ss->module == NULL);
3753
3754 /*
3755 * we shouldn't be called if the subsystem is in use, and the use of
3756 * try_module_get in parse_cgroupfs_options should ensure that it
3757 * doesn't start being used while we're killing it off.
3758 */
3759 BUG_ON(ss->root != &rootnode);
3760
3761 mutex_lock(&cgroup_mutex);
3762 /* deassign the subsys_id */
3763 BUG_ON(ss->subsys_id < CGROUP_BUILTIN_SUBSYS_COUNT);
3764 subsys[ss->subsys_id] = NULL;
3765
3766 /* remove subsystem from rootnode's list of subsystems */
3767 list_del(&ss->sibling);
3768
3769 /*
3770 * disentangle the css from all css_sets attached to the dummytop. as
3771 * in loading, we need to pay our respects to the hashtable gods.
3772 */
3773 write_lock(&css_set_lock);
3774 list_for_each_entry(link, &dummytop->css_sets, cgrp_link_list) {
3775 struct css_set *cg = link->cg;
3776
3777 hlist_del(&cg->hlist);
3778 BUG_ON(!cg->subsys[ss->subsys_id]);
3779 cg->subsys[ss->subsys_id] = NULL;
3780 hhead = css_set_hash(cg->subsys);
3781 hlist_add_head(&cg->hlist, hhead);
3782 }
3783 write_unlock(&css_set_lock);
3784
3785 /*
3786 * remove subsystem's css from the dummytop and free it - need to free
3787 * before marking as null because ss->destroy needs the cgrp->subsys
3788 * pointer to find their state. note that this also takes care of
3789 * freeing the css_id.
3790 */
3791 ss->destroy(ss, dummytop);
3792 dummytop->subsys[ss->subsys_id] = NULL;
3793
3794 mutex_unlock(&cgroup_mutex);
3795}
3796EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
3797
3798/**
3799 * cgroup_init_early - cgroup initialization at system boot
3800 *
3801 * Initialize cgroups at system boot, and initialize any
3802 * subsystems that request early init.
3803 */
3804int __init cgroup_init_early(void)
3805{
3806 int i;
3807 atomic_set(&init_css_set.refcount, 1);
3808 INIT_LIST_HEAD(&init_css_set.cg_links);
3809 INIT_LIST_HEAD(&init_css_set.tasks);
3810 INIT_HLIST_NODE(&init_css_set.hlist);
3811 css_set_count = 1;
3812 init_cgroup_root(&rootnode);
3813 root_count = 1;
3814 init_task.cgroups = &init_css_set;
3815
3816 init_css_set_link.cg = &init_css_set;
3817 init_css_set_link.cgrp = dummytop;
3818 list_add(&init_css_set_link.cgrp_link_list,
3819 &rootnode.top_cgroup.css_sets);
3820 list_add(&init_css_set_link.cg_link_list,
3821 &init_css_set.cg_links);
3822
3823 for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
3824 INIT_HLIST_HEAD(&css_set_table[i]);
3825
3826 /* at bootup time, we don't worry about modular subsystems */
3827 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3828 struct cgroup_subsys *ss = subsys[i];
3829
3830 BUG_ON(!ss->name);
3831 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
3832 BUG_ON(!ss->create);
3833 BUG_ON(!ss->destroy);
3834 if (ss->subsys_id != i) {
3835 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
3836 ss->name, ss->subsys_id);
3837 BUG();
3838 }
3839
3840 if (ss->early_init)
3841 cgroup_init_subsys(ss);
3842 }
3843 return 0;
3844}
3845
3846/**
3847 * cgroup_init - cgroup initialization
3848 *
3849 * Register cgroup filesystem and /proc file, and initialize
3850 * any subsystems that didn't request early init.
3851 */
3852int __init cgroup_init(void)
3853{
3854 int err;
3855 int i;
3856 struct hlist_head *hhead;
3857
3858 err = bdi_init(&cgroup_backing_dev_info);
3859 if (err)
3860 return err;
3861
3862 /* at bootup time, we don't worry about modular subsystems */
3863 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
3864 struct cgroup_subsys *ss = subsys[i];
3865 if (!ss->early_init)
3866 cgroup_init_subsys(ss);
3867 if (ss->use_id)
3868 cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
3869 }
3870
3871 /* Add init_css_set to the hash table */
3872 hhead = css_set_hash(init_css_set.subsys);
3873 hlist_add_head(&init_css_set.hlist, hhead);
3874 BUG_ON(!init_root_id(&rootnode));
3875 err = register_filesystem(&cgroup_fs_type);
3876 if (err < 0)
3877 goto out;
3878
3879 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
3880
3881out:
3882 if (err)
3883 bdi_destroy(&cgroup_backing_dev_info);
3884
3885 return err;
3886}
3887
3888/*
3889 * proc_cgroup_show()
3890 * - Print task's cgroup paths into seq_file, one line for each hierarchy
3891 * - Used for /proc/<pid>/cgroup.
3892 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
3893 * doesn't really matter if tsk->cgroup changes after we read it,
3894 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
3895 * anyway. No need to check that tsk->cgroup != NULL, thanks to
3896 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
3897 * cgroup to top_cgroup.
3898 */
3899
3900/* TODO: Use a proper seq_file iterator */
3901static int proc_cgroup_show(struct seq_file *m, void *v)
3902{
3903 struct pid *pid;
3904 struct task_struct *tsk;
3905 char *buf;
3906 int retval;
3907 struct cgroupfs_root *root;
3908
3909 retval = -ENOMEM;
3910 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
3911 if (!buf)
3912 goto out;
3913
3914 retval = -ESRCH;
3915 pid = m->private;
3916 tsk = get_pid_task(pid, PIDTYPE_PID);
3917 if (!tsk)
3918 goto out_free;
3919
3920 retval = 0;
3921
3922 mutex_lock(&cgroup_mutex);
3923
3924 for_each_active_root(root) {
3925 struct cgroup_subsys *ss;
3926 struct cgroup *cgrp;
3927 int count = 0;
3928
3929 seq_printf(m, "%d:", root->hierarchy_id);
3930 for_each_subsys(root, ss)
3931 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
3932 if (strlen(root->name))
3933 seq_printf(m, "%sname=%s", count ? "," : "",
3934 root->name);
3935 seq_putc(m, ':');
3936 cgrp = task_cgroup_from_root(tsk, root);
3937 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
3938 if (retval < 0)
3939 goto out_unlock;
3940 seq_puts(m, buf);
3941 seq_putc(m, '\n');
3942 }
3943
3944out_unlock:
3945 mutex_unlock(&cgroup_mutex);
3946 put_task_struct(tsk);
3947out_free:
3948 kfree(buf);
3949out:
3950 return retval;
3951}
3952
3953static int cgroup_open(struct inode *inode, struct file *file)
3954{
3955 struct pid *pid = PROC_I(inode)->pid;
3956 return single_open(file, proc_cgroup_show, pid);
3957}
3958
3959const struct file_operations proc_cgroup_operations = {
3960 .open = cgroup_open,
3961 .read = seq_read,
3962 .llseek = seq_lseek,
3963 .release = single_release,
3964};
3965
3966/* Display information about each subsystem and each hierarchy */
3967static int proc_cgroupstats_show(struct seq_file *m, void *v)
3968{
3969 int i;
3970
3971 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
3972 /*
3973 * ideally we don't want subsystems moving around while we do this.
3974 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
3975 * subsys/hierarchy state.
3976 */
3977 mutex_lock(&cgroup_mutex);
3978 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3979 struct cgroup_subsys *ss = subsys[i];
3980 if (ss == NULL)
3981 continue;
3982 seq_printf(m, "%s\t%d\t%d\t%d\n",
3983 ss->name, ss->root->hierarchy_id,
3984 ss->root->number_of_cgroups, !ss->disabled);
3985 }
3986 mutex_unlock(&cgroup_mutex);
3987 return 0;
3988}
3989
3990static int cgroupstats_open(struct inode *inode, struct file *file)
3991{
3992 return single_open(file, proc_cgroupstats_show, NULL);
3993}
3994
3995static const struct file_operations proc_cgroupstats_operations = {
3996 .open = cgroupstats_open,
3997 .read = seq_read,
3998 .llseek = seq_lseek,
3999 .release = single_release,
4000};
4001
4002/**
4003 * cgroup_fork - attach newly forked task to its parents cgroup.
4004 * @child: pointer to task_struct of forking parent process.
4005 *
4006 * Description: A task inherits its parent's cgroup at fork().
4007 *
4008 * A pointer to the shared css_set was automatically copied in
4009 * fork.c by dup_task_struct(). However, we ignore that copy, since
4010 * it was not made under the protection of RCU or cgroup_mutex, so
4011 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4012 * have already changed current->cgroups, allowing the previously
4013 * referenced cgroup group to be removed and freed.
4014 *
4015 * At the point that cgroup_fork() is called, 'current' is the parent
4016 * task, and the passed argument 'child' points to the child task.
4017 */
4018void cgroup_fork(struct task_struct *child)
4019{
4020 task_lock(current);
4021 child->cgroups = current->cgroups;
4022 get_css_set(child->cgroups);
4023 task_unlock(current);
4024 INIT_LIST_HEAD(&child->cg_list);
4025}
4026
4027/**
4028 * cgroup_fork_callbacks - run fork callbacks
4029 * @child: the new task
4030 *
4031 * Called on a new task very soon before adding it to the
4032 * tasklist. No need to take any locks since no-one can
4033 * be operating on this task.
4034 */
4035void cgroup_fork_callbacks(struct task_struct *child)
4036{
4037 if (need_forkexit_callback) {
4038 int i;
4039 /*
4040 * forkexit callbacks are only supported for builtin
4041 * subsystems, and the builtin section of the subsys array is
4042 * immutable, so we don't need to lock the subsys array here.
4043 */
4044 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4045 struct cgroup_subsys *ss = subsys[i];
4046 if (ss->fork)
4047 ss->fork(ss, child);
4048 }
4049 }
4050}
4051
4052/**
4053 * cgroup_post_fork - called on a new task after adding it to the task list
4054 * @child: the task in question
4055 *
4056 * Adds the task to the list running through its css_set if necessary.
4057 * Has to be after the task is visible on the task list in case we race
4058 * with the first call to cgroup_iter_start() - to guarantee that the
4059 * new task ends up on its list.
4060 */
4061void cgroup_post_fork(struct task_struct *child)
4062{
4063 if (use_task_css_set_links) {
4064 write_lock(&css_set_lock);
4065 task_lock(child);
4066 if (list_empty(&child->cg_list))
4067 list_add(&child->cg_list, &child->cgroups->tasks);
4068 task_unlock(child);
4069 write_unlock(&css_set_lock);
4070 }
4071}
4072/**
4073 * cgroup_exit - detach cgroup from exiting task
4074 * @tsk: pointer to task_struct of exiting process
4075 * @run_callback: run exit callbacks?
4076 *
4077 * Description: Detach cgroup from @tsk and release it.
4078 *
4079 * Note that cgroups marked notify_on_release force every task in
4080 * them to take the global cgroup_mutex mutex when exiting.
4081 * This could impact scaling on very large systems. Be reluctant to
4082 * use notify_on_release cgroups where very high task exit scaling
4083 * is required on large systems.
4084 *
4085 * the_top_cgroup_hack:
4086 *
4087 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4088 *
4089 * We call cgroup_exit() while the task is still competent to
4090 * handle notify_on_release(), then leave the task attached to the
4091 * root cgroup in each hierarchy for the remainder of its exit.
4092 *
4093 * To do this properly, we would increment the reference count on
4094 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4095 * code we would add a second cgroup function call, to drop that
4096 * reference. This would just create an unnecessary hot spot on
4097 * the top_cgroup reference count, to no avail.
4098 *
4099 * Normally, holding a reference to a cgroup without bumping its
4100 * count is unsafe. The cgroup could go away, or someone could
4101 * attach us to a different cgroup, decrementing the count on
4102 * the first cgroup that we never incremented. But in this case,
4103 * top_cgroup isn't going away, and either task has PF_EXITING set,
4104 * which wards off any cgroup_attach_task() attempts, or task is a failed
4105 * fork, never visible to cgroup_attach_task.
4106 */
4107void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4108{
4109 int i;
4110 struct css_set *cg;
4111
4112 if (run_callbacks && need_forkexit_callback) {
4113 /*
4114 * modular subsystems can't use callbacks, so no need to lock
4115 * the subsys array
4116 */
4117 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4118 struct cgroup_subsys *ss = subsys[i];
4119 if (ss->exit)
4120 ss->exit(ss, tsk);
4121 }
4122 }
4123
4124 /*
4125 * Unlink from the css_set task list if necessary.
4126 * Optimistically check cg_list before taking
4127 * css_set_lock
4128 */
4129 if (!list_empty(&tsk->cg_list)) {
4130 write_lock(&css_set_lock);
4131 if (!list_empty(&tsk->cg_list))
4132 list_del(&tsk->cg_list);
4133 write_unlock(&css_set_lock);
4134 }
4135
4136 /* Reassign the task to the init_css_set. */
4137 task_lock(tsk);
4138 cg = tsk->cgroups;
4139 tsk->cgroups = &init_css_set;
4140 task_unlock(tsk);
4141 if (cg)
4142 put_css_set_taskexit(cg);
4143}
4144
4145/**
4146 * cgroup_clone - clone the cgroup the given subsystem is attached to
4147 * @tsk: the task to be moved
4148 * @subsys: the given subsystem
4149 * @nodename: the name for the new cgroup
4150 *
4151 * Duplicate the current cgroup in the hierarchy that the given
4152 * subsystem is attached to, and move this task into the new
4153 * child.
4154 */
4155int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
4156 char *nodename)
4157{
4158 struct dentry *dentry;
4159 int ret = 0;
4160 struct cgroup *parent, *child;
4161 struct inode *inode;
4162 struct css_set *cg;
4163 struct cgroupfs_root *root;
4164 struct cgroup_subsys *ss;
4165
4166 /* We shouldn't be called by an unregistered subsystem */
4167 BUG_ON(!subsys->active);
4168
4169 /* First figure out what hierarchy and cgroup we're dealing
4170 * with, and pin them so we can drop cgroup_mutex */
4171 mutex_lock(&cgroup_mutex);
4172 again:
4173 root = subsys->root;
4174 if (root == &rootnode) {
4175 mutex_unlock(&cgroup_mutex);
4176 return 0;
4177 }
4178
4179 /* Pin the hierarchy */
4180 if (!atomic_inc_not_zero(&root->sb->s_active)) {
4181 /* We race with the final deactivate_super() */
4182 mutex_unlock(&cgroup_mutex);
4183 return 0;
4184 }
4185
4186 /* Keep the cgroup alive */
4187 task_lock(tsk);
4188 parent = task_cgroup(tsk, subsys->subsys_id);
4189 cg = tsk->cgroups;
4190 get_css_set(cg);
4191 task_unlock(tsk);
4192
4193 mutex_unlock(&cgroup_mutex);
4194
4195 /* Now do the VFS work to create a cgroup */
4196 inode = parent->dentry->d_inode;
4197
4198 /* Hold the parent directory mutex across this operation to
4199 * stop anyone else deleting the new cgroup */
4200 mutex_lock(&inode->i_mutex);
4201 dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
4202 if (IS_ERR(dentry)) {
4203 printk(KERN_INFO
4204 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
4205 PTR_ERR(dentry));
4206 ret = PTR_ERR(dentry);
4207 goto out_release;
4208 }
4209
4210 /* Create the cgroup directory, which also creates the cgroup */
4211 ret = vfs_mkdir(inode, dentry, 0755);
4212 child = __d_cgrp(dentry);
4213 dput(dentry);
4214 if (ret) {
4215 printk(KERN_INFO
4216 "Failed to create cgroup %s: %d\n", nodename,
4217 ret);
4218 goto out_release;
4219 }
4220
4221 /* The cgroup now exists. Retake cgroup_mutex and check
4222 * that we're still in the same state that we thought we
4223 * were. */
4224 mutex_lock(&cgroup_mutex);
4225 if ((root != subsys->root) ||
4226 (parent != task_cgroup(tsk, subsys->subsys_id))) {
4227 /* Aargh, we raced ... */
4228 mutex_unlock(&inode->i_mutex);
4229 put_css_set(cg);
4230
4231 deactivate_super(root->sb);
4232 /* The cgroup is still accessible in the VFS, but
4233 * we're not going to try to rmdir() it at this
4234 * point. */
4235 printk(KERN_INFO
4236 "Race in cgroup_clone() - leaking cgroup %s\n",
4237 nodename);
4238 goto again;
4239 }
4240
4241 /* do any required auto-setup */
4242 for_each_subsys(root, ss) {
4243 if (ss->post_clone)
4244 ss->post_clone(ss, child);
4245 }
4246
4247 /* All seems fine. Finish by moving the task into the new cgroup */
4248 ret = cgroup_attach_task(child, tsk);
4249 mutex_unlock(&cgroup_mutex);
4250
4251 out_release:
4252 mutex_unlock(&inode->i_mutex);
4253
4254 mutex_lock(&cgroup_mutex);
4255 put_css_set(cg);
4256 mutex_unlock(&cgroup_mutex);
4257 deactivate_super(root->sb);
4258 return ret;
4259}
4260
4261/**
4262 * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
4263 * @cgrp: the cgroup in question
4264 * @task: the task in question
4265 *
4266 * See if @cgrp is a descendant of @task's cgroup in the appropriate
4267 * hierarchy.
4268 *
4269 * If we are sending in dummytop, then presumably we are creating
4270 * the top cgroup in the subsystem.
4271 *
4272 * Called only by the ns (nsproxy) cgroup.
4273 */
4274int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
4275{
4276 int ret;
4277 struct cgroup *target;
4278
4279 if (cgrp == dummytop)
4280 return 1;
4281
4282 target = task_cgroup_from_root(task, cgrp->root);
4283 while (cgrp != target && cgrp!= cgrp->top_cgroup)
4284 cgrp = cgrp->parent;
4285 ret = (cgrp == target);
4286 return ret;
4287}
4288
4289static void check_for_release(struct cgroup *cgrp)
4290{
4291 /* All of these checks rely on RCU to keep the cgroup
4292 * structure alive */
4293 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
4294 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
4295 /* Control Group is currently removeable. If it's not
4296 * already queued for a userspace notification, queue
4297 * it now */
4298 int need_schedule_work = 0;
4299 spin_lock(&release_list_lock);
4300 if (!cgroup_is_removed(cgrp) &&
4301 list_empty(&cgrp->release_list)) {
4302 list_add(&cgrp->release_list, &release_list);
4303 need_schedule_work = 1;
4304 }
4305 spin_unlock(&release_list_lock);
4306 if (need_schedule_work)
4307 schedule_work(&release_agent_work);
4308 }
4309}
4310
4311/* Caller must verify that the css is not for root cgroup */
4312void __css_put(struct cgroup_subsys_state *css, int count)
4313{
4314 struct cgroup *cgrp = css->cgroup;
4315 int val;
4316 rcu_read_lock();
4317 val = atomic_sub_return(count, &css->refcnt);
4318 if (val == 1) {
4319 if (notify_on_release(cgrp)) {
4320 set_bit(CGRP_RELEASABLE, &cgrp->flags);
4321 check_for_release(cgrp);
4322 }
4323 cgroup_wakeup_rmdir_waiter(cgrp);
4324 }
4325 rcu_read_unlock();
4326 WARN_ON_ONCE(val < 1);
4327}
4328EXPORT_SYMBOL_GPL(__css_put);
4329
4330/*
4331 * Notify userspace when a cgroup is released, by running the
4332 * configured release agent with the name of the cgroup (path
4333 * relative to the root of cgroup file system) as the argument.
4334 *
4335 * Most likely, this user command will try to rmdir this cgroup.
4336 *
4337 * This races with the possibility that some other task will be
4338 * attached to this cgroup before it is removed, or that some other
4339 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4340 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4341 * unused, and this cgroup will be reprieved from its death sentence,
4342 * to continue to serve a useful existence. Next time it's released,
4343 * we will get notified again, if it still has 'notify_on_release' set.
4344 *
4345 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4346 * means only wait until the task is successfully execve()'d. The
4347 * separate release agent task is forked by call_usermodehelper(),
4348 * then control in this thread returns here, without waiting for the
4349 * release agent task. We don't bother to wait because the caller of
4350 * this routine has no use for the exit status of the release agent
4351 * task, so no sense holding our caller up for that.
4352 */
4353static void cgroup_release_agent(struct work_struct *work)
4354{
4355 BUG_ON(work != &release_agent_work);
4356 mutex_lock(&cgroup_mutex);
4357 spin_lock(&release_list_lock);
4358 while (!list_empty(&release_list)) {
4359 char *argv[3], *envp[3];
4360 int i;
4361 char *pathbuf = NULL, *agentbuf = NULL;
4362 struct cgroup *cgrp = list_entry(release_list.next,
4363 struct cgroup,
4364 release_list);
4365 list_del_init(&cgrp->release_list);
4366 spin_unlock(&release_list_lock);
4367 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4368 if (!pathbuf)
4369 goto continue_free;
4370 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4371 goto continue_free;
4372 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4373 if (!agentbuf)
4374 goto continue_free;
4375
4376 i = 0;
4377 argv[i++] = agentbuf;
4378 argv[i++] = pathbuf;
4379 argv[i] = NULL;
4380
4381 i = 0;
4382 /* minimal command environment */
4383 envp[i++] = "HOME=/";
4384 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4385 envp[i] = NULL;
4386
4387 /* Drop the lock while we invoke the usermode helper,
4388 * since the exec could involve hitting disk and hence
4389 * be a slow process */
4390 mutex_unlock(&cgroup_mutex);
4391 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4392 mutex_lock(&cgroup_mutex);
4393 continue_free:
4394 kfree(pathbuf);
4395 kfree(agentbuf);
4396 spin_lock(&release_list_lock);
4397 }
4398 spin_unlock(&release_list_lock);
4399 mutex_unlock(&cgroup_mutex);
4400}
4401
4402static int __init cgroup_disable(char *str)
4403{
4404 int i;
4405 char *token;
4406
4407 while ((token = strsep(&str, ",")) != NULL) {
4408 if (!*token)
4409 continue;
4410 /*
4411 * cgroup_disable, being at boot time, can't know about module
4412 * subsystems, so we don't worry about them.
4413 */
4414 for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4415 struct cgroup_subsys *ss = subsys[i];
4416
4417 if (!strcmp(token, ss->name)) {
4418 ss->disabled = 1;
4419 printk(KERN_INFO "Disabling %s control group"
4420 " subsystem\n", ss->name);
4421 break;
4422 }
4423 }
4424 }
4425 return 1;
4426}
4427__setup("cgroup_disable=", cgroup_disable);
4428
4429/*
4430 * Functons for CSS ID.
4431 */
4432
4433/*
4434 *To get ID other than 0, this should be called when !cgroup_is_removed().
4435 */
4436unsigned short css_id(struct cgroup_subsys_state *css)
4437{
4438 struct css_id *cssid;
4439
4440 /*
4441 * This css_id() can return correct value when somone has refcnt
4442 * on this or this is under rcu_read_lock(). Once css->id is allocated,
4443 * it's unchanged until freed.
4444 */
4445 cssid = rcu_dereference_check(css->id,
4446 rcu_read_lock_held() || atomic_read(&css->refcnt));
4447
4448 if (cssid)
4449 return cssid->id;
4450 return 0;
4451}
4452EXPORT_SYMBOL_GPL(css_id);
4453
4454unsigned short css_depth(struct cgroup_subsys_state *css)
4455{
4456 struct css_id *cssid;
4457
4458 cssid = rcu_dereference_check(css->id,
4459 rcu_read_lock_held() || atomic_read(&css->refcnt));
4460
4461 if (cssid)
4462 return cssid->depth;
4463 return 0;
4464}
4465EXPORT_SYMBOL_GPL(css_depth);
4466
4467/**
4468 * css_is_ancestor - test "root" css is an ancestor of "child"
4469 * @child: the css to be tested.
4470 * @root: the css supporsed to be an ancestor of the child.
4471 *
4472 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
4473 * this function reads css->id, this use rcu_dereference() and rcu_read_lock().
4474 * But, considering usual usage, the csses should be valid objects after test.
4475 * Assuming that the caller will do some action to the child if this returns
4476 * returns true, the caller must take "child";s reference count.
4477 * If "child" is valid object and this returns true, "root" is valid, too.
4478 */
4479
4480bool css_is_ancestor(struct cgroup_subsys_state *child,
4481 const struct cgroup_subsys_state *root)
4482{
4483 struct css_id *child_id;
4484 struct css_id *root_id;
4485 bool ret = true;
4486
4487 rcu_read_lock();
4488 child_id = rcu_dereference(child->id);
4489 root_id = rcu_dereference(root->id);
4490 if (!child_id
4491 || !root_id
4492 || (child_id->depth < root_id->depth)
4493 || (child_id->stack[root_id->depth] != root_id->id))
4494 ret = false;
4495 rcu_read_unlock();
4496 return ret;
4497}
4498
4499static void __free_css_id_cb(struct rcu_head *head)
4500{
4501 struct css_id *id;
4502
4503 id = container_of(head, struct css_id, rcu_head);
4504 kfree(id);
4505}
4506
4507void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
4508{
4509 struct css_id *id = css->id;
4510 /* When this is called before css_id initialization, id can be NULL */
4511 if (!id)
4512 return;
4513
4514 BUG_ON(!ss->use_id);
4515
4516 rcu_assign_pointer(id->css, NULL);
4517 rcu_assign_pointer(css->id, NULL);
4518 spin_lock(&ss->id_lock);
4519 idr_remove(&ss->idr, id->id);
4520 spin_unlock(&ss->id_lock);
4521 call_rcu(&id->rcu_head, __free_css_id_cb);
4522}
4523EXPORT_SYMBOL_GPL(free_css_id);
4524
4525/*
4526 * This is called by init or create(). Then, calls to this function are
4527 * always serialized (By cgroup_mutex() at create()).
4528 */
4529
4530static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
4531{
4532 struct css_id *newid;
4533 int myid, error, size;
4534
4535 BUG_ON(!ss->use_id);
4536
4537 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
4538 newid = kzalloc(size, GFP_KERNEL);
4539 if (!newid)
4540 return ERR_PTR(-ENOMEM);
4541 /* get id */
4542 if (unlikely(!idr_pre_get(&ss->idr, GFP_KERNEL))) {
4543 error = -ENOMEM;
4544 goto err_out;
4545 }
4546 spin_lock(&ss->id_lock);
4547 /* Don't use 0. allocates an ID of 1-65535 */
4548 error = idr_get_new_above(&ss->idr, newid, 1, &myid);
4549 spin_unlock(&ss->id_lock);
4550
4551 /* Returns error when there are no free spaces for new ID.*/
4552 if (error) {
4553 error = -ENOSPC;
4554 goto err_out;
4555 }
4556 if (myid > CSS_ID_MAX)
4557 goto remove_idr;
4558
4559 newid->id = myid;
4560 newid->depth = depth;
4561 return newid;
4562remove_idr:
4563 error = -ENOSPC;
4564 spin_lock(&ss->id_lock);
4565 idr_remove(&ss->idr, myid);
4566 spin_unlock(&ss->id_lock);
4567err_out:
4568 kfree(newid);
4569 return ERR_PTR(error);
4570
4571}
4572
4573static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
4574 struct cgroup_subsys_state *rootcss)
4575{
4576 struct css_id *newid;
4577
4578 spin_lock_init(&ss->id_lock);
4579 idr_init(&ss->idr);
4580
4581 newid = get_new_cssid(ss, 0);
4582 if (IS_ERR(newid))
4583 return PTR_ERR(newid);
4584
4585 newid->stack[0] = newid->id;
4586 newid->css = rootcss;
4587 rootcss->id = newid;
4588 return 0;
4589}
4590
4591static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
4592 struct cgroup *child)
4593{
4594 int subsys_id, i, depth = 0;
4595 struct cgroup_subsys_state *parent_css, *child_css;
4596 struct css_id *child_id, *parent_id;
4597
4598 subsys_id = ss->subsys_id;
4599 parent_css = parent->subsys[subsys_id];
4600 child_css = child->subsys[subsys_id];
4601 parent_id = parent_css->id;
4602 depth = parent_id->depth;
4603
4604 child_id = get_new_cssid(ss, depth);
4605 if (IS_ERR(child_id))
4606 return PTR_ERR(child_id);
4607
4608 for (i = 0; i < depth; i++)
4609 child_id->stack[i] = parent_id->stack[i];
4610 child_id->stack[depth] = child_id->id;
4611 /*
4612 * child_id->css pointer will be set after this cgroup is available
4613 * see cgroup_populate_dir()
4614 */
4615 rcu_assign_pointer(child_css->id, child_id);
4616
4617 return 0;
4618}
4619
4620/**
4621 * css_lookup - lookup css by id
4622 * @ss: cgroup subsys to be looked into.
4623 * @id: the id
4624 *
4625 * Returns pointer to cgroup_subsys_state if there is valid one with id.
4626 * NULL if not. Should be called under rcu_read_lock()
4627 */
4628struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
4629{
4630 struct css_id *cssid = NULL;
4631
4632 BUG_ON(!ss->use_id);
4633 cssid = idr_find(&ss->idr, id);
4634
4635 if (unlikely(!cssid))
4636 return NULL;
4637
4638 return rcu_dereference(cssid->css);
4639}
4640EXPORT_SYMBOL_GPL(css_lookup);
4641
4642/**
4643 * css_get_next - lookup next cgroup under specified hierarchy.
4644 * @ss: pointer to subsystem
4645 * @id: current position of iteration.
4646 * @root: pointer to css. search tree under this.
4647 * @foundid: position of found object.
4648 *
4649 * Search next css under the specified hierarchy of rootid. Calling under
4650 * rcu_read_lock() is necessary. Returns NULL if it reaches the end.
4651 */
4652struct cgroup_subsys_state *
4653css_get_next(struct cgroup_subsys *ss, int id,
4654 struct cgroup_subsys_state *root, int *foundid)
4655{
4656 struct cgroup_subsys_state *ret = NULL;
4657 struct css_id *tmp;
4658 int tmpid;
4659 int rootid = css_id(root);
4660 int depth = css_depth(root);
4661
4662 if (!rootid)
4663 return NULL;
4664
4665 BUG_ON(!ss->use_id);
4666 /* fill start point for scan */
4667 tmpid = id;
4668 while (1) {
4669 /*
4670 * scan next entry from bitmap(tree), tmpid is updated after
4671 * idr_get_next().
4672 */
4673 spin_lock(&ss->id_lock);
4674 tmp = idr_get_next(&ss->idr, &tmpid);
4675 spin_unlock(&ss->id_lock);
4676
4677 if (!tmp)
4678 break;
4679 if (tmp->depth >= depth && tmp->stack[depth] == rootid) {
4680 ret = rcu_dereference(tmp->css);
4681 if (ret) {
4682 *foundid = tmpid;
4683 break;
4684 }
4685 }
4686 /* continue to scan from next id */
4687 tmpid = tmpid + 1;
4688 }
4689 return ret;
4690}
4691
4692#ifdef CONFIG_CGROUP_DEBUG
4693static struct cgroup_subsys_state *debug_create(struct cgroup_subsys *ss,
4694 struct cgroup *cont)
4695{
4696 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4697
4698 if (!css)
4699 return ERR_PTR(-ENOMEM);
4700
4701 return css;
4702}
4703
4704static void debug_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
4705{
4706 kfree(cont->subsys[debug_subsys_id]);
4707}
4708
4709static u64 cgroup_refcount_read(struct cgroup *cont, struct cftype *cft)
4710{
4711 return atomic_read(&cont->count);
4712}
4713
4714static u64 debug_taskcount_read(struct cgroup *cont, struct cftype *cft)
4715{
4716 return cgroup_task_count(cont);
4717}
4718
4719static u64 current_css_set_read(struct cgroup *cont, struct cftype *cft)
4720{
4721 return (u64)(unsigned long)current->cgroups;
4722}
4723
4724static u64 current_css_set_refcount_read(struct cgroup *cont,
4725 struct cftype *cft)
4726{
4727 u64 count;
4728
4729 rcu_read_lock();
4730 count = atomic_read(&current->cgroups->refcount);
4731 rcu_read_unlock();
4732 return count;
4733}
4734
4735static int current_css_set_cg_links_read(struct cgroup *cont,
4736 struct cftype *cft,
4737 struct seq_file *seq)
4738{
4739 struct cg_cgroup_link *link;
4740 struct css_set *cg;
4741
4742 read_lock(&css_set_lock);
4743 rcu_read_lock();
4744 cg = rcu_dereference(current->cgroups);
4745 list_for_each_entry(link, &cg->cg_links, cg_link_list) {
4746 struct cgroup *c = link->cgrp;
4747 const char *name;
4748
4749 if (c->dentry)
4750 name = c->dentry->d_name.name;
4751 else
4752 name = "?";
4753 seq_printf(seq, "Root %d group %s\n",
4754 c->root->hierarchy_id, name);
4755 }
4756 rcu_read_unlock();
4757 read_unlock(&css_set_lock);
4758 return 0;
4759}
4760
4761#define MAX_TASKS_SHOWN_PER_CSS 25
4762static int cgroup_css_links_read(struct cgroup *cont,
4763 struct cftype *cft,
4764 struct seq_file *seq)
4765{
4766 struct cg_cgroup_link *link;
4767
4768 read_lock(&css_set_lock);
4769 list_for_each_entry(link, &cont->css_sets, cgrp_link_list) {
4770 struct css_set *cg = link->cg;
4771 struct task_struct *task;
4772 int count = 0;
4773 seq_printf(seq, "css_set %p\n", cg);
4774 list_for_each_entry(task, &cg->tasks, cg_list) {
4775 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4776 seq_puts(seq, " ...\n");
4777 break;
4778 } else {
4779 seq_printf(seq, " task %d\n",
4780 task_pid_vnr(task));
4781 }
4782 }
4783 }
4784 read_unlock(&css_set_lock);
4785 return 0;
4786}
4787
4788static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
4789{
4790 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
4791}
4792
4793static struct cftype debug_files[] = {
4794 {
4795 .name = "cgroup_refcount",
4796 .read_u64 = cgroup_refcount_read,
4797 },
4798 {
4799 .name = "taskcount",
4800 .read_u64 = debug_taskcount_read,
4801 },
4802
4803 {
4804 .name = "current_css_set",
4805 .read_u64 = current_css_set_read,
4806 },
4807
4808 {
4809 .name = "current_css_set_refcount",
4810 .read_u64 = current_css_set_refcount_read,
4811 },
4812
4813 {
4814 .name = "current_css_set_cg_links",
4815 .read_seq_string = current_css_set_cg_links_read,
4816 },
4817
4818 {
4819 .name = "cgroup_css_links",
4820 .read_seq_string = cgroup_css_links_read,
4821 },
4822
4823 {
4824 .name = "releasable",
4825 .read_u64 = releasable_read,
4826 },
4827};
4828
4829static int debug_populate(struct cgroup_subsys *ss, struct cgroup *cont)
4830{
4831 return cgroup_add_files(cont, ss, debug_files,
4832 ARRAY_SIZE(debug_files));
4833}
4834
4835struct cgroup_subsys debug_subsys = {
4836 .name = "debug",
4837 .create = debug_create,
4838 .destroy = debug_destroy,
4839 .populate = debug_populate,
4840 .subsys_id = debug_subsys_id,
4841};
4842#endif /* CONFIG_CGROUP_DEBUG */