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/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/cgroupstats.h>
53 #include <linux/hashtable.h>
54 #include <linux/namei.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/flex_array.h> /* used in cgroup_attach_task */
59 #include <linux/kthread.h>
61 #include <linux/atomic.h>
64 * pidlists linger the following amount before being destroyed. The goal
65 * is avoiding frequent destruction in the middle of consecutive read calls
66 * Expiring in the middle is a performance problem not a correctness one.
67 * 1 sec should be enough.
69 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
72 * cgroup_mutex is the master lock. Any modification to cgroup or its
73 * hierarchy must be performed while holding it.
75 * cgroup_root_mutex nests inside cgroup_mutex and should be held to modify
76 * cgroupfs_root of any cgroup hierarchy - subsys list, flags,
77 * release_agent_path and so on. Modifying requires both cgroup_mutex and
78 * cgroup_root_mutex. Readers can acquire either of the two. This is to
79 * break the following locking order cycle.
81 * A. cgroup_mutex -> cred_guard_mutex -> s_type->i_mutex_key -> namespace_sem
82 * B. namespace_sem -> cgroup_mutex
84 * B happens only through cgroup_show_options() and using cgroup_root_mutex
87 #ifdef CONFIG_PROVE_RCU
88 DEFINE_MUTEX(cgroup_mutex);
89 EXPORT_SYMBOL_GPL(cgroup_mutex); /* only for lockdep */
91 static DEFINE_MUTEX(cgroup_mutex);
94 static DEFINE_MUTEX(cgroup_root_mutex);
96 #define cgroup_assert_mutex_or_rcu_locked() \
97 rcu_lockdep_assert(rcu_read_lock_held() || \
98 lockdep_is_held(&cgroup_mutex), \
99 "cgroup_mutex or RCU read lock required");
101 #ifdef CONFIG_LOCKDEP
102 #define cgroup_assert_mutex_or_root_locked() \
103 WARN_ON_ONCE(debug_locks && (!lockdep_is_held(&cgroup_mutex) && \
104 !lockdep_is_held(&cgroup_root_mutex)))
106 #define cgroup_assert_mutex_or_root_locked() do { } while (0)
110 * cgroup destruction makes heavy use of work items and there can be a lot
111 * of concurrent destructions. Use a separate workqueue so that cgroup
112 * destruction work items don't end up filling up max_active of system_wq
113 * which may lead to deadlock.
115 static struct workqueue_struct *cgroup_destroy_wq;
118 * pidlist destructions need to be flushed on cgroup destruction. Use a
119 * separate workqueue as flush domain.
121 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
124 * Generate an array of cgroup subsystem pointers. At boot time, this is
125 * populated with the built in subsystems, and modular subsystems are
126 * registered after that. The mutable section of this array is protected by
129 #define SUBSYS(_x) [_x ## _subsys_id] = &_x ## _subsys,
130 #define IS_SUBSYS_ENABLED(option) IS_BUILTIN(option)
131 static struct cgroup_subsys *cgroup_subsys[CGROUP_SUBSYS_COUNT] = {
132 #include <linux/cgroup_subsys.h>
136 * The dummy hierarchy, reserved for the subsystems that are otherwise
137 * unattached - it never has more than a single cgroup, and all tasks are
138 * part of that cgroup.
140 static struct cgroupfs_root cgroup_dummy_root;
142 /* dummy_top is a shorthand for the dummy hierarchy's top cgroup */
143 static struct cgroup * const cgroup_dummy_top = &cgroup_dummy_root.top_cgroup;
145 /* The list of hierarchy roots */
147 static LIST_HEAD(cgroup_roots);
148 static int cgroup_root_count;
151 * Hierarchy ID allocation and mapping. It follows the same exclusion
152 * rules as other root ops - both cgroup_mutex and cgroup_root_mutex for
153 * writes, either for reads.
155 static DEFINE_IDR(cgroup_hierarchy_idr);
157 static struct cgroup_name root_cgroup_name = { .name = "/" };
160 * Assign a monotonically increasing serial number to cgroups. It
161 * guarantees cgroups with bigger numbers are newer than those with smaller
162 * numbers. Also, as cgroups are always appended to the parent's
163 * ->children list, it guarantees that sibling cgroups are always sorted in
164 * the ascending serial number order on the list. Protected by
167 static u64 cgroup_serial_nr_next = 1;
169 /* This flag indicates whether tasks in the fork and exit paths should
170 * check for fork/exit handlers to call. This avoids us having to do
171 * extra work in the fork/exit path if none of the subsystems need to
174 static int need_forkexit_callback __read_mostly;
176 static struct cftype cgroup_base_files[];
178 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
179 static int cgroup_destroy_locked(struct cgroup *cgrp);
180 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
182 static int cgroup_file_release(struct inode *inode, struct file *file);
183 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
186 * cgroup_css - obtain a cgroup's css for the specified subsystem
187 * @cgrp: the cgroup of interest
188 * @ss: the subsystem of interest (%NULL returns the dummy_css)
190 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
191 * function must be called either under cgroup_mutex or rcu_read_lock() and
192 * the caller is responsible for pinning the returned css if it wants to
193 * keep accessing it outside the said locks. This function may return
194 * %NULL if @cgrp doesn't have @subsys_id enabled.
196 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
197 struct cgroup_subsys *ss)
200 return rcu_dereference_check(cgrp->subsys[ss->subsys_id],
201 lockdep_is_held(&cgroup_mutex));
203 return &cgrp->dummy_css;
206 /* convenient tests for these bits */
207 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
209 return test_bit(CGRP_DEAD, &cgrp->flags);
213 * cgroup_is_descendant - test ancestry
214 * @cgrp: the cgroup to be tested
215 * @ancestor: possible ancestor of @cgrp
217 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
218 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
219 * and @ancestor are accessible.
221 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
224 if (cgrp == ancestor)
230 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
232 static int cgroup_is_releasable(const struct cgroup *cgrp)
235 (1 << CGRP_RELEASABLE) |
236 (1 << CGRP_NOTIFY_ON_RELEASE);
237 return (cgrp->flags & bits) == bits;
240 static int notify_on_release(const struct cgroup *cgrp)
242 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
246 * for_each_subsys - iterate all loaded cgroup subsystems
247 * @ss: the iteration cursor
248 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
250 * Iterates through all loaded subsystems. Should be called under
251 * cgroup_mutex or cgroup_root_mutex.
253 #define for_each_subsys(ss, ssid) \
254 for (({ cgroup_assert_mutex_or_root_locked(); (ssid) = 0; }); \
255 (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
256 if (!((ss) = cgroup_subsys[(ssid)])) { } \
260 * for_each_builtin_subsys - iterate all built-in cgroup subsystems
261 * @ss: the iteration cursor
262 * @i: the index of @ss, CGROUP_BUILTIN_SUBSYS_COUNT after reaching the end
264 * Bulit-in subsystems are always present and iteration itself doesn't
265 * require any synchronization.
267 #define for_each_builtin_subsys(ss, i) \
268 for ((i) = 0; (i) < CGROUP_BUILTIN_SUBSYS_COUNT && \
269 (((ss) = cgroup_subsys[i]) || true); (i)++)
271 /* iterate each subsystem attached to a hierarchy */
272 #define for_each_root_subsys(root, ss) \
273 list_for_each_entry((ss), &(root)->subsys_list, sibling)
275 /* iterate across the active hierarchies */
276 #define for_each_active_root(root) \
277 list_for_each_entry((root), &cgroup_roots, root_list)
279 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
281 return dentry->d_fsdata;
284 static inline struct cfent *__d_cfe(struct dentry *dentry)
286 return dentry->d_fsdata;
289 static inline struct cftype *__d_cft(struct dentry *dentry)
291 return __d_cfe(dentry)->type;
295 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
296 * @cgrp: the cgroup to be checked for liveness
298 * On success, returns true; the mutex should be later unlocked. On
299 * failure returns false with no lock held.
301 static bool cgroup_lock_live_group(struct cgroup *cgrp)
303 mutex_lock(&cgroup_mutex);
304 if (cgroup_is_dead(cgrp)) {
305 mutex_unlock(&cgroup_mutex);
311 /* the list of cgroups eligible for automatic release. Protected by
312 * release_list_lock */
313 static LIST_HEAD(release_list);
314 static DEFINE_RAW_SPINLOCK(release_list_lock);
315 static void cgroup_release_agent(struct work_struct *work);
316 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
317 static void check_for_release(struct cgroup *cgrp);
320 * A cgroup can be associated with multiple css_sets as different tasks may
321 * belong to different cgroups on different hierarchies. In the other
322 * direction, a css_set is naturally associated with multiple cgroups.
323 * This M:N relationship is represented by the following link structure
324 * which exists for each association and allows traversing the associations
327 struct cgrp_cset_link {
328 /* the cgroup and css_set this link associates */
330 struct css_set *cset;
332 /* list of cgrp_cset_links anchored at cgrp->cset_links */
333 struct list_head cset_link;
335 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
336 struct list_head cgrp_link;
339 /* The default css_set - used by init and its children prior to any
340 * hierarchies being mounted. It contains a pointer to the root state
341 * for each subsystem. Also used to anchor the list of css_sets. Not
342 * reference-counted, to improve performance when child cgroups
343 * haven't been created.
346 static struct css_set init_css_set;
347 static struct cgrp_cset_link init_cgrp_cset_link;
350 * css_set_lock protects the list of css_set objects, and the chain of
351 * tasks off each css_set. Nests outside task->alloc_lock due to
352 * css_task_iter_start().
354 static DEFINE_RWLOCK(css_set_lock);
355 static int css_set_count;
358 * hash table for cgroup groups. This improves the performance to find
359 * an existing css_set. This hash doesn't (currently) take into
360 * account cgroups in empty hierarchies.
362 #define CSS_SET_HASH_BITS 7
363 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
365 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
367 unsigned long key = 0UL;
368 struct cgroup_subsys *ss;
371 for_each_subsys(ss, i)
372 key += (unsigned long)css[i];
373 key = (key >> 16) ^ key;
379 * We don't maintain the lists running through each css_set to its task
380 * until after the first call to css_task_iter_start(). This reduces the
381 * fork()/exit() overhead for people who have cgroups compiled into their
382 * kernel but not actually in use.
384 static int use_task_css_set_links __read_mostly;
386 static void __put_css_set(struct css_set *cset, int taskexit)
388 struct cgrp_cset_link *link, *tmp_link;
391 * Ensure that the refcount doesn't hit zero while any readers
392 * can see it. Similar to atomic_dec_and_lock(), but for an
395 if (atomic_add_unless(&cset->refcount, -1, 1))
397 write_lock(&css_set_lock);
398 if (!atomic_dec_and_test(&cset->refcount)) {
399 write_unlock(&css_set_lock);
403 /* This css_set is dead. unlink it and release cgroup refcounts */
404 hash_del(&cset->hlist);
407 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
408 struct cgroup *cgrp = link->cgrp;
410 list_del(&link->cset_link);
411 list_del(&link->cgrp_link);
413 /* @cgrp can't go away while we're holding css_set_lock */
414 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
416 set_bit(CGRP_RELEASABLE, &cgrp->flags);
417 check_for_release(cgrp);
423 write_unlock(&css_set_lock);
424 kfree_rcu(cset, rcu_head);
428 * refcounted get/put for css_set objects
430 static inline void get_css_set(struct css_set *cset)
432 atomic_inc(&cset->refcount);
435 static inline void put_css_set(struct css_set *cset)
437 __put_css_set(cset, 0);
440 static inline void put_css_set_taskexit(struct css_set *cset)
442 __put_css_set(cset, 1);
446 * compare_css_sets - helper function for find_existing_css_set().
447 * @cset: candidate css_set being tested
448 * @old_cset: existing css_set for a task
449 * @new_cgrp: cgroup that's being entered by the task
450 * @template: desired set of css pointers in css_set (pre-calculated)
452 * Returns true if "cset" matches "old_cset" except for the hierarchy
453 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
455 static bool compare_css_sets(struct css_set *cset,
456 struct css_set *old_cset,
457 struct cgroup *new_cgrp,
458 struct cgroup_subsys_state *template[])
460 struct list_head *l1, *l2;
462 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
463 /* Not all subsystems matched */
468 * Compare cgroup pointers in order to distinguish between
469 * different cgroups in heirarchies with no subsystems. We
470 * could get by with just this check alone (and skip the
471 * memcmp above) but on most setups the memcmp check will
472 * avoid the need for this more expensive check on almost all
476 l1 = &cset->cgrp_links;
477 l2 = &old_cset->cgrp_links;
479 struct cgrp_cset_link *link1, *link2;
480 struct cgroup *cgrp1, *cgrp2;
484 /* See if we reached the end - both lists are equal length. */
485 if (l1 == &cset->cgrp_links) {
486 BUG_ON(l2 != &old_cset->cgrp_links);
489 BUG_ON(l2 == &old_cset->cgrp_links);
491 /* Locate the cgroups associated with these links. */
492 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
493 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
496 /* Hierarchies should be linked in the same order. */
497 BUG_ON(cgrp1->root != cgrp2->root);
500 * If this hierarchy is the hierarchy of the cgroup
501 * that's changing, then we need to check that this
502 * css_set points to the new cgroup; if it's any other
503 * hierarchy, then this css_set should point to the
504 * same cgroup as the old css_set.
506 if (cgrp1->root == new_cgrp->root) {
507 if (cgrp1 != new_cgrp)
518 * find_existing_css_set - init css array and find the matching css_set
519 * @old_cset: the css_set that we're using before the cgroup transition
520 * @cgrp: the cgroup that we're moving into
521 * @template: out param for the new set of csses, should be clear on entry
523 static struct css_set *find_existing_css_set(struct css_set *old_cset,
525 struct cgroup_subsys_state *template[])
527 struct cgroupfs_root *root = cgrp->root;
528 struct cgroup_subsys *ss;
529 struct css_set *cset;
534 * Build the set of subsystem state objects that we want to see in the
535 * new css_set. while subsystems can change globally, the entries here
536 * won't change, so no need for locking.
538 for_each_subsys(ss, i) {
539 if (root->subsys_mask & (1UL << i)) {
540 /* Subsystem is in this hierarchy. So we want
541 * the subsystem state from the new
543 template[i] = cgroup_css(cgrp, ss);
545 /* Subsystem is not in this hierarchy, so we
546 * don't want to change the subsystem state */
547 template[i] = old_cset->subsys[i];
551 key = css_set_hash(template);
552 hash_for_each_possible(css_set_table, cset, hlist, key) {
553 if (!compare_css_sets(cset, old_cset, cgrp, template))
556 /* This css_set matches what we need */
560 /* No existing cgroup group matched */
564 static void free_cgrp_cset_links(struct list_head *links_to_free)
566 struct cgrp_cset_link *link, *tmp_link;
568 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
569 list_del(&link->cset_link);
575 * allocate_cgrp_cset_links - allocate cgrp_cset_links
576 * @count: the number of links to allocate
577 * @tmp_links: list_head the allocated links are put on
579 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
580 * through ->cset_link. Returns 0 on success or -errno.
582 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
584 struct cgrp_cset_link *link;
587 INIT_LIST_HEAD(tmp_links);
589 for (i = 0; i < count; i++) {
590 link = kzalloc(sizeof(*link), GFP_KERNEL);
592 free_cgrp_cset_links(tmp_links);
595 list_add(&link->cset_link, tmp_links);
601 * link_css_set - a helper function to link a css_set to a cgroup
602 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
603 * @cset: the css_set to be linked
604 * @cgrp: the destination cgroup
606 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
609 struct cgrp_cset_link *link;
611 BUG_ON(list_empty(tmp_links));
612 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
615 list_move(&link->cset_link, &cgrp->cset_links);
617 * Always add links to the tail of the list so that the list
618 * is sorted by order of hierarchy creation
620 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
624 * find_css_set - return a new css_set with one cgroup updated
625 * @old_cset: the baseline css_set
626 * @cgrp: the cgroup to be updated
628 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
629 * substituted into the appropriate hierarchy.
631 static struct css_set *find_css_set(struct css_set *old_cset,
634 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
635 struct css_set *cset;
636 struct list_head tmp_links;
637 struct cgrp_cset_link *link;
640 lockdep_assert_held(&cgroup_mutex);
642 /* First see if we already have a cgroup group that matches
644 read_lock(&css_set_lock);
645 cset = find_existing_css_set(old_cset, cgrp, template);
648 read_unlock(&css_set_lock);
653 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
657 /* Allocate all the cgrp_cset_link objects that we'll need */
658 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
663 atomic_set(&cset->refcount, 1);
664 INIT_LIST_HEAD(&cset->cgrp_links);
665 INIT_LIST_HEAD(&cset->tasks);
666 INIT_HLIST_NODE(&cset->hlist);
668 /* Copy the set of subsystem state objects generated in
669 * find_existing_css_set() */
670 memcpy(cset->subsys, template, sizeof(cset->subsys));
672 write_lock(&css_set_lock);
673 /* Add reference counts and links from the new css_set. */
674 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
675 struct cgroup *c = link->cgrp;
677 if (c->root == cgrp->root)
679 link_css_set(&tmp_links, cset, c);
682 BUG_ON(!list_empty(&tmp_links));
686 /* Add this cgroup group to the hash table */
687 key = css_set_hash(cset->subsys);
688 hash_add(css_set_table, &cset->hlist, key);
690 write_unlock(&css_set_lock);
696 * Return the cgroup for "task" from the given hierarchy. Must be
697 * called with cgroup_mutex held.
699 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
700 struct cgroupfs_root *root)
702 struct css_set *cset;
703 struct cgroup *res = NULL;
705 BUG_ON(!mutex_is_locked(&cgroup_mutex));
706 read_lock(&css_set_lock);
708 * No need to lock the task - since we hold cgroup_mutex the
709 * task can't change groups, so the only thing that can happen
710 * is that it exits and its css is set back to init_css_set.
712 cset = task_css_set(task);
713 if (cset == &init_css_set) {
714 res = &root->top_cgroup;
716 struct cgrp_cset_link *link;
718 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
719 struct cgroup *c = link->cgrp;
721 if (c->root == root) {
727 read_unlock(&css_set_lock);
733 * There is one global cgroup mutex. We also require taking
734 * task_lock() when dereferencing a task's cgroup subsys pointers.
735 * See "The task_lock() exception", at the end of this comment.
737 * A task must hold cgroup_mutex to modify cgroups.
739 * Any task can increment and decrement the count field without lock.
740 * So in general, code holding cgroup_mutex can't rely on the count
741 * field not changing. However, if the count goes to zero, then only
742 * cgroup_attach_task() can increment it again. Because a count of zero
743 * means that no tasks are currently attached, therefore there is no
744 * way a task attached to that cgroup can fork (the other way to
745 * increment the count). So code holding cgroup_mutex can safely
746 * assume that if the count is zero, it will stay zero. Similarly, if
747 * a task holds cgroup_mutex on a cgroup with zero count, it
748 * knows that the cgroup won't be removed, as cgroup_rmdir()
751 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
752 * (usually) take cgroup_mutex. These are the two most performance
753 * critical pieces of code here. The exception occurs on cgroup_exit(),
754 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
755 * is taken, and if the cgroup count is zero, a usermode call made
756 * to the release agent with the name of the cgroup (path relative to
757 * the root of cgroup file system) as the argument.
759 * A cgroup can only be deleted if both its 'count' of using tasks
760 * is zero, and its list of 'children' cgroups is empty. Since all
761 * tasks in the system use _some_ cgroup, and since there is always at
762 * least one task in the system (init, pid == 1), therefore, top_cgroup
763 * always has either children cgroups and/or using tasks. So we don't
764 * need a special hack to ensure that top_cgroup cannot be deleted.
766 * The task_lock() exception
768 * The need for this exception arises from the action of
769 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
770 * another. It does so using cgroup_mutex, however there are
771 * several performance critical places that need to reference
772 * task->cgroup without the expense of grabbing a system global
773 * mutex. Therefore except as noted below, when dereferencing or, as
774 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
775 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
776 * the task_struct routinely used for such matters.
778 * P.S. One more locking exception. RCU is used to guard the
779 * update of a tasks cgroup pointer by cgroup_attach_task()
783 * A couple of forward declarations required, due to cyclic reference loop:
784 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
785 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
789 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode);
790 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
791 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
792 static const struct inode_operations cgroup_dir_inode_operations;
793 static const struct file_operations proc_cgroupstats_operations;
795 static struct backing_dev_info cgroup_backing_dev_info = {
797 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
800 static struct inode *cgroup_new_inode(umode_t mode, struct super_block *sb)
802 struct inode *inode = new_inode(sb);
805 inode->i_ino = get_next_ino();
806 inode->i_mode = mode;
807 inode->i_uid = current_fsuid();
808 inode->i_gid = current_fsgid();
809 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
810 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
815 static struct cgroup_name *cgroup_alloc_name(struct dentry *dentry)
817 struct cgroup_name *name;
819 name = kmalloc(sizeof(*name) + dentry->d_name.len + 1, GFP_KERNEL);
822 strcpy(name->name, dentry->d_name.name);
826 static void cgroup_free_fn(struct work_struct *work)
828 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
830 mutex_lock(&cgroup_mutex);
831 cgrp->root->number_of_cgroups--;
832 mutex_unlock(&cgroup_mutex);
835 * We get a ref to the parent's dentry, and put the ref when
836 * this cgroup is being freed, so it's guaranteed that the
837 * parent won't be destroyed before its children.
839 dput(cgrp->parent->dentry);
842 * Drop the active superblock reference that we took when we
843 * created the cgroup. This will free cgrp->root, if we are
844 * holding the last reference to @sb.
846 deactivate_super(cgrp->root->sb);
848 cgroup_pidlist_destroy_all(cgrp);
850 simple_xattrs_free(&cgrp->xattrs);
852 kfree(rcu_dereference_raw(cgrp->name));
856 static void cgroup_free_rcu(struct rcu_head *head)
858 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
860 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
861 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
864 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
866 /* is dentry a directory ? if so, kfree() associated cgroup */
867 if (S_ISDIR(inode->i_mode)) {
868 struct cgroup *cgrp = dentry->d_fsdata;
870 BUG_ON(!(cgroup_is_dead(cgrp)));
871 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
873 struct cfent *cfe = __d_cfe(dentry);
874 struct cgroup *cgrp = dentry->d_parent->d_fsdata;
876 WARN_ONCE(!list_empty(&cfe->node) &&
877 cgrp != &cgrp->root->top_cgroup,
878 "cfe still linked for %s\n", cfe->type->name);
879 simple_xattrs_free(&cfe->xattrs);
885 static void remove_dir(struct dentry *d)
887 struct dentry *parent = dget(d->d_parent);
890 simple_rmdir(parent->d_inode, d);
894 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
898 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
899 lockdep_assert_held(&cgroup_mutex);
902 * If we're doing cleanup due to failure of cgroup_create(),
903 * the corresponding @cfe may not exist.
905 list_for_each_entry(cfe, &cgrp->files, node) {
906 struct dentry *d = cfe->dentry;
908 if (cft && cfe->type != cft)
913 simple_unlink(cgrp->dentry->d_inode, d);
914 list_del_init(&cfe->node);
922 * cgroup_clear_dir - remove subsys files in a cgroup directory
923 * @cgrp: target cgroup
924 * @subsys_mask: mask of the subsystem ids whose files should be removed
926 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
928 struct cgroup_subsys *ss;
931 for_each_subsys(ss, i) {
932 struct cftype_set *set;
934 if (!test_bit(i, &subsys_mask))
936 list_for_each_entry(set, &ss->cftsets, node)
937 cgroup_addrm_files(cgrp, set->cfts, false);
942 * NOTE : the dentry must have been dget()'ed
944 static void cgroup_d_remove_dir(struct dentry *dentry)
946 struct dentry *parent;
948 parent = dentry->d_parent;
949 spin_lock(&parent->d_lock);
950 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
951 list_del_init(&dentry->d_u.d_child);
952 spin_unlock(&dentry->d_lock);
953 spin_unlock(&parent->d_lock);
958 * Call with cgroup_mutex held. Drops reference counts on modules, including
959 * any duplicate ones that parse_cgroupfs_options took. If this function
960 * returns an error, no reference counts are touched.
962 static int rebind_subsystems(struct cgroupfs_root *root,
963 unsigned long added_mask, unsigned removed_mask)
965 struct cgroup *cgrp = &root->top_cgroup;
966 struct cgroup_subsys *ss;
967 unsigned long pinned = 0;
970 BUG_ON(!mutex_is_locked(&cgroup_mutex));
971 BUG_ON(!mutex_is_locked(&cgroup_root_mutex));
973 /* Check that any added subsystems are currently free */
974 for_each_subsys(ss, i) {
975 if (!(added_mask & (1 << i)))
978 /* is the subsystem mounted elsewhere? */
979 if (ss->root != &cgroup_dummy_root) {
985 if (!try_module_get(ss->module)) {
992 /* subsys could be missing if unloaded between parsing and here */
993 if (added_mask != pinned) {
998 ret = cgroup_populate_dir(cgrp, added_mask);
1003 * Nothing can fail from this point on. Remove files for the
1004 * removed subsystems and rebind each subsystem.
1006 cgroup_clear_dir(cgrp, removed_mask);
1008 for_each_subsys(ss, i) {
1009 unsigned long bit = 1UL << i;
1011 if (bit & added_mask) {
1012 /* We're binding this subsystem to this hierarchy */
1013 BUG_ON(cgroup_css(cgrp, ss));
1014 BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
1015 BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
1017 rcu_assign_pointer(cgrp->subsys[i],
1018 cgroup_css(cgroup_dummy_top, ss));
1019 cgroup_css(cgrp, ss)->cgroup = cgrp;
1021 list_move(&ss->sibling, &root->subsys_list);
1024 ss->bind(cgroup_css(cgrp, ss));
1026 /* refcount was already taken, and we're keeping it */
1027 root->subsys_mask |= bit;
1028 } else if (bit & removed_mask) {
1029 /* We're removing this subsystem */
1030 BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1031 BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1034 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1036 cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1037 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1039 cgroup_subsys[i]->root = &cgroup_dummy_root;
1040 list_move(&ss->sibling, &cgroup_dummy_root.subsys_list);
1042 /* subsystem is now free - drop reference on module */
1043 module_put(ss->module);
1044 root->subsys_mask &= ~bit;
1049 * Mark @root has finished binding subsystems. @root->subsys_mask
1050 * now matches the bound subsystems.
1052 root->flags |= CGRP_ROOT_SUBSYS_BOUND;
1057 for_each_subsys(ss, i)
1058 if (pinned & (1 << i))
1059 module_put(ss->module);
1063 static int cgroup_show_options(struct seq_file *seq, struct dentry *dentry)
1065 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
1066 struct cgroup_subsys *ss;
1068 mutex_lock(&cgroup_root_mutex);
1069 for_each_root_subsys(root, ss)
1070 seq_printf(seq, ",%s", ss->name);
1071 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1072 seq_puts(seq, ",sane_behavior");
1073 if (root->flags & CGRP_ROOT_NOPREFIX)
1074 seq_puts(seq, ",noprefix");
1075 if (root->flags & CGRP_ROOT_XATTR)
1076 seq_puts(seq, ",xattr");
1077 if (strlen(root->release_agent_path))
1078 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1079 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1080 seq_puts(seq, ",clone_children");
1081 if (strlen(root->name))
1082 seq_printf(seq, ",name=%s", root->name);
1083 mutex_unlock(&cgroup_root_mutex);
1087 struct cgroup_sb_opts {
1088 unsigned long subsys_mask;
1089 unsigned long flags;
1090 char *release_agent;
1091 bool cpuset_clone_children;
1093 /* User explicitly requested empty subsystem */
1096 struct cgroupfs_root *new_root;
1101 * Convert a hierarchy specifier into a bitmask of subsystems and
1102 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1103 * array. This function takes refcounts on subsystems to be used, unless it
1104 * returns error, in which case no refcounts are taken.
1106 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1108 char *token, *o = data;
1109 bool all_ss = false, one_ss = false;
1110 unsigned long mask = (unsigned long)-1;
1111 struct cgroup_subsys *ss;
1114 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1116 #ifdef CONFIG_CPUSETS
1117 mask = ~(1UL << cpuset_subsys_id);
1120 memset(opts, 0, sizeof(*opts));
1122 while ((token = strsep(&o, ",")) != NULL) {
1125 if (!strcmp(token, "none")) {
1126 /* Explicitly have no subsystems */
1130 if (!strcmp(token, "all")) {
1131 /* Mutually exclusive option 'all' + subsystem name */
1137 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1138 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1141 if (!strcmp(token, "noprefix")) {
1142 opts->flags |= CGRP_ROOT_NOPREFIX;
1145 if (!strcmp(token, "clone_children")) {
1146 opts->cpuset_clone_children = true;
1149 if (!strcmp(token, "xattr")) {
1150 opts->flags |= CGRP_ROOT_XATTR;
1153 if (!strncmp(token, "release_agent=", 14)) {
1154 /* Specifying two release agents is forbidden */
1155 if (opts->release_agent)
1157 opts->release_agent =
1158 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1159 if (!opts->release_agent)
1163 if (!strncmp(token, "name=", 5)) {
1164 const char *name = token + 5;
1165 /* Can't specify an empty name */
1168 /* Must match [\w.-]+ */
1169 for (i = 0; i < strlen(name); i++) {
1173 if ((c == '.') || (c == '-') || (c == '_'))
1177 /* Specifying two names is forbidden */
1180 opts->name = kstrndup(name,
1181 MAX_CGROUP_ROOT_NAMELEN - 1,
1189 for_each_subsys(ss, i) {
1190 if (strcmp(token, ss->name))
1195 /* Mutually exclusive option 'all' + subsystem name */
1198 set_bit(i, &opts->subsys_mask);
1203 if (i == CGROUP_SUBSYS_COUNT)
1208 * If the 'all' option was specified select all the subsystems,
1209 * otherwise if 'none', 'name=' and a subsystem name options
1210 * were not specified, let's default to 'all'
1212 if (all_ss || (!one_ss && !opts->none && !opts->name))
1213 for_each_subsys(ss, i)
1215 set_bit(i, &opts->subsys_mask);
1217 /* Consistency checks */
1219 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1220 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1222 if (opts->flags & CGRP_ROOT_NOPREFIX) {
1223 pr_err("cgroup: sane_behavior: noprefix is not allowed\n");
1227 if (opts->cpuset_clone_children) {
1228 pr_err("cgroup: sane_behavior: clone_children is not allowed\n");
1234 * Option noprefix was introduced just for backward compatibility
1235 * with the old cpuset, so we allow noprefix only if mounting just
1236 * the cpuset subsystem.
1238 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1242 /* Can't specify "none" and some subsystems */
1243 if (opts->subsys_mask && opts->none)
1247 * We either have to specify by name or by subsystems. (So all
1248 * empty hierarchies must have a name).
1250 if (!opts->subsys_mask && !opts->name)
1256 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1259 struct cgroupfs_root *root = sb->s_fs_info;
1260 struct cgroup *cgrp = &root->top_cgroup;
1261 struct cgroup_sb_opts opts;
1262 unsigned long added_mask, removed_mask;
1264 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1265 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1269 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1270 mutex_lock(&cgroup_mutex);
1271 mutex_lock(&cgroup_root_mutex);
1273 /* See what subsystems are wanted */
1274 ret = parse_cgroupfs_options(data, &opts);
1278 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1279 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1280 task_tgid_nr(current), current->comm);
1282 added_mask = opts.subsys_mask & ~root->subsys_mask;
1283 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1285 /* Don't allow flags or name to change at remount */
1286 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1287 (opts.name && strcmp(opts.name, root->name))) {
1288 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1289 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1290 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1295 /* remounting is not allowed for populated hierarchies */
1296 if (root->number_of_cgroups > 1) {
1301 ret = rebind_subsystems(root, added_mask, removed_mask);
1305 if (opts.release_agent)
1306 strcpy(root->release_agent_path, opts.release_agent);
1308 kfree(opts.release_agent);
1310 mutex_unlock(&cgroup_root_mutex);
1311 mutex_unlock(&cgroup_mutex);
1312 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1316 static const struct super_operations cgroup_ops = {
1317 .statfs = simple_statfs,
1318 .drop_inode = generic_delete_inode,
1319 .show_options = cgroup_show_options,
1320 .remount_fs = cgroup_remount,
1323 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1325 INIT_LIST_HEAD(&cgrp->sibling);
1326 INIT_LIST_HEAD(&cgrp->children);
1327 INIT_LIST_HEAD(&cgrp->files);
1328 INIT_LIST_HEAD(&cgrp->cset_links);
1329 INIT_LIST_HEAD(&cgrp->release_list);
1330 INIT_LIST_HEAD(&cgrp->pidlists);
1331 mutex_init(&cgrp->pidlist_mutex);
1332 cgrp->dummy_css.cgroup = cgrp;
1333 simple_xattrs_init(&cgrp->xattrs);
1336 static void init_cgroup_root(struct cgroupfs_root *root)
1338 struct cgroup *cgrp = &root->top_cgroup;
1340 INIT_LIST_HEAD(&root->subsys_list);
1341 INIT_LIST_HEAD(&root->root_list);
1342 root->number_of_cgroups = 1;
1344 RCU_INIT_POINTER(cgrp->name, &root_cgroup_name);
1345 init_cgroup_housekeeping(cgrp);
1346 idr_init(&root->cgroup_idr);
1349 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
1353 lockdep_assert_held(&cgroup_mutex);
1354 lockdep_assert_held(&cgroup_root_mutex);
1356 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
1361 root->hierarchy_id = id;
1365 static void cgroup_exit_root_id(struct cgroupfs_root *root)
1367 lockdep_assert_held(&cgroup_mutex);
1368 lockdep_assert_held(&cgroup_root_mutex);
1370 if (root->hierarchy_id) {
1371 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1372 root->hierarchy_id = 0;
1376 static int cgroup_test_super(struct super_block *sb, void *data)
1378 struct cgroup_sb_opts *opts = data;
1379 struct cgroupfs_root *root = sb->s_fs_info;
1381 /* If we asked for a name then it must match */
1382 if (opts->name && strcmp(opts->name, root->name))
1386 * If we asked for subsystems (or explicitly for no
1387 * subsystems) then they must match
1389 if ((opts->subsys_mask || opts->none)
1390 && (opts->subsys_mask != root->subsys_mask))
1396 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1398 struct cgroupfs_root *root;
1400 if (!opts->subsys_mask && !opts->none)
1403 root = kzalloc(sizeof(*root), GFP_KERNEL);
1405 return ERR_PTR(-ENOMEM);
1407 init_cgroup_root(root);
1410 * We need to set @root->subsys_mask now so that @root can be
1411 * matched by cgroup_test_super() before it finishes
1412 * initialization; otherwise, competing mounts with the same
1413 * options may try to bind the same subsystems instead of waiting
1414 * for the first one leading to unexpected mount errors.
1415 * SUBSYS_BOUND will be set once actual binding is complete.
1417 root->subsys_mask = opts->subsys_mask;
1418 root->flags = opts->flags;
1419 if (opts->release_agent)
1420 strcpy(root->release_agent_path, opts->release_agent);
1422 strcpy(root->name, opts->name);
1423 if (opts->cpuset_clone_children)
1424 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1428 static void cgroup_free_root(struct cgroupfs_root *root)
1431 /* hierarhcy ID shoulid already have been released */
1432 WARN_ON_ONCE(root->hierarchy_id);
1434 idr_destroy(&root->cgroup_idr);
1439 static int cgroup_set_super(struct super_block *sb, void *data)
1442 struct cgroup_sb_opts *opts = data;
1444 /* If we don't have a new root, we can't set up a new sb */
1445 if (!opts->new_root)
1448 BUG_ON(!opts->subsys_mask && !opts->none);
1450 ret = set_anon_super(sb, NULL);
1454 sb->s_fs_info = opts->new_root;
1455 opts->new_root->sb = sb;
1457 sb->s_blocksize = PAGE_CACHE_SIZE;
1458 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1459 sb->s_magic = CGROUP_SUPER_MAGIC;
1460 sb->s_op = &cgroup_ops;
1465 static int cgroup_get_rootdir(struct super_block *sb)
1467 static const struct dentry_operations cgroup_dops = {
1468 .d_iput = cgroup_diput,
1469 .d_delete = always_delete_dentry,
1472 struct inode *inode =
1473 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1478 inode->i_fop = &simple_dir_operations;
1479 inode->i_op = &cgroup_dir_inode_operations;
1480 /* directories start off with i_nlink == 2 (for "." entry) */
1482 sb->s_root = d_make_root(inode);
1485 /* for everything else we want ->d_op set */
1486 sb->s_d_op = &cgroup_dops;
1490 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1491 int flags, const char *unused_dev_name,
1494 struct cgroup_sb_opts opts;
1495 struct cgroupfs_root *root;
1497 struct super_block *sb;
1498 struct cgroupfs_root *new_root;
1499 struct list_head tmp_links;
1500 struct inode *inode;
1501 const struct cred *cred;
1503 /* First find the desired set of subsystems */
1504 mutex_lock(&cgroup_mutex);
1505 ret = parse_cgroupfs_options(data, &opts);
1506 mutex_unlock(&cgroup_mutex);
1511 * Allocate a new cgroup root. We may not need it if we're
1512 * reusing an existing hierarchy.
1514 new_root = cgroup_root_from_opts(&opts);
1515 if (IS_ERR(new_root)) {
1516 ret = PTR_ERR(new_root);
1519 opts.new_root = new_root;
1521 /* Locate an existing or new sb for this hierarchy */
1522 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, 0, &opts);
1525 cgroup_free_root(opts.new_root);
1529 root = sb->s_fs_info;
1531 if (root == opts.new_root) {
1532 /* We used the new root structure, so this is a new hierarchy */
1533 struct cgroup *root_cgrp = &root->top_cgroup;
1534 struct cgroupfs_root *existing_root;
1536 struct css_set *cset;
1538 BUG_ON(sb->s_root != NULL);
1540 ret = cgroup_get_rootdir(sb);
1542 goto drop_new_super;
1543 inode = sb->s_root->d_inode;
1545 mutex_lock(&inode->i_mutex);
1546 mutex_lock(&cgroup_mutex);
1547 mutex_lock(&cgroup_root_mutex);
1549 root_cgrp->id = idr_alloc(&root->cgroup_idr, root_cgrp,
1551 if (root_cgrp->id < 0)
1554 /* Check for name clashes with existing mounts */
1556 if (strlen(root->name))
1557 for_each_active_root(existing_root)
1558 if (!strcmp(existing_root->name, root->name))
1562 * We're accessing css_set_count without locking
1563 * css_set_lock here, but that's OK - it can only be
1564 * increased by someone holding cgroup_lock, and
1565 * that's us. The worst that can happen is that we
1566 * have some link structures left over
1568 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1572 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1573 ret = cgroup_init_root_id(root, 2, 0);
1577 sb->s_root->d_fsdata = root_cgrp;
1578 root_cgrp->dentry = sb->s_root;
1581 * We're inside get_sb() and will call lookup_one_len() to
1582 * create the root files, which doesn't work if SELinux is
1583 * in use. The following cred dancing somehow works around
1584 * it. See 2ce9738ba ("cgroupfs: use init_cred when
1585 * populating new cgroupfs mount") for more details.
1587 cred = override_creds(&init_cred);
1589 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1593 ret = rebind_subsystems(root, root->subsys_mask, 0);
1600 * There must be no failure case after here, since rebinding
1601 * takes care of subsystems' refcounts, which are explicitly
1602 * dropped in the failure exit path.
1605 list_add(&root->root_list, &cgroup_roots);
1606 cgroup_root_count++;
1608 /* Link the top cgroup in this hierarchy into all
1609 * the css_set objects */
1610 write_lock(&css_set_lock);
1611 hash_for_each(css_set_table, i, cset, hlist)
1612 link_css_set(&tmp_links, cset, root_cgrp);
1613 write_unlock(&css_set_lock);
1615 free_cgrp_cset_links(&tmp_links);
1617 BUG_ON(!list_empty(&root_cgrp->children));
1618 BUG_ON(root->number_of_cgroups != 1);
1620 mutex_unlock(&cgroup_root_mutex);
1621 mutex_unlock(&cgroup_mutex);
1622 mutex_unlock(&inode->i_mutex);
1625 * We re-used an existing hierarchy - the new root (if
1626 * any) is not needed
1628 cgroup_free_root(opts.new_root);
1630 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1631 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1632 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1634 goto drop_new_super;
1636 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1641 kfree(opts.release_agent);
1643 return dget(sb->s_root);
1646 free_cgrp_cset_links(&tmp_links);
1647 cgroup_addrm_files(&root->top_cgroup, cgroup_base_files, false);
1650 cgroup_exit_root_id(root);
1651 mutex_unlock(&cgroup_root_mutex);
1652 mutex_unlock(&cgroup_mutex);
1653 mutex_unlock(&inode->i_mutex);
1655 deactivate_locked_super(sb);
1657 kfree(opts.release_agent);
1659 return ERR_PTR(ret);
1662 static void cgroup_kill_sb(struct super_block *sb) {
1663 struct cgroupfs_root *root = sb->s_fs_info;
1664 struct cgroup *cgrp = &root->top_cgroup;
1665 struct cgrp_cset_link *link, *tmp_link;
1670 BUG_ON(root->number_of_cgroups != 1);
1671 BUG_ON(!list_empty(&cgrp->children));
1673 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1674 mutex_lock(&cgroup_mutex);
1675 mutex_lock(&cgroup_root_mutex);
1677 /* Rebind all subsystems back to the default hierarchy */
1678 if (root->flags & CGRP_ROOT_SUBSYS_BOUND) {
1679 ret = rebind_subsystems(root, 0, root->subsys_mask);
1680 /* Shouldn't be able to fail ... */
1685 * Release all the links from cset_links to this hierarchy's
1688 write_lock(&css_set_lock);
1690 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1691 list_del(&link->cset_link);
1692 list_del(&link->cgrp_link);
1695 write_unlock(&css_set_lock);
1697 if (!list_empty(&root->root_list)) {
1698 list_del(&root->root_list);
1699 cgroup_root_count--;
1702 cgroup_exit_root_id(root);
1704 mutex_unlock(&cgroup_root_mutex);
1705 mutex_unlock(&cgroup_mutex);
1706 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1708 simple_xattrs_free(&cgrp->xattrs);
1710 kill_litter_super(sb);
1711 cgroup_free_root(root);
1714 static struct file_system_type cgroup_fs_type = {
1716 .mount = cgroup_mount,
1717 .kill_sb = cgroup_kill_sb,
1720 static struct kobject *cgroup_kobj;
1723 * cgroup_path - generate the path of a cgroup
1724 * @cgrp: the cgroup in question
1725 * @buf: the buffer to write the path into
1726 * @buflen: the length of the buffer
1728 * Writes path of cgroup into buf. Returns 0 on success, -errno on error.
1730 * We can't generate cgroup path using dentry->d_name, as accessing
1731 * dentry->name must be protected by irq-unsafe dentry->d_lock or parent
1732 * inode's i_mutex, while on the other hand cgroup_path() can be called
1733 * with some irq-safe spinlocks held.
1735 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1737 int ret = -ENAMETOOLONG;
1740 if (!cgrp->parent) {
1741 if (strlcpy(buf, "/", buflen) >= buflen)
1742 return -ENAMETOOLONG;
1746 start = buf + buflen - 1;
1751 const char *name = cgroup_name(cgrp);
1755 if ((start -= len) < buf)
1757 memcpy(start, name, len);
1763 cgrp = cgrp->parent;
1764 } while (cgrp->parent);
1766 memmove(buf, start, buf + buflen - start);
1771 EXPORT_SYMBOL_GPL(cgroup_path);
1774 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1775 * @task: target task
1776 * @buf: the buffer to write the path into
1777 * @buflen: the length of the buffer
1779 * Determine @task's cgroup on the first (the one with the lowest non-zero
1780 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1781 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1782 * cgroup controller callbacks.
1784 * Returns 0 on success, fails with -%ENAMETOOLONG if @buflen is too short.
1786 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1788 struct cgroupfs_root *root;
1789 struct cgroup *cgrp;
1790 int hierarchy_id = 1, ret = 0;
1793 return -ENAMETOOLONG;
1795 mutex_lock(&cgroup_mutex);
1797 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1800 cgrp = task_cgroup_from_root(task, root);
1801 ret = cgroup_path(cgrp, buf, buflen);
1803 /* if no hierarchy exists, everyone is in "/" */
1804 memcpy(buf, "/", 2);
1807 mutex_unlock(&cgroup_mutex);
1810 EXPORT_SYMBOL_GPL(task_cgroup_path);
1813 * Control Group taskset
1815 struct task_and_cgroup {
1816 struct task_struct *task;
1817 struct cgroup *cgrp;
1818 struct css_set *cset;
1821 struct cgroup_taskset {
1822 struct task_and_cgroup single;
1823 struct flex_array *tc_array;
1826 struct cgroup *cur_cgrp;
1830 * cgroup_taskset_first - reset taskset and return the first task
1831 * @tset: taskset of interest
1833 * @tset iteration is initialized and the first task is returned.
1835 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1837 if (tset->tc_array) {
1839 return cgroup_taskset_next(tset);
1841 tset->cur_cgrp = tset->single.cgrp;
1842 return tset->single.task;
1845 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1848 * cgroup_taskset_next - iterate to the next task in taskset
1849 * @tset: taskset of interest
1851 * Return the next task in @tset. Iteration must have been initialized
1852 * with cgroup_taskset_first().
1854 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1856 struct task_and_cgroup *tc;
1858 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1861 tc = flex_array_get(tset->tc_array, tset->idx++);
1862 tset->cur_cgrp = tc->cgrp;
1865 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1868 * cgroup_taskset_cur_css - return the matching css for the current task
1869 * @tset: taskset of interest
1870 * @subsys_id: the ID of the target subsystem
1872 * Return the css for the current (last returned) task of @tset for
1873 * subsystem specified by @subsys_id. This function must be preceded by
1874 * either cgroup_taskset_first() or cgroup_taskset_next().
1876 struct cgroup_subsys_state *cgroup_taskset_cur_css(struct cgroup_taskset *tset,
1879 return cgroup_css(tset->cur_cgrp, cgroup_subsys[subsys_id]);
1881 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_css);
1884 * cgroup_taskset_size - return the number of tasks in taskset
1885 * @tset: taskset of interest
1887 int cgroup_taskset_size(struct cgroup_taskset *tset)
1889 return tset->tc_array ? tset->tc_array_len : 1;
1891 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1895 * cgroup_task_migrate - move a task from one cgroup to another.
1897 * Must be called with cgroup_mutex and threadgroup locked.
1899 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1900 struct task_struct *tsk,
1901 struct css_set *new_cset)
1903 struct css_set *old_cset;
1906 * We are synchronized through threadgroup_lock() against PF_EXITING
1907 * setting such that we can't race against cgroup_exit() changing the
1908 * css_set to init_css_set and dropping the old one.
1910 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1911 old_cset = task_css_set(tsk);
1914 rcu_assign_pointer(tsk->cgroups, new_cset);
1917 /* Update the css_set linked lists if we're using them */
1918 write_lock(&css_set_lock);
1919 if (!list_empty(&tsk->cg_list))
1920 list_move(&tsk->cg_list, &new_cset->tasks);
1921 write_unlock(&css_set_lock);
1924 * We just gained a reference on old_cset by taking it from the
1925 * task. As trading it for new_cset is protected by cgroup_mutex,
1926 * we're safe to drop it here; it will be freed under RCU.
1928 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1929 put_css_set(old_cset);
1933 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1934 * @cgrp: the cgroup to attach to
1935 * @tsk: the task or the leader of the threadgroup to be attached
1936 * @threadgroup: attach the whole threadgroup?
1938 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1939 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1941 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1944 int retval, i, group_size;
1945 struct cgroup_subsys *ss, *failed_ss = NULL;
1946 struct cgroupfs_root *root = cgrp->root;
1947 /* threadgroup list cursor and array */
1948 struct task_struct *leader = tsk;
1949 struct task_and_cgroup *tc;
1950 struct flex_array *group;
1951 struct cgroup_taskset tset = { };
1954 * step 0: in order to do expensive, possibly blocking operations for
1955 * every thread, we cannot iterate the thread group list, since it needs
1956 * rcu or tasklist locked. instead, build an array of all threads in the
1957 * group - group_rwsem prevents new threads from appearing, and if
1958 * threads exit, this will just be an over-estimate.
1961 group_size = get_nr_threads(tsk);
1964 /* flex_array supports very large thread-groups better than kmalloc. */
1965 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1968 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1969 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1971 goto out_free_group_list;
1975 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1976 * already PF_EXITING could be freed from underneath us unless we
1977 * take an rcu_read_lock.
1981 struct task_and_cgroup ent;
1983 /* @tsk either already exited or can't exit until the end */
1984 if (tsk->flags & PF_EXITING)
1987 /* as per above, nr_threads may decrease, but not increase. */
1988 BUG_ON(i >= group_size);
1990 ent.cgrp = task_cgroup_from_root(tsk, root);
1991 /* nothing to do if this task is already in the cgroup */
1992 if (ent.cgrp == cgrp)
1995 * saying GFP_ATOMIC has no effect here because we did prealloc
1996 * earlier, but it's good form to communicate our expectations.
1998 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
1999 BUG_ON(retval != 0);
2004 } while_each_thread(leader, tsk);
2006 /* remember the number of threads in the array for later. */
2008 tset.tc_array = group;
2009 tset.tc_array_len = group_size;
2011 /* methods shouldn't be called if no task is actually migrating */
2014 goto out_free_group_list;
2017 * step 1: check that we can legitimately attach to the cgroup.
2019 for_each_root_subsys(root, ss) {
2020 struct cgroup_subsys_state *css = cgroup_css(cgrp, ss);
2022 if (ss->can_attach) {
2023 retval = ss->can_attach(css, &tset);
2026 goto out_cancel_attach;
2032 * step 2: make sure css_sets exist for all threads to be migrated.
2033 * we use find_css_set, which allocates a new one if necessary.
2035 for (i = 0; i < group_size; i++) {
2036 struct css_set *old_cset;
2038 tc = flex_array_get(group, i);
2039 old_cset = task_css_set(tc->task);
2040 tc->cset = find_css_set(old_cset, cgrp);
2043 goto out_put_css_set_refs;
2048 * step 3: now that we're guaranteed success wrt the css_sets,
2049 * proceed to move all tasks to the new cgroup. There are no
2050 * failure cases after here, so this is the commit point.
2052 for (i = 0; i < group_size; i++) {
2053 tc = flex_array_get(group, i);
2054 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
2056 /* nothing is sensitive to fork() after this point. */
2059 * step 4: do subsystem attach callbacks.
2061 for_each_root_subsys(root, ss) {
2062 struct cgroup_subsys_state *css = cgroup_css(cgrp, ss);
2065 ss->attach(css, &tset);
2069 * step 5: success! and cleanup
2072 out_put_css_set_refs:
2074 for (i = 0; i < group_size; i++) {
2075 tc = flex_array_get(group, i);
2078 put_css_set(tc->cset);
2083 for_each_root_subsys(root, ss) {
2084 struct cgroup_subsys_state *css = cgroup_css(cgrp, ss);
2086 if (ss == failed_ss)
2088 if (ss->cancel_attach)
2089 ss->cancel_attach(css, &tset);
2092 out_free_group_list:
2093 flex_array_free(group);
2098 * Find the task_struct of the task to attach by vpid and pass it along to the
2099 * function to attach either it or all tasks in its threadgroup. Will lock
2100 * cgroup_mutex and threadgroup; may take task_lock of task.
2102 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2104 struct task_struct *tsk;
2105 const struct cred *cred = current_cred(), *tcred;
2108 if (!cgroup_lock_live_group(cgrp))
2114 tsk = find_task_by_vpid(pid);
2118 goto out_unlock_cgroup;
2121 * even if we're attaching all tasks in the thread group, we
2122 * only need to check permissions on one of them.
2124 tcred = __task_cred(tsk);
2125 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2126 !uid_eq(cred->euid, tcred->uid) &&
2127 !uid_eq(cred->euid, tcred->suid)) {
2130 goto out_unlock_cgroup;
2136 tsk = tsk->group_leader;
2139 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2140 * trapped in a cpuset, or RT worker may be born in a cgroup
2141 * with no rt_runtime allocated. Just say no.
2143 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2146 goto out_unlock_cgroup;
2149 get_task_struct(tsk);
2152 threadgroup_lock(tsk);
2154 if (!thread_group_leader(tsk)) {
2156 * a race with de_thread from another thread's exec()
2157 * may strip us of our leadership, if this happens,
2158 * there is no choice but to throw this task away and
2159 * try again; this is
2160 * "double-double-toil-and-trouble-check locking".
2162 threadgroup_unlock(tsk);
2163 put_task_struct(tsk);
2164 goto retry_find_task;
2168 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2170 threadgroup_unlock(tsk);
2172 put_task_struct(tsk);
2174 mutex_unlock(&cgroup_mutex);
2179 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2180 * @from: attach to all cgroups of a given task
2181 * @tsk: the task to be attached
2183 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2185 struct cgroupfs_root *root;
2188 mutex_lock(&cgroup_mutex);
2189 for_each_active_root(root) {
2190 struct cgroup *from_cgrp = task_cgroup_from_root(from, root);
2192 retval = cgroup_attach_task(from_cgrp, tsk, false);
2196 mutex_unlock(&cgroup_mutex);
2200 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2202 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2203 struct cftype *cft, u64 pid)
2205 return attach_task_by_pid(css->cgroup, pid, false);
2208 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2209 struct cftype *cft, u64 tgid)
2211 return attach_task_by_pid(css->cgroup, tgid, true);
2214 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2215 struct cftype *cft, const char *buffer)
2217 BUILD_BUG_ON(sizeof(css->cgroup->root->release_agent_path) < PATH_MAX);
2218 if (strlen(buffer) >= PATH_MAX)
2220 if (!cgroup_lock_live_group(css->cgroup))
2222 mutex_lock(&cgroup_root_mutex);
2223 strcpy(css->cgroup->root->release_agent_path, buffer);
2224 mutex_unlock(&cgroup_root_mutex);
2225 mutex_unlock(&cgroup_mutex);
2229 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2231 struct cgroup *cgrp = seq_css(seq)->cgroup;
2233 if (!cgroup_lock_live_group(cgrp))
2235 seq_puts(seq, cgrp->root->release_agent_path);
2236 seq_putc(seq, '\n');
2237 mutex_unlock(&cgroup_mutex);
2241 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2243 struct cgroup *cgrp = seq_css(seq)->cgroup;
2245 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2249 /* A buffer size big enough for numbers or short strings */
2250 #define CGROUP_LOCAL_BUFFER_SIZE 64
2252 static ssize_t cgroup_file_write(struct file *file, const char __user *userbuf,
2253 size_t nbytes, loff_t *ppos)
2255 struct cfent *cfe = __d_cfe(file->f_dentry);
2256 struct cftype *cft = __d_cft(file->f_dentry);
2257 struct cgroup_subsys_state *css = cfe->css;
2258 size_t max_bytes = cft->max_write_len ?: CGROUP_LOCAL_BUFFER_SIZE - 1;
2262 if (nbytes >= max_bytes)
2265 buf = kmalloc(nbytes + 1, GFP_KERNEL);
2269 if (copy_from_user(buf, userbuf, nbytes)) {
2276 if (cft->write_string) {
2277 ret = cft->write_string(css, cft, strstrip(buf));
2278 } else if (cft->write_u64) {
2279 unsigned long long v;
2280 ret = kstrtoull(buf, 0, &v);
2282 ret = cft->write_u64(css, cft, v);
2283 } else if (cft->write_s64) {
2285 ret = kstrtoll(buf, 0, &v);
2287 ret = cft->write_s64(css, cft, v);
2288 } else if (cft->trigger) {
2289 ret = cft->trigger(css, (unsigned int)cft->private);
2295 return ret ?: nbytes;
2299 * seqfile ops/methods for returning structured data. Currently just
2300 * supports string->u64 maps, but can be extended in future.
2303 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2305 struct cftype *cft = seq_cft(seq);
2307 if (cft->seq_start) {
2308 return cft->seq_start(seq, ppos);
2311 * The same behavior and code as single_open(). Returns
2312 * !NULL if pos is at the beginning; otherwise, NULL.
2314 return NULL + !*ppos;
2318 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2320 struct cftype *cft = seq_cft(seq);
2322 if (cft->seq_next) {
2323 return cft->seq_next(seq, v, ppos);
2326 * The same behavior and code as single_open(), always
2327 * terminate after the initial read.
2334 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2336 struct cftype *cft = seq_cft(seq);
2339 cft->seq_stop(seq, v);
2342 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2344 struct cftype *cft = seq_cft(m);
2345 struct cgroup_subsys_state *css = seq_css(m);
2348 return cft->seq_show(m, arg);
2351 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2352 else if (cft->read_s64)
2353 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2359 static struct seq_operations cgroup_seq_operations = {
2360 .start = cgroup_seqfile_start,
2361 .next = cgroup_seqfile_next,
2362 .stop = cgroup_seqfile_stop,
2363 .show = cgroup_seqfile_show,
2366 static int cgroup_file_open(struct inode *inode, struct file *file)
2368 struct cfent *cfe = __d_cfe(file->f_dentry);
2369 struct cftype *cft = __d_cft(file->f_dentry);
2370 struct cgroup *cgrp = __d_cgrp(cfe->dentry->d_parent);
2371 struct cgroup_subsys_state *css;
2372 struct cgroup_open_file *of;
2375 err = generic_file_open(inode, file);
2380 * If the file belongs to a subsystem, pin the css. Will be
2381 * unpinned either on open failure or release. This ensures that
2382 * @css stays alive for all file operations.
2385 css = cgroup_css(cgrp, cft->ss);
2386 if (cft->ss && !css_tryget(css))
2394 * @cfe->css is used by read/write/close to determine the
2395 * associated css. @file->private_data would be a better place but
2396 * that's already used by seqfile. Multiple accessors may use it
2397 * simultaneously which is okay as the association never changes.
2399 WARN_ON_ONCE(cfe->css && cfe->css != css);
2402 of = __seq_open_private(file, &cgroup_seq_operations,
2403 sizeof(struct cgroup_open_file));
2414 static int cgroup_file_release(struct inode *inode, struct file *file)
2416 struct cfent *cfe = __d_cfe(file->f_dentry);
2417 struct cgroup_subsys_state *css = cfe->css;
2421 return seq_release_private(inode, file);
2425 * cgroup_rename - Only allow simple rename of directories in place.
2427 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2428 struct inode *new_dir, struct dentry *new_dentry)
2431 struct cgroup_name *name, *old_name;
2432 struct cgroup *cgrp;
2435 * It's convinient to use parent dir's i_mutex to protected
2438 lockdep_assert_held(&old_dir->i_mutex);
2440 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2442 if (new_dentry->d_inode)
2444 if (old_dir != new_dir)
2447 cgrp = __d_cgrp(old_dentry);
2450 * This isn't a proper migration and its usefulness is very
2451 * limited. Disallow if sane_behavior.
2453 if (cgroup_sane_behavior(cgrp))
2456 name = cgroup_alloc_name(new_dentry);
2460 ret = simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2466 old_name = rcu_dereference_protected(cgrp->name, true);
2467 rcu_assign_pointer(cgrp->name, name);
2469 kfree_rcu(old_name, rcu_head);
2473 static struct simple_xattrs *__d_xattrs(struct dentry *dentry)
2475 if (S_ISDIR(dentry->d_inode->i_mode))
2476 return &__d_cgrp(dentry)->xattrs;
2478 return &__d_cfe(dentry)->xattrs;
2481 static inline int xattr_enabled(struct dentry *dentry)
2483 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
2484 return root->flags & CGRP_ROOT_XATTR;
2487 static bool is_valid_xattr(const char *name)
2489 if (!strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
2490 !strncmp(name, XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN))
2495 static int cgroup_setxattr(struct dentry *dentry, const char *name,
2496 const void *val, size_t size, int flags)
2498 if (!xattr_enabled(dentry))
2500 if (!is_valid_xattr(name))
2502 return simple_xattr_set(__d_xattrs(dentry), name, val, size, flags);
2505 static int cgroup_removexattr(struct dentry *dentry, const char *name)
2507 if (!xattr_enabled(dentry))
2509 if (!is_valid_xattr(name))
2511 return simple_xattr_remove(__d_xattrs(dentry), name);
2514 static ssize_t cgroup_getxattr(struct dentry *dentry, const char *name,
2515 void *buf, size_t size)
2517 if (!xattr_enabled(dentry))
2519 if (!is_valid_xattr(name))
2521 return simple_xattr_get(__d_xattrs(dentry), name, buf, size);
2524 static ssize_t cgroup_listxattr(struct dentry *dentry, char *buf, size_t size)
2526 if (!xattr_enabled(dentry))
2528 return simple_xattr_list(__d_xattrs(dentry), buf, size);
2531 static const struct file_operations cgroup_file_operations = {
2533 .write = cgroup_file_write,
2534 .llseek = generic_file_llseek,
2535 .open = cgroup_file_open,
2536 .release = cgroup_file_release,
2539 static const struct inode_operations cgroup_file_inode_operations = {
2540 .setxattr = cgroup_setxattr,
2541 .getxattr = cgroup_getxattr,
2542 .listxattr = cgroup_listxattr,
2543 .removexattr = cgroup_removexattr,
2546 static const struct inode_operations cgroup_dir_inode_operations = {
2547 .lookup = simple_lookup,
2548 .mkdir = cgroup_mkdir,
2549 .rmdir = cgroup_rmdir,
2550 .rename = cgroup_rename,
2551 .setxattr = cgroup_setxattr,
2552 .getxattr = cgroup_getxattr,
2553 .listxattr = cgroup_listxattr,
2554 .removexattr = cgroup_removexattr,
2557 static int cgroup_create_file(struct dentry *dentry, umode_t mode,
2558 struct super_block *sb)
2560 struct inode *inode;
2564 if (dentry->d_inode)
2567 inode = cgroup_new_inode(mode, sb);
2571 if (S_ISDIR(mode)) {
2572 inode->i_op = &cgroup_dir_inode_operations;
2573 inode->i_fop = &simple_dir_operations;
2575 /* start off with i_nlink == 2 (for "." entry) */
2577 inc_nlink(dentry->d_parent->d_inode);
2580 * Control reaches here with cgroup_mutex held.
2581 * @inode->i_mutex should nest outside cgroup_mutex but we
2582 * want to populate it immediately without releasing
2583 * cgroup_mutex. As @inode isn't visible to anyone else
2584 * yet, trylock will always succeed without affecting
2587 WARN_ON_ONCE(!mutex_trylock(&inode->i_mutex));
2588 } else if (S_ISREG(mode)) {
2590 inode->i_fop = &cgroup_file_operations;
2591 inode->i_op = &cgroup_file_inode_operations;
2593 d_instantiate(dentry, inode);
2594 dget(dentry); /* Extra count - pin the dentry in core */
2599 * cgroup_file_mode - deduce file mode of a control file
2600 * @cft: the control file in question
2602 * returns cft->mode if ->mode is not 0
2603 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2604 * returns S_IRUGO if it has only a read handler
2605 * returns S_IWUSR if it has only a write hander
2607 static umode_t cgroup_file_mode(const struct cftype *cft)
2614 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
2617 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
2624 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2626 struct dentry *dir = cgrp->dentry;
2627 struct cgroup *parent = __d_cgrp(dir);
2628 struct dentry *dentry;
2632 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2634 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
2635 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
2636 strcpy(name, cft->ss->name);
2639 strcat(name, cft->name);
2641 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2643 cfe = kzalloc(sizeof(*cfe), GFP_KERNEL);
2647 dentry = lookup_one_len(name, dir, strlen(name));
2648 if (IS_ERR(dentry)) {
2649 error = PTR_ERR(dentry);
2653 cfe->type = (void *)cft;
2654 cfe->dentry = dentry;
2655 dentry->d_fsdata = cfe;
2656 simple_xattrs_init(&cfe->xattrs);
2658 mode = cgroup_file_mode(cft);
2659 error = cgroup_create_file(dentry, mode | S_IFREG, cgrp->root->sb);
2661 list_add_tail(&cfe->node, &parent->files);
2671 * cgroup_addrm_files - add or remove files to a cgroup directory
2672 * @cgrp: the target cgroup
2673 * @cfts: array of cftypes to be added
2674 * @is_add: whether to add or remove
2676 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2677 * For removals, this function never fails. If addition fails, this
2678 * function doesn't remove files already added. The caller is responsible
2681 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2687 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
2688 lockdep_assert_held(&cgroup_mutex);
2690 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2691 /* does cft->flags tell us to skip this file on @cgrp? */
2692 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2694 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2696 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2700 ret = cgroup_add_file(cgrp, cft);
2702 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2707 cgroup_rm_file(cgrp, cft);
2713 static void cgroup_cfts_prepare(void)
2714 __acquires(&cgroup_mutex)
2717 * Thanks to the entanglement with vfs inode locking, we can't walk
2718 * the existing cgroups under cgroup_mutex and create files.
2719 * Instead, we use css_for_each_descendant_pre() and drop RCU read
2720 * lock before calling cgroup_addrm_files().
2722 mutex_lock(&cgroup_mutex);
2725 static int cgroup_cfts_commit(struct cftype *cfts, bool is_add)
2726 __releases(&cgroup_mutex)
2729 struct cgroup_subsys *ss = cfts[0].ss;
2730 struct cgroup *root = &ss->root->top_cgroup;
2731 struct super_block *sb = ss->root->sb;
2732 struct dentry *prev = NULL;
2733 struct inode *inode;
2734 struct cgroup_subsys_state *css;
2738 /* %NULL @cfts indicates abort and don't bother if @ss isn't attached */
2739 if (!cfts || ss->root == &cgroup_dummy_root ||
2740 !atomic_inc_not_zero(&sb->s_active)) {
2741 mutex_unlock(&cgroup_mutex);
2746 * All cgroups which are created after we drop cgroup_mutex will
2747 * have the updated set of files, so we only need to update the
2748 * cgroups created before the current @cgroup_serial_nr_next.
2750 update_before = cgroup_serial_nr_next;
2752 mutex_unlock(&cgroup_mutex);
2754 /* add/rm files for all cgroups created before */
2756 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2757 struct cgroup *cgrp = css->cgroup;
2759 if (cgroup_is_dead(cgrp))
2762 inode = cgrp->dentry->d_inode;
2767 prev = cgrp->dentry;
2769 mutex_lock(&inode->i_mutex);
2770 mutex_lock(&cgroup_mutex);
2771 if (cgrp->serial_nr < update_before && !cgroup_is_dead(cgrp))
2772 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2773 mutex_unlock(&cgroup_mutex);
2774 mutex_unlock(&inode->i_mutex);
2782 deactivate_super(sb);
2787 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2788 * @ss: target cgroup subsystem
2789 * @cfts: zero-length name terminated array of cftypes
2791 * Register @cfts to @ss. Files described by @cfts are created for all
2792 * existing cgroups to which @ss is attached and all future cgroups will
2793 * have them too. This function can be called anytime whether @ss is
2796 * Returns 0 on successful registration, -errno on failure. Note that this
2797 * function currently returns 0 as long as @cfts registration is successful
2798 * even if some file creation attempts on existing cgroups fail.
2800 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2802 struct cftype_set *set;
2806 set = kzalloc(sizeof(*set), GFP_KERNEL);
2810 for (cft = cfts; cft->name[0] != '\0'; cft++)
2813 cgroup_cfts_prepare();
2815 list_add_tail(&set->node, &ss->cftsets);
2816 ret = cgroup_cfts_commit(cfts, true);
2818 cgroup_rm_cftypes(cfts);
2821 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2824 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2825 * @cfts: zero-length name terminated array of cftypes
2827 * Unregister @cfts. Files described by @cfts are removed from all
2828 * existing cgroups and all future cgroups won't have them either. This
2829 * function can be called anytime whether @cfts' subsys is attached or not.
2831 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2834 int cgroup_rm_cftypes(struct cftype *cfts)
2836 struct cftype_set *set;
2838 if (!cfts || !cfts[0].ss)
2841 cgroup_cfts_prepare();
2843 list_for_each_entry(set, &cfts[0].ss->cftsets, node) {
2844 if (set->cfts == cfts) {
2845 list_del(&set->node);
2847 cgroup_cfts_commit(cfts, false);
2852 cgroup_cfts_commit(NULL, false);
2857 * cgroup_task_count - count the number of tasks in a cgroup.
2858 * @cgrp: the cgroup in question
2860 * Return the number of tasks in the cgroup.
2862 int cgroup_task_count(const struct cgroup *cgrp)
2865 struct cgrp_cset_link *link;
2867 read_lock(&css_set_lock);
2868 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2869 count += atomic_read(&link->cset->refcount);
2870 read_unlock(&css_set_lock);
2875 * To reduce the fork() overhead for systems that are not actually using
2876 * their cgroups capability, we don't maintain the lists running through
2877 * each css_set to its tasks until we see the list actually used - in other
2878 * words after the first call to css_task_iter_start().
2880 static void cgroup_enable_task_cg_lists(void)
2882 struct task_struct *p, *g;
2883 write_lock(&css_set_lock);
2884 use_task_css_set_links = 1;
2886 * We need tasklist_lock because RCU is not safe against
2887 * while_each_thread(). Besides, a forking task that has passed
2888 * cgroup_post_fork() without seeing use_task_css_set_links = 1
2889 * is not guaranteed to have its child immediately visible in the
2890 * tasklist if we walk through it with RCU.
2892 read_lock(&tasklist_lock);
2893 do_each_thread(g, p) {
2896 * We should check if the process is exiting, otherwise
2897 * it will race with cgroup_exit() in that the list
2898 * entry won't be deleted though the process has exited.
2900 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2901 list_add(&p->cg_list, &task_css_set(p)->tasks);
2903 } while_each_thread(g, p);
2904 read_unlock(&tasklist_lock);
2905 write_unlock(&css_set_lock);
2909 * css_next_child - find the next child of a given css
2910 * @pos_css: the current position (%NULL to initiate traversal)
2911 * @parent_css: css whose children to walk
2913 * This function returns the next child of @parent_css and should be called
2914 * under either cgroup_mutex or RCU read lock. The only requirement is
2915 * that @parent_css and @pos_css are accessible. The next sibling is
2916 * guaranteed to be returned regardless of their states.
2918 struct cgroup_subsys_state *
2919 css_next_child(struct cgroup_subsys_state *pos_css,
2920 struct cgroup_subsys_state *parent_css)
2922 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2923 struct cgroup *cgrp = parent_css->cgroup;
2924 struct cgroup *next;
2926 cgroup_assert_mutex_or_rcu_locked();
2929 * @pos could already have been removed. Once a cgroup is removed,
2930 * its ->sibling.next is no longer updated when its next sibling
2931 * changes. As CGRP_DEAD assertion is serialized and happens
2932 * before the cgroup is taken off the ->sibling list, if we see it
2933 * unasserted, it's guaranteed that the next sibling hasn't
2934 * finished its grace period even if it's already removed, and thus
2935 * safe to dereference from this RCU critical section. If
2936 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2937 * to be visible as %true here.
2939 * If @pos is dead, its next pointer can't be dereferenced;
2940 * however, as each cgroup is given a monotonically increasing
2941 * unique serial number and always appended to the sibling list,
2942 * the next one can be found by walking the parent's children until
2943 * we see a cgroup with higher serial number than @pos's. While
2944 * this path can be slower, it's taken only when either the current
2945 * cgroup is removed or iteration and removal race.
2948 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2949 } else if (likely(!cgroup_is_dead(pos))) {
2950 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2952 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2953 if (next->serial_nr > pos->serial_nr)
2957 if (&next->sibling == &cgrp->children)
2960 return cgroup_css(next, parent_css->ss);
2962 EXPORT_SYMBOL_GPL(css_next_child);
2965 * css_next_descendant_pre - find the next descendant for pre-order walk
2966 * @pos: the current position (%NULL to initiate traversal)
2967 * @root: css whose descendants to walk
2969 * To be used by css_for_each_descendant_pre(). Find the next descendant
2970 * to visit for pre-order traversal of @root's descendants. @root is
2971 * included in the iteration and the first node to be visited.
2973 * While this function requires cgroup_mutex or RCU read locking, it
2974 * doesn't require the whole traversal to be contained in a single critical
2975 * section. This function will return the correct next descendant as long
2976 * as both @pos and @root are accessible and @pos is a descendant of @root.
2978 struct cgroup_subsys_state *
2979 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2980 struct cgroup_subsys_state *root)
2982 struct cgroup_subsys_state *next;
2984 cgroup_assert_mutex_or_rcu_locked();
2986 /* if first iteration, visit @root */
2990 /* visit the first child if exists */
2991 next = css_next_child(NULL, pos);
2995 /* no child, visit my or the closest ancestor's next sibling */
2996 while (pos != root) {
2997 next = css_next_child(pos, css_parent(pos));
3000 pos = css_parent(pos);
3005 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
3008 * css_rightmost_descendant - return the rightmost descendant of a css
3009 * @pos: css of interest
3011 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3012 * is returned. This can be used during pre-order traversal to skip
3015 * While this function requires cgroup_mutex or RCU read locking, it
3016 * doesn't require the whole traversal to be contained in a single critical
3017 * section. This function will return the correct rightmost descendant as
3018 * long as @pos is accessible.
3020 struct cgroup_subsys_state *
3021 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3023 struct cgroup_subsys_state *last, *tmp;
3025 cgroup_assert_mutex_or_rcu_locked();
3029 /* ->prev isn't RCU safe, walk ->next till the end */
3031 css_for_each_child(tmp, last)
3037 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
3039 static struct cgroup_subsys_state *
3040 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3042 struct cgroup_subsys_state *last;
3046 pos = css_next_child(NULL, pos);
3053 * css_next_descendant_post - find the next descendant for post-order walk
3054 * @pos: the current position (%NULL to initiate traversal)
3055 * @root: css whose descendants to walk
3057 * To be used by css_for_each_descendant_post(). Find the next descendant
3058 * to visit for post-order traversal of @root's descendants. @root is
3059 * included in the iteration and the last node to be visited.
3061 * While this function requires cgroup_mutex or RCU read locking, it
3062 * doesn't require the whole traversal to be contained in a single critical
3063 * section. This function will return the correct next descendant as long
3064 * as both @pos and @cgroup are accessible and @pos is a descendant of
3067 struct cgroup_subsys_state *
3068 css_next_descendant_post(struct cgroup_subsys_state *pos,
3069 struct cgroup_subsys_state *root)
3071 struct cgroup_subsys_state *next;
3073 cgroup_assert_mutex_or_rcu_locked();
3075 /* if first iteration, visit leftmost descendant which may be @root */
3077 return css_leftmost_descendant(root);
3079 /* if we visited @root, we're done */
3083 /* if there's an unvisited sibling, visit its leftmost descendant */
3084 next = css_next_child(pos, css_parent(pos));
3086 return css_leftmost_descendant(next);
3088 /* no sibling left, visit parent */
3089 return css_parent(pos);
3091 EXPORT_SYMBOL_GPL(css_next_descendant_post);
3094 * css_advance_task_iter - advance a task itererator to the next css_set
3095 * @it: the iterator to advance
3097 * Advance @it to the next css_set to walk.
3099 static void css_advance_task_iter(struct css_task_iter *it)
3101 struct list_head *l = it->cset_link;
3102 struct cgrp_cset_link *link;
3103 struct css_set *cset;
3105 /* Advance to the next non-empty css_set */
3108 if (l == &it->origin_css->cgroup->cset_links) {
3109 it->cset_link = NULL;
3112 link = list_entry(l, struct cgrp_cset_link, cset_link);
3114 } while (list_empty(&cset->tasks));
3116 it->task = cset->tasks.next;
3120 * css_task_iter_start - initiate task iteration
3121 * @css: the css to walk tasks of
3122 * @it: the task iterator to use
3124 * Initiate iteration through the tasks of @css. The caller can call
3125 * css_task_iter_next() to walk through the tasks until the function
3126 * returns NULL. On completion of iteration, css_task_iter_end() must be
3129 * Note that this function acquires a lock which is released when the
3130 * iteration finishes. The caller can't sleep while iteration is in
3133 void css_task_iter_start(struct cgroup_subsys_state *css,
3134 struct css_task_iter *it)
3135 __acquires(css_set_lock)
3138 * The first time anyone tries to iterate across a css, we need to
3139 * enable the list linking each css_set to its tasks, and fix up
3140 * all existing tasks.
3142 if (!use_task_css_set_links)
3143 cgroup_enable_task_cg_lists();
3145 read_lock(&css_set_lock);
3147 it->origin_css = css;
3148 it->cset_link = &css->cgroup->cset_links;
3150 css_advance_task_iter(it);
3154 * css_task_iter_next - return the next task for the iterator
3155 * @it: the task iterator being iterated
3157 * The "next" function for task iteration. @it should have been
3158 * initialized via css_task_iter_start(). Returns NULL when the iteration
3161 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3163 struct task_struct *res;
3164 struct list_head *l = it->task;
3165 struct cgrp_cset_link *link;
3167 /* If the iterator cg is NULL, we have no tasks */
3170 res = list_entry(l, struct task_struct, cg_list);
3171 /* Advance iterator to find next entry */
3173 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
3174 if (l == &link->cset->tasks) {
3176 * We reached the end of this task list - move on to the
3177 * next cgrp_cset_link.
3179 css_advance_task_iter(it);
3187 * css_task_iter_end - finish task iteration
3188 * @it: the task iterator to finish
3190 * Finish task iteration started by css_task_iter_start().
3192 void css_task_iter_end(struct css_task_iter *it)
3193 __releases(css_set_lock)
3195 read_unlock(&css_set_lock);
3198 static inline int started_after_time(struct task_struct *t1,
3199 struct timespec *time,
3200 struct task_struct *t2)
3202 int start_diff = timespec_compare(&t1->start_time, time);
3203 if (start_diff > 0) {
3205 } else if (start_diff < 0) {
3209 * Arbitrarily, if two processes started at the same
3210 * time, we'll say that the lower pointer value
3211 * started first. Note that t2 may have exited by now
3212 * so this may not be a valid pointer any longer, but
3213 * that's fine - it still serves to distinguish
3214 * between two tasks started (effectively) simultaneously.
3221 * This function is a callback from heap_insert() and is used to order
3223 * In this case we order the heap in descending task start time.
3225 static inline int started_after(void *p1, void *p2)
3227 struct task_struct *t1 = p1;
3228 struct task_struct *t2 = p2;
3229 return started_after_time(t1, &t2->start_time, t2);
3233 * css_scan_tasks - iterate though all the tasks in a css
3234 * @css: the css to iterate tasks of
3235 * @test: optional test callback
3236 * @process: process callback
3237 * @data: data passed to @test and @process
3238 * @heap: optional pre-allocated heap used for task iteration
3240 * Iterate through all the tasks in @css, calling @test for each, and if it
3241 * returns %true, call @process for it also.
3243 * @test may be NULL, meaning always true (select all tasks), which
3244 * effectively duplicates css_task_iter_{start,next,end}() but does not
3245 * lock css_set_lock for the call to @process.
3247 * It is guaranteed that @process will act on every task that is a member
3248 * of @css for the duration of this call. This function may or may not
3249 * call @process for tasks that exit or move to a different css during the
3250 * call, or are forked or move into the css during the call.
3252 * Note that @test may be called with locks held, and may in some
3253 * situations be called multiple times for the same task, so it should be
3256 * If @heap is non-NULL, a heap has been pre-allocated and will be used for
3257 * heap operations (and its "gt" member will be overwritten), else a
3258 * temporary heap will be used (allocation of which may cause this function
3261 int css_scan_tasks(struct cgroup_subsys_state *css,
3262 bool (*test)(struct task_struct *, void *),
3263 void (*process)(struct task_struct *, void *),
3264 void *data, struct ptr_heap *heap)
3267 struct css_task_iter it;
3268 struct task_struct *p, *dropped;
3269 /* Never dereference latest_task, since it's not refcounted */
3270 struct task_struct *latest_task = NULL;
3271 struct ptr_heap tmp_heap;
3272 struct timespec latest_time = { 0, 0 };
3275 /* The caller supplied our heap and pre-allocated its memory */
3276 heap->gt = &started_after;
3278 /* We need to allocate our own heap memory */
3280 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
3282 /* cannot allocate the heap */
3288 * Scan tasks in the css, using the @test callback to determine
3289 * which are of interest, and invoking @process callback on the
3290 * ones which need an update. Since we don't want to hold any
3291 * locks during the task updates, gather tasks to be processed in a
3292 * heap structure. The heap is sorted by descending task start
3293 * time. If the statically-sized heap fills up, we overflow tasks
3294 * that started later, and in future iterations only consider tasks
3295 * that started after the latest task in the previous pass. This
3296 * guarantees forward progress and that we don't miss any tasks.
3299 css_task_iter_start(css, &it);
3300 while ((p = css_task_iter_next(&it))) {
3302 * Only affect tasks that qualify per the caller's callback,
3303 * if he provided one
3305 if (test && !test(p, data))
3308 * Only process tasks that started after the last task
3311 if (!started_after_time(p, &latest_time, latest_task))
3313 dropped = heap_insert(heap, p);
3314 if (dropped == NULL) {
3316 * The new task was inserted; the heap wasn't
3320 } else if (dropped != p) {
3322 * The new task was inserted, and pushed out a
3326 put_task_struct(dropped);
3329 * Else the new task was newer than anything already in
3330 * the heap and wasn't inserted
3333 css_task_iter_end(&it);
3336 for (i = 0; i < heap->size; i++) {
3337 struct task_struct *q = heap->ptrs[i];
3339 latest_time = q->start_time;
3342 /* Process the task per the caller's callback */
3347 * If we had to process any tasks at all, scan again
3348 * in case some of them were in the middle of forking
3349 * children that didn't get processed.
3350 * Not the most efficient way to do it, but it avoids
3351 * having to take callback_mutex in the fork path
3355 if (heap == &tmp_heap)
3356 heap_free(&tmp_heap);
3360 static void cgroup_transfer_one_task(struct task_struct *task, void *data)
3362 struct cgroup *new_cgroup = data;
3364 mutex_lock(&cgroup_mutex);
3365 cgroup_attach_task(new_cgroup, task, false);
3366 mutex_unlock(&cgroup_mutex);
3370 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3371 * @to: cgroup to which the tasks will be moved
3372 * @from: cgroup in which the tasks currently reside
3374 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3376 return css_scan_tasks(&from->dummy_css, NULL, cgroup_transfer_one_task,
3381 * Stuff for reading the 'tasks'/'procs' files.
3383 * Reading this file can return large amounts of data if a cgroup has
3384 * *lots* of attached tasks. So it may need several calls to read(),
3385 * but we cannot guarantee that the information we produce is correct
3386 * unless we produce it entirely atomically.
3390 /* which pidlist file are we talking about? */
3391 enum cgroup_filetype {
3397 * A pidlist is a list of pids that virtually represents the contents of one
3398 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3399 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3402 struct cgroup_pidlist {
3404 * used to find which pidlist is wanted. doesn't change as long as
3405 * this particular list stays in the list.
3407 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3410 /* how many elements the above list has */
3412 /* each of these stored in a list by its cgroup */
3413 struct list_head links;
3414 /* pointer to the cgroup we belong to, for list removal purposes */
3415 struct cgroup *owner;
3416 /* for delayed destruction */
3417 struct delayed_work destroy_dwork;
3421 * The following two functions "fix" the issue where there are more pids
3422 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3423 * TODO: replace with a kernel-wide solution to this problem
3425 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3426 static void *pidlist_allocate(int count)
3428 if (PIDLIST_TOO_LARGE(count))
3429 return vmalloc(count * sizeof(pid_t));
3431 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3434 static void pidlist_free(void *p)
3436 if (is_vmalloc_addr(p))
3443 * Used to destroy all pidlists lingering waiting for destroy timer. None
3444 * should be left afterwards.
3446 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3448 struct cgroup_pidlist *l, *tmp_l;
3450 mutex_lock(&cgrp->pidlist_mutex);
3451 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3452 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3453 mutex_unlock(&cgrp->pidlist_mutex);
3455 flush_workqueue(cgroup_pidlist_destroy_wq);
3456 BUG_ON(!list_empty(&cgrp->pidlists));
3459 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3461 struct delayed_work *dwork = to_delayed_work(work);
3462 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3464 struct cgroup_pidlist *tofree = NULL;
3466 mutex_lock(&l->owner->pidlist_mutex);
3469 * Destroy iff we didn't get queued again. The state won't change
3470 * as destroy_dwork can only be queued while locked.
3472 if (!delayed_work_pending(dwork)) {
3473 list_del(&l->links);
3474 pidlist_free(l->list);
3475 put_pid_ns(l->key.ns);
3479 mutex_unlock(&l->owner->pidlist_mutex);
3484 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3485 * Returns the number of unique elements.
3487 static int pidlist_uniq(pid_t *list, int length)
3492 * we presume the 0th element is unique, so i starts at 1. trivial
3493 * edge cases first; no work needs to be done for either
3495 if (length == 0 || length == 1)
3497 /* src and dest walk down the list; dest counts unique elements */
3498 for (src = 1; src < length; src++) {
3499 /* find next unique element */
3500 while (list[src] == list[src-1]) {
3505 /* dest always points to where the next unique element goes */
3506 list[dest] = list[src];
3514 * The two pid files - task and cgroup.procs - guaranteed that the result
3515 * is sorted, which forced this whole pidlist fiasco. As pid order is
3516 * different per namespace, each namespace needs differently sorted list,
3517 * making it impossible to use, for example, single rbtree of member tasks
3518 * sorted by task pointer. As pidlists can be fairly large, allocating one
3519 * per open file is dangerous, so cgroup had to implement shared pool of
3520 * pidlists keyed by cgroup and namespace.
3522 * All this extra complexity was caused by the original implementation
3523 * committing to an entirely unnecessary property. In the long term, we
3524 * want to do away with it. Explicitly scramble sort order if
3525 * sane_behavior so that no such expectation exists in the new interface.
3527 * Scrambling is done by swapping every two consecutive bits, which is
3528 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3530 static pid_t pid_fry(pid_t pid)
3532 unsigned a = pid & 0x55555555;
3533 unsigned b = pid & 0xAAAAAAAA;
3535 return (a << 1) | (b >> 1);
3538 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3540 if (cgroup_sane_behavior(cgrp))
3541 return pid_fry(pid);
3546 static int cmppid(const void *a, const void *b)
3548 return *(pid_t *)a - *(pid_t *)b;
3551 static int fried_cmppid(const void *a, const void *b)
3553 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3556 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3557 enum cgroup_filetype type)
3559 struct cgroup_pidlist *l;
3560 /* don't need task_nsproxy() if we're looking at ourself */
3561 struct pid_namespace *ns = task_active_pid_ns(current);
3563 lockdep_assert_held(&cgrp->pidlist_mutex);
3565 list_for_each_entry(l, &cgrp->pidlists, links)
3566 if (l->key.type == type && l->key.ns == ns)
3572 * find the appropriate pidlist for our purpose (given procs vs tasks)
3573 * returns with the lock on that pidlist already held, and takes care
3574 * of the use count, or returns NULL with no locks held if we're out of
3577 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3578 enum cgroup_filetype type)
3580 struct cgroup_pidlist *l;
3582 lockdep_assert_held(&cgrp->pidlist_mutex);
3584 l = cgroup_pidlist_find(cgrp, type);
3588 /* entry not found; create a new one */
3589 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3593 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3595 /* don't need task_nsproxy() if we're looking at ourself */
3596 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3598 list_add(&l->links, &cgrp->pidlists);
3603 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3605 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3606 struct cgroup_pidlist **lp)
3610 int pid, n = 0; /* used for populating the array */
3611 struct css_task_iter it;
3612 struct task_struct *tsk;
3613 struct cgroup_pidlist *l;
3615 lockdep_assert_held(&cgrp->pidlist_mutex);
3618 * If cgroup gets more users after we read count, we won't have
3619 * enough space - tough. This race is indistinguishable to the
3620 * caller from the case that the additional cgroup users didn't
3621 * show up until sometime later on.
3623 length = cgroup_task_count(cgrp);
3624 array = pidlist_allocate(length);
3627 /* now, populate the array */
3628 css_task_iter_start(&cgrp->dummy_css, &it);
3629 while ((tsk = css_task_iter_next(&it))) {
3630 if (unlikely(n == length))
3632 /* get tgid or pid for procs or tasks file respectively */
3633 if (type == CGROUP_FILE_PROCS)
3634 pid = task_tgid_vnr(tsk);
3636 pid = task_pid_vnr(tsk);
3637 if (pid > 0) /* make sure to only use valid results */
3640 css_task_iter_end(&it);
3642 /* now sort & (if procs) strip out duplicates */
3643 if (cgroup_sane_behavior(cgrp))
3644 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3646 sort(array, length, sizeof(pid_t), cmppid, NULL);
3647 if (type == CGROUP_FILE_PROCS)
3648 length = pidlist_uniq(array, length);
3650 l = cgroup_pidlist_find_create(cgrp, type);
3652 mutex_unlock(&cgrp->pidlist_mutex);
3653 pidlist_free(array);
3657 /* store array, freeing old if necessary */
3658 pidlist_free(l->list);
3666 * cgroupstats_build - build and fill cgroupstats
3667 * @stats: cgroupstats to fill information into
3668 * @dentry: A dentry entry belonging to the cgroup for which stats have
3671 * Build and fill cgroupstats so that taskstats can export it to user
3674 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3677 struct cgroup *cgrp;
3678 struct css_task_iter it;
3679 struct task_struct *tsk;
3682 * Validate dentry by checking the superblock operations,
3683 * and make sure it's a directory.
3685 if (dentry->d_sb->s_op != &cgroup_ops ||
3686 !S_ISDIR(dentry->d_inode->i_mode))
3690 cgrp = dentry->d_fsdata;
3692 css_task_iter_start(&cgrp->dummy_css, &it);
3693 while ((tsk = css_task_iter_next(&it))) {
3694 switch (tsk->state) {
3696 stats->nr_running++;
3698 case TASK_INTERRUPTIBLE:
3699 stats->nr_sleeping++;
3701 case TASK_UNINTERRUPTIBLE:
3702 stats->nr_uninterruptible++;
3705 stats->nr_stopped++;
3708 if (delayacct_is_task_waiting_on_io(tsk))
3709 stats->nr_io_wait++;
3713 css_task_iter_end(&it);
3721 * seq_file methods for the tasks/procs files. The seq_file position is the
3722 * next pid to display; the seq_file iterator is a pointer to the pid
3723 * in the cgroup->l->list array.
3726 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3729 * Initially we receive a position value that corresponds to
3730 * one more than the last pid shown (or 0 on the first call or
3731 * after a seek to the start). Use a binary-search to find the
3732 * next pid to display, if any
3734 struct cgroup_open_file *of = s->private;
3735 struct cgroup *cgrp = seq_css(s)->cgroup;
3736 struct cgroup_pidlist *l;
3737 enum cgroup_filetype type = seq_cft(s)->private;
3738 int index = 0, pid = *pos;
3741 mutex_lock(&cgrp->pidlist_mutex);
3744 * !NULL @of->priv indicates that this isn't the first start()
3745 * after open. If the matching pidlist is around, we can use that.
3746 * Look for it. Note that @of->priv can't be used directly. It
3747 * could already have been destroyed.
3750 of->priv = cgroup_pidlist_find(cgrp, type);
3753 * Either this is the first start() after open or the matching
3754 * pidlist has been destroyed inbetween. Create a new one.
3757 ret = pidlist_array_load(cgrp, type,
3758 (struct cgroup_pidlist **)&of->priv);
3760 return ERR_PTR(ret);
3765 int end = l->length;
3767 while (index < end) {
3768 int mid = (index + end) / 2;
3769 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3772 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3778 /* If we're off the end of the array, we're done */
3779 if (index >= l->length)
3781 /* Update the abstract position to be the actual pid that we found */
3782 iter = l->list + index;
3783 *pos = cgroup_pid_fry(cgrp, *iter);
3787 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3789 struct cgroup_open_file *of = s->private;
3790 struct cgroup_pidlist *l = of->priv;
3793 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3794 CGROUP_PIDLIST_DESTROY_DELAY);
3795 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3798 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3800 struct cgroup_open_file *of = s->private;
3801 struct cgroup_pidlist *l = of->priv;
3803 pid_t *end = l->list + l->length;
3805 * Advance to the next pid in the array. If this goes off the
3812 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3817 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3819 return seq_printf(s, "%d\n", *(int *)v);
3823 * seq_operations functions for iterating on pidlists through seq_file -
3824 * independent of whether it's tasks or procs
3826 static const struct seq_operations cgroup_pidlist_seq_operations = {
3827 .start = cgroup_pidlist_start,
3828 .stop = cgroup_pidlist_stop,
3829 .next = cgroup_pidlist_next,
3830 .show = cgroup_pidlist_show,
3833 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3836 return notify_on_release(css->cgroup);
3839 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3840 struct cftype *cft, u64 val)
3842 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3844 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3846 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3851 * When dput() is called asynchronously, if umount has been done and
3852 * then deactivate_super() in cgroup_free_fn() kills the superblock,
3853 * there's a small window that vfs will see the root dentry with non-zero
3854 * refcnt and trigger BUG().
3856 * That's why we hold a reference before dput() and drop it right after.
3858 static void cgroup_dput(struct cgroup *cgrp)
3860 struct super_block *sb = cgrp->root->sb;
3862 atomic_inc(&sb->s_active);
3864 deactivate_super(sb);
3867 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3870 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3873 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3874 struct cftype *cft, u64 val)
3877 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3879 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3883 static struct cftype cgroup_base_files[] = {
3885 .name = "cgroup.procs",
3886 .seq_start = cgroup_pidlist_start,
3887 .seq_next = cgroup_pidlist_next,
3888 .seq_stop = cgroup_pidlist_stop,
3889 .seq_show = cgroup_pidlist_show,
3890 .private = CGROUP_FILE_PROCS,
3891 .write_u64 = cgroup_procs_write,
3892 .mode = S_IRUGO | S_IWUSR,
3895 .name = "cgroup.clone_children",
3896 .flags = CFTYPE_INSANE,
3897 .read_u64 = cgroup_clone_children_read,
3898 .write_u64 = cgroup_clone_children_write,
3901 .name = "cgroup.sane_behavior",
3902 .flags = CFTYPE_ONLY_ON_ROOT,
3903 .seq_show = cgroup_sane_behavior_show,
3907 * Historical crazy stuff. These don't have "cgroup." prefix and
3908 * don't exist if sane_behavior. If you're depending on these, be
3909 * prepared to be burned.
3913 .flags = CFTYPE_INSANE, /* use "procs" instead */
3914 .seq_start = cgroup_pidlist_start,
3915 .seq_next = cgroup_pidlist_next,
3916 .seq_stop = cgroup_pidlist_stop,
3917 .seq_show = cgroup_pidlist_show,
3918 .private = CGROUP_FILE_TASKS,
3919 .write_u64 = cgroup_tasks_write,
3920 .mode = S_IRUGO | S_IWUSR,
3923 .name = "notify_on_release",
3924 .flags = CFTYPE_INSANE,
3925 .read_u64 = cgroup_read_notify_on_release,
3926 .write_u64 = cgroup_write_notify_on_release,
3929 .name = "release_agent",
3930 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3931 .seq_show = cgroup_release_agent_show,
3932 .write_string = cgroup_release_agent_write,
3933 .max_write_len = PATH_MAX,
3939 * cgroup_populate_dir - create subsys files in a cgroup directory
3940 * @cgrp: target cgroup
3941 * @subsys_mask: mask of the subsystem ids whose files should be added
3943 * On failure, no file is added.
3945 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3947 struct cgroup_subsys *ss;
3950 /* process cftsets of each subsystem */
3951 for_each_subsys(ss, i) {
3952 struct cftype_set *set;
3954 if (!test_bit(i, &subsys_mask))
3957 list_for_each_entry(set, &ss->cftsets, node) {
3958 ret = cgroup_addrm_files(cgrp, set->cfts, true);
3965 cgroup_clear_dir(cgrp, subsys_mask);
3970 * css destruction is four-stage process.
3972 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3973 * Implemented in kill_css().
3975 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3976 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3977 * by invoking offline_css(). After offlining, the base ref is put.
3978 * Implemented in css_killed_work_fn().
3980 * 3. When the percpu_ref reaches zero, the only possible remaining
3981 * accessors are inside RCU read sections. css_release() schedules the
3984 * 4. After the grace period, the css can be freed. Implemented in
3985 * css_free_work_fn().
3987 * It is actually hairier because both step 2 and 4 require process context
3988 * and thus involve punting to css->destroy_work adding two additional
3989 * steps to the already complex sequence.
3991 static void css_free_work_fn(struct work_struct *work)
3993 struct cgroup_subsys_state *css =
3994 container_of(work, struct cgroup_subsys_state, destroy_work);
3995 struct cgroup *cgrp = css->cgroup;
3998 css_put(css->parent);
4000 css->ss->css_free(css);
4004 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4006 struct cgroup_subsys_state *css =
4007 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4010 * css holds an extra ref to @cgrp->dentry which is put on the last
4011 * css_put(). dput() requires process context which we don't have.
4013 INIT_WORK(&css->destroy_work, css_free_work_fn);
4014 queue_work(cgroup_destroy_wq, &css->destroy_work);
4017 static void css_release(struct percpu_ref *ref)
4019 struct cgroup_subsys_state *css =
4020 container_of(ref, struct cgroup_subsys_state, refcnt);
4022 call_rcu(&css->rcu_head, css_free_rcu_fn);
4025 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
4026 struct cgroup *cgrp)
4033 css->parent = cgroup_css(cgrp->parent, ss);
4035 css->flags |= CSS_ROOT;
4037 BUG_ON(cgroup_css(cgrp, ss));
4040 /* invoke ->css_online() on a new CSS and mark it online if successful */
4041 static int online_css(struct cgroup_subsys_state *css)
4043 struct cgroup_subsys *ss = css->ss;
4046 lockdep_assert_held(&cgroup_mutex);
4049 ret = ss->css_online(css);
4051 css->flags |= CSS_ONLINE;
4052 css->cgroup->nr_css++;
4053 rcu_assign_pointer(css->cgroup->subsys[ss->subsys_id], css);
4058 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4059 static void offline_css(struct cgroup_subsys_state *css)
4061 struct cgroup_subsys *ss = css->ss;
4063 lockdep_assert_held(&cgroup_mutex);
4065 if (!(css->flags & CSS_ONLINE))
4068 if (ss->css_offline)
4069 ss->css_offline(css);
4071 css->flags &= ~CSS_ONLINE;
4072 css->cgroup->nr_css--;
4073 RCU_INIT_POINTER(css->cgroup->subsys[ss->subsys_id], css);
4077 * cgroup_create - create a cgroup
4078 * @parent: cgroup that will be parent of the new cgroup
4079 * @dentry: dentry of the new cgroup
4080 * @mode: mode to set on new inode
4082 * Must be called with the mutex on the parent inode held
4084 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
4087 struct cgroup_subsys_state *css = NULL;
4088 struct cgroup *cgrp;
4089 struct cgroup_name *name;
4090 struct cgroupfs_root *root = parent->root;
4092 struct cgroup_subsys *ss;
4093 struct super_block *sb = root->sb;
4095 /* allocate the cgroup and its ID, 0 is reserved for the root */
4096 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4100 name = cgroup_alloc_name(dentry);
4103 rcu_assign_pointer(cgrp->name, name);
4106 * Temporarily set the pointer to NULL, so idr_find() won't return
4107 * a half-baked cgroup.
4109 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
4114 * Only live parents can have children. Note that the liveliness
4115 * check isn't strictly necessary because cgroup_mkdir() and
4116 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
4117 * anyway so that locking is contained inside cgroup proper and we
4118 * don't get nasty surprises if we ever grow another caller.
4120 if (!cgroup_lock_live_group(parent)) {
4125 /* Grab a reference on the superblock so the hierarchy doesn't
4126 * get deleted on unmount if there are child cgroups. This
4127 * can be done outside cgroup_mutex, since the sb can't
4128 * disappear while someone has an open control file on the
4130 atomic_inc(&sb->s_active);
4132 init_cgroup_housekeeping(cgrp);
4134 dentry->d_fsdata = cgrp;
4135 cgrp->dentry = dentry;
4137 cgrp->parent = parent;
4138 cgrp->dummy_css.parent = &parent->dummy_css;
4139 cgrp->root = parent->root;
4141 if (notify_on_release(parent))
4142 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4144 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4145 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4148 * Create directory. cgroup_create_file() returns with the new
4149 * directory locked on success so that it can be populated without
4150 * dropping cgroup_mutex.
4152 err = cgroup_create_file(dentry, S_IFDIR | mode, sb);
4155 lockdep_assert_held(&dentry->d_inode->i_mutex);
4157 cgrp->serial_nr = cgroup_serial_nr_next++;
4159 /* allocation complete, commit to creation */
4160 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
4161 root->number_of_cgroups++;
4163 /* hold a ref to the parent's dentry */
4164 dget(parent->dentry);
4167 * @cgrp is now fully operational. If something fails after this
4168 * point, it'll be released via the normal destruction path.
4170 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4172 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
4176 /* let's create and online css's */
4177 for_each_root_subsys(root, ss) {
4178 css = ss->css_alloc(cgroup_css(parent, ss));
4185 err = percpu_ref_init(&css->refcnt, css_release);
4189 init_css(css, ss, cgrp);
4191 err = cgroup_populate_dir(cgrp, 1 << ss->subsys_id);
4195 err = online_css(css);
4200 css_get(css->parent);
4202 /* mark it consumed for error path */
4205 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4207 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",
4208 current->comm, current->pid, ss->name);
4209 if (!strcmp(ss->name, "memory"))
4210 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
4211 ss->warned_broken_hierarchy = true;
4215 mutex_unlock(&cgroup_mutex);
4216 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
4221 mutex_unlock(&cgroup_mutex);
4222 /* Release the reference count that we took on the superblock */
4223 deactivate_super(sb);
4225 idr_remove(&root->cgroup_idr, cgrp->id);
4227 kfree(rcu_dereference_raw(cgrp->name));
4234 percpu_ref_cancel_init(&css->refcnt);
4235 css->ss->css_free(css);
4237 cgroup_destroy_locked(cgrp);
4238 mutex_unlock(&cgroup_mutex);
4239 mutex_unlock(&dentry->d_inode->i_mutex);
4243 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
4245 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
4247 /* the vfs holds inode->i_mutex already */
4248 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
4252 * This is called when the refcnt of a css is confirmed to be killed.
4253 * css_tryget() is now guaranteed to fail.
4255 static void css_killed_work_fn(struct work_struct *work)
4257 struct cgroup_subsys_state *css =
4258 container_of(work, struct cgroup_subsys_state, destroy_work);
4259 struct cgroup *cgrp = css->cgroup;
4261 mutex_lock(&cgroup_mutex);
4264 * css_tryget() is guaranteed to fail now. Tell subsystems to
4265 * initate destruction.
4270 * If @cgrp is marked dead, it's waiting for refs of all css's to
4271 * be disabled before proceeding to the second phase of cgroup
4272 * destruction. If we are the last one, kick it off.
4274 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
4275 cgroup_destroy_css_killed(cgrp);
4277 mutex_unlock(&cgroup_mutex);
4280 * Put the css refs from kill_css(). Each css holds an extra
4281 * reference to the cgroup's dentry and cgroup removal proceeds
4282 * regardless of css refs. On the last put of each css, whenever
4283 * that may be, the extra dentry ref is put so that dentry
4284 * destruction happens only after all css's are released.
4289 /* css kill confirmation processing requires process context, bounce */
4290 static void css_killed_ref_fn(struct percpu_ref *ref)
4292 struct cgroup_subsys_state *css =
4293 container_of(ref, struct cgroup_subsys_state, refcnt);
4295 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4296 queue_work(cgroup_destroy_wq, &css->destroy_work);
4300 * kill_css - destroy a css
4301 * @css: css to destroy
4303 * This function initiates destruction of @css by removing cgroup interface
4304 * files and putting its base reference. ->css_offline() will be invoked
4305 * asynchronously once css_tryget() is guaranteed to fail and when the
4306 * reference count reaches zero, @css will be released.
4308 static void kill_css(struct cgroup_subsys_state *css)
4310 cgroup_clear_dir(css->cgroup, 1 << css->ss->subsys_id);
4313 * Killing would put the base ref, but we need to keep it alive
4314 * until after ->css_offline().
4319 * cgroup core guarantees that, by the time ->css_offline() is
4320 * invoked, no new css reference will be given out via
4321 * css_tryget(). We can't simply call percpu_ref_kill() and
4322 * proceed to offlining css's because percpu_ref_kill() doesn't
4323 * guarantee that the ref is seen as killed on all CPUs on return.
4325 * Use percpu_ref_kill_and_confirm() to get notifications as each
4326 * css is confirmed to be seen as killed on all CPUs.
4328 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4332 * cgroup_destroy_locked - the first stage of cgroup destruction
4333 * @cgrp: cgroup to be destroyed
4335 * css's make use of percpu refcnts whose killing latency shouldn't be
4336 * exposed to userland and are RCU protected. Also, cgroup core needs to
4337 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
4338 * invoked. To satisfy all the requirements, destruction is implemented in
4339 * the following two steps.
4341 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4342 * userland visible parts and start killing the percpu refcnts of
4343 * css's. Set up so that the next stage will be kicked off once all
4344 * the percpu refcnts are confirmed to be killed.
4346 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4347 * rest of destruction. Once all cgroup references are gone, the
4348 * cgroup is RCU-freed.
4350 * This function implements s1. After this step, @cgrp is gone as far as
4351 * the userland is concerned and a new cgroup with the same name may be
4352 * created. As cgroup doesn't care about the names internally, this
4353 * doesn't cause any problem.
4355 static int cgroup_destroy_locked(struct cgroup *cgrp)
4356 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4358 struct dentry *d = cgrp->dentry;
4359 struct cgroup_subsys *ss;
4360 struct cgroup *child;
4363 lockdep_assert_held(&d->d_inode->i_mutex);
4364 lockdep_assert_held(&cgroup_mutex);
4367 * css_set_lock synchronizes access to ->cset_links and prevents
4368 * @cgrp from being removed while __put_css_set() is in progress.
4370 read_lock(&css_set_lock);
4371 empty = list_empty(&cgrp->cset_links);
4372 read_unlock(&css_set_lock);
4377 * Make sure there's no live children. We can't test ->children
4378 * emptiness as dead children linger on it while being destroyed;
4379 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
4383 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
4384 empty = cgroup_is_dead(child);
4393 * Initiate massacre of all css's. cgroup_destroy_css_killed()
4394 * will be invoked to perform the rest of destruction once the
4395 * percpu refs of all css's are confirmed to be killed.
4397 for_each_root_subsys(cgrp->root, ss) {
4398 struct cgroup_subsys_state *css = cgroup_css(cgrp, ss);
4405 * Mark @cgrp dead. This prevents further task migration and child
4406 * creation by disabling cgroup_lock_live_group(). Note that
4407 * CGRP_DEAD assertion is depended upon by css_next_child() to
4408 * resume iteration after dropping RCU read lock. See
4409 * css_next_child() for details.
4411 set_bit(CGRP_DEAD, &cgrp->flags);
4413 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4414 raw_spin_lock(&release_list_lock);
4415 if (!list_empty(&cgrp->release_list))
4416 list_del_init(&cgrp->release_list);
4417 raw_spin_unlock(&release_list_lock);
4420 * If @cgrp has css's attached, the second stage of cgroup
4421 * destruction is kicked off from css_killed_work_fn() after the
4422 * refs of all attached css's are killed. If @cgrp doesn't have
4423 * any css, we kick it off here.
4426 cgroup_destroy_css_killed(cgrp);
4429 * Clear the base files and remove @cgrp directory. The removal
4430 * puts the base ref but we aren't quite done with @cgrp yet, so
4433 cgroup_addrm_files(cgrp, cgroup_base_files, false);
4435 cgroup_d_remove_dir(d);
4441 * cgroup_destroy_css_killed - the second step of cgroup destruction
4442 * @work: cgroup->destroy_free_work
4444 * This function is invoked from a work item for a cgroup which is being
4445 * destroyed after all css's are offlined and performs the rest of
4446 * destruction. This is the second step of destruction described in the
4447 * comment above cgroup_destroy_locked().
4449 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
4451 struct cgroup *parent = cgrp->parent;
4452 struct dentry *d = cgrp->dentry;
4454 lockdep_assert_held(&cgroup_mutex);
4456 /* delete this cgroup from parent->children */
4457 list_del_rcu(&cgrp->sibling);
4460 * We should remove the cgroup object from idr before its grace
4461 * period starts, so we won't be looking up a cgroup while the
4462 * cgroup is being freed.
4464 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4469 set_bit(CGRP_RELEASABLE, &parent->flags);
4470 check_for_release(parent);
4473 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
4477 mutex_lock(&cgroup_mutex);
4478 ret = cgroup_destroy_locked(dentry->d_fsdata);
4479 mutex_unlock(&cgroup_mutex);
4484 static void __init_or_module cgroup_init_cftsets(struct cgroup_subsys *ss)
4486 INIT_LIST_HEAD(&ss->cftsets);
4489 * base_cftset is embedded in subsys itself, no need to worry about
4492 if (ss->base_cftypes) {
4495 for (cft = ss->base_cftypes; cft->name[0] != '\0'; cft++)
4498 ss->base_cftset.cfts = ss->base_cftypes;
4499 list_add_tail(&ss->base_cftset.node, &ss->cftsets);
4503 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4505 struct cgroup_subsys_state *css;
4507 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4509 mutex_lock(&cgroup_mutex);
4511 /* init base cftset */
4512 cgroup_init_cftsets(ss);
4514 /* Create the top cgroup state for this subsystem */
4515 list_add(&ss->sibling, &cgroup_dummy_root.subsys_list);
4516 ss->root = &cgroup_dummy_root;
4517 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4518 /* We don't handle early failures gracefully */
4519 BUG_ON(IS_ERR(css));
4520 init_css(css, ss, cgroup_dummy_top);
4522 /* Update the init_css_set to contain a subsys
4523 * pointer to this state - since the subsystem is
4524 * newly registered, all tasks and hence the
4525 * init_css_set is in the subsystem's top cgroup. */
4526 init_css_set.subsys[ss->subsys_id] = css;
4528 need_forkexit_callback |= ss->fork || ss->exit;
4530 /* At system boot, before all subsystems have been
4531 * registered, no tasks have been forked, so we don't
4532 * need to invoke fork callbacks here. */
4533 BUG_ON(!list_empty(&init_task.tasks));
4535 BUG_ON(online_css(css));
4537 mutex_unlock(&cgroup_mutex);
4539 /* this function shouldn't be used with modular subsystems, since they
4540 * need to register a subsys_id, among other things */
4545 * cgroup_load_subsys: load and register a modular subsystem at runtime
4546 * @ss: the subsystem to load
4548 * This function should be called in a modular subsystem's initcall. If the
4549 * subsystem is built as a module, it will be assigned a new subsys_id and set
4550 * up for use. If the subsystem is built-in anyway, work is delegated to the
4551 * simpler cgroup_init_subsys.
4553 int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
4555 struct cgroup_subsys_state *css;
4557 struct hlist_node *tmp;
4558 struct css_set *cset;
4561 /* check name and function validity */
4562 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
4563 ss->css_alloc == NULL || ss->css_free == NULL)
4567 * we don't support callbacks in modular subsystems. this check is
4568 * before the ss->module check for consistency; a subsystem that could
4569 * be a module should still have no callbacks even if the user isn't
4570 * compiling it as one.
4572 if (ss->fork || ss->exit)
4576 * an optionally modular subsystem is built-in: we want to do nothing,
4577 * since cgroup_init_subsys will have already taken care of it.
4579 if (ss->module == NULL) {
4580 /* a sanity check */
4581 BUG_ON(cgroup_subsys[ss->subsys_id] != ss);
4585 /* init base cftset */
4586 cgroup_init_cftsets(ss);
4588 mutex_lock(&cgroup_mutex);
4589 mutex_lock(&cgroup_root_mutex);
4590 cgroup_subsys[ss->subsys_id] = ss;
4593 * no ss->css_alloc seems to need anything important in the ss
4594 * struct, so this can happen first (i.e. before the dummy root
4597 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4599 /* failure case - need to deassign the cgroup_subsys[] slot. */
4600 cgroup_subsys[ss->subsys_id] = NULL;
4601 mutex_unlock(&cgroup_mutex);
4602 return PTR_ERR(css);
4605 list_add(&ss->sibling, &cgroup_dummy_root.subsys_list);
4606 ss->root = &cgroup_dummy_root;
4608 /* our new subsystem will be attached to the dummy hierarchy. */
4609 init_css(css, ss, cgroup_dummy_top);
4612 * Now we need to entangle the css into the existing css_sets. unlike
4613 * in cgroup_init_subsys, there are now multiple css_sets, so each one
4614 * will need a new pointer to it; done by iterating the css_set_table.
4615 * furthermore, modifying the existing css_sets will corrupt the hash
4616 * table state, so each changed css_set will need its hash recomputed.
4617 * this is all done under the css_set_lock.
4619 write_lock(&css_set_lock);
4620 hash_for_each_safe(css_set_table, i, tmp, cset, hlist) {
4621 /* skip entries that we already rehashed */
4622 if (cset->subsys[ss->subsys_id])
4624 /* remove existing entry */
4625 hash_del(&cset->hlist);
4627 cset->subsys[ss->subsys_id] = css;
4628 /* recompute hash and restore entry */
4629 key = css_set_hash(cset->subsys);
4630 hash_add(css_set_table, &cset->hlist, key);
4632 write_unlock(&css_set_lock);
4634 ret = online_css(css);
4639 mutex_unlock(&cgroup_root_mutex);
4640 mutex_unlock(&cgroup_mutex);
4644 mutex_unlock(&cgroup_root_mutex);
4645 mutex_unlock(&cgroup_mutex);
4646 /* @ss can't be mounted here as try_module_get() would fail */
4647 cgroup_unload_subsys(ss);
4650 EXPORT_SYMBOL_GPL(cgroup_load_subsys);
4653 * cgroup_unload_subsys: unload a modular subsystem
4654 * @ss: the subsystem to unload
4656 * This function should be called in a modular subsystem's exitcall. When this
4657 * function is invoked, the refcount on the subsystem's module will be 0, so
4658 * the subsystem will not be attached to any hierarchy.
4660 void cgroup_unload_subsys(struct cgroup_subsys *ss)
4662 struct cgrp_cset_link *link;
4664 BUG_ON(ss->module == NULL);
4667 * we shouldn't be called if the subsystem is in use, and the use of
4668 * try_module_get() in rebind_subsystems() should ensure that it
4669 * doesn't start being used while we're killing it off.
4671 BUG_ON(ss->root != &cgroup_dummy_root);
4673 mutex_lock(&cgroup_mutex);
4674 mutex_lock(&cgroup_root_mutex);
4676 offline_css(cgroup_css(cgroup_dummy_top, ss));
4678 /* deassign the subsys_id */
4679 cgroup_subsys[ss->subsys_id] = NULL;
4681 /* remove subsystem from the dummy root's list of subsystems */
4682 list_del_init(&ss->sibling);
4685 * disentangle the css from all css_sets attached to the dummy
4686 * top. as in loading, we need to pay our respects to the hashtable
4689 write_lock(&css_set_lock);
4690 list_for_each_entry(link, &cgroup_dummy_top->cset_links, cset_link) {
4691 struct css_set *cset = link->cset;
4694 hash_del(&cset->hlist);
4695 cset->subsys[ss->subsys_id] = NULL;
4696 key = css_set_hash(cset->subsys);
4697 hash_add(css_set_table, &cset->hlist, key);
4699 write_unlock(&css_set_lock);
4702 * remove subsystem's css from the cgroup_dummy_top and free it -
4703 * need to free before marking as null because ss->css_free needs
4704 * the cgrp->subsys pointer to find their state.
4706 ss->css_free(cgroup_css(cgroup_dummy_top, ss));
4707 RCU_INIT_POINTER(cgroup_dummy_top->subsys[ss->subsys_id], NULL);
4709 mutex_unlock(&cgroup_root_mutex);
4710 mutex_unlock(&cgroup_mutex);
4712 EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
4715 * cgroup_init_early - cgroup initialization at system boot
4717 * Initialize cgroups at system boot, and initialize any
4718 * subsystems that request early init.
4720 int __init cgroup_init_early(void)
4722 struct cgroup_subsys *ss;
4725 atomic_set(&init_css_set.refcount, 1);
4726 INIT_LIST_HEAD(&init_css_set.cgrp_links);
4727 INIT_LIST_HEAD(&init_css_set.tasks);
4728 INIT_HLIST_NODE(&init_css_set.hlist);
4730 init_cgroup_root(&cgroup_dummy_root);
4731 cgroup_root_count = 1;
4732 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4734 init_cgrp_cset_link.cset = &init_css_set;
4735 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4736 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4737 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4739 /* at bootup time, we don't worry about modular subsystems */
4740 for_each_builtin_subsys(ss, i) {
4742 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
4743 BUG_ON(!ss->css_alloc);
4744 BUG_ON(!ss->css_free);
4745 if (ss->subsys_id != i) {
4746 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
4747 ss->name, ss->subsys_id);
4752 cgroup_init_subsys(ss);
4758 * cgroup_init - cgroup initialization
4760 * Register cgroup filesystem and /proc file, and initialize
4761 * any subsystems that didn't request early init.
4763 int __init cgroup_init(void)
4765 struct cgroup_subsys *ss;
4769 err = bdi_init(&cgroup_backing_dev_info);
4773 for_each_builtin_subsys(ss, i) {
4774 if (!ss->early_init)
4775 cgroup_init_subsys(ss);
4778 /* allocate id for the dummy hierarchy */
4779 mutex_lock(&cgroup_mutex);
4780 mutex_lock(&cgroup_root_mutex);
4782 /* Add init_css_set to the hash table */
4783 key = css_set_hash(init_css_set.subsys);
4784 hash_add(css_set_table, &init_css_set.hlist, key);
4786 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4788 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
4792 mutex_unlock(&cgroup_root_mutex);
4793 mutex_unlock(&cgroup_mutex);
4795 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4801 err = register_filesystem(&cgroup_fs_type);
4803 kobject_put(cgroup_kobj);
4807 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4811 bdi_destroy(&cgroup_backing_dev_info);
4816 static int __init cgroup_wq_init(void)
4819 * There isn't much point in executing destruction path in
4820 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4821 * Use 1 for @max_active.
4823 * We would prefer to do this in cgroup_init() above, but that
4824 * is called before init_workqueues(): so leave this until after.
4826 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
4827 BUG_ON(!cgroup_destroy_wq);
4830 * Used to destroy pidlists and separate to serve as flush domain.
4831 * Cap @max_active to 1 too.
4833 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4835 BUG_ON(!cgroup_pidlist_destroy_wq);
4839 core_initcall(cgroup_wq_init);
4842 * proc_cgroup_show()
4843 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4844 * - Used for /proc/<pid>/cgroup.
4845 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4846 * doesn't really matter if tsk->cgroup changes after we read it,
4847 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4848 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4849 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4850 * cgroup to top_cgroup.
4853 /* TODO: Use a proper seq_file iterator */
4854 int proc_cgroup_show(struct seq_file *m, void *v)
4857 struct task_struct *tsk;
4860 struct cgroupfs_root *root;
4863 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4869 tsk = get_pid_task(pid, PIDTYPE_PID);
4875 mutex_lock(&cgroup_mutex);
4877 for_each_active_root(root) {
4878 struct cgroup_subsys *ss;
4879 struct cgroup *cgrp;
4882 seq_printf(m, "%d:", root->hierarchy_id);
4883 for_each_root_subsys(root, ss)
4884 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4885 if (strlen(root->name))
4886 seq_printf(m, "%sname=%s", count ? "," : "",
4889 cgrp = task_cgroup_from_root(tsk, root);
4890 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4898 mutex_unlock(&cgroup_mutex);
4899 put_task_struct(tsk);
4906 /* Display information about each subsystem and each hierarchy */
4907 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4909 struct cgroup_subsys *ss;
4912 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4914 * ideally we don't want subsystems moving around while we do this.
4915 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4916 * subsys/hierarchy state.
4918 mutex_lock(&cgroup_mutex);
4920 for_each_subsys(ss, i)
4921 seq_printf(m, "%s\t%d\t%d\t%d\n",
4922 ss->name, ss->root->hierarchy_id,
4923 ss->root->number_of_cgroups, !ss->disabled);
4925 mutex_unlock(&cgroup_mutex);
4929 static int cgroupstats_open(struct inode *inode, struct file *file)
4931 return single_open(file, proc_cgroupstats_show, NULL);
4934 static const struct file_operations proc_cgroupstats_operations = {
4935 .open = cgroupstats_open,
4937 .llseek = seq_lseek,
4938 .release = single_release,
4942 * cgroup_fork - attach newly forked task to its parents cgroup.
4943 * @child: pointer to task_struct of forking parent process.
4945 * Description: A task inherits its parent's cgroup at fork().
4947 * A pointer to the shared css_set was automatically copied in
4948 * fork.c by dup_task_struct(). However, we ignore that copy, since
4949 * it was not made under the protection of RCU or cgroup_mutex, so
4950 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4951 * have already changed current->cgroups, allowing the previously
4952 * referenced cgroup group to be removed and freed.
4954 * At the point that cgroup_fork() is called, 'current' is the parent
4955 * task, and the passed argument 'child' points to the child task.
4957 void cgroup_fork(struct task_struct *child)
4960 get_css_set(task_css_set(current));
4961 child->cgroups = current->cgroups;
4962 task_unlock(current);
4963 INIT_LIST_HEAD(&child->cg_list);
4967 * cgroup_post_fork - called on a new task after adding it to the task list
4968 * @child: the task in question
4970 * Adds the task to the list running through its css_set if necessary and
4971 * call the subsystem fork() callbacks. Has to be after the task is
4972 * visible on the task list in case we race with the first call to
4973 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4976 void cgroup_post_fork(struct task_struct *child)
4978 struct cgroup_subsys *ss;
4982 * use_task_css_set_links is set to 1 before we walk the tasklist
4983 * under the tasklist_lock and we read it here after we added the child
4984 * to the tasklist under the tasklist_lock as well. If the child wasn't
4985 * yet in the tasklist when we walked through it from
4986 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4987 * should be visible now due to the paired locking and barriers implied
4988 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4989 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4992 if (use_task_css_set_links) {
4993 write_lock(&css_set_lock);
4995 if (list_empty(&child->cg_list))
4996 list_add(&child->cg_list, &task_css_set(child)->tasks);
4998 write_unlock(&css_set_lock);
5002 * Call ss->fork(). This must happen after @child is linked on
5003 * css_set; otherwise, @child might change state between ->fork()
5004 * and addition to css_set.
5006 if (need_forkexit_callback) {
5008 * fork/exit callbacks are supported only for builtin
5009 * subsystems, and the builtin section of the subsys
5010 * array is immutable, so we don't need to lock the
5011 * subsys array here. On the other hand, modular section
5012 * of the array can be freed at module unload, so we
5015 for_each_builtin_subsys(ss, i)
5022 * cgroup_exit - detach cgroup from exiting task
5023 * @tsk: pointer to task_struct of exiting process
5024 * @run_callback: run exit callbacks?
5026 * Description: Detach cgroup from @tsk and release it.
5028 * Note that cgroups marked notify_on_release force every task in
5029 * them to take the global cgroup_mutex mutex when exiting.
5030 * This could impact scaling on very large systems. Be reluctant to
5031 * use notify_on_release cgroups where very high task exit scaling
5032 * is required on large systems.
5034 * the_top_cgroup_hack:
5036 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
5038 * We call cgroup_exit() while the task is still competent to
5039 * handle notify_on_release(), then leave the task attached to the
5040 * root cgroup in each hierarchy for the remainder of its exit.
5042 * To do this properly, we would increment the reference count on
5043 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
5044 * code we would add a second cgroup function call, to drop that
5045 * reference. This would just create an unnecessary hot spot on
5046 * the top_cgroup reference count, to no avail.
5048 * Normally, holding a reference to a cgroup without bumping its
5049 * count is unsafe. The cgroup could go away, or someone could
5050 * attach us to a different cgroup, decrementing the count on
5051 * the first cgroup that we never incremented. But in this case,
5052 * top_cgroup isn't going away, and either task has PF_EXITING set,
5053 * which wards off any cgroup_attach_task() attempts, or task is a failed
5054 * fork, never visible to cgroup_attach_task.
5056 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
5058 struct cgroup_subsys *ss;
5059 struct css_set *cset;
5063 * Unlink from the css_set task list if necessary.
5064 * Optimistically check cg_list before taking
5067 if (!list_empty(&tsk->cg_list)) {
5068 write_lock(&css_set_lock);
5069 if (!list_empty(&tsk->cg_list))
5070 list_del_init(&tsk->cg_list);
5071 write_unlock(&css_set_lock);
5074 /* Reassign the task to the init_css_set. */
5076 cset = task_css_set(tsk);
5077 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5079 if (run_callbacks && need_forkexit_callback) {
5081 * fork/exit callbacks are supported only for builtin
5082 * subsystems, see cgroup_post_fork() for details.
5084 for_each_builtin_subsys(ss, i) {
5086 struct cgroup_subsys_state *old_css = cset->subsys[i];
5087 struct cgroup_subsys_state *css = task_css(tsk, i);
5089 ss->exit(css, old_css, tsk);
5095 put_css_set_taskexit(cset);
5098 static void check_for_release(struct cgroup *cgrp)
5100 if (cgroup_is_releasable(cgrp) &&
5101 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
5103 * Control Group is currently removeable. If it's not
5104 * already queued for a userspace notification, queue
5107 int need_schedule_work = 0;
5109 raw_spin_lock(&release_list_lock);
5110 if (!cgroup_is_dead(cgrp) &&
5111 list_empty(&cgrp->release_list)) {
5112 list_add(&cgrp->release_list, &release_list);
5113 need_schedule_work = 1;
5115 raw_spin_unlock(&release_list_lock);
5116 if (need_schedule_work)
5117 schedule_work(&release_agent_work);
5122 * Notify userspace when a cgroup is released, by running the
5123 * configured release agent with the name of the cgroup (path
5124 * relative to the root of cgroup file system) as the argument.
5126 * Most likely, this user command will try to rmdir this cgroup.
5128 * This races with the possibility that some other task will be
5129 * attached to this cgroup before it is removed, or that some other
5130 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5131 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5132 * unused, and this cgroup will be reprieved from its death sentence,
5133 * to continue to serve a useful existence. Next time it's released,
5134 * we will get notified again, if it still has 'notify_on_release' set.
5136 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5137 * means only wait until the task is successfully execve()'d. The
5138 * separate release agent task is forked by call_usermodehelper(),
5139 * then control in this thread returns here, without waiting for the
5140 * release agent task. We don't bother to wait because the caller of
5141 * this routine has no use for the exit status of the release agent
5142 * task, so no sense holding our caller up for that.
5144 static void cgroup_release_agent(struct work_struct *work)
5146 BUG_ON(work != &release_agent_work);
5147 mutex_lock(&cgroup_mutex);
5148 raw_spin_lock(&release_list_lock);
5149 while (!list_empty(&release_list)) {
5150 char *argv[3], *envp[3];
5152 char *pathbuf = NULL, *agentbuf = NULL;
5153 struct cgroup *cgrp = list_entry(release_list.next,
5156 list_del_init(&cgrp->release_list);
5157 raw_spin_unlock(&release_list_lock);
5158 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
5161 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
5163 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5168 argv[i++] = agentbuf;
5169 argv[i++] = pathbuf;
5173 /* minimal command environment */
5174 envp[i++] = "HOME=/";
5175 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5178 /* Drop the lock while we invoke the usermode helper,
5179 * since the exec could involve hitting disk and hence
5180 * be a slow process */
5181 mutex_unlock(&cgroup_mutex);
5182 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5183 mutex_lock(&cgroup_mutex);
5187 raw_spin_lock(&release_list_lock);
5189 raw_spin_unlock(&release_list_lock);
5190 mutex_unlock(&cgroup_mutex);
5193 static int __init cgroup_disable(char *str)
5195 struct cgroup_subsys *ss;
5199 while ((token = strsep(&str, ",")) != NULL) {
5204 * cgroup_disable, being at boot time, can't know about
5205 * module subsystems, so we don't worry about them.
5207 for_each_builtin_subsys(ss, i) {
5208 if (!strcmp(token, ss->name)) {
5210 printk(KERN_INFO "Disabling %s control group"
5211 " subsystem\n", ss->name);
5218 __setup("cgroup_disable=", cgroup_disable);
5221 * css_from_dir - get corresponding css from the dentry of a cgroup dir
5222 * @dentry: directory dentry of interest
5223 * @ss: subsystem of interest
5225 * Must be called under cgroup_mutex or RCU read lock. The caller is
5226 * responsible for pinning the returned css if it needs to be accessed
5227 * outside the critical section.
5229 struct cgroup_subsys_state *css_from_dir(struct dentry *dentry,
5230 struct cgroup_subsys *ss)
5232 struct cgroup *cgrp;
5234 cgroup_assert_mutex_or_rcu_locked();
5236 /* is @dentry a cgroup dir? */
5237 if (!dentry->d_inode ||
5238 dentry->d_inode->i_op != &cgroup_dir_inode_operations)
5239 return ERR_PTR(-EBADF);
5241 cgrp = __d_cgrp(dentry);
5242 return cgroup_css(cgrp, ss) ?: ERR_PTR(-ENOENT);
5246 * css_from_id - lookup css by id
5247 * @id: the cgroup id
5248 * @ss: cgroup subsys to be looked into
5250 * Returns the css if there's valid one with @id, otherwise returns NULL.
5251 * Should be called under rcu_read_lock().
5253 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5255 struct cgroup *cgrp;
5257 cgroup_assert_mutex_or_rcu_locked();
5259 cgrp = idr_find(&ss->root->cgroup_idr, id);
5261 return cgroup_css(cgrp, ss);
5265 #ifdef CONFIG_CGROUP_DEBUG
5266 static struct cgroup_subsys_state *
5267 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5269 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5272 return ERR_PTR(-ENOMEM);
5277 static void debug_css_free(struct cgroup_subsys_state *css)
5282 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5285 return cgroup_task_count(css->cgroup);
5288 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5291 return (u64)(unsigned long)current->cgroups;
5294 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5300 count = atomic_read(&task_css_set(current)->refcount);
5305 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5307 struct cgrp_cset_link *link;
5308 struct css_set *cset;
5310 read_lock(&css_set_lock);
5312 cset = rcu_dereference(current->cgroups);
5313 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5314 struct cgroup *c = link->cgrp;
5318 name = c->dentry->d_name.name;
5321 seq_printf(seq, "Root %d group %s\n",
5322 c->root->hierarchy_id, name);
5325 read_unlock(&css_set_lock);
5329 #define MAX_TASKS_SHOWN_PER_CSS 25
5330 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5332 struct cgroup_subsys_state *css = seq_css(seq);
5333 struct cgrp_cset_link *link;
5335 read_lock(&css_set_lock);
5336 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5337 struct css_set *cset = link->cset;
5338 struct task_struct *task;
5340 seq_printf(seq, "css_set %p\n", cset);
5341 list_for_each_entry(task, &cset->tasks, cg_list) {
5342 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
5343 seq_puts(seq, " ...\n");
5346 seq_printf(seq, " task %d\n",
5347 task_pid_vnr(task));
5351 read_unlock(&css_set_lock);
5355 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5357 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5360 static struct cftype debug_files[] = {
5362 .name = "taskcount",
5363 .read_u64 = debug_taskcount_read,
5367 .name = "current_css_set",
5368 .read_u64 = current_css_set_read,
5372 .name = "current_css_set_refcount",
5373 .read_u64 = current_css_set_refcount_read,
5377 .name = "current_css_set_cg_links",
5378 .seq_show = current_css_set_cg_links_read,
5382 .name = "cgroup_css_links",
5383 .seq_show = cgroup_css_links_read,
5387 .name = "releasable",
5388 .read_u64 = releasable_read,
5394 struct cgroup_subsys debug_subsys = {
5396 .css_alloc = debug_css_alloc,
5397 .css_free = debug_css_free,
5398 .subsys_id = debug_subsys_id,
5399 .base_cftypes = debug_files,
5401 #endif /* CONFIG_CGROUP_DEBUG */