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/seq_file.h>
45 #include <linux/slab.h>
46 #include <linux/magic.h>
47 #include <linux/spinlock.h>
48 #include <linux/string.h>
49 #include <linux/sort.h>
50 #include <linux/kmod.h>
51 #include <linux/module.h>
52 #include <linux/delayacct.h>
53 #include <linux/cgroupstats.h>
54 #include <linux/hashtable.h>
55 #include <linux/namei.h>
56 #include <linux/pid_namespace.h>
57 #include <linux/idr.h>
58 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
59 #include <linux/eventfd.h>
60 #include <linux/poll.h>
61 #include <linux/flex_array.h> /* used in cgroup_attach_task */
62 #include <linux/kthread.h>
64 #include <linux/atomic.h>
67 * cgroup_mutex is the master lock. Any modification to cgroup or its
68 * hierarchy must be performed while holding it.
70 * cgroup_root_mutex nests inside cgroup_mutex and should be held to modify
71 * cgroupfs_root of any cgroup hierarchy - subsys list, flags,
72 * release_agent_path and so on. Modifying requires both cgroup_mutex and
73 * cgroup_root_mutex. Readers can acquire either of the two. This is to
74 * break the following locking order cycle.
76 * A. cgroup_mutex -> cred_guard_mutex -> s_type->i_mutex_key -> namespace_sem
77 * B. namespace_sem -> cgroup_mutex
79 * B happens only through cgroup_show_options() and using cgroup_root_mutex
82 #ifdef CONFIG_PROVE_RCU
83 DEFINE_MUTEX(cgroup_mutex);
84 EXPORT_SYMBOL_GPL(cgroup_mutex); /* only for lockdep */
86 static DEFINE_MUTEX(cgroup_mutex);
89 static DEFINE_MUTEX(cgroup_root_mutex);
92 * Generate an array of cgroup subsystem pointers. At boot time, this is
93 * populated with the built in subsystems, and modular subsystems are
94 * registered after that. The mutable section of this array is protected by
97 #define SUBSYS(_x) [_x ## _subsys_id] = &_x ## _subsys,
98 #define IS_SUBSYS_ENABLED(option) IS_BUILTIN(option)
99 static struct cgroup_subsys *cgroup_subsys[CGROUP_SUBSYS_COUNT] = {
100 #include <linux/cgroup_subsys.h>
104 * The dummy hierarchy, reserved for the subsystems that are otherwise
105 * unattached - it never has more than a single cgroup, and all tasks are
106 * part of that cgroup.
108 static struct cgroupfs_root cgroup_dummy_root;
110 /* dummy_top is a shorthand for the dummy hierarchy's top cgroup */
111 static struct cgroup * const cgroup_dummy_top = &cgroup_dummy_root.top_cgroup;
114 * cgroupfs file entry, pointed to from leaf dentry->d_fsdata.
117 struct list_head node;
118 struct dentry *dentry;
120 struct cgroup_subsys_state *css;
123 struct simple_xattrs xattrs;
127 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
128 * cgroup_subsys->use_id != 0.
130 #define CSS_ID_MAX (65535)
133 * The css to which this ID points. This pointer is set to valid value
134 * after cgroup is populated. If cgroup is removed, this will be NULL.
135 * This pointer is expected to be RCU-safe because destroy()
136 * is called after synchronize_rcu(). But for safe use, css_tryget()
137 * should be used for avoiding race.
139 struct cgroup_subsys_state __rcu *css;
145 * Depth in hierarchy which this ID belongs to.
147 unsigned short depth;
149 * ID is freed by RCU. (and lookup routine is RCU safe.)
151 struct rcu_head rcu_head;
153 * Hierarchy of CSS ID belongs to.
155 unsigned short stack[0]; /* Array of Length (depth+1) */
159 * cgroup_event represents events which userspace want to receive.
161 struct cgroup_event {
163 * css which the event belongs to.
165 struct cgroup_subsys_state *css;
167 * Control file which the event associated.
171 * eventfd to signal userspace about the event.
173 struct eventfd_ctx *eventfd;
175 * Each of these stored in a list by the cgroup.
177 struct list_head list;
179 * All fields below needed to unregister event when
180 * userspace closes eventfd.
183 wait_queue_head_t *wqh;
185 struct work_struct remove;
188 /* The list of hierarchy roots */
190 static LIST_HEAD(cgroup_roots);
191 static int cgroup_root_count;
194 * Hierarchy ID allocation and mapping. It follows the same exclusion
195 * rules as other root ops - both cgroup_mutex and cgroup_root_mutex for
196 * writes, either for reads.
198 static DEFINE_IDR(cgroup_hierarchy_idr);
200 static struct cgroup_name root_cgroup_name = { .name = "/" };
203 * Assign a monotonically increasing serial number to cgroups. It
204 * guarantees cgroups with bigger numbers are newer than those with smaller
205 * numbers. Also, as cgroups are always appended to the parent's
206 * ->children list, it guarantees that sibling cgroups are always sorted in
207 * the ascending serial number order on the list. Protected by
210 static u64 cgroup_serial_nr_next = 1;
212 /* This flag indicates whether tasks in the fork and exit paths should
213 * check for fork/exit handlers to call. This avoids us having to do
214 * extra work in the fork/exit path if none of the subsystems need to
217 static int need_forkexit_callback __read_mostly;
219 static struct cftype cgroup_base_files[];
221 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
222 static int cgroup_destroy_locked(struct cgroup *cgrp);
223 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
227 * cgroup_css - obtain a cgroup's css for the specified subsystem
228 * @cgrp: the cgroup of interest
229 * @subsys_id: the subsystem of interest
231 * Return @cgrp's css (cgroup_subsys_state) associated with @subsys_id.
232 * This function must be called either under cgroup_mutex or
233 * rcu_read_lock() and the caller is responsible for pinning the returned
234 * css if it wants to keep accessing it outside the said locks. This
235 * function may return %NULL if @cgrp doesn't have @subsys_id enabled.
237 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
240 return rcu_dereference_check(cgrp->subsys[subsys_id],
241 lockdep_is_held(&cgroup_mutex));
244 /* convenient tests for these bits */
245 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
247 return test_bit(CGRP_DEAD, &cgrp->flags);
251 * cgroup_is_descendant - test ancestry
252 * @cgrp: the cgroup to be tested
253 * @ancestor: possible ancestor of @cgrp
255 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
256 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
257 * and @ancestor are accessible.
259 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
262 if (cgrp == ancestor)
268 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
270 static int cgroup_is_releasable(const struct cgroup *cgrp)
273 (1 << CGRP_RELEASABLE) |
274 (1 << CGRP_NOTIFY_ON_RELEASE);
275 return (cgrp->flags & bits) == bits;
278 static int notify_on_release(const struct cgroup *cgrp)
280 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
284 * for_each_subsys - iterate all loaded cgroup subsystems
285 * @ss: the iteration cursor
286 * @i: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
288 * Should be called under cgroup_mutex.
290 #define for_each_subsys(ss, i) \
291 for ((i) = 0; (i) < CGROUP_SUBSYS_COUNT; (i)++) \
292 if (({ lockdep_assert_held(&cgroup_mutex); \
293 !((ss) = cgroup_subsys[i]); })) { } \
297 * for_each_builtin_subsys - iterate all built-in cgroup subsystems
298 * @ss: the iteration cursor
299 * @i: the index of @ss, CGROUP_BUILTIN_SUBSYS_COUNT after reaching the end
301 * Bulit-in subsystems are always present and iteration itself doesn't
302 * require any synchronization.
304 #define for_each_builtin_subsys(ss, i) \
305 for ((i) = 0; (i) < CGROUP_BUILTIN_SUBSYS_COUNT && \
306 (((ss) = cgroup_subsys[i]) || true); (i)++)
308 /* iterate each subsystem attached to a hierarchy */
309 #define for_each_root_subsys(root, ss) \
310 list_for_each_entry((ss), &(root)->subsys_list, sibling)
312 /* iterate across the active hierarchies */
313 #define for_each_active_root(root) \
314 list_for_each_entry((root), &cgroup_roots, root_list)
316 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
318 return dentry->d_fsdata;
321 static inline struct cfent *__d_cfe(struct dentry *dentry)
323 return dentry->d_fsdata;
326 static inline struct cftype *__d_cft(struct dentry *dentry)
328 return __d_cfe(dentry)->type;
332 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
333 * @cgrp: the cgroup to be checked for liveness
335 * On success, returns true; the mutex should be later unlocked. On
336 * failure returns false with no lock held.
338 static bool cgroup_lock_live_group(struct cgroup *cgrp)
340 mutex_lock(&cgroup_mutex);
341 if (cgroup_is_dead(cgrp)) {
342 mutex_unlock(&cgroup_mutex);
348 /* the list of cgroups eligible for automatic release. Protected by
349 * release_list_lock */
350 static LIST_HEAD(release_list);
351 static DEFINE_RAW_SPINLOCK(release_list_lock);
352 static void cgroup_release_agent(struct work_struct *work);
353 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
354 static void check_for_release(struct cgroup *cgrp);
357 * A cgroup can be associated with multiple css_sets as different tasks may
358 * belong to different cgroups on different hierarchies. In the other
359 * direction, a css_set is naturally associated with multiple cgroups.
360 * This M:N relationship is represented by the following link structure
361 * which exists for each association and allows traversing the associations
364 struct cgrp_cset_link {
365 /* the cgroup and css_set this link associates */
367 struct css_set *cset;
369 /* list of cgrp_cset_links anchored at cgrp->cset_links */
370 struct list_head cset_link;
372 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
373 struct list_head cgrp_link;
376 /* The default css_set - used by init and its children prior to any
377 * hierarchies being mounted. It contains a pointer to the root state
378 * for each subsystem. Also used to anchor the list of css_sets. Not
379 * reference-counted, to improve performance when child cgroups
380 * haven't been created.
383 static struct css_set init_css_set;
384 static struct cgrp_cset_link init_cgrp_cset_link;
386 static int cgroup_init_idr(struct cgroup_subsys *ss,
387 struct cgroup_subsys_state *css);
390 * css_set_lock protects the list of css_set objects, and the chain of
391 * tasks off each css_set. Nests outside task->alloc_lock due to
392 * css_task_iter_start().
394 static DEFINE_RWLOCK(css_set_lock);
395 static int css_set_count;
398 * hash table for cgroup groups. This improves the performance to find
399 * an existing css_set. This hash doesn't (currently) take into
400 * account cgroups in empty hierarchies.
402 #define CSS_SET_HASH_BITS 7
403 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
405 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
407 unsigned long key = 0UL;
408 struct cgroup_subsys *ss;
411 for_each_subsys(ss, i)
412 key += (unsigned long)css[i];
413 key = (key >> 16) ^ key;
419 * We don't maintain the lists running through each css_set to its task
420 * until after the first call to css_task_iter_start(). This reduces the
421 * fork()/exit() overhead for people who have cgroups compiled into their
422 * kernel but not actually in use.
424 static int use_task_css_set_links __read_mostly;
426 static void __put_css_set(struct css_set *cset, int taskexit)
428 struct cgrp_cset_link *link, *tmp_link;
431 * Ensure that the refcount doesn't hit zero while any readers
432 * can see it. Similar to atomic_dec_and_lock(), but for an
435 if (atomic_add_unless(&cset->refcount, -1, 1))
437 write_lock(&css_set_lock);
438 if (!atomic_dec_and_test(&cset->refcount)) {
439 write_unlock(&css_set_lock);
443 /* This css_set is dead. unlink it and release cgroup refcounts */
444 hash_del(&cset->hlist);
447 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
448 struct cgroup *cgrp = link->cgrp;
450 list_del(&link->cset_link);
451 list_del(&link->cgrp_link);
453 /* @cgrp can't go away while we're holding css_set_lock */
454 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
456 set_bit(CGRP_RELEASABLE, &cgrp->flags);
457 check_for_release(cgrp);
463 write_unlock(&css_set_lock);
464 kfree_rcu(cset, rcu_head);
468 * refcounted get/put for css_set objects
470 static inline void get_css_set(struct css_set *cset)
472 atomic_inc(&cset->refcount);
475 static inline void put_css_set(struct css_set *cset)
477 __put_css_set(cset, 0);
480 static inline void put_css_set_taskexit(struct css_set *cset)
482 __put_css_set(cset, 1);
486 * compare_css_sets - helper function for find_existing_css_set().
487 * @cset: candidate css_set being tested
488 * @old_cset: existing css_set for a task
489 * @new_cgrp: cgroup that's being entered by the task
490 * @template: desired set of css pointers in css_set (pre-calculated)
492 * Returns true if "cset" matches "old_cset" except for the hierarchy
493 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
495 static bool compare_css_sets(struct css_set *cset,
496 struct css_set *old_cset,
497 struct cgroup *new_cgrp,
498 struct cgroup_subsys_state *template[])
500 struct list_head *l1, *l2;
502 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
503 /* Not all subsystems matched */
508 * Compare cgroup pointers in order to distinguish between
509 * different cgroups in heirarchies with no subsystems. We
510 * could get by with just this check alone (and skip the
511 * memcmp above) but on most setups the memcmp check will
512 * avoid the need for this more expensive check on almost all
516 l1 = &cset->cgrp_links;
517 l2 = &old_cset->cgrp_links;
519 struct cgrp_cset_link *link1, *link2;
520 struct cgroup *cgrp1, *cgrp2;
524 /* See if we reached the end - both lists are equal length. */
525 if (l1 == &cset->cgrp_links) {
526 BUG_ON(l2 != &old_cset->cgrp_links);
529 BUG_ON(l2 == &old_cset->cgrp_links);
531 /* Locate the cgroups associated with these links. */
532 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
533 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
536 /* Hierarchies should be linked in the same order. */
537 BUG_ON(cgrp1->root != cgrp2->root);
540 * If this hierarchy is the hierarchy of the cgroup
541 * that's changing, then we need to check that this
542 * css_set points to the new cgroup; if it's any other
543 * hierarchy, then this css_set should point to the
544 * same cgroup as the old css_set.
546 if (cgrp1->root == new_cgrp->root) {
547 if (cgrp1 != new_cgrp)
558 * find_existing_css_set - init css array and find the matching css_set
559 * @old_cset: the css_set that we're using before the cgroup transition
560 * @cgrp: the cgroup that we're moving into
561 * @template: out param for the new set of csses, should be clear on entry
563 static struct css_set *find_existing_css_set(struct css_set *old_cset,
565 struct cgroup_subsys_state *template[])
567 struct cgroupfs_root *root = cgrp->root;
568 struct cgroup_subsys *ss;
569 struct css_set *cset;
574 * Build the set of subsystem state objects that we want to see in the
575 * new css_set. while subsystems can change globally, the entries here
576 * won't change, so no need for locking.
578 for_each_subsys(ss, i) {
579 if (root->subsys_mask & (1UL << i)) {
580 /* Subsystem is in this hierarchy. So we want
581 * the subsystem state from the new
583 template[i] = cgroup_css(cgrp, i);
585 /* Subsystem is not in this hierarchy, so we
586 * don't want to change the subsystem state */
587 template[i] = old_cset->subsys[i];
591 key = css_set_hash(template);
592 hash_for_each_possible(css_set_table, cset, hlist, key) {
593 if (!compare_css_sets(cset, old_cset, cgrp, template))
596 /* This css_set matches what we need */
600 /* No existing cgroup group matched */
604 static void free_cgrp_cset_links(struct list_head *links_to_free)
606 struct cgrp_cset_link *link, *tmp_link;
608 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
609 list_del(&link->cset_link);
615 * allocate_cgrp_cset_links - allocate cgrp_cset_links
616 * @count: the number of links to allocate
617 * @tmp_links: list_head the allocated links are put on
619 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
620 * through ->cset_link. Returns 0 on success or -errno.
622 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
624 struct cgrp_cset_link *link;
627 INIT_LIST_HEAD(tmp_links);
629 for (i = 0; i < count; i++) {
630 link = kzalloc(sizeof(*link), GFP_KERNEL);
632 free_cgrp_cset_links(tmp_links);
635 list_add(&link->cset_link, tmp_links);
641 * link_css_set - a helper function to link a css_set to a cgroup
642 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
643 * @cset: the css_set to be linked
644 * @cgrp: the destination cgroup
646 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
649 struct cgrp_cset_link *link;
651 BUG_ON(list_empty(tmp_links));
652 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
655 list_move(&link->cset_link, &cgrp->cset_links);
657 * Always add links to the tail of the list so that the list
658 * is sorted by order of hierarchy creation
660 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
664 * find_css_set - return a new css_set with one cgroup updated
665 * @old_cset: the baseline css_set
666 * @cgrp: the cgroup to be updated
668 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
669 * substituted into the appropriate hierarchy.
671 static struct css_set *find_css_set(struct css_set *old_cset,
674 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
675 struct css_set *cset;
676 struct list_head tmp_links;
677 struct cgrp_cset_link *link;
680 lockdep_assert_held(&cgroup_mutex);
682 /* First see if we already have a cgroup group that matches
684 read_lock(&css_set_lock);
685 cset = find_existing_css_set(old_cset, cgrp, template);
688 read_unlock(&css_set_lock);
693 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
697 /* Allocate all the cgrp_cset_link objects that we'll need */
698 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
703 atomic_set(&cset->refcount, 1);
704 INIT_LIST_HEAD(&cset->cgrp_links);
705 INIT_LIST_HEAD(&cset->tasks);
706 INIT_HLIST_NODE(&cset->hlist);
708 /* Copy the set of subsystem state objects generated in
709 * find_existing_css_set() */
710 memcpy(cset->subsys, template, sizeof(cset->subsys));
712 write_lock(&css_set_lock);
713 /* Add reference counts and links from the new css_set. */
714 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
715 struct cgroup *c = link->cgrp;
717 if (c->root == cgrp->root)
719 link_css_set(&tmp_links, cset, c);
722 BUG_ON(!list_empty(&tmp_links));
726 /* Add this cgroup group to the hash table */
727 key = css_set_hash(cset->subsys);
728 hash_add(css_set_table, &cset->hlist, key);
730 write_unlock(&css_set_lock);
736 * Return the cgroup for "task" from the given hierarchy. Must be
737 * called with cgroup_mutex held.
739 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
740 struct cgroupfs_root *root)
742 struct css_set *cset;
743 struct cgroup *res = NULL;
745 BUG_ON(!mutex_is_locked(&cgroup_mutex));
746 read_lock(&css_set_lock);
748 * No need to lock the task - since we hold cgroup_mutex the
749 * task can't change groups, so the only thing that can happen
750 * is that it exits and its css is set back to init_css_set.
752 cset = task_css_set(task);
753 if (cset == &init_css_set) {
754 res = &root->top_cgroup;
756 struct cgrp_cset_link *link;
758 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
759 struct cgroup *c = link->cgrp;
761 if (c->root == root) {
767 read_unlock(&css_set_lock);
773 * There is one global cgroup mutex. We also require taking
774 * task_lock() when dereferencing a task's cgroup subsys pointers.
775 * See "The task_lock() exception", at the end of this comment.
777 * A task must hold cgroup_mutex to modify cgroups.
779 * Any task can increment and decrement the count field without lock.
780 * So in general, code holding cgroup_mutex can't rely on the count
781 * field not changing. However, if the count goes to zero, then only
782 * cgroup_attach_task() can increment it again. Because a count of zero
783 * means that no tasks are currently attached, therefore there is no
784 * way a task attached to that cgroup can fork (the other way to
785 * increment the count). So code holding cgroup_mutex can safely
786 * assume that if the count is zero, it will stay zero. Similarly, if
787 * a task holds cgroup_mutex on a cgroup with zero count, it
788 * knows that the cgroup won't be removed, as cgroup_rmdir()
791 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
792 * (usually) take cgroup_mutex. These are the two most performance
793 * critical pieces of code here. The exception occurs on cgroup_exit(),
794 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
795 * is taken, and if the cgroup count is zero, a usermode call made
796 * to the release agent with the name of the cgroup (path relative to
797 * the root of cgroup file system) as the argument.
799 * A cgroup can only be deleted if both its 'count' of using tasks
800 * is zero, and its list of 'children' cgroups is empty. Since all
801 * tasks in the system use _some_ cgroup, and since there is always at
802 * least one task in the system (init, pid == 1), therefore, top_cgroup
803 * always has either children cgroups and/or using tasks. So we don't
804 * need a special hack to ensure that top_cgroup cannot be deleted.
806 * The task_lock() exception
808 * The need for this exception arises from the action of
809 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
810 * another. It does so using cgroup_mutex, however there are
811 * several performance critical places that need to reference
812 * task->cgroup without the expense of grabbing a system global
813 * mutex. Therefore except as noted below, when dereferencing or, as
814 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
815 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
816 * the task_struct routinely used for such matters.
818 * P.S. One more locking exception. RCU is used to guard the
819 * update of a tasks cgroup pointer by cgroup_attach_task()
823 * A couple of forward declarations required, due to cyclic reference loop:
824 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
825 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
829 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode);
830 static struct dentry *cgroup_lookup(struct inode *, struct dentry *, unsigned int);
831 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
832 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
833 static const struct inode_operations cgroup_dir_inode_operations;
834 static const struct file_operations proc_cgroupstats_operations;
836 static struct backing_dev_info cgroup_backing_dev_info = {
838 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
841 static int alloc_css_id(struct cgroup_subsys_state *child_css);
843 static struct inode *cgroup_new_inode(umode_t mode, struct super_block *sb)
845 struct inode *inode = new_inode(sb);
848 inode->i_ino = get_next_ino();
849 inode->i_mode = mode;
850 inode->i_uid = current_fsuid();
851 inode->i_gid = current_fsgid();
852 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
853 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
858 static struct cgroup_name *cgroup_alloc_name(struct dentry *dentry)
860 struct cgroup_name *name;
862 name = kmalloc(sizeof(*name) + dentry->d_name.len + 1, GFP_KERNEL);
865 strcpy(name->name, dentry->d_name.name);
869 static void cgroup_free_fn(struct work_struct *work)
871 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
873 mutex_lock(&cgroup_mutex);
874 cgrp->root->number_of_cgroups--;
875 mutex_unlock(&cgroup_mutex);
878 * We get a ref to the parent's dentry, and put the ref when
879 * this cgroup is being freed, so it's guaranteed that the
880 * parent won't be destroyed before its children.
882 dput(cgrp->parent->dentry);
885 * Drop the active superblock reference that we took when we
886 * created the cgroup. This will free cgrp->root, if we are
887 * holding the last reference to @sb.
889 deactivate_super(cgrp->root->sb);
892 * if we're getting rid of the cgroup, refcount should ensure
893 * that there are no pidlists left.
895 BUG_ON(!list_empty(&cgrp->pidlists));
897 simple_xattrs_free(&cgrp->xattrs);
899 kfree(rcu_dereference_raw(cgrp->name));
903 static void cgroup_free_rcu(struct rcu_head *head)
905 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
907 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
908 schedule_work(&cgrp->destroy_work);
911 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
913 /* is dentry a directory ? if so, kfree() associated cgroup */
914 if (S_ISDIR(inode->i_mode)) {
915 struct cgroup *cgrp = dentry->d_fsdata;
917 BUG_ON(!(cgroup_is_dead(cgrp)));
918 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
920 struct cfent *cfe = __d_cfe(dentry);
921 struct cgroup *cgrp = dentry->d_parent->d_fsdata;
923 WARN_ONCE(!list_empty(&cfe->node) &&
924 cgrp != &cgrp->root->top_cgroup,
925 "cfe still linked for %s\n", cfe->type->name);
926 simple_xattrs_free(&cfe->xattrs);
932 static int cgroup_delete(const struct dentry *d)
937 static void remove_dir(struct dentry *d)
939 struct dentry *parent = dget(d->d_parent);
942 simple_rmdir(parent->d_inode, d);
946 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
950 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
951 lockdep_assert_held(&cgroup_mutex);
954 * If we're doing cleanup due to failure of cgroup_create(),
955 * the corresponding @cfe may not exist.
957 list_for_each_entry(cfe, &cgrp->files, node) {
958 struct dentry *d = cfe->dentry;
960 if (cft && cfe->type != cft)
965 simple_unlink(cgrp->dentry->d_inode, d);
966 list_del_init(&cfe->node);
974 * cgroup_clear_dir - remove subsys files in a cgroup directory
975 * @cgrp: target cgroup
976 * @subsys_mask: mask of the subsystem ids whose files should be removed
978 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
980 struct cgroup_subsys *ss;
983 for_each_subsys(ss, i) {
984 struct cftype_set *set;
986 if (!test_bit(i, &subsys_mask))
988 list_for_each_entry(set, &ss->cftsets, node)
989 cgroup_addrm_files(cgrp, set->cfts, false);
994 * NOTE : the dentry must have been dget()'ed
996 static void cgroup_d_remove_dir(struct dentry *dentry)
998 struct dentry *parent;
1000 parent = dentry->d_parent;
1001 spin_lock(&parent->d_lock);
1002 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1003 list_del_init(&dentry->d_u.d_child);
1004 spin_unlock(&dentry->d_lock);
1005 spin_unlock(&parent->d_lock);
1010 * Call with cgroup_mutex held. Drops reference counts on modules, including
1011 * any duplicate ones that parse_cgroupfs_options took. If this function
1012 * returns an error, no reference counts are touched.
1014 static int rebind_subsystems(struct cgroupfs_root *root,
1015 unsigned long added_mask, unsigned removed_mask)
1017 struct cgroup *cgrp = &root->top_cgroup;
1018 struct cgroup_subsys *ss;
1019 unsigned long pinned = 0;
1022 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1023 BUG_ON(!mutex_is_locked(&cgroup_root_mutex));
1025 /* Check that any added subsystems are currently free */
1026 for_each_subsys(ss, i) {
1027 if (!(added_mask & (1 << i)))
1030 /* is the subsystem mounted elsewhere? */
1031 if (ss->root != &cgroup_dummy_root) {
1036 /* pin the module */
1037 if (!try_module_get(ss->module)) {
1044 /* subsys could be missing if unloaded between parsing and here */
1045 if (added_mask != pinned) {
1050 ret = cgroup_populate_dir(cgrp, added_mask);
1055 * Nothing can fail from this point on. Remove files for the
1056 * removed subsystems and rebind each subsystem.
1058 cgroup_clear_dir(cgrp, removed_mask);
1060 for_each_subsys(ss, i) {
1061 unsigned long bit = 1UL << i;
1063 if (bit & added_mask) {
1064 /* We're binding this subsystem to this hierarchy */
1065 BUG_ON(cgroup_css(cgrp, i));
1066 BUG_ON(!cgroup_css(cgroup_dummy_top, i));
1067 BUG_ON(cgroup_css(cgroup_dummy_top, i)->cgroup != cgroup_dummy_top);
1069 rcu_assign_pointer(cgrp->subsys[i],
1070 cgroup_css(cgroup_dummy_top, i));
1071 cgroup_css(cgrp, i)->cgroup = cgrp;
1073 list_move(&ss->sibling, &root->subsys_list);
1076 ss->bind(cgroup_css(cgrp, i));
1078 /* refcount was already taken, and we're keeping it */
1079 root->subsys_mask |= bit;
1080 } else if (bit & removed_mask) {
1081 /* We're removing this subsystem */
1082 BUG_ON(cgroup_css(cgrp, i) != cgroup_css(cgroup_dummy_top, i));
1083 BUG_ON(cgroup_css(cgrp, i)->cgroup != cgrp);
1086 ss->bind(cgroup_css(cgroup_dummy_top, i));
1088 cgroup_css(cgroup_dummy_top, i)->cgroup = cgroup_dummy_top;
1089 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1091 cgroup_subsys[i]->root = &cgroup_dummy_root;
1092 list_move(&ss->sibling, &cgroup_dummy_root.subsys_list);
1094 /* subsystem is now free - drop reference on module */
1095 module_put(ss->module);
1096 root->subsys_mask &= ~bit;
1101 * Mark @root has finished binding subsystems. @root->subsys_mask
1102 * now matches the bound subsystems.
1104 root->flags |= CGRP_ROOT_SUBSYS_BOUND;
1109 for_each_subsys(ss, i)
1110 if (pinned & (1 << i))
1111 module_put(ss->module);
1115 static int cgroup_show_options(struct seq_file *seq, struct dentry *dentry)
1117 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
1118 struct cgroup_subsys *ss;
1120 mutex_lock(&cgroup_root_mutex);
1121 for_each_root_subsys(root, ss)
1122 seq_printf(seq, ",%s", ss->name);
1123 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1124 seq_puts(seq, ",sane_behavior");
1125 if (root->flags & CGRP_ROOT_NOPREFIX)
1126 seq_puts(seq, ",noprefix");
1127 if (root->flags & CGRP_ROOT_XATTR)
1128 seq_puts(seq, ",xattr");
1129 if (strlen(root->release_agent_path))
1130 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1131 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1132 seq_puts(seq, ",clone_children");
1133 if (strlen(root->name))
1134 seq_printf(seq, ",name=%s", root->name);
1135 mutex_unlock(&cgroup_root_mutex);
1139 struct cgroup_sb_opts {
1140 unsigned long subsys_mask;
1141 unsigned long flags;
1142 char *release_agent;
1143 bool cpuset_clone_children;
1145 /* User explicitly requested empty subsystem */
1148 struct cgroupfs_root *new_root;
1153 * Convert a hierarchy specifier into a bitmask of subsystems and
1154 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1155 * array. This function takes refcounts on subsystems to be used, unless it
1156 * returns error, in which case no refcounts are taken.
1158 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1160 char *token, *o = data;
1161 bool all_ss = false, one_ss = false;
1162 unsigned long mask = (unsigned long)-1;
1163 struct cgroup_subsys *ss;
1166 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1168 #ifdef CONFIG_CPUSETS
1169 mask = ~(1UL << cpuset_subsys_id);
1172 memset(opts, 0, sizeof(*opts));
1174 while ((token = strsep(&o, ",")) != NULL) {
1177 if (!strcmp(token, "none")) {
1178 /* Explicitly have no subsystems */
1182 if (!strcmp(token, "all")) {
1183 /* Mutually exclusive option 'all' + subsystem name */
1189 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1190 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1193 if (!strcmp(token, "noprefix")) {
1194 opts->flags |= CGRP_ROOT_NOPREFIX;
1197 if (!strcmp(token, "clone_children")) {
1198 opts->cpuset_clone_children = true;
1201 if (!strcmp(token, "xattr")) {
1202 opts->flags |= CGRP_ROOT_XATTR;
1205 if (!strncmp(token, "release_agent=", 14)) {
1206 /* Specifying two release agents is forbidden */
1207 if (opts->release_agent)
1209 opts->release_agent =
1210 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1211 if (!opts->release_agent)
1215 if (!strncmp(token, "name=", 5)) {
1216 const char *name = token + 5;
1217 /* Can't specify an empty name */
1220 /* Must match [\w.-]+ */
1221 for (i = 0; i < strlen(name); i++) {
1225 if ((c == '.') || (c == '-') || (c == '_'))
1229 /* Specifying two names is forbidden */
1232 opts->name = kstrndup(name,
1233 MAX_CGROUP_ROOT_NAMELEN - 1,
1241 for_each_subsys(ss, i) {
1242 if (strcmp(token, ss->name))
1247 /* Mutually exclusive option 'all' + subsystem name */
1250 set_bit(i, &opts->subsys_mask);
1255 if (i == CGROUP_SUBSYS_COUNT)
1260 * If the 'all' option was specified select all the subsystems,
1261 * otherwise if 'none', 'name=' and a subsystem name options
1262 * were not specified, let's default to 'all'
1264 if (all_ss || (!one_ss && !opts->none && !opts->name))
1265 for_each_subsys(ss, i)
1267 set_bit(i, &opts->subsys_mask);
1269 /* Consistency checks */
1271 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1272 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1274 if (opts->flags & CGRP_ROOT_NOPREFIX) {
1275 pr_err("cgroup: sane_behavior: noprefix is not allowed\n");
1279 if (opts->cpuset_clone_children) {
1280 pr_err("cgroup: sane_behavior: clone_children is not allowed\n");
1286 * Option noprefix was introduced just for backward compatibility
1287 * with the old cpuset, so we allow noprefix only if mounting just
1288 * the cpuset subsystem.
1290 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1294 /* Can't specify "none" and some subsystems */
1295 if (opts->subsys_mask && opts->none)
1299 * We either have to specify by name or by subsystems. (So all
1300 * empty hierarchies must have a name).
1302 if (!opts->subsys_mask && !opts->name)
1308 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1311 struct cgroupfs_root *root = sb->s_fs_info;
1312 struct cgroup *cgrp = &root->top_cgroup;
1313 struct cgroup_sb_opts opts;
1314 unsigned long added_mask, removed_mask;
1316 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1317 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1321 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1322 mutex_lock(&cgroup_mutex);
1323 mutex_lock(&cgroup_root_mutex);
1325 /* See what subsystems are wanted */
1326 ret = parse_cgroupfs_options(data, &opts);
1330 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1331 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1332 task_tgid_nr(current), current->comm);
1334 added_mask = opts.subsys_mask & ~root->subsys_mask;
1335 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1337 /* Don't allow flags or name to change at remount */
1338 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1339 (opts.name && strcmp(opts.name, root->name))) {
1340 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1341 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1342 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1347 /* remounting is not allowed for populated hierarchies */
1348 if (root->number_of_cgroups > 1) {
1353 ret = rebind_subsystems(root, added_mask, removed_mask);
1357 if (opts.release_agent)
1358 strcpy(root->release_agent_path, opts.release_agent);
1360 kfree(opts.release_agent);
1362 mutex_unlock(&cgroup_root_mutex);
1363 mutex_unlock(&cgroup_mutex);
1364 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1368 static const struct super_operations cgroup_ops = {
1369 .statfs = simple_statfs,
1370 .drop_inode = generic_delete_inode,
1371 .show_options = cgroup_show_options,
1372 .remount_fs = cgroup_remount,
1375 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1377 INIT_LIST_HEAD(&cgrp->sibling);
1378 INIT_LIST_HEAD(&cgrp->children);
1379 INIT_LIST_HEAD(&cgrp->files);
1380 INIT_LIST_HEAD(&cgrp->cset_links);
1381 INIT_LIST_HEAD(&cgrp->release_list);
1382 INIT_LIST_HEAD(&cgrp->pidlists);
1383 mutex_init(&cgrp->pidlist_mutex);
1384 cgrp->dummy_css.cgroup = cgrp;
1385 INIT_LIST_HEAD(&cgrp->event_list);
1386 spin_lock_init(&cgrp->event_list_lock);
1387 simple_xattrs_init(&cgrp->xattrs);
1390 static void init_cgroup_root(struct cgroupfs_root *root)
1392 struct cgroup *cgrp = &root->top_cgroup;
1394 INIT_LIST_HEAD(&root->subsys_list);
1395 INIT_LIST_HEAD(&root->root_list);
1396 root->number_of_cgroups = 1;
1398 RCU_INIT_POINTER(cgrp->name, &root_cgroup_name);
1399 init_cgroup_housekeeping(cgrp);
1400 idr_init(&root->cgroup_idr);
1403 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
1407 lockdep_assert_held(&cgroup_mutex);
1408 lockdep_assert_held(&cgroup_root_mutex);
1410 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
1415 root->hierarchy_id = id;
1419 static void cgroup_exit_root_id(struct cgroupfs_root *root)
1421 lockdep_assert_held(&cgroup_mutex);
1422 lockdep_assert_held(&cgroup_root_mutex);
1424 if (root->hierarchy_id) {
1425 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1426 root->hierarchy_id = 0;
1430 static int cgroup_test_super(struct super_block *sb, void *data)
1432 struct cgroup_sb_opts *opts = data;
1433 struct cgroupfs_root *root = sb->s_fs_info;
1435 /* If we asked for a name then it must match */
1436 if (opts->name && strcmp(opts->name, root->name))
1440 * If we asked for subsystems (or explicitly for no
1441 * subsystems) then they must match
1443 if ((opts->subsys_mask || opts->none)
1444 && (opts->subsys_mask != root->subsys_mask))
1450 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1452 struct cgroupfs_root *root;
1454 if (!opts->subsys_mask && !opts->none)
1457 root = kzalloc(sizeof(*root), GFP_KERNEL);
1459 return ERR_PTR(-ENOMEM);
1461 init_cgroup_root(root);
1464 * We need to set @root->subsys_mask now so that @root can be
1465 * matched by cgroup_test_super() before it finishes
1466 * initialization; otherwise, competing mounts with the same
1467 * options may try to bind the same subsystems instead of waiting
1468 * for the first one leading to unexpected mount errors.
1469 * SUBSYS_BOUND will be set once actual binding is complete.
1471 root->subsys_mask = opts->subsys_mask;
1472 root->flags = opts->flags;
1473 if (opts->release_agent)
1474 strcpy(root->release_agent_path, opts->release_agent);
1476 strcpy(root->name, opts->name);
1477 if (opts->cpuset_clone_children)
1478 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1482 static void cgroup_free_root(struct cgroupfs_root *root)
1485 /* hierarhcy ID shoulid already have been released */
1486 WARN_ON_ONCE(root->hierarchy_id);
1488 idr_destroy(&root->cgroup_idr);
1493 static int cgroup_set_super(struct super_block *sb, void *data)
1496 struct cgroup_sb_opts *opts = data;
1498 /* If we don't have a new root, we can't set up a new sb */
1499 if (!opts->new_root)
1502 BUG_ON(!opts->subsys_mask && !opts->none);
1504 ret = set_anon_super(sb, NULL);
1508 sb->s_fs_info = opts->new_root;
1509 opts->new_root->sb = sb;
1511 sb->s_blocksize = PAGE_CACHE_SIZE;
1512 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1513 sb->s_magic = CGROUP_SUPER_MAGIC;
1514 sb->s_op = &cgroup_ops;
1519 static int cgroup_get_rootdir(struct super_block *sb)
1521 static const struct dentry_operations cgroup_dops = {
1522 .d_iput = cgroup_diput,
1523 .d_delete = cgroup_delete,
1526 struct inode *inode =
1527 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1532 inode->i_fop = &simple_dir_operations;
1533 inode->i_op = &cgroup_dir_inode_operations;
1534 /* directories start off with i_nlink == 2 (for "." entry) */
1536 sb->s_root = d_make_root(inode);
1539 /* for everything else we want ->d_op set */
1540 sb->s_d_op = &cgroup_dops;
1544 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1545 int flags, const char *unused_dev_name,
1548 struct cgroup_sb_opts opts;
1549 struct cgroupfs_root *root;
1551 struct super_block *sb;
1552 struct cgroupfs_root *new_root;
1553 struct list_head tmp_links;
1554 struct inode *inode;
1555 const struct cred *cred;
1557 /* First find the desired set of subsystems */
1558 mutex_lock(&cgroup_mutex);
1559 ret = parse_cgroupfs_options(data, &opts);
1560 mutex_unlock(&cgroup_mutex);
1565 * Allocate a new cgroup root. We may not need it if we're
1566 * reusing an existing hierarchy.
1568 new_root = cgroup_root_from_opts(&opts);
1569 if (IS_ERR(new_root)) {
1570 ret = PTR_ERR(new_root);
1573 opts.new_root = new_root;
1575 /* Locate an existing or new sb for this hierarchy */
1576 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, 0, &opts);
1579 cgroup_free_root(opts.new_root);
1583 root = sb->s_fs_info;
1585 if (root == opts.new_root) {
1586 /* We used the new root structure, so this is a new hierarchy */
1587 struct cgroup *root_cgrp = &root->top_cgroup;
1588 struct cgroupfs_root *existing_root;
1590 struct css_set *cset;
1592 BUG_ON(sb->s_root != NULL);
1594 ret = cgroup_get_rootdir(sb);
1596 goto drop_new_super;
1597 inode = sb->s_root->d_inode;
1599 mutex_lock(&inode->i_mutex);
1600 mutex_lock(&cgroup_mutex);
1601 mutex_lock(&cgroup_root_mutex);
1603 root_cgrp->id = idr_alloc(&root->cgroup_idr, root_cgrp,
1605 if (root_cgrp->id < 0)
1608 /* Check for name clashes with existing mounts */
1610 if (strlen(root->name))
1611 for_each_active_root(existing_root)
1612 if (!strcmp(existing_root->name, root->name))
1616 * We're accessing css_set_count without locking
1617 * css_set_lock here, but that's OK - it can only be
1618 * increased by someone holding cgroup_lock, and
1619 * that's us. The worst that can happen is that we
1620 * have some link structures left over
1622 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1626 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1627 ret = cgroup_init_root_id(root, 2, 0);
1631 sb->s_root->d_fsdata = root_cgrp;
1632 root_cgrp->dentry = sb->s_root;
1635 * We're inside get_sb() and will call lookup_one_len() to
1636 * create the root files, which doesn't work if SELinux is
1637 * in use. The following cred dancing somehow works around
1638 * it. See 2ce9738ba ("cgroupfs: use init_cred when
1639 * populating new cgroupfs mount") for more details.
1641 cred = override_creds(&init_cred);
1643 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1647 ret = rebind_subsystems(root, root->subsys_mask, 0);
1654 * There must be no failure case after here, since rebinding
1655 * takes care of subsystems' refcounts, which are explicitly
1656 * dropped in the failure exit path.
1659 list_add(&root->root_list, &cgroup_roots);
1660 cgroup_root_count++;
1662 /* Link the top cgroup in this hierarchy into all
1663 * the css_set objects */
1664 write_lock(&css_set_lock);
1665 hash_for_each(css_set_table, i, cset, hlist)
1666 link_css_set(&tmp_links, cset, root_cgrp);
1667 write_unlock(&css_set_lock);
1669 free_cgrp_cset_links(&tmp_links);
1671 BUG_ON(!list_empty(&root_cgrp->children));
1672 BUG_ON(root->number_of_cgroups != 1);
1674 mutex_unlock(&cgroup_root_mutex);
1675 mutex_unlock(&cgroup_mutex);
1676 mutex_unlock(&inode->i_mutex);
1679 * We re-used an existing hierarchy - the new root (if
1680 * any) is not needed
1682 cgroup_free_root(opts.new_root);
1684 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1685 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1686 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1688 goto drop_new_super;
1690 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1695 kfree(opts.release_agent);
1697 return dget(sb->s_root);
1700 free_cgrp_cset_links(&tmp_links);
1701 cgroup_addrm_files(&root->top_cgroup, cgroup_base_files, false);
1704 cgroup_exit_root_id(root);
1705 mutex_unlock(&cgroup_root_mutex);
1706 mutex_unlock(&cgroup_mutex);
1707 mutex_unlock(&inode->i_mutex);
1709 deactivate_locked_super(sb);
1711 kfree(opts.release_agent);
1713 return ERR_PTR(ret);
1716 static void cgroup_kill_sb(struct super_block *sb) {
1717 struct cgroupfs_root *root = sb->s_fs_info;
1718 struct cgroup *cgrp = &root->top_cgroup;
1719 struct cgrp_cset_link *link, *tmp_link;
1724 BUG_ON(root->number_of_cgroups != 1);
1725 BUG_ON(!list_empty(&cgrp->children));
1727 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1728 mutex_lock(&cgroup_mutex);
1729 mutex_lock(&cgroup_root_mutex);
1731 /* Rebind all subsystems back to the default hierarchy */
1732 if (root->flags & CGRP_ROOT_SUBSYS_BOUND) {
1733 ret = rebind_subsystems(root, 0, root->subsys_mask);
1734 /* Shouldn't be able to fail ... */
1739 * Release all the links from cset_links to this hierarchy's
1742 write_lock(&css_set_lock);
1744 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1745 list_del(&link->cset_link);
1746 list_del(&link->cgrp_link);
1749 write_unlock(&css_set_lock);
1751 if (!list_empty(&root->root_list)) {
1752 list_del(&root->root_list);
1753 cgroup_root_count--;
1756 cgroup_exit_root_id(root);
1758 mutex_unlock(&cgroup_root_mutex);
1759 mutex_unlock(&cgroup_mutex);
1760 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1762 simple_xattrs_free(&cgrp->xattrs);
1764 kill_litter_super(sb);
1765 cgroup_free_root(root);
1768 static struct file_system_type cgroup_fs_type = {
1770 .mount = cgroup_mount,
1771 .kill_sb = cgroup_kill_sb,
1774 static struct kobject *cgroup_kobj;
1777 * cgroup_path - generate the path of a cgroup
1778 * @cgrp: the cgroup in question
1779 * @buf: the buffer to write the path into
1780 * @buflen: the length of the buffer
1782 * Writes path of cgroup into buf. Returns 0 on success, -errno on error.
1784 * We can't generate cgroup path using dentry->d_name, as accessing
1785 * dentry->name must be protected by irq-unsafe dentry->d_lock or parent
1786 * inode's i_mutex, while on the other hand cgroup_path() can be called
1787 * with some irq-safe spinlocks held.
1789 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1791 int ret = -ENAMETOOLONG;
1794 if (!cgrp->parent) {
1795 if (strlcpy(buf, "/", buflen) >= buflen)
1796 return -ENAMETOOLONG;
1800 start = buf + buflen - 1;
1805 const char *name = cgroup_name(cgrp);
1809 if ((start -= len) < buf)
1811 memcpy(start, name, len);
1817 cgrp = cgrp->parent;
1818 } while (cgrp->parent);
1820 memmove(buf, start, buf + buflen - start);
1825 EXPORT_SYMBOL_GPL(cgroup_path);
1828 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1829 * @task: target task
1830 * @buf: the buffer to write the path into
1831 * @buflen: the length of the buffer
1833 * Determine @task's cgroup on the first (the one with the lowest non-zero
1834 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1835 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1836 * cgroup controller callbacks.
1838 * Returns 0 on success, fails with -%ENAMETOOLONG if @buflen is too short.
1840 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1842 struct cgroupfs_root *root;
1843 struct cgroup *cgrp;
1844 int hierarchy_id = 1, ret = 0;
1847 return -ENAMETOOLONG;
1849 mutex_lock(&cgroup_mutex);
1851 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1854 cgrp = task_cgroup_from_root(task, root);
1855 ret = cgroup_path(cgrp, buf, buflen);
1857 /* if no hierarchy exists, everyone is in "/" */
1858 memcpy(buf, "/", 2);
1861 mutex_unlock(&cgroup_mutex);
1864 EXPORT_SYMBOL_GPL(task_cgroup_path);
1867 * Control Group taskset
1869 struct task_and_cgroup {
1870 struct task_struct *task;
1871 struct cgroup *cgrp;
1872 struct css_set *cset;
1875 struct cgroup_taskset {
1876 struct task_and_cgroup single;
1877 struct flex_array *tc_array;
1880 struct cgroup *cur_cgrp;
1884 * cgroup_taskset_first - reset taskset and return the first task
1885 * @tset: taskset of interest
1887 * @tset iteration is initialized and the first task is returned.
1889 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1891 if (tset->tc_array) {
1893 return cgroup_taskset_next(tset);
1895 tset->cur_cgrp = tset->single.cgrp;
1896 return tset->single.task;
1899 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1902 * cgroup_taskset_next - iterate to the next task in taskset
1903 * @tset: taskset of interest
1905 * Return the next task in @tset. Iteration must have been initialized
1906 * with cgroup_taskset_first().
1908 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1910 struct task_and_cgroup *tc;
1912 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1915 tc = flex_array_get(tset->tc_array, tset->idx++);
1916 tset->cur_cgrp = tc->cgrp;
1919 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1922 * cgroup_taskset_cur_css - return the matching css for the current task
1923 * @tset: taskset of interest
1924 * @subsys_id: the ID of the target subsystem
1926 * Return the css for the current (last returned) task of @tset for
1927 * subsystem specified by @subsys_id. This function must be preceded by
1928 * either cgroup_taskset_first() or cgroup_taskset_next().
1930 struct cgroup_subsys_state *cgroup_taskset_cur_css(struct cgroup_taskset *tset,
1933 return cgroup_css(tset->cur_cgrp, subsys_id);
1935 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_css);
1938 * cgroup_taskset_size - return the number of tasks in taskset
1939 * @tset: taskset of interest
1941 int cgroup_taskset_size(struct cgroup_taskset *tset)
1943 return tset->tc_array ? tset->tc_array_len : 1;
1945 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1949 * cgroup_task_migrate - move a task from one cgroup to another.
1951 * Must be called with cgroup_mutex and threadgroup locked.
1953 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1954 struct task_struct *tsk,
1955 struct css_set *new_cset)
1957 struct css_set *old_cset;
1960 * We are synchronized through threadgroup_lock() against PF_EXITING
1961 * setting such that we can't race against cgroup_exit() changing the
1962 * css_set to init_css_set and dropping the old one.
1964 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1965 old_cset = task_css_set(tsk);
1968 rcu_assign_pointer(tsk->cgroups, new_cset);
1971 /* Update the css_set linked lists if we're using them */
1972 write_lock(&css_set_lock);
1973 if (!list_empty(&tsk->cg_list))
1974 list_move(&tsk->cg_list, &new_cset->tasks);
1975 write_unlock(&css_set_lock);
1978 * We just gained a reference on old_cset by taking it from the
1979 * task. As trading it for new_cset is protected by cgroup_mutex,
1980 * we're safe to drop it here; it will be freed under RCU.
1982 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1983 put_css_set(old_cset);
1987 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1988 * @cgrp: the cgroup to attach to
1989 * @tsk: the task or the leader of the threadgroup to be attached
1990 * @threadgroup: attach the whole threadgroup?
1992 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1993 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1995 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1998 int retval, i, group_size;
1999 struct cgroup_subsys *ss, *failed_ss = NULL;
2000 struct cgroupfs_root *root = cgrp->root;
2001 /* threadgroup list cursor and array */
2002 struct task_struct *leader = tsk;
2003 struct task_and_cgroup *tc;
2004 struct flex_array *group;
2005 struct cgroup_taskset tset = { };
2008 * step 0: in order to do expensive, possibly blocking operations for
2009 * every thread, we cannot iterate the thread group list, since it needs
2010 * rcu or tasklist locked. instead, build an array of all threads in the
2011 * group - group_rwsem prevents new threads from appearing, and if
2012 * threads exit, this will just be an over-estimate.
2015 group_size = get_nr_threads(tsk);
2018 /* flex_array supports very large thread-groups better than kmalloc. */
2019 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
2022 /* pre-allocate to guarantee space while iterating in rcu read-side. */
2023 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
2025 goto out_free_group_list;
2029 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2030 * already PF_EXITING could be freed from underneath us unless we
2031 * take an rcu_read_lock.
2035 struct task_and_cgroup ent;
2037 /* @tsk either already exited or can't exit until the end */
2038 if (tsk->flags & PF_EXITING)
2041 /* as per above, nr_threads may decrease, but not increase. */
2042 BUG_ON(i >= group_size);
2044 ent.cgrp = task_cgroup_from_root(tsk, root);
2045 /* nothing to do if this task is already in the cgroup */
2046 if (ent.cgrp == cgrp)
2049 * saying GFP_ATOMIC has no effect here because we did prealloc
2050 * earlier, but it's good form to communicate our expectations.
2052 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
2053 BUG_ON(retval != 0);
2058 } while_each_thread(leader, tsk);
2060 /* remember the number of threads in the array for later. */
2062 tset.tc_array = group;
2063 tset.tc_array_len = group_size;
2065 /* methods shouldn't be called if no task is actually migrating */
2068 goto out_free_group_list;
2071 * step 1: check that we can legitimately attach to the cgroup.
2073 for_each_root_subsys(root, ss) {
2074 struct cgroup_subsys_state *css = cgroup_css(cgrp, ss->subsys_id);
2076 if (ss->can_attach) {
2077 retval = ss->can_attach(css, &tset);
2080 goto out_cancel_attach;
2086 * step 2: make sure css_sets exist for all threads to be migrated.
2087 * we use find_css_set, which allocates a new one if necessary.
2089 for (i = 0; i < group_size; i++) {
2090 struct css_set *old_cset;
2092 tc = flex_array_get(group, i);
2093 old_cset = task_css_set(tc->task);
2094 tc->cset = find_css_set(old_cset, cgrp);
2097 goto out_put_css_set_refs;
2102 * step 3: now that we're guaranteed success wrt the css_sets,
2103 * proceed to move all tasks to the new cgroup. There are no
2104 * failure cases after here, so this is the commit point.
2106 for (i = 0; i < group_size; i++) {
2107 tc = flex_array_get(group, i);
2108 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
2110 /* nothing is sensitive to fork() after this point. */
2113 * step 4: do subsystem attach callbacks.
2115 for_each_root_subsys(root, ss) {
2116 struct cgroup_subsys_state *css = cgroup_css(cgrp, ss->subsys_id);
2119 ss->attach(css, &tset);
2123 * step 5: success! and cleanup
2126 out_put_css_set_refs:
2128 for (i = 0; i < group_size; i++) {
2129 tc = flex_array_get(group, i);
2132 put_css_set(tc->cset);
2137 for_each_root_subsys(root, ss) {
2138 struct cgroup_subsys_state *css = cgroup_css(cgrp, ss->subsys_id);
2140 if (ss == failed_ss)
2142 if (ss->cancel_attach)
2143 ss->cancel_attach(css, &tset);
2146 out_free_group_list:
2147 flex_array_free(group);
2152 * Find the task_struct of the task to attach by vpid and pass it along to the
2153 * function to attach either it or all tasks in its threadgroup. Will lock
2154 * cgroup_mutex and threadgroup; may take task_lock of task.
2156 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2158 struct task_struct *tsk;
2159 const struct cred *cred = current_cred(), *tcred;
2162 if (!cgroup_lock_live_group(cgrp))
2168 tsk = find_task_by_vpid(pid);
2172 goto out_unlock_cgroup;
2175 * even if we're attaching all tasks in the thread group, we
2176 * only need to check permissions on one of them.
2178 tcred = __task_cred(tsk);
2179 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2180 !uid_eq(cred->euid, tcred->uid) &&
2181 !uid_eq(cred->euid, tcred->suid)) {
2184 goto out_unlock_cgroup;
2190 tsk = tsk->group_leader;
2193 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2194 * trapped in a cpuset, or RT worker may be born in a cgroup
2195 * with no rt_runtime allocated. Just say no.
2197 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2200 goto out_unlock_cgroup;
2203 get_task_struct(tsk);
2206 threadgroup_lock(tsk);
2208 if (!thread_group_leader(tsk)) {
2210 * a race with de_thread from another thread's exec()
2211 * may strip us of our leadership, if this happens,
2212 * there is no choice but to throw this task away and
2213 * try again; this is
2214 * "double-double-toil-and-trouble-check locking".
2216 threadgroup_unlock(tsk);
2217 put_task_struct(tsk);
2218 goto retry_find_task;
2222 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2224 threadgroup_unlock(tsk);
2226 put_task_struct(tsk);
2228 mutex_unlock(&cgroup_mutex);
2233 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2234 * @from: attach to all cgroups of a given task
2235 * @tsk: the task to be attached
2237 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2239 struct cgroupfs_root *root;
2242 mutex_lock(&cgroup_mutex);
2243 for_each_active_root(root) {
2244 struct cgroup *from_cgrp = task_cgroup_from_root(from, root);
2246 retval = cgroup_attach_task(from_cgrp, tsk, false);
2250 mutex_unlock(&cgroup_mutex);
2254 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2256 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2257 struct cftype *cft, u64 pid)
2259 return attach_task_by_pid(css->cgroup, pid, false);
2262 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2263 struct cftype *cft, u64 tgid)
2265 return attach_task_by_pid(css->cgroup, tgid, true);
2268 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2269 struct cftype *cft, const char *buffer)
2271 BUILD_BUG_ON(sizeof(css->cgroup->root->release_agent_path) < PATH_MAX);
2272 if (strlen(buffer) >= PATH_MAX)
2274 if (!cgroup_lock_live_group(css->cgroup))
2276 mutex_lock(&cgroup_root_mutex);
2277 strcpy(css->cgroup->root->release_agent_path, buffer);
2278 mutex_unlock(&cgroup_root_mutex);
2279 mutex_unlock(&cgroup_mutex);
2283 static int cgroup_release_agent_show(struct cgroup_subsys_state *css,
2284 struct cftype *cft, struct seq_file *seq)
2286 struct cgroup *cgrp = css->cgroup;
2288 if (!cgroup_lock_live_group(cgrp))
2290 seq_puts(seq, cgrp->root->release_agent_path);
2291 seq_putc(seq, '\n');
2292 mutex_unlock(&cgroup_mutex);
2296 static int cgroup_sane_behavior_show(struct cgroup_subsys_state *css,
2297 struct cftype *cft, struct seq_file *seq)
2299 seq_printf(seq, "%d\n", cgroup_sane_behavior(css->cgroup));
2303 /* A buffer size big enough for numbers or short strings */
2304 #define CGROUP_LOCAL_BUFFER_SIZE 64
2306 static ssize_t cgroup_write_X64(struct cgroup_subsys_state *css,
2307 struct cftype *cft, struct file *file,
2308 const char __user *userbuf, size_t nbytes,
2309 loff_t *unused_ppos)
2311 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
2317 if (nbytes >= sizeof(buffer))
2319 if (copy_from_user(buffer, userbuf, nbytes))
2322 buffer[nbytes] = 0; /* nul-terminate */
2323 if (cft->write_u64) {
2324 u64 val = simple_strtoull(strstrip(buffer), &end, 0);
2327 retval = cft->write_u64(css, cft, val);
2329 s64 val = simple_strtoll(strstrip(buffer), &end, 0);
2332 retval = cft->write_s64(css, cft, val);
2339 static ssize_t cgroup_write_string(struct cgroup_subsys_state *css,
2340 struct cftype *cft, struct file *file,
2341 const char __user *userbuf, size_t nbytes,
2342 loff_t *unused_ppos)
2344 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
2346 size_t max_bytes = cft->max_write_len;
2347 char *buffer = local_buffer;
2350 max_bytes = sizeof(local_buffer) - 1;
2351 if (nbytes >= max_bytes)
2353 /* Allocate a dynamic buffer if we need one */
2354 if (nbytes >= sizeof(local_buffer)) {
2355 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
2359 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
2364 buffer[nbytes] = 0; /* nul-terminate */
2365 retval = cft->write_string(css, cft, strstrip(buffer));
2369 if (buffer != local_buffer)
2374 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
2375 size_t nbytes, loff_t *ppos)
2377 struct cfent *cfe = __d_cfe(file->f_dentry);
2378 struct cftype *cft = __d_cft(file->f_dentry);
2379 struct cgroup_subsys_state *css = cfe->css;
2382 return cft->write(css, cft, file, buf, nbytes, ppos);
2383 if (cft->write_u64 || cft->write_s64)
2384 return cgroup_write_X64(css, cft, file, buf, nbytes, ppos);
2385 if (cft->write_string)
2386 return cgroup_write_string(css, cft, file, buf, nbytes, ppos);
2388 int ret = cft->trigger(css, (unsigned int)cft->private);
2389 return ret ? ret : nbytes;
2394 static ssize_t cgroup_read_u64(struct cgroup_subsys_state *css,
2395 struct cftype *cft, struct file *file,
2396 char __user *buf, size_t nbytes, loff_t *ppos)
2398 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2399 u64 val = cft->read_u64(css, cft);
2400 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
2402 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2405 static ssize_t cgroup_read_s64(struct cgroup_subsys_state *css,
2406 struct cftype *cft, struct file *file,
2407 char __user *buf, size_t nbytes, loff_t *ppos)
2409 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2410 s64 val = cft->read_s64(css, cft);
2411 int len = sprintf(tmp, "%lld\n", (long long) val);
2413 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2416 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2417 size_t nbytes, loff_t *ppos)
2419 struct cfent *cfe = __d_cfe(file->f_dentry);
2420 struct cftype *cft = __d_cft(file->f_dentry);
2421 struct cgroup_subsys_state *css = cfe->css;
2424 return cft->read(css, cft, file, buf, nbytes, ppos);
2426 return cgroup_read_u64(css, cft, file, buf, nbytes, ppos);
2428 return cgroup_read_s64(css, cft, file, buf, nbytes, ppos);
2433 * seqfile ops/methods for returning structured data. Currently just
2434 * supports string->u64 maps, but can be extended in future.
2437 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2439 struct seq_file *sf = cb->state;
2440 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2443 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2445 struct cfent *cfe = m->private;
2446 struct cftype *cft = cfe->type;
2447 struct cgroup_subsys_state *css = cfe->css;
2449 if (cft->read_map) {
2450 struct cgroup_map_cb cb = {
2451 .fill = cgroup_map_add,
2454 return cft->read_map(css, cft, &cb);
2456 return cft->read_seq_string(css, cft, m);
2459 static const struct file_operations cgroup_seqfile_operations = {
2461 .write = cgroup_file_write,
2462 .llseek = seq_lseek,
2463 .release = single_release,
2466 static int cgroup_file_open(struct inode *inode, struct file *file)
2468 struct cfent *cfe = __d_cfe(file->f_dentry);
2469 struct cftype *cft = __d_cft(file->f_dentry);
2470 struct cgroup *cgrp = __d_cgrp(cfe->dentry->d_parent);
2471 struct cgroup_subsys_state *css;
2474 err = generic_file_open(inode, file);
2479 * If the file belongs to a subsystem, pin the css. Will be
2480 * unpinned either on open failure or release. This ensures that
2481 * @css stays alive for all file operations.
2485 css = cgroup_css(cgrp, cft->ss->subsys_id);
2486 if (!css_tryget(css))
2489 css = &cgrp->dummy_css;
2493 /* css should match @cfe->css, see cgroup_add_file() for details */
2494 if (!css || WARN_ON_ONCE(css != cfe->css))
2497 if (cft->read_map || cft->read_seq_string) {
2498 file->f_op = &cgroup_seqfile_operations;
2499 err = single_open(file, cgroup_seqfile_show, cfe);
2500 } else if (cft->open) {
2501 err = cft->open(inode, file);
2509 static int cgroup_file_release(struct inode *inode, struct file *file)
2511 struct cfent *cfe = __d_cfe(file->f_dentry);
2512 struct cftype *cft = __d_cft(file->f_dentry);
2513 struct cgroup_subsys_state *css = cfe->css;
2517 ret = cft->release(inode, file);
2524 * cgroup_rename - Only allow simple rename of directories in place.
2526 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2527 struct inode *new_dir, struct dentry *new_dentry)
2530 struct cgroup_name *name, *old_name;
2531 struct cgroup *cgrp;
2534 * It's convinient to use parent dir's i_mutex to protected
2537 lockdep_assert_held(&old_dir->i_mutex);
2539 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2541 if (new_dentry->d_inode)
2543 if (old_dir != new_dir)
2546 cgrp = __d_cgrp(old_dentry);
2549 * This isn't a proper migration and its usefulness is very
2550 * limited. Disallow if sane_behavior.
2552 if (cgroup_sane_behavior(cgrp))
2555 name = cgroup_alloc_name(new_dentry);
2559 ret = simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2565 old_name = rcu_dereference_protected(cgrp->name, true);
2566 rcu_assign_pointer(cgrp->name, name);
2568 kfree_rcu(old_name, rcu_head);
2572 static struct simple_xattrs *__d_xattrs(struct dentry *dentry)
2574 if (S_ISDIR(dentry->d_inode->i_mode))
2575 return &__d_cgrp(dentry)->xattrs;
2577 return &__d_cfe(dentry)->xattrs;
2580 static inline int xattr_enabled(struct dentry *dentry)
2582 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
2583 return root->flags & CGRP_ROOT_XATTR;
2586 static bool is_valid_xattr(const char *name)
2588 if (!strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
2589 !strncmp(name, XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN))
2594 static int cgroup_setxattr(struct dentry *dentry, const char *name,
2595 const void *val, size_t size, int flags)
2597 if (!xattr_enabled(dentry))
2599 if (!is_valid_xattr(name))
2601 return simple_xattr_set(__d_xattrs(dentry), name, val, size, flags);
2604 static int cgroup_removexattr(struct dentry *dentry, const char *name)
2606 if (!xattr_enabled(dentry))
2608 if (!is_valid_xattr(name))
2610 return simple_xattr_remove(__d_xattrs(dentry), name);
2613 static ssize_t cgroup_getxattr(struct dentry *dentry, const char *name,
2614 void *buf, size_t size)
2616 if (!xattr_enabled(dentry))
2618 if (!is_valid_xattr(name))
2620 return simple_xattr_get(__d_xattrs(dentry), name, buf, size);
2623 static ssize_t cgroup_listxattr(struct dentry *dentry, char *buf, size_t size)
2625 if (!xattr_enabled(dentry))
2627 return simple_xattr_list(__d_xattrs(dentry), buf, size);
2630 static const struct file_operations cgroup_file_operations = {
2631 .read = cgroup_file_read,
2632 .write = cgroup_file_write,
2633 .llseek = generic_file_llseek,
2634 .open = cgroup_file_open,
2635 .release = cgroup_file_release,
2638 static const struct inode_operations cgroup_file_inode_operations = {
2639 .setxattr = cgroup_setxattr,
2640 .getxattr = cgroup_getxattr,
2641 .listxattr = cgroup_listxattr,
2642 .removexattr = cgroup_removexattr,
2645 static const struct inode_operations cgroup_dir_inode_operations = {
2646 .lookup = cgroup_lookup,
2647 .mkdir = cgroup_mkdir,
2648 .rmdir = cgroup_rmdir,
2649 .rename = cgroup_rename,
2650 .setxattr = cgroup_setxattr,
2651 .getxattr = cgroup_getxattr,
2652 .listxattr = cgroup_listxattr,
2653 .removexattr = cgroup_removexattr,
2656 static struct dentry *cgroup_lookup(struct inode *dir, struct dentry *dentry, unsigned int flags)
2658 if (dentry->d_name.len > NAME_MAX)
2659 return ERR_PTR(-ENAMETOOLONG);
2660 d_add(dentry, NULL);
2665 * Check if a file is a control file
2667 static inline struct cftype *__file_cft(struct file *file)
2669 if (file_inode(file)->i_fop != &cgroup_file_operations)
2670 return ERR_PTR(-EINVAL);
2671 return __d_cft(file->f_dentry);
2674 static int cgroup_create_file(struct dentry *dentry, umode_t mode,
2675 struct super_block *sb)
2677 struct inode *inode;
2681 if (dentry->d_inode)
2684 inode = cgroup_new_inode(mode, sb);
2688 if (S_ISDIR(mode)) {
2689 inode->i_op = &cgroup_dir_inode_operations;
2690 inode->i_fop = &simple_dir_operations;
2692 /* start off with i_nlink == 2 (for "." entry) */
2694 inc_nlink(dentry->d_parent->d_inode);
2697 * Control reaches here with cgroup_mutex held.
2698 * @inode->i_mutex should nest outside cgroup_mutex but we
2699 * want to populate it immediately without releasing
2700 * cgroup_mutex. As @inode isn't visible to anyone else
2701 * yet, trylock will always succeed without affecting
2704 WARN_ON_ONCE(!mutex_trylock(&inode->i_mutex));
2705 } else if (S_ISREG(mode)) {
2707 inode->i_fop = &cgroup_file_operations;
2708 inode->i_op = &cgroup_file_inode_operations;
2710 d_instantiate(dentry, inode);
2711 dget(dentry); /* Extra count - pin the dentry in core */
2716 * cgroup_file_mode - deduce file mode of a control file
2717 * @cft: the control file in question
2719 * returns cft->mode if ->mode is not 0
2720 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2721 * returns S_IRUGO if it has only a read handler
2722 * returns S_IWUSR if it has only a write hander
2724 static umode_t cgroup_file_mode(const struct cftype *cft)
2731 if (cft->read || cft->read_u64 || cft->read_s64 ||
2732 cft->read_map || cft->read_seq_string)
2735 if (cft->write || cft->write_u64 || cft->write_s64 ||
2736 cft->write_string || cft->trigger)
2742 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2744 struct dentry *dir = cgrp->dentry;
2745 struct cgroup *parent = __d_cgrp(dir);
2746 struct dentry *dentry;
2750 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2752 if (cft->ss && !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
2753 strcpy(name, cft->ss->name);
2756 strcat(name, cft->name);
2758 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2760 cfe = kzalloc(sizeof(*cfe), GFP_KERNEL);
2764 dentry = lookup_one_len(name, dir, strlen(name));
2765 if (IS_ERR(dentry)) {
2766 error = PTR_ERR(dentry);
2770 cfe->type = (void *)cft;
2771 cfe->dentry = dentry;
2772 dentry->d_fsdata = cfe;
2773 simple_xattrs_init(&cfe->xattrs);
2776 * cfe->css is used by read/write/close to determine the associated
2777 * css. file->private_data would be a better place but that's
2778 * already used by seqfile. Note that open will use the usual
2779 * cgroup_css() and css_tryget() to acquire the css and this
2780 * caching doesn't affect css lifetime management.
2783 cfe->css = cgroup_css(cgrp, cft->ss->subsys_id);
2785 cfe->css = &cgrp->dummy_css;
2787 mode = cgroup_file_mode(cft);
2788 error = cgroup_create_file(dentry, mode | S_IFREG, cgrp->root->sb);
2790 list_add_tail(&cfe->node, &parent->files);
2800 * cgroup_addrm_files - add or remove files to a cgroup directory
2801 * @cgrp: the target cgroup
2802 * @cfts: array of cftypes to be added
2803 * @is_add: whether to add or remove
2805 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2806 * For removals, this function never fails. If addition fails, this
2807 * function doesn't remove files already added. The caller is responsible
2810 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2816 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
2817 lockdep_assert_held(&cgroup_mutex);
2819 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2820 /* does cft->flags tell us to skip this file on @cgrp? */
2821 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2823 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2825 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2829 ret = cgroup_add_file(cgrp, cft);
2831 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2836 cgroup_rm_file(cgrp, cft);
2842 static void cgroup_cfts_prepare(void)
2843 __acquires(&cgroup_mutex)
2846 * Thanks to the entanglement with vfs inode locking, we can't walk
2847 * the existing cgroups under cgroup_mutex and create files.
2848 * Instead, we use css_for_each_descendant_pre() and drop RCU read
2849 * lock before calling cgroup_addrm_files().
2851 mutex_lock(&cgroup_mutex);
2854 static int cgroup_cfts_commit(struct cftype *cfts, bool is_add)
2855 __releases(&cgroup_mutex)
2858 struct cgroup_subsys *ss = cfts[0].ss;
2859 struct cgroup *root = &ss->root->top_cgroup;
2860 struct super_block *sb = ss->root->sb;
2861 struct dentry *prev = NULL;
2862 struct inode *inode;
2863 struct cgroup_subsys_state *css;
2867 /* %NULL @cfts indicates abort and don't bother if @ss isn't attached */
2868 if (!cfts || ss->root == &cgroup_dummy_root ||
2869 !atomic_inc_not_zero(&sb->s_active)) {
2870 mutex_unlock(&cgroup_mutex);
2875 * All cgroups which are created after we drop cgroup_mutex will
2876 * have the updated set of files, so we only need to update the
2877 * cgroups created before the current @cgroup_serial_nr_next.
2879 update_before = cgroup_serial_nr_next;
2881 mutex_unlock(&cgroup_mutex);
2883 /* add/rm files for all cgroups created before */
2885 css_for_each_descendant_pre(css, cgroup_css(root, ss->subsys_id)) {
2886 struct cgroup *cgrp = css->cgroup;
2888 if (cgroup_is_dead(cgrp))
2891 inode = cgrp->dentry->d_inode;
2896 prev = cgrp->dentry;
2898 mutex_lock(&inode->i_mutex);
2899 mutex_lock(&cgroup_mutex);
2900 if (cgrp->serial_nr < update_before && !cgroup_is_dead(cgrp))
2901 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2902 mutex_unlock(&cgroup_mutex);
2903 mutex_unlock(&inode->i_mutex);
2911 deactivate_super(sb);
2916 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2917 * @ss: target cgroup subsystem
2918 * @cfts: zero-length name terminated array of cftypes
2920 * Register @cfts to @ss. Files described by @cfts are created for all
2921 * existing cgroups to which @ss is attached and all future cgroups will
2922 * have them too. This function can be called anytime whether @ss is
2925 * Returns 0 on successful registration, -errno on failure. Note that this
2926 * function currently returns 0 as long as @cfts registration is successful
2927 * even if some file creation attempts on existing cgroups fail.
2929 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2931 struct cftype_set *set;
2935 set = kzalloc(sizeof(*set), GFP_KERNEL);
2939 for (cft = cfts; cft->name[0] != '\0'; cft++)
2942 cgroup_cfts_prepare();
2944 list_add_tail(&set->node, &ss->cftsets);
2945 ret = cgroup_cfts_commit(cfts, true);
2947 cgroup_rm_cftypes(cfts);
2950 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2953 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2954 * @cfts: zero-length name terminated array of cftypes
2956 * Unregister @cfts. Files described by @cfts are removed from all
2957 * existing cgroups and all future cgroups won't have them either. This
2958 * function can be called anytime whether @cfts' subsys is attached or not.
2960 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2963 int cgroup_rm_cftypes(struct cftype *cfts)
2965 struct cftype_set *set;
2967 if (!cfts || !cfts[0].ss)
2970 cgroup_cfts_prepare();
2972 list_for_each_entry(set, &cfts[0].ss->cftsets, node) {
2973 if (set->cfts == cfts) {
2974 list_del(&set->node);
2976 cgroup_cfts_commit(cfts, false);
2981 cgroup_cfts_commit(NULL, false);
2986 * cgroup_task_count - count the number of tasks in a cgroup.
2987 * @cgrp: the cgroup in question
2989 * Return the number of tasks in the cgroup.
2991 int cgroup_task_count(const struct cgroup *cgrp)
2994 struct cgrp_cset_link *link;
2996 read_lock(&css_set_lock);
2997 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2998 count += atomic_read(&link->cset->refcount);
2999 read_unlock(&css_set_lock);
3004 * To reduce the fork() overhead for systems that are not actually using
3005 * their cgroups capability, we don't maintain the lists running through
3006 * each css_set to its tasks until we see the list actually used - in other
3007 * words after the first call to css_task_iter_start().
3009 static void cgroup_enable_task_cg_lists(void)
3011 struct task_struct *p, *g;
3012 write_lock(&css_set_lock);
3013 use_task_css_set_links = 1;
3015 * We need tasklist_lock because RCU is not safe against
3016 * while_each_thread(). Besides, a forking task that has passed
3017 * cgroup_post_fork() without seeing use_task_css_set_links = 1
3018 * is not guaranteed to have its child immediately visible in the
3019 * tasklist if we walk through it with RCU.
3021 read_lock(&tasklist_lock);
3022 do_each_thread(g, p) {
3025 * We should check if the process is exiting, otherwise
3026 * it will race with cgroup_exit() in that the list
3027 * entry won't be deleted though the process has exited.
3029 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
3030 list_add(&p->cg_list, &task_css_set(p)->tasks);
3032 } while_each_thread(g, p);
3033 read_unlock(&tasklist_lock);
3034 write_unlock(&css_set_lock);
3038 * css_next_child - find the next child of a given css
3039 * @pos_css: the current position (%NULL to initiate traversal)
3040 * @parent_css: css whose children to walk
3042 * This function returns the next child of @parent_css and should be called
3043 * under RCU read lock. The only requirement is that @parent_css and
3044 * @pos_css are accessible. The next sibling is guaranteed to be returned
3045 * regardless of their states.
3047 struct cgroup_subsys_state *
3048 css_next_child(struct cgroup_subsys_state *pos_css,
3049 struct cgroup_subsys_state *parent_css)
3051 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
3052 struct cgroup *cgrp = parent_css->cgroup;
3053 struct cgroup *next;
3055 WARN_ON_ONCE(!rcu_read_lock_held());
3058 * @pos could already have been removed. Once a cgroup is removed,
3059 * its ->sibling.next is no longer updated when its next sibling
3060 * changes. As CGRP_DEAD assertion is serialized and happens
3061 * before the cgroup is taken off the ->sibling list, if we see it
3062 * unasserted, it's guaranteed that the next sibling hasn't
3063 * finished its grace period even if it's already removed, and thus
3064 * safe to dereference from this RCU critical section. If
3065 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
3066 * to be visible as %true here.
3068 * If @pos is dead, its next pointer can't be dereferenced;
3069 * however, as each cgroup is given a monotonically increasing
3070 * unique serial number and always appended to the sibling list,
3071 * the next one can be found by walking the parent's children until
3072 * we see a cgroup with higher serial number than @pos's. While
3073 * this path can be slower, it's taken only when either the current
3074 * cgroup is removed or iteration and removal race.
3077 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
3078 } else if (likely(!cgroup_is_dead(pos))) {
3079 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
3081 list_for_each_entry_rcu(next, &cgrp->children, sibling)
3082 if (next->serial_nr > pos->serial_nr)
3086 if (&next->sibling == &cgrp->children)
3090 return cgroup_css(next, parent_css->ss->subsys_id);
3092 return &next->dummy_css;
3094 EXPORT_SYMBOL_GPL(css_next_child);
3097 * css_next_descendant_pre - find the next descendant for pre-order walk
3098 * @pos: the current position (%NULL to initiate traversal)
3099 * @root: css whose descendants to walk
3101 * To be used by css_for_each_descendant_pre(). Find the next descendant
3102 * to visit for pre-order traversal of @root's descendants. @root is
3103 * included in the iteration and the first node to be visited.
3105 * While this function requires RCU read locking, it doesn't require the
3106 * whole traversal to be contained in a single RCU critical section. This
3107 * function will return the correct next descendant as long as both @pos
3108 * and @root are accessible and @pos is a descendant of @root.
3110 struct cgroup_subsys_state *
3111 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3112 struct cgroup_subsys_state *root)
3114 struct cgroup_subsys_state *next;
3116 WARN_ON_ONCE(!rcu_read_lock_held());
3118 /* if first iteration, visit @root */
3122 /* visit the first child if exists */
3123 next = css_next_child(NULL, pos);
3127 /* no child, visit my or the closest ancestor's next sibling */
3128 while (pos != root) {
3129 next = css_next_child(pos, css_parent(pos));
3132 pos = css_parent(pos);
3137 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
3140 * css_rightmost_descendant - return the rightmost descendant of a css
3141 * @pos: css of interest
3143 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3144 * is returned. This can be used during pre-order traversal to skip
3147 * While this function requires RCU read locking, it doesn't require the
3148 * whole traversal to be contained in a single RCU critical section. This
3149 * function will return the correct rightmost descendant as long as @pos is
3152 struct cgroup_subsys_state *
3153 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3155 struct cgroup_subsys_state *last, *tmp;
3157 WARN_ON_ONCE(!rcu_read_lock_held());
3161 /* ->prev isn't RCU safe, walk ->next till the end */
3163 css_for_each_child(tmp, last)
3169 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
3171 static struct cgroup_subsys_state *
3172 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3174 struct cgroup_subsys_state *last;
3178 pos = css_next_child(NULL, pos);
3185 * css_next_descendant_post - find the next descendant for post-order walk
3186 * @pos: the current position (%NULL to initiate traversal)
3187 * @root: css whose descendants to walk
3189 * To be used by css_for_each_descendant_post(). Find the next descendant
3190 * to visit for post-order traversal of @root's descendants. @root is
3191 * included in the iteration and the last node to be visited.
3193 * While this function requires RCU read locking, it doesn't require the
3194 * whole traversal to be contained in a single RCU critical section. This
3195 * function will return the correct next descendant as long as both @pos
3196 * and @cgroup are accessible and @pos is a descendant of @cgroup.
3198 struct cgroup_subsys_state *
3199 css_next_descendant_post(struct cgroup_subsys_state *pos,
3200 struct cgroup_subsys_state *root)
3202 struct cgroup_subsys_state *next;
3204 WARN_ON_ONCE(!rcu_read_lock_held());
3206 /* if first iteration, visit the leftmost descendant */
3208 next = css_leftmost_descendant(root);
3209 return next != root ? next : NULL;
3212 /* if we visited @root, we're done */
3216 /* if there's an unvisited sibling, visit its leftmost descendant */
3217 next = css_next_child(pos, css_parent(pos));
3219 return css_leftmost_descendant(next);
3221 /* no sibling left, visit parent */
3222 return css_parent(pos);
3224 EXPORT_SYMBOL_GPL(css_next_descendant_post);
3227 * css_advance_task_iter - advance a task itererator to the next css_set
3228 * @it: the iterator to advance
3230 * Advance @it to the next css_set to walk.
3232 static void css_advance_task_iter(struct css_task_iter *it)
3234 struct list_head *l = it->cset_link;
3235 struct cgrp_cset_link *link;
3236 struct css_set *cset;
3238 /* Advance to the next non-empty css_set */
3241 if (l == &it->origin_css->cgroup->cset_links) {
3242 it->cset_link = NULL;
3245 link = list_entry(l, struct cgrp_cset_link, cset_link);
3247 } while (list_empty(&cset->tasks));
3249 it->task = cset->tasks.next;
3253 * css_task_iter_start - initiate task iteration
3254 * @css: the css to walk tasks of
3255 * @it: the task iterator to use
3257 * Initiate iteration through the tasks of @css. The caller can call
3258 * css_task_iter_next() to walk through the tasks until the function
3259 * returns NULL. On completion of iteration, css_task_iter_end() must be
3262 * Note that this function acquires a lock which is released when the
3263 * iteration finishes. The caller can't sleep while iteration is in
3266 void css_task_iter_start(struct cgroup_subsys_state *css,
3267 struct css_task_iter *it)
3268 __acquires(css_set_lock)
3271 * The first time anyone tries to iterate across a css, we need to
3272 * enable the list linking each css_set to its tasks, and fix up
3273 * all existing tasks.
3275 if (!use_task_css_set_links)
3276 cgroup_enable_task_cg_lists();
3278 read_lock(&css_set_lock);
3280 it->origin_css = css;
3281 it->cset_link = &css->cgroup->cset_links;
3283 css_advance_task_iter(it);
3287 * css_task_iter_next - return the next task for the iterator
3288 * @it: the task iterator being iterated
3290 * The "next" function for task iteration. @it should have been
3291 * initialized via css_task_iter_start(). Returns NULL when the iteration
3294 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3296 struct task_struct *res;
3297 struct list_head *l = it->task;
3298 struct cgrp_cset_link *link;
3300 /* If the iterator cg is NULL, we have no tasks */
3303 res = list_entry(l, struct task_struct, cg_list);
3304 /* Advance iterator to find next entry */
3306 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
3307 if (l == &link->cset->tasks) {
3309 * We reached the end of this task list - move on to the
3310 * next cgrp_cset_link.
3312 css_advance_task_iter(it);
3320 * css_task_iter_end - finish task iteration
3321 * @it: the task iterator to finish
3323 * Finish task iteration started by css_task_iter_start().
3325 void css_task_iter_end(struct css_task_iter *it)
3326 __releases(css_set_lock)
3328 read_unlock(&css_set_lock);
3331 static inline int started_after_time(struct task_struct *t1,
3332 struct timespec *time,
3333 struct task_struct *t2)
3335 int start_diff = timespec_compare(&t1->start_time, time);
3336 if (start_diff > 0) {
3338 } else if (start_diff < 0) {
3342 * Arbitrarily, if two processes started at the same
3343 * time, we'll say that the lower pointer value
3344 * started first. Note that t2 may have exited by now
3345 * so this may not be a valid pointer any longer, but
3346 * that's fine - it still serves to distinguish
3347 * between two tasks started (effectively) simultaneously.
3354 * This function is a callback from heap_insert() and is used to order
3356 * In this case we order the heap in descending task start time.
3358 static inline int started_after(void *p1, void *p2)
3360 struct task_struct *t1 = p1;
3361 struct task_struct *t2 = p2;
3362 return started_after_time(t1, &t2->start_time, t2);
3366 * css_scan_tasks - iterate though all the tasks in a css
3367 * @css: the css to iterate tasks of
3368 * @test: optional test callback
3369 * @process: process callback
3370 * @data: data passed to @test and @process
3371 * @heap: optional pre-allocated heap used for task iteration
3373 * Iterate through all the tasks in @css, calling @test for each, and if it
3374 * returns %true, call @process for it also.
3376 * @test may be NULL, meaning always true (select all tasks), which
3377 * effectively duplicates css_task_iter_{start,next,end}() but does not
3378 * lock css_set_lock for the call to @process.
3380 * It is guaranteed that @process will act on every task that is a member
3381 * of @css for the duration of this call. This function may or may not
3382 * call @process for tasks that exit or move to a different css during the
3383 * call, or are forked or move into the css during the call.
3385 * Note that @test may be called with locks held, and may in some
3386 * situations be called multiple times for the same task, so it should be
3389 * If @heap is non-NULL, a heap has been pre-allocated and will be used for
3390 * heap operations (and its "gt" member will be overwritten), else a
3391 * temporary heap will be used (allocation of which may cause this function
3394 int css_scan_tasks(struct cgroup_subsys_state *css,
3395 bool (*test)(struct task_struct *, void *),
3396 void (*process)(struct task_struct *, void *),
3397 void *data, struct ptr_heap *heap)
3400 struct css_task_iter it;
3401 struct task_struct *p, *dropped;
3402 /* Never dereference latest_task, since it's not refcounted */
3403 struct task_struct *latest_task = NULL;
3404 struct ptr_heap tmp_heap;
3405 struct timespec latest_time = { 0, 0 };
3408 /* The caller supplied our heap and pre-allocated its memory */
3409 heap->gt = &started_after;
3411 /* We need to allocate our own heap memory */
3413 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
3415 /* cannot allocate the heap */
3421 * Scan tasks in the css, using the @test callback to determine
3422 * which are of interest, and invoking @process callback on the
3423 * ones which need an update. Since we don't want to hold any
3424 * locks during the task updates, gather tasks to be processed in a
3425 * heap structure. The heap is sorted by descending task start
3426 * time. If the statically-sized heap fills up, we overflow tasks
3427 * that started later, and in future iterations only consider tasks
3428 * that started after the latest task in the previous pass. This
3429 * guarantees forward progress and that we don't miss any tasks.
3432 css_task_iter_start(css, &it);
3433 while ((p = css_task_iter_next(&it))) {
3435 * Only affect tasks that qualify per the caller's callback,
3436 * if he provided one
3438 if (test && !test(p, data))
3441 * Only process tasks that started after the last task
3444 if (!started_after_time(p, &latest_time, latest_task))
3446 dropped = heap_insert(heap, p);
3447 if (dropped == NULL) {
3449 * The new task was inserted; the heap wasn't
3453 } else if (dropped != p) {
3455 * The new task was inserted, and pushed out a
3459 put_task_struct(dropped);
3462 * Else the new task was newer than anything already in
3463 * the heap and wasn't inserted
3466 css_task_iter_end(&it);
3469 for (i = 0; i < heap->size; i++) {
3470 struct task_struct *q = heap->ptrs[i];
3472 latest_time = q->start_time;
3475 /* Process the task per the caller's callback */
3480 * If we had to process any tasks at all, scan again
3481 * in case some of them were in the middle of forking
3482 * children that didn't get processed.
3483 * Not the most efficient way to do it, but it avoids
3484 * having to take callback_mutex in the fork path
3488 if (heap == &tmp_heap)
3489 heap_free(&tmp_heap);
3493 static void cgroup_transfer_one_task(struct task_struct *task, void *data)
3495 struct cgroup *new_cgroup = data;
3497 mutex_lock(&cgroup_mutex);
3498 cgroup_attach_task(new_cgroup, task, false);
3499 mutex_unlock(&cgroup_mutex);
3503 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3504 * @to: cgroup to which the tasks will be moved
3505 * @from: cgroup in which the tasks currently reside
3507 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3509 return css_scan_tasks(&from->dummy_css, NULL, cgroup_transfer_one_task,
3514 * Stuff for reading the 'tasks'/'procs' files.
3516 * Reading this file can return large amounts of data if a cgroup has
3517 * *lots* of attached tasks. So it may need several calls to read(),
3518 * but we cannot guarantee that the information we produce is correct
3519 * unless we produce it entirely atomically.
3523 /* which pidlist file are we talking about? */
3524 enum cgroup_filetype {
3530 * A pidlist is a list of pids that virtually represents the contents of one
3531 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3532 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3535 struct cgroup_pidlist {
3537 * used to find which pidlist is wanted. doesn't change as long as
3538 * this particular list stays in the list.
3540 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3543 /* how many elements the above list has */
3545 /* how many files are using the current array */
3547 /* each of these stored in a list by its cgroup */
3548 struct list_head links;
3549 /* pointer to the cgroup we belong to, for list removal purposes */
3550 struct cgroup *owner;
3551 /* protects the other fields */
3552 struct rw_semaphore rwsem;
3556 * The following two functions "fix" the issue where there are more pids
3557 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3558 * TODO: replace with a kernel-wide solution to this problem
3560 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3561 static void *pidlist_allocate(int count)
3563 if (PIDLIST_TOO_LARGE(count))
3564 return vmalloc(count * sizeof(pid_t));
3566 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3568 static void pidlist_free(void *p)
3570 if (is_vmalloc_addr(p))
3577 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3578 * Returns the number of unique elements.
3580 static int pidlist_uniq(pid_t *list, int length)
3585 * we presume the 0th element is unique, so i starts at 1. trivial
3586 * edge cases first; no work needs to be done for either
3588 if (length == 0 || length == 1)
3590 /* src and dest walk down the list; dest counts unique elements */
3591 for (src = 1; src < length; src++) {
3592 /* find next unique element */
3593 while (list[src] == list[src-1]) {
3598 /* dest always points to where the next unique element goes */
3599 list[dest] = list[src];
3606 static int cmppid(const void *a, const void *b)
3608 return *(pid_t *)a - *(pid_t *)b;
3612 * find the appropriate pidlist for our purpose (given procs vs tasks)
3613 * returns with the lock on that pidlist already held, and takes care
3614 * of the use count, or returns NULL with no locks held if we're out of
3617 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3618 enum cgroup_filetype type)
3620 struct cgroup_pidlist *l;
3621 /* don't need task_nsproxy() if we're looking at ourself */
3622 struct pid_namespace *ns = task_active_pid_ns(current);
3625 * We can't drop the pidlist_mutex before taking the l->rwsem in case
3626 * the last ref-holder is trying to remove l from the list at the same
3627 * time. Holding the pidlist_mutex precludes somebody taking whichever
3628 * list we find out from under us - compare release_pid_array().
3630 mutex_lock(&cgrp->pidlist_mutex);
3631 list_for_each_entry(l, &cgrp->pidlists, links) {
3632 if (l->key.type == type && l->key.ns == ns) {
3633 /* make sure l doesn't vanish out from under us */
3634 down_write(&l->rwsem);
3635 mutex_unlock(&cgrp->pidlist_mutex);
3639 /* entry not found; create a new one */
3640 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3642 mutex_unlock(&cgrp->pidlist_mutex);
3645 init_rwsem(&l->rwsem);
3646 down_write(&l->rwsem);
3648 l->key.ns = get_pid_ns(ns);
3650 list_add(&l->links, &cgrp->pidlists);
3651 mutex_unlock(&cgrp->pidlist_mutex);
3656 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3658 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3659 struct cgroup_pidlist **lp)
3663 int pid, n = 0; /* used for populating the array */
3664 struct css_task_iter it;
3665 struct task_struct *tsk;
3666 struct cgroup_pidlist *l;
3669 * If cgroup gets more users after we read count, we won't have
3670 * enough space - tough. This race is indistinguishable to the
3671 * caller from the case that the additional cgroup users didn't
3672 * show up until sometime later on.
3674 length = cgroup_task_count(cgrp);
3675 array = pidlist_allocate(length);
3678 /* now, populate the array */
3679 css_task_iter_start(&cgrp->dummy_css, &it);
3680 while ((tsk = css_task_iter_next(&it))) {
3681 if (unlikely(n == length))
3683 /* get tgid or pid for procs or tasks file respectively */
3684 if (type == CGROUP_FILE_PROCS)
3685 pid = task_tgid_vnr(tsk);
3687 pid = task_pid_vnr(tsk);
3688 if (pid > 0) /* make sure to only use valid results */
3691 css_task_iter_end(&it);
3693 /* now sort & (if procs) strip out duplicates */
3694 sort(array, length, sizeof(pid_t), cmppid, NULL);
3695 if (type == CGROUP_FILE_PROCS)
3696 length = pidlist_uniq(array, length);
3697 l = cgroup_pidlist_find(cgrp, type);
3699 pidlist_free(array);
3702 /* store array, freeing old if necessary - lock already held */
3703 pidlist_free(l->list);
3707 up_write(&l->rwsem);
3713 * cgroupstats_build - build and fill cgroupstats
3714 * @stats: cgroupstats to fill information into
3715 * @dentry: A dentry entry belonging to the cgroup for which stats have
3718 * Build and fill cgroupstats so that taskstats can export it to user
3721 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3724 struct cgroup *cgrp;
3725 struct css_task_iter it;
3726 struct task_struct *tsk;
3729 * Validate dentry by checking the superblock operations,
3730 * and make sure it's a directory.
3732 if (dentry->d_sb->s_op != &cgroup_ops ||
3733 !S_ISDIR(dentry->d_inode->i_mode))
3737 cgrp = dentry->d_fsdata;
3739 css_task_iter_start(&cgrp->dummy_css, &it);
3740 while ((tsk = css_task_iter_next(&it))) {
3741 switch (tsk->state) {
3743 stats->nr_running++;
3745 case TASK_INTERRUPTIBLE:
3746 stats->nr_sleeping++;
3748 case TASK_UNINTERRUPTIBLE:
3749 stats->nr_uninterruptible++;
3752 stats->nr_stopped++;
3755 if (delayacct_is_task_waiting_on_io(tsk))
3756 stats->nr_io_wait++;
3760 css_task_iter_end(&it);
3768 * seq_file methods for the tasks/procs files. The seq_file position is the
3769 * next pid to display; the seq_file iterator is a pointer to the pid
3770 * in the cgroup->l->list array.
3773 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3776 * Initially we receive a position value that corresponds to
3777 * one more than the last pid shown (or 0 on the first call or
3778 * after a seek to the start). Use a binary-search to find the
3779 * next pid to display, if any
3781 struct cgroup_pidlist *l = s->private;
3782 int index = 0, pid = *pos;
3785 down_read(&l->rwsem);
3787 int end = l->length;
3789 while (index < end) {
3790 int mid = (index + end) / 2;
3791 if (l->list[mid] == pid) {
3794 } else if (l->list[mid] <= pid)
3800 /* If we're off the end of the array, we're done */
3801 if (index >= l->length)
3803 /* Update the abstract position to be the actual pid that we found */
3804 iter = l->list + index;
3809 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3811 struct cgroup_pidlist *l = s->private;
3815 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3817 struct cgroup_pidlist *l = s->private;
3819 pid_t *end = l->list + l->length;
3821 * Advance to the next pid in the array. If this goes off the
3833 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3835 return seq_printf(s, "%d\n", *(int *)v);
3839 * seq_operations functions for iterating on pidlists through seq_file -
3840 * independent of whether it's tasks or procs
3842 static const struct seq_operations cgroup_pidlist_seq_operations = {
3843 .start = cgroup_pidlist_start,
3844 .stop = cgroup_pidlist_stop,
3845 .next = cgroup_pidlist_next,
3846 .show = cgroup_pidlist_show,
3849 static void cgroup_release_pid_array(struct cgroup_pidlist *l)
3852 * the case where we're the last user of this particular pidlist will
3853 * have us remove it from the cgroup's list, which entails taking the
3854 * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
3855 * pidlist_mutex, we have to take pidlist_mutex first.
3857 mutex_lock(&l->owner->pidlist_mutex);
3858 down_write(&l->rwsem);
3859 BUG_ON(!l->use_count);
3860 if (!--l->use_count) {
3861 /* we're the last user if refcount is 0; remove and free */
3862 list_del(&l->links);
3863 mutex_unlock(&l->owner->pidlist_mutex);
3864 pidlist_free(l->list);
3865 put_pid_ns(l->key.ns);
3866 up_write(&l->rwsem);
3870 mutex_unlock(&l->owner->pidlist_mutex);
3871 up_write(&l->rwsem);
3874 static int cgroup_pidlist_release(struct inode *inode, struct file *file)
3876 struct cgroup_pidlist *l;
3877 if (!(file->f_mode & FMODE_READ))
3880 * the seq_file will only be initialized if the file was opened for
3881 * reading; hence we check if it's not null only in that case.
3883 l = ((struct seq_file *)file->private_data)->private;
3884 cgroup_release_pid_array(l);
3885 return seq_release(inode, file);
3888 static const struct file_operations cgroup_pidlist_operations = {
3890 .llseek = seq_lseek,
3891 .write = cgroup_file_write,
3892 .release = cgroup_pidlist_release,
3896 * The following functions handle opens on a file that displays a pidlist
3897 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
3900 /* helper function for the two below it */
3901 static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
3903 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
3904 struct cgroup_pidlist *l;
3907 /* Nothing to do for write-only files */
3908 if (!(file->f_mode & FMODE_READ))
3911 /* have the array populated */
3912 retval = pidlist_array_load(cgrp, type, &l);
3915 /* configure file information */
3916 file->f_op = &cgroup_pidlist_operations;
3918 retval = seq_open(file, &cgroup_pidlist_seq_operations);
3920 cgroup_release_pid_array(l);
3923 ((struct seq_file *)file->private_data)->private = l;
3926 static int cgroup_tasks_open(struct inode *unused, struct file *file)
3928 return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
3930 static int cgroup_procs_open(struct inode *unused, struct file *file)
3932 return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
3935 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3938 return notify_on_release(css->cgroup);
3941 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3942 struct cftype *cft, u64 val)
3944 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3946 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3948 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3953 * When dput() is called asynchronously, if umount has been done and
3954 * then deactivate_super() in cgroup_free_fn() kills the superblock,
3955 * there's a small window that vfs will see the root dentry with non-zero
3956 * refcnt and trigger BUG().
3958 * That's why we hold a reference before dput() and drop it right after.
3960 static void cgroup_dput(struct cgroup *cgrp)
3962 struct super_block *sb = cgrp->root->sb;
3964 atomic_inc(&sb->s_active);
3966 deactivate_super(sb);
3970 * Unregister event and free resources.
3972 * Gets called from workqueue.
3974 static void cgroup_event_remove(struct work_struct *work)
3976 struct cgroup_event *event = container_of(work, struct cgroup_event,
3978 struct cgroup_subsys_state *css = event->css;
3979 struct cgroup *cgrp = css->cgroup;
3981 remove_wait_queue(event->wqh, &event->wait);
3983 event->cft->unregister_event(css, event->cft, event->eventfd);
3985 /* Notify userspace the event is going away. */
3986 eventfd_signal(event->eventfd, 1);
3988 eventfd_ctx_put(event->eventfd);
3994 * Gets called on POLLHUP on eventfd when user closes it.
3996 * Called with wqh->lock held and interrupts disabled.
3998 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3999 int sync, void *key)
4001 struct cgroup_event *event = container_of(wait,
4002 struct cgroup_event, wait);
4003 struct cgroup *cgrp = event->css->cgroup;
4004 unsigned long flags = (unsigned long)key;
4006 if (flags & POLLHUP) {
4008 * If the event has been detached at cgroup removal, we
4009 * can simply return knowing the other side will cleanup
4012 * We can't race against event freeing since the other
4013 * side will require wqh->lock via remove_wait_queue(),
4016 spin_lock(&cgrp->event_list_lock);
4017 if (!list_empty(&event->list)) {
4018 list_del_init(&event->list);
4020 * We are in atomic context, but cgroup_event_remove()
4021 * may sleep, so we have to call it in workqueue.
4023 schedule_work(&event->remove);
4025 spin_unlock(&cgrp->event_list_lock);
4031 static void cgroup_event_ptable_queue_proc(struct file *file,
4032 wait_queue_head_t *wqh, poll_table *pt)
4034 struct cgroup_event *event = container_of(pt,
4035 struct cgroup_event, pt);
4038 add_wait_queue(wqh, &event->wait);
4042 * Parse input and register new cgroup event handler.
4044 * Input must be in format '<event_fd> <control_fd> <args>'.
4045 * Interpretation of args is defined by control file implementation.
4047 static int cgroup_write_event_control(struct cgroup_subsys_state *css,
4048 struct cftype *cft, const char *buffer)
4050 struct cgroup *cgrp = css->cgroup;
4051 struct cgroup_event *event;
4052 struct cgroup *cgrp_cfile;
4053 unsigned int efd, cfd;
4059 efd = simple_strtoul(buffer, &endp, 10);
4064 cfd = simple_strtoul(buffer, &endp, 10);
4065 if ((*endp != ' ') && (*endp != '\0'))
4069 event = kzalloc(sizeof(*event), GFP_KERNEL);
4073 INIT_LIST_HEAD(&event->list);
4074 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
4075 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
4076 INIT_WORK(&event->remove, cgroup_event_remove);
4078 efile = eventfd_fget(efd);
4079 if (IS_ERR(efile)) {
4080 ret = PTR_ERR(efile);
4084 event->eventfd = eventfd_ctx_fileget(efile);
4085 if (IS_ERR(event->eventfd)) {
4086 ret = PTR_ERR(event->eventfd);
4093 goto out_put_eventfd;
4096 /* the process need read permission on control file */
4097 /* AV: shouldn't we check that it's been opened for read instead? */
4098 ret = inode_permission(file_inode(cfile), MAY_READ);
4102 event->cft = __file_cft(cfile);
4103 if (IS_ERR(event->cft)) {
4104 ret = PTR_ERR(event->cft);
4109 * The file to be monitored must be in the same cgroup as
4110 * cgroup.event_control is.
4112 cgrp_cfile = __d_cgrp(cfile->f_dentry->d_parent);
4113 if (cgrp_cfile != cgrp) {
4118 if (!event->cft->register_event || !event->cft->unregister_event) {
4123 ret = event->cft->register_event(css, event->cft,
4124 event->eventfd, buffer);
4128 efile->f_op->poll(efile, &event->pt);
4131 * Events should be removed after rmdir of cgroup directory, but before
4132 * destroying subsystem state objects. Let's take reference to cgroup
4133 * directory dentry to do that.
4137 spin_lock(&cgrp->event_list_lock);
4138 list_add(&event->list, &cgrp->event_list);
4139 spin_unlock(&cgrp->event_list_lock);
4149 eventfd_ctx_put(event->eventfd);
4158 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4161 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4164 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4165 struct cftype *cft, u64 val)
4168 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4170 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4174 static struct cftype cgroup_base_files[] = {
4176 .name = "cgroup.procs",
4177 .open = cgroup_procs_open,
4178 .write_u64 = cgroup_procs_write,
4179 .release = cgroup_pidlist_release,
4180 .mode = S_IRUGO | S_IWUSR,
4183 .name = "cgroup.event_control",
4184 .write_string = cgroup_write_event_control,
4188 .name = "cgroup.clone_children",
4189 .flags = CFTYPE_INSANE,
4190 .read_u64 = cgroup_clone_children_read,
4191 .write_u64 = cgroup_clone_children_write,
4194 .name = "cgroup.sane_behavior",
4195 .flags = CFTYPE_ONLY_ON_ROOT,
4196 .read_seq_string = cgroup_sane_behavior_show,
4200 * Historical crazy stuff. These don't have "cgroup." prefix and
4201 * don't exist if sane_behavior. If you're depending on these, be
4202 * prepared to be burned.
4206 .flags = CFTYPE_INSANE, /* use "procs" instead */
4207 .open = cgroup_tasks_open,
4208 .write_u64 = cgroup_tasks_write,
4209 .release = cgroup_pidlist_release,
4210 .mode = S_IRUGO | S_IWUSR,
4213 .name = "notify_on_release",
4214 .flags = CFTYPE_INSANE,
4215 .read_u64 = cgroup_read_notify_on_release,
4216 .write_u64 = cgroup_write_notify_on_release,
4219 .name = "release_agent",
4220 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
4221 .read_seq_string = cgroup_release_agent_show,
4222 .write_string = cgroup_release_agent_write,
4223 .max_write_len = PATH_MAX,
4229 * cgroup_populate_dir - create subsys files in a cgroup directory
4230 * @cgrp: target cgroup
4231 * @subsys_mask: mask of the subsystem ids whose files should be added
4233 * On failure, no file is added.
4235 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
4237 struct cgroup_subsys *ss;
4240 /* process cftsets of each subsystem */
4241 for_each_subsys(ss, i) {
4242 struct cftype_set *set;
4244 if (!test_bit(i, &subsys_mask))
4247 list_for_each_entry(set, &ss->cftsets, node) {
4248 ret = cgroup_addrm_files(cgrp, set->cfts, true);
4254 /* This cgroup is ready now */
4255 for_each_root_subsys(cgrp->root, ss) {
4256 struct cgroup_subsys_state *css = cgroup_css(cgrp, ss->subsys_id);
4257 struct css_id *id = rcu_dereference_protected(css->id, true);
4260 * Update id->css pointer and make this css visible from
4261 * CSS ID functions. This pointer will be dereferened
4262 * from RCU-read-side without locks.
4265 rcu_assign_pointer(id->css, css);
4270 cgroup_clear_dir(cgrp, subsys_mask);
4275 * css destruction is four-stage process.
4277 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4278 * Implemented in kill_css().
4280 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4281 * and thus css_tryget() is guaranteed to fail, the css can be offlined
4282 * by invoking offline_css(). After offlining, the base ref is put.
4283 * Implemented in css_killed_work_fn().
4285 * 3. When the percpu_ref reaches zero, the only possible remaining
4286 * accessors are inside RCU read sections. css_release() schedules the
4289 * 4. After the grace period, the css can be freed. Implemented in
4290 * css_free_work_fn().
4292 * It is actually hairier because both step 2 and 4 require process context
4293 * and thus involve punting to css->destroy_work adding two additional
4294 * steps to the already complex sequence.
4296 static void css_free_work_fn(struct work_struct *work)
4298 struct cgroup_subsys_state *css =
4299 container_of(work, struct cgroup_subsys_state, destroy_work);
4300 struct cgroup *cgrp = css->cgroup;
4303 css_put(css->parent);
4305 css->ss->css_free(css);
4309 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4311 struct cgroup_subsys_state *css =
4312 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4315 * css holds an extra ref to @cgrp->dentry which is put on the last
4316 * css_put(). dput() requires process context which we don't have.
4318 INIT_WORK(&css->destroy_work, css_free_work_fn);
4319 schedule_work(&css->destroy_work);
4322 static void css_release(struct percpu_ref *ref)
4324 struct cgroup_subsys_state *css =
4325 container_of(ref, struct cgroup_subsys_state, refcnt);
4327 call_rcu(&css->rcu_head, css_free_rcu_fn);
4330 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
4331 struct cgroup *cgrp)
4339 css->parent = cgroup_css(cgrp->parent, ss->subsys_id);
4341 css->flags |= CSS_ROOT;
4343 BUG_ON(cgroup_css(cgrp, ss->subsys_id));
4346 /* invoke ->css_online() on a new CSS and mark it online if successful */
4347 static int online_css(struct cgroup_subsys_state *css)
4349 struct cgroup_subsys *ss = css->ss;
4352 lockdep_assert_held(&cgroup_mutex);
4355 ret = ss->css_online(css);
4357 css->flags |= CSS_ONLINE;
4358 css->cgroup->nr_css++;
4359 rcu_assign_pointer(css->cgroup->subsys[ss->subsys_id], css);
4364 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4365 static void offline_css(struct cgroup_subsys_state *css)
4367 struct cgroup_subsys *ss = css->ss;
4369 lockdep_assert_held(&cgroup_mutex);
4371 if (!(css->flags & CSS_ONLINE))
4374 if (ss->css_offline)
4375 ss->css_offline(css);
4377 css->flags &= ~CSS_ONLINE;
4378 css->cgroup->nr_css--;
4379 RCU_INIT_POINTER(css->cgroup->subsys[ss->subsys_id], css);
4383 * cgroup_create - create a cgroup
4384 * @parent: cgroup that will be parent of the new cgroup
4385 * @dentry: dentry of the new cgroup
4386 * @mode: mode to set on new inode
4388 * Must be called with the mutex on the parent inode held
4390 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
4393 struct cgroup_subsys_state *css_ar[CGROUP_SUBSYS_COUNT] = { };
4394 struct cgroup *cgrp;
4395 struct cgroup_name *name;
4396 struct cgroupfs_root *root = parent->root;
4398 struct cgroup_subsys *ss;
4399 struct super_block *sb = root->sb;
4401 /* allocate the cgroup and its ID, 0 is reserved for the root */
4402 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4406 name = cgroup_alloc_name(dentry);
4409 rcu_assign_pointer(cgrp->name, name);
4412 * Temporarily set the pointer to NULL, so idr_find() won't return
4413 * a half-baked cgroup.
4415 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
4420 * Only live parents can have children. Note that the liveliness
4421 * check isn't strictly necessary because cgroup_mkdir() and
4422 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
4423 * anyway so that locking is contained inside cgroup proper and we
4424 * don't get nasty surprises if we ever grow another caller.
4426 if (!cgroup_lock_live_group(parent)) {
4431 /* Grab a reference on the superblock so the hierarchy doesn't
4432 * get deleted on unmount if there are child cgroups. This
4433 * can be done outside cgroup_mutex, since the sb can't
4434 * disappear while someone has an open control file on the
4436 atomic_inc(&sb->s_active);
4438 init_cgroup_housekeeping(cgrp);
4440 dentry->d_fsdata = cgrp;
4441 cgrp->dentry = dentry;
4443 cgrp->parent = parent;
4444 cgrp->dummy_css.parent = &parent->dummy_css;
4445 cgrp->root = parent->root;
4447 if (notify_on_release(parent))
4448 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4450 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4451 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4453 for_each_root_subsys(root, ss) {
4454 struct cgroup_subsys_state *css;
4456 css = ss->css_alloc(cgroup_css(parent, ss->subsys_id));
4461 css_ar[ss->subsys_id] = css;
4463 err = percpu_ref_init(&css->refcnt, css_release);
4467 init_css(css, ss, cgrp);
4470 err = alloc_css_id(css);
4477 * Create directory. cgroup_create_file() returns with the new
4478 * directory locked on success so that it can be populated without
4479 * dropping cgroup_mutex.
4481 err = cgroup_create_file(dentry, S_IFDIR | mode, sb);
4484 lockdep_assert_held(&dentry->d_inode->i_mutex);
4486 cgrp->serial_nr = cgroup_serial_nr_next++;
4488 /* allocation complete, commit to creation */
4489 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
4490 root->number_of_cgroups++;
4492 /* each css holds a ref to the cgroup's dentry and the parent css */
4493 for_each_root_subsys(root, ss) {
4494 struct cgroup_subsys_state *css = css_ar[ss->subsys_id];
4497 percpu_ref_get(&css->parent->refcnt);
4500 /* hold a ref to the parent's dentry */
4501 dget(parent->dentry);
4503 /* creation succeeded, notify subsystems */
4504 for_each_root_subsys(root, ss) {
4505 struct cgroup_subsys_state *css = css_ar[ss->subsys_id];
4507 err = online_css(css);
4511 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4513 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",
4514 current->comm, current->pid, ss->name);
4515 if (!strcmp(ss->name, "memory"))
4516 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
4517 ss->warned_broken_hierarchy = true;
4521 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4523 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
4527 err = cgroup_populate_dir(cgrp, root->subsys_mask);
4531 mutex_unlock(&cgroup_mutex);
4532 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
4537 for_each_root_subsys(root, ss) {
4538 struct cgroup_subsys_state *css = css_ar[ss->subsys_id];
4541 percpu_ref_cancel_init(&css->refcnt);
4545 mutex_unlock(&cgroup_mutex);
4546 /* Release the reference count that we took on the superblock */
4547 deactivate_super(sb);
4549 idr_remove(&root->cgroup_idr, cgrp->id);
4551 kfree(rcu_dereference_raw(cgrp->name));
4557 cgroup_destroy_locked(cgrp);
4558 mutex_unlock(&cgroup_mutex);
4559 mutex_unlock(&dentry->d_inode->i_mutex);
4563 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
4565 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
4567 /* the vfs holds inode->i_mutex already */
4568 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
4572 * This is called when the refcnt of a css is confirmed to be killed.
4573 * css_tryget() is now guaranteed to fail.
4575 static void css_killed_work_fn(struct work_struct *work)
4577 struct cgroup_subsys_state *css =
4578 container_of(work, struct cgroup_subsys_state, destroy_work);
4579 struct cgroup *cgrp = css->cgroup;
4581 mutex_lock(&cgroup_mutex);
4584 * css_tryget() is guaranteed to fail now. Tell subsystems to
4585 * initate destruction.
4590 * If @cgrp is marked dead, it's waiting for refs of all css's to
4591 * be disabled before proceeding to the second phase of cgroup
4592 * destruction. If we are the last one, kick it off.
4594 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
4595 cgroup_destroy_css_killed(cgrp);
4597 mutex_unlock(&cgroup_mutex);
4600 * Put the css refs from kill_css(). Each css holds an extra
4601 * reference to the cgroup's dentry and cgroup removal proceeds
4602 * regardless of css refs. On the last put of each css, whenever
4603 * that may be, the extra dentry ref is put so that dentry
4604 * destruction happens only after all css's are released.
4609 /* css kill confirmation processing requires process context, bounce */
4610 static void css_killed_ref_fn(struct percpu_ref *ref)
4612 struct cgroup_subsys_state *css =
4613 container_of(ref, struct cgroup_subsys_state, refcnt);
4615 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4616 schedule_work(&css->destroy_work);
4620 * kill_css - destroy a css
4621 * @css: css to destroy
4623 * This function initiates destruction of @css by removing cgroup interface
4624 * files and putting its base reference. ->css_offline() will be invoked
4625 * asynchronously once css_tryget() is guaranteed to fail and when the
4626 * reference count reaches zero, @css will be released.
4628 static void kill_css(struct cgroup_subsys_state *css)
4630 cgroup_clear_dir(css->cgroup, 1 << css->ss->subsys_id);
4633 * Killing would put the base ref, but we need to keep it alive
4634 * until after ->css_offline().
4639 * cgroup core guarantees that, by the time ->css_offline() is
4640 * invoked, no new css reference will be given out via
4641 * css_tryget(). We can't simply call percpu_ref_kill() and
4642 * proceed to offlining css's because percpu_ref_kill() doesn't
4643 * guarantee that the ref is seen as killed on all CPUs on return.
4645 * Use percpu_ref_kill_and_confirm() to get notifications as each
4646 * css is confirmed to be seen as killed on all CPUs.
4648 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4652 * cgroup_destroy_locked - the first stage of cgroup destruction
4653 * @cgrp: cgroup to be destroyed
4655 * css's make use of percpu refcnts whose killing latency shouldn't be
4656 * exposed to userland and are RCU protected. Also, cgroup core needs to
4657 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
4658 * invoked. To satisfy all the requirements, destruction is implemented in
4659 * the following two steps.
4661 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4662 * userland visible parts and start killing the percpu refcnts of
4663 * css's. Set up so that the next stage will be kicked off once all
4664 * the percpu refcnts are confirmed to be killed.
4666 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4667 * rest of destruction. Once all cgroup references are gone, the
4668 * cgroup is RCU-freed.
4670 * This function implements s1. After this step, @cgrp is gone as far as
4671 * the userland is concerned and a new cgroup with the same name may be
4672 * created. As cgroup doesn't care about the names internally, this
4673 * doesn't cause any problem.
4675 static int cgroup_destroy_locked(struct cgroup *cgrp)
4676 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4678 struct dentry *d = cgrp->dentry;
4679 struct cgroup_event *event, *tmp;
4680 struct cgroup_subsys *ss;
4683 lockdep_assert_held(&d->d_inode->i_mutex);
4684 lockdep_assert_held(&cgroup_mutex);
4687 * css_set_lock synchronizes access to ->cset_links and prevents
4688 * @cgrp from being removed while __put_css_set() is in progress.
4690 read_lock(&css_set_lock);
4691 empty = list_empty(&cgrp->cset_links) && list_empty(&cgrp->children);
4692 read_unlock(&css_set_lock);
4697 * Initiate massacre of all css's. cgroup_destroy_css_killed()
4698 * will be invoked to perform the rest of destruction once the
4699 * percpu refs of all css's are confirmed to be killed.
4701 for_each_root_subsys(cgrp->root, ss)
4702 kill_css(cgroup_css(cgrp, ss->subsys_id));
4705 * Mark @cgrp dead. This prevents further task migration and child
4706 * creation by disabling cgroup_lock_live_group(). Note that
4707 * CGRP_DEAD assertion is depended upon by css_next_child() to
4708 * resume iteration after dropping RCU read lock. See
4709 * css_next_child() for details.
4711 set_bit(CGRP_DEAD, &cgrp->flags);
4713 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4714 raw_spin_lock(&release_list_lock);
4715 if (!list_empty(&cgrp->release_list))
4716 list_del_init(&cgrp->release_list);
4717 raw_spin_unlock(&release_list_lock);
4720 * If @cgrp has css's attached, the second stage of cgroup
4721 * destruction is kicked off from css_killed_work_fn() after the
4722 * refs of all attached css's are killed. If @cgrp doesn't have
4723 * any css, we kick it off here.
4726 cgroup_destroy_css_killed(cgrp);
4729 * Clear the base files and remove @cgrp directory. The removal
4730 * puts the base ref but we aren't quite done with @cgrp yet, so
4733 cgroup_addrm_files(cgrp, cgroup_base_files, false);
4735 cgroup_d_remove_dir(d);
4738 * Unregister events and notify userspace.
4739 * Notify userspace about cgroup removing only after rmdir of cgroup
4740 * directory to avoid race between userspace and kernelspace.
4742 spin_lock(&cgrp->event_list_lock);
4743 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
4744 list_del_init(&event->list);
4745 schedule_work(&event->remove);
4747 spin_unlock(&cgrp->event_list_lock);
4753 * cgroup_destroy_css_killed - the second step of cgroup destruction
4754 * @work: cgroup->destroy_free_work
4756 * This function is invoked from a work item for a cgroup which is being
4757 * destroyed after all css's are offlined and performs the rest of
4758 * destruction. This is the second step of destruction described in the
4759 * comment above cgroup_destroy_locked().
4761 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
4763 struct cgroup *parent = cgrp->parent;
4764 struct dentry *d = cgrp->dentry;
4766 lockdep_assert_held(&cgroup_mutex);
4768 /* delete this cgroup from parent->children */
4769 list_del_rcu(&cgrp->sibling);
4772 * We should remove the cgroup object from idr before its grace
4773 * period starts, so we won't be looking up a cgroup while the
4774 * cgroup is being freed.
4776 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4781 set_bit(CGRP_RELEASABLE, &parent->flags);
4782 check_for_release(parent);
4785 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
4789 mutex_lock(&cgroup_mutex);
4790 ret = cgroup_destroy_locked(dentry->d_fsdata);
4791 mutex_unlock(&cgroup_mutex);
4796 static void __init_or_module cgroup_init_cftsets(struct cgroup_subsys *ss)
4798 INIT_LIST_HEAD(&ss->cftsets);
4801 * base_cftset is embedded in subsys itself, no need to worry about
4804 if (ss->base_cftypes) {
4807 for (cft = ss->base_cftypes; cft->name[0] != '\0'; cft++)
4810 ss->base_cftset.cfts = ss->base_cftypes;
4811 list_add_tail(&ss->base_cftset.node, &ss->cftsets);
4815 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4817 struct cgroup_subsys_state *css;
4819 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4821 mutex_lock(&cgroup_mutex);
4823 /* init base cftset */
4824 cgroup_init_cftsets(ss);
4826 /* Create the top cgroup state for this subsystem */
4827 list_add(&ss->sibling, &cgroup_dummy_root.subsys_list);
4828 ss->root = &cgroup_dummy_root;
4829 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss->subsys_id));
4830 /* We don't handle early failures gracefully */
4831 BUG_ON(IS_ERR(css));
4832 init_css(css, ss, cgroup_dummy_top);
4834 /* Update the init_css_set to contain a subsys
4835 * pointer to this state - since the subsystem is
4836 * newly registered, all tasks and hence the
4837 * init_css_set is in the subsystem's top cgroup. */
4838 init_css_set.subsys[ss->subsys_id] = css;
4840 need_forkexit_callback |= ss->fork || ss->exit;
4842 /* At system boot, before all subsystems have been
4843 * registered, no tasks have been forked, so we don't
4844 * need to invoke fork callbacks here. */
4845 BUG_ON(!list_empty(&init_task.tasks));
4847 BUG_ON(online_css(css));
4849 mutex_unlock(&cgroup_mutex);
4851 /* this function shouldn't be used with modular subsystems, since they
4852 * need to register a subsys_id, among other things */
4857 * cgroup_load_subsys: load and register a modular subsystem at runtime
4858 * @ss: the subsystem to load
4860 * This function should be called in a modular subsystem's initcall. If the
4861 * subsystem is built as a module, it will be assigned a new subsys_id and set
4862 * up for use. If the subsystem is built-in anyway, work is delegated to the
4863 * simpler cgroup_init_subsys.
4865 int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
4867 struct cgroup_subsys_state *css;
4869 struct hlist_node *tmp;
4870 struct css_set *cset;
4873 /* check name and function validity */
4874 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
4875 ss->css_alloc == NULL || ss->css_free == NULL)
4879 * we don't support callbacks in modular subsystems. this check is
4880 * before the ss->module check for consistency; a subsystem that could
4881 * be a module should still have no callbacks even if the user isn't
4882 * compiling it as one.
4884 if (ss->fork || ss->exit)
4888 * an optionally modular subsystem is built-in: we want to do nothing,
4889 * since cgroup_init_subsys will have already taken care of it.
4891 if (ss->module == NULL) {
4892 /* a sanity check */
4893 BUG_ON(cgroup_subsys[ss->subsys_id] != ss);
4897 /* init base cftset */
4898 cgroup_init_cftsets(ss);
4900 mutex_lock(&cgroup_mutex);
4901 cgroup_subsys[ss->subsys_id] = ss;
4904 * no ss->css_alloc seems to need anything important in the ss
4905 * struct, so this can happen first (i.e. before the dummy root
4908 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss->subsys_id));
4910 /* failure case - need to deassign the cgroup_subsys[] slot. */
4911 cgroup_subsys[ss->subsys_id] = NULL;
4912 mutex_unlock(&cgroup_mutex);
4913 return PTR_ERR(css);
4916 list_add(&ss->sibling, &cgroup_dummy_root.subsys_list);
4917 ss->root = &cgroup_dummy_root;
4919 /* our new subsystem will be attached to the dummy hierarchy. */
4920 init_css(css, ss, cgroup_dummy_top);
4921 /* init_idr must be after init_css() because it sets css->id. */
4923 ret = cgroup_init_idr(ss, css);
4929 * Now we need to entangle the css into the existing css_sets. unlike
4930 * in cgroup_init_subsys, there are now multiple css_sets, so each one
4931 * will need a new pointer to it; done by iterating the css_set_table.
4932 * furthermore, modifying the existing css_sets will corrupt the hash
4933 * table state, so each changed css_set will need its hash recomputed.
4934 * this is all done under the css_set_lock.
4936 write_lock(&css_set_lock);
4937 hash_for_each_safe(css_set_table, i, tmp, cset, hlist) {
4938 /* skip entries that we already rehashed */
4939 if (cset->subsys[ss->subsys_id])
4941 /* remove existing entry */
4942 hash_del(&cset->hlist);
4944 cset->subsys[ss->subsys_id] = css;
4945 /* recompute hash and restore entry */
4946 key = css_set_hash(cset->subsys);
4947 hash_add(css_set_table, &cset->hlist, key);
4949 write_unlock(&css_set_lock);
4951 ret = online_css(css);
4956 mutex_unlock(&cgroup_mutex);
4960 mutex_unlock(&cgroup_mutex);
4961 /* @ss can't be mounted here as try_module_get() would fail */
4962 cgroup_unload_subsys(ss);
4965 EXPORT_SYMBOL_GPL(cgroup_load_subsys);
4968 * cgroup_unload_subsys: unload a modular subsystem
4969 * @ss: the subsystem to unload
4971 * This function should be called in a modular subsystem's exitcall. When this
4972 * function is invoked, the refcount on the subsystem's module will be 0, so
4973 * the subsystem will not be attached to any hierarchy.
4975 void cgroup_unload_subsys(struct cgroup_subsys *ss)
4977 struct cgrp_cset_link *link;
4979 BUG_ON(ss->module == NULL);
4982 * we shouldn't be called if the subsystem is in use, and the use of
4983 * try_module_get() in rebind_subsystems() should ensure that it
4984 * doesn't start being used while we're killing it off.
4986 BUG_ON(ss->root != &cgroup_dummy_root);
4988 mutex_lock(&cgroup_mutex);
4990 offline_css(cgroup_css(cgroup_dummy_top, ss->subsys_id));
4993 idr_destroy(&ss->idr);
4995 /* deassign the subsys_id */
4996 cgroup_subsys[ss->subsys_id] = NULL;
4998 /* remove subsystem from the dummy root's list of subsystems */
4999 list_del_init(&ss->sibling);
5002 * disentangle the css from all css_sets attached to the dummy
5003 * top. as in loading, we need to pay our respects to the hashtable
5006 write_lock(&css_set_lock);
5007 list_for_each_entry(link, &cgroup_dummy_top->cset_links, cset_link) {
5008 struct css_set *cset = link->cset;
5011 hash_del(&cset->hlist);
5012 cset->subsys[ss->subsys_id] = NULL;
5013 key = css_set_hash(cset->subsys);
5014 hash_add(css_set_table, &cset->hlist, key);
5016 write_unlock(&css_set_lock);
5019 * remove subsystem's css from the cgroup_dummy_top and free it -
5020 * need to free before marking as null because ss->css_free needs
5021 * the cgrp->subsys pointer to find their state. note that this
5022 * also takes care of freeing the css_id.
5024 ss->css_free(cgroup_css(cgroup_dummy_top, ss->subsys_id));
5025 RCU_INIT_POINTER(cgroup_dummy_top->subsys[ss->subsys_id], NULL);
5027 mutex_unlock(&cgroup_mutex);
5029 EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
5032 * cgroup_init_early - cgroup initialization at system boot
5034 * Initialize cgroups at system boot, and initialize any
5035 * subsystems that request early init.
5037 int __init cgroup_init_early(void)
5039 struct cgroup_subsys *ss;
5042 atomic_set(&init_css_set.refcount, 1);
5043 INIT_LIST_HEAD(&init_css_set.cgrp_links);
5044 INIT_LIST_HEAD(&init_css_set.tasks);
5045 INIT_HLIST_NODE(&init_css_set.hlist);
5047 init_cgroup_root(&cgroup_dummy_root);
5048 cgroup_root_count = 1;
5049 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5051 init_cgrp_cset_link.cset = &init_css_set;
5052 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
5053 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
5054 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
5056 /* at bootup time, we don't worry about modular subsystems */
5057 for_each_builtin_subsys(ss, i) {
5059 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
5060 BUG_ON(!ss->css_alloc);
5061 BUG_ON(!ss->css_free);
5062 if (ss->subsys_id != i) {
5063 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
5064 ss->name, ss->subsys_id);
5069 cgroup_init_subsys(ss);
5075 * cgroup_init - cgroup initialization
5077 * Register cgroup filesystem and /proc file, and initialize
5078 * any subsystems that didn't request early init.
5080 int __init cgroup_init(void)
5082 struct cgroup_subsys *ss;
5086 err = bdi_init(&cgroup_backing_dev_info);
5090 for_each_builtin_subsys(ss, i) {
5091 if (!ss->early_init)
5092 cgroup_init_subsys(ss);
5094 cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
5097 /* allocate id for the dummy hierarchy */
5098 mutex_lock(&cgroup_mutex);
5099 mutex_lock(&cgroup_root_mutex);
5101 /* Add init_css_set to the hash table */
5102 key = css_set_hash(init_css_set.subsys);
5103 hash_add(css_set_table, &init_css_set.hlist, key);
5105 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
5107 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
5111 mutex_unlock(&cgroup_root_mutex);
5112 mutex_unlock(&cgroup_mutex);
5114 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
5120 err = register_filesystem(&cgroup_fs_type);
5122 kobject_put(cgroup_kobj);
5126 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
5130 bdi_destroy(&cgroup_backing_dev_info);
5136 * proc_cgroup_show()
5137 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5138 * - Used for /proc/<pid>/cgroup.
5139 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
5140 * doesn't really matter if tsk->cgroup changes after we read it,
5141 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
5142 * anyway. No need to check that tsk->cgroup != NULL, thanks to
5143 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
5144 * cgroup to top_cgroup.
5147 /* TODO: Use a proper seq_file iterator */
5148 int proc_cgroup_show(struct seq_file *m, void *v)
5151 struct task_struct *tsk;
5154 struct cgroupfs_root *root;
5157 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
5163 tsk = get_pid_task(pid, PIDTYPE_PID);
5169 mutex_lock(&cgroup_mutex);
5171 for_each_active_root(root) {
5172 struct cgroup_subsys *ss;
5173 struct cgroup *cgrp;
5176 seq_printf(m, "%d:", root->hierarchy_id);
5177 for_each_root_subsys(root, ss)
5178 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
5179 if (strlen(root->name))
5180 seq_printf(m, "%sname=%s", count ? "," : "",
5183 cgrp = task_cgroup_from_root(tsk, root);
5184 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
5192 mutex_unlock(&cgroup_mutex);
5193 put_task_struct(tsk);
5200 /* Display information about each subsystem and each hierarchy */
5201 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5203 struct cgroup_subsys *ss;
5206 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5208 * ideally we don't want subsystems moving around while we do this.
5209 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5210 * subsys/hierarchy state.
5212 mutex_lock(&cgroup_mutex);
5214 for_each_subsys(ss, i)
5215 seq_printf(m, "%s\t%d\t%d\t%d\n",
5216 ss->name, ss->root->hierarchy_id,
5217 ss->root->number_of_cgroups, !ss->disabled);
5219 mutex_unlock(&cgroup_mutex);
5223 static int cgroupstats_open(struct inode *inode, struct file *file)
5225 return single_open(file, proc_cgroupstats_show, NULL);
5228 static const struct file_operations proc_cgroupstats_operations = {
5229 .open = cgroupstats_open,
5231 .llseek = seq_lseek,
5232 .release = single_release,
5236 * cgroup_fork - attach newly forked task to its parents cgroup.
5237 * @child: pointer to task_struct of forking parent process.
5239 * Description: A task inherits its parent's cgroup at fork().
5241 * A pointer to the shared css_set was automatically copied in
5242 * fork.c by dup_task_struct(). However, we ignore that copy, since
5243 * it was not made under the protection of RCU or cgroup_mutex, so
5244 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
5245 * have already changed current->cgroups, allowing the previously
5246 * referenced cgroup group to be removed and freed.
5248 * At the point that cgroup_fork() is called, 'current' is the parent
5249 * task, and the passed argument 'child' points to the child task.
5251 void cgroup_fork(struct task_struct *child)
5254 get_css_set(task_css_set(current));
5255 child->cgroups = current->cgroups;
5256 task_unlock(current);
5257 INIT_LIST_HEAD(&child->cg_list);
5261 * cgroup_post_fork - called on a new task after adding it to the task list
5262 * @child: the task in question
5264 * Adds the task to the list running through its css_set if necessary and
5265 * call the subsystem fork() callbacks. Has to be after the task is
5266 * visible on the task list in case we race with the first call to
5267 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5270 void cgroup_post_fork(struct task_struct *child)
5272 struct cgroup_subsys *ss;
5276 * use_task_css_set_links is set to 1 before we walk the tasklist
5277 * under the tasklist_lock and we read it here after we added the child
5278 * to the tasklist under the tasklist_lock as well. If the child wasn't
5279 * yet in the tasklist when we walked through it from
5280 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
5281 * should be visible now due to the paired locking and barriers implied
5282 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
5283 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
5286 if (use_task_css_set_links) {
5287 write_lock(&css_set_lock);
5289 if (list_empty(&child->cg_list))
5290 list_add(&child->cg_list, &task_css_set(child)->tasks);
5292 write_unlock(&css_set_lock);
5296 * Call ss->fork(). This must happen after @child is linked on
5297 * css_set; otherwise, @child might change state between ->fork()
5298 * and addition to css_set.
5300 if (need_forkexit_callback) {
5302 * fork/exit callbacks are supported only for builtin
5303 * subsystems, and the builtin section of the subsys
5304 * array is immutable, so we don't need to lock the
5305 * subsys array here. On the other hand, modular section
5306 * of the array can be freed at module unload, so we
5309 for_each_builtin_subsys(ss, i)
5316 * cgroup_exit - detach cgroup from exiting task
5317 * @tsk: pointer to task_struct of exiting process
5318 * @run_callback: run exit callbacks?
5320 * Description: Detach cgroup from @tsk and release it.
5322 * Note that cgroups marked notify_on_release force every task in
5323 * them to take the global cgroup_mutex mutex when exiting.
5324 * This could impact scaling on very large systems. Be reluctant to
5325 * use notify_on_release cgroups where very high task exit scaling
5326 * is required on large systems.
5328 * the_top_cgroup_hack:
5330 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
5332 * We call cgroup_exit() while the task is still competent to
5333 * handle notify_on_release(), then leave the task attached to the
5334 * root cgroup in each hierarchy for the remainder of its exit.
5336 * To do this properly, we would increment the reference count on
5337 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
5338 * code we would add a second cgroup function call, to drop that
5339 * reference. This would just create an unnecessary hot spot on
5340 * the top_cgroup reference count, to no avail.
5342 * Normally, holding a reference to a cgroup without bumping its
5343 * count is unsafe. The cgroup could go away, or someone could
5344 * attach us to a different cgroup, decrementing the count on
5345 * the first cgroup that we never incremented. But in this case,
5346 * top_cgroup isn't going away, and either task has PF_EXITING set,
5347 * which wards off any cgroup_attach_task() attempts, or task is a failed
5348 * fork, never visible to cgroup_attach_task.
5350 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
5352 struct cgroup_subsys *ss;
5353 struct css_set *cset;
5357 * Unlink from the css_set task list if necessary.
5358 * Optimistically check cg_list before taking
5361 if (!list_empty(&tsk->cg_list)) {
5362 write_lock(&css_set_lock);
5363 if (!list_empty(&tsk->cg_list))
5364 list_del_init(&tsk->cg_list);
5365 write_unlock(&css_set_lock);
5368 /* Reassign the task to the init_css_set. */
5370 cset = task_css_set(tsk);
5371 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5373 if (run_callbacks && need_forkexit_callback) {
5375 * fork/exit callbacks are supported only for builtin
5376 * subsystems, see cgroup_post_fork() for details.
5378 for_each_builtin_subsys(ss, i) {
5380 struct cgroup_subsys_state *old_css = cset->subsys[i];
5381 struct cgroup_subsys_state *css = task_css(tsk, i);
5383 ss->exit(css, old_css, tsk);
5389 put_css_set_taskexit(cset);
5392 static void check_for_release(struct cgroup *cgrp)
5394 if (cgroup_is_releasable(cgrp) &&
5395 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
5397 * Control Group is currently removeable. If it's not
5398 * already queued for a userspace notification, queue
5401 int need_schedule_work = 0;
5403 raw_spin_lock(&release_list_lock);
5404 if (!cgroup_is_dead(cgrp) &&
5405 list_empty(&cgrp->release_list)) {
5406 list_add(&cgrp->release_list, &release_list);
5407 need_schedule_work = 1;
5409 raw_spin_unlock(&release_list_lock);
5410 if (need_schedule_work)
5411 schedule_work(&release_agent_work);
5416 * Notify userspace when a cgroup is released, by running the
5417 * configured release agent with the name of the cgroup (path
5418 * relative to the root of cgroup file system) as the argument.
5420 * Most likely, this user command will try to rmdir this cgroup.
5422 * This races with the possibility that some other task will be
5423 * attached to this cgroup before it is removed, or that some other
5424 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5425 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5426 * unused, and this cgroup will be reprieved from its death sentence,
5427 * to continue to serve a useful existence. Next time it's released,
5428 * we will get notified again, if it still has 'notify_on_release' set.
5430 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5431 * means only wait until the task is successfully execve()'d. The
5432 * separate release agent task is forked by call_usermodehelper(),
5433 * then control in this thread returns here, without waiting for the
5434 * release agent task. We don't bother to wait because the caller of
5435 * this routine has no use for the exit status of the release agent
5436 * task, so no sense holding our caller up for that.
5438 static void cgroup_release_agent(struct work_struct *work)
5440 BUG_ON(work != &release_agent_work);
5441 mutex_lock(&cgroup_mutex);
5442 raw_spin_lock(&release_list_lock);
5443 while (!list_empty(&release_list)) {
5444 char *argv[3], *envp[3];
5446 char *pathbuf = NULL, *agentbuf = NULL;
5447 struct cgroup *cgrp = list_entry(release_list.next,
5450 list_del_init(&cgrp->release_list);
5451 raw_spin_unlock(&release_list_lock);
5452 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
5455 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
5457 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5462 argv[i++] = agentbuf;
5463 argv[i++] = pathbuf;
5467 /* minimal command environment */
5468 envp[i++] = "HOME=/";
5469 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5472 /* Drop the lock while we invoke the usermode helper,
5473 * since the exec could involve hitting disk and hence
5474 * be a slow process */
5475 mutex_unlock(&cgroup_mutex);
5476 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5477 mutex_lock(&cgroup_mutex);
5481 raw_spin_lock(&release_list_lock);
5483 raw_spin_unlock(&release_list_lock);
5484 mutex_unlock(&cgroup_mutex);
5487 static int __init cgroup_disable(char *str)
5489 struct cgroup_subsys *ss;
5493 while ((token = strsep(&str, ",")) != NULL) {
5498 * cgroup_disable, being at boot time, can't know about
5499 * module subsystems, so we don't worry about them.
5501 for_each_builtin_subsys(ss, i) {
5502 if (!strcmp(token, ss->name)) {
5504 printk(KERN_INFO "Disabling %s control group"
5505 " subsystem\n", ss->name);
5512 __setup("cgroup_disable=", cgroup_disable);
5515 * Functons for CSS ID.
5518 /* to get ID other than 0, this should be called when !cgroup_is_dead() */
5519 unsigned short css_id(struct cgroup_subsys_state *css)
5521 struct css_id *cssid;
5524 * This css_id() can return correct value when somone has refcnt
5525 * on this or this is under rcu_read_lock(). Once css->id is allocated,
5526 * it's unchanged until freed.
5528 cssid = rcu_dereference_raw(css->id);
5534 EXPORT_SYMBOL_GPL(css_id);
5537 * css_is_ancestor - test "root" css is an ancestor of "child"
5538 * @child: the css to be tested.
5539 * @root: the css supporsed to be an ancestor of the child.
5541 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
5542 * this function reads css->id, the caller must hold rcu_read_lock().
5543 * But, considering usual usage, the csses should be valid objects after test.
5544 * Assuming that the caller will do some action to the child if this returns
5545 * returns true, the caller must take "child";s reference count.
5546 * If "child" is valid object and this returns true, "root" is valid, too.
5549 bool css_is_ancestor(struct cgroup_subsys_state *child,
5550 const struct cgroup_subsys_state *root)
5552 struct css_id *child_id;
5553 struct css_id *root_id;
5555 child_id = rcu_dereference(child->id);
5558 root_id = rcu_dereference(root->id);
5561 if (child_id->depth < root_id->depth)
5563 if (child_id->stack[root_id->depth] != root_id->id)
5568 void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
5570 struct css_id *id = rcu_dereference_protected(css->id, true);
5572 /* When this is called before css_id initialization, id can be NULL */
5576 BUG_ON(!ss->use_id);
5578 rcu_assign_pointer(id->css, NULL);
5579 rcu_assign_pointer(css->id, NULL);
5580 spin_lock(&ss->id_lock);
5581 idr_remove(&ss->idr, id->id);
5582 spin_unlock(&ss->id_lock);
5583 kfree_rcu(id, rcu_head);
5585 EXPORT_SYMBOL_GPL(free_css_id);
5588 * This is called by init or create(). Then, calls to this function are
5589 * always serialized (By cgroup_mutex() at create()).
5592 static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
5594 struct css_id *newid;
5597 BUG_ON(!ss->use_id);
5599 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
5600 newid = kzalloc(size, GFP_KERNEL);
5602 return ERR_PTR(-ENOMEM);
5604 idr_preload(GFP_KERNEL);
5605 spin_lock(&ss->id_lock);
5606 /* Don't use 0. allocates an ID of 1-65535 */
5607 ret = idr_alloc(&ss->idr, newid, 1, CSS_ID_MAX + 1, GFP_NOWAIT);
5608 spin_unlock(&ss->id_lock);
5611 /* Returns error when there are no free spaces for new ID.*/
5616 newid->depth = depth;
5620 return ERR_PTR(ret);
5624 static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
5625 struct cgroup_subsys_state *rootcss)
5627 struct css_id *newid;
5629 spin_lock_init(&ss->id_lock);
5632 newid = get_new_cssid(ss, 0);
5634 return PTR_ERR(newid);
5636 newid->stack[0] = newid->id;
5637 RCU_INIT_POINTER(newid->css, rootcss);
5638 RCU_INIT_POINTER(rootcss->id, newid);
5642 static int alloc_css_id(struct cgroup_subsys_state *child_css)
5644 struct cgroup_subsys_state *parent_css = css_parent(child_css);
5645 struct css_id *child_id, *parent_id;
5648 parent_id = rcu_dereference_protected(parent_css->id, true);
5649 depth = parent_id->depth + 1;
5651 child_id = get_new_cssid(child_css->ss, depth);
5652 if (IS_ERR(child_id))
5653 return PTR_ERR(child_id);
5655 for (i = 0; i < depth; i++)
5656 child_id->stack[i] = parent_id->stack[i];
5657 child_id->stack[depth] = child_id->id;
5659 * child_id->css pointer will be set after this cgroup is available
5660 * see cgroup_populate_dir()
5662 rcu_assign_pointer(child_css->id, child_id);
5668 * css_lookup - lookup css by id
5669 * @ss: cgroup subsys to be looked into.
5672 * Returns pointer to cgroup_subsys_state if there is valid one with id.
5673 * NULL if not. Should be called under rcu_read_lock()
5675 struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
5677 struct css_id *cssid = NULL;
5679 BUG_ON(!ss->use_id);
5680 cssid = idr_find(&ss->idr, id);
5682 if (unlikely(!cssid))
5685 return rcu_dereference(cssid->css);
5687 EXPORT_SYMBOL_GPL(css_lookup);
5690 * cgroup_css_from_dir - get corresponding css from file open on cgroup dir
5691 * @f: directory file of interest
5692 * @id: subsystem id of interest
5694 * Must be called under RCU read lock. The caller is responsible for
5695 * pinning the returned css if it needs to be accessed outside the RCU
5698 struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id)
5700 struct cgroup *cgrp;
5701 struct inode *inode;
5702 struct cgroup_subsys_state *css;
5704 WARN_ON_ONCE(!rcu_read_lock_held());
5706 inode = file_inode(f);
5707 /* check in cgroup filesystem dir */
5708 if (inode->i_op != &cgroup_dir_inode_operations)
5709 return ERR_PTR(-EBADF);
5711 if (id < 0 || id >= CGROUP_SUBSYS_COUNT)
5712 return ERR_PTR(-EINVAL);
5715 cgrp = __d_cgrp(f->f_dentry);
5716 css = cgroup_css(cgrp, id);
5717 return css ? css : ERR_PTR(-ENOENT);
5720 #ifdef CONFIG_CGROUP_DEBUG
5721 static struct cgroup_subsys_state *
5722 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5724 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5727 return ERR_PTR(-ENOMEM);
5732 static void debug_css_free(struct cgroup_subsys_state *css)
5737 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5740 return cgroup_task_count(css->cgroup);
5743 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5746 return (u64)(unsigned long)current->cgroups;
5749 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5755 count = atomic_read(&task_css_set(current)->refcount);
5760 static int current_css_set_cg_links_read(struct cgroup_subsys_state *css,
5762 struct seq_file *seq)
5764 struct cgrp_cset_link *link;
5765 struct css_set *cset;
5767 read_lock(&css_set_lock);
5769 cset = rcu_dereference(current->cgroups);
5770 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5771 struct cgroup *c = link->cgrp;
5775 name = c->dentry->d_name.name;
5778 seq_printf(seq, "Root %d group %s\n",
5779 c->root->hierarchy_id, name);
5782 read_unlock(&css_set_lock);
5786 #define MAX_TASKS_SHOWN_PER_CSS 25
5787 static int cgroup_css_links_read(struct cgroup_subsys_state *css,
5788 struct cftype *cft, struct seq_file *seq)
5790 struct cgrp_cset_link *link;
5792 read_lock(&css_set_lock);
5793 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5794 struct css_set *cset = link->cset;
5795 struct task_struct *task;
5797 seq_printf(seq, "css_set %p\n", cset);
5798 list_for_each_entry(task, &cset->tasks, cg_list) {
5799 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
5800 seq_puts(seq, " ...\n");
5803 seq_printf(seq, " task %d\n",
5804 task_pid_vnr(task));
5808 read_unlock(&css_set_lock);
5812 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5814 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5817 static struct cftype debug_files[] = {
5819 .name = "taskcount",
5820 .read_u64 = debug_taskcount_read,
5824 .name = "current_css_set",
5825 .read_u64 = current_css_set_read,
5829 .name = "current_css_set_refcount",
5830 .read_u64 = current_css_set_refcount_read,
5834 .name = "current_css_set_cg_links",
5835 .read_seq_string = current_css_set_cg_links_read,
5839 .name = "cgroup_css_links",
5840 .read_seq_string = cgroup_css_links_read,
5844 .name = "releasable",
5845 .read_u64 = releasable_read,
5851 struct cgroup_subsys debug_subsys = {
5853 .css_alloc = debug_css_alloc,
5854 .css_free = debug_css_free,
5855 .subsys_id = debug_subsys_id,
5856 .base_cftypes = debug_files,
5858 #endif /* CONFIG_CGROUP_DEBUG */