2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #include <linux/cgroup.h>
30 #include <linux/cred.h>
31 #include <linux/ctype.h>
32 #include <linux/errno.h>
33 #include <linux/init_task.h>
34 #include <linux/kernel.h>
35 #include <linux/list.h>
37 #include <linux/mutex.h>
38 #include <linux/mount.h>
39 #include <linux/pagemap.h>
40 #include <linux/proc_fs.h>
41 #include <linux/rcupdate.h>
42 #include <linux/sched.h>
43 #include <linux/backing-dev.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/delayacct.h>
51 #include <linux/cgroupstats.h>
52 #include <linux/hashtable.h>
53 #include <linux/namei.h>
54 #include <linux/pid_namespace.h>
55 #include <linux/idr.h>
56 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
57 #include <linux/flex_array.h> /* used in cgroup_attach_task */
58 #include <linux/kthread.h>
60 #include <linux/atomic.h>
63 * pidlists linger the following amount before being destroyed. The goal
64 * is avoiding frequent destruction in the middle of consecutive read calls
65 * Expiring in the middle is a performance problem not a correctness one.
66 * 1 sec should be enough.
68 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
70 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
74 * cgroup_tree_mutex nests above cgroup_mutex and protects cftypes, file
75 * creation/removal and hierarchy changing operations including cgroup
76 * creation, removal, css association and controller rebinding. This outer
77 * lock is needed mainly to resolve the circular dependency between kernfs
78 * active ref and cgroup_mutex. cgroup_tree_mutex nests above both.
80 static DEFINE_MUTEX(cgroup_tree_mutex);
83 * cgroup_mutex is the master lock. Any modification to cgroup or its
84 * hierarchy must be performed while holding it.
86 #ifdef CONFIG_PROVE_RCU
87 DEFINE_MUTEX(cgroup_mutex);
88 EXPORT_SYMBOL_GPL(cgroup_mutex); /* only for lockdep */
90 static DEFINE_MUTEX(cgroup_mutex);
94 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
95 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
97 static DEFINE_SPINLOCK(release_agent_path_lock);
99 #define cgroup_assert_mutexes_or_rcu_locked() \
100 rcu_lockdep_assert(rcu_read_lock_held() || \
101 lockdep_is_held(&cgroup_tree_mutex) || \
102 lockdep_is_held(&cgroup_mutex), \
103 "cgroup_[tree_]mutex or RCU read lock required");
106 * cgroup destruction makes heavy use of work items and there can be a lot
107 * of concurrent destructions. Use a separate workqueue so that cgroup
108 * destruction work items don't end up filling up max_active of system_wq
109 * which may lead to deadlock.
111 static struct workqueue_struct *cgroup_destroy_wq;
114 * pidlist destructions need to be flushed on cgroup destruction. Use a
115 * separate workqueue as flush domain.
117 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
119 /* generate an array of cgroup subsystem pointers */
120 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
121 static struct cgroup_subsys *cgroup_subsys[] = {
122 #include <linux/cgroup_subsys.h>
126 /* array of cgroup subsystem names */
127 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
128 static const char *cgroup_subsys_name[] = {
129 #include <linux/cgroup_subsys.h>
134 * The dummy hierarchy, reserved for the subsystems that are otherwise
135 * unattached - it never has more than a single cgroup, and all tasks are
136 * part of that cgroup.
138 static struct cgroupfs_root cgroup_dummy_root;
140 /* dummy_top is a shorthand for the dummy hierarchy's top cgroup */
141 static struct cgroup * const cgroup_dummy_top = &cgroup_dummy_root.top_cgroup;
143 /* The list of hierarchy roots */
145 static LIST_HEAD(cgroup_roots);
146 static int cgroup_root_count;
148 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
149 static DEFINE_IDR(cgroup_hierarchy_idr);
151 static struct cgroup_name root_cgroup_name = { .name = "/" };
154 * Assign a monotonically increasing serial number to cgroups. It
155 * guarantees cgroups with bigger numbers are newer than those with smaller
156 * numbers. Also, as cgroups are always appended to the parent's
157 * ->children list, it guarantees that sibling cgroups are always sorted in
158 * the ascending serial number order on the list. Protected by
161 static u64 cgroup_serial_nr_next = 1;
163 /* This flag indicates whether tasks in the fork and exit paths should
164 * check for fork/exit handlers to call. This avoids us having to do
165 * extra work in the fork/exit path if none of the subsystems need to
168 static int need_forkexit_callback __read_mostly;
170 static struct cftype cgroup_base_files[];
172 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
173 static int cgroup_destroy_locked(struct cgroup *cgrp);
174 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
176 static int cgroup_file_release(struct inode *inode, struct file *file);
177 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
180 * cgroup_css - obtain a cgroup's css for the specified subsystem
181 * @cgrp: the cgroup of interest
182 * @ss: the subsystem of interest (%NULL returns the dummy_css)
184 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
185 * function must be called either under cgroup_mutex or rcu_read_lock() and
186 * the caller is responsible for pinning the returned css if it wants to
187 * keep accessing it outside the said locks. This function may return
188 * %NULL if @cgrp doesn't have @subsys_id enabled.
190 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
191 struct cgroup_subsys *ss)
194 return rcu_dereference_check(cgrp->subsys[ss->id],
195 lockdep_is_held(&cgroup_tree_mutex) ||
196 lockdep_is_held(&cgroup_mutex));
198 return &cgrp->dummy_css;
201 /* convenient tests for these bits */
202 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
204 return test_bit(CGRP_DEAD, &cgrp->flags);
208 * cgroup_is_descendant - test ancestry
209 * @cgrp: the cgroup to be tested
210 * @ancestor: possible ancestor of @cgrp
212 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
213 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
214 * and @ancestor are accessible.
216 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
219 if (cgrp == ancestor)
225 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
227 static int cgroup_is_releasable(const struct cgroup *cgrp)
230 (1 << CGRP_RELEASABLE) |
231 (1 << CGRP_NOTIFY_ON_RELEASE);
232 return (cgrp->flags & bits) == bits;
235 static int notify_on_release(const struct cgroup *cgrp)
237 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
241 * for_each_css - iterate all css's of a cgroup
242 * @css: the iteration cursor
243 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
244 * @cgrp: the target cgroup to iterate css's of
246 * Should be called under cgroup_mutex.
248 #define for_each_css(css, ssid, cgrp) \
249 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
250 if (!((css) = rcu_dereference_check( \
251 (cgrp)->subsys[(ssid)], \
252 lockdep_is_held(&cgroup_tree_mutex) || \
253 lockdep_is_held(&cgroup_mutex)))) { } \
257 * for_each_subsys - iterate all enabled cgroup subsystems
258 * @ss: the iteration cursor
259 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
261 #define for_each_subsys(ss, ssid) \
262 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
263 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
265 /* iterate across the active hierarchies */
266 #define for_each_active_root(root) \
267 list_for_each_entry((root), &cgroup_roots, root_list)
269 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
271 return dentry->d_fsdata;
274 static inline struct cfent *__d_cfe(struct dentry *dentry)
276 return dentry->d_fsdata;
279 static inline struct cftype *__d_cft(struct dentry *dentry)
281 return __d_cfe(dentry)->type;
285 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
286 * @cgrp: the cgroup to be checked for liveness
288 * On success, returns true; the mutex should be later unlocked. On
289 * failure returns false with no lock held.
291 static bool cgroup_lock_live_group(struct cgroup *cgrp)
293 mutex_lock(&cgroup_mutex);
294 if (cgroup_is_dead(cgrp)) {
295 mutex_unlock(&cgroup_mutex);
301 /* the list of cgroups eligible for automatic release. Protected by
302 * release_list_lock */
303 static LIST_HEAD(release_list);
304 static DEFINE_RAW_SPINLOCK(release_list_lock);
305 static void cgroup_release_agent(struct work_struct *work);
306 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
307 static void check_for_release(struct cgroup *cgrp);
310 * A cgroup can be associated with multiple css_sets as different tasks may
311 * belong to different cgroups on different hierarchies. In the other
312 * direction, a css_set is naturally associated with multiple cgroups.
313 * This M:N relationship is represented by the following link structure
314 * which exists for each association and allows traversing the associations
317 struct cgrp_cset_link {
318 /* the cgroup and css_set this link associates */
320 struct css_set *cset;
322 /* list of cgrp_cset_links anchored at cgrp->cset_links */
323 struct list_head cset_link;
325 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
326 struct list_head cgrp_link;
329 /* The default css_set - used by init and its children prior to any
330 * hierarchies being mounted. It contains a pointer to the root state
331 * for each subsystem. Also used to anchor the list of css_sets. Not
332 * reference-counted, to improve performance when child cgroups
333 * haven't been created.
336 static struct css_set init_css_set;
337 static struct cgrp_cset_link init_cgrp_cset_link;
340 * css_set_lock protects the list of css_set objects, and the chain of
341 * tasks off each css_set. Nests outside task->alloc_lock due to
342 * css_task_iter_start().
344 static DEFINE_RWLOCK(css_set_lock);
345 static int css_set_count;
348 * hash table for cgroup groups. This improves the performance to find
349 * an existing css_set. This hash doesn't (currently) take into
350 * account cgroups in empty hierarchies.
352 #define CSS_SET_HASH_BITS 7
353 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
355 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
357 unsigned long key = 0UL;
358 struct cgroup_subsys *ss;
361 for_each_subsys(ss, i)
362 key += (unsigned long)css[i];
363 key = (key >> 16) ^ key;
369 * We don't maintain the lists running through each css_set to its task
370 * until after the first call to css_task_iter_start(). This reduces the
371 * fork()/exit() overhead for people who have cgroups compiled into their
372 * kernel but not actually in use.
374 static int use_task_css_set_links __read_mostly;
376 static void __put_css_set(struct css_set *cset, int taskexit)
378 struct cgrp_cset_link *link, *tmp_link;
381 * Ensure that the refcount doesn't hit zero while any readers
382 * can see it. Similar to atomic_dec_and_lock(), but for an
385 if (atomic_add_unless(&cset->refcount, -1, 1))
387 write_lock(&css_set_lock);
388 if (!atomic_dec_and_test(&cset->refcount)) {
389 write_unlock(&css_set_lock);
393 /* This css_set is dead. unlink it and release cgroup refcounts */
394 hash_del(&cset->hlist);
397 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
398 struct cgroup *cgrp = link->cgrp;
400 list_del(&link->cset_link);
401 list_del(&link->cgrp_link);
403 /* @cgrp can't go away while we're holding css_set_lock */
404 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
406 set_bit(CGRP_RELEASABLE, &cgrp->flags);
407 check_for_release(cgrp);
413 write_unlock(&css_set_lock);
414 kfree_rcu(cset, rcu_head);
418 * refcounted get/put for css_set objects
420 static inline void get_css_set(struct css_set *cset)
422 atomic_inc(&cset->refcount);
425 static inline void put_css_set(struct css_set *cset)
427 __put_css_set(cset, 0);
430 static inline void put_css_set_taskexit(struct css_set *cset)
432 __put_css_set(cset, 1);
436 * compare_css_sets - helper function for find_existing_css_set().
437 * @cset: candidate css_set being tested
438 * @old_cset: existing css_set for a task
439 * @new_cgrp: cgroup that's being entered by the task
440 * @template: desired set of css pointers in css_set (pre-calculated)
442 * Returns true if "cset" matches "old_cset" except for the hierarchy
443 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
445 static bool compare_css_sets(struct css_set *cset,
446 struct css_set *old_cset,
447 struct cgroup *new_cgrp,
448 struct cgroup_subsys_state *template[])
450 struct list_head *l1, *l2;
452 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
453 /* Not all subsystems matched */
458 * Compare cgroup pointers in order to distinguish between
459 * different cgroups in heirarchies with no subsystems. We
460 * could get by with just this check alone (and skip the
461 * memcmp above) but on most setups the memcmp check will
462 * avoid the need for this more expensive check on almost all
466 l1 = &cset->cgrp_links;
467 l2 = &old_cset->cgrp_links;
469 struct cgrp_cset_link *link1, *link2;
470 struct cgroup *cgrp1, *cgrp2;
474 /* See if we reached the end - both lists are equal length. */
475 if (l1 == &cset->cgrp_links) {
476 BUG_ON(l2 != &old_cset->cgrp_links);
479 BUG_ON(l2 == &old_cset->cgrp_links);
481 /* Locate the cgroups associated with these links. */
482 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
483 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
486 /* Hierarchies should be linked in the same order. */
487 BUG_ON(cgrp1->root != cgrp2->root);
490 * If this hierarchy is the hierarchy of the cgroup
491 * that's changing, then we need to check that this
492 * css_set points to the new cgroup; if it's any other
493 * hierarchy, then this css_set should point to the
494 * same cgroup as the old css_set.
496 if (cgrp1->root == new_cgrp->root) {
497 if (cgrp1 != new_cgrp)
508 * find_existing_css_set - init css array and find the matching css_set
509 * @old_cset: the css_set that we're using before the cgroup transition
510 * @cgrp: the cgroup that we're moving into
511 * @template: out param for the new set of csses, should be clear on entry
513 static struct css_set *find_existing_css_set(struct css_set *old_cset,
515 struct cgroup_subsys_state *template[])
517 struct cgroupfs_root *root = cgrp->root;
518 struct cgroup_subsys *ss;
519 struct css_set *cset;
524 * Build the set of subsystem state objects that we want to see in the
525 * new css_set. while subsystems can change globally, the entries here
526 * won't change, so no need for locking.
528 for_each_subsys(ss, i) {
529 if (root->subsys_mask & (1UL << i)) {
530 /* Subsystem is in this hierarchy. So we want
531 * the subsystem state from the new
533 template[i] = cgroup_css(cgrp, ss);
535 /* Subsystem is not in this hierarchy, so we
536 * don't want to change the subsystem state */
537 template[i] = old_cset->subsys[i];
541 key = css_set_hash(template);
542 hash_for_each_possible(css_set_table, cset, hlist, key) {
543 if (!compare_css_sets(cset, old_cset, cgrp, template))
546 /* This css_set matches what we need */
550 /* No existing cgroup group matched */
554 static void free_cgrp_cset_links(struct list_head *links_to_free)
556 struct cgrp_cset_link *link, *tmp_link;
558 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
559 list_del(&link->cset_link);
565 * allocate_cgrp_cset_links - allocate cgrp_cset_links
566 * @count: the number of links to allocate
567 * @tmp_links: list_head the allocated links are put on
569 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
570 * through ->cset_link. Returns 0 on success or -errno.
572 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
574 struct cgrp_cset_link *link;
577 INIT_LIST_HEAD(tmp_links);
579 for (i = 0; i < count; i++) {
580 link = kzalloc(sizeof(*link), GFP_KERNEL);
582 free_cgrp_cset_links(tmp_links);
585 list_add(&link->cset_link, tmp_links);
591 * link_css_set - a helper function to link a css_set to a cgroup
592 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
593 * @cset: the css_set to be linked
594 * @cgrp: the destination cgroup
596 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
599 struct cgrp_cset_link *link;
601 BUG_ON(list_empty(tmp_links));
602 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
605 list_move(&link->cset_link, &cgrp->cset_links);
607 * Always add links to the tail of the list so that the list
608 * is sorted by order of hierarchy creation
610 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
614 * find_css_set - return a new css_set with one cgroup updated
615 * @old_cset: the baseline css_set
616 * @cgrp: the cgroup to be updated
618 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
619 * substituted into the appropriate hierarchy.
621 static struct css_set *find_css_set(struct css_set *old_cset,
624 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
625 struct css_set *cset;
626 struct list_head tmp_links;
627 struct cgrp_cset_link *link;
630 lockdep_assert_held(&cgroup_mutex);
632 /* First see if we already have a cgroup group that matches
634 read_lock(&css_set_lock);
635 cset = find_existing_css_set(old_cset, cgrp, template);
638 read_unlock(&css_set_lock);
643 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
647 /* Allocate all the cgrp_cset_link objects that we'll need */
648 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
653 atomic_set(&cset->refcount, 1);
654 INIT_LIST_HEAD(&cset->cgrp_links);
655 INIT_LIST_HEAD(&cset->tasks);
656 INIT_HLIST_NODE(&cset->hlist);
658 /* Copy the set of subsystem state objects generated in
659 * find_existing_css_set() */
660 memcpy(cset->subsys, template, sizeof(cset->subsys));
662 write_lock(&css_set_lock);
663 /* Add reference counts and links from the new css_set. */
664 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
665 struct cgroup *c = link->cgrp;
667 if (c->root == cgrp->root)
669 link_css_set(&tmp_links, cset, c);
672 BUG_ON(!list_empty(&tmp_links));
676 /* Add this cgroup group to the hash table */
677 key = css_set_hash(cset->subsys);
678 hash_add(css_set_table, &cset->hlist, key);
680 write_unlock(&css_set_lock);
686 * Return the cgroup for "task" from the given hierarchy. Must be
687 * called with cgroup_mutex held.
689 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
690 struct cgroupfs_root *root)
692 struct css_set *cset;
693 struct cgroup *res = NULL;
695 BUG_ON(!mutex_is_locked(&cgroup_mutex));
696 read_lock(&css_set_lock);
698 * No need to lock the task - since we hold cgroup_mutex the
699 * task can't change groups, so the only thing that can happen
700 * is that it exits and its css is set back to init_css_set.
702 cset = task_css_set(task);
703 if (cset == &init_css_set) {
704 res = &root->top_cgroup;
706 struct cgrp_cset_link *link;
708 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
709 struct cgroup *c = link->cgrp;
711 if (c->root == root) {
717 read_unlock(&css_set_lock);
723 * There is one global cgroup mutex. We also require taking
724 * task_lock() when dereferencing a task's cgroup subsys pointers.
725 * See "The task_lock() exception", at the end of this comment.
727 * A task must hold cgroup_mutex to modify cgroups.
729 * Any task can increment and decrement the count field without lock.
730 * So in general, code holding cgroup_mutex can't rely on the count
731 * field not changing. However, if the count goes to zero, then only
732 * cgroup_attach_task() can increment it again. Because a count of zero
733 * means that no tasks are currently attached, therefore there is no
734 * way a task attached to that cgroup can fork (the other way to
735 * increment the count). So code holding cgroup_mutex can safely
736 * assume that if the count is zero, it will stay zero. Similarly, if
737 * a task holds cgroup_mutex on a cgroup with zero count, it
738 * knows that the cgroup won't be removed, as cgroup_rmdir()
741 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
742 * (usually) take cgroup_mutex. These are the two most performance
743 * critical pieces of code here. The exception occurs on cgroup_exit(),
744 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
745 * is taken, and if the cgroup count is zero, a usermode call made
746 * to the release agent with the name of the cgroup (path relative to
747 * the root of cgroup file system) as the argument.
749 * A cgroup can only be deleted if both its 'count' of using tasks
750 * is zero, and its list of 'children' cgroups is empty. Since all
751 * tasks in the system use _some_ cgroup, and since there is always at
752 * least one task in the system (init, pid == 1), therefore, top_cgroup
753 * always has either children cgroups and/or using tasks. So we don't
754 * need a special hack to ensure that top_cgroup cannot be deleted.
756 * The task_lock() exception
758 * The need for this exception arises from the action of
759 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
760 * another. It does so using cgroup_mutex, however there are
761 * several performance critical places that need to reference
762 * task->cgroup without the expense of grabbing a system global
763 * mutex. Therefore except as noted below, when dereferencing or, as
764 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
765 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
766 * the task_struct routinely used for such matters.
768 * P.S. One more locking exception. RCU is used to guard the
769 * update of a tasks cgroup pointer by cgroup_attach_task()
773 * A couple of forward declarations required, due to cyclic reference loop:
774 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
775 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
779 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode);
780 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
781 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
782 static const struct inode_operations cgroup_dir_inode_operations;
783 static const struct file_operations proc_cgroupstats_operations;
785 static struct backing_dev_info cgroup_backing_dev_info = {
787 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
790 static struct inode *cgroup_new_inode(umode_t mode, struct super_block *sb)
792 struct inode *inode = new_inode(sb);
795 inode->i_ino = get_next_ino();
796 inode->i_mode = mode;
797 inode->i_uid = current_fsuid();
798 inode->i_gid = current_fsgid();
799 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
800 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
805 static struct cgroup_name *cgroup_alloc_name(const char *name_str)
807 struct cgroup_name *name;
809 name = kmalloc(sizeof(*name) + strlen(name_str) + 1, GFP_KERNEL);
812 strcpy(name->name, name_str);
816 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
819 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
820 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
821 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
822 cft->ss->name, cft->name);
824 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
828 static void cgroup_free_fn(struct work_struct *work)
830 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
832 mutex_lock(&cgroup_mutex);
833 cgrp->root->number_of_cgroups--;
834 mutex_unlock(&cgroup_mutex);
837 * We get a ref to the parent's dentry, and put the ref when
838 * this cgroup is being freed, so it's guaranteed that the
839 * parent won't be destroyed before its children.
841 dput(cgrp->parent->dentry);
844 * Drop the active superblock reference that we took when we
845 * created the cgroup. This will free cgrp->root, if we are
846 * holding the last reference to @sb.
848 deactivate_super(cgrp->root->sb);
850 cgroup_pidlist_destroy_all(cgrp);
852 simple_xattrs_free(&cgrp->xattrs);
854 kfree(rcu_dereference_raw(cgrp->name));
858 static void cgroup_free_rcu(struct rcu_head *head)
860 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
862 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
863 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
866 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
868 /* is dentry a directory ? if so, kfree() associated cgroup */
869 if (S_ISDIR(inode->i_mode)) {
870 struct cgroup *cgrp = dentry->d_fsdata;
872 BUG_ON(!(cgroup_is_dead(cgrp)));
875 * XXX: cgrp->id is only used to look up css's. As cgroup
876 * and css's lifetimes will be decoupled, it should be made
877 * per-subsystem and moved to css->id so that lookups are
878 * successful until the target css is released.
880 mutex_lock(&cgroup_mutex);
881 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
882 mutex_unlock(&cgroup_mutex);
885 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
887 struct cfent *cfe = __d_cfe(dentry);
888 struct cgroup *cgrp = dentry->d_parent->d_fsdata;
890 WARN_ONCE(!list_empty(&cfe->node) &&
891 cgrp != &cgrp->root->top_cgroup,
892 "cfe still linked for %s\n", cfe->type->name);
893 simple_xattrs_free(&cfe->xattrs);
899 static void remove_dir(struct dentry *d)
901 struct dentry *parent = dget(d->d_parent);
904 simple_rmdir(parent->d_inode, d);
908 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
912 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
913 lockdep_assert_held(&cgroup_tree_mutex);
916 * If we're doing cleanup due to failure of cgroup_create(),
917 * the corresponding @cfe may not exist.
919 list_for_each_entry(cfe, &cgrp->files, node) {
920 struct dentry *d = cfe->dentry;
922 if (cft && cfe->type != cft)
927 simple_unlink(cgrp->dentry->d_inode, d);
928 list_del_init(&cfe->node);
936 * cgroup_clear_dir - remove subsys files in a cgroup directory
937 * @cgrp: target cgroup
938 * @subsys_mask: mask of the subsystem ids whose files should be removed
940 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
942 struct cgroup_subsys *ss;
945 for_each_subsys(ss, i) {
946 struct cftype_set *set;
948 if (!test_bit(i, &subsys_mask))
950 list_for_each_entry(set, &ss->cftsets, node)
951 cgroup_addrm_files(cgrp, set->cfts, false);
956 * NOTE : the dentry must have been dget()'ed
958 static void cgroup_d_remove_dir(struct dentry *dentry)
960 struct dentry *parent;
962 parent = dentry->d_parent;
963 spin_lock(&parent->d_lock);
964 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
965 list_del_init(&dentry->d_u.d_child);
966 spin_unlock(&dentry->d_lock);
967 spin_unlock(&parent->d_lock);
971 static int rebind_subsystems(struct cgroupfs_root *root,
972 unsigned long added_mask, unsigned removed_mask)
974 struct cgroup *cgrp = &root->top_cgroup;
975 struct cgroup_subsys *ss;
978 lockdep_assert_held(&cgroup_tree_mutex);
979 lockdep_assert_held(&cgroup_mutex);
981 /* Check that any added subsystems are currently free */
982 for_each_subsys(ss, i)
983 if ((added_mask & (1 << i)) && ss->root != &cgroup_dummy_root)
986 ret = cgroup_populate_dir(cgrp, added_mask);
991 * Nothing can fail from this point on. Remove files for the
992 * removed subsystems and rebind each subsystem.
994 mutex_unlock(&cgroup_mutex);
995 cgroup_clear_dir(cgrp, removed_mask);
996 mutex_lock(&cgroup_mutex);
998 for_each_subsys(ss, i) {
999 unsigned long bit = 1UL << i;
1001 if (bit & added_mask) {
1002 /* We're binding this subsystem to this hierarchy */
1003 BUG_ON(cgroup_css(cgrp, ss));
1004 BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
1005 BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
1007 rcu_assign_pointer(cgrp->subsys[i],
1008 cgroup_css(cgroup_dummy_top, ss));
1009 cgroup_css(cgrp, ss)->cgroup = cgrp;
1013 ss->bind(cgroup_css(cgrp, ss));
1015 /* refcount was already taken, and we're keeping it */
1016 root->subsys_mask |= bit;
1017 } else if (bit & removed_mask) {
1018 /* We're removing this subsystem */
1019 BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1020 BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1023 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1025 cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1026 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1028 cgroup_subsys[i]->root = &cgroup_dummy_root;
1029 root->subsys_mask &= ~bit;
1034 * Mark @root has finished binding subsystems. @root->subsys_mask
1035 * now matches the bound subsystems.
1037 root->flags |= CGRP_ROOT_SUBSYS_BOUND;
1042 static int cgroup_show_options(struct seq_file *seq, struct dentry *dentry)
1044 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
1045 struct cgroup_subsys *ss;
1048 for_each_subsys(ss, ssid)
1049 if (root->subsys_mask & (1 << ssid))
1050 seq_printf(seq, ",%s", ss->name);
1051 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1052 seq_puts(seq, ",sane_behavior");
1053 if (root->flags & CGRP_ROOT_NOPREFIX)
1054 seq_puts(seq, ",noprefix");
1055 if (root->flags & CGRP_ROOT_XATTR)
1056 seq_puts(seq, ",xattr");
1058 spin_lock(&release_agent_path_lock);
1059 if (strlen(root->release_agent_path))
1060 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1061 spin_unlock(&release_agent_path_lock);
1063 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1064 seq_puts(seq, ",clone_children");
1065 if (strlen(root->name))
1066 seq_printf(seq, ",name=%s", root->name);
1070 struct cgroup_sb_opts {
1071 unsigned long subsys_mask;
1072 unsigned long flags;
1073 char *release_agent;
1074 bool cpuset_clone_children;
1076 /* User explicitly requested empty subsystem */
1079 struct cgroupfs_root *new_root;
1084 * Convert a hierarchy specifier into a bitmask of subsystems and
1085 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1086 * array. This function takes refcounts on subsystems to be used, unless it
1087 * returns error, in which case no refcounts are taken.
1089 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1091 char *token, *o = data;
1092 bool all_ss = false, one_ss = false;
1093 unsigned long mask = (unsigned long)-1;
1094 struct cgroup_subsys *ss;
1097 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1099 #ifdef CONFIG_CPUSETS
1100 mask = ~(1UL << cpuset_cgrp_id);
1103 memset(opts, 0, sizeof(*opts));
1105 while ((token = strsep(&o, ",")) != NULL) {
1108 if (!strcmp(token, "none")) {
1109 /* Explicitly have no subsystems */
1113 if (!strcmp(token, "all")) {
1114 /* Mutually exclusive option 'all' + subsystem name */
1120 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1121 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1124 if (!strcmp(token, "noprefix")) {
1125 opts->flags |= CGRP_ROOT_NOPREFIX;
1128 if (!strcmp(token, "clone_children")) {
1129 opts->cpuset_clone_children = true;
1132 if (!strcmp(token, "xattr")) {
1133 opts->flags |= CGRP_ROOT_XATTR;
1136 if (!strncmp(token, "release_agent=", 14)) {
1137 /* Specifying two release agents is forbidden */
1138 if (opts->release_agent)
1140 opts->release_agent =
1141 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1142 if (!opts->release_agent)
1146 if (!strncmp(token, "name=", 5)) {
1147 const char *name = token + 5;
1148 /* Can't specify an empty name */
1151 /* Must match [\w.-]+ */
1152 for (i = 0; i < strlen(name); i++) {
1156 if ((c == '.') || (c == '-') || (c == '_'))
1160 /* Specifying two names is forbidden */
1163 opts->name = kstrndup(name,
1164 MAX_CGROUP_ROOT_NAMELEN - 1,
1172 for_each_subsys(ss, i) {
1173 if (strcmp(token, ss->name))
1178 /* Mutually exclusive option 'all' + subsystem name */
1181 set_bit(i, &opts->subsys_mask);
1186 if (i == CGROUP_SUBSYS_COUNT)
1191 * If the 'all' option was specified select all the subsystems,
1192 * otherwise if 'none', 'name=' and a subsystem name options
1193 * were not specified, let's default to 'all'
1195 if (all_ss || (!one_ss && !opts->none && !opts->name))
1196 for_each_subsys(ss, i)
1198 set_bit(i, &opts->subsys_mask);
1200 /* Consistency checks */
1202 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1203 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1205 if (opts->flags & CGRP_ROOT_NOPREFIX) {
1206 pr_err("cgroup: sane_behavior: noprefix is not allowed\n");
1210 if (opts->cpuset_clone_children) {
1211 pr_err("cgroup: sane_behavior: clone_children is not allowed\n");
1217 * Option noprefix was introduced just for backward compatibility
1218 * with the old cpuset, so we allow noprefix only if mounting just
1219 * the cpuset subsystem.
1221 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1225 /* Can't specify "none" and some subsystems */
1226 if (opts->subsys_mask && opts->none)
1230 * We either have to specify by name or by subsystems. (So all
1231 * empty hierarchies must have a name).
1233 if (!opts->subsys_mask && !opts->name)
1239 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1242 struct cgroupfs_root *root = sb->s_fs_info;
1243 struct cgroup *cgrp = &root->top_cgroup;
1244 struct cgroup_sb_opts opts;
1245 unsigned long added_mask, removed_mask;
1247 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1248 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1252 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1253 mutex_lock(&cgroup_tree_mutex);
1254 mutex_lock(&cgroup_mutex);
1256 /* See what subsystems are wanted */
1257 ret = parse_cgroupfs_options(data, &opts);
1261 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1262 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1263 task_tgid_nr(current), current->comm);
1265 added_mask = opts.subsys_mask & ~root->subsys_mask;
1266 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1268 /* Don't allow flags or name to change at remount */
1269 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1270 (opts.name && strcmp(opts.name, root->name))) {
1271 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1272 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1273 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1278 /* remounting is not allowed for populated hierarchies */
1279 if (root->number_of_cgroups > 1) {
1284 ret = rebind_subsystems(root, added_mask, removed_mask);
1288 if (opts.release_agent) {
1289 spin_lock(&release_agent_path_lock);
1290 strcpy(root->release_agent_path, opts.release_agent);
1291 spin_unlock(&release_agent_path_lock);
1294 kfree(opts.release_agent);
1296 mutex_unlock(&cgroup_mutex);
1297 mutex_unlock(&cgroup_tree_mutex);
1298 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1302 static const struct super_operations cgroup_ops = {
1303 .statfs = simple_statfs,
1304 .drop_inode = generic_delete_inode,
1305 .show_options = cgroup_show_options,
1306 .remount_fs = cgroup_remount,
1309 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1311 INIT_LIST_HEAD(&cgrp->sibling);
1312 INIT_LIST_HEAD(&cgrp->children);
1313 INIT_LIST_HEAD(&cgrp->files);
1314 INIT_LIST_HEAD(&cgrp->cset_links);
1315 INIT_LIST_HEAD(&cgrp->release_list);
1316 INIT_LIST_HEAD(&cgrp->pidlists);
1317 mutex_init(&cgrp->pidlist_mutex);
1318 cgrp->dummy_css.cgroup = cgrp;
1319 simple_xattrs_init(&cgrp->xattrs);
1322 static void init_cgroup_root(struct cgroupfs_root *root)
1324 struct cgroup *cgrp = &root->top_cgroup;
1326 INIT_LIST_HEAD(&root->root_list);
1327 root->number_of_cgroups = 1;
1329 RCU_INIT_POINTER(cgrp->name, &root_cgroup_name);
1330 init_cgroup_housekeeping(cgrp);
1331 idr_init(&root->cgroup_idr);
1334 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
1338 lockdep_assert_held(&cgroup_mutex);
1340 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
1345 root->hierarchy_id = id;
1349 static void cgroup_exit_root_id(struct cgroupfs_root *root)
1351 lockdep_assert_held(&cgroup_mutex);
1353 if (root->hierarchy_id) {
1354 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1355 root->hierarchy_id = 0;
1359 static int cgroup_test_super(struct super_block *sb, void *data)
1361 struct cgroup_sb_opts *opts = data;
1362 struct cgroupfs_root *root = sb->s_fs_info;
1364 /* If we asked for a name then it must match */
1365 if (opts->name && strcmp(opts->name, root->name))
1369 * If we asked for subsystems (or explicitly for no
1370 * subsystems) then they must match
1372 if ((opts->subsys_mask || opts->none)
1373 && (opts->subsys_mask != root->subsys_mask))
1379 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1381 struct cgroupfs_root *root;
1383 if (!opts->subsys_mask && !opts->none)
1386 root = kzalloc(sizeof(*root), GFP_KERNEL);
1388 return ERR_PTR(-ENOMEM);
1390 init_cgroup_root(root);
1393 * We need to set @root->subsys_mask now so that @root can be
1394 * matched by cgroup_test_super() before it finishes
1395 * initialization; otherwise, competing mounts with the same
1396 * options may try to bind the same subsystems instead of waiting
1397 * for the first one leading to unexpected mount errors.
1398 * SUBSYS_BOUND will be set once actual binding is complete.
1400 root->subsys_mask = opts->subsys_mask;
1401 root->flags = opts->flags;
1402 if (opts->release_agent)
1403 strcpy(root->release_agent_path, opts->release_agent);
1405 strcpy(root->name, opts->name);
1406 if (opts->cpuset_clone_children)
1407 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1411 static void cgroup_free_root(struct cgroupfs_root *root)
1414 /* hierarhcy ID shoulid already have been released */
1415 WARN_ON_ONCE(root->hierarchy_id);
1417 idr_destroy(&root->cgroup_idr);
1422 static int cgroup_set_super(struct super_block *sb, void *data)
1425 struct cgroup_sb_opts *opts = data;
1427 /* If we don't have a new root, we can't set up a new sb */
1428 if (!opts->new_root)
1431 BUG_ON(!opts->subsys_mask && !opts->none);
1433 ret = set_anon_super(sb, NULL);
1437 sb->s_fs_info = opts->new_root;
1438 opts->new_root->sb = sb;
1440 sb->s_blocksize = PAGE_CACHE_SIZE;
1441 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1442 sb->s_magic = CGROUP_SUPER_MAGIC;
1443 sb->s_op = &cgroup_ops;
1448 static int cgroup_get_rootdir(struct super_block *sb)
1450 static const struct dentry_operations cgroup_dops = {
1451 .d_iput = cgroup_diput,
1452 .d_delete = always_delete_dentry,
1455 struct inode *inode =
1456 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1461 inode->i_fop = &simple_dir_operations;
1462 inode->i_op = &cgroup_dir_inode_operations;
1463 /* directories start off with i_nlink == 2 (for "." entry) */
1465 sb->s_root = d_make_root(inode);
1468 /* for everything else we want ->d_op set */
1469 sb->s_d_op = &cgroup_dops;
1473 static int cgroup_setup_root(struct cgroupfs_root *root)
1475 LIST_HEAD(tmp_links);
1476 struct super_block *sb = root->sb;
1477 struct cgroup *root_cgrp = &root->top_cgroup;
1478 struct cgroupfs_root *existing_root;
1479 struct css_set *cset;
1480 struct inode *inode;
1481 const struct cred *cred;
1484 lockdep_assert_held(&cgroup_tree_mutex);
1485 lockdep_assert_held(&cgroup_mutex);
1486 BUG_ON(sb->s_root != NULL);
1488 mutex_unlock(&cgroup_mutex);
1489 mutex_unlock(&cgroup_tree_mutex);
1491 ret = cgroup_get_rootdir(sb);
1493 mutex_lock(&cgroup_tree_mutex);
1494 mutex_lock(&cgroup_mutex);
1497 inode = sb->s_root->d_inode;
1499 mutex_lock(&inode->i_mutex);
1500 mutex_lock(&cgroup_tree_mutex);
1501 mutex_lock(&cgroup_mutex);
1503 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1506 root_cgrp->id = ret;
1508 /* check for name clashes with existing mounts */
1510 if (strlen(root->name))
1511 for_each_active_root(existing_root)
1512 if (!strcmp(existing_root->name, root->name))
1516 * We're accessing css_set_count without locking css_set_lock here,
1517 * but that's OK - it can only be increased by someone holding
1518 * cgroup_lock, and that's us. The worst that can happen is that we
1519 * have some link structures left over
1521 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1525 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1526 ret = cgroup_init_root_id(root, 2, 0);
1530 sb->s_root->d_fsdata = root_cgrp;
1531 root_cgrp->dentry = sb->s_root;
1534 * We're inside get_sb() and will call lookup_one_len() to create
1535 * the root files, which doesn't work if SELinux is in use. The
1536 * following cred dancing somehow works around it. See 2ce9738ba
1537 * ("cgroupfs: use init_cred when populating new cgroupfs mount")
1540 cred = override_creds(&init_cred);
1542 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1546 ret = rebind_subsystems(root, root->subsys_mask, 0);
1553 * There must be no failure case after here, since rebinding takes
1554 * care of subsystems' refcounts, which are explicitly dropped in
1555 * the failure exit path.
1557 list_add(&root->root_list, &cgroup_roots);
1558 cgroup_root_count++;
1561 * Link the top cgroup in this hierarchy into all the css_set
1564 write_lock(&css_set_lock);
1565 hash_for_each(css_set_table, i, cset, hlist)
1566 link_css_set(&tmp_links, cset, root_cgrp);
1567 write_unlock(&css_set_lock);
1569 BUG_ON(!list_empty(&root_cgrp->children));
1570 BUG_ON(root->number_of_cgroups != 1);
1576 cgroup_addrm_files(&root->top_cgroup, cgroup_base_files, false);
1578 cgroup_exit_root_id(root);
1580 mutex_unlock(&inode->i_mutex);
1581 free_cgrp_cset_links(&tmp_links);
1585 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1586 int flags, const char *unused_dev_name,
1589 struct super_block *sb = NULL;
1590 struct cgroupfs_root *root = NULL;
1591 struct cgroup_sb_opts opts;
1592 struct cgroupfs_root *new_root;
1595 mutex_lock(&cgroup_tree_mutex);
1596 mutex_lock(&cgroup_mutex);
1598 /* First find the desired set of subsystems */
1599 ret = parse_cgroupfs_options(data, &opts);
1604 * Allocate a new cgroup root. We may not need it if we're
1605 * reusing an existing hierarchy.
1607 new_root = cgroup_root_from_opts(&opts);
1608 if (IS_ERR(new_root)) {
1609 ret = PTR_ERR(new_root);
1612 opts.new_root = new_root;
1614 /* Locate an existing or new sb for this hierarchy */
1615 mutex_unlock(&cgroup_mutex);
1616 mutex_unlock(&cgroup_tree_mutex);
1617 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, 0, &opts);
1618 mutex_lock(&cgroup_tree_mutex);
1619 mutex_lock(&cgroup_mutex);
1622 cgroup_free_root(opts.new_root);
1626 root = sb->s_fs_info;
1628 if (root == opts.new_root) {
1629 ret = cgroup_setup_root(root);
1634 * We re-used an existing hierarchy - the new root (if
1635 * any) is not needed
1637 cgroup_free_root(opts.new_root);
1639 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1640 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1641 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1645 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1652 mutex_unlock(&cgroup_mutex);
1653 mutex_unlock(&cgroup_tree_mutex);
1655 if (ret && !IS_ERR_OR_NULL(sb))
1656 deactivate_locked_super(sb);
1658 kfree(opts.release_agent);
1662 return dget(sb->s_root);
1664 return ERR_PTR(ret);
1667 static void cgroup_kill_sb(struct super_block *sb)
1669 struct cgroupfs_root *root = sb->s_fs_info;
1670 struct cgroup *cgrp = &root->top_cgroup;
1671 struct cgrp_cset_link *link, *tmp_link;
1676 BUG_ON(root->number_of_cgroups != 1);
1677 BUG_ON(!list_empty(&cgrp->children));
1679 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1680 mutex_lock(&cgroup_tree_mutex);
1681 mutex_lock(&cgroup_mutex);
1683 /* Rebind all subsystems back to the default hierarchy */
1684 if (root->flags & CGRP_ROOT_SUBSYS_BOUND) {
1685 ret = rebind_subsystems(root, 0, root->subsys_mask);
1686 /* Shouldn't be able to fail ... */
1691 * Release all the links from cset_links to this hierarchy's
1694 write_lock(&css_set_lock);
1696 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1697 list_del(&link->cset_link);
1698 list_del(&link->cgrp_link);
1701 write_unlock(&css_set_lock);
1703 if (!list_empty(&root->root_list)) {
1704 list_del(&root->root_list);
1705 cgroup_root_count--;
1708 cgroup_exit_root_id(root);
1710 mutex_unlock(&cgroup_mutex);
1711 mutex_unlock(&cgroup_tree_mutex);
1712 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1714 simple_xattrs_free(&cgrp->xattrs);
1716 kill_litter_super(sb);
1717 cgroup_free_root(root);
1720 static struct file_system_type cgroup_fs_type = {
1722 .mount = cgroup_mount,
1723 .kill_sb = cgroup_kill_sb,
1726 static struct kobject *cgroup_kobj;
1729 * cgroup_path - generate the path of a cgroup
1730 * @cgrp: the cgroup in question
1731 * @buf: the buffer to write the path into
1732 * @buflen: the length of the buffer
1734 * Writes path of cgroup into buf. Returns 0 on success, -errno on error.
1736 * We can't generate cgroup path using dentry->d_name, as accessing
1737 * dentry->name must be protected by irq-unsafe dentry->d_lock or parent
1738 * inode's i_mutex, while on the other hand cgroup_path() can be called
1739 * with some irq-safe spinlocks held.
1741 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1743 int ret = -ENAMETOOLONG;
1746 if (!cgrp->parent) {
1747 if (strlcpy(buf, "/", buflen) >= buflen)
1748 return -ENAMETOOLONG;
1752 start = buf + buflen - 1;
1757 const char *name = cgroup_name(cgrp);
1761 if ((start -= len) < buf)
1763 memcpy(start, name, len);
1769 cgrp = cgrp->parent;
1770 } while (cgrp->parent);
1772 memmove(buf, start, buf + buflen - start);
1777 EXPORT_SYMBOL_GPL(cgroup_path);
1780 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1781 * @task: target task
1782 * @buf: the buffer to write the path into
1783 * @buflen: the length of the buffer
1785 * Determine @task's cgroup on the first (the one with the lowest non-zero
1786 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1787 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1788 * cgroup controller callbacks.
1790 * Returns 0 on success, fails with -%ENAMETOOLONG if @buflen is too short.
1792 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1794 struct cgroupfs_root *root;
1795 struct cgroup *cgrp;
1796 int hierarchy_id = 1, ret = 0;
1799 return -ENAMETOOLONG;
1801 mutex_lock(&cgroup_mutex);
1803 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1806 cgrp = task_cgroup_from_root(task, root);
1807 ret = cgroup_path(cgrp, buf, buflen);
1809 /* if no hierarchy exists, everyone is in "/" */
1810 memcpy(buf, "/", 2);
1813 mutex_unlock(&cgroup_mutex);
1816 EXPORT_SYMBOL_GPL(task_cgroup_path);
1819 * Control Group taskset
1821 struct task_and_cgroup {
1822 struct task_struct *task;
1823 struct cgroup *cgrp;
1824 struct css_set *cset;
1827 struct cgroup_taskset {
1828 struct task_and_cgroup single;
1829 struct flex_array *tc_array;
1832 struct cgroup *cur_cgrp;
1836 * cgroup_taskset_first - reset taskset and return the first task
1837 * @tset: taskset of interest
1839 * @tset iteration is initialized and the first task is returned.
1841 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1843 if (tset->tc_array) {
1845 return cgroup_taskset_next(tset);
1847 tset->cur_cgrp = tset->single.cgrp;
1848 return tset->single.task;
1851 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1854 * cgroup_taskset_next - iterate to the next task in taskset
1855 * @tset: taskset of interest
1857 * Return the next task in @tset. Iteration must have been initialized
1858 * with cgroup_taskset_first().
1860 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1862 struct task_and_cgroup *tc;
1864 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1867 tc = flex_array_get(tset->tc_array, tset->idx++);
1868 tset->cur_cgrp = tc->cgrp;
1871 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1874 * cgroup_taskset_cur_css - return the matching css for the current task
1875 * @tset: taskset of interest
1876 * @subsys_id: the ID of the target subsystem
1878 * Return the css for the current (last returned) task of @tset for
1879 * subsystem specified by @subsys_id. This function must be preceded by
1880 * either cgroup_taskset_first() or cgroup_taskset_next().
1882 struct cgroup_subsys_state *cgroup_taskset_cur_css(struct cgroup_taskset *tset,
1885 return cgroup_css(tset->cur_cgrp, cgroup_subsys[subsys_id]);
1887 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_css);
1890 * cgroup_taskset_size - return the number of tasks in taskset
1891 * @tset: taskset of interest
1893 int cgroup_taskset_size(struct cgroup_taskset *tset)
1895 return tset->tc_array ? tset->tc_array_len : 1;
1897 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1901 * cgroup_task_migrate - move a task from one cgroup to another.
1903 * Must be called with cgroup_mutex and threadgroup locked.
1905 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1906 struct task_struct *tsk,
1907 struct css_set *new_cset)
1909 struct css_set *old_cset;
1912 * We are synchronized through threadgroup_lock() against PF_EXITING
1913 * setting such that we can't race against cgroup_exit() changing the
1914 * css_set to init_css_set and dropping the old one.
1916 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1917 old_cset = task_css_set(tsk);
1920 rcu_assign_pointer(tsk->cgroups, new_cset);
1923 /* Update the css_set linked lists if we're using them */
1924 write_lock(&css_set_lock);
1925 if (!list_empty(&tsk->cg_list))
1926 list_move(&tsk->cg_list, &new_cset->tasks);
1927 write_unlock(&css_set_lock);
1930 * We just gained a reference on old_cset by taking it from the
1931 * task. As trading it for new_cset is protected by cgroup_mutex,
1932 * we're safe to drop it here; it will be freed under RCU.
1934 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1935 put_css_set(old_cset);
1939 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1940 * @cgrp: the cgroup to attach to
1941 * @tsk: the task or the leader of the threadgroup to be attached
1942 * @threadgroup: attach the whole threadgroup?
1944 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1945 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1947 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1950 int retval, i, group_size;
1951 struct cgroupfs_root *root = cgrp->root;
1952 struct cgroup_subsys_state *css, *failed_css = NULL;
1953 /* threadgroup list cursor and array */
1954 struct task_struct *leader = tsk;
1955 struct task_and_cgroup *tc;
1956 struct flex_array *group;
1957 struct cgroup_taskset tset = { };
1960 * step 0: in order to do expensive, possibly blocking operations for
1961 * every thread, we cannot iterate the thread group list, since it needs
1962 * rcu or tasklist locked. instead, build an array of all threads in the
1963 * group - group_rwsem prevents new threads from appearing, and if
1964 * threads exit, this will just be an over-estimate.
1967 group_size = get_nr_threads(tsk);
1970 /* flex_array supports very large thread-groups better than kmalloc. */
1971 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1974 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1975 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1977 goto out_free_group_list;
1981 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1982 * already PF_EXITING could be freed from underneath us unless we
1983 * take an rcu_read_lock.
1987 struct task_and_cgroup ent;
1989 /* @tsk either already exited or can't exit until the end */
1990 if (tsk->flags & PF_EXITING)
1993 /* as per above, nr_threads may decrease, but not increase. */
1994 BUG_ON(i >= group_size);
1996 ent.cgrp = task_cgroup_from_root(tsk, root);
1997 /* nothing to do if this task is already in the cgroup */
1998 if (ent.cgrp == cgrp)
2001 * saying GFP_ATOMIC has no effect here because we did prealloc
2002 * earlier, but it's good form to communicate our expectations.
2004 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
2005 BUG_ON(retval != 0);
2010 } while_each_thread(leader, tsk);
2012 /* remember the number of threads in the array for later. */
2014 tset.tc_array = group;
2015 tset.tc_array_len = group_size;
2017 /* methods shouldn't be called if no task is actually migrating */
2020 goto out_free_group_list;
2023 * step 1: check that we can legitimately attach to the cgroup.
2025 for_each_css(css, i, cgrp) {
2026 if (css->ss->can_attach) {
2027 retval = css->ss->can_attach(css, &tset);
2030 goto out_cancel_attach;
2036 * step 2: make sure css_sets exist for all threads to be migrated.
2037 * we use find_css_set, which allocates a new one if necessary.
2039 for (i = 0; i < group_size; i++) {
2040 struct css_set *old_cset;
2042 tc = flex_array_get(group, i);
2043 old_cset = task_css_set(tc->task);
2044 tc->cset = find_css_set(old_cset, cgrp);
2047 goto out_put_css_set_refs;
2052 * step 3: now that we're guaranteed success wrt the css_sets,
2053 * proceed to move all tasks to the new cgroup. There are no
2054 * failure cases after here, so this is the commit point.
2056 for (i = 0; i < group_size; i++) {
2057 tc = flex_array_get(group, i);
2058 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
2060 /* nothing is sensitive to fork() after this point. */
2063 * step 4: do subsystem attach callbacks.
2065 for_each_css(css, i, cgrp)
2066 if (css->ss->attach)
2067 css->ss->attach(css, &tset);
2070 * step 5: success! and cleanup
2073 out_put_css_set_refs:
2075 for (i = 0; i < group_size; i++) {
2076 tc = flex_array_get(group, i);
2079 put_css_set(tc->cset);
2084 for_each_css(css, i, cgrp) {
2085 if (css == failed_css)
2087 if (css->ss->cancel_attach)
2088 css->ss->cancel_attach(css, &tset);
2091 out_free_group_list:
2092 flex_array_free(group);
2097 * Find the task_struct of the task to attach by vpid and pass it along to the
2098 * function to attach either it or all tasks in its threadgroup. Will lock
2099 * cgroup_mutex and threadgroup; may take task_lock of task.
2101 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2103 struct task_struct *tsk;
2104 const struct cred *cred = current_cred(), *tcred;
2107 if (!cgroup_lock_live_group(cgrp))
2113 tsk = find_task_by_vpid(pid);
2117 goto out_unlock_cgroup;
2120 * even if we're attaching all tasks in the thread group, we
2121 * only need to check permissions on one of them.
2123 tcred = __task_cred(tsk);
2124 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2125 !uid_eq(cred->euid, tcred->uid) &&
2126 !uid_eq(cred->euid, tcred->suid)) {
2129 goto out_unlock_cgroup;
2135 tsk = tsk->group_leader;
2138 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2139 * trapped in a cpuset, or RT worker may be born in a cgroup
2140 * with no rt_runtime allocated. Just say no.
2142 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2145 goto out_unlock_cgroup;
2148 get_task_struct(tsk);
2151 threadgroup_lock(tsk);
2153 if (!thread_group_leader(tsk)) {
2155 * a race with de_thread from another thread's exec()
2156 * may strip us of our leadership, if this happens,
2157 * there is no choice but to throw this task away and
2158 * try again; this is
2159 * "double-double-toil-and-trouble-check locking".
2161 threadgroup_unlock(tsk);
2162 put_task_struct(tsk);
2163 goto retry_find_task;
2167 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2169 threadgroup_unlock(tsk);
2171 put_task_struct(tsk);
2173 mutex_unlock(&cgroup_mutex);
2178 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2179 * @from: attach to all cgroups of a given task
2180 * @tsk: the task to be attached
2182 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2184 struct cgroupfs_root *root;
2187 mutex_lock(&cgroup_mutex);
2188 for_each_active_root(root) {
2189 struct cgroup *from_cgrp = task_cgroup_from_root(from, root);
2191 retval = cgroup_attach_task(from_cgrp, tsk, false);
2195 mutex_unlock(&cgroup_mutex);
2199 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2201 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2202 struct cftype *cft, u64 pid)
2204 return attach_task_by_pid(css->cgroup, pid, false);
2207 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2208 struct cftype *cft, u64 tgid)
2210 return attach_task_by_pid(css->cgroup, tgid, true);
2213 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2214 struct cftype *cft, const char *buffer)
2216 BUILD_BUG_ON(sizeof(css->cgroup->root->release_agent_path) < PATH_MAX);
2217 if (strlen(buffer) >= PATH_MAX)
2219 if (!cgroup_lock_live_group(css->cgroup))
2221 spin_lock(&release_agent_path_lock);
2222 strcpy(css->cgroup->root->release_agent_path, buffer);
2223 spin_unlock(&release_agent_path_lock);
2224 mutex_unlock(&cgroup_mutex);
2228 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2230 struct cgroup *cgrp = seq_css(seq)->cgroup;
2232 if (!cgroup_lock_live_group(cgrp))
2234 seq_puts(seq, cgrp->root->release_agent_path);
2235 seq_putc(seq, '\n');
2236 mutex_unlock(&cgroup_mutex);
2240 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2242 struct cgroup *cgrp = seq_css(seq)->cgroup;
2244 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2248 /* A buffer size big enough for numbers or short strings */
2249 #define CGROUP_LOCAL_BUFFER_SIZE 64
2251 static ssize_t cgroup_file_write(struct file *file, const char __user *userbuf,
2252 size_t nbytes, loff_t *ppos)
2254 struct cfent *cfe = __d_cfe(file->f_dentry);
2255 struct cftype *cft = __d_cft(file->f_dentry);
2256 struct cgroup_subsys_state *css = cfe->css;
2257 size_t max_bytes = cft->max_write_len ?: CGROUP_LOCAL_BUFFER_SIZE - 1;
2261 if (nbytes >= max_bytes)
2264 buf = kmalloc(nbytes + 1, GFP_KERNEL);
2268 if (copy_from_user(buf, userbuf, nbytes)) {
2275 if (cft->write_string) {
2276 ret = cft->write_string(css, cft, strstrip(buf));
2277 } else if (cft->write_u64) {
2278 unsigned long long v;
2279 ret = kstrtoull(buf, 0, &v);
2281 ret = cft->write_u64(css, cft, v);
2282 } else if (cft->write_s64) {
2284 ret = kstrtoll(buf, 0, &v);
2286 ret = cft->write_s64(css, cft, v);
2287 } else if (cft->trigger) {
2288 ret = cft->trigger(css, (unsigned int)cft->private);
2294 return ret ?: nbytes;
2298 * seqfile ops/methods for returning structured data. Currently just
2299 * supports string->u64 maps, but can be extended in future.
2302 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2304 struct cftype *cft = seq_cft(seq);
2306 if (cft->seq_start) {
2307 return cft->seq_start(seq, ppos);
2310 * The same behavior and code as single_open(). Returns
2311 * !NULL if pos is at the beginning; otherwise, NULL.
2313 return NULL + !*ppos;
2317 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2319 struct cftype *cft = seq_cft(seq);
2321 if (cft->seq_next) {
2322 return cft->seq_next(seq, v, ppos);
2325 * The same behavior and code as single_open(), always
2326 * terminate after the initial read.
2333 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2335 struct cftype *cft = seq_cft(seq);
2338 cft->seq_stop(seq, v);
2341 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2343 struct cftype *cft = seq_cft(m);
2344 struct cgroup_subsys_state *css = seq_css(m);
2347 return cft->seq_show(m, arg);
2350 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2351 else if (cft->read_s64)
2352 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2358 static struct seq_operations cgroup_seq_operations = {
2359 .start = cgroup_seqfile_start,
2360 .next = cgroup_seqfile_next,
2361 .stop = cgroup_seqfile_stop,
2362 .show = cgroup_seqfile_show,
2365 static int cgroup_file_open(struct inode *inode, struct file *file)
2367 struct cfent *cfe = __d_cfe(file->f_dentry);
2368 struct cftype *cft = __d_cft(file->f_dentry);
2369 struct cgroup *cgrp = __d_cgrp(cfe->dentry->d_parent);
2370 struct cgroup_subsys_state *css;
2371 struct cgroup_open_file *of;
2374 err = generic_file_open(inode, file);
2379 * If the file belongs to a subsystem, pin the css. Will be
2380 * unpinned either on open failure or release. This ensures that
2381 * @css stays alive for all file operations.
2384 css = cgroup_css(cgrp, cft->ss);
2385 if (cft->ss && !css_tryget(css))
2393 * @cfe->css is used by read/write/close to determine the
2394 * associated css. @file->private_data would be a better place but
2395 * that's already used by seqfile. Multiple accessors may use it
2396 * simultaneously which is okay as the association never changes.
2398 WARN_ON_ONCE(cfe->css && cfe->css != css);
2401 of = __seq_open_private(file, &cgroup_seq_operations,
2402 sizeof(struct cgroup_open_file));
2413 static int cgroup_file_release(struct inode *inode, struct file *file)
2415 struct cfent *cfe = __d_cfe(file->f_dentry);
2416 struct cgroup_subsys_state *css = cfe->css;
2420 return seq_release_private(inode, file);
2424 * cgroup_rename - Only allow simple rename of directories in place.
2426 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2427 struct inode *new_dir, struct dentry *new_dentry)
2430 struct cgroup_name *name, *old_name;
2431 struct cgroup *cgrp;
2434 * It's convinient to use parent dir's i_mutex to protected
2437 lockdep_assert_held(&old_dir->i_mutex);
2439 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2441 if (new_dentry->d_inode)
2443 if (old_dir != new_dir)
2446 cgrp = __d_cgrp(old_dentry);
2449 * This isn't a proper migration and its usefulness is very
2450 * limited. Disallow if sane_behavior.
2452 if (cgroup_sane_behavior(cgrp))
2455 name = cgroup_alloc_name(new_dentry->d_name.name);
2459 ret = simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2465 old_name = rcu_dereference_protected(cgrp->name, true);
2466 rcu_assign_pointer(cgrp->name, name);
2468 kfree_rcu(old_name, rcu_head);
2472 static struct simple_xattrs *__d_xattrs(struct dentry *dentry)
2474 if (S_ISDIR(dentry->d_inode->i_mode))
2475 return &__d_cgrp(dentry)->xattrs;
2477 return &__d_cfe(dentry)->xattrs;
2480 static inline int xattr_enabled(struct dentry *dentry)
2482 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
2483 return root->flags & CGRP_ROOT_XATTR;
2486 static bool is_valid_xattr(const char *name)
2488 if (!strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
2489 !strncmp(name, XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN))
2494 static int cgroup_setxattr(struct dentry *dentry, const char *name,
2495 const void *val, size_t size, int flags)
2497 if (!xattr_enabled(dentry))
2499 if (!is_valid_xattr(name))
2501 return simple_xattr_set(__d_xattrs(dentry), name, val, size, flags);
2504 static int cgroup_removexattr(struct dentry *dentry, const char *name)
2506 if (!xattr_enabled(dentry))
2508 if (!is_valid_xattr(name))
2510 return simple_xattr_remove(__d_xattrs(dentry), name);
2513 static ssize_t cgroup_getxattr(struct dentry *dentry, const char *name,
2514 void *buf, size_t size)
2516 if (!xattr_enabled(dentry))
2518 if (!is_valid_xattr(name))
2520 return simple_xattr_get(__d_xattrs(dentry), name, buf, size);
2523 static ssize_t cgroup_listxattr(struct dentry *dentry, char *buf, size_t size)
2525 if (!xattr_enabled(dentry))
2527 return simple_xattr_list(__d_xattrs(dentry), buf, size);
2530 static const struct file_operations cgroup_file_operations = {
2532 .write = cgroup_file_write,
2533 .llseek = generic_file_llseek,
2534 .open = cgroup_file_open,
2535 .release = cgroup_file_release,
2538 static const struct inode_operations cgroup_file_inode_operations = {
2539 .setxattr = cgroup_setxattr,
2540 .getxattr = cgroup_getxattr,
2541 .listxattr = cgroup_listxattr,
2542 .removexattr = cgroup_removexattr,
2545 static const struct inode_operations cgroup_dir_inode_operations = {
2546 .lookup = simple_lookup,
2547 .mkdir = cgroup_mkdir,
2548 .rmdir = cgroup_rmdir,
2549 .rename = cgroup_rename,
2550 .setxattr = cgroup_setxattr,
2551 .getxattr = cgroup_getxattr,
2552 .listxattr = cgroup_listxattr,
2553 .removexattr = cgroup_removexattr,
2556 static int cgroup_create_file(struct dentry *dentry, umode_t mode,
2557 struct super_block *sb)
2559 struct inode *inode;
2563 if (dentry->d_inode)
2566 inode = cgroup_new_inode(mode, sb);
2570 if (S_ISDIR(mode)) {
2571 inode->i_op = &cgroup_dir_inode_operations;
2572 inode->i_fop = &simple_dir_operations;
2574 /* start off with i_nlink == 2 (for "." entry) */
2576 inc_nlink(dentry->d_parent->d_inode);
2579 * Control reaches here with cgroup_mutex held.
2580 * @inode->i_mutex should nest outside cgroup_mutex but we
2581 * want to populate it immediately without releasing
2582 * cgroup_mutex. As @inode isn't visible to anyone else
2583 * yet, trylock will always succeed without affecting
2586 WARN_ON_ONCE(!mutex_trylock(&inode->i_mutex));
2587 } else if (S_ISREG(mode)) {
2589 inode->i_fop = &cgroup_file_operations;
2590 inode->i_op = &cgroup_file_inode_operations;
2592 d_instantiate(dentry, inode);
2593 dget(dentry); /* Extra count - pin the dentry in core */
2598 * cgroup_file_mode - deduce file mode of a control file
2599 * @cft: the control file in question
2601 * returns cft->mode if ->mode is not 0
2602 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2603 * returns S_IRUGO if it has only a read handler
2604 * returns S_IWUSR if it has only a write hander
2606 static umode_t cgroup_file_mode(const struct cftype *cft)
2613 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
2616 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
2623 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2625 struct dentry *dir = cgrp->dentry;
2626 struct cgroup *parent = __d_cgrp(dir);
2627 struct dentry *dentry;
2631 char name[CGROUP_FILE_NAME_MAX];
2633 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2635 cfe = kzalloc(sizeof(*cfe), GFP_KERNEL);
2639 cgroup_file_name(cgrp, cft, name);
2640 dentry = lookup_one_len(name, dir, strlen(name));
2641 if (IS_ERR(dentry)) {
2642 error = PTR_ERR(dentry);
2646 cfe->type = (void *)cft;
2647 cfe->dentry = dentry;
2648 dentry->d_fsdata = cfe;
2649 simple_xattrs_init(&cfe->xattrs);
2651 mode = cgroup_file_mode(cft);
2652 error = cgroup_create_file(dentry, mode | S_IFREG, cgrp->root->sb);
2654 list_add_tail(&cfe->node, &parent->files);
2664 * cgroup_addrm_files - add or remove files to a cgroup directory
2665 * @cgrp: the target cgroup
2666 * @cfts: array of cftypes to be added
2667 * @is_add: whether to add or remove
2669 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2670 * For removals, this function never fails. If addition fails, this
2671 * function doesn't remove files already added. The caller is responsible
2674 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2680 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
2681 lockdep_assert_held(&cgroup_tree_mutex);
2683 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2684 /* does cft->flags tell us to skip this file on @cgrp? */
2685 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2687 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2689 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2693 ret = cgroup_add_file(cgrp, cft);
2695 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2700 cgroup_rm_file(cgrp, cft);
2706 static void cgroup_cfts_prepare(void)
2707 __acquires(&cgroup_mutex)
2710 * Thanks to the entanglement with vfs inode locking, we can't walk
2711 * the existing cgroups under cgroup_mutex and create files.
2712 * Instead, we use css_for_each_descendant_pre() and drop RCU read
2713 * lock before calling cgroup_addrm_files().
2715 mutex_lock(&cgroup_tree_mutex);
2716 mutex_lock(&cgroup_mutex);
2719 static int cgroup_cfts_commit(struct cftype *cfts, bool is_add)
2720 __releases(&cgroup_mutex)
2723 struct cgroup_subsys *ss = cfts[0].ss;
2724 struct cgroup *root = &ss->root->top_cgroup;
2725 struct super_block *sb = ss->root->sb;
2726 struct dentry *prev = NULL;
2727 struct inode *inode;
2728 struct cgroup_subsys_state *css;
2732 mutex_unlock(&cgroup_mutex);
2734 /* %NULL @cfts indicates abort and don't bother if @ss isn't attached */
2735 if (!cfts || ss->root == &cgroup_dummy_root ||
2736 !atomic_inc_not_zero(&sb->s_active)) {
2737 mutex_unlock(&cgroup_tree_mutex);
2742 * All cgroups which are created after we drop cgroup_mutex will
2743 * have the updated set of files, so we only need to update the
2744 * cgroups created before the current @cgroup_serial_nr_next.
2746 update_before = cgroup_serial_nr_next;
2748 /* add/rm files for all cgroups created before */
2749 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2750 struct cgroup *cgrp = css->cgroup;
2752 if (cgroup_is_dead(cgrp))
2755 inode = cgrp->dentry->d_inode;
2758 prev = cgrp->dentry;
2760 mutex_unlock(&cgroup_tree_mutex);
2761 mutex_lock(&inode->i_mutex);
2762 mutex_lock(&cgroup_tree_mutex);
2763 if (cgrp->serial_nr < update_before && !cgroup_is_dead(cgrp))
2764 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2765 mutex_unlock(&inode->i_mutex);
2769 mutex_unlock(&cgroup_tree_mutex);
2771 deactivate_super(sb);
2776 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2777 * @ss: target cgroup subsystem
2778 * @cfts: zero-length name terminated array of cftypes
2780 * Register @cfts to @ss. Files described by @cfts are created for all
2781 * existing cgroups to which @ss is attached and all future cgroups will
2782 * have them too. This function can be called anytime whether @ss is
2785 * Returns 0 on successful registration, -errno on failure. Note that this
2786 * function currently returns 0 as long as @cfts registration is successful
2787 * even if some file creation attempts on existing cgroups fail.
2789 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2791 struct cftype_set *set;
2795 set = kzalloc(sizeof(*set), GFP_KERNEL);
2799 for (cft = cfts; cft->name[0] != '\0'; cft++)
2802 cgroup_cfts_prepare();
2804 list_add_tail(&set->node, &ss->cftsets);
2805 ret = cgroup_cfts_commit(cfts, true);
2807 cgroup_rm_cftypes(cfts);
2810 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2813 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2814 * @cfts: zero-length name terminated array of cftypes
2816 * Unregister @cfts. Files described by @cfts are removed from all
2817 * existing cgroups and all future cgroups won't have them either. This
2818 * function can be called anytime whether @cfts' subsys is attached or not.
2820 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2823 int cgroup_rm_cftypes(struct cftype *cfts)
2825 struct cftype_set *set;
2827 if (!cfts || !cfts[0].ss)
2830 cgroup_cfts_prepare();
2832 list_for_each_entry(set, &cfts[0].ss->cftsets, node) {
2833 if (set->cfts == cfts) {
2834 list_del(&set->node);
2836 cgroup_cfts_commit(cfts, false);
2841 cgroup_cfts_commit(NULL, false);
2846 * cgroup_task_count - count the number of tasks in a cgroup.
2847 * @cgrp: the cgroup in question
2849 * Return the number of tasks in the cgroup.
2851 int cgroup_task_count(const struct cgroup *cgrp)
2854 struct cgrp_cset_link *link;
2856 read_lock(&css_set_lock);
2857 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2858 count += atomic_read(&link->cset->refcount);
2859 read_unlock(&css_set_lock);
2864 * To reduce the fork() overhead for systems that are not actually using
2865 * their cgroups capability, we don't maintain the lists running through
2866 * each css_set to its tasks until we see the list actually used - in other
2867 * words after the first call to css_task_iter_start().
2869 static void cgroup_enable_task_cg_lists(void)
2871 struct task_struct *p, *g;
2872 write_lock(&css_set_lock);
2873 use_task_css_set_links = 1;
2875 * We need tasklist_lock because RCU is not safe against
2876 * while_each_thread(). Besides, a forking task that has passed
2877 * cgroup_post_fork() without seeing use_task_css_set_links = 1
2878 * is not guaranteed to have its child immediately visible in the
2879 * tasklist if we walk through it with RCU.
2881 read_lock(&tasklist_lock);
2882 do_each_thread(g, p) {
2885 * We should check if the process is exiting, otherwise
2886 * it will race with cgroup_exit() in that the list
2887 * entry won't be deleted though the process has exited.
2889 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2890 list_add(&p->cg_list, &task_css_set(p)->tasks);
2892 } while_each_thread(g, p);
2893 read_unlock(&tasklist_lock);
2894 write_unlock(&css_set_lock);
2898 * css_next_child - find the next child of a given css
2899 * @pos_css: the current position (%NULL to initiate traversal)
2900 * @parent_css: css whose children to walk
2902 * This function returns the next child of @parent_css and should be called
2903 * under either cgroup_mutex or RCU read lock. The only requirement is
2904 * that @parent_css and @pos_css are accessible. The next sibling is
2905 * guaranteed to be returned regardless of their states.
2907 struct cgroup_subsys_state *
2908 css_next_child(struct cgroup_subsys_state *pos_css,
2909 struct cgroup_subsys_state *parent_css)
2911 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2912 struct cgroup *cgrp = parent_css->cgroup;
2913 struct cgroup *next;
2915 cgroup_assert_mutexes_or_rcu_locked();
2918 * @pos could already have been removed. Once a cgroup is removed,
2919 * its ->sibling.next is no longer updated when its next sibling
2920 * changes. As CGRP_DEAD assertion is serialized and happens
2921 * before the cgroup is taken off the ->sibling list, if we see it
2922 * unasserted, it's guaranteed that the next sibling hasn't
2923 * finished its grace period even if it's already removed, and thus
2924 * safe to dereference from this RCU critical section. If
2925 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2926 * to be visible as %true here.
2928 * If @pos is dead, its next pointer can't be dereferenced;
2929 * however, as each cgroup is given a monotonically increasing
2930 * unique serial number and always appended to the sibling list,
2931 * the next one can be found by walking the parent's children until
2932 * we see a cgroup with higher serial number than @pos's. While
2933 * this path can be slower, it's taken only when either the current
2934 * cgroup is removed or iteration and removal race.
2937 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2938 } else if (likely(!cgroup_is_dead(pos))) {
2939 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2941 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2942 if (next->serial_nr > pos->serial_nr)
2946 if (&next->sibling == &cgrp->children)
2949 return cgroup_css(next, parent_css->ss);
2951 EXPORT_SYMBOL_GPL(css_next_child);
2954 * css_next_descendant_pre - find the next descendant for pre-order walk
2955 * @pos: the current position (%NULL to initiate traversal)
2956 * @root: css whose descendants to walk
2958 * To be used by css_for_each_descendant_pre(). Find the next descendant
2959 * to visit for pre-order traversal of @root's descendants. @root is
2960 * included in the iteration and the first node to be visited.
2962 * While this function requires cgroup_mutex or RCU read locking, it
2963 * doesn't require the whole traversal to be contained in a single critical
2964 * section. This function will return the correct next descendant as long
2965 * as both @pos and @root are accessible and @pos is a descendant of @root.
2967 struct cgroup_subsys_state *
2968 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2969 struct cgroup_subsys_state *root)
2971 struct cgroup_subsys_state *next;
2973 cgroup_assert_mutexes_or_rcu_locked();
2975 /* if first iteration, visit @root */
2979 /* visit the first child if exists */
2980 next = css_next_child(NULL, pos);
2984 /* no child, visit my or the closest ancestor's next sibling */
2985 while (pos != root) {
2986 next = css_next_child(pos, css_parent(pos));
2989 pos = css_parent(pos);
2994 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
2997 * css_rightmost_descendant - return the rightmost descendant of a css
2998 * @pos: css of interest
3000 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3001 * is returned. This can be used during pre-order traversal to skip
3004 * While this function requires cgroup_mutex or RCU read locking, it
3005 * doesn't require the whole traversal to be contained in a single critical
3006 * section. This function will return the correct rightmost descendant as
3007 * long as @pos is accessible.
3009 struct cgroup_subsys_state *
3010 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3012 struct cgroup_subsys_state *last, *tmp;
3014 cgroup_assert_mutexes_or_rcu_locked();
3018 /* ->prev isn't RCU safe, walk ->next till the end */
3020 css_for_each_child(tmp, last)
3026 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
3028 static struct cgroup_subsys_state *
3029 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3031 struct cgroup_subsys_state *last;
3035 pos = css_next_child(NULL, pos);
3042 * css_next_descendant_post - find the next descendant for post-order walk
3043 * @pos: the current position (%NULL to initiate traversal)
3044 * @root: css whose descendants to walk
3046 * To be used by css_for_each_descendant_post(). Find the next descendant
3047 * to visit for post-order traversal of @root's descendants. @root is
3048 * included in the iteration and the last node to be visited.
3050 * While this function requires cgroup_mutex or RCU read locking, it
3051 * doesn't require the whole traversal to be contained in a single critical
3052 * section. This function will return the correct next descendant as long
3053 * as both @pos and @cgroup are accessible and @pos is a descendant of
3056 struct cgroup_subsys_state *
3057 css_next_descendant_post(struct cgroup_subsys_state *pos,
3058 struct cgroup_subsys_state *root)
3060 struct cgroup_subsys_state *next;
3062 cgroup_assert_mutexes_or_rcu_locked();
3064 /* if first iteration, visit leftmost descendant which may be @root */
3066 return css_leftmost_descendant(root);
3068 /* if we visited @root, we're done */
3072 /* if there's an unvisited sibling, visit its leftmost descendant */
3073 next = css_next_child(pos, css_parent(pos));
3075 return css_leftmost_descendant(next);
3077 /* no sibling left, visit parent */
3078 return css_parent(pos);
3080 EXPORT_SYMBOL_GPL(css_next_descendant_post);
3083 * css_advance_task_iter - advance a task itererator to the next css_set
3084 * @it: the iterator to advance
3086 * Advance @it to the next css_set to walk.
3088 static void css_advance_task_iter(struct css_task_iter *it)
3090 struct list_head *l = it->cset_link;
3091 struct cgrp_cset_link *link;
3092 struct css_set *cset;
3094 /* Advance to the next non-empty css_set */
3097 if (l == &it->origin_css->cgroup->cset_links) {
3098 it->cset_link = NULL;
3101 link = list_entry(l, struct cgrp_cset_link, cset_link);
3103 } while (list_empty(&cset->tasks));
3105 it->task = cset->tasks.next;
3109 * css_task_iter_start - initiate task iteration
3110 * @css: the css to walk tasks of
3111 * @it: the task iterator to use
3113 * Initiate iteration through the tasks of @css. The caller can call
3114 * css_task_iter_next() to walk through the tasks until the function
3115 * returns NULL. On completion of iteration, css_task_iter_end() must be
3118 * Note that this function acquires a lock which is released when the
3119 * iteration finishes. The caller can't sleep while iteration is in
3122 void css_task_iter_start(struct cgroup_subsys_state *css,
3123 struct css_task_iter *it)
3124 __acquires(css_set_lock)
3127 * The first time anyone tries to iterate across a css, we need to
3128 * enable the list linking each css_set to its tasks, and fix up
3129 * all existing tasks.
3131 if (!use_task_css_set_links)
3132 cgroup_enable_task_cg_lists();
3134 read_lock(&css_set_lock);
3136 it->origin_css = css;
3137 it->cset_link = &css->cgroup->cset_links;
3139 css_advance_task_iter(it);
3143 * css_task_iter_next - return the next task for the iterator
3144 * @it: the task iterator being iterated
3146 * The "next" function for task iteration. @it should have been
3147 * initialized via css_task_iter_start(). Returns NULL when the iteration
3150 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3152 struct task_struct *res;
3153 struct list_head *l = it->task;
3154 struct cgrp_cset_link *link;
3156 /* If the iterator cg is NULL, we have no tasks */
3159 res = list_entry(l, struct task_struct, cg_list);
3160 /* Advance iterator to find next entry */
3162 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
3163 if (l == &link->cset->tasks) {
3165 * We reached the end of this task list - move on to the
3166 * next cgrp_cset_link.
3168 css_advance_task_iter(it);
3176 * css_task_iter_end - finish task iteration
3177 * @it: the task iterator to finish
3179 * Finish task iteration started by css_task_iter_start().
3181 void css_task_iter_end(struct css_task_iter *it)
3182 __releases(css_set_lock)
3184 read_unlock(&css_set_lock);
3187 static inline int started_after_time(struct task_struct *t1,
3188 struct timespec *time,
3189 struct task_struct *t2)
3191 int start_diff = timespec_compare(&t1->start_time, time);
3192 if (start_diff > 0) {
3194 } else if (start_diff < 0) {
3198 * Arbitrarily, if two processes started at the same
3199 * time, we'll say that the lower pointer value
3200 * started first. Note that t2 may have exited by now
3201 * so this may not be a valid pointer any longer, but
3202 * that's fine - it still serves to distinguish
3203 * between two tasks started (effectively) simultaneously.
3210 * This function is a callback from heap_insert() and is used to order
3212 * In this case we order the heap in descending task start time.
3214 static inline int started_after(void *p1, void *p2)
3216 struct task_struct *t1 = p1;
3217 struct task_struct *t2 = p2;
3218 return started_after_time(t1, &t2->start_time, t2);
3222 * css_scan_tasks - iterate though all the tasks in a css
3223 * @css: the css to iterate tasks of
3224 * @test: optional test callback
3225 * @process: process callback
3226 * @data: data passed to @test and @process
3227 * @heap: optional pre-allocated heap used for task iteration
3229 * Iterate through all the tasks in @css, calling @test for each, and if it
3230 * returns %true, call @process for it also.
3232 * @test may be NULL, meaning always true (select all tasks), which
3233 * effectively duplicates css_task_iter_{start,next,end}() but does not
3234 * lock css_set_lock for the call to @process.
3236 * It is guaranteed that @process will act on every task that is a member
3237 * of @css for the duration of this call. This function may or may not
3238 * call @process for tasks that exit or move to a different css during the
3239 * call, or are forked or move into the css during the call.
3241 * Note that @test may be called with locks held, and may in some
3242 * situations be called multiple times for the same task, so it should be
3245 * If @heap is non-NULL, a heap has been pre-allocated and will be used for
3246 * heap operations (and its "gt" member will be overwritten), else a
3247 * temporary heap will be used (allocation of which may cause this function
3250 int css_scan_tasks(struct cgroup_subsys_state *css,
3251 bool (*test)(struct task_struct *, void *),
3252 void (*process)(struct task_struct *, void *),
3253 void *data, struct ptr_heap *heap)
3256 struct css_task_iter it;
3257 struct task_struct *p, *dropped;
3258 /* Never dereference latest_task, since it's not refcounted */
3259 struct task_struct *latest_task = NULL;
3260 struct ptr_heap tmp_heap;
3261 struct timespec latest_time = { 0, 0 };
3264 /* The caller supplied our heap and pre-allocated its memory */
3265 heap->gt = &started_after;
3267 /* We need to allocate our own heap memory */
3269 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
3271 /* cannot allocate the heap */
3277 * Scan tasks in the css, using the @test callback to determine
3278 * which are of interest, and invoking @process callback on the
3279 * ones which need an update. Since we don't want to hold any
3280 * locks during the task updates, gather tasks to be processed in a
3281 * heap structure. The heap is sorted by descending task start
3282 * time. If the statically-sized heap fills up, we overflow tasks
3283 * that started later, and in future iterations only consider tasks
3284 * that started after the latest task in the previous pass. This
3285 * guarantees forward progress and that we don't miss any tasks.
3288 css_task_iter_start(css, &it);
3289 while ((p = css_task_iter_next(&it))) {
3291 * Only affect tasks that qualify per the caller's callback,
3292 * if he provided one
3294 if (test && !test(p, data))
3297 * Only process tasks that started after the last task
3300 if (!started_after_time(p, &latest_time, latest_task))
3302 dropped = heap_insert(heap, p);
3303 if (dropped == NULL) {
3305 * The new task was inserted; the heap wasn't
3309 } else if (dropped != p) {
3311 * The new task was inserted, and pushed out a
3315 put_task_struct(dropped);
3318 * Else the new task was newer than anything already in
3319 * the heap and wasn't inserted
3322 css_task_iter_end(&it);
3325 for (i = 0; i < heap->size; i++) {
3326 struct task_struct *q = heap->ptrs[i];
3328 latest_time = q->start_time;
3331 /* Process the task per the caller's callback */
3336 * If we had to process any tasks at all, scan again
3337 * in case some of them were in the middle of forking
3338 * children that didn't get processed.
3339 * Not the most efficient way to do it, but it avoids
3340 * having to take callback_mutex in the fork path
3344 if (heap == &tmp_heap)
3345 heap_free(&tmp_heap);
3349 static void cgroup_transfer_one_task(struct task_struct *task, void *data)
3351 struct cgroup *new_cgroup = data;
3353 mutex_lock(&cgroup_mutex);
3354 cgroup_attach_task(new_cgroup, task, false);
3355 mutex_unlock(&cgroup_mutex);
3359 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3360 * @to: cgroup to which the tasks will be moved
3361 * @from: cgroup in which the tasks currently reside
3363 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3365 return css_scan_tasks(&from->dummy_css, NULL, cgroup_transfer_one_task,
3370 * Stuff for reading the 'tasks'/'procs' files.
3372 * Reading this file can return large amounts of data if a cgroup has
3373 * *lots* of attached tasks. So it may need several calls to read(),
3374 * but we cannot guarantee that the information we produce is correct
3375 * unless we produce it entirely atomically.
3379 /* which pidlist file are we talking about? */
3380 enum cgroup_filetype {
3386 * A pidlist is a list of pids that virtually represents the contents of one
3387 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3388 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3391 struct cgroup_pidlist {
3393 * used to find which pidlist is wanted. doesn't change as long as
3394 * this particular list stays in the list.
3396 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3399 /* how many elements the above list has */
3401 /* each of these stored in a list by its cgroup */
3402 struct list_head links;
3403 /* pointer to the cgroup we belong to, for list removal purposes */
3404 struct cgroup *owner;
3405 /* for delayed destruction */
3406 struct delayed_work destroy_dwork;
3410 * The following two functions "fix" the issue where there are more pids
3411 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3412 * TODO: replace with a kernel-wide solution to this problem
3414 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3415 static void *pidlist_allocate(int count)
3417 if (PIDLIST_TOO_LARGE(count))
3418 return vmalloc(count * sizeof(pid_t));
3420 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3423 static void pidlist_free(void *p)
3425 if (is_vmalloc_addr(p))
3432 * Used to destroy all pidlists lingering waiting for destroy timer. None
3433 * should be left afterwards.
3435 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3437 struct cgroup_pidlist *l, *tmp_l;
3439 mutex_lock(&cgrp->pidlist_mutex);
3440 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3441 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3442 mutex_unlock(&cgrp->pidlist_mutex);
3444 flush_workqueue(cgroup_pidlist_destroy_wq);
3445 BUG_ON(!list_empty(&cgrp->pidlists));
3448 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3450 struct delayed_work *dwork = to_delayed_work(work);
3451 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3453 struct cgroup_pidlist *tofree = NULL;
3455 mutex_lock(&l->owner->pidlist_mutex);
3458 * Destroy iff we didn't get queued again. The state won't change
3459 * as destroy_dwork can only be queued while locked.
3461 if (!delayed_work_pending(dwork)) {
3462 list_del(&l->links);
3463 pidlist_free(l->list);
3464 put_pid_ns(l->key.ns);
3468 mutex_unlock(&l->owner->pidlist_mutex);
3473 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3474 * Returns the number of unique elements.
3476 static int pidlist_uniq(pid_t *list, int length)
3481 * we presume the 0th element is unique, so i starts at 1. trivial
3482 * edge cases first; no work needs to be done for either
3484 if (length == 0 || length == 1)
3486 /* src and dest walk down the list; dest counts unique elements */
3487 for (src = 1; src < length; src++) {
3488 /* find next unique element */
3489 while (list[src] == list[src-1]) {
3494 /* dest always points to where the next unique element goes */
3495 list[dest] = list[src];
3503 * The two pid files - task and cgroup.procs - guaranteed that the result
3504 * is sorted, which forced this whole pidlist fiasco. As pid order is
3505 * different per namespace, each namespace needs differently sorted list,
3506 * making it impossible to use, for example, single rbtree of member tasks
3507 * sorted by task pointer. As pidlists can be fairly large, allocating one
3508 * per open file is dangerous, so cgroup had to implement shared pool of
3509 * pidlists keyed by cgroup and namespace.
3511 * All this extra complexity was caused by the original implementation
3512 * committing to an entirely unnecessary property. In the long term, we
3513 * want to do away with it. Explicitly scramble sort order if
3514 * sane_behavior so that no such expectation exists in the new interface.
3516 * Scrambling is done by swapping every two consecutive bits, which is
3517 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3519 static pid_t pid_fry(pid_t pid)
3521 unsigned a = pid & 0x55555555;
3522 unsigned b = pid & 0xAAAAAAAA;
3524 return (a << 1) | (b >> 1);
3527 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3529 if (cgroup_sane_behavior(cgrp))
3530 return pid_fry(pid);
3535 static int cmppid(const void *a, const void *b)
3537 return *(pid_t *)a - *(pid_t *)b;
3540 static int fried_cmppid(const void *a, const void *b)
3542 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3545 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3546 enum cgroup_filetype type)
3548 struct cgroup_pidlist *l;
3549 /* don't need task_nsproxy() if we're looking at ourself */
3550 struct pid_namespace *ns = task_active_pid_ns(current);
3552 lockdep_assert_held(&cgrp->pidlist_mutex);
3554 list_for_each_entry(l, &cgrp->pidlists, links)
3555 if (l->key.type == type && l->key.ns == ns)
3561 * find the appropriate pidlist for our purpose (given procs vs tasks)
3562 * returns with the lock on that pidlist already held, and takes care
3563 * of the use count, or returns NULL with no locks held if we're out of
3566 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3567 enum cgroup_filetype type)
3569 struct cgroup_pidlist *l;
3571 lockdep_assert_held(&cgrp->pidlist_mutex);
3573 l = cgroup_pidlist_find(cgrp, type);
3577 /* entry not found; create a new one */
3578 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3582 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3584 /* don't need task_nsproxy() if we're looking at ourself */
3585 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3587 list_add(&l->links, &cgrp->pidlists);
3592 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3594 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3595 struct cgroup_pidlist **lp)
3599 int pid, n = 0; /* used for populating the array */
3600 struct css_task_iter it;
3601 struct task_struct *tsk;
3602 struct cgroup_pidlist *l;
3604 lockdep_assert_held(&cgrp->pidlist_mutex);
3607 * If cgroup gets more users after we read count, we won't have
3608 * enough space - tough. This race is indistinguishable to the
3609 * caller from the case that the additional cgroup users didn't
3610 * show up until sometime later on.
3612 length = cgroup_task_count(cgrp);
3613 array = pidlist_allocate(length);
3616 /* now, populate the array */
3617 css_task_iter_start(&cgrp->dummy_css, &it);
3618 while ((tsk = css_task_iter_next(&it))) {
3619 if (unlikely(n == length))
3621 /* get tgid or pid for procs or tasks file respectively */
3622 if (type == CGROUP_FILE_PROCS)
3623 pid = task_tgid_vnr(tsk);
3625 pid = task_pid_vnr(tsk);
3626 if (pid > 0) /* make sure to only use valid results */
3629 css_task_iter_end(&it);
3631 /* now sort & (if procs) strip out duplicates */
3632 if (cgroup_sane_behavior(cgrp))
3633 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3635 sort(array, length, sizeof(pid_t), cmppid, NULL);
3636 if (type == CGROUP_FILE_PROCS)
3637 length = pidlist_uniq(array, length);
3639 l = cgroup_pidlist_find_create(cgrp, type);
3641 mutex_unlock(&cgrp->pidlist_mutex);
3642 pidlist_free(array);
3646 /* store array, freeing old if necessary */
3647 pidlist_free(l->list);
3655 * cgroupstats_build - build and fill cgroupstats
3656 * @stats: cgroupstats to fill information into
3657 * @dentry: A dentry entry belonging to the cgroup for which stats have
3660 * Build and fill cgroupstats so that taskstats can export it to user
3663 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3666 struct cgroup *cgrp;
3667 struct css_task_iter it;
3668 struct task_struct *tsk;
3671 * Validate dentry by checking the superblock operations,
3672 * and make sure it's a directory.
3674 if (dentry->d_sb->s_op != &cgroup_ops ||
3675 !S_ISDIR(dentry->d_inode->i_mode))
3679 cgrp = dentry->d_fsdata;
3681 css_task_iter_start(&cgrp->dummy_css, &it);
3682 while ((tsk = css_task_iter_next(&it))) {
3683 switch (tsk->state) {
3685 stats->nr_running++;
3687 case TASK_INTERRUPTIBLE:
3688 stats->nr_sleeping++;
3690 case TASK_UNINTERRUPTIBLE:
3691 stats->nr_uninterruptible++;
3694 stats->nr_stopped++;
3697 if (delayacct_is_task_waiting_on_io(tsk))
3698 stats->nr_io_wait++;
3702 css_task_iter_end(&it);
3710 * seq_file methods for the tasks/procs files. The seq_file position is the
3711 * next pid to display; the seq_file iterator is a pointer to the pid
3712 * in the cgroup->l->list array.
3715 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3718 * Initially we receive a position value that corresponds to
3719 * one more than the last pid shown (or 0 on the first call or
3720 * after a seek to the start). Use a binary-search to find the
3721 * next pid to display, if any
3723 struct cgroup_open_file *of = s->private;
3724 struct cgroup *cgrp = seq_css(s)->cgroup;
3725 struct cgroup_pidlist *l;
3726 enum cgroup_filetype type = seq_cft(s)->private;
3727 int index = 0, pid = *pos;
3730 mutex_lock(&cgrp->pidlist_mutex);
3733 * !NULL @of->priv indicates that this isn't the first start()
3734 * after open. If the matching pidlist is around, we can use that.
3735 * Look for it. Note that @of->priv can't be used directly. It
3736 * could already have been destroyed.
3739 of->priv = cgroup_pidlist_find(cgrp, type);
3742 * Either this is the first start() after open or the matching
3743 * pidlist has been destroyed inbetween. Create a new one.
3746 ret = pidlist_array_load(cgrp, type,
3747 (struct cgroup_pidlist **)&of->priv);
3749 return ERR_PTR(ret);
3754 int end = l->length;
3756 while (index < end) {
3757 int mid = (index + end) / 2;
3758 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3761 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3767 /* If we're off the end of the array, we're done */
3768 if (index >= l->length)
3770 /* Update the abstract position to be the actual pid that we found */
3771 iter = l->list + index;
3772 *pos = cgroup_pid_fry(cgrp, *iter);
3776 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3778 struct cgroup_open_file *of = s->private;
3779 struct cgroup_pidlist *l = of->priv;
3782 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3783 CGROUP_PIDLIST_DESTROY_DELAY);
3784 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3787 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3789 struct cgroup_open_file *of = s->private;
3790 struct cgroup_pidlist *l = of->priv;
3792 pid_t *end = l->list + l->length;
3794 * Advance to the next pid in the array. If this goes off the
3801 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3806 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3808 return seq_printf(s, "%d\n", *(int *)v);
3812 * seq_operations functions for iterating on pidlists through seq_file -
3813 * independent of whether it's tasks or procs
3815 static const struct seq_operations cgroup_pidlist_seq_operations = {
3816 .start = cgroup_pidlist_start,
3817 .stop = cgroup_pidlist_stop,
3818 .next = cgroup_pidlist_next,
3819 .show = cgroup_pidlist_show,
3822 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3825 return notify_on_release(css->cgroup);
3828 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3829 struct cftype *cft, u64 val)
3831 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3833 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3835 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3840 * When dput() is called asynchronously, if umount has been done and
3841 * then deactivate_super() in cgroup_free_fn() kills the superblock,
3842 * there's a small window that vfs will see the root dentry with non-zero
3843 * refcnt and trigger BUG().
3845 * That's why we hold a reference before dput() and drop it right after.
3847 static void cgroup_dput(struct cgroup *cgrp)
3849 struct super_block *sb = cgrp->root->sb;
3851 atomic_inc(&sb->s_active);
3853 deactivate_super(sb);
3856 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3859 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3862 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3863 struct cftype *cft, u64 val)
3866 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3868 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3872 static struct cftype cgroup_base_files[] = {
3874 .name = "cgroup.procs",
3875 .seq_start = cgroup_pidlist_start,
3876 .seq_next = cgroup_pidlist_next,
3877 .seq_stop = cgroup_pidlist_stop,
3878 .seq_show = cgroup_pidlist_show,
3879 .private = CGROUP_FILE_PROCS,
3880 .write_u64 = cgroup_procs_write,
3881 .mode = S_IRUGO | S_IWUSR,
3884 .name = "cgroup.clone_children",
3885 .flags = CFTYPE_INSANE,
3886 .read_u64 = cgroup_clone_children_read,
3887 .write_u64 = cgroup_clone_children_write,
3890 .name = "cgroup.sane_behavior",
3891 .flags = CFTYPE_ONLY_ON_ROOT,
3892 .seq_show = cgroup_sane_behavior_show,
3896 * Historical crazy stuff. These don't have "cgroup." prefix and
3897 * don't exist if sane_behavior. If you're depending on these, be
3898 * prepared to be burned.
3902 .flags = CFTYPE_INSANE, /* use "procs" instead */
3903 .seq_start = cgroup_pidlist_start,
3904 .seq_next = cgroup_pidlist_next,
3905 .seq_stop = cgroup_pidlist_stop,
3906 .seq_show = cgroup_pidlist_show,
3907 .private = CGROUP_FILE_TASKS,
3908 .write_u64 = cgroup_tasks_write,
3909 .mode = S_IRUGO | S_IWUSR,
3912 .name = "notify_on_release",
3913 .flags = CFTYPE_INSANE,
3914 .read_u64 = cgroup_read_notify_on_release,
3915 .write_u64 = cgroup_write_notify_on_release,
3918 .name = "release_agent",
3919 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3920 .seq_show = cgroup_release_agent_show,
3921 .write_string = cgroup_release_agent_write,
3922 .max_write_len = PATH_MAX,
3928 * cgroup_populate_dir - create subsys files in a cgroup directory
3929 * @cgrp: target cgroup
3930 * @subsys_mask: mask of the subsystem ids whose files should be added
3932 * On failure, no file is added.
3934 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3936 struct cgroup_subsys *ss;
3939 /* process cftsets of each subsystem */
3940 for_each_subsys(ss, i) {
3941 struct cftype_set *set;
3943 if (!test_bit(i, &subsys_mask))
3946 list_for_each_entry(set, &ss->cftsets, node) {
3947 ret = cgroup_addrm_files(cgrp, set->cfts, true);
3954 cgroup_clear_dir(cgrp, subsys_mask);
3959 * css destruction is four-stage process.
3961 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3962 * Implemented in kill_css().
3964 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3965 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3966 * by invoking offline_css(). After offlining, the base ref is put.
3967 * Implemented in css_killed_work_fn().
3969 * 3. When the percpu_ref reaches zero, the only possible remaining
3970 * accessors are inside RCU read sections. css_release() schedules the
3973 * 4. After the grace period, the css can be freed. Implemented in
3974 * css_free_work_fn().
3976 * It is actually hairier because both step 2 and 4 require process context
3977 * and thus involve punting to css->destroy_work adding two additional
3978 * steps to the already complex sequence.
3980 static void css_free_work_fn(struct work_struct *work)
3982 struct cgroup_subsys_state *css =
3983 container_of(work, struct cgroup_subsys_state, destroy_work);
3984 struct cgroup *cgrp = css->cgroup;
3987 css_put(css->parent);
3989 css->ss->css_free(css);
3993 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3995 struct cgroup_subsys_state *css =
3996 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3999 * css holds an extra ref to @cgrp->dentry which is put on the last
4000 * css_put(). dput() requires process context which we don't have.
4002 INIT_WORK(&css->destroy_work, css_free_work_fn);
4003 queue_work(cgroup_destroy_wq, &css->destroy_work);
4006 static void css_release(struct percpu_ref *ref)
4008 struct cgroup_subsys_state *css =
4009 container_of(ref, struct cgroup_subsys_state, refcnt);
4011 rcu_assign_pointer(css->cgroup->subsys[css->ss->id], NULL);
4012 call_rcu(&css->rcu_head, css_free_rcu_fn);
4015 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
4016 struct cgroup *cgrp)
4023 css->parent = cgroup_css(cgrp->parent, ss);
4025 css->flags |= CSS_ROOT;
4027 BUG_ON(cgroup_css(cgrp, ss));
4030 /* invoke ->css_online() on a new CSS and mark it online if successful */
4031 static int online_css(struct cgroup_subsys_state *css)
4033 struct cgroup_subsys *ss = css->ss;
4036 lockdep_assert_held(&cgroup_tree_mutex);
4037 lockdep_assert_held(&cgroup_mutex);
4040 ret = ss->css_online(css);
4042 css->flags |= CSS_ONLINE;
4043 css->cgroup->nr_css++;
4044 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4049 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4050 static void offline_css(struct cgroup_subsys_state *css)
4052 struct cgroup_subsys *ss = css->ss;
4054 lockdep_assert_held(&cgroup_tree_mutex);
4055 lockdep_assert_held(&cgroup_mutex);
4057 if (!(css->flags & CSS_ONLINE))
4060 if (ss->css_offline)
4061 ss->css_offline(css);
4063 css->flags &= ~CSS_ONLINE;
4064 css->cgroup->nr_css--;
4065 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
4069 * create_css - create a cgroup_subsys_state
4070 * @cgrp: the cgroup new css will be associated with
4071 * @ss: the subsys of new css
4073 * Create a new css associated with @cgrp - @ss pair. On success, the new
4074 * css is online and installed in @cgrp with all interface files created.
4075 * Returns 0 on success, -errno on failure.
4077 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
4079 struct cgroup *parent = cgrp->parent;
4080 struct cgroup_subsys_state *css;
4083 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
4084 lockdep_assert_held(&cgroup_mutex);
4086 css = ss->css_alloc(cgroup_css(parent, ss));
4088 return PTR_ERR(css);
4090 err = percpu_ref_init(&css->refcnt, css_release);
4094 init_css(css, ss, cgrp);
4096 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4100 err = online_css(css);
4105 css_get(css->parent);
4107 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4109 pr_warning("cgroup: %s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4110 current->comm, current->pid, ss->name);
4111 if (!strcmp(ss->name, "memory"))
4112 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
4113 ss->warned_broken_hierarchy = true;
4119 percpu_ref_cancel_init(&css->refcnt);
4125 * cgroup_create - create a cgroup
4126 * @parent: cgroup that will be parent of the new cgroup
4127 * @dentry: dentry of the new cgroup
4128 * @mode: mode to set on new inode
4130 * Must be called with the mutex on the parent inode held
4132 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
4135 struct cgroup *cgrp;
4136 struct cgroup_name *name;
4137 struct cgroupfs_root *root = parent->root;
4139 struct cgroup_subsys *ss;
4140 struct super_block *sb = root->sb;
4142 /* allocate the cgroup and its ID, 0 is reserved for the root */
4143 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4147 name = cgroup_alloc_name(dentry->d_name.name);
4152 rcu_assign_pointer(cgrp->name, name);
4154 mutex_lock(&cgroup_tree_mutex);
4157 * Only live parents can have children. Note that the liveliness
4158 * check isn't strictly necessary because cgroup_mkdir() and
4159 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
4160 * anyway so that locking is contained inside cgroup proper and we
4161 * don't get nasty surprises if we ever grow another caller.
4163 if (!cgroup_lock_live_group(parent)) {
4165 goto err_unlock_tree;
4169 * Temporarily set the pointer to NULL, so idr_find() won't return
4170 * a half-baked cgroup.
4172 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
4178 /* Grab a reference on the superblock so the hierarchy doesn't
4179 * get deleted on unmount if there are child cgroups. This
4180 * can be done outside cgroup_mutex, since the sb can't
4181 * disappear while someone has an open control file on the
4183 atomic_inc(&sb->s_active);
4185 init_cgroup_housekeeping(cgrp);
4187 dentry->d_fsdata = cgrp;
4188 cgrp->dentry = dentry;
4190 cgrp->parent = parent;
4191 cgrp->dummy_css.parent = &parent->dummy_css;
4192 cgrp->root = parent->root;
4194 if (notify_on_release(parent))
4195 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4197 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4198 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4201 * Create directory. cgroup_create_file() returns with the new
4202 * directory locked on success so that it can be populated without
4203 * dropping cgroup_mutex.
4205 err = cgroup_create_file(dentry, S_IFDIR | mode, sb);
4208 lockdep_assert_held(&dentry->d_inode->i_mutex);
4210 cgrp->serial_nr = cgroup_serial_nr_next++;
4212 /* allocation complete, commit to creation */
4213 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
4214 root->number_of_cgroups++;
4216 /* hold a ref to the parent's dentry */
4217 dget(parent->dentry);
4220 * @cgrp is now fully operational. If something fails after this
4221 * point, it'll be released via the normal destruction path.
4223 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4225 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
4229 /* let's create and online css's */
4230 for_each_subsys(ss, ssid) {
4231 if (root->subsys_mask & (1 << ssid)) {
4232 err = create_css(cgrp, ss);
4238 mutex_unlock(&cgroup_mutex);
4239 mutex_unlock(&cgroup_tree_mutex);
4240 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
4245 idr_remove(&root->cgroup_idr, cgrp->id);
4246 /* Release the reference count that we took on the superblock */
4247 deactivate_super(sb);
4249 mutex_unlock(&cgroup_mutex);
4251 mutex_unlock(&cgroup_tree_mutex);
4252 kfree(rcu_dereference_raw(cgrp->name));
4258 cgroup_destroy_locked(cgrp);
4259 mutex_unlock(&cgroup_mutex);
4260 mutex_unlock(&cgroup_tree_mutex);
4261 mutex_unlock(&dentry->d_inode->i_mutex);
4265 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
4267 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
4269 /* the vfs holds inode->i_mutex already */
4270 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
4274 * This is called when the refcnt of a css is confirmed to be killed.
4275 * css_tryget() is now guaranteed to fail.
4277 static void css_killed_work_fn(struct work_struct *work)
4279 struct cgroup_subsys_state *css =
4280 container_of(work, struct cgroup_subsys_state, destroy_work);
4281 struct cgroup *cgrp = css->cgroup;
4283 mutex_lock(&cgroup_tree_mutex);
4284 mutex_lock(&cgroup_mutex);
4287 * css_tryget() is guaranteed to fail now. Tell subsystems to
4288 * initate destruction.
4293 * If @cgrp is marked dead, it's waiting for refs of all css's to
4294 * be disabled before proceeding to the second phase of cgroup
4295 * destruction. If we are the last one, kick it off.
4297 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
4298 cgroup_destroy_css_killed(cgrp);
4300 mutex_unlock(&cgroup_mutex);
4301 mutex_unlock(&cgroup_tree_mutex);
4304 * Put the css refs from kill_css(). Each css holds an extra
4305 * reference to the cgroup's dentry and cgroup removal proceeds
4306 * regardless of css refs. On the last put of each css, whenever
4307 * that may be, the extra dentry ref is put so that dentry
4308 * destruction happens only after all css's are released.
4313 /* css kill confirmation processing requires process context, bounce */
4314 static void css_killed_ref_fn(struct percpu_ref *ref)
4316 struct cgroup_subsys_state *css =
4317 container_of(ref, struct cgroup_subsys_state, refcnt);
4319 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4320 queue_work(cgroup_destroy_wq, &css->destroy_work);
4324 * kill_css - destroy a css
4325 * @css: css to destroy
4327 * This function initiates destruction of @css by removing cgroup interface
4328 * files and putting its base reference. ->css_offline() will be invoked
4329 * asynchronously once css_tryget() is guaranteed to fail and when the
4330 * reference count reaches zero, @css will be released.
4332 static void kill_css(struct cgroup_subsys_state *css)
4334 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4337 * Killing would put the base ref, but we need to keep it alive
4338 * until after ->css_offline().
4343 * cgroup core guarantees that, by the time ->css_offline() is
4344 * invoked, no new css reference will be given out via
4345 * css_tryget(). We can't simply call percpu_ref_kill() and
4346 * proceed to offlining css's because percpu_ref_kill() doesn't
4347 * guarantee that the ref is seen as killed on all CPUs on return.
4349 * Use percpu_ref_kill_and_confirm() to get notifications as each
4350 * css is confirmed to be seen as killed on all CPUs.
4352 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4356 * cgroup_destroy_locked - the first stage of cgroup destruction
4357 * @cgrp: cgroup to be destroyed
4359 * css's make use of percpu refcnts whose killing latency shouldn't be
4360 * exposed to userland and are RCU protected. Also, cgroup core needs to
4361 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
4362 * invoked. To satisfy all the requirements, destruction is implemented in
4363 * the following two steps.
4365 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4366 * userland visible parts and start killing the percpu refcnts of
4367 * css's. Set up so that the next stage will be kicked off once all
4368 * the percpu refcnts are confirmed to be killed.
4370 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4371 * rest of destruction. Once all cgroup references are gone, the
4372 * cgroup is RCU-freed.
4374 * This function implements s1. After this step, @cgrp is gone as far as
4375 * the userland is concerned and a new cgroup with the same name may be
4376 * created. As cgroup doesn't care about the names internally, this
4377 * doesn't cause any problem.
4379 static int cgroup_destroy_locked(struct cgroup *cgrp)
4380 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4382 struct dentry *d = cgrp->dentry;
4383 struct cgroup_subsys_state *css;
4384 struct cgroup *child;
4388 lockdep_assert_held(&d->d_inode->i_mutex);
4389 lockdep_assert_held(&cgroup_tree_mutex);
4390 lockdep_assert_held(&cgroup_mutex);
4393 * css_set_lock synchronizes access to ->cset_links and prevents
4394 * @cgrp from being removed while __put_css_set() is in progress.
4396 read_lock(&css_set_lock);
4397 empty = list_empty(&cgrp->cset_links);
4398 read_unlock(&css_set_lock);
4403 * Make sure there's no live children. We can't test ->children
4404 * emptiness as dead children linger on it while being destroyed;
4405 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
4409 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
4410 empty = cgroup_is_dead(child);
4419 * Initiate massacre of all css's. cgroup_destroy_css_killed()
4420 * will be invoked to perform the rest of destruction once the
4421 * percpu refs of all css's are confirmed to be killed. This
4422 * involves removing the subsystem's files, drop cgroup_mutex.
4424 mutex_unlock(&cgroup_mutex);
4425 for_each_css(css, ssid, cgrp)
4427 mutex_lock(&cgroup_mutex);
4430 * Mark @cgrp dead. This prevents further task migration and child
4431 * creation by disabling cgroup_lock_live_group(). Note that
4432 * CGRP_DEAD assertion is depended upon by css_next_child() to
4433 * resume iteration after dropping RCU read lock. See
4434 * css_next_child() for details.
4436 set_bit(CGRP_DEAD, &cgrp->flags);
4438 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4439 raw_spin_lock(&release_list_lock);
4440 if (!list_empty(&cgrp->release_list))
4441 list_del_init(&cgrp->release_list);
4442 raw_spin_unlock(&release_list_lock);
4445 * If @cgrp has css's attached, the second stage of cgroup
4446 * destruction is kicked off from css_killed_work_fn() after the
4447 * refs of all attached css's are killed. If @cgrp doesn't have
4448 * any css, we kick it off here.
4451 cgroup_destroy_css_killed(cgrp);
4454 * Clear the base files and remove @cgrp directory. The removal
4455 * puts the base ref but we aren't quite done with @cgrp yet, so
4458 mutex_unlock(&cgroup_mutex);
4459 cgroup_addrm_files(cgrp, cgroup_base_files, false);
4461 cgroup_d_remove_dir(d);
4462 mutex_lock(&cgroup_mutex);
4468 * cgroup_destroy_css_killed - the second step of cgroup destruction
4469 * @work: cgroup->destroy_free_work
4471 * This function is invoked from a work item for a cgroup which is being
4472 * destroyed after all css's are offlined and performs the rest of
4473 * destruction. This is the second step of destruction described in the
4474 * comment above cgroup_destroy_locked().
4476 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
4478 struct cgroup *parent = cgrp->parent;
4479 struct dentry *d = cgrp->dentry;
4481 lockdep_assert_held(&cgroup_tree_mutex);
4482 lockdep_assert_held(&cgroup_mutex);
4484 /* delete this cgroup from parent->children */
4485 list_del_rcu(&cgrp->sibling);
4489 set_bit(CGRP_RELEASABLE, &parent->flags);
4490 check_for_release(parent);
4493 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
4497 mutex_lock(&cgroup_tree_mutex);
4498 mutex_lock(&cgroup_mutex);
4499 ret = cgroup_destroy_locked(dentry->d_fsdata);
4500 mutex_unlock(&cgroup_mutex);
4501 mutex_unlock(&cgroup_tree_mutex);
4506 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4508 struct cgroup_subsys_state *css;
4510 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4512 mutex_lock(&cgroup_tree_mutex);
4513 mutex_lock(&cgroup_mutex);
4515 INIT_LIST_HEAD(&ss->cftsets);
4517 /* Create the top cgroup state for this subsystem */
4518 ss->root = &cgroup_dummy_root;
4519 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4520 /* We don't handle early failures gracefully */
4521 BUG_ON(IS_ERR(css));
4522 init_css(css, ss, cgroup_dummy_top);
4524 /* Update the init_css_set to contain a subsys
4525 * pointer to this state - since the subsystem is
4526 * newly registered, all tasks and hence the
4527 * init_css_set is in the subsystem's top cgroup. */
4528 init_css_set.subsys[ss->id] = css;
4530 need_forkexit_callback |= ss->fork || ss->exit;
4532 /* At system boot, before all subsystems have been
4533 * registered, no tasks have been forked, so we don't
4534 * need to invoke fork callbacks here. */
4535 BUG_ON(!list_empty(&init_task.tasks));
4537 BUG_ON(online_css(css));
4539 mutex_unlock(&cgroup_mutex);
4540 mutex_unlock(&cgroup_tree_mutex);
4544 * cgroup_init_early - cgroup initialization at system boot
4546 * Initialize cgroups at system boot, and initialize any
4547 * subsystems that request early init.
4549 int __init cgroup_init_early(void)
4551 struct cgroup_subsys *ss;
4554 atomic_set(&init_css_set.refcount, 1);
4555 INIT_LIST_HEAD(&init_css_set.cgrp_links);
4556 INIT_LIST_HEAD(&init_css_set.tasks);
4557 INIT_HLIST_NODE(&init_css_set.hlist);
4559 init_cgroup_root(&cgroup_dummy_root);
4560 cgroup_root_count = 1;
4561 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4563 init_cgrp_cset_link.cset = &init_css_set;
4564 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4565 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4566 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4568 for_each_subsys(ss, i) {
4569 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4570 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4571 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4573 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4574 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4577 ss->name = cgroup_subsys_name[i];
4580 cgroup_init_subsys(ss);
4586 * cgroup_init - cgroup initialization
4588 * Register cgroup filesystem and /proc file, and initialize
4589 * any subsystems that didn't request early init.
4591 int __init cgroup_init(void)
4593 struct cgroup_subsys *ss;
4597 err = bdi_init(&cgroup_backing_dev_info);
4601 for_each_subsys(ss, i) {
4602 if (!ss->early_init)
4603 cgroup_init_subsys(ss);
4606 * cftype registration needs kmalloc and can't be done
4607 * during early_init. Register base cftypes separately.
4609 if (ss->base_cftypes)
4610 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4613 /* allocate id for the dummy hierarchy */
4614 mutex_lock(&cgroup_mutex);
4616 /* Add init_css_set to the hash table */
4617 key = css_set_hash(init_css_set.subsys);
4618 hash_add(css_set_table, &init_css_set.hlist, key);
4620 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4622 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
4626 mutex_unlock(&cgroup_mutex);
4628 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4634 err = register_filesystem(&cgroup_fs_type);
4636 kobject_put(cgroup_kobj);
4640 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4644 bdi_destroy(&cgroup_backing_dev_info);
4649 static int __init cgroup_wq_init(void)
4652 * There isn't much point in executing destruction path in
4653 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4655 * XXX: Must be ordered to make sure parent is offlined after
4656 * children. The ordering requirement is for memcg where a
4657 * parent's offline may wait for a child's leading to deadlock. In
4658 * the long term, this should be fixed from memcg side.
4660 * We would prefer to do this in cgroup_init() above, but that
4661 * is called before init_workqueues(): so leave this until after.
4663 cgroup_destroy_wq = alloc_ordered_workqueue("cgroup_destroy", 0);
4664 BUG_ON(!cgroup_destroy_wq);
4667 * Used to destroy pidlists and separate to serve as flush domain.
4668 * Cap @max_active to 1 too.
4670 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4672 BUG_ON(!cgroup_pidlist_destroy_wq);
4676 core_initcall(cgroup_wq_init);
4679 * proc_cgroup_show()
4680 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4681 * - Used for /proc/<pid>/cgroup.
4682 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4683 * doesn't really matter if tsk->cgroup changes after we read it,
4684 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4685 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4686 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4687 * cgroup to top_cgroup.
4690 /* TODO: Use a proper seq_file iterator */
4691 int proc_cgroup_show(struct seq_file *m, void *v)
4694 struct task_struct *tsk;
4697 struct cgroupfs_root *root;
4700 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4706 tsk = get_pid_task(pid, PIDTYPE_PID);
4712 mutex_lock(&cgroup_mutex);
4714 for_each_active_root(root) {
4715 struct cgroup_subsys *ss;
4716 struct cgroup *cgrp;
4717 int ssid, count = 0;
4719 seq_printf(m, "%d:", root->hierarchy_id);
4720 for_each_subsys(ss, ssid)
4721 if (root->subsys_mask & (1 << ssid))
4722 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4723 if (strlen(root->name))
4724 seq_printf(m, "%sname=%s", count ? "," : "",
4727 cgrp = task_cgroup_from_root(tsk, root);
4728 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4736 mutex_unlock(&cgroup_mutex);
4737 put_task_struct(tsk);
4744 /* Display information about each subsystem and each hierarchy */
4745 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4747 struct cgroup_subsys *ss;
4750 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4752 * ideally we don't want subsystems moving around while we do this.
4753 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4754 * subsys/hierarchy state.
4756 mutex_lock(&cgroup_mutex);
4758 for_each_subsys(ss, i)
4759 seq_printf(m, "%s\t%d\t%d\t%d\n",
4760 ss->name, ss->root->hierarchy_id,
4761 ss->root->number_of_cgroups, !ss->disabled);
4763 mutex_unlock(&cgroup_mutex);
4767 static int cgroupstats_open(struct inode *inode, struct file *file)
4769 return single_open(file, proc_cgroupstats_show, NULL);
4772 static const struct file_operations proc_cgroupstats_operations = {
4773 .open = cgroupstats_open,
4775 .llseek = seq_lseek,
4776 .release = single_release,
4780 * cgroup_fork - attach newly forked task to its parents cgroup.
4781 * @child: pointer to task_struct of forking parent process.
4783 * Description: A task inherits its parent's cgroup at fork().
4785 * A pointer to the shared css_set was automatically copied in
4786 * fork.c by dup_task_struct(). However, we ignore that copy, since
4787 * it was not made under the protection of RCU or cgroup_mutex, so
4788 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4789 * have already changed current->cgroups, allowing the previously
4790 * referenced cgroup group to be removed and freed.
4792 * At the point that cgroup_fork() is called, 'current' is the parent
4793 * task, and the passed argument 'child' points to the child task.
4795 void cgroup_fork(struct task_struct *child)
4798 get_css_set(task_css_set(current));
4799 child->cgroups = current->cgroups;
4800 task_unlock(current);
4801 INIT_LIST_HEAD(&child->cg_list);
4805 * cgroup_post_fork - called on a new task after adding it to the task list
4806 * @child: the task in question
4808 * Adds the task to the list running through its css_set if necessary and
4809 * call the subsystem fork() callbacks. Has to be after the task is
4810 * visible on the task list in case we race with the first call to
4811 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4814 void cgroup_post_fork(struct task_struct *child)
4816 struct cgroup_subsys *ss;
4820 * use_task_css_set_links is set to 1 before we walk the tasklist
4821 * under the tasklist_lock and we read it here after we added the child
4822 * to the tasklist under the tasklist_lock as well. If the child wasn't
4823 * yet in the tasklist when we walked through it from
4824 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4825 * should be visible now due to the paired locking and barriers implied
4826 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4827 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4830 if (use_task_css_set_links) {
4831 write_lock(&css_set_lock);
4833 if (list_empty(&child->cg_list))
4834 list_add(&child->cg_list, &task_css_set(child)->tasks);
4836 write_unlock(&css_set_lock);
4840 * Call ss->fork(). This must happen after @child is linked on
4841 * css_set; otherwise, @child might change state between ->fork()
4842 * and addition to css_set.
4844 if (need_forkexit_callback) {
4845 for_each_subsys(ss, i)
4852 * cgroup_exit - detach cgroup from exiting task
4853 * @tsk: pointer to task_struct of exiting process
4854 * @run_callback: run exit callbacks?
4856 * Description: Detach cgroup from @tsk and release it.
4858 * Note that cgroups marked notify_on_release force every task in
4859 * them to take the global cgroup_mutex mutex when exiting.
4860 * This could impact scaling on very large systems. Be reluctant to
4861 * use notify_on_release cgroups where very high task exit scaling
4862 * is required on large systems.
4864 * the_top_cgroup_hack:
4866 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4868 * We call cgroup_exit() while the task is still competent to
4869 * handle notify_on_release(), then leave the task attached to the
4870 * root cgroup in each hierarchy for the remainder of its exit.
4872 * To do this properly, we would increment the reference count on
4873 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4874 * code we would add a second cgroup function call, to drop that
4875 * reference. This would just create an unnecessary hot spot on
4876 * the top_cgroup reference count, to no avail.
4878 * Normally, holding a reference to a cgroup without bumping its
4879 * count is unsafe. The cgroup could go away, or someone could
4880 * attach us to a different cgroup, decrementing the count on
4881 * the first cgroup that we never incremented. But in this case,
4882 * top_cgroup isn't going away, and either task has PF_EXITING set,
4883 * which wards off any cgroup_attach_task() attempts, or task is a failed
4884 * fork, never visible to cgroup_attach_task.
4886 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4888 struct cgroup_subsys *ss;
4889 struct css_set *cset;
4893 * Unlink from the css_set task list if necessary.
4894 * Optimistically check cg_list before taking
4897 if (!list_empty(&tsk->cg_list)) {
4898 write_lock(&css_set_lock);
4899 if (!list_empty(&tsk->cg_list))
4900 list_del_init(&tsk->cg_list);
4901 write_unlock(&css_set_lock);
4904 /* Reassign the task to the init_css_set. */
4906 cset = task_css_set(tsk);
4907 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4909 if (run_callbacks && need_forkexit_callback) {
4910 /* see cgroup_post_fork() for details */
4911 for_each_subsys(ss, i) {
4913 struct cgroup_subsys_state *old_css = cset->subsys[i];
4914 struct cgroup_subsys_state *css = task_css(tsk, i);
4916 ss->exit(css, old_css, tsk);
4922 put_css_set_taskexit(cset);
4925 static void check_for_release(struct cgroup *cgrp)
4927 if (cgroup_is_releasable(cgrp) &&
4928 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4930 * Control Group is currently removeable. If it's not
4931 * already queued for a userspace notification, queue
4934 int need_schedule_work = 0;
4936 raw_spin_lock(&release_list_lock);
4937 if (!cgroup_is_dead(cgrp) &&
4938 list_empty(&cgrp->release_list)) {
4939 list_add(&cgrp->release_list, &release_list);
4940 need_schedule_work = 1;
4942 raw_spin_unlock(&release_list_lock);
4943 if (need_schedule_work)
4944 schedule_work(&release_agent_work);
4949 * Notify userspace when a cgroup is released, by running the
4950 * configured release agent with the name of the cgroup (path
4951 * relative to the root of cgroup file system) as the argument.
4953 * Most likely, this user command will try to rmdir this cgroup.
4955 * This races with the possibility that some other task will be
4956 * attached to this cgroup before it is removed, or that some other
4957 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4958 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4959 * unused, and this cgroup will be reprieved from its death sentence,
4960 * to continue to serve a useful existence. Next time it's released,
4961 * we will get notified again, if it still has 'notify_on_release' set.
4963 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4964 * means only wait until the task is successfully execve()'d. The
4965 * separate release agent task is forked by call_usermodehelper(),
4966 * then control in this thread returns here, without waiting for the
4967 * release agent task. We don't bother to wait because the caller of
4968 * this routine has no use for the exit status of the release agent
4969 * task, so no sense holding our caller up for that.
4971 static void cgroup_release_agent(struct work_struct *work)
4973 BUG_ON(work != &release_agent_work);
4974 mutex_lock(&cgroup_mutex);
4975 raw_spin_lock(&release_list_lock);
4976 while (!list_empty(&release_list)) {
4977 char *argv[3], *envp[3];
4979 char *pathbuf = NULL, *agentbuf = NULL;
4980 struct cgroup *cgrp = list_entry(release_list.next,
4983 list_del_init(&cgrp->release_list);
4984 raw_spin_unlock(&release_list_lock);
4985 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4988 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4990 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4995 argv[i++] = agentbuf;
4996 argv[i++] = pathbuf;
5000 /* minimal command environment */
5001 envp[i++] = "HOME=/";
5002 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5005 /* Drop the lock while we invoke the usermode helper,
5006 * since the exec could involve hitting disk and hence
5007 * be a slow process */
5008 mutex_unlock(&cgroup_mutex);
5009 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5010 mutex_lock(&cgroup_mutex);
5014 raw_spin_lock(&release_list_lock);
5016 raw_spin_unlock(&release_list_lock);
5017 mutex_unlock(&cgroup_mutex);
5020 static int __init cgroup_disable(char *str)
5022 struct cgroup_subsys *ss;
5026 while ((token = strsep(&str, ",")) != NULL) {
5030 for_each_subsys(ss, i) {
5031 if (!strcmp(token, ss->name)) {
5033 printk(KERN_INFO "Disabling %s control group"
5034 " subsystem\n", ss->name);
5041 __setup("cgroup_disable=", cgroup_disable);
5044 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
5045 * @dentry: directory dentry of interest
5046 * @ss: subsystem of interest
5048 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5049 * to get the corresponding css and return it. If such css doesn't exist
5050 * or can't be pinned, an ERR_PTR value is returned.
5052 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
5053 struct cgroup_subsys *ss)
5055 struct cgroup *cgrp;
5056 struct cgroup_subsys_state *css;
5058 /* is @dentry a cgroup dir? */
5059 if (!dentry->d_inode ||
5060 dentry->d_inode->i_op != &cgroup_dir_inode_operations)
5061 return ERR_PTR(-EBADF);
5065 cgrp = __d_cgrp(dentry);
5066 css = cgroup_css(cgrp, ss);
5068 if (!css || !css_tryget(css))
5069 css = ERR_PTR(-ENOENT);
5076 * css_from_id - lookup css by id
5077 * @id: the cgroup id
5078 * @ss: cgroup subsys to be looked into
5080 * Returns the css if there's valid one with @id, otherwise returns NULL.
5081 * Should be called under rcu_read_lock().
5083 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5085 struct cgroup *cgrp;
5087 cgroup_assert_mutexes_or_rcu_locked();
5089 cgrp = idr_find(&ss->root->cgroup_idr, id);
5091 return cgroup_css(cgrp, ss);
5095 #ifdef CONFIG_CGROUP_DEBUG
5096 static struct cgroup_subsys_state *
5097 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5099 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5102 return ERR_PTR(-ENOMEM);
5107 static void debug_css_free(struct cgroup_subsys_state *css)
5112 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5115 return cgroup_task_count(css->cgroup);
5118 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5121 return (u64)(unsigned long)current->cgroups;
5124 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5130 count = atomic_read(&task_css_set(current)->refcount);
5135 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5137 struct cgrp_cset_link *link;
5138 struct css_set *cset;
5140 read_lock(&css_set_lock);
5142 cset = rcu_dereference(current->cgroups);
5143 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5144 struct cgroup *c = link->cgrp;
5148 name = c->dentry->d_name.name;
5151 seq_printf(seq, "Root %d group %s\n",
5152 c->root->hierarchy_id, name);
5155 read_unlock(&css_set_lock);
5159 #define MAX_TASKS_SHOWN_PER_CSS 25
5160 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5162 struct cgroup_subsys_state *css = seq_css(seq);
5163 struct cgrp_cset_link *link;
5165 read_lock(&css_set_lock);
5166 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5167 struct css_set *cset = link->cset;
5168 struct task_struct *task;
5170 seq_printf(seq, "css_set %p\n", cset);
5171 list_for_each_entry(task, &cset->tasks, cg_list) {
5172 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
5173 seq_puts(seq, " ...\n");
5176 seq_printf(seq, " task %d\n",
5177 task_pid_vnr(task));
5181 read_unlock(&css_set_lock);
5185 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5187 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5190 static struct cftype debug_files[] = {
5192 .name = "taskcount",
5193 .read_u64 = debug_taskcount_read,
5197 .name = "current_css_set",
5198 .read_u64 = current_css_set_read,
5202 .name = "current_css_set_refcount",
5203 .read_u64 = current_css_set_refcount_read,
5207 .name = "current_css_set_cg_links",
5208 .seq_show = current_css_set_cg_links_read,
5212 .name = "cgroup_css_links",
5213 .seq_show = cgroup_css_links_read,
5217 .name = "releasable",
5218 .read_u64 = releasable_read,
5224 struct cgroup_subsys debug_cgrp_subsys = {
5225 .css_alloc = debug_css_alloc,
5226 .css_free = debug_css_free,
5227 .base_cftypes = debug_files,
5229 #endif /* CONFIG_CGROUP_DEBUG */