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 task_subsys_state_check() */
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;
122 struct simple_xattrs xattrs;
126 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
127 * cgroup_subsys->use_id != 0.
129 #define CSS_ID_MAX (65535)
132 * The css to which this ID points. This pointer is set to valid value
133 * after cgroup is populated. If cgroup is removed, this will be NULL.
134 * This pointer is expected to be RCU-safe because destroy()
135 * is called after synchronize_rcu(). But for safe use, css_tryget()
136 * should be used for avoiding race.
138 struct cgroup_subsys_state __rcu *css;
144 * Depth in hierarchy which this ID belongs to.
146 unsigned short depth;
148 * ID is freed by RCU. (and lookup routine is RCU safe.)
150 struct rcu_head rcu_head;
152 * Hierarchy of CSS ID belongs to.
154 unsigned short stack[0]; /* Array of Length (depth+1) */
158 * cgroup_event represents events which userspace want to receive.
160 struct cgroup_event {
162 * Cgroup which the event belongs to.
166 * Control file which the event associated.
170 * eventfd to signal userspace about the event.
172 struct eventfd_ctx *eventfd;
174 * Each of these stored in a list by the cgroup.
176 struct list_head list;
178 * All fields below needed to unregister event when
179 * userspace closes eventfd.
182 wait_queue_head_t *wqh;
184 struct work_struct remove;
187 /* The list of hierarchy roots */
189 static LIST_HEAD(cgroup_roots);
190 static int cgroup_root_count;
193 * Hierarchy ID allocation and mapping. It follows the same exclusion
194 * rules as other root ops - both cgroup_mutex and cgroup_root_mutex for
195 * writes, either for reads.
197 static DEFINE_IDR(cgroup_hierarchy_idr);
199 static struct cgroup_name root_cgroup_name = { .name = "/" };
202 * Assign a monotonically increasing serial number to cgroups. It
203 * guarantees cgroups with bigger numbers are newer than those with smaller
204 * numbers. Also, as cgroups are always appended to the parent's
205 * ->children list, it guarantees that sibling cgroups are always sorted in
206 * the ascending serial number order on the list. Protected by
209 static u64 cgroup_serial_nr_next = 1;
211 /* This flag indicates whether tasks in the fork and exit paths should
212 * check for fork/exit handlers to call. This avoids us having to do
213 * extra work in the fork/exit path if none of the subsystems need to
216 static int need_forkexit_callback __read_mostly;
218 static struct cftype cgroup_base_files[];
220 static void cgroup_offline_fn(struct work_struct *work);
221 static int cgroup_destroy_locked(struct cgroup *cgrp);
222 static int cgroup_addrm_files(struct cgroup *cgrp, struct cgroup_subsys *subsys,
223 struct cftype cfts[], bool is_add);
225 /* convenient tests for these bits */
226 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
228 return test_bit(CGRP_DEAD, &cgrp->flags);
232 * cgroup_is_descendant - test ancestry
233 * @cgrp: the cgroup to be tested
234 * @ancestor: possible ancestor of @cgrp
236 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
237 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
238 * and @ancestor are accessible.
240 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
243 if (cgrp == ancestor)
249 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
251 static int cgroup_is_releasable(const struct cgroup *cgrp)
254 (1 << CGRP_RELEASABLE) |
255 (1 << CGRP_NOTIFY_ON_RELEASE);
256 return (cgrp->flags & bits) == bits;
259 static int notify_on_release(const struct cgroup *cgrp)
261 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
265 * for_each_subsys - iterate all loaded cgroup subsystems
266 * @ss: the iteration cursor
267 * @i: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
269 * Should be called under cgroup_mutex.
271 #define for_each_subsys(ss, i) \
272 for ((i) = 0; (i) < CGROUP_SUBSYS_COUNT; (i)++) \
273 if (({ lockdep_assert_held(&cgroup_mutex); \
274 !((ss) = cgroup_subsys[i]); })) { } \
278 * for_each_builtin_subsys - iterate all built-in cgroup subsystems
279 * @ss: the iteration cursor
280 * @i: the index of @ss, CGROUP_BUILTIN_SUBSYS_COUNT after reaching the end
282 * Bulit-in subsystems are always present and iteration itself doesn't
283 * require any synchronization.
285 #define for_each_builtin_subsys(ss, i) \
286 for ((i) = 0; (i) < CGROUP_BUILTIN_SUBSYS_COUNT && \
287 (((ss) = cgroup_subsys[i]) || true); (i)++)
289 /* iterate each subsystem attached to a hierarchy */
290 #define for_each_root_subsys(root, ss) \
291 list_for_each_entry((ss), &(root)->subsys_list, sibling)
293 /* iterate across the active hierarchies */
294 #define for_each_active_root(root) \
295 list_for_each_entry((root), &cgroup_roots, root_list)
297 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
299 return dentry->d_fsdata;
302 static inline struct cfent *__d_cfe(struct dentry *dentry)
304 return dentry->d_fsdata;
307 static inline struct cftype *__d_cft(struct dentry *dentry)
309 return __d_cfe(dentry)->type;
313 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
314 * @cgrp: the cgroup to be checked for liveness
316 * On success, returns true; the mutex should be later unlocked. On
317 * failure returns false with no lock held.
319 static bool cgroup_lock_live_group(struct cgroup *cgrp)
321 mutex_lock(&cgroup_mutex);
322 if (cgroup_is_dead(cgrp)) {
323 mutex_unlock(&cgroup_mutex);
329 /* the list of cgroups eligible for automatic release. Protected by
330 * release_list_lock */
331 static LIST_HEAD(release_list);
332 static DEFINE_RAW_SPINLOCK(release_list_lock);
333 static void cgroup_release_agent(struct work_struct *work);
334 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
335 static void check_for_release(struct cgroup *cgrp);
338 * A cgroup can be associated with multiple css_sets as different tasks may
339 * belong to different cgroups on different hierarchies. In the other
340 * direction, a css_set is naturally associated with multiple cgroups.
341 * This M:N relationship is represented by the following link structure
342 * which exists for each association and allows traversing the associations
345 struct cgrp_cset_link {
346 /* the cgroup and css_set this link associates */
348 struct css_set *cset;
350 /* list of cgrp_cset_links anchored at cgrp->cset_links */
351 struct list_head cset_link;
353 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
354 struct list_head cgrp_link;
357 /* The default css_set - used by init and its children prior to any
358 * hierarchies being mounted. It contains a pointer to the root state
359 * for each subsystem. Also used to anchor the list of css_sets. Not
360 * reference-counted, to improve performance when child cgroups
361 * haven't been created.
364 static struct css_set init_css_set;
365 static struct cgrp_cset_link init_cgrp_cset_link;
367 static int cgroup_init_idr(struct cgroup_subsys *ss,
368 struct cgroup_subsys_state *css);
370 /* css_set_lock protects the list of css_set objects, and the
371 * chain of tasks off each css_set. Nests outside task->alloc_lock
372 * due to cgroup_iter_start() */
373 static DEFINE_RWLOCK(css_set_lock);
374 static int css_set_count;
377 * hash table for cgroup groups. This improves the performance to find
378 * an existing css_set. This hash doesn't (currently) take into
379 * account cgroups in empty hierarchies.
381 #define CSS_SET_HASH_BITS 7
382 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
384 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
386 unsigned long key = 0UL;
387 struct cgroup_subsys *ss;
390 for_each_subsys(ss, i)
391 key += (unsigned long)css[i];
392 key = (key >> 16) ^ key;
397 /* We don't maintain the lists running through each css_set to its
398 * task until after the first call to cgroup_iter_start(). This
399 * reduces the fork()/exit() overhead for people who have cgroups
400 * compiled into their kernel but not actually in use */
401 static int use_task_css_set_links __read_mostly;
403 static void __put_css_set(struct css_set *cset, int taskexit)
405 struct cgrp_cset_link *link, *tmp_link;
408 * Ensure that the refcount doesn't hit zero while any readers
409 * can see it. Similar to atomic_dec_and_lock(), but for an
412 if (atomic_add_unless(&cset->refcount, -1, 1))
414 write_lock(&css_set_lock);
415 if (!atomic_dec_and_test(&cset->refcount)) {
416 write_unlock(&css_set_lock);
420 /* This css_set is dead. unlink it and release cgroup refcounts */
421 hash_del(&cset->hlist);
424 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
425 struct cgroup *cgrp = link->cgrp;
427 list_del(&link->cset_link);
428 list_del(&link->cgrp_link);
430 /* @cgrp can't go away while we're holding css_set_lock */
431 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
433 set_bit(CGRP_RELEASABLE, &cgrp->flags);
434 check_for_release(cgrp);
440 write_unlock(&css_set_lock);
441 kfree_rcu(cset, rcu_head);
445 * refcounted get/put for css_set objects
447 static inline void get_css_set(struct css_set *cset)
449 atomic_inc(&cset->refcount);
452 static inline void put_css_set(struct css_set *cset)
454 __put_css_set(cset, 0);
457 static inline void put_css_set_taskexit(struct css_set *cset)
459 __put_css_set(cset, 1);
463 * compare_css_sets - helper function for find_existing_css_set().
464 * @cset: candidate css_set being tested
465 * @old_cset: existing css_set for a task
466 * @new_cgrp: cgroup that's being entered by the task
467 * @template: desired set of css pointers in css_set (pre-calculated)
469 * Returns true if "cg" matches "old_cg" except for the hierarchy
470 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
472 static bool compare_css_sets(struct css_set *cset,
473 struct css_set *old_cset,
474 struct cgroup *new_cgrp,
475 struct cgroup_subsys_state *template[])
477 struct list_head *l1, *l2;
479 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
480 /* Not all subsystems matched */
485 * Compare cgroup pointers in order to distinguish between
486 * different cgroups in heirarchies with no subsystems. We
487 * could get by with just this check alone (and skip the
488 * memcmp above) but on most setups the memcmp check will
489 * avoid the need for this more expensive check on almost all
493 l1 = &cset->cgrp_links;
494 l2 = &old_cset->cgrp_links;
496 struct cgrp_cset_link *link1, *link2;
497 struct cgroup *cgrp1, *cgrp2;
501 /* See if we reached the end - both lists are equal length. */
502 if (l1 == &cset->cgrp_links) {
503 BUG_ON(l2 != &old_cset->cgrp_links);
506 BUG_ON(l2 == &old_cset->cgrp_links);
508 /* Locate the cgroups associated with these links. */
509 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
510 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
513 /* Hierarchies should be linked in the same order. */
514 BUG_ON(cgrp1->root != cgrp2->root);
517 * If this hierarchy is the hierarchy of the cgroup
518 * that's changing, then we need to check that this
519 * css_set points to the new cgroup; if it's any other
520 * hierarchy, then this css_set should point to the
521 * same cgroup as the old css_set.
523 if (cgrp1->root == new_cgrp->root) {
524 if (cgrp1 != new_cgrp)
535 * find_existing_css_set - init css array and find the matching css_set
536 * @old_cset: the css_set that we're using before the cgroup transition
537 * @cgrp: the cgroup that we're moving into
538 * @template: out param for the new set of csses, should be clear on entry
540 static struct css_set *find_existing_css_set(struct css_set *old_cset,
542 struct cgroup_subsys_state *template[])
544 struct cgroupfs_root *root = cgrp->root;
545 struct cgroup_subsys *ss;
546 struct css_set *cset;
551 * Build the set of subsystem state objects that we want to see in the
552 * new css_set. while subsystems can change globally, the entries here
553 * won't change, so no need for locking.
555 for_each_subsys(ss, i) {
556 if (root->subsys_mask & (1UL << i)) {
557 /* Subsystem is in this hierarchy. So we want
558 * the subsystem state from the new
560 template[i] = cgrp->subsys[i];
562 /* Subsystem is not in this hierarchy, so we
563 * don't want to change the subsystem state */
564 template[i] = old_cset->subsys[i];
568 key = css_set_hash(template);
569 hash_for_each_possible(css_set_table, cset, hlist, key) {
570 if (!compare_css_sets(cset, old_cset, cgrp, template))
573 /* This css_set matches what we need */
577 /* No existing cgroup group matched */
581 static void free_cgrp_cset_links(struct list_head *links_to_free)
583 struct cgrp_cset_link *link, *tmp_link;
585 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
586 list_del(&link->cset_link);
592 * allocate_cgrp_cset_links - allocate cgrp_cset_links
593 * @count: the number of links to allocate
594 * @tmp_links: list_head the allocated links are put on
596 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
597 * through ->cset_link. Returns 0 on success or -errno.
599 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
601 struct cgrp_cset_link *link;
604 INIT_LIST_HEAD(tmp_links);
606 for (i = 0; i < count; i++) {
607 link = kzalloc(sizeof(*link), GFP_KERNEL);
609 free_cgrp_cset_links(tmp_links);
612 list_add(&link->cset_link, tmp_links);
618 * link_css_set - a helper function to link a css_set to a cgroup
619 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
620 * @cset: the css_set to be linked
621 * @cgrp: the destination cgroup
623 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
626 struct cgrp_cset_link *link;
628 BUG_ON(list_empty(tmp_links));
629 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
632 list_move(&link->cset_link, &cgrp->cset_links);
634 * Always add links to the tail of the list so that the list
635 * is sorted by order of hierarchy creation
637 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
641 * find_css_set - return a new css_set with one cgroup updated
642 * @old_cset: the baseline css_set
643 * @cgrp: the cgroup to be updated
645 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
646 * substituted into the appropriate hierarchy.
648 static struct css_set *find_css_set(struct css_set *old_cset,
651 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
652 struct css_set *cset;
653 struct list_head tmp_links;
654 struct cgrp_cset_link *link;
657 lockdep_assert_held(&cgroup_mutex);
659 /* First see if we already have a cgroup group that matches
661 read_lock(&css_set_lock);
662 cset = find_existing_css_set(old_cset, cgrp, template);
665 read_unlock(&css_set_lock);
670 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
674 /* Allocate all the cgrp_cset_link objects that we'll need */
675 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
680 atomic_set(&cset->refcount, 1);
681 INIT_LIST_HEAD(&cset->cgrp_links);
682 INIT_LIST_HEAD(&cset->tasks);
683 INIT_HLIST_NODE(&cset->hlist);
685 /* Copy the set of subsystem state objects generated in
686 * find_existing_css_set() */
687 memcpy(cset->subsys, template, sizeof(cset->subsys));
689 write_lock(&css_set_lock);
690 /* Add reference counts and links from the new css_set. */
691 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
692 struct cgroup *c = link->cgrp;
694 if (c->root == cgrp->root)
696 link_css_set(&tmp_links, cset, c);
699 BUG_ON(!list_empty(&tmp_links));
703 /* Add this cgroup group to the hash table */
704 key = css_set_hash(cset->subsys);
705 hash_add(css_set_table, &cset->hlist, key);
707 write_unlock(&css_set_lock);
713 * Return the cgroup for "task" from the given hierarchy. Must be
714 * called with cgroup_mutex held.
716 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
717 struct cgroupfs_root *root)
719 struct css_set *cset;
720 struct cgroup *res = NULL;
722 BUG_ON(!mutex_is_locked(&cgroup_mutex));
723 read_lock(&css_set_lock);
725 * No need to lock the task - since we hold cgroup_mutex the
726 * task can't change groups, so the only thing that can happen
727 * is that it exits and its css is set back to init_css_set.
729 cset = task_css_set(task);
730 if (cset == &init_css_set) {
731 res = &root->top_cgroup;
733 struct cgrp_cset_link *link;
735 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
736 struct cgroup *c = link->cgrp;
738 if (c->root == root) {
744 read_unlock(&css_set_lock);
750 * There is one global cgroup mutex. We also require taking
751 * task_lock() when dereferencing a task's cgroup subsys pointers.
752 * See "The task_lock() exception", at the end of this comment.
754 * A task must hold cgroup_mutex to modify cgroups.
756 * Any task can increment and decrement the count field without lock.
757 * So in general, code holding cgroup_mutex can't rely on the count
758 * field not changing. However, if the count goes to zero, then only
759 * cgroup_attach_task() can increment it again. Because a count of zero
760 * means that no tasks are currently attached, therefore there is no
761 * way a task attached to that cgroup can fork (the other way to
762 * increment the count). So code holding cgroup_mutex can safely
763 * assume that if the count is zero, it will stay zero. Similarly, if
764 * a task holds cgroup_mutex on a cgroup with zero count, it
765 * knows that the cgroup won't be removed, as cgroup_rmdir()
768 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
769 * (usually) take cgroup_mutex. These are the two most performance
770 * critical pieces of code here. The exception occurs on cgroup_exit(),
771 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
772 * is taken, and if the cgroup count is zero, a usermode call made
773 * to the release agent with the name of the cgroup (path relative to
774 * the root of cgroup file system) as the argument.
776 * A cgroup can only be deleted if both its 'count' of using tasks
777 * is zero, and its list of 'children' cgroups is empty. Since all
778 * tasks in the system use _some_ cgroup, and since there is always at
779 * least one task in the system (init, pid == 1), therefore, top_cgroup
780 * always has either children cgroups and/or using tasks. So we don't
781 * need a special hack to ensure that top_cgroup cannot be deleted.
783 * The task_lock() exception
785 * The need for this exception arises from the action of
786 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
787 * another. It does so using cgroup_mutex, however there are
788 * several performance critical places that need to reference
789 * task->cgroup without the expense of grabbing a system global
790 * mutex. Therefore except as noted below, when dereferencing or, as
791 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
792 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
793 * the task_struct routinely used for such matters.
795 * P.S. One more locking exception. RCU is used to guard the
796 * update of a tasks cgroup pointer by cgroup_attach_task()
800 * A couple of forward declarations required, due to cyclic reference loop:
801 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
802 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
806 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode);
807 static struct dentry *cgroup_lookup(struct inode *, struct dentry *, unsigned int);
808 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
809 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
810 static const struct inode_operations cgroup_dir_inode_operations;
811 static const struct file_operations proc_cgroupstats_operations;
813 static struct backing_dev_info cgroup_backing_dev_info = {
815 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
818 static int alloc_css_id(struct cgroup_subsys *ss,
819 struct cgroup *parent, struct cgroup *child);
821 static struct inode *cgroup_new_inode(umode_t mode, struct super_block *sb)
823 struct inode *inode = new_inode(sb);
826 inode->i_ino = get_next_ino();
827 inode->i_mode = mode;
828 inode->i_uid = current_fsuid();
829 inode->i_gid = current_fsgid();
830 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
831 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
836 static struct cgroup_name *cgroup_alloc_name(struct dentry *dentry)
838 struct cgroup_name *name;
840 name = kmalloc(sizeof(*name) + dentry->d_name.len + 1, GFP_KERNEL);
843 strcpy(name->name, dentry->d_name.name);
847 static void cgroup_free_fn(struct work_struct *work)
849 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
850 struct cgroup_subsys *ss;
852 mutex_lock(&cgroup_mutex);
854 * Release the subsystem state objects.
856 for_each_root_subsys(cgrp->root, ss)
859 cgrp->root->number_of_cgroups--;
860 mutex_unlock(&cgroup_mutex);
863 * We get a ref to the parent's dentry, and put the ref when
864 * this cgroup is being freed, so it's guaranteed that the
865 * parent won't be destroyed before its children.
867 dput(cgrp->parent->dentry);
869 ida_simple_remove(&cgrp->root->cgroup_ida, cgrp->id);
872 * Drop the active superblock reference that we took when we
873 * created the cgroup. This will free cgrp->root, if we are
874 * holding the last reference to @sb.
876 deactivate_super(cgrp->root->sb);
879 * if we're getting rid of the cgroup, refcount should ensure
880 * that there are no pidlists left.
882 BUG_ON(!list_empty(&cgrp->pidlists));
884 simple_xattrs_free(&cgrp->xattrs);
886 kfree(rcu_dereference_raw(cgrp->name));
890 static void cgroup_free_rcu(struct rcu_head *head)
892 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
894 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
895 schedule_work(&cgrp->destroy_work);
898 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
900 /* is dentry a directory ? if so, kfree() associated cgroup */
901 if (S_ISDIR(inode->i_mode)) {
902 struct cgroup *cgrp = dentry->d_fsdata;
904 BUG_ON(!(cgroup_is_dead(cgrp)));
905 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
907 struct cfent *cfe = __d_cfe(dentry);
908 struct cgroup *cgrp = dentry->d_parent->d_fsdata;
910 WARN_ONCE(!list_empty(&cfe->node) &&
911 cgrp != &cgrp->root->top_cgroup,
912 "cfe still linked for %s\n", cfe->type->name);
913 simple_xattrs_free(&cfe->xattrs);
919 static int cgroup_delete(const struct dentry *d)
924 static void remove_dir(struct dentry *d)
926 struct dentry *parent = dget(d->d_parent);
929 simple_rmdir(parent->d_inode, d);
933 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
937 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
938 lockdep_assert_held(&cgroup_mutex);
941 * If we're doing cleanup due to failure of cgroup_create(),
942 * the corresponding @cfe may not exist.
944 list_for_each_entry(cfe, &cgrp->files, node) {
945 struct dentry *d = cfe->dentry;
947 if (cft && cfe->type != cft)
952 simple_unlink(cgrp->dentry->d_inode, d);
953 list_del_init(&cfe->node);
961 * cgroup_clear_dir - remove subsys files in a cgroup directory
962 * @cgrp: target cgroup
963 * @subsys_mask: mask of the subsystem ids whose files should be removed
965 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
967 struct cgroup_subsys *ss;
970 for_each_subsys(ss, i) {
971 struct cftype_set *set;
973 if (!test_bit(i, &subsys_mask))
975 list_for_each_entry(set, &ss->cftsets, node)
976 cgroup_addrm_files(cgrp, NULL, set->cfts, false);
981 * NOTE : the dentry must have been dget()'ed
983 static void cgroup_d_remove_dir(struct dentry *dentry)
985 struct dentry *parent;
987 parent = dentry->d_parent;
988 spin_lock(&parent->d_lock);
989 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
990 list_del_init(&dentry->d_u.d_child);
991 spin_unlock(&dentry->d_lock);
992 spin_unlock(&parent->d_lock);
997 * Call with cgroup_mutex held. Drops reference counts on modules, including
998 * any duplicate ones that parse_cgroupfs_options took. If this function
999 * returns an error, no reference counts are touched.
1001 static int rebind_subsystems(struct cgroupfs_root *root,
1002 unsigned long added_mask, unsigned removed_mask)
1004 struct cgroup *cgrp = &root->top_cgroup;
1005 struct cgroup_subsys *ss;
1006 unsigned long pinned = 0;
1009 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1010 BUG_ON(!mutex_is_locked(&cgroup_root_mutex));
1012 /* Check that any added subsystems are currently free */
1013 for_each_subsys(ss, i) {
1014 if (!(added_mask & (1 << i)))
1017 /* is the subsystem mounted elsewhere? */
1018 if (ss->root != &cgroup_dummy_root) {
1023 /* pin the module */
1024 if (!try_module_get(ss->module)) {
1031 /* subsys could be missing if unloaded between parsing and here */
1032 if (added_mask != pinned) {
1037 ret = cgroup_populate_dir(cgrp, added_mask);
1042 * Nothing can fail from this point on. Remove files for the
1043 * removed subsystems and rebind each subsystem.
1045 cgroup_clear_dir(cgrp, removed_mask);
1047 for_each_subsys(ss, i) {
1048 unsigned long bit = 1UL << i;
1050 if (bit & added_mask) {
1051 /* We're binding this subsystem to this hierarchy */
1052 BUG_ON(cgrp->subsys[i]);
1053 BUG_ON(!cgroup_dummy_top->subsys[i]);
1054 BUG_ON(cgroup_dummy_top->subsys[i]->cgroup != cgroup_dummy_top);
1056 cgrp->subsys[i] = cgroup_dummy_top->subsys[i];
1057 cgrp->subsys[i]->cgroup = cgrp;
1058 list_move(&ss->sibling, &root->subsys_list);
1063 /* refcount was already taken, and we're keeping it */
1064 root->subsys_mask |= bit;
1065 } else if (bit & removed_mask) {
1066 /* We're removing this subsystem */
1067 BUG_ON(cgrp->subsys[i] != cgroup_dummy_top->subsys[i]);
1068 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
1071 ss->bind(cgroup_dummy_top);
1072 cgroup_dummy_top->subsys[i]->cgroup = cgroup_dummy_top;
1073 cgrp->subsys[i] = NULL;
1074 cgroup_subsys[i]->root = &cgroup_dummy_root;
1075 list_move(&ss->sibling, &cgroup_dummy_root.subsys_list);
1077 /* subsystem is now free - drop reference on module */
1078 module_put(ss->module);
1079 root->subsys_mask &= ~bit;
1084 * Mark @root has finished binding subsystems. @root->subsys_mask
1085 * now matches the bound subsystems.
1087 root->flags |= CGRP_ROOT_SUBSYS_BOUND;
1092 for_each_subsys(ss, i)
1093 if (pinned & (1 << i))
1094 module_put(ss->module);
1098 static int cgroup_show_options(struct seq_file *seq, struct dentry *dentry)
1100 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
1101 struct cgroup_subsys *ss;
1103 mutex_lock(&cgroup_root_mutex);
1104 for_each_root_subsys(root, ss)
1105 seq_printf(seq, ",%s", ss->name);
1106 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1107 seq_puts(seq, ",sane_behavior");
1108 if (root->flags & CGRP_ROOT_NOPREFIX)
1109 seq_puts(seq, ",noprefix");
1110 if (root->flags & CGRP_ROOT_XATTR)
1111 seq_puts(seq, ",xattr");
1112 if (strlen(root->release_agent_path))
1113 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1114 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1115 seq_puts(seq, ",clone_children");
1116 if (strlen(root->name))
1117 seq_printf(seq, ",name=%s", root->name);
1118 mutex_unlock(&cgroup_root_mutex);
1122 struct cgroup_sb_opts {
1123 unsigned long subsys_mask;
1124 unsigned long flags;
1125 char *release_agent;
1126 bool cpuset_clone_children;
1128 /* User explicitly requested empty subsystem */
1131 struct cgroupfs_root *new_root;
1136 * Convert a hierarchy specifier into a bitmask of subsystems and
1137 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1138 * array. This function takes refcounts on subsystems to be used, unless it
1139 * returns error, in which case no refcounts are taken.
1141 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1143 char *token, *o = data;
1144 bool all_ss = false, one_ss = false;
1145 unsigned long mask = (unsigned long)-1;
1146 struct cgroup_subsys *ss;
1149 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1151 #ifdef CONFIG_CPUSETS
1152 mask = ~(1UL << cpuset_subsys_id);
1155 memset(opts, 0, sizeof(*opts));
1157 while ((token = strsep(&o, ",")) != NULL) {
1160 if (!strcmp(token, "none")) {
1161 /* Explicitly have no subsystems */
1165 if (!strcmp(token, "all")) {
1166 /* Mutually exclusive option 'all' + subsystem name */
1172 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1173 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1176 if (!strcmp(token, "noprefix")) {
1177 opts->flags |= CGRP_ROOT_NOPREFIX;
1180 if (!strcmp(token, "clone_children")) {
1181 opts->cpuset_clone_children = true;
1184 if (!strcmp(token, "xattr")) {
1185 opts->flags |= CGRP_ROOT_XATTR;
1188 if (!strncmp(token, "release_agent=", 14)) {
1189 /* Specifying two release agents is forbidden */
1190 if (opts->release_agent)
1192 opts->release_agent =
1193 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1194 if (!opts->release_agent)
1198 if (!strncmp(token, "name=", 5)) {
1199 const char *name = token + 5;
1200 /* Can't specify an empty name */
1203 /* Must match [\w.-]+ */
1204 for (i = 0; i < strlen(name); i++) {
1208 if ((c == '.') || (c == '-') || (c == '_'))
1212 /* Specifying two names is forbidden */
1215 opts->name = kstrndup(name,
1216 MAX_CGROUP_ROOT_NAMELEN - 1,
1224 for_each_subsys(ss, i) {
1225 if (strcmp(token, ss->name))
1230 /* Mutually exclusive option 'all' + subsystem name */
1233 set_bit(i, &opts->subsys_mask);
1238 if (i == CGROUP_SUBSYS_COUNT)
1243 * If the 'all' option was specified select all the subsystems,
1244 * otherwise if 'none', 'name=' and a subsystem name options
1245 * were not specified, let's default to 'all'
1247 if (all_ss || (!one_ss && !opts->none && !opts->name))
1248 for_each_subsys(ss, i)
1250 set_bit(i, &opts->subsys_mask);
1252 /* Consistency checks */
1254 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1255 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1257 if (opts->flags & CGRP_ROOT_NOPREFIX) {
1258 pr_err("cgroup: sane_behavior: noprefix is not allowed\n");
1262 if (opts->cpuset_clone_children) {
1263 pr_err("cgroup: sane_behavior: clone_children is not allowed\n");
1269 * Option noprefix was introduced just for backward compatibility
1270 * with the old cpuset, so we allow noprefix only if mounting just
1271 * the cpuset subsystem.
1273 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1277 /* Can't specify "none" and some subsystems */
1278 if (opts->subsys_mask && opts->none)
1282 * We either have to specify by name or by subsystems. (So all
1283 * empty hierarchies must have a name).
1285 if (!opts->subsys_mask && !opts->name)
1291 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1294 struct cgroupfs_root *root = sb->s_fs_info;
1295 struct cgroup *cgrp = &root->top_cgroup;
1296 struct cgroup_sb_opts opts;
1297 unsigned long added_mask, removed_mask;
1299 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1300 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1304 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1305 mutex_lock(&cgroup_mutex);
1306 mutex_lock(&cgroup_root_mutex);
1308 /* See what subsystems are wanted */
1309 ret = parse_cgroupfs_options(data, &opts);
1313 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1314 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1315 task_tgid_nr(current), current->comm);
1317 added_mask = opts.subsys_mask & ~root->subsys_mask;
1318 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1320 /* Don't allow flags or name to change at remount */
1321 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1322 (opts.name && strcmp(opts.name, root->name))) {
1323 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1324 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1325 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1330 /* remounting is not allowed for populated hierarchies */
1331 if (root->number_of_cgroups > 1) {
1336 ret = rebind_subsystems(root, added_mask, removed_mask);
1340 if (opts.release_agent)
1341 strcpy(root->release_agent_path, opts.release_agent);
1343 kfree(opts.release_agent);
1345 mutex_unlock(&cgroup_root_mutex);
1346 mutex_unlock(&cgroup_mutex);
1347 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1351 static const struct super_operations cgroup_ops = {
1352 .statfs = simple_statfs,
1353 .drop_inode = generic_delete_inode,
1354 .show_options = cgroup_show_options,
1355 .remount_fs = cgroup_remount,
1358 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1360 INIT_LIST_HEAD(&cgrp->sibling);
1361 INIT_LIST_HEAD(&cgrp->children);
1362 INIT_LIST_HEAD(&cgrp->files);
1363 INIT_LIST_HEAD(&cgrp->cset_links);
1364 INIT_LIST_HEAD(&cgrp->release_list);
1365 INIT_LIST_HEAD(&cgrp->pidlists);
1366 mutex_init(&cgrp->pidlist_mutex);
1367 INIT_LIST_HEAD(&cgrp->event_list);
1368 spin_lock_init(&cgrp->event_list_lock);
1369 simple_xattrs_init(&cgrp->xattrs);
1372 static void init_cgroup_root(struct cgroupfs_root *root)
1374 struct cgroup *cgrp = &root->top_cgroup;
1376 INIT_LIST_HEAD(&root->subsys_list);
1377 INIT_LIST_HEAD(&root->root_list);
1378 root->number_of_cgroups = 1;
1380 RCU_INIT_POINTER(cgrp->name, &root_cgroup_name);
1381 init_cgroup_housekeeping(cgrp);
1384 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
1388 lockdep_assert_held(&cgroup_mutex);
1389 lockdep_assert_held(&cgroup_root_mutex);
1391 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
1396 root->hierarchy_id = id;
1400 static void cgroup_exit_root_id(struct cgroupfs_root *root)
1402 lockdep_assert_held(&cgroup_mutex);
1403 lockdep_assert_held(&cgroup_root_mutex);
1405 if (root->hierarchy_id) {
1406 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1407 root->hierarchy_id = 0;
1411 static int cgroup_test_super(struct super_block *sb, void *data)
1413 struct cgroup_sb_opts *opts = data;
1414 struct cgroupfs_root *root = sb->s_fs_info;
1416 /* If we asked for a name then it must match */
1417 if (opts->name && strcmp(opts->name, root->name))
1421 * If we asked for subsystems (or explicitly for no
1422 * subsystems) then they must match
1424 if ((opts->subsys_mask || opts->none)
1425 && (opts->subsys_mask != root->subsys_mask))
1431 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1433 struct cgroupfs_root *root;
1435 if (!opts->subsys_mask && !opts->none)
1438 root = kzalloc(sizeof(*root), GFP_KERNEL);
1440 return ERR_PTR(-ENOMEM);
1442 init_cgroup_root(root);
1445 * We need to set @root->subsys_mask now so that @root can be
1446 * matched by cgroup_test_super() before it finishes
1447 * initialization; otherwise, competing mounts with the same
1448 * options may try to bind the same subsystems instead of waiting
1449 * for the first one leading to unexpected mount errors.
1450 * SUBSYS_BOUND will be set once actual binding is complete.
1452 root->subsys_mask = opts->subsys_mask;
1453 root->flags = opts->flags;
1454 ida_init(&root->cgroup_ida);
1455 if (opts->release_agent)
1456 strcpy(root->release_agent_path, opts->release_agent);
1458 strcpy(root->name, opts->name);
1459 if (opts->cpuset_clone_children)
1460 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1464 static void cgroup_free_root(struct cgroupfs_root *root)
1467 /* hierarhcy ID shoulid already have been released */
1468 WARN_ON_ONCE(root->hierarchy_id);
1470 ida_destroy(&root->cgroup_ida);
1475 static int cgroup_set_super(struct super_block *sb, void *data)
1478 struct cgroup_sb_opts *opts = data;
1480 /* If we don't have a new root, we can't set up a new sb */
1481 if (!opts->new_root)
1484 BUG_ON(!opts->subsys_mask && !opts->none);
1486 ret = set_anon_super(sb, NULL);
1490 sb->s_fs_info = opts->new_root;
1491 opts->new_root->sb = sb;
1493 sb->s_blocksize = PAGE_CACHE_SIZE;
1494 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1495 sb->s_magic = CGROUP_SUPER_MAGIC;
1496 sb->s_op = &cgroup_ops;
1501 static int cgroup_get_rootdir(struct super_block *sb)
1503 static const struct dentry_operations cgroup_dops = {
1504 .d_iput = cgroup_diput,
1505 .d_delete = cgroup_delete,
1508 struct inode *inode =
1509 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1514 inode->i_fop = &simple_dir_operations;
1515 inode->i_op = &cgroup_dir_inode_operations;
1516 /* directories start off with i_nlink == 2 (for "." entry) */
1518 sb->s_root = d_make_root(inode);
1521 /* for everything else we want ->d_op set */
1522 sb->s_d_op = &cgroup_dops;
1526 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1527 int flags, const char *unused_dev_name,
1530 struct cgroup_sb_opts opts;
1531 struct cgroupfs_root *root;
1533 struct super_block *sb;
1534 struct cgroupfs_root *new_root;
1535 struct list_head tmp_links;
1536 struct inode *inode;
1537 const struct cred *cred;
1539 /* First find the desired set of subsystems */
1540 mutex_lock(&cgroup_mutex);
1541 ret = parse_cgroupfs_options(data, &opts);
1542 mutex_unlock(&cgroup_mutex);
1547 * Allocate a new cgroup root. We may not need it if we're
1548 * reusing an existing hierarchy.
1550 new_root = cgroup_root_from_opts(&opts);
1551 if (IS_ERR(new_root)) {
1552 ret = PTR_ERR(new_root);
1555 opts.new_root = new_root;
1557 /* Locate an existing or new sb for this hierarchy */
1558 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, 0, &opts);
1561 cgroup_free_root(opts.new_root);
1565 root = sb->s_fs_info;
1567 if (root == opts.new_root) {
1568 /* We used the new root structure, so this is a new hierarchy */
1569 struct cgroup *root_cgrp = &root->top_cgroup;
1570 struct cgroupfs_root *existing_root;
1572 struct css_set *cset;
1574 BUG_ON(sb->s_root != NULL);
1576 ret = cgroup_get_rootdir(sb);
1578 goto drop_new_super;
1579 inode = sb->s_root->d_inode;
1581 mutex_lock(&inode->i_mutex);
1582 mutex_lock(&cgroup_mutex);
1583 mutex_lock(&cgroup_root_mutex);
1585 /* Check for name clashes with existing mounts */
1587 if (strlen(root->name))
1588 for_each_active_root(existing_root)
1589 if (!strcmp(existing_root->name, root->name))
1593 * We're accessing css_set_count without locking
1594 * css_set_lock here, but that's OK - it can only be
1595 * increased by someone holding cgroup_lock, and
1596 * that's us. The worst that can happen is that we
1597 * have some link structures left over
1599 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1603 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1604 ret = cgroup_init_root_id(root, 2, 0);
1608 sb->s_root->d_fsdata = root_cgrp;
1609 root_cgrp->dentry = sb->s_root;
1612 * We're inside get_sb() and will call lookup_one_len() to
1613 * create the root files, which doesn't work if SELinux is
1614 * in use. The following cred dancing somehow works around
1615 * it. See 2ce9738ba ("cgroupfs: use init_cred when
1616 * populating new cgroupfs mount") for more details.
1618 cred = override_creds(&init_cred);
1620 ret = cgroup_addrm_files(root_cgrp, NULL, cgroup_base_files, true);
1624 ret = rebind_subsystems(root, root->subsys_mask, 0);
1631 * There must be no failure case after here, since rebinding
1632 * takes care of subsystems' refcounts, which are explicitly
1633 * dropped in the failure exit path.
1636 list_add(&root->root_list, &cgroup_roots);
1637 cgroup_root_count++;
1639 /* Link the top cgroup in this hierarchy into all
1640 * the css_set objects */
1641 write_lock(&css_set_lock);
1642 hash_for_each(css_set_table, i, cset, hlist)
1643 link_css_set(&tmp_links, cset, root_cgrp);
1644 write_unlock(&css_set_lock);
1646 free_cgrp_cset_links(&tmp_links);
1648 BUG_ON(!list_empty(&root_cgrp->children));
1649 BUG_ON(root->number_of_cgroups != 1);
1651 mutex_unlock(&cgroup_root_mutex);
1652 mutex_unlock(&cgroup_mutex);
1653 mutex_unlock(&inode->i_mutex);
1656 * We re-used an existing hierarchy - the new root (if
1657 * any) is not needed
1659 cgroup_free_root(opts.new_root);
1661 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1662 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1663 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1665 goto drop_new_super;
1667 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1672 kfree(opts.release_agent);
1674 return dget(sb->s_root);
1677 free_cgrp_cset_links(&tmp_links);
1678 cgroup_addrm_files(&root->top_cgroup, NULL, cgroup_base_files, false);
1681 cgroup_exit_root_id(root);
1682 mutex_unlock(&cgroup_root_mutex);
1683 mutex_unlock(&cgroup_mutex);
1684 mutex_unlock(&inode->i_mutex);
1686 deactivate_locked_super(sb);
1688 kfree(opts.release_agent);
1690 return ERR_PTR(ret);
1693 static void cgroup_kill_sb(struct super_block *sb) {
1694 struct cgroupfs_root *root = sb->s_fs_info;
1695 struct cgroup *cgrp = &root->top_cgroup;
1696 struct cgrp_cset_link *link, *tmp_link;
1701 BUG_ON(root->number_of_cgroups != 1);
1702 BUG_ON(!list_empty(&cgrp->children));
1704 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1705 mutex_lock(&cgroup_mutex);
1706 mutex_lock(&cgroup_root_mutex);
1708 /* Rebind all subsystems back to the default hierarchy */
1709 if (root->flags & CGRP_ROOT_SUBSYS_BOUND) {
1710 ret = rebind_subsystems(root, 0, root->subsys_mask);
1711 /* Shouldn't be able to fail ... */
1716 * Release all the links from cset_links to this hierarchy's
1719 write_lock(&css_set_lock);
1721 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1722 list_del(&link->cset_link);
1723 list_del(&link->cgrp_link);
1726 write_unlock(&css_set_lock);
1728 if (!list_empty(&root->root_list)) {
1729 list_del(&root->root_list);
1730 cgroup_root_count--;
1733 cgroup_exit_root_id(root);
1735 mutex_unlock(&cgroup_root_mutex);
1736 mutex_unlock(&cgroup_mutex);
1737 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1739 simple_xattrs_free(&cgrp->xattrs);
1741 kill_litter_super(sb);
1742 cgroup_free_root(root);
1745 static struct file_system_type cgroup_fs_type = {
1747 .mount = cgroup_mount,
1748 .kill_sb = cgroup_kill_sb,
1751 static struct kobject *cgroup_kobj;
1754 * cgroup_path - generate the path of a cgroup
1755 * @cgrp: the cgroup in question
1756 * @buf: the buffer to write the path into
1757 * @buflen: the length of the buffer
1759 * Writes path of cgroup into buf. Returns 0 on success, -errno on error.
1761 * We can't generate cgroup path using dentry->d_name, as accessing
1762 * dentry->name must be protected by irq-unsafe dentry->d_lock or parent
1763 * inode's i_mutex, while on the other hand cgroup_path() can be called
1764 * with some irq-safe spinlocks held.
1766 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1768 int ret = -ENAMETOOLONG;
1771 if (!cgrp->parent) {
1772 if (strlcpy(buf, "/", buflen) >= buflen)
1773 return -ENAMETOOLONG;
1777 start = buf + buflen - 1;
1782 const char *name = cgroup_name(cgrp);
1786 if ((start -= len) < buf)
1788 memcpy(start, name, len);
1794 cgrp = cgrp->parent;
1795 } while (cgrp->parent);
1797 memmove(buf, start, buf + buflen - start);
1802 EXPORT_SYMBOL_GPL(cgroup_path);
1805 * task_cgroup_path_from_hierarchy - cgroup path of a task on a hierarchy
1806 * @task: target task
1807 * @hierarchy_id: the hierarchy to look up @task's cgroup from
1808 * @buf: the buffer to write the path into
1809 * @buflen: the length of the buffer
1811 * Determine @task's cgroup on the hierarchy specified by @hierarchy_id and
1812 * copy its path into @buf. This function grabs cgroup_mutex and shouldn't
1813 * be used inside locks used by cgroup controller callbacks.
1815 int task_cgroup_path_from_hierarchy(struct task_struct *task, int hierarchy_id,
1816 char *buf, size_t buflen)
1818 struct cgroupfs_root *root;
1819 struct cgroup *cgrp = NULL;
1822 mutex_lock(&cgroup_mutex);
1824 root = idr_find(&cgroup_hierarchy_idr, hierarchy_id);
1826 cgrp = task_cgroup_from_root(task, root);
1827 ret = cgroup_path(cgrp, buf, buflen);
1830 mutex_unlock(&cgroup_mutex);
1834 EXPORT_SYMBOL_GPL(task_cgroup_path_from_hierarchy);
1837 * Control Group taskset
1839 struct task_and_cgroup {
1840 struct task_struct *task;
1841 struct cgroup *cgrp;
1845 struct cgroup_taskset {
1846 struct task_and_cgroup single;
1847 struct flex_array *tc_array;
1850 struct cgroup *cur_cgrp;
1854 * cgroup_taskset_first - reset taskset and return the first task
1855 * @tset: taskset of interest
1857 * @tset iteration is initialized and the first task is returned.
1859 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1861 if (tset->tc_array) {
1863 return cgroup_taskset_next(tset);
1865 tset->cur_cgrp = tset->single.cgrp;
1866 return tset->single.task;
1869 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1872 * cgroup_taskset_next - iterate to the next task in taskset
1873 * @tset: taskset of interest
1875 * Return the next task in @tset. Iteration must have been initialized
1876 * with cgroup_taskset_first().
1878 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1880 struct task_and_cgroup *tc;
1882 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1885 tc = flex_array_get(tset->tc_array, tset->idx++);
1886 tset->cur_cgrp = tc->cgrp;
1889 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1892 * cgroup_taskset_cur_cgroup - return the matching cgroup for the current task
1893 * @tset: taskset of interest
1895 * Return the cgroup for the current (last returned) task of @tset. This
1896 * function must be preceded by either cgroup_taskset_first() or
1897 * cgroup_taskset_next().
1899 struct cgroup *cgroup_taskset_cur_cgroup(struct cgroup_taskset *tset)
1901 return tset->cur_cgrp;
1903 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_cgroup);
1906 * cgroup_taskset_size - return the number of tasks in taskset
1907 * @tset: taskset of interest
1909 int cgroup_taskset_size(struct cgroup_taskset *tset)
1911 return tset->tc_array ? tset->tc_array_len : 1;
1913 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1917 * cgroup_task_migrate - move a task from one cgroup to another.
1919 * Must be called with cgroup_mutex and threadgroup locked.
1921 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1922 struct task_struct *tsk,
1923 struct css_set *new_cset)
1925 struct css_set *old_cset;
1928 * We are synchronized through threadgroup_lock() against PF_EXITING
1929 * setting such that we can't race against cgroup_exit() changing the
1930 * css_set to init_css_set and dropping the old one.
1932 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1933 old_cset = task_css_set(tsk);
1936 rcu_assign_pointer(tsk->cgroups, new_cset);
1939 /* Update the css_set linked lists if we're using them */
1940 write_lock(&css_set_lock);
1941 if (!list_empty(&tsk->cg_list))
1942 list_move(&tsk->cg_list, &new_cset->tasks);
1943 write_unlock(&css_set_lock);
1946 * We just gained a reference on old_cset by taking it from the
1947 * task. As trading it for new_cset is protected by cgroup_mutex,
1948 * we're safe to drop it here; it will be freed under RCU.
1950 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1951 put_css_set(old_cset);
1955 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1956 * @cgrp: the cgroup to attach to
1957 * @tsk: the task or the leader of the threadgroup to be attached
1958 * @threadgroup: attach the whole threadgroup?
1960 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1961 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1963 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1966 int retval, i, group_size;
1967 struct cgroup_subsys *ss, *failed_ss = NULL;
1968 struct cgroupfs_root *root = cgrp->root;
1969 /* threadgroup list cursor and array */
1970 struct task_struct *leader = tsk;
1971 struct task_and_cgroup *tc;
1972 struct flex_array *group;
1973 struct cgroup_taskset tset = { };
1976 * step 0: in order to do expensive, possibly blocking operations for
1977 * every thread, we cannot iterate the thread group list, since it needs
1978 * rcu or tasklist locked. instead, build an array of all threads in the
1979 * group - group_rwsem prevents new threads from appearing, and if
1980 * threads exit, this will just be an over-estimate.
1983 group_size = get_nr_threads(tsk);
1986 /* flex_array supports very large thread-groups better than kmalloc. */
1987 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1990 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1991 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1993 goto out_free_group_list;
1997 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1998 * already PF_EXITING could be freed from underneath us unless we
1999 * take an rcu_read_lock.
2003 struct task_and_cgroup ent;
2005 /* @tsk either already exited or can't exit until the end */
2006 if (tsk->flags & PF_EXITING)
2009 /* as per above, nr_threads may decrease, but not increase. */
2010 BUG_ON(i >= group_size);
2012 ent.cgrp = task_cgroup_from_root(tsk, root);
2013 /* nothing to do if this task is already in the cgroup */
2014 if (ent.cgrp == cgrp)
2017 * saying GFP_ATOMIC has no effect here because we did prealloc
2018 * earlier, but it's good form to communicate our expectations.
2020 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
2021 BUG_ON(retval != 0);
2026 } while_each_thread(leader, tsk);
2028 /* remember the number of threads in the array for later. */
2030 tset.tc_array = group;
2031 tset.tc_array_len = group_size;
2033 /* methods shouldn't be called if no task is actually migrating */
2036 goto out_free_group_list;
2039 * step 1: check that we can legitimately attach to the cgroup.
2041 for_each_root_subsys(root, ss) {
2042 if (ss->can_attach) {
2043 retval = ss->can_attach(cgrp, &tset);
2046 goto out_cancel_attach;
2052 * step 2: make sure css_sets exist for all threads to be migrated.
2053 * we use find_css_set, which allocates a new one if necessary.
2055 for (i = 0; i < group_size; i++) {
2056 struct css_set *old_cset;
2058 tc = flex_array_get(group, i);
2059 old_cset = task_css_set(tc->task);
2060 tc->cg = find_css_set(old_cset, cgrp);
2063 goto out_put_css_set_refs;
2068 * step 3: now that we're guaranteed success wrt the css_sets,
2069 * proceed to move all tasks to the new cgroup. There are no
2070 * failure cases after here, so this is the commit point.
2072 for (i = 0; i < group_size; i++) {
2073 tc = flex_array_get(group, i);
2074 cgroup_task_migrate(tc->cgrp, tc->task, tc->cg);
2076 /* nothing is sensitive to fork() after this point. */
2079 * step 4: do subsystem attach callbacks.
2081 for_each_root_subsys(root, ss) {
2083 ss->attach(cgrp, &tset);
2087 * step 5: success! and cleanup
2090 out_put_css_set_refs:
2092 for (i = 0; i < group_size; i++) {
2093 tc = flex_array_get(group, i);
2096 put_css_set(tc->cg);
2101 for_each_root_subsys(root, ss) {
2102 if (ss == failed_ss)
2104 if (ss->cancel_attach)
2105 ss->cancel_attach(cgrp, &tset);
2108 out_free_group_list:
2109 flex_array_free(group);
2114 * Find the task_struct of the task to attach by vpid and pass it along to the
2115 * function to attach either it or all tasks in its threadgroup. Will lock
2116 * cgroup_mutex and threadgroup; may take task_lock of task.
2118 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2120 struct task_struct *tsk;
2121 const struct cred *cred = current_cred(), *tcred;
2124 if (!cgroup_lock_live_group(cgrp))
2130 tsk = find_task_by_vpid(pid);
2134 goto out_unlock_cgroup;
2137 * even if we're attaching all tasks in the thread group, we
2138 * only need to check permissions on one of them.
2140 tcred = __task_cred(tsk);
2141 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2142 !uid_eq(cred->euid, tcred->uid) &&
2143 !uid_eq(cred->euid, tcred->suid)) {
2146 goto out_unlock_cgroup;
2152 tsk = tsk->group_leader;
2155 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2156 * trapped in a cpuset, or RT worker may be born in a cgroup
2157 * with no rt_runtime allocated. Just say no.
2159 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2162 goto out_unlock_cgroup;
2165 get_task_struct(tsk);
2168 threadgroup_lock(tsk);
2170 if (!thread_group_leader(tsk)) {
2172 * a race with de_thread from another thread's exec()
2173 * may strip us of our leadership, if this happens,
2174 * there is no choice but to throw this task away and
2175 * try again; this is
2176 * "double-double-toil-and-trouble-check locking".
2178 threadgroup_unlock(tsk);
2179 put_task_struct(tsk);
2180 goto retry_find_task;
2184 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2186 threadgroup_unlock(tsk);
2188 put_task_struct(tsk);
2190 mutex_unlock(&cgroup_mutex);
2195 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2196 * @from: attach to all cgroups of a given task
2197 * @tsk: the task to be attached
2199 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2201 struct cgroupfs_root *root;
2204 mutex_lock(&cgroup_mutex);
2205 for_each_active_root(root) {
2206 struct cgroup *from_cg = task_cgroup_from_root(from, root);
2208 retval = cgroup_attach_task(from_cg, tsk, false);
2212 mutex_unlock(&cgroup_mutex);
2216 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2218 static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
2220 return attach_task_by_pid(cgrp, pid, false);
2223 static int cgroup_procs_write(struct cgroup *cgrp, struct cftype *cft, u64 tgid)
2225 return attach_task_by_pid(cgrp, tgid, true);
2228 static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
2231 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2232 if (strlen(buffer) >= PATH_MAX)
2234 if (!cgroup_lock_live_group(cgrp))
2236 mutex_lock(&cgroup_root_mutex);
2237 strcpy(cgrp->root->release_agent_path, buffer);
2238 mutex_unlock(&cgroup_root_mutex);
2239 mutex_unlock(&cgroup_mutex);
2243 static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
2244 struct seq_file *seq)
2246 if (!cgroup_lock_live_group(cgrp))
2248 seq_puts(seq, cgrp->root->release_agent_path);
2249 seq_putc(seq, '\n');
2250 mutex_unlock(&cgroup_mutex);
2254 static int cgroup_sane_behavior_show(struct cgroup *cgrp, struct cftype *cft,
2255 struct seq_file *seq)
2257 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2261 /* A buffer size big enough for numbers or short strings */
2262 #define CGROUP_LOCAL_BUFFER_SIZE 64
2264 static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
2266 const char __user *userbuf,
2267 size_t nbytes, loff_t *unused_ppos)
2269 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
2275 if (nbytes >= sizeof(buffer))
2277 if (copy_from_user(buffer, userbuf, nbytes))
2280 buffer[nbytes] = 0; /* nul-terminate */
2281 if (cft->write_u64) {
2282 u64 val = simple_strtoull(strstrip(buffer), &end, 0);
2285 retval = cft->write_u64(cgrp, cft, val);
2287 s64 val = simple_strtoll(strstrip(buffer), &end, 0);
2290 retval = cft->write_s64(cgrp, cft, val);
2297 static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
2299 const char __user *userbuf,
2300 size_t nbytes, loff_t *unused_ppos)
2302 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
2304 size_t max_bytes = cft->max_write_len;
2305 char *buffer = local_buffer;
2308 max_bytes = sizeof(local_buffer) - 1;
2309 if (nbytes >= max_bytes)
2311 /* Allocate a dynamic buffer if we need one */
2312 if (nbytes >= sizeof(local_buffer)) {
2313 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
2317 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
2322 buffer[nbytes] = 0; /* nul-terminate */
2323 retval = cft->write_string(cgrp, cft, strstrip(buffer));
2327 if (buffer != local_buffer)
2332 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
2333 size_t nbytes, loff_t *ppos)
2335 struct cftype *cft = __d_cft(file->f_dentry);
2336 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2338 if (cgroup_is_dead(cgrp))
2341 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
2342 if (cft->write_u64 || cft->write_s64)
2343 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
2344 if (cft->write_string)
2345 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
2347 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
2348 return ret ? ret : nbytes;
2353 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
2355 char __user *buf, size_t nbytes,
2358 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2359 u64 val = cft->read_u64(cgrp, cft);
2360 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
2362 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2365 static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
2367 char __user *buf, size_t nbytes,
2370 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2371 s64 val = cft->read_s64(cgrp, cft);
2372 int len = sprintf(tmp, "%lld\n", (long long) val);
2374 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2377 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2378 size_t nbytes, loff_t *ppos)
2380 struct cftype *cft = __d_cft(file->f_dentry);
2381 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2383 if (cgroup_is_dead(cgrp))
2387 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2389 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2391 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2396 * seqfile ops/methods for returning structured data. Currently just
2397 * supports string->u64 maps, but can be extended in future.
2400 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2402 struct seq_file *sf = cb->state;
2403 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2406 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2408 struct cfent *cfe = m->private;
2409 struct cftype *cft = cfe->type;
2410 struct cgroup *cgrp = __d_cgrp(cfe->dentry->d_parent);
2412 if (cft->read_map) {
2413 struct cgroup_map_cb cb = {
2414 .fill = cgroup_map_add,
2417 return cft->read_map(cgrp, cft, &cb);
2419 return cft->read_seq_string(cgrp, cft, m);
2422 static const struct file_operations cgroup_seqfile_operations = {
2424 .write = cgroup_file_write,
2425 .llseek = seq_lseek,
2426 .release = single_release,
2429 static int cgroup_file_open(struct inode *inode, struct file *file)
2435 err = generic_file_open(inode, file);
2438 cfe = __d_cfe(file->f_dentry);
2441 if (cft->read_map || cft->read_seq_string) {
2442 file->f_op = &cgroup_seqfile_operations;
2443 err = single_open(file, cgroup_seqfile_show, cfe);
2444 } else if (cft->open) {
2445 err = cft->open(inode, file);
2451 static int cgroup_file_release(struct inode *inode, struct file *file)
2453 struct cftype *cft = __d_cft(file->f_dentry);
2455 return cft->release(inode, file);
2460 * cgroup_rename - Only allow simple rename of directories in place.
2462 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2463 struct inode *new_dir, struct dentry *new_dentry)
2466 struct cgroup_name *name, *old_name;
2467 struct cgroup *cgrp;
2470 * It's convinient to use parent dir's i_mutex to protected
2473 lockdep_assert_held(&old_dir->i_mutex);
2475 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2477 if (new_dentry->d_inode)
2479 if (old_dir != new_dir)
2482 cgrp = __d_cgrp(old_dentry);
2485 * This isn't a proper migration and its usefulness is very
2486 * limited. Disallow if sane_behavior.
2488 if (cgroup_sane_behavior(cgrp))
2491 name = cgroup_alloc_name(new_dentry);
2495 ret = simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2501 old_name = rcu_dereference_protected(cgrp->name, true);
2502 rcu_assign_pointer(cgrp->name, name);
2504 kfree_rcu(old_name, rcu_head);
2508 static struct simple_xattrs *__d_xattrs(struct dentry *dentry)
2510 if (S_ISDIR(dentry->d_inode->i_mode))
2511 return &__d_cgrp(dentry)->xattrs;
2513 return &__d_cfe(dentry)->xattrs;
2516 static inline int xattr_enabled(struct dentry *dentry)
2518 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
2519 return root->flags & CGRP_ROOT_XATTR;
2522 static bool is_valid_xattr(const char *name)
2524 if (!strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
2525 !strncmp(name, XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN))
2530 static int cgroup_setxattr(struct dentry *dentry, const char *name,
2531 const void *val, size_t size, int flags)
2533 if (!xattr_enabled(dentry))
2535 if (!is_valid_xattr(name))
2537 return simple_xattr_set(__d_xattrs(dentry), name, val, size, flags);
2540 static int cgroup_removexattr(struct dentry *dentry, const char *name)
2542 if (!xattr_enabled(dentry))
2544 if (!is_valid_xattr(name))
2546 return simple_xattr_remove(__d_xattrs(dentry), name);
2549 static ssize_t cgroup_getxattr(struct dentry *dentry, const char *name,
2550 void *buf, size_t size)
2552 if (!xattr_enabled(dentry))
2554 if (!is_valid_xattr(name))
2556 return simple_xattr_get(__d_xattrs(dentry), name, buf, size);
2559 static ssize_t cgroup_listxattr(struct dentry *dentry, char *buf, size_t size)
2561 if (!xattr_enabled(dentry))
2563 return simple_xattr_list(__d_xattrs(dentry), buf, size);
2566 static const struct file_operations cgroup_file_operations = {
2567 .read = cgroup_file_read,
2568 .write = cgroup_file_write,
2569 .llseek = generic_file_llseek,
2570 .open = cgroup_file_open,
2571 .release = cgroup_file_release,
2574 static const struct inode_operations cgroup_file_inode_operations = {
2575 .setxattr = cgroup_setxattr,
2576 .getxattr = cgroup_getxattr,
2577 .listxattr = cgroup_listxattr,
2578 .removexattr = cgroup_removexattr,
2581 static const struct inode_operations cgroup_dir_inode_operations = {
2582 .lookup = cgroup_lookup,
2583 .mkdir = cgroup_mkdir,
2584 .rmdir = cgroup_rmdir,
2585 .rename = cgroup_rename,
2586 .setxattr = cgroup_setxattr,
2587 .getxattr = cgroup_getxattr,
2588 .listxattr = cgroup_listxattr,
2589 .removexattr = cgroup_removexattr,
2592 static struct dentry *cgroup_lookup(struct inode *dir, struct dentry *dentry, unsigned int flags)
2594 if (dentry->d_name.len > NAME_MAX)
2595 return ERR_PTR(-ENAMETOOLONG);
2596 d_add(dentry, NULL);
2601 * Check if a file is a control file
2603 static inline struct cftype *__file_cft(struct file *file)
2605 if (file_inode(file)->i_fop != &cgroup_file_operations)
2606 return ERR_PTR(-EINVAL);
2607 return __d_cft(file->f_dentry);
2610 static int cgroup_create_file(struct dentry *dentry, umode_t mode,
2611 struct super_block *sb)
2613 struct inode *inode;
2617 if (dentry->d_inode)
2620 inode = cgroup_new_inode(mode, sb);
2624 if (S_ISDIR(mode)) {
2625 inode->i_op = &cgroup_dir_inode_operations;
2626 inode->i_fop = &simple_dir_operations;
2628 /* start off with i_nlink == 2 (for "." entry) */
2630 inc_nlink(dentry->d_parent->d_inode);
2633 * Control reaches here with cgroup_mutex held.
2634 * @inode->i_mutex should nest outside cgroup_mutex but we
2635 * want to populate it immediately without releasing
2636 * cgroup_mutex. As @inode isn't visible to anyone else
2637 * yet, trylock will always succeed without affecting
2640 WARN_ON_ONCE(!mutex_trylock(&inode->i_mutex));
2641 } else if (S_ISREG(mode)) {
2643 inode->i_fop = &cgroup_file_operations;
2644 inode->i_op = &cgroup_file_inode_operations;
2646 d_instantiate(dentry, inode);
2647 dget(dentry); /* Extra count - pin the dentry in core */
2652 * cgroup_file_mode - deduce file mode of a control file
2653 * @cft: the control file in question
2655 * returns cft->mode if ->mode is not 0
2656 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2657 * returns S_IRUGO if it has only a read handler
2658 * returns S_IWUSR if it has only a write hander
2660 static umode_t cgroup_file_mode(const struct cftype *cft)
2667 if (cft->read || cft->read_u64 || cft->read_s64 ||
2668 cft->read_map || cft->read_seq_string)
2671 if (cft->write || cft->write_u64 || cft->write_s64 ||
2672 cft->write_string || cft->trigger)
2678 static int cgroup_add_file(struct cgroup *cgrp, struct cgroup_subsys *subsys,
2681 struct dentry *dir = cgrp->dentry;
2682 struct cgroup *parent = __d_cgrp(dir);
2683 struct dentry *dentry;
2687 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2689 if (subsys && !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
2690 strcpy(name, subsys->name);
2693 strcat(name, cft->name);
2695 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2697 cfe = kzalloc(sizeof(*cfe), GFP_KERNEL);
2701 dentry = lookup_one_len(name, dir, strlen(name));
2702 if (IS_ERR(dentry)) {
2703 error = PTR_ERR(dentry);
2707 cfe->type = (void *)cft;
2708 cfe->dentry = dentry;
2709 dentry->d_fsdata = cfe;
2710 simple_xattrs_init(&cfe->xattrs);
2712 mode = cgroup_file_mode(cft);
2713 error = cgroup_create_file(dentry, mode | S_IFREG, cgrp->root->sb);
2715 list_add_tail(&cfe->node, &parent->files);
2725 * cgroup_addrm_files - add or remove files to a cgroup directory
2726 * @cgrp: the target cgroup
2727 * @subsys: the subsystem of files to be added
2728 * @cfts: array of cftypes to be added
2729 * @is_add: whether to add or remove
2731 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2732 * All @cfts should belong to @subsys. For removals, this function never
2733 * fails. If addition fails, this function doesn't remove files already
2734 * added. The caller is responsible for cleaning up.
2736 static int cgroup_addrm_files(struct cgroup *cgrp, struct cgroup_subsys *subsys,
2737 struct cftype cfts[], bool is_add)
2742 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
2743 lockdep_assert_held(&cgroup_mutex);
2745 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2746 /* does cft->flags tell us to skip this file on @cgrp? */
2747 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2749 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2751 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2755 ret = cgroup_add_file(cgrp, subsys, cft);
2757 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2762 cgroup_rm_file(cgrp, cft);
2768 static void cgroup_cfts_prepare(void)
2769 __acquires(&cgroup_mutex)
2772 * Thanks to the entanglement with vfs inode locking, we can't walk
2773 * the existing cgroups under cgroup_mutex and create files.
2774 * Instead, we use cgroup_for_each_descendant_pre() and drop RCU
2775 * read lock before calling cgroup_addrm_files().
2777 mutex_lock(&cgroup_mutex);
2780 static int cgroup_cfts_commit(struct cgroup_subsys *ss,
2781 struct cftype *cfts, bool is_add)
2782 __releases(&cgroup_mutex)
2785 struct cgroup *cgrp, *root = &ss->root->top_cgroup;
2786 struct super_block *sb = ss->root->sb;
2787 struct dentry *prev = NULL;
2788 struct inode *inode;
2792 /* %NULL @cfts indicates abort and don't bother if @ss isn't attached */
2793 if (!cfts || ss->root == &cgroup_dummy_root ||
2794 !atomic_inc_not_zero(&sb->s_active)) {
2795 mutex_unlock(&cgroup_mutex);
2800 * All cgroups which are created after we drop cgroup_mutex will
2801 * have the updated set of files, so we only need to update the
2802 * cgroups created before the current @cgroup_serial_nr_next.
2804 update_before = cgroup_serial_nr_next;
2806 mutex_unlock(&cgroup_mutex);
2808 /* @root always needs to be updated */
2809 inode = root->dentry->d_inode;
2810 mutex_lock(&inode->i_mutex);
2811 mutex_lock(&cgroup_mutex);
2812 ret = cgroup_addrm_files(root, ss, cfts, is_add);
2813 mutex_unlock(&cgroup_mutex);
2814 mutex_unlock(&inode->i_mutex);
2819 /* add/rm files for all cgroups created before */
2821 cgroup_for_each_descendant_pre(cgrp, root) {
2822 if (cgroup_is_dead(cgrp))
2825 inode = cgrp->dentry->d_inode;
2830 prev = cgrp->dentry;
2832 mutex_lock(&inode->i_mutex);
2833 mutex_lock(&cgroup_mutex);
2834 if (cgrp->serial_nr < update_before && !cgroup_is_dead(cgrp))
2835 ret = cgroup_addrm_files(cgrp, ss, cfts, is_add);
2836 mutex_unlock(&cgroup_mutex);
2837 mutex_unlock(&inode->i_mutex);
2846 deactivate_super(sb);
2851 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2852 * @ss: target cgroup subsystem
2853 * @cfts: zero-length name terminated array of cftypes
2855 * Register @cfts to @ss. Files described by @cfts are created for all
2856 * existing cgroups to which @ss is attached and all future cgroups will
2857 * have them too. This function can be called anytime whether @ss is
2860 * Returns 0 on successful registration, -errno on failure. Note that this
2861 * function currently returns 0 as long as @cfts registration is successful
2862 * even if some file creation attempts on existing cgroups fail.
2864 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2866 struct cftype_set *set;
2869 set = kzalloc(sizeof(*set), GFP_KERNEL);
2873 cgroup_cfts_prepare();
2875 list_add_tail(&set->node, &ss->cftsets);
2876 ret = cgroup_cfts_commit(ss, cfts, true);
2878 cgroup_rm_cftypes(ss, cfts);
2881 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2884 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2885 * @ss: target cgroup subsystem
2886 * @cfts: zero-length name terminated array of cftypes
2888 * Unregister @cfts from @ss. Files described by @cfts are removed from
2889 * all existing cgroups to which @ss is attached and all future cgroups
2890 * won't have them either. This function can be called anytime whether @ss
2891 * is attached or not.
2893 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2894 * registered with @ss.
2896 int cgroup_rm_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2898 struct cftype_set *set;
2900 cgroup_cfts_prepare();
2902 list_for_each_entry(set, &ss->cftsets, node) {
2903 if (set->cfts == cfts) {
2904 list_del(&set->node);
2906 cgroup_cfts_commit(ss, cfts, false);
2911 cgroup_cfts_commit(ss, NULL, false);
2916 * cgroup_task_count - count the number of tasks in a cgroup.
2917 * @cgrp: the cgroup in question
2919 * Return the number of tasks in the cgroup.
2921 int cgroup_task_count(const struct cgroup *cgrp)
2924 struct cgrp_cset_link *link;
2926 read_lock(&css_set_lock);
2927 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2928 count += atomic_read(&link->cset->refcount);
2929 read_unlock(&css_set_lock);
2934 * Advance a list_head iterator. The iterator should be positioned at
2935 * the start of a css_set
2937 static void cgroup_advance_iter(struct cgroup *cgrp, struct cgroup_iter *it)
2939 struct list_head *l = it->cset_link;
2940 struct cgrp_cset_link *link;
2941 struct css_set *cset;
2943 /* Advance to the next non-empty css_set */
2946 if (l == &cgrp->cset_links) {
2947 it->cset_link = NULL;
2950 link = list_entry(l, struct cgrp_cset_link, cset_link);
2952 } while (list_empty(&cset->tasks));
2954 it->task = cset->tasks.next;
2958 * To reduce the fork() overhead for systems that are not actually
2959 * using their cgroups capability, we don't maintain the lists running
2960 * through each css_set to its tasks until we see the list actually
2961 * used - in other words after the first call to cgroup_iter_start().
2963 static void cgroup_enable_task_cg_lists(void)
2965 struct task_struct *p, *g;
2966 write_lock(&css_set_lock);
2967 use_task_css_set_links = 1;
2969 * We need tasklist_lock because RCU is not safe against
2970 * while_each_thread(). Besides, a forking task that has passed
2971 * cgroup_post_fork() without seeing use_task_css_set_links = 1
2972 * is not guaranteed to have its child immediately visible in the
2973 * tasklist if we walk through it with RCU.
2975 read_lock(&tasklist_lock);
2976 do_each_thread(g, p) {
2979 * We should check if the process is exiting, otherwise
2980 * it will race with cgroup_exit() in that the list
2981 * entry won't be deleted though the process has exited.
2983 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2984 list_add(&p->cg_list, &task_css_set(p)->tasks);
2986 } while_each_thread(g, p);
2987 read_unlock(&tasklist_lock);
2988 write_unlock(&css_set_lock);
2992 * cgroup_next_sibling - find the next sibling of a given cgroup
2993 * @pos: the current cgroup
2995 * This function returns the next sibling of @pos and should be called
2996 * under RCU read lock. The only requirement is that @pos is accessible.
2997 * The next sibling is guaranteed to be returned regardless of @pos's
3000 struct cgroup *cgroup_next_sibling(struct cgroup *pos)
3002 struct cgroup *next;
3004 WARN_ON_ONCE(!rcu_read_lock_held());
3007 * @pos could already have been removed. Once a cgroup is removed,
3008 * its ->sibling.next is no longer updated when its next sibling
3009 * changes. As CGRP_DEAD assertion is serialized and happens
3010 * before the cgroup is taken off the ->sibling list, if we see it
3011 * unasserted, it's guaranteed that the next sibling hasn't
3012 * finished its grace period even if it's already removed, and thus
3013 * safe to dereference from this RCU critical section. If
3014 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
3015 * to be visible as %true here.
3017 if (likely(!cgroup_is_dead(pos))) {
3018 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
3019 if (&next->sibling != &pos->parent->children)
3025 * Can't dereference the next pointer. Each cgroup is given a
3026 * monotonically increasing unique serial number and always
3027 * appended to the sibling list, so the next one can be found by
3028 * walking the parent's children until we see a cgroup with higher
3029 * serial number than @pos's.
3031 * While this path can be slow, it's taken only when either the
3032 * current cgroup is removed or iteration and removal race.
3034 list_for_each_entry_rcu(next, &pos->parent->children, sibling)
3035 if (next->serial_nr > pos->serial_nr)
3039 EXPORT_SYMBOL_GPL(cgroup_next_sibling);
3042 * cgroup_next_descendant_pre - find the next descendant for pre-order walk
3043 * @pos: the current position (%NULL to initiate traversal)
3044 * @cgroup: cgroup whose descendants to walk
3046 * To be used by cgroup_for_each_descendant_pre(). Find the next
3047 * descendant to visit for pre-order traversal of @cgroup's descendants.
3049 * While this function requires RCU read locking, it doesn't require the
3050 * whole traversal to be contained in a single RCU critical section. This
3051 * function will return the correct next descendant as long as both @pos
3052 * and @cgroup are accessible and @pos is a descendant of @cgroup.
3054 struct cgroup *cgroup_next_descendant_pre(struct cgroup *pos,
3055 struct cgroup *cgroup)
3057 struct cgroup *next;
3059 WARN_ON_ONCE(!rcu_read_lock_held());
3061 /* if first iteration, pretend we just visited @cgroup */
3065 /* visit the first child if exists */
3066 next = list_first_or_null_rcu(&pos->children, struct cgroup, sibling);
3070 /* no child, visit my or the closest ancestor's next sibling */
3071 while (pos != cgroup) {
3072 next = cgroup_next_sibling(pos);
3080 EXPORT_SYMBOL_GPL(cgroup_next_descendant_pre);
3083 * cgroup_rightmost_descendant - return the rightmost descendant of a cgroup
3084 * @pos: cgroup of interest
3086 * Return the rightmost descendant of @pos. If there's no descendant,
3087 * @pos is returned. This can be used during pre-order traversal to skip
3090 * While this function requires RCU read locking, it doesn't require the
3091 * whole traversal to be contained in a single RCU critical section. This
3092 * function will return the correct rightmost descendant as long as @pos is
3095 struct cgroup *cgroup_rightmost_descendant(struct cgroup *pos)
3097 struct cgroup *last, *tmp;
3099 WARN_ON_ONCE(!rcu_read_lock_held());
3103 /* ->prev isn't RCU safe, walk ->next till the end */
3105 list_for_each_entry_rcu(tmp, &last->children, sibling)
3111 EXPORT_SYMBOL_GPL(cgroup_rightmost_descendant);
3113 static struct cgroup *cgroup_leftmost_descendant(struct cgroup *pos)
3115 struct cgroup *last;
3119 pos = list_first_or_null_rcu(&pos->children, struct cgroup,
3127 * cgroup_next_descendant_post - find the next descendant for post-order walk
3128 * @pos: the current position (%NULL to initiate traversal)
3129 * @cgroup: cgroup whose descendants to walk
3131 * To be used by cgroup_for_each_descendant_post(). Find the next
3132 * descendant to visit for post-order traversal of @cgroup's descendants.
3134 * While this function requires RCU read locking, it doesn't require the
3135 * whole traversal to be contained in a single RCU critical section. This
3136 * function will return the correct next descendant as long as both @pos
3137 * and @cgroup are accessible and @pos is a descendant of @cgroup.
3139 struct cgroup *cgroup_next_descendant_post(struct cgroup *pos,
3140 struct cgroup *cgroup)
3142 struct cgroup *next;
3144 WARN_ON_ONCE(!rcu_read_lock_held());
3146 /* if first iteration, visit the leftmost descendant */
3148 next = cgroup_leftmost_descendant(cgroup);
3149 return next != cgroup ? next : NULL;
3152 /* if there's an unvisited sibling, visit its leftmost descendant */
3153 next = cgroup_next_sibling(pos);
3155 return cgroup_leftmost_descendant(next);
3157 /* no sibling left, visit parent */
3159 return next != cgroup ? next : NULL;
3161 EXPORT_SYMBOL_GPL(cgroup_next_descendant_post);
3163 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
3164 __acquires(css_set_lock)
3167 * The first time anyone tries to iterate across a cgroup,
3168 * we need to enable the list linking each css_set to its
3169 * tasks, and fix up all existing tasks.
3171 if (!use_task_css_set_links)
3172 cgroup_enable_task_cg_lists();
3174 read_lock(&css_set_lock);
3175 it->cset_link = &cgrp->cset_links;
3176 cgroup_advance_iter(cgrp, it);
3179 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
3180 struct cgroup_iter *it)
3182 struct task_struct *res;
3183 struct list_head *l = it->task;
3184 struct cgrp_cset_link *link;
3186 /* If the iterator cg is NULL, we have no tasks */
3189 res = list_entry(l, struct task_struct, cg_list);
3190 /* Advance iterator to find next entry */
3192 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
3193 if (l == &link->cset->tasks) {
3194 /* We reached the end of this task list - move on to
3195 * the next cg_cgroup_link */
3196 cgroup_advance_iter(cgrp, it);
3203 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
3204 __releases(css_set_lock)
3206 read_unlock(&css_set_lock);
3209 static inline int started_after_time(struct task_struct *t1,
3210 struct timespec *time,
3211 struct task_struct *t2)
3213 int start_diff = timespec_compare(&t1->start_time, time);
3214 if (start_diff > 0) {
3216 } else if (start_diff < 0) {
3220 * Arbitrarily, if two processes started at the same
3221 * time, we'll say that the lower pointer value
3222 * started first. Note that t2 may have exited by now
3223 * so this may not be a valid pointer any longer, but
3224 * that's fine - it still serves to distinguish
3225 * between two tasks started (effectively) simultaneously.
3232 * This function is a callback from heap_insert() and is used to order
3234 * In this case we order the heap in descending task start time.
3236 static inline int started_after(void *p1, void *p2)
3238 struct task_struct *t1 = p1;
3239 struct task_struct *t2 = p2;
3240 return started_after_time(t1, &t2->start_time, t2);
3244 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
3245 * @scan: struct cgroup_scanner containing arguments for the scan
3247 * Arguments include pointers to callback functions test_task() and
3249 * Iterate through all the tasks in a cgroup, calling test_task() for each,
3250 * and if it returns true, call process_task() for it also.
3251 * The test_task pointer may be NULL, meaning always true (select all tasks).
3252 * Effectively duplicates cgroup_iter_{start,next,end}()
3253 * but does not lock css_set_lock for the call to process_task().
3254 * The struct cgroup_scanner may be embedded in any structure of the caller's
3256 * It is guaranteed that process_task() will act on every task that
3257 * is a member of the cgroup for the duration of this call. This
3258 * function may or may not call process_task() for tasks that exit
3259 * or move to a different cgroup during the call, or are forked or
3260 * move into the cgroup during the call.
3262 * Note that test_task() may be called with locks held, and may in some
3263 * situations be called multiple times for the same task, so it should
3265 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
3266 * pre-allocated and will be used for heap operations (and its "gt" member will
3267 * be overwritten), else a temporary heap will be used (allocation of which
3268 * may cause this function to fail).
3270 int cgroup_scan_tasks(struct cgroup_scanner *scan)
3273 struct cgroup_iter it;
3274 struct task_struct *p, *dropped;
3275 /* Never dereference latest_task, since it's not refcounted */
3276 struct task_struct *latest_task = NULL;
3277 struct ptr_heap tmp_heap;
3278 struct ptr_heap *heap;
3279 struct timespec latest_time = { 0, 0 };
3282 /* The caller supplied our heap and pre-allocated its memory */
3284 heap->gt = &started_after;
3286 /* We need to allocate our own heap memory */
3288 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
3290 /* cannot allocate the heap */
3296 * Scan tasks in the cgroup, using the scanner's "test_task" callback
3297 * to determine which are of interest, and using the scanner's
3298 * "process_task" callback to process any of them that need an update.
3299 * Since we don't want to hold any locks during the task updates,
3300 * gather tasks to be processed in a heap structure.
3301 * The heap is sorted by descending task start time.
3302 * If the statically-sized heap fills up, we overflow tasks that
3303 * started later, and in future iterations only consider tasks that
3304 * started after the latest task in the previous pass. This
3305 * guarantees forward progress and that we don't miss any tasks.
3308 cgroup_iter_start(scan->cg, &it);
3309 while ((p = cgroup_iter_next(scan->cg, &it))) {
3311 * Only affect tasks that qualify per the caller's callback,
3312 * if he provided one
3314 if (scan->test_task && !scan->test_task(p, scan))
3317 * Only process tasks that started after the last task
3320 if (!started_after_time(p, &latest_time, latest_task))
3322 dropped = heap_insert(heap, p);
3323 if (dropped == NULL) {
3325 * The new task was inserted; the heap wasn't
3329 } else if (dropped != p) {
3331 * The new task was inserted, and pushed out a
3335 put_task_struct(dropped);
3338 * Else the new task was newer than anything already in
3339 * the heap and wasn't inserted
3342 cgroup_iter_end(scan->cg, &it);
3345 for (i = 0; i < heap->size; i++) {
3346 struct task_struct *q = heap->ptrs[i];
3348 latest_time = q->start_time;
3351 /* Process the task per the caller's callback */
3352 scan->process_task(q, scan);
3356 * If we had to process any tasks at all, scan again
3357 * in case some of them were in the middle of forking
3358 * children that didn't get processed.
3359 * Not the most efficient way to do it, but it avoids
3360 * having to take callback_mutex in the fork path
3364 if (heap == &tmp_heap)
3365 heap_free(&tmp_heap);
3369 static void cgroup_transfer_one_task(struct task_struct *task,
3370 struct cgroup_scanner *scan)
3372 struct cgroup *new_cgroup = scan->data;
3374 mutex_lock(&cgroup_mutex);
3375 cgroup_attach_task(new_cgroup, task, false);
3376 mutex_unlock(&cgroup_mutex);
3380 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3381 * @to: cgroup to which the tasks will be moved
3382 * @from: cgroup in which the tasks currently reside
3384 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3386 struct cgroup_scanner scan;
3389 scan.test_task = NULL; /* select all tasks in cgroup */
3390 scan.process_task = cgroup_transfer_one_task;
3394 return cgroup_scan_tasks(&scan);
3398 * Stuff for reading the 'tasks'/'procs' files.
3400 * Reading this file can return large amounts of data if a cgroup has
3401 * *lots* of attached tasks. So it may need several calls to read(),
3402 * but we cannot guarantee that the information we produce is correct
3403 * unless we produce it entirely atomically.
3407 /* which pidlist file are we talking about? */
3408 enum cgroup_filetype {
3414 * A pidlist is a list of pids that virtually represents the contents of one
3415 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3416 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3419 struct cgroup_pidlist {
3421 * used to find which pidlist is wanted. doesn't change as long as
3422 * this particular list stays in the list.
3424 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3427 /* how many elements the above list has */
3429 /* how many files are using the current array */
3431 /* each of these stored in a list by its cgroup */
3432 struct list_head links;
3433 /* pointer to the cgroup we belong to, for list removal purposes */
3434 struct cgroup *owner;
3435 /* protects the other fields */
3436 struct rw_semaphore mutex;
3440 * The following two functions "fix" the issue where there are more pids
3441 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3442 * TODO: replace with a kernel-wide solution to this problem
3444 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3445 static void *pidlist_allocate(int count)
3447 if (PIDLIST_TOO_LARGE(count))
3448 return vmalloc(count * sizeof(pid_t));
3450 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3452 static void pidlist_free(void *p)
3454 if (is_vmalloc_addr(p))
3461 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3462 * Returns the number of unique elements.
3464 static int pidlist_uniq(pid_t *list, int length)
3469 * we presume the 0th element is unique, so i starts at 1. trivial
3470 * edge cases first; no work needs to be done for either
3472 if (length == 0 || length == 1)
3474 /* src and dest walk down the list; dest counts unique elements */
3475 for (src = 1; src < length; src++) {
3476 /* find next unique element */
3477 while (list[src] == list[src-1]) {
3482 /* dest always points to where the next unique element goes */
3483 list[dest] = list[src];
3490 static int cmppid(const void *a, const void *b)
3492 return *(pid_t *)a - *(pid_t *)b;
3496 * find the appropriate pidlist for our purpose (given procs vs tasks)
3497 * returns with the lock on that pidlist already held, and takes care
3498 * of the use count, or returns NULL with no locks held if we're out of
3501 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3502 enum cgroup_filetype type)
3504 struct cgroup_pidlist *l;
3505 /* don't need task_nsproxy() if we're looking at ourself */
3506 struct pid_namespace *ns = task_active_pid_ns(current);
3509 * We can't drop the pidlist_mutex before taking the l->mutex in case
3510 * the last ref-holder is trying to remove l from the list at the same
3511 * time. Holding the pidlist_mutex precludes somebody taking whichever
3512 * list we find out from under us - compare release_pid_array().
3514 mutex_lock(&cgrp->pidlist_mutex);
3515 list_for_each_entry(l, &cgrp->pidlists, links) {
3516 if (l->key.type == type && l->key.ns == ns) {
3517 /* make sure l doesn't vanish out from under us */
3518 down_write(&l->mutex);
3519 mutex_unlock(&cgrp->pidlist_mutex);
3523 /* entry not found; create a new one */
3524 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3526 mutex_unlock(&cgrp->pidlist_mutex);
3529 init_rwsem(&l->mutex);
3530 down_write(&l->mutex);
3532 l->key.ns = get_pid_ns(ns);
3534 list_add(&l->links, &cgrp->pidlists);
3535 mutex_unlock(&cgrp->pidlist_mutex);
3540 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3542 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3543 struct cgroup_pidlist **lp)
3547 int pid, n = 0; /* used for populating the array */
3548 struct cgroup_iter it;
3549 struct task_struct *tsk;
3550 struct cgroup_pidlist *l;
3553 * If cgroup gets more users after we read count, we won't have
3554 * enough space - tough. This race is indistinguishable to the
3555 * caller from the case that the additional cgroup users didn't
3556 * show up until sometime later on.
3558 length = cgroup_task_count(cgrp);
3559 array = pidlist_allocate(length);
3562 /* now, populate the array */
3563 cgroup_iter_start(cgrp, &it);
3564 while ((tsk = cgroup_iter_next(cgrp, &it))) {
3565 if (unlikely(n == length))
3567 /* get tgid or pid for procs or tasks file respectively */
3568 if (type == CGROUP_FILE_PROCS)
3569 pid = task_tgid_vnr(tsk);
3571 pid = task_pid_vnr(tsk);
3572 if (pid > 0) /* make sure to only use valid results */
3575 cgroup_iter_end(cgrp, &it);
3577 /* now sort & (if procs) strip out duplicates */
3578 sort(array, length, sizeof(pid_t), cmppid, NULL);
3579 if (type == CGROUP_FILE_PROCS)
3580 length = pidlist_uniq(array, length);
3581 l = cgroup_pidlist_find(cgrp, type);
3583 pidlist_free(array);
3586 /* store array, freeing old if necessary - lock already held */
3587 pidlist_free(l->list);
3591 up_write(&l->mutex);
3597 * cgroupstats_build - build and fill cgroupstats
3598 * @stats: cgroupstats to fill information into
3599 * @dentry: A dentry entry belonging to the cgroup for which stats have
3602 * Build and fill cgroupstats so that taskstats can export it to user
3605 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3608 struct cgroup *cgrp;
3609 struct cgroup_iter it;
3610 struct task_struct *tsk;
3613 * Validate dentry by checking the superblock operations,
3614 * and make sure it's a directory.
3616 if (dentry->d_sb->s_op != &cgroup_ops ||
3617 !S_ISDIR(dentry->d_inode->i_mode))
3621 cgrp = dentry->d_fsdata;
3623 cgroup_iter_start(cgrp, &it);
3624 while ((tsk = cgroup_iter_next(cgrp, &it))) {
3625 switch (tsk->state) {
3627 stats->nr_running++;
3629 case TASK_INTERRUPTIBLE:
3630 stats->nr_sleeping++;
3632 case TASK_UNINTERRUPTIBLE:
3633 stats->nr_uninterruptible++;
3636 stats->nr_stopped++;
3639 if (delayacct_is_task_waiting_on_io(tsk))
3640 stats->nr_io_wait++;
3644 cgroup_iter_end(cgrp, &it);
3652 * seq_file methods for the tasks/procs files. The seq_file position is the
3653 * next pid to display; the seq_file iterator is a pointer to the pid
3654 * in the cgroup->l->list array.
3657 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3660 * Initially we receive a position value that corresponds to
3661 * one more than the last pid shown (or 0 on the first call or
3662 * after a seek to the start). Use a binary-search to find the
3663 * next pid to display, if any
3665 struct cgroup_pidlist *l = s->private;
3666 int index = 0, pid = *pos;
3669 down_read(&l->mutex);
3671 int end = l->length;
3673 while (index < end) {
3674 int mid = (index + end) / 2;
3675 if (l->list[mid] == pid) {
3678 } else if (l->list[mid] <= pid)
3684 /* If we're off the end of the array, we're done */
3685 if (index >= l->length)
3687 /* Update the abstract position to be the actual pid that we found */
3688 iter = l->list + index;
3693 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3695 struct cgroup_pidlist *l = s->private;
3699 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3701 struct cgroup_pidlist *l = s->private;
3703 pid_t *end = l->list + l->length;
3705 * Advance to the next pid in the array. If this goes off the
3717 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3719 return seq_printf(s, "%d\n", *(int *)v);
3723 * seq_operations functions for iterating on pidlists through seq_file -
3724 * independent of whether it's tasks or procs
3726 static const struct seq_operations cgroup_pidlist_seq_operations = {
3727 .start = cgroup_pidlist_start,
3728 .stop = cgroup_pidlist_stop,
3729 .next = cgroup_pidlist_next,
3730 .show = cgroup_pidlist_show,
3733 static void cgroup_release_pid_array(struct cgroup_pidlist *l)
3736 * the case where we're the last user of this particular pidlist will
3737 * have us remove it from the cgroup's list, which entails taking the
3738 * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
3739 * pidlist_mutex, we have to take pidlist_mutex first.
3741 mutex_lock(&l->owner->pidlist_mutex);
3742 down_write(&l->mutex);
3743 BUG_ON(!l->use_count);
3744 if (!--l->use_count) {
3745 /* we're the last user if refcount is 0; remove and free */
3746 list_del(&l->links);
3747 mutex_unlock(&l->owner->pidlist_mutex);
3748 pidlist_free(l->list);
3749 put_pid_ns(l->key.ns);
3750 up_write(&l->mutex);
3754 mutex_unlock(&l->owner->pidlist_mutex);
3755 up_write(&l->mutex);
3758 static int cgroup_pidlist_release(struct inode *inode, struct file *file)
3760 struct cgroup_pidlist *l;
3761 if (!(file->f_mode & FMODE_READ))
3764 * the seq_file will only be initialized if the file was opened for
3765 * reading; hence we check if it's not null only in that case.
3767 l = ((struct seq_file *)file->private_data)->private;
3768 cgroup_release_pid_array(l);
3769 return seq_release(inode, file);
3772 static const struct file_operations cgroup_pidlist_operations = {
3774 .llseek = seq_lseek,
3775 .write = cgroup_file_write,
3776 .release = cgroup_pidlist_release,
3780 * The following functions handle opens on a file that displays a pidlist
3781 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
3784 /* helper function for the two below it */
3785 static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
3787 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
3788 struct cgroup_pidlist *l;
3791 /* Nothing to do for write-only files */
3792 if (!(file->f_mode & FMODE_READ))
3795 /* have the array populated */
3796 retval = pidlist_array_load(cgrp, type, &l);
3799 /* configure file information */
3800 file->f_op = &cgroup_pidlist_operations;
3802 retval = seq_open(file, &cgroup_pidlist_seq_operations);
3804 cgroup_release_pid_array(l);
3807 ((struct seq_file *)file->private_data)->private = l;
3810 static int cgroup_tasks_open(struct inode *unused, struct file *file)
3812 return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
3814 static int cgroup_procs_open(struct inode *unused, struct file *file)
3816 return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
3819 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
3822 return notify_on_release(cgrp);
3825 static int cgroup_write_notify_on_release(struct cgroup *cgrp,
3829 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
3831 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3833 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3838 * When dput() is called asynchronously, if umount has been done and
3839 * then deactivate_super() in cgroup_free_fn() kills the superblock,
3840 * there's a small window that vfs will see the root dentry with non-zero
3841 * refcnt and trigger BUG().
3843 * That's why we hold a reference before dput() and drop it right after.
3845 static void cgroup_dput(struct cgroup *cgrp)
3847 struct super_block *sb = cgrp->root->sb;
3849 atomic_inc(&sb->s_active);
3851 deactivate_super(sb);
3855 * Unregister event and free resources.
3857 * Gets called from workqueue.
3859 static void cgroup_event_remove(struct work_struct *work)
3861 struct cgroup_event *event = container_of(work, struct cgroup_event,
3863 struct cgroup *cgrp = event->cgrp;
3865 remove_wait_queue(event->wqh, &event->wait);
3867 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3869 /* Notify userspace the event is going away. */
3870 eventfd_signal(event->eventfd, 1);
3872 eventfd_ctx_put(event->eventfd);
3878 * Gets called on POLLHUP on eventfd when user closes it.
3880 * Called with wqh->lock held and interrupts disabled.
3882 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3883 int sync, void *key)
3885 struct cgroup_event *event = container_of(wait,
3886 struct cgroup_event, wait);
3887 struct cgroup *cgrp = event->cgrp;
3888 unsigned long flags = (unsigned long)key;
3890 if (flags & POLLHUP) {
3892 * If the event has been detached at cgroup removal, we
3893 * can simply return knowing the other side will cleanup
3896 * We can't race against event freeing since the other
3897 * side will require wqh->lock via remove_wait_queue(),
3900 spin_lock(&cgrp->event_list_lock);
3901 if (!list_empty(&event->list)) {
3902 list_del_init(&event->list);
3904 * We are in atomic context, but cgroup_event_remove()
3905 * may sleep, so we have to call it in workqueue.
3907 schedule_work(&event->remove);
3909 spin_unlock(&cgrp->event_list_lock);
3915 static void cgroup_event_ptable_queue_proc(struct file *file,
3916 wait_queue_head_t *wqh, poll_table *pt)
3918 struct cgroup_event *event = container_of(pt,
3919 struct cgroup_event, pt);
3922 add_wait_queue(wqh, &event->wait);
3926 * Parse input and register new cgroup event handler.
3928 * Input must be in format '<event_fd> <control_fd> <args>'.
3929 * Interpretation of args is defined by control file implementation.
3931 static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3934 struct cgroup_event *event = NULL;
3935 struct cgroup *cgrp_cfile;
3936 unsigned int efd, cfd;
3937 struct file *efile = NULL;
3938 struct file *cfile = NULL;
3942 efd = simple_strtoul(buffer, &endp, 10);
3947 cfd = simple_strtoul(buffer, &endp, 10);
3948 if ((*endp != ' ') && (*endp != '\0'))
3952 event = kzalloc(sizeof(*event), GFP_KERNEL);
3956 INIT_LIST_HEAD(&event->list);
3957 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
3958 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
3959 INIT_WORK(&event->remove, cgroup_event_remove);
3961 efile = eventfd_fget(efd);
3962 if (IS_ERR(efile)) {
3963 ret = PTR_ERR(efile);
3967 event->eventfd = eventfd_ctx_fileget(efile);
3968 if (IS_ERR(event->eventfd)) {
3969 ret = PTR_ERR(event->eventfd);
3979 /* the process need read permission on control file */
3980 /* AV: shouldn't we check that it's been opened for read instead? */
3981 ret = inode_permission(file_inode(cfile), MAY_READ);
3985 event->cft = __file_cft(cfile);
3986 if (IS_ERR(event->cft)) {
3987 ret = PTR_ERR(event->cft);
3992 * The file to be monitored must be in the same cgroup as
3993 * cgroup.event_control is.
3995 cgrp_cfile = __d_cgrp(cfile->f_dentry->d_parent);
3996 if (cgrp_cfile != cgrp) {
4001 if (!event->cft->register_event || !event->cft->unregister_event) {
4006 ret = event->cft->register_event(cgrp, event->cft,
4007 event->eventfd, buffer);
4011 efile->f_op->poll(efile, &event->pt);
4014 * Events should be removed after rmdir of cgroup directory, but before
4015 * destroying subsystem state objects. Let's take reference to cgroup
4016 * directory dentry to do that.
4020 spin_lock(&cgrp->event_list_lock);
4021 list_add(&event->list, &cgrp->event_list);
4022 spin_unlock(&cgrp->event_list_lock);
4033 if (event && event->eventfd && !IS_ERR(event->eventfd))
4034 eventfd_ctx_put(event->eventfd);
4036 if (!IS_ERR_OR_NULL(efile))
4044 static u64 cgroup_clone_children_read(struct cgroup *cgrp,
4047 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4050 static int cgroup_clone_children_write(struct cgroup *cgrp,
4055 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4057 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4061 static struct cftype cgroup_base_files[] = {
4063 .name = "cgroup.procs",
4064 .open = cgroup_procs_open,
4065 .write_u64 = cgroup_procs_write,
4066 .release = cgroup_pidlist_release,
4067 .mode = S_IRUGO | S_IWUSR,
4070 .name = "cgroup.event_control",
4071 .write_string = cgroup_write_event_control,
4075 .name = "cgroup.clone_children",
4076 .flags = CFTYPE_INSANE,
4077 .read_u64 = cgroup_clone_children_read,
4078 .write_u64 = cgroup_clone_children_write,
4081 .name = "cgroup.sane_behavior",
4082 .flags = CFTYPE_ONLY_ON_ROOT,
4083 .read_seq_string = cgroup_sane_behavior_show,
4087 * Historical crazy stuff. These don't have "cgroup." prefix and
4088 * don't exist if sane_behavior. If you're depending on these, be
4089 * prepared to be burned.
4093 .flags = CFTYPE_INSANE, /* use "procs" instead */
4094 .open = cgroup_tasks_open,
4095 .write_u64 = cgroup_tasks_write,
4096 .release = cgroup_pidlist_release,
4097 .mode = S_IRUGO | S_IWUSR,
4100 .name = "notify_on_release",
4101 .flags = CFTYPE_INSANE,
4102 .read_u64 = cgroup_read_notify_on_release,
4103 .write_u64 = cgroup_write_notify_on_release,
4106 .name = "release_agent",
4107 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
4108 .read_seq_string = cgroup_release_agent_show,
4109 .write_string = cgroup_release_agent_write,
4110 .max_write_len = PATH_MAX,
4116 * cgroup_populate_dir - create subsys files in a cgroup directory
4117 * @cgrp: target cgroup
4118 * @subsys_mask: mask of the subsystem ids whose files should be added
4120 * On failure, no file is added.
4122 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
4124 struct cgroup_subsys *ss;
4127 /* process cftsets of each subsystem */
4128 for_each_subsys(ss, i) {
4129 struct cftype_set *set;
4131 if (!test_bit(i, &subsys_mask))
4134 list_for_each_entry(set, &ss->cftsets, node) {
4135 ret = cgroup_addrm_files(cgrp, ss, set->cfts, true);
4141 /* This cgroup is ready now */
4142 for_each_root_subsys(cgrp->root, ss) {
4143 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
4144 struct css_id *id = rcu_dereference_protected(css->id, true);
4147 * Update id->css pointer and make this css visible from
4148 * CSS ID functions. This pointer will be dereferened
4149 * from RCU-read-side without locks.
4152 rcu_assign_pointer(id->css, css);
4157 cgroup_clear_dir(cgrp, subsys_mask);
4161 static void css_dput_fn(struct work_struct *work)
4163 struct cgroup_subsys_state *css =
4164 container_of(work, struct cgroup_subsys_state, dput_work);
4166 cgroup_dput(css->cgroup);
4169 static void css_release(struct percpu_ref *ref)
4171 struct cgroup_subsys_state *css =
4172 container_of(ref, struct cgroup_subsys_state, refcnt);
4174 schedule_work(&css->dput_work);
4177 static void init_cgroup_css(struct cgroup_subsys_state *css,
4178 struct cgroup_subsys *ss,
4179 struct cgroup *cgrp)
4184 if (cgrp == cgroup_dummy_top)
4185 css->flags |= CSS_ROOT;
4186 BUG_ON(cgrp->subsys[ss->subsys_id]);
4187 cgrp->subsys[ss->subsys_id] = css;
4190 * css holds an extra ref to @cgrp->dentry which is put on the last
4191 * css_put(). dput() requires process context, which css_put() may
4192 * be called without. @css->dput_work will be used to invoke
4193 * dput() asynchronously from css_put().
4195 INIT_WORK(&css->dput_work, css_dput_fn);
4198 /* invoke ->css_online() on a new CSS and mark it online if successful */
4199 static int online_css(struct cgroup_subsys *ss, struct cgroup *cgrp)
4203 lockdep_assert_held(&cgroup_mutex);
4206 ret = ss->css_online(cgrp);
4208 cgrp->subsys[ss->subsys_id]->flags |= CSS_ONLINE;
4212 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4213 static void offline_css(struct cgroup_subsys *ss, struct cgroup *cgrp)
4215 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
4217 lockdep_assert_held(&cgroup_mutex);
4219 if (!(css->flags & CSS_ONLINE))
4222 if (ss->css_offline)
4223 ss->css_offline(cgrp);
4225 cgrp->subsys[ss->subsys_id]->flags &= ~CSS_ONLINE;
4229 * cgroup_create - create a cgroup
4230 * @parent: cgroup that will be parent of the new cgroup
4231 * @dentry: dentry of the new cgroup
4232 * @mode: mode to set on new inode
4234 * Must be called with the mutex on the parent inode held
4236 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
4239 struct cgroup *cgrp;
4240 struct cgroup_name *name;
4241 struct cgroupfs_root *root = parent->root;
4243 struct cgroup_subsys *ss;
4244 struct super_block *sb = root->sb;
4246 /* allocate the cgroup and its ID, 0 is reserved for the root */
4247 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4251 name = cgroup_alloc_name(dentry);
4254 rcu_assign_pointer(cgrp->name, name);
4256 cgrp->id = ida_simple_get(&root->cgroup_ida, 1, 0, GFP_KERNEL);
4261 * Only live parents can have children. Note that the liveliness
4262 * check isn't strictly necessary because cgroup_mkdir() and
4263 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
4264 * anyway so that locking is contained inside cgroup proper and we
4265 * don't get nasty surprises if we ever grow another caller.
4267 if (!cgroup_lock_live_group(parent)) {
4272 /* Grab a reference on the superblock so the hierarchy doesn't
4273 * get deleted on unmount if there are child cgroups. This
4274 * can be done outside cgroup_mutex, since the sb can't
4275 * disappear while someone has an open control file on the
4277 atomic_inc(&sb->s_active);
4279 init_cgroup_housekeeping(cgrp);
4281 dentry->d_fsdata = cgrp;
4282 cgrp->dentry = dentry;
4284 cgrp->parent = parent;
4285 cgrp->root = parent->root;
4287 if (notify_on_release(parent))
4288 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4290 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4291 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4293 for_each_root_subsys(root, ss) {
4294 struct cgroup_subsys_state *css;
4296 css = ss->css_alloc(cgrp);
4302 err = percpu_ref_init(&css->refcnt, css_release);
4306 init_cgroup_css(css, ss, cgrp);
4309 err = alloc_css_id(ss, parent, cgrp);
4316 * Create directory. cgroup_create_file() returns with the new
4317 * directory locked on success so that it can be populated without
4318 * dropping cgroup_mutex.
4320 err = cgroup_create_file(dentry, S_IFDIR | mode, sb);
4323 lockdep_assert_held(&dentry->d_inode->i_mutex);
4325 cgrp->serial_nr = cgroup_serial_nr_next++;
4327 /* allocation complete, commit to creation */
4328 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
4329 root->number_of_cgroups++;
4331 /* each css holds a ref to the cgroup's dentry */
4332 for_each_root_subsys(root, ss)
4335 /* hold a ref to the parent's dentry */
4336 dget(parent->dentry);
4338 /* creation succeeded, notify subsystems */
4339 for_each_root_subsys(root, ss) {
4340 err = online_css(ss, cgrp);
4344 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4346 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",
4347 current->comm, current->pid, ss->name);
4348 if (!strcmp(ss->name, "memory"))
4349 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
4350 ss->warned_broken_hierarchy = true;
4354 err = cgroup_addrm_files(cgrp, NULL, cgroup_base_files, true);
4358 err = cgroup_populate_dir(cgrp, root->subsys_mask);
4362 mutex_unlock(&cgroup_mutex);
4363 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
4368 for_each_root_subsys(root, ss) {
4369 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
4372 percpu_ref_cancel_init(&css->refcnt);
4376 mutex_unlock(&cgroup_mutex);
4377 /* Release the reference count that we took on the superblock */
4378 deactivate_super(sb);
4380 ida_simple_remove(&root->cgroup_ida, cgrp->id);
4382 kfree(rcu_dereference_raw(cgrp->name));
4388 cgroup_destroy_locked(cgrp);
4389 mutex_unlock(&cgroup_mutex);
4390 mutex_unlock(&dentry->d_inode->i_mutex);
4394 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
4396 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
4398 /* the vfs holds inode->i_mutex already */
4399 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
4402 static void cgroup_css_killed(struct cgroup *cgrp)
4404 if (!atomic_dec_and_test(&cgrp->css_kill_cnt))
4407 /* percpu ref's of all css's are killed, kick off the next step */
4408 INIT_WORK(&cgrp->destroy_work, cgroup_offline_fn);
4409 schedule_work(&cgrp->destroy_work);
4412 static void css_ref_killed_fn(struct percpu_ref *ref)
4414 struct cgroup_subsys_state *css =
4415 container_of(ref, struct cgroup_subsys_state, refcnt);
4417 cgroup_css_killed(css->cgroup);
4421 * cgroup_destroy_locked - the first stage of cgroup destruction
4422 * @cgrp: cgroup to be destroyed
4424 * css's make use of percpu refcnts whose killing latency shouldn't be
4425 * exposed to userland and are RCU protected. Also, cgroup core needs to
4426 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
4427 * invoked. To satisfy all the requirements, destruction is implemented in
4428 * the following two steps.
4430 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4431 * userland visible parts and start killing the percpu refcnts of
4432 * css's. Set up so that the next stage will be kicked off once all
4433 * the percpu refcnts are confirmed to be killed.
4435 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4436 * rest of destruction. Once all cgroup references are gone, the
4437 * cgroup is RCU-freed.
4439 * This function implements s1. After this step, @cgrp is gone as far as
4440 * the userland is concerned and a new cgroup with the same name may be
4441 * created. As cgroup doesn't care about the names internally, this
4442 * doesn't cause any problem.
4444 static int cgroup_destroy_locked(struct cgroup *cgrp)
4445 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4447 struct dentry *d = cgrp->dentry;
4448 struct cgroup_event *event, *tmp;
4449 struct cgroup_subsys *ss;
4452 lockdep_assert_held(&d->d_inode->i_mutex);
4453 lockdep_assert_held(&cgroup_mutex);
4456 * css_set_lock synchronizes access to ->cset_links and prevents
4457 * @cgrp from being removed while __put_css_set() is in progress.
4459 read_lock(&css_set_lock);
4460 empty = list_empty(&cgrp->cset_links) && list_empty(&cgrp->children);
4461 read_unlock(&css_set_lock);
4466 * Block new css_tryget() by killing css refcnts. cgroup core
4467 * guarantees that, by the time ->css_offline() is invoked, no new
4468 * css reference will be given out via css_tryget(). We can't
4469 * simply call percpu_ref_kill() and proceed to offlining css's
4470 * because percpu_ref_kill() doesn't guarantee that the ref is seen
4471 * as killed on all CPUs on return.
4473 * Use percpu_ref_kill_and_confirm() to get notifications as each
4474 * css is confirmed to be seen as killed on all CPUs. The
4475 * notification callback keeps track of the number of css's to be
4476 * killed and schedules cgroup_offline_fn() to perform the rest of
4477 * destruction once the percpu refs of all css's are confirmed to
4480 atomic_set(&cgrp->css_kill_cnt, 1);
4481 for_each_root_subsys(cgrp->root, ss) {
4482 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
4485 * Killing would put the base ref, but we need to keep it
4486 * alive until after ->css_offline.
4488 percpu_ref_get(&css->refcnt);
4490 atomic_inc(&cgrp->css_kill_cnt);
4491 percpu_ref_kill_and_confirm(&css->refcnt, css_ref_killed_fn);
4493 cgroup_css_killed(cgrp);
4496 * Mark @cgrp dead. This prevents further task migration and child
4497 * creation by disabling cgroup_lock_live_group(). Note that
4498 * CGRP_DEAD assertion is depended upon by cgroup_next_sibling() to
4499 * resume iteration after dropping RCU read lock. See
4500 * cgroup_next_sibling() for details.
4502 set_bit(CGRP_DEAD, &cgrp->flags);
4504 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4505 raw_spin_lock(&release_list_lock);
4506 if (!list_empty(&cgrp->release_list))
4507 list_del_init(&cgrp->release_list);
4508 raw_spin_unlock(&release_list_lock);
4511 * Clear and remove @cgrp directory. The removal puts the base ref
4512 * but we aren't quite done with @cgrp yet, so hold onto it.
4514 cgroup_clear_dir(cgrp, cgrp->root->subsys_mask);
4515 cgroup_addrm_files(cgrp, NULL, cgroup_base_files, false);
4517 cgroup_d_remove_dir(d);
4520 * Unregister events and notify userspace.
4521 * Notify userspace about cgroup removing only after rmdir of cgroup
4522 * directory to avoid race between userspace and kernelspace.
4524 spin_lock(&cgrp->event_list_lock);
4525 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
4526 list_del_init(&event->list);
4527 schedule_work(&event->remove);
4529 spin_unlock(&cgrp->event_list_lock);
4535 * cgroup_offline_fn - the second step of cgroup destruction
4536 * @work: cgroup->destroy_free_work
4538 * This function is invoked from a work item for a cgroup which is being
4539 * destroyed after the percpu refcnts of all css's are guaranteed to be
4540 * seen as killed on all CPUs, and performs the rest of destruction. This
4541 * is the second step of destruction described in the comment above
4542 * cgroup_destroy_locked().
4544 static void cgroup_offline_fn(struct work_struct *work)
4546 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
4547 struct cgroup *parent = cgrp->parent;
4548 struct dentry *d = cgrp->dentry;
4549 struct cgroup_subsys *ss;
4551 mutex_lock(&cgroup_mutex);
4554 * css_tryget() is guaranteed to fail now. Tell subsystems to
4555 * initate destruction.
4557 for_each_root_subsys(cgrp->root, ss)
4558 offline_css(ss, cgrp);
4561 * Put the css refs from cgroup_destroy_locked(). Each css holds
4562 * an extra reference to the cgroup's dentry and cgroup removal
4563 * proceeds regardless of css refs. On the last put of each css,
4564 * whenever that may be, the extra dentry ref is put so that dentry
4565 * destruction happens only after all css's are released.
4567 for_each_root_subsys(cgrp->root, ss)
4568 css_put(cgrp->subsys[ss->subsys_id]);
4570 /* delete this cgroup from parent->children */
4571 list_del_rcu(&cgrp->sibling);
4575 set_bit(CGRP_RELEASABLE, &parent->flags);
4576 check_for_release(parent);
4578 mutex_unlock(&cgroup_mutex);
4581 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
4585 mutex_lock(&cgroup_mutex);
4586 ret = cgroup_destroy_locked(dentry->d_fsdata);
4587 mutex_unlock(&cgroup_mutex);
4592 static void __init_or_module cgroup_init_cftsets(struct cgroup_subsys *ss)
4594 INIT_LIST_HEAD(&ss->cftsets);
4597 * base_cftset is embedded in subsys itself, no need to worry about
4600 if (ss->base_cftypes) {
4601 ss->base_cftset.cfts = ss->base_cftypes;
4602 list_add_tail(&ss->base_cftset.node, &ss->cftsets);
4606 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4608 struct cgroup_subsys_state *css;
4610 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4612 mutex_lock(&cgroup_mutex);
4614 /* init base cftset */
4615 cgroup_init_cftsets(ss);
4617 /* Create the top cgroup state for this subsystem */
4618 list_add(&ss->sibling, &cgroup_dummy_root.subsys_list);
4619 ss->root = &cgroup_dummy_root;
4620 css = ss->css_alloc(cgroup_dummy_top);
4621 /* We don't handle early failures gracefully */
4622 BUG_ON(IS_ERR(css));
4623 init_cgroup_css(css, ss, cgroup_dummy_top);
4625 /* Update the init_css_set to contain a subsys
4626 * pointer to this state - since the subsystem is
4627 * newly registered, all tasks and hence the
4628 * init_css_set is in the subsystem's top cgroup. */
4629 init_css_set.subsys[ss->subsys_id] = css;
4631 need_forkexit_callback |= ss->fork || ss->exit;
4633 /* At system boot, before all subsystems have been
4634 * registered, no tasks have been forked, so we don't
4635 * need to invoke fork callbacks here. */
4636 BUG_ON(!list_empty(&init_task.tasks));
4638 BUG_ON(online_css(ss, cgroup_dummy_top));
4640 mutex_unlock(&cgroup_mutex);
4642 /* this function shouldn't be used with modular subsystems, since they
4643 * need to register a subsys_id, among other things */
4648 * cgroup_load_subsys: load and register a modular subsystem at runtime
4649 * @ss: the subsystem to load
4651 * This function should be called in a modular subsystem's initcall. If the
4652 * subsystem is built as a module, it will be assigned a new subsys_id and set
4653 * up for use. If the subsystem is built-in anyway, work is delegated to the
4654 * simpler cgroup_init_subsys.
4656 int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
4658 struct cgroup_subsys_state *css;
4660 struct hlist_node *tmp;
4661 struct css_set *cset;
4664 /* check name and function validity */
4665 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
4666 ss->css_alloc == NULL || ss->css_free == NULL)
4670 * we don't support callbacks in modular subsystems. this check is
4671 * before the ss->module check for consistency; a subsystem that could
4672 * be a module should still have no callbacks even if the user isn't
4673 * compiling it as one.
4675 if (ss->fork || ss->exit)
4679 * an optionally modular subsystem is built-in: we want to do nothing,
4680 * since cgroup_init_subsys will have already taken care of it.
4682 if (ss->module == NULL) {
4683 /* a sanity check */
4684 BUG_ON(cgroup_subsys[ss->subsys_id] != ss);
4688 /* init base cftset */
4689 cgroup_init_cftsets(ss);
4691 mutex_lock(&cgroup_mutex);
4692 cgroup_subsys[ss->subsys_id] = ss;
4695 * no ss->css_alloc seems to need anything important in the ss
4696 * struct, so this can happen first (i.e. before the dummy root
4699 css = ss->css_alloc(cgroup_dummy_top);
4701 /* failure case - need to deassign the cgroup_subsys[] slot. */
4702 cgroup_subsys[ss->subsys_id] = NULL;
4703 mutex_unlock(&cgroup_mutex);
4704 return PTR_ERR(css);
4707 list_add(&ss->sibling, &cgroup_dummy_root.subsys_list);
4708 ss->root = &cgroup_dummy_root;
4710 /* our new subsystem will be attached to the dummy hierarchy. */
4711 init_cgroup_css(css, ss, cgroup_dummy_top);
4712 /* init_idr must be after init_cgroup_css because it sets css->id. */
4714 ret = cgroup_init_idr(ss, css);
4720 * Now we need to entangle the css into the existing css_sets. unlike
4721 * in cgroup_init_subsys, there are now multiple css_sets, so each one
4722 * will need a new pointer to it; done by iterating the css_set_table.
4723 * furthermore, modifying the existing css_sets will corrupt the hash
4724 * table state, so each changed css_set will need its hash recomputed.
4725 * this is all done under the css_set_lock.
4727 write_lock(&css_set_lock);
4728 hash_for_each_safe(css_set_table, i, tmp, cset, hlist) {
4729 /* skip entries that we already rehashed */
4730 if (cset->subsys[ss->subsys_id])
4732 /* remove existing entry */
4733 hash_del(&cset->hlist);
4735 cset->subsys[ss->subsys_id] = css;
4736 /* recompute hash and restore entry */
4737 key = css_set_hash(cset->subsys);
4738 hash_add(css_set_table, &cset->hlist, key);
4740 write_unlock(&css_set_lock);
4742 ret = online_css(ss, cgroup_dummy_top);
4747 mutex_unlock(&cgroup_mutex);
4751 mutex_unlock(&cgroup_mutex);
4752 /* @ss can't be mounted here as try_module_get() would fail */
4753 cgroup_unload_subsys(ss);
4756 EXPORT_SYMBOL_GPL(cgroup_load_subsys);
4759 * cgroup_unload_subsys: unload a modular subsystem
4760 * @ss: the subsystem to unload
4762 * This function should be called in a modular subsystem's exitcall. When this
4763 * function is invoked, the refcount on the subsystem's module will be 0, so
4764 * the subsystem will not be attached to any hierarchy.
4766 void cgroup_unload_subsys(struct cgroup_subsys *ss)
4768 struct cgrp_cset_link *link;
4770 BUG_ON(ss->module == NULL);
4773 * we shouldn't be called if the subsystem is in use, and the use of
4774 * try_module_get() in rebind_subsystems() should ensure that it
4775 * doesn't start being used while we're killing it off.
4777 BUG_ON(ss->root != &cgroup_dummy_root);
4779 mutex_lock(&cgroup_mutex);
4781 offline_css(ss, cgroup_dummy_top);
4784 idr_destroy(&ss->idr);
4786 /* deassign the subsys_id */
4787 cgroup_subsys[ss->subsys_id] = NULL;
4789 /* remove subsystem from the dummy root's list of subsystems */
4790 list_del_init(&ss->sibling);
4793 * disentangle the css from all css_sets attached to the dummy
4794 * top. as in loading, we need to pay our respects to the hashtable
4797 write_lock(&css_set_lock);
4798 list_for_each_entry(link, &cgroup_dummy_top->cset_links, cset_link) {
4799 struct css_set *cset = link->cset;
4802 hash_del(&cset->hlist);
4803 cset->subsys[ss->subsys_id] = NULL;
4804 key = css_set_hash(cset->subsys);
4805 hash_add(css_set_table, &cset->hlist, key);
4807 write_unlock(&css_set_lock);
4810 * remove subsystem's css from the cgroup_dummy_top and free it -
4811 * need to free before marking as null because ss->css_free needs
4812 * the cgrp->subsys pointer to find their state. note that this
4813 * also takes care of freeing the css_id.
4815 ss->css_free(cgroup_dummy_top);
4816 cgroup_dummy_top->subsys[ss->subsys_id] = NULL;
4818 mutex_unlock(&cgroup_mutex);
4820 EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
4823 * cgroup_init_early - cgroup initialization at system boot
4825 * Initialize cgroups at system boot, and initialize any
4826 * subsystems that request early init.
4828 int __init cgroup_init_early(void)
4830 struct cgroup_subsys *ss;
4833 atomic_set(&init_css_set.refcount, 1);
4834 INIT_LIST_HEAD(&init_css_set.cgrp_links);
4835 INIT_LIST_HEAD(&init_css_set.tasks);
4836 INIT_HLIST_NODE(&init_css_set.hlist);
4838 init_cgroup_root(&cgroup_dummy_root);
4839 cgroup_root_count = 1;
4840 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4842 init_cgrp_cset_link.cset = &init_css_set;
4843 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4844 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4845 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4847 /* at bootup time, we don't worry about modular subsystems */
4848 for_each_builtin_subsys(ss, i) {
4850 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
4851 BUG_ON(!ss->css_alloc);
4852 BUG_ON(!ss->css_free);
4853 if (ss->subsys_id != i) {
4854 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
4855 ss->name, ss->subsys_id);
4860 cgroup_init_subsys(ss);
4866 * cgroup_init - cgroup initialization
4868 * Register cgroup filesystem and /proc file, and initialize
4869 * any subsystems that didn't request early init.
4871 int __init cgroup_init(void)
4873 struct cgroup_subsys *ss;
4877 err = bdi_init(&cgroup_backing_dev_info);
4881 for_each_builtin_subsys(ss, i) {
4882 if (!ss->early_init)
4883 cgroup_init_subsys(ss);
4885 cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
4888 /* allocate id for the dummy hierarchy */
4889 mutex_lock(&cgroup_mutex);
4890 mutex_lock(&cgroup_root_mutex);
4892 /* Add init_css_set to the hash table */
4893 key = css_set_hash(init_css_set.subsys);
4894 hash_add(css_set_table, &init_css_set.hlist, key);
4896 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4898 mutex_unlock(&cgroup_root_mutex);
4899 mutex_unlock(&cgroup_mutex);
4901 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4907 err = register_filesystem(&cgroup_fs_type);
4909 kobject_put(cgroup_kobj);
4913 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4917 bdi_destroy(&cgroup_backing_dev_info);
4923 * proc_cgroup_show()
4924 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4925 * - Used for /proc/<pid>/cgroup.
4926 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4927 * doesn't really matter if tsk->cgroup changes after we read it,
4928 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4929 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4930 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4931 * cgroup to top_cgroup.
4934 /* TODO: Use a proper seq_file iterator */
4935 int proc_cgroup_show(struct seq_file *m, void *v)
4938 struct task_struct *tsk;
4941 struct cgroupfs_root *root;
4944 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4950 tsk = get_pid_task(pid, PIDTYPE_PID);
4956 mutex_lock(&cgroup_mutex);
4958 for_each_active_root(root) {
4959 struct cgroup_subsys *ss;
4960 struct cgroup *cgrp;
4963 seq_printf(m, "%d:", root->hierarchy_id);
4964 for_each_root_subsys(root, ss)
4965 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4966 if (strlen(root->name))
4967 seq_printf(m, "%sname=%s", count ? "," : "",
4970 cgrp = task_cgroup_from_root(tsk, root);
4971 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4979 mutex_unlock(&cgroup_mutex);
4980 put_task_struct(tsk);
4987 /* Display information about each subsystem and each hierarchy */
4988 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4990 struct cgroup_subsys *ss;
4993 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4995 * ideally we don't want subsystems moving around while we do this.
4996 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4997 * subsys/hierarchy state.
4999 mutex_lock(&cgroup_mutex);
5001 for_each_subsys(ss, i)
5002 seq_printf(m, "%s\t%d\t%d\t%d\n",
5003 ss->name, ss->root->hierarchy_id,
5004 ss->root->number_of_cgroups, !ss->disabled);
5006 mutex_unlock(&cgroup_mutex);
5010 static int cgroupstats_open(struct inode *inode, struct file *file)
5012 return single_open(file, proc_cgroupstats_show, NULL);
5015 static const struct file_operations proc_cgroupstats_operations = {
5016 .open = cgroupstats_open,
5018 .llseek = seq_lseek,
5019 .release = single_release,
5023 * cgroup_fork - attach newly forked task to its parents cgroup.
5024 * @child: pointer to task_struct of forking parent process.
5026 * Description: A task inherits its parent's cgroup at fork().
5028 * A pointer to the shared css_set was automatically copied in
5029 * fork.c by dup_task_struct(). However, we ignore that copy, since
5030 * it was not made under the protection of RCU or cgroup_mutex, so
5031 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
5032 * have already changed current->cgroups, allowing the previously
5033 * referenced cgroup group to be removed and freed.
5035 * At the point that cgroup_fork() is called, 'current' is the parent
5036 * task, and the passed argument 'child' points to the child task.
5038 void cgroup_fork(struct task_struct *child)
5041 get_css_set(task_css_set(current));
5042 child->cgroups = current->cgroups;
5043 task_unlock(current);
5044 INIT_LIST_HEAD(&child->cg_list);
5048 * cgroup_post_fork - called on a new task after adding it to the task list
5049 * @child: the task in question
5051 * Adds the task to the list running through its css_set if necessary and
5052 * call the subsystem fork() callbacks. Has to be after the task is
5053 * visible on the task list in case we race with the first call to
5054 * cgroup_iter_start() - to guarantee that the new task ends up on its
5057 void cgroup_post_fork(struct task_struct *child)
5059 struct cgroup_subsys *ss;
5063 * use_task_css_set_links is set to 1 before we walk the tasklist
5064 * under the tasklist_lock and we read it here after we added the child
5065 * to the tasklist under the tasklist_lock as well. If the child wasn't
5066 * yet in the tasklist when we walked through it from
5067 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
5068 * should be visible now due to the paired locking and barriers implied
5069 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
5070 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
5073 if (use_task_css_set_links) {
5074 write_lock(&css_set_lock);
5076 if (list_empty(&child->cg_list))
5077 list_add(&child->cg_list, &task_css_set(child)->tasks);
5079 write_unlock(&css_set_lock);
5083 * Call ss->fork(). This must happen after @child is linked on
5084 * css_set; otherwise, @child might change state between ->fork()
5085 * and addition to css_set.
5087 if (need_forkexit_callback) {
5089 * fork/exit callbacks are supported only for builtin
5090 * subsystems, and the builtin section of the subsys
5091 * array is immutable, so we don't need to lock the
5092 * subsys array here. On the other hand, modular section
5093 * of the array can be freed at module unload, so we
5096 for_each_builtin_subsys(ss, i)
5103 * cgroup_exit - detach cgroup from exiting task
5104 * @tsk: pointer to task_struct of exiting process
5105 * @run_callback: run exit callbacks?
5107 * Description: Detach cgroup from @tsk and release it.
5109 * Note that cgroups marked notify_on_release force every task in
5110 * them to take the global cgroup_mutex mutex when exiting.
5111 * This could impact scaling on very large systems. Be reluctant to
5112 * use notify_on_release cgroups where very high task exit scaling
5113 * is required on large systems.
5115 * the_top_cgroup_hack:
5117 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
5119 * We call cgroup_exit() while the task is still competent to
5120 * handle notify_on_release(), then leave the task attached to the
5121 * root cgroup in each hierarchy for the remainder of its exit.
5123 * To do this properly, we would increment the reference count on
5124 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
5125 * code we would add a second cgroup function call, to drop that
5126 * reference. This would just create an unnecessary hot spot on
5127 * the top_cgroup reference count, to no avail.
5129 * Normally, holding a reference to a cgroup without bumping its
5130 * count is unsafe. The cgroup could go away, or someone could
5131 * attach us to a different cgroup, decrementing the count on
5132 * the first cgroup that we never incremented. But in this case,
5133 * top_cgroup isn't going away, and either task has PF_EXITING set,
5134 * which wards off any cgroup_attach_task() attempts, or task is a failed
5135 * fork, never visible to cgroup_attach_task.
5137 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
5139 struct cgroup_subsys *ss;
5140 struct css_set *cset;
5144 * Unlink from the css_set task list if necessary.
5145 * Optimistically check cg_list before taking
5148 if (!list_empty(&tsk->cg_list)) {
5149 write_lock(&css_set_lock);
5150 if (!list_empty(&tsk->cg_list))
5151 list_del_init(&tsk->cg_list);
5152 write_unlock(&css_set_lock);
5155 /* Reassign the task to the init_css_set. */
5157 cset = task_css_set(tsk);
5158 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5160 if (run_callbacks && need_forkexit_callback) {
5162 * fork/exit callbacks are supported only for builtin
5163 * subsystems, see cgroup_post_fork() for details.
5165 for_each_builtin_subsys(ss, i) {
5167 struct cgroup *old_cgrp = cset->subsys[i]->cgroup;
5168 struct cgroup *cgrp = task_cgroup(tsk, i);
5170 ss->exit(cgrp, old_cgrp, tsk);
5176 put_css_set_taskexit(cset);
5179 static void check_for_release(struct cgroup *cgrp)
5181 if (cgroup_is_releasable(cgrp) &&
5182 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
5184 * Control Group is currently removeable. If it's not
5185 * already queued for a userspace notification, queue
5188 int need_schedule_work = 0;
5190 raw_spin_lock(&release_list_lock);
5191 if (!cgroup_is_dead(cgrp) &&
5192 list_empty(&cgrp->release_list)) {
5193 list_add(&cgrp->release_list, &release_list);
5194 need_schedule_work = 1;
5196 raw_spin_unlock(&release_list_lock);
5197 if (need_schedule_work)
5198 schedule_work(&release_agent_work);
5203 * Notify userspace when a cgroup is released, by running the
5204 * configured release agent with the name of the cgroup (path
5205 * relative to the root of cgroup file system) as the argument.
5207 * Most likely, this user command will try to rmdir this cgroup.
5209 * This races with the possibility that some other task will be
5210 * attached to this cgroup before it is removed, or that some other
5211 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5212 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5213 * unused, and this cgroup will be reprieved from its death sentence,
5214 * to continue to serve a useful existence. Next time it's released,
5215 * we will get notified again, if it still has 'notify_on_release' set.
5217 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5218 * means only wait until the task is successfully execve()'d. The
5219 * separate release agent task is forked by call_usermodehelper(),
5220 * then control in this thread returns here, without waiting for the
5221 * release agent task. We don't bother to wait because the caller of
5222 * this routine has no use for the exit status of the release agent
5223 * task, so no sense holding our caller up for that.
5225 static void cgroup_release_agent(struct work_struct *work)
5227 BUG_ON(work != &release_agent_work);
5228 mutex_lock(&cgroup_mutex);
5229 raw_spin_lock(&release_list_lock);
5230 while (!list_empty(&release_list)) {
5231 char *argv[3], *envp[3];
5233 char *pathbuf = NULL, *agentbuf = NULL;
5234 struct cgroup *cgrp = list_entry(release_list.next,
5237 list_del_init(&cgrp->release_list);
5238 raw_spin_unlock(&release_list_lock);
5239 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
5242 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
5244 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5249 argv[i++] = agentbuf;
5250 argv[i++] = pathbuf;
5254 /* minimal command environment */
5255 envp[i++] = "HOME=/";
5256 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5259 /* Drop the lock while we invoke the usermode helper,
5260 * since the exec could involve hitting disk and hence
5261 * be a slow process */
5262 mutex_unlock(&cgroup_mutex);
5263 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5264 mutex_lock(&cgroup_mutex);
5268 raw_spin_lock(&release_list_lock);
5270 raw_spin_unlock(&release_list_lock);
5271 mutex_unlock(&cgroup_mutex);
5274 static int __init cgroup_disable(char *str)
5276 struct cgroup_subsys *ss;
5280 while ((token = strsep(&str, ",")) != NULL) {
5285 * cgroup_disable, being at boot time, can't know about
5286 * module subsystems, so we don't worry about them.
5288 for_each_builtin_subsys(ss, i) {
5289 if (!strcmp(token, ss->name)) {
5291 printk(KERN_INFO "Disabling %s control group"
5292 " subsystem\n", ss->name);
5299 __setup("cgroup_disable=", cgroup_disable);
5302 * Functons for CSS ID.
5305 /* to get ID other than 0, this should be called when !cgroup_is_dead() */
5306 unsigned short css_id(struct cgroup_subsys_state *css)
5308 struct css_id *cssid;
5311 * This css_id() can return correct value when somone has refcnt
5312 * on this or this is under rcu_read_lock(). Once css->id is allocated,
5313 * it's unchanged until freed.
5315 cssid = rcu_dereference_raw(css->id);
5321 EXPORT_SYMBOL_GPL(css_id);
5324 * css_is_ancestor - test "root" css is an ancestor of "child"
5325 * @child: the css to be tested.
5326 * @root: the css supporsed to be an ancestor of the child.
5328 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
5329 * this function reads css->id, the caller must hold rcu_read_lock().
5330 * But, considering usual usage, the csses should be valid objects after test.
5331 * Assuming that the caller will do some action to the child if this returns
5332 * returns true, the caller must take "child";s reference count.
5333 * If "child" is valid object and this returns true, "root" is valid, too.
5336 bool css_is_ancestor(struct cgroup_subsys_state *child,
5337 const struct cgroup_subsys_state *root)
5339 struct css_id *child_id;
5340 struct css_id *root_id;
5342 child_id = rcu_dereference(child->id);
5345 root_id = rcu_dereference(root->id);
5348 if (child_id->depth < root_id->depth)
5350 if (child_id->stack[root_id->depth] != root_id->id)
5355 void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
5357 struct css_id *id = rcu_dereference_protected(css->id, true);
5359 /* When this is called before css_id initialization, id can be NULL */
5363 BUG_ON(!ss->use_id);
5365 rcu_assign_pointer(id->css, NULL);
5366 rcu_assign_pointer(css->id, NULL);
5367 spin_lock(&ss->id_lock);
5368 idr_remove(&ss->idr, id->id);
5369 spin_unlock(&ss->id_lock);
5370 kfree_rcu(id, rcu_head);
5372 EXPORT_SYMBOL_GPL(free_css_id);
5375 * This is called by init or create(). Then, calls to this function are
5376 * always serialized (By cgroup_mutex() at create()).
5379 static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
5381 struct css_id *newid;
5384 BUG_ON(!ss->use_id);
5386 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
5387 newid = kzalloc(size, GFP_KERNEL);
5389 return ERR_PTR(-ENOMEM);
5391 idr_preload(GFP_KERNEL);
5392 spin_lock(&ss->id_lock);
5393 /* Don't use 0. allocates an ID of 1-65535 */
5394 ret = idr_alloc(&ss->idr, newid, 1, CSS_ID_MAX + 1, GFP_NOWAIT);
5395 spin_unlock(&ss->id_lock);
5398 /* Returns error when there are no free spaces for new ID.*/
5403 newid->depth = depth;
5407 return ERR_PTR(ret);
5411 static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
5412 struct cgroup_subsys_state *rootcss)
5414 struct css_id *newid;
5416 spin_lock_init(&ss->id_lock);
5419 newid = get_new_cssid(ss, 0);
5421 return PTR_ERR(newid);
5423 newid->stack[0] = newid->id;
5424 RCU_INIT_POINTER(newid->css, rootcss);
5425 RCU_INIT_POINTER(rootcss->id, newid);
5429 static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
5430 struct cgroup *child)
5432 int subsys_id, i, depth = 0;
5433 struct cgroup_subsys_state *parent_css, *child_css;
5434 struct css_id *child_id, *parent_id;
5436 subsys_id = ss->subsys_id;
5437 parent_css = parent->subsys[subsys_id];
5438 child_css = child->subsys[subsys_id];
5439 parent_id = rcu_dereference_protected(parent_css->id, true);
5440 depth = parent_id->depth + 1;
5442 child_id = get_new_cssid(ss, depth);
5443 if (IS_ERR(child_id))
5444 return PTR_ERR(child_id);
5446 for (i = 0; i < depth; i++)
5447 child_id->stack[i] = parent_id->stack[i];
5448 child_id->stack[depth] = child_id->id;
5450 * child_id->css pointer will be set after this cgroup is available
5451 * see cgroup_populate_dir()
5453 rcu_assign_pointer(child_css->id, child_id);
5459 * css_lookup - lookup css by id
5460 * @ss: cgroup subsys to be looked into.
5463 * Returns pointer to cgroup_subsys_state if there is valid one with id.
5464 * NULL if not. Should be called under rcu_read_lock()
5466 struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
5468 struct css_id *cssid = NULL;
5470 BUG_ON(!ss->use_id);
5471 cssid = idr_find(&ss->idr, id);
5473 if (unlikely(!cssid))
5476 return rcu_dereference(cssid->css);
5478 EXPORT_SYMBOL_GPL(css_lookup);
5481 * get corresponding css from file open on cgroupfs directory
5483 struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id)
5485 struct cgroup *cgrp;
5486 struct inode *inode;
5487 struct cgroup_subsys_state *css;
5489 inode = file_inode(f);
5490 /* check in cgroup filesystem dir */
5491 if (inode->i_op != &cgroup_dir_inode_operations)
5492 return ERR_PTR(-EBADF);
5494 if (id < 0 || id >= CGROUP_SUBSYS_COUNT)
5495 return ERR_PTR(-EINVAL);
5498 cgrp = __d_cgrp(f->f_dentry);
5499 css = cgrp->subsys[id];
5500 return css ? css : ERR_PTR(-ENOENT);
5503 #ifdef CONFIG_CGROUP_DEBUG
5504 static struct cgroup_subsys_state *debug_css_alloc(struct cgroup *cgrp)
5506 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5509 return ERR_PTR(-ENOMEM);
5514 static void debug_css_free(struct cgroup *cgrp)
5516 kfree(cgrp->subsys[debug_subsys_id]);
5519 static u64 debug_taskcount_read(struct cgroup *cgrp, struct cftype *cft)
5521 return cgroup_task_count(cgrp);
5524 static u64 current_css_set_read(struct cgroup *cgrp, struct cftype *cft)
5526 return (u64)(unsigned long)current->cgroups;
5529 static u64 current_css_set_refcount_read(struct cgroup *cgrp,
5535 count = atomic_read(&task_css_set(current)->refcount);
5540 static int current_css_set_cg_links_read(struct cgroup *cgrp,
5542 struct seq_file *seq)
5544 struct cgrp_cset_link *link;
5545 struct css_set *cset;
5547 read_lock(&css_set_lock);
5549 cset = rcu_dereference(current->cgroups);
5550 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5551 struct cgroup *c = link->cgrp;
5555 name = c->dentry->d_name.name;
5558 seq_printf(seq, "Root %d group %s\n",
5559 c->root->hierarchy_id, name);
5562 read_unlock(&css_set_lock);
5566 #define MAX_TASKS_SHOWN_PER_CSS 25
5567 static int cgroup_css_links_read(struct cgroup *cgrp,
5569 struct seq_file *seq)
5571 struct cgrp_cset_link *link;
5573 read_lock(&css_set_lock);
5574 list_for_each_entry(link, &cgrp->cset_links, cset_link) {
5575 struct css_set *cset = link->cset;
5576 struct task_struct *task;
5578 seq_printf(seq, "css_set %p\n", cset);
5579 list_for_each_entry(task, &cset->tasks, cg_list) {
5580 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
5581 seq_puts(seq, " ...\n");
5584 seq_printf(seq, " task %d\n",
5585 task_pid_vnr(task));
5589 read_unlock(&css_set_lock);
5593 static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
5595 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
5598 static struct cftype debug_files[] = {
5600 .name = "taskcount",
5601 .read_u64 = debug_taskcount_read,
5605 .name = "current_css_set",
5606 .read_u64 = current_css_set_read,
5610 .name = "current_css_set_refcount",
5611 .read_u64 = current_css_set_refcount_read,
5615 .name = "current_css_set_cg_links",
5616 .read_seq_string = current_css_set_cg_links_read,
5620 .name = "cgroup_css_links",
5621 .read_seq_string = cgroup_css_links_read,
5625 .name = "releasable",
5626 .read_u64 = releasable_read,
5632 struct cgroup_subsys debug_subsys = {
5634 .css_alloc = debug_css_alloc,
5635 .css_free = debug_css_free,
5636 .subsys_id = debug_subsys_id,
5637 .base_cftypes = debug_files,
5639 #endif /* CONFIG_CGROUP_DEBUG */