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;
1008 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1009 BUG_ON(!mutex_is_locked(&cgroup_root_mutex));
1011 /* Check that any added subsystems are currently free */
1012 for_each_subsys(ss, i) {
1013 unsigned long bit = 1UL << i;
1015 if (!(bit & added_mask))
1018 if (ss->root != &cgroup_dummy_root) {
1019 /* Subsystem isn't free */
1024 ret = cgroup_populate_dir(cgrp, added_mask);
1029 * Nothing can fail from this point on. Remove files for the
1030 * removed subsystems and rebind each subsystem.
1032 cgroup_clear_dir(cgrp, removed_mask);
1034 for_each_subsys(ss, i) {
1035 unsigned long bit = 1UL << i;
1037 if (bit & added_mask) {
1038 /* We're binding this subsystem to this hierarchy */
1039 BUG_ON(cgrp->subsys[i]);
1040 BUG_ON(!cgroup_dummy_top->subsys[i]);
1041 BUG_ON(cgroup_dummy_top->subsys[i]->cgroup != cgroup_dummy_top);
1043 cgrp->subsys[i] = cgroup_dummy_top->subsys[i];
1044 cgrp->subsys[i]->cgroup = cgrp;
1045 list_move(&ss->sibling, &root->subsys_list);
1050 /* refcount was already taken, and we're keeping it */
1051 root->subsys_mask |= bit;
1052 } else if (bit & removed_mask) {
1053 /* We're removing this subsystem */
1054 BUG_ON(cgrp->subsys[i] != cgroup_dummy_top->subsys[i]);
1055 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
1058 ss->bind(cgroup_dummy_top);
1059 cgroup_dummy_top->subsys[i]->cgroup = cgroup_dummy_top;
1060 cgrp->subsys[i] = NULL;
1061 cgroup_subsys[i]->root = &cgroup_dummy_root;
1062 list_move(&ss->sibling, &cgroup_dummy_root.subsys_list);
1064 /* subsystem is now free - drop reference on module */
1065 module_put(ss->module);
1066 root->subsys_mask &= ~bit;
1067 } else if (bit & root->subsys_mask) {
1068 /* Subsystem state should already exist */
1069 BUG_ON(!cgrp->subsys[i]);
1071 * a refcount was taken, but we already had one, so
1072 * drop the extra reference.
1074 module_put(ss->module);
1075 #ifdef CONFIG_MODULE_UNLOAD
1076 BUG_ON(ss->module && !module_refcount(ss->module));
1079 /* Subsystem state shouldn't exist */
1080 BUG_ON(cgrp->subsys[i]);
1085 * Mark @root has finished binding subsystems. @root->subsys_mask
1086 * now matches the bound subsystems.
1088 root->flags |= CGRP_ROOT_SUBSYS_BOUND;
1093 static int cgroup_show_options(struct seq_file *seq, struct dentry *dentry)
1095 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
1096 struct cgroup_subsys *ss;
1098 mutex_lock(&cgroup_root_mutex);
1099 for_each_root_subsys(root, ss)
1100 seq_printf(seq, ",%s", ss->name);
1101 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1102 seq_puts(seq, ",sane_behavior");
1103 if (root->flags & CGRP_ROOT_NOPREFIX)
1104 seq_puts(seq, ",noprefix");
1105 if (root->flags & CGRP_ROOT_XATTR)
1106 seq_puts(seq, ",xattr");
1107 if (strlen(root->release_agent_path))
1108 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1109 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1110 seq_puts(seq, ",clone_children");
1111 if (strlen(root->name))
1112 seq_printf(seq, ",name=%s", root->name);
1113 mutex_unlock(&cgroup_root_mutex);
1117 struct cgroup_sb_opts {
1118 unsigned long subsys_mask;
1119 unsigned long flags;
1120 char *release_agent;
1121 bool cpuset_clone_children;
1123 /* User explicitly requested empty subsystem */
1126 struct cgroupfs_root *new_root;
1131 * Convert a hierarchy specifier into a bitmask of subsystems and
1132 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1133 * array. This function takes refcounts on subsystems to be used, unless it
1134 * returns error, in which case no refcounts are taken.
1136 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1138 char *token, *o = data;
1139 bool all_ss = false, one_ss = false;
1140 unsigned long mask = (unsigned long)-1;
1141 bool module_pin_failed = false;
1142 struct cgroup_subsys *ss;
1145 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1147 #ifdef CONFIG_CPUSETS
1148 mask = ~(1UL << cpuset_subsys_id);
1151 memset(opts, 0, sizeof(*opts));
1153 while ((token = strsep(&o, ",")) != NULL) {
1156 if (!strcmp(token, "none")) {
1157 /* Explicitly have no subsystems */
1161 if (!strcmp(token, "all")) {
1162 /* Mutually exclusive option 'all' + subsystem name */
1168 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1169 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1172 if (!strcmp(token, "noprefix")) {
1173 opts->flags |= CGRP_ROOT_NOPREFIX;
1176 if (!strcmp(token, "clone_children")) {
1177 opts->cpuset_clone_children = true;
1180 if (!strcmp(token, "xattr")) {
1181 opts->flags |= CGRP_ROOT_XATTR;
1184 if (!strncmp(token, "release_agent=", 14)) {
1185 /* Specifying two release agents is forbidden */
1186 if (opts->release_agent)
1188 opts->release_agent =
1189 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1190 if (!opts->release_agent)
1194 if (!strncmp(token, "name=", 5)) {
1195 const char *name = token + 5;
1196 /* Can't specify an empty name */
1199 /* Must match [\w.-]+ */
1200 for (i = 0; i < strlen(name); i++) {
1204 if ((c == '.') || (c == '-') || (c == '_'))
1208 /* Specifying two names is forbidden */
1211 opts->name = kstrndup(name,
1212 MAX_CGROUP_ROOT_NAMELEN - 1,
1220 for_each_subsys(ss, i) {
1221 if (strcmp(token, ss->name))
1226 /* Mutually exclusive option 'all' + subsystem name */
1229 set_bit(i, &opts->subsys_mask);
1234 if (i == CGROUP_SUBSYS_COUNT)
1239 * If the 'all' option was specified select all the subsystems,
1240 * otherwise if 'none', 'name=' and a subsystem name options
1241 * were not specified, let's default to 'all'
1243 if (all_ss || (!one_ss && !opts->none && !opts->name))
1244 for_each_subsys(ss, i)
1246 set_bit(i, &opts->subsys_mask);
1248 /* Consistency checks */
1250 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1251 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1253 if (opts->flags & CGRP_ROOT_NOPREFIX) {
1254 pr_err("cgroup: sane_behavior: noprefix is not allowed\n");
1258 if (opts->cpuset_clone_children) {
1259 pr_err("cgroup: sane_behavior: clone_children is not allowed\n");
1265 * Option noprefix was introduced just for backward compatibility
1266 * with the old cpuset, so we allow noprefix only if mounting just
1267 * the cpuset subsystem.
1269 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1273 /* Can't specify "none" and some subsystems */
1274 if (opts->subsys_mask && opts->none)
1278 * We either have to specify by name or by subsystems. (So all
1279 * empty hierarchies must have a name).
1281 if (!opts->subsys_mask && !opts->name)
1285 * Grab references on all the modules we'll need, so the subsystems
1286 * don't dance around before rebind_subsystems attaches them. This may
1287 * take duplicate reference counts on a subsystem that's already used,
1288 * but rebind_subsystems handles this case.
1290 for_each_subsys(ss, i) {
1291 if (!(opts->subsys_mask & (1UL << i)))
1293 if (!try_module_get(cgroup_subsys[i]->module)) {
1294 module_pin_failed = true;
1298 if (module_pin_failed) {
1300 * oops, one of the modules was going away. this means that we
1301 * raced with a module_delete call, and to the user this is
1302 * essentially a "subsystem doesn't exist" case.
1304 for (i--; i >= 0; i--) {
1305 /* drop refcounts only on the ones we took */
1306 unsigned long bit = 1UL << i;
1308 if (!(bit & opts->subsys_mask))
1310 module_put(cgroup_subsys[i]->module);
1318 static void drop_parsed_module_refcounts(unsigned long subsys_mask)
1320 struct cgroup_subsys *ss;
1323 mutex_lock(&cgroup_mutex);
1324 for_each_subsys(ss, i)
1325 if (subsys_mask & (1UL << i))
1326 module_put(cgroup_subsys[i]->module);
1327 mutex_unlock(&cgroup_mutex);
1330 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1333 struct cgroupfs_root *root = sb->s_fs_info;
1334 struct cgroup *cgrp = &root->top_cgroup;
1335 struct cgroup_sb_opts opts;
1336 unsigned long added_mask, removed_mask;
1338 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1339 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1343 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1344 mutex_lock(&cgroup_mutex);
1345 mutex_lock(&cgroup_root_mutex);
1347 /* See what subsystems are wanted */
1348 ret = parse_cgroupfs_options(data, &opts);
1352 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1353 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1354 task_tgid_nr(current), current->comm);
1356 added_mask = opts.subsys_mask & ~root->subsys_mask;
1357 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1359 /* Don't allow flags or name to change at remount */
1360 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1361 (opts.name && strcmp(opts.name, root->name))) {
1362 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1363 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1364 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1369 /* remounting is not allowed for populated hierarchies */
1370 if (root->number_of_cgroups > 1) {
1375 ret = rebind_subsystems(root, added_mask, removed_mask);
1379 if (opts.release_agent)
1380 strcpy(root->release_agent_path, opts.release_agent);
1382 kfree(opts.release_agent);
1384 mutex_unlock(&cgroup_root_mutex);
1385 mutex_unlock(&cgroup_mutex);
1386 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1388 drop_parsed_module_refcounts(opts.subsys_mask);
1392 static const struct super_operations cgroup_ops = {
1393 .statfs = simple_statfs,
1394 .drop_inode = generic_delete_inode,
1395 .show_options = cgroup_show_options,
1396 .remount_fs = cgroup_remount,
1399 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1401 INIT_LIST_HEAD(&cgrp->sibling);
1402 INIT_LIST_HEAD(&cgrp->children);
1403 INIT_LIST_HEAD(&cgrp->files);
1404 INIT_LIST_HEAD(&cgrp->cset_links);
1405 INIT_LIST_HEAD(&cgrp->release_list);
1406 INIT_LIST_HEAD(&cgrp->pidlists);
1407 mutex_init(&cgrp->pidlist_mutex);
1408 INIT_LIST_HEAD(&cgrp->event_list);
1409 spin_lock_init(&cgrp->event_list_lock);
1410 simple_xattrs_init(&cgrp->xattrs);
1413 static void init_cgroup_root(struct cgroupfs_root *root)
1415 struct cgroup *cgrp = &root->top_cgroup;
1417 INIT_LIST_HEAD(&root->subsys_list);
1418 INIT_LIST_HEAD(&root->root_list);
1419 root->number_of_cgroups = 1;
1421 RCU_INIT_POINTER(cgrp->name, &root_cgroup_name);
1422 init_cgroup_housekeeping(cgrp);
1425 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
1429 lockdep_assert_held(&cgroup_mutex);
1430 lockdep_assert_held(&cgroup_root_mutex);
1432 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
1437 root->hierarchy_id = id;
1441 static void cgroup_exit_root_id(struct cgroupfs_root *root)
1443 lockdep_assert_held(&cgroup_mutex);
1444 lockdep_assert_held(&cgroup_root_mutex);
1446 if (root->hierarchy_id) {
1447 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1448 root->hierarchy_id = 0;
1452 static int cgroup_test_super(struct super_block *sb, void *data)
1454 struct cgroup_sb_opts *opts = data;
1455 struct cgroupfs_root *root = sb->s_fs_info;
1457 /* If we asked for a name then it must match */
1458 if (opts->name && strcmp(opts->name, root->name))
1462 * If we asked for subsystems (or explicitly for no
1463 * subsystems) then they must match
1465 if ((opts->subsys_mask || opts->none)
1466 && (opts->subsys_mask != root->subsys_mask))
1472 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1474 struct cgroupfs_root *root;
1476 if (!opts->subsys_mask && !opts->none)
1479 root = kzalloc(sizeof(*root), GFP_KERNEL);
1481 return ERR_PTR(-ENOMEM);
1483 init_cgroup_root(root);
1486 * We need to set @root->subsys_mask now so that @root can be
1487 * matched by cgroup_test_super() before it finishes
1488 * initialization; otherwise, competing mounts with the same
1489 * options may try to bind the same subsystems instead of waiting
1490 * for the first one leading to unexpected mount errors.
1491 * SUBSYS_BOUND will be set once actual binding is complete.
1493 root->subsys_mask = opts->subsys_mask;
1494 root->flags = opts->flags;
1495 ida_init(&root->cgroup_ida);
1496 if (opts->release_agent)
1497 strcpy(root->release_agent_path, opts->release_agent);
1499 strcpy(root->name, opts->name);
1500 if (opts->cpuset_clone_children)
1501 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1505 static void cgroup_free_root(struct cgroupfs_root *root)
1508 /* hierarhcy ID shoulid already have been released */
1509 WARN_ON_ONCE(root->hierarchy_id);
1511 ida_destroy(&root->cgroup_ida);
1516 static int cgroup_set_super(struct super_block *sb, void *data)
1519 struct cgroup_sb_opts *opts = data;
1521 /* If we don't have a new root, we can't set up a new sb */
1522 if (!opts->new_root)
1525 BUG_ON(!opts->subsys_mask && !opts->none);
1527 ret = set_anon_super(sb, NULL);
1531 sb->s_fs_info = opts->new_root;
1532 opts->new_root->sb = sb;
1534 sb->s_blocksize = PAGE_CACHE_SIZE;
1535 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1536 sb->s_magic = CGROUP_SUPER_MAGIC;
1537 sb->s_op = &cgroup_ops;
1542 static int cgroup_get_rootdir(struct super_block *sb)
1544 static const struct dentry_operations cgroup_dops = {
1545 .d_iput = cgroup_diput,
1546 .d_delete = cgroup_delete,
1549 struct inode *inode =
1550 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1555 inode->i_fop = &simple_dir_operations;
1556 inode->i_op = &cgroup_dir_inode_operations;
1557 /* directories start off with i_nlink == 2 (for "." entry) */
1559 sb->s_root = d_make_root(inode);
1562 /* for everything else we want ->d_op set */
1563 sb->s_d_op = &cgroup_dops;
1567 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1568 int flags, const char *unused_dev_name,
1571 struct cgroup_sb_opts opts;
1572 struct cgroupfs_root *root;
1574 struct super_block *sb;
1575 struct cgroupfs_root *new_root;
1576 struct list_head tmp_links;
1577 struct inode *inode;
1578 const struct cred *cred;
1580 /* First find the desired set of subsystems */
1581 mutex_lock(&cgroup_mutex);
1582 ret = parse_cgroupfs_options(data, &opts);
1583 mutex_unlock(&cgroup_mutex);
1588 * Allocate a new cgroup root. We may not need it if we're
1589 * reusing an existing hierarchy.
1591 new_root = cgroup_root_from_opts(&opts);
1592 if (IS_ERR(new_root)) {
1593 ret = PTR_ERR(new_root);
1596 opts.new_root = new_root;
1598 /* Locate an existing or new sb for this hierarchy */
1599 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, 0, &opts);
1602 cgroup_free_root(opts.new_root);
1606 root = sb->s_fs_info;
1608 if (root == opts.new_root) {
1609 /* We used the new root structure, so this is a new hierarchy */
1610 struct cgroup *root_cgrp = &root->top_cgroup;
1611 struct cgroupfs_root *existing_root;
1613 struct css_set *cset;
1615 BUG_ON(sb->s_root != NULL);
1617 ret = cgroup_get_rootdir(sb);
1619 goto drop_new_super;
1620 inode = sb->s_root->d_inode;
1622 mutex_lock(&inode->i_mutex);
1623 mutex_lock(&cgroup_mutex);
1624 mutex_lock(&cgroup_root_mutex);
1626 /* Check for name clashes with existing mounts */
1628 if (strlen(root->name))
1629 for_each_active_root(existing_root)
1630 if (!strcmp(existing_root->name, root->name))
1634 * We're accessing css_set_count without locking
1635 * css_set_lock here, but that's OK - it can only be
1636 * increased by someone holding cgroup_lock, and
1637 * that's us. The worst that can happen is that we
1638 * have some link structures left over
1640 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1644 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1645 ret = cgroup_init_root_id(root, 2, 0);
1649 sb->s_root->d_fsdata = root_cgrp;
1650 root_cgrp->dentry = sb->s_root;
1653 * We're inside get_sb() and will call lookup_one_len() to
1654 * create the root files, which doesn't work if SELinux is
1655 * in use. The following cred dancing somehow works around
1656 * it. See 2ce9738ba ("cgroupfs: use init_cred when
1657 * populating new cgroupfs mount") for more details.
1659 cred = override_creds(&init_cred);
1661 ret = cgroup_addrm_files(root_cgrp, NULL, cgroup_base_files, true);
1665 ret = rebind_subsystems(root, root->subsys_mask, 0);
1672 * There must be no failure case after here, since rebinding
1673 * takes care of subsystems' refcounts, which are explicitly
1674 * dropped in the failure exit path.
1677 list_add(&root->root_list, &cgroup_roots);
1678 cgroup_root_count++;
1680 /* Link the top cgroup in this hierarchy into all
1681 * the css_set objects */
1682 write_lock(&css_set_lock);
1683 hash_for_each(css_set_table, i, cset, hlist)
1684 link_css_set(&tmp_links, cset, root_cgrp);
1685 write_unlock(&css_set_lock);
1687 free_cgrp_cset_links(&tmp_links);
1689 BUG_ON(!list_empty(&root_cgrp->children));
1690 BUG_ON(root->number_of_cgroups != 1);
1692 mutex_unlock(&cgroup_root_mutex);
1693 mutex_unlock(&cgroup_mutex);
1694 mutex_unlock(&inode->i_mutex);
1697 * We re-used an existing hierarchy - the new root (if
1698 * any) is not needed
1700 cgroup_free_root(opts.new_root);
1702 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1703 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1704 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1706 goto drop_new_super;
1708 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1712 /* no subsys rebinding, so refcounts don't change */
1713 drop_parsed_module_refcounts(opts.subsys_mask);
1716 kfree(opts.release_agent);
1718 return dget(sb->s_root);
1721 free_cgrp_cset_links(&tmp_links);
1722 cgroup_addrm_files(&root->top_cgroup, NULL, cgroup_base_files, false);
1725 cgroup_exit_root_id(root);
1726 mutex_unlock(&cgroup_root_mutex);
1727 mutex_unlock(&cgroup_mutex);
1728 mutex_unlock(&inode->i_mutex);
1730 deactivate_locked_super(sb);
1732 drop_parsed_module_refcounts(opts.subsys_mask);
1734 kfree(opts.release_agent);
1736 return ERR_PTR(ret);
1739 static void cgroup_kill_sb(struct super_block *sb) {
1740 struct cgroupfs_root *root = sb->s_fs_info;
1741 struct cgroup *cgrp = &root->top_cgroup;
1742 struct cgrp_cset_link *link, *tmp_link;
1747 BUG_ON(root->number_of_cgroups != 1);
1748 BUG_ON(!list_empty(&cgrp->children));
1750 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1751 mutex_lock(&cgroup_mutex);
1752 mutex_lock(&cgroup_root_mutex);
1754 /* Rebind all subsystems back to the default hierarchy */
1755 if (root->flags & CGRP_ROOT_SUBSYS_BOUND) {
1756 ret = rebind_subsystems(root, 0, root->subsys_mask);
1757 /* Shouldn't be able to fail ... */
1762 * Release all the links from cset_links to this hierarchy's
1765 write_lock(&css_set_lock);
1767 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1768 list_del(&link->cset_link);
1769 list_del(&link->cgrp_link);
1772 write_unlock(&css_set_lock);
1774 if (!list_empty(&root->root_list)) {
1775 list_del(&root->root_list);
1776 cgroup_root_count--;
1779 cgroup_exit_root_id(root);
1781 mutex_unlock(&cgroup_root_mutex);
1782 mutex_unlock(&cgroup_mutex);
1783 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1785 simple_xattrs_free(&cgrp->xattrs);
1787 kill_litter_super(sb);
1788 cgroup_free_root(root);
1791 static struct file_system_type cgroup_fs_type = {
1793 .mount = cgroup_mount,
1794 .kill_sb = cgroup_kill_sb,
1797 static struct kobject *cgroup_kobj;
1800 * cgroup_path - generate the path of a cgroup
1801 * @cgrp: the cgroup in question
1802 * @buf: the buffer to write the path into
1803 * @buflen: the length of the buffer
1805 * Writes path of cgroup into buf. Returns 0 on success, -errno on error.
1807 * We can't generate cgroup path using dentry->d_name, as accessing
1808 * dentry->name must be protected by irq-unsafe dentry->d_lock or parent
1809 * inode's i_mutex, while on the other hand cgroup_path() can be called
1810 * with some irq-safe spinlocks held.
1812 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1814 int ret = -ENAMETOOLONG;
1817 if (!cgrp->parent) {
1818 if (strlcpy(buf, "/", buflen) >= buflen)
1819 return -ENAMETOOLONG;
1823 start = buf + buflen - 1;
1828 const char *name = cgroup_name(cgrp);
1832 if ((start -= len) < buf)
1834 memcpy(start, name, len);
1840 cgrp = cgrp->parent;
1841 } while (cgrp->parent);
1843 memmove(buf, start, buf + buflen - start);
1848 EXPORT_SYMBOL_GPL(cgroup_path);
1851 * task_cgroup_path_from_hierarchy - cgroup path of a task on a hierarchy
1852 * @task: target task
1853 * @hierarchy_id: the hierarchy to look up @task's cgroup from
1854 * @buf: the buffer to write the path into
1855 * @buflen: the length of the buffer
1857 * Determine @task's cgroup on the hierarchy specified by @hierarchy_id and
1858 * copy its path into @buf. This function grabs cgroup_mutex and shouldn't
1859 * be used inside locks used by cgroup controller callbacks.
1861 int task_cgroup_path_from_hierarchy(struct task_struct *task, int hierarchy_id,
1862 char *buf, size_t buflen)
1864 struct cgroupfs_root *root;
1865 struct cgroup *cgrp = NULL;
1868 mutex_lock(&cgroup_mutex);
1870 root = idr_find(&cgroup_hierarchy_idr, hierarchy_id);
1872 cgrp = task_cgroup_from_root(task, root);
1873 ret = cgroup_path(cgrp, buf, buflen);
1876 mutex_unlock(&cgroup_mutex);
1880 EXPORT_SYMBOL_GPL(task_cgroup_path_from_hierarchy);
1883 * Control Group taskset
1885 struct task_and_cgroup {
1886 struct task_struct *task;
1887 struct cgroup *cgrp;
1891 struct cgroup_taskset {
1892 struct task_and_cgroup single;
1893 struct flex_array *tc_array;
1896 struct cgroup *cur_cgrp;
1900 * cgroup_taskset_first - reset taskset and return the first task
1901 * @tset: taskset of interest
1903 * @tset iteration is initialized and the first task is returned.
1905 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1907 if (tset->tc_array) {
1909 return cgroup_taskset_next(tset);
1911 tset->cur_cgrp = tset->single.cgrp;
1912 return tset->single.task;
1915 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1918 * cgroup_taskset_next - iterate to the next task in taskset
1919 * @tset: taskset of interest
1921 * Return the next task in @tset. Iteration must have been initialized
1922 * with cgroup_taskset_first().
1924 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1926 struct task_and_cgroup *tc;
1928 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1931 tc = flex_array_get(tset->tc_array, tset->idx++);
1932 tset->cur_cgrp = tc->cgrp;
1935 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1938 * cgroup_taskset_cur_cgroup - return the matching cgroup for the current task
1939 * @tset: taskset of interest
1941 * Return the cgroup for the current (last returned) task of @tset. This
1942 * function must be preceded by either cgroup_taskset_first() or
1943 * cgroup_taskset_next().
1945 struct cgroup *cgroup_taskset_cur_cgroup(struct cgroup_taskset *tset)
1947 return tset->cur_cgrp;
1949 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_cgroup);
1952 * cgroup_taskset_size - return the number of tasks in taskset
1953 * @tset: taskset of interest
1955 int cgroup_taskset_size(struct cgroup_taskset *tset)
1957 return tset->tc_array ? tset->tc_array_len : 1;
1959 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1963 * cgroup_task_migrate - move a task from one cgroup to another.
1965 * Must be called with cgroup_mutex and threadgroup locked.
1967 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1968 struct task_struct *tsk,
1969 struct css_set *new_cset)
1971 struct css_set *old_cset;
1974 * We are synchronized through threadgroup_lock() against PF_EXITING
1975 * setting such that we can't race against cgroup_exit() changing the
1976 * css_set to init_css_set and dropping the old one.
1978 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1979 old_cset = task_css_set(tsk);
1982 rcu_assign_pointer(tsk->cgroups, new_cset);
1985 /* Update the css_set linked lists if we're using them */
1986 write_lock(&css_set_lock);
1987 if (!list_empty(&tsk->cg_list))
1988 list_move(&tsk->cg_list, &new_cset->tasks);
1989 write_unlock(&css_set_lock);
1992 * We just gained a reference on old_cset by taking it from the
1993 * task. As trading it for new_cset is protected by cgroup_mutex,
1994 * we're safe to drop it here; it will be freed under RCU.
1996 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1997 put_css_set(old_cset);
2001 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2002 * @cgrp: the cgroup to attach to
2003 * @tsk: the task or the leader of the threadgroup to be attached
2004 * @threadgroup: attach the whole threadgroup?
2006 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
2007 * task_lock of @tsk or each thread in the threadgroup individually in turn.
2009 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
2012 int retval, i, group_size;
2013 struct cgroup_subsys *ss, *failed_ss = NULL;
2014 struct cgroupfs_root *root = cgrp->root;
2015 /* threadgroup list cursor and array */
2016 struct task_struct *leader = tsk;
2017 struct task_and_cgroup *tc;
2018 struct flex_array *group;
2019 struct cgroup_taskset tset = { };
2022 * step 0: in order to do expensive, possibly blocking operations for
2023 * every thread, we cannot iterate the thread group list, since it needs
2024 * rcu or tasklist locked. instead, build an array of all threads in the
2025 * group - group_rwsem prevents new threads from appearing, and if
2026 * threads exit, this will just be an over-estimate.
2029 group_size = get_nr_threads(tsk);
2032 /* flex_array supports very large thread-groups better than kmalloc. */
2033 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
2036 /* pre-allocate to guarantee space while iterating in rcu read-side. */
2037 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
2039 goto out_free_group_list;
2043 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2044 * already PF_EXITING could be freed from underneath us unless we
2045 * take an rcu_read_lock.
2049 struct task_and_cgroup ent;
2051 /* @tsk either already exited or can't exit until the end */
2052 if (tsk->flags & PF_EXITING)
2055 /* as per above, nr_threads may decrease, but not increase. */
2056 BUG_ON(i >= group_size);
2058 ent.cgrp = task_cgroup_from_root(tsk, root);
2059 /* nothing to do if this task is already in the cgroup */
2060 if (ent.cgrp == cgrp)
2063 * saying GFP_ATOMIC has no effect here because we did prealloc
2064 * earlier, but it's good form to communicate our expectations.
2066 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
2067 BUG_ON(retval != 0);
2072 } while_each_thread(leader, tsk);
2074 /* remember the number of threads in the array for later. */
2076 tset.tc_array = group;
2077 tset.tc_array_len = group_size;
2079 /* methods shouldn't be called if no task is actually migrating */
2082 goto out_free_group_list;
2085 * step 1: check that we can legitimately attach to the cgroup.
2087 for_each_root_subsys(root, ss) {
2088 if (ss->can_attach) {
2089 retval = ss->can_attach(cgrp, &tset);
2092 goto out_cancel_attach;
2098 * step 2: make sure css_sets exist for all threads to be migrated.
2099 * we use find_css_set, which allocates a new one if necessary.
2101 for (i = 0; i < group_size; i++) {
2102 struct css_set *old_cset;
2104 tc = flex_array_get(group, i);
2105 old_cset = task_css_set(tc->task);
2106 tc->cg = find_css_set(old_cset, cgrp);
2109 goto out_put_css_set_refs;
2114 * step 3: now that we're guaranteed success wrt the css_sets,
2115 * proceed to move all tasks to the new cgroup. There are no
2116 * failure cases after here, so this is the commit point.
2118 for (i = 0; i < group_size; i++) {
2119 tc = flex_array_get(group, i);
2120 cgroup_task_migrate(tc->cgrp, tc->task, tc->cg);
2122 /* nothing is sensitive to fork() after this point. */
2125 * step 4: do subsystem attach callbacks.
2127 for_each_root_subsys(root, ss) {
2129 ss->attach(cgrp, &tset);
2133 * step 5: success! and cleanup
2136 out_put_css_set_refs:
2138 for (i = 0; i < group_size; i++) {
2139 tc = flex_array_get(group, i);
2142 put_css_set(tc->cg);
2147 for_each_root_subsys(root, ss) {
2148 if (ss == failed_ss)
2150 if (ss->cancel_attach)
2151 ss->cancel_attach(cgrp, &tset);
2154 out_free_group_list:
2155 flex_array_free(group);
2160 * Find the task_struct of the task to attach by vpid and pass it along to the
2161 * function to attach either it or all tasks in its threadgroup. Will lock
2162 * cgroup_mutex and threadgroup; may take task_lock of task.
2164 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2166 struct task_struct *tsk;
2167 const struct cred *cred = current_cred(), *tcred;
2170 if (!cgroup_lock_live_group(cgrp))
2176 tsk = find_task_by_vpid(pid);
2180 goto out_unlock_cgroup;
2183 * even if we're attaching all tasks in the thread group, we
2184 * only need to check permissions on one of them.
2186 tcred = __task_cred(tsk);
2187 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2188 !uid_eq(cred->euid, tcred->uid) &&
2189 !uid_eq(cred->euid, tcred->suid)) {
2192 goto out_unlock_cgroup;
2198 tsk = tsk->group_leader;
2201 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2202 * trapped in a cpuset, or RT worker may be born in a cgroup
2203 * with no rt_runtime allocated. Just say no.
2205 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2208 goto out_unlock_cgroup;
2211 get_task_struct(tsk);
2214 threadgroup_lock(tsk);
2216 if (!thread_group_leader(tsk)) {
2218 * a race with de_thread from another thread's exec()
2219 * may strip us of our leadership, if this happens,
2220 * there is no choice but to throw this task away and
2221 * try again; this is
2222 * "double-double-toil-and-trouble-check locking".
2224 threadgroup_unlock(tsk);
2225 put_task_struct(tsk);
2226 goto retry_find_task;
2230 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2232 threadgroup_unlock(tsk);
2234 put_task_struct(tsk);
2236 mutex_unlock(&cgroup_mutex);
2241 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2242 * @from: attach to all cgroups of a given task
2243 * @tsk: the task to be attached
2245 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2247 struct cgroupfs_root *root;
2250 mutex_lock(&cgroup_mutex);
2251 for_each_active_root(root) {
2252 struct cgroup *from_cg = task_cgroup_from_root(from, root);
2254 retval = cgroup_attach_task(from_cg, tsk, false);
2258 mutex_unlock(&cgroup_mutex);
2262 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2264 static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
2266 return attach_task_by_pid(cgrp, pid, false);
2269 static int cgroup_procs_write(struct cgroup *cgrp, struct cftype *cft, u64 tgid)
2271 return attach_task_by_pid(cgrp, tgid, true);
2274 static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
2277 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2278 if (strlen(buffer) >= PATH_MAX)
2280 if (!cgroup_lock_live_group(cgrp))
2282 mutex_lock(&cgroup_root_mutex);
2283 strcpy(cgrp->root->release_agent_path, buffer);
2284 mutex_unlock(&cgroup_root_mutex);
2285 mutex_unlock(&cgroup_mutex);
2289 static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
2290 struct seq_file *seq)
2292 if (!cgroup_lock_live_group(cgrp))
2294 seq_puts(seq, cgrp->root->release_agent_path);
2295 seq_putc(seq, '\n');
2296 mutex_unlock(&cgroup_mutex);
2300 static int cgroup_sane_behavior_show(struct cgroup *cgrp, struct cftype *cft,
2301 struct seq_file *seq)
2303 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2307 /* A buffer size big enough for numbers or short strings */
2308 #define CGROUP_LOCAL_BUFFER_SIZE 64
2310 static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
2312 const char __user *userbuf,
2313 size_t nbytes, loff_t *unused_ppos)
2315 char buffer[CGROUP_LOCAL_BUFFER_SIZE];
2321 if (nbytes >= sizeof(buffer))
2323 if (copy_from_user(buffer, userbuf, nbytes))
2326 buffer[nbytes] = 0; /* nul-terminate */
2327 if (cft->write_u64) {
2328 u64 val = simple_strtoull(strstrip(buffer), &end, 0);
2331 retval = cft->write_u64(cgrp, cft, val);
2333 s64 val = simple_strtoll(strstrip(buffer), &end, 0);
2336 retval = cft->write_s64(cgrp, cft, val);
2343 static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
2345 const char __user *userbuf,
2346 size_t nbytes, loff_t *unused_ppos)
2348 char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
2350 size_t max_bytes = cft->max_write_len;
2351 char *buffer = local_buffer;
2354 max_bytes = sizeof(local_buffer) - 1;
2355 if (nbytes >= max_bytes)
2357 /* Allocate a dynamic buffer if we need one */
2358 if (nbytes >= sizeof(local_buffer)) {
2359 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
2363 if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
2368 buffer[nbytes] = 0; /* nul-terminate */
2369 retval = cft->write_string(cgrp, cft, strstrip(buffer));
2373 if (buffer != local_buffer)
2378 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
2379 size_t nbytes, loff_t *ppos)
2381 struct cftype *cft = __d_cft(file->f_dentry);
2382 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2384 if (cgroup_is_dead(cgrp))
2387 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
2388 if (cft->write_u64 || cft->write_s64)
2389 return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
2390 if (cft->write_string)
2391 return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
2393 int ret = cft->trigger(cgrp, (unsigned int)cft->private);
2394 return ret ? ret : nbytes;
2399 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
2401 char __user *buf, size_t nbytes,
2404 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2405 u64 val = cft->read_u64(cgrp, cft);
2406 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
2408 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2411 static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
2413 char __user *buf, size_t nbytes,
2416 char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2417 s64 val = cft->read_s64(cgrp, cft);
2418 int len = sprintf(tmp, "%lld\n", (long long) val);
2420 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2423 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2424 size_t nbytes, loff_t *ppos)
2426 struct cftype *cft = __d_cft(file->f_dentry);
2427 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2429 if (cgroup_is_dead(cgrp))
2433 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2435 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2437 return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2442 * seqfile ops/methods for returning structured data. Currently just
2443 * supports string->u64 maps, but can be extended in future.
2446 struct cgroup_seqfile_state {
2448 struct cgroup *cgroup;
2451 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2453 struct seq_file *sf = cb->state;
2454 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2457 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2459 struct cgroup_seqfile_state *state = m->private;
2460 struct cftype *cft = state->cft;
2461 if (cft->read_map) {
2462 struct cgroup_map_cb cb = {
2463 .fill = cgroup_map_add,
2466 return cft->read_map(state->cgroup, cft, &cb);
2468 return cft->read_seq_string(state->cgroup, cft, m);
2471 static int cgroup_seqfile_release(struct inode *inode, struct file *file)
2473 struct seq_file *seq = file->private_data;
2474 kfree(seq->private);
2475 return single_release(inode, file);
2478 static const struct file_operations cgroup_seqfile_operations = {
2480 .write = cgroup_file_write,
2481 .llseek = seq_lseek,
2482 .release = cgroup_seqfile_release,
2485 static int cgroup_file_open(struct inode *inode, struct file *file)
2490 err = generic_file_open(inode, file);
2493 cft = __d_cft(file->f_dentry);
2495 if (cft->read_map || cft->read_seq_string) {
2496 struct cgroup_seqfile_state *state;
2498 state = kzalloc(sizeof(*state), GFP_USER);
2503 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
2504 file->f_op = &cgroup_seqfile_operations;
2505 err = single_open(file, cgroup_seqfile_show, state);
2508 } else if (cft->open)
2509 err = cft->open(inode, file);
2516 static int cgroup_file_release(struct inode *inode, struct file *file)
2518 struct cftype *cft = __d_cft(file->f_dentry);
2520 return cft->release(inode, file);
2525 * cgroup_rename - Only allow simple rename of directories in place.
2527 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2528 struct inode *new_dir, struct dentry *new_dentry)
2531 struct cgroup_name *name, *old_name;
2532 struct cgroup *cgrp;
2535 * It's convinient to use parent dir's i_mutex to protected
2538 lockdep_assert_held(&old_dir->i_mutex);
2540 if (!S_ISDIR(old_dentry->d_inode->i_mode))
2542 if (new_dentry->d_inode)
2544 if (old_dir != new_dir)
2547 cgrp = __d_cgrp(old_dentry);
2550 * This isn't a proper migration and its usefulness is very
2551 * limited. Disallow if sane_behavior.
2553 if (cgroup_sane_behavior(cgrp))
2556 name = cgroup_alloc_name(new_dentry);
2560 ret = simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2566 old_name = rcu_dereference_protected(cgrp->name, true);
2567 rcu_assign_pointer(cgrp->name, name);
2569 kfree_rcu(old_name, rcu_head);
2573 static struct simple_xattrs *__d_xattrs(struct dentry *dentry)
2575 if (S_ISDIR(dentry->d_inode->i_mode))
2576 return &__d_cgrp(dentry)->xattrs;
2578 return &__d_cfe(dentry)->xattrs;
2581 static inline int xattr_enabled(struct dentry *dentry)
2583 struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
2584 return root->flags & CGRP_ROOT_XATTR;
2587 static bool is_valid_xattr(const char *name)
2589 if (!strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
2590 !strncmp(name, XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN))
2595 static int cgroup_setxattr(struct dentry *dentry, const char *name,
2596 const void *val, size_t size, int flags)
2598 if (!xattr_enabled(dentry))
2600 if (!is_valid_xattr(name))
2602 return simple_xattr_set(__d_xattrs(dentry), name, val, size, flags);
2605 static int cgroup_removexattr(struct dentry *dentry, const char *name)
2607 if (!xattr_enabled(dentry))
2609 if (!is_valid_xattr(name))
2611 return simple_xattr_remove(__d_xattrs(dentry), name);
2614 static ssize_t cgroup_getxattr(struct dentry *dentry, const char *name,
2615 void *buf, size_t size)
2617 if (!xattr_enabled(dentry))
2619 if (!is_valid_xattr(name))
2621 return simple_xattr_get(__d_xattrs(dentry), name, buf, size);
2624 static ssize_t cgroup_listxattr(struct dentry *dentry, char *buf, size_t size)
2626 if (!xattr_enabled(dentry))
2628 return simple_xattr_list(__d_xattrs(dentry), buf, size);
2631 static const struct file_operations cgroup_file_operations = {
2632 .read = cgroup_file_read,
2633 .write = cgroup_file_write,
2634 .llseek = generic_file_llseek,
2635 .open = cgroup_file_open,
2636 .release = cgroup_file_release,
2639 static const struct inode_operations cgroup_file_inode_operations = {
2640 .setxattr = cgroup_setxattr,
2641 .getxattr = cgroup_getxattr,
2642 .listxattr = cgroup_listxattr,
2643 .removexattr = cgroup_removexattr,
2646 static const struct inode_operations cgroup_dir_inode_operations = {
2647 .lookup = cgroup_lookup,
2648 .mkdir = cgroup_mkdir,
2649 .rmdir = cgroup_rmdir,
2650 .rename = cgroup_rename,
2651 .setxattr = cgroup_setxattr,
2652 .getxattr = cgroup_getxattr,
2653 .listxattr = cgroup_listxattr,
2654 .removexattr = cgroup_removexattr,
2657 static struct dentry *cgroup_lookup(struct inode *dir, struct dentry *dentry, unsigned int flags)
2659 if (dentry->d_name.len > NAME_MAX)
2660 return ERR_PTR(-ENAMETOOLONG);
2661 d_add(dentry, NULL);
2666 * Check if a file is a control file
2668 static inline struct cftype *__file_cft(struct file *file)
2670 if (file_inode(file)->i_fop != &cgroup_file_operations)
2671 return ERR_PTR(-EINVAL);
2672 return __d_cft(file->f_dentry);
2675 static int cgroup_create_file(struct dentry *dentry, umode_t mode,
2676 struct super_block *sb)
2678 struct inode *inode;
2682 if (dentry->d_inode)
2685 inode = cgroup_new_inode(mode, sb);
2689 if (S_ISDIR(mode)) {
2690 inode->i_op = &cgroup_dir_inode_operations;
2691 inode->i_fop = &simple_dir_operations;
2693 /* start off with i_nlink == 2 (for "." entry) */
2695 inc_nlink(dentry->d_parent->d_inode);
2698 * Control reaches here with cgroup_mutex held.
2699 * @inode->i_mutex should nest outside cgroup_mutex but we
2700 * want to populate it immediately without releasing
2701 * cgroup_mutex. As @inode isn't visible to anyone else
2702 * yet, trylock will always succeed without affecting
2705 WARN_ON_ONCE(!mutex_trylock(&inode->i_mutex));
2706 } else if (S_ISREG(mode)) {
2708 inode->i_fop = &cgroup_file_operations;
2709 inode->i_op = &cgroup_file_inode_operations;
2711 d_instantiate(dentry, inode);
2712 dget(dentry); /* Extra count - pin the dentry in core */
2717 * cgroup_file_mode - deduce file mode of a control file
2718 * @cft: the control file in question
2720 * returns cft->mode if ->mode is not 0
2721 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2722 * returns S_IRUGO if it has only a read handler
2723 * returns S_IWUSR if it has only a write hander
2725 static umode_t cgroup_file_mode(const struct cftype *cft)
2732 if (cft->read || cft->read_u64 || cft->read_s64 ||
2733 cft->read_map || cft->read_seq_string)
2736 if (cft->write || cft->write_u64 || cft->write_s64 ||
2737 cft->write_string || cft->trigger)
2743 static int cgroup_add_file(struct cgroup *cgrp, struct cgroup_subsys *subsys,
2746 struct dentry *dir = cgrp->dentry;
2747 struct cgroup *parent = __d_cgrp(dir);
2748 struct dentry *dentry;
2752 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2754 if (subsys && !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
2755 strcpy(name, subsys->name);
2758 strcat(name, cft->name);
2760 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2762 cfe = kzalloc(sizeof(*cfe), GFP_KERNEL);
2766 dentry = lookup_one_len(name, dir, strlen(name));
2767 if (IS_ERR(dentry)) {
2768 error = PTR_ERR(dentry);
2772 cfe->type = (void *)cft;
2773 cfe->dentry = dentry;
2774 dentry->d_fsdata = cfe;
2775 simple_xattrs_init(&cfe->xattrs);
2777 mode = cgroup_file_mode(cft);
2778 error = cgroup_create_file(dentry, mode | S_IFREG, cgrp->root->sb);
2780 list_add_tail(&cfe->node, &parent->files);
2790 * cgroup_addrm_files - add or remove files to a cgroup directory
2791 * @cgrp: the target cgroup
2792 * @subsys: the subsystem of files to be added
2793 * @cfts: array of cftypes to be added
2794 * @is_add: whether to add or remove
2796 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2797 * All @cfts should belong to @subsys. For removals, this function never
2798 * fails. If addition fails, this function doesn't remove files already
2799 * added. The caller is responsible for cleaning up.
2801 static int cgroup_addrm_files(struct cgroup *cgrp, struct cgroup_subsys *subsys,
2802 struct cftype cfts[], bool is_add)
2807 lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
2808 lockdep_assert_held(&cgroup_mutex);
2810 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2811 /* does cft->flags tell us to skip this file on @cgrp? */
2812 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2814 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2816 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2820 ret = cgroup_add_file(cgrp, subsys, cft);
2822 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2827 cgroup_rm_file(cgrp, cft);
2833 static void cgroup_cfts_prepare(void)
2834 __acquires(&cgroup_mutex)
2837 * Thanks to the entanglement with vfs inode locking, we can't walk
2838 * the existing cgroups under cgroup_mutex and create files.
2839 * Instead, we use cgroup_for_each_descendant_pre() and drop RCU
2840 * read lock before calling cgroup_addrm_files().
2842 mutex_lock(&cgroup_mutex);
2845 static int cgroup_cfts_commit(struct cgroup_subsys *ss,
2846 struct cftype *cfts, bool is_add)
2847 __releases(&cgroup_mutex)
2850 struct cgroup *cgrp, *root = &ss->root->top_cgroup;
2851 struct super_block *sb = ss->root->sb;
2852 struct dentry *prev = NULL;
2853 struct inode *inode;
2857 /* %NULL @cfts indicates abort and don't bother if @ss isn't attached */
2858 if (!cfts || ss->root == &cgroup_dummy_root ||
2859 !atomic_inc_not_zero(&sb->s_active)) {
2860 mutex_unlock(&cgroup_mutex);
2865 * All cgroups which are created after we drop cgroup_mutex will
2866 * have the updated set of files, so we only need to update the
2867 * cgroups created before the current @cgroup_serial_nr_next.
2869 update_before = cgroup_serial_nr_next;
2871 mutex_unlock(&cgroup_mutex);
2873 /* @root always needs to be updated */
2874 inode = root->dentry->d_inode;
2875 mutex_lock(&inode->i_mutex);
2876 mutex_lock(&cgroup_mutex);
2877 ret = cgroup_addrm_files(root, ss, cfts, is_add);
2878 mutex_unlock(&cgroup_mutex);
2879 mutex_unlock(&inode->i_mutex);
2884 /* add/rm files for all cgroups created before */
2886 cgroup_for_each_descendant_pre(cgrp, root) {
2887 if (cgroup_is_dead(cgrp))
2890 inode = cgrp->dentry->d_inode;
2895 prev = cgrp->dentry;
2897 mutex_lock(&inode->i_mutex);
2898 mutex_lock(&cgroup_mutex);
2899 if (cgrp->serial_nr < update_before && !cgroup_is_dead(cgrp))
2900 ret = cgroup_addrm_files(cgrp, ss, cfts, is_add);
2901 mutex_unlock(&cgroup_mutex);
2902 mutex_unlock(&inode->i_mutex);
2911 deactivate_super(sb);
2916 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2917 * @ss: target cgroup subsystem
2918 * @cfts: zero-length name terminated array of cftypes
2920 * Register @cfts to @ss. Files described by @cfts are created for all
2921 * existing cgroups to which @ss is attached and all future cgroups will
2922 * have them too. This function can be called anytime whether @ss is
2925 * Returns 0 on successful registration, -errno on failure. Note that this
2926 * function currently returns 0 as long as @cfts registration is successful
2927 * even if some file creation attempts on existing cgroups fail.
2929 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2931 struct cftype_set *set;
2934 set = kzalloc(sizeof(*set), GFP_KERNEL);
2938 cgroup_cfts_prepare();
2940 list_add_tail(&set->node, &ss->cftsets);
2941 ret = cgroup_cfts_commit(ss, cfts, true);
2943 cgroup_rm_cftypes(ss, cfts);
2946 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2949 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2950 * @ss: target cgroup subsystem
2951 * @cfts: zero-length name terminated array of cftypes
2953 * Unregister @cfts from @ss. Files described by @cfts are removed from
2954 * all existing cgroups to which @ss is attached and all future cgroups
2955 * won't have them either. This function can be called anytime whether @ss
2956 * is attached or not.
2958 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2959 * registered with @ss.
2961 int cgroup_rm_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2963 struct cftype_set *set;
2965 cgroup_cfts_prepare();
2967 list_for_each_entry(set, &ss->cftsets, node) {
2968 if (set->cfts == cfts) {
2969 list_del(&set->node);
2971 cgroup_cfts_commit(ss, cfts, false);
2976 cgroup_cfts_commit(ss, NULL, false);
2981 * cgroup_task_count - count the number of tasks in a cgroup.
2982 * @cgrp: the cgroup in question
2984 * Return the number of tasks in the cgroup.
2986 int cgroup_task_count(const struct cgroup *cgrp)
2989 struct cgrp_cset_link *link;
2991 read_lock(&css_set_lock);
2992 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2993 count += atomic_read(&link->cset->refcount);
2994 read_unlock(&css_set_lock);
2999 * Advance a list_head iterator. The iterator should be positioned at
3000 * the start of a css_set
3002 static void cgroup_advance_iter(struct cgroup *cgrp, struct cgroup_iter *it)
3004 struct list_head *l = it->cset_link;
3005 struct cgrp_cset_link *link;
3006 struct css_set *cset;
3008 /* Advance to the next non-empty css_set */
3011 if (l == &cgrp->cset_links) {
3012 it->cset_link = NULL;
3015 link = list_entry(l, struct cgrp_cset_link, cset_link);
3017 } while (list_empty(&cset->tasks));
3019 it->task = cset->tasks.next;
3023 * To reduce the fork() overhead for systems that are not actually
3024 * using their cgroups capability, we don't maintain the lists running
3025 * through each css_set to its tasks until we see the list actually
3026 * used - in other words after the first call to cgroup_iter_start().
3028 static void cgroup_enable_task_cg_lists(void)
3030 struct task_struct *p, *g;
3031 write_lock(&css_set_lock);
3032 use_task_css_set_links = 1;
3034 * We need tasklist_lock because RCU is not safe against
3035 * while_each_thread(). Besides, a forking task that has passed
3036 * cgroup_post_fork() without seeing use_task_css_set_links = 1
3037 * is not guaranteed to have its child immediately visible in the
3038 * tasklist if we walk through it with RCU.
3040 read_lock(&tasklist_lock);
3041 do_each_thread(g, p) {
3044 * We should check if the process is exiting, otherwise
3045 * it will race with cgroup_exit() in that the list
3046 * entry won't be deleted though the process has exited.
3048 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
3049 list_add(&p->cg_list, &task_css_set(p)->tasks);
3051 } while_each_thread(g, p);
3052 read_unlock(&tasklist_lock);
3053 write_unlock(&css_set_lock);
3057 * cgroup_next_sibling - find the next sibling of a given cgroup
3058 * @pos: the current cgroup
3060 * This function returns the next sibling of @pos and should be called
3061 * under RCU read lock. The only requirement is that @pos is accessible.
3062 * The next sibling is guaranteed to be returned regardless of @pos's
3065 struct cgroup *cgroup_next_sibling(struct cgroup *pos)
3067 struct cgroup *next;
3069 WARN_ON_ONCE(!rcu_read_lock_held());
3072 * @pos could already have been removed. Once a cgroup is removed,
3073 * its ->sibling.next is no longer updated when its next sibling
3074 * changes. As CGRP_DEAD assertion is serialized and happens
3075 * before the cgroup is taken off the ->sibling list, if we see it
3076 * unasserted, it's guaranteed that the next sibling hasn't
3077 * finished its grace period even if it's already removed, and thus
3078 * safe to dereference from this RCU critical section. If
3079 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
3080 * to be visible as %true here.
3082 if (likely(!cgroup_is_dead(pos))) {
3083 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
3084 if (&next->sibling != &pos->parent->children)
3090 * Can't dereference the next pointer. Each cgroup is given a
3091 * monotonically increasing unique serial number and always
3092 * appended to the sibling list, so the next one can be found by
3093 * walking the parent's children until we see a cgroup with higher
3094 * serial number than @pos's.
3096 * While this path can be slow, it's taken only when either the
3097 * current cgroup is removed or iteration and removal race.
3099 list_for_each_entry_rcu(next, &pos->parent->children, sibling)
3100 if (next->serial_nr > pos->serial_nr)
3104 EXPORT_SYMBOL_GPL(cgroup_next_sibling);
3107 * cgroup_next_descendant_pre - find the next descendant for pre-order walk
3108 * @pos: the current position (%NULL to initiate traversal)
3109 * @cgroup: cgroup whose descendants to walk
3111 * To be used by cgroup_for_each_descendant_pre(). Find the next
3112 * descendant to visit for pre-order traversal of @cgroup's descendants.
3114 * While this function requires RCU read locking, it doesn't require the
3115 * whole traversal to be contained in a single RCU critical section. This
3116 * function will return the correct next descendant as long as both @pos
3117 * and @cgroup are accessible and @pos is a descendant of @cgroup.
3119 struct cgroup *cgroup_next_descendant_pre(struct cgroup *pos,
3120 struct cgroup *cgroup)
3122 struct cgroup *next;
3124 WARN_ON_ONCE(!rcu_read_lock_held());
3126 /* if first iteration, pretend we just visited @cgroup */
3130 /* visit the first child if exists */
3131 next = list_first_or_null_rcu(&pos->children, struct cgroup, sibling);
3135 /* no child, visit my or the closest ancestor's next sibling */
3136 while (pos != cgroup) {
3137 next = cgroup_next_sibling(pos);
3145 EXPORT_SYMBOL_GPL(cgroup_next_descendant_pre);
3148 * cgroup_rightmost_descendant - return the rightmost descendant of a cgroup
3149 * @pos: cgroup of interest
3151 * Return the rightmost descendant of @pos. If there's no descendant,
3152 * @pos is returned. This can be used during pre-order traversal to skip
3155 * While this function requires RCU read locking, it doesn't require the
3156 * whole traversal to be contained in a single RCU critical section. This
3157 * function will return the correct rightmost descendant as long as @pos is
3160 struct cgroup *cgroup_rightmost_descendant(struct cgroup *pos)
3162 struct cgroup *last, *tmp;
3164 WARN_ON_ONCE(!rcu_read_lock_held());
3168 /* ->prev isn't RCU safe, walk ->next till the end */
3170 list_for_each_entry_rcu(tmp, &last->children, sibling)
3176 EXPORT_SYMBOL_GPL(cgroup_rightmost_descendant);
3178 static struct cgroup *cgroup_leftmost_descendant(struct cgroup *pos)
3180 struct cgroup *last;
3184 pos = list_first_or_null_rcu(&pos->children, struct cgroup,
3192 * cgroup_next_descendant_post - find the next descendant for post-order walk
3193 * @pos: the current position (%NULL to initiate traversal)
3194 * @cgroup: cgroup whose descendants to walk
3196 * To be used by cgroup_for_each_descendant_post(). Find the next
3197 * descendant to visit for post-order traversal of @cgroup's descendants.
3199 * While this function requires RCU read locking, it doesn't require the
3200 * whole traversal to be contained in a single RCU critical section. This
3201 * function will return the correct next descendant as long as both @pos
3202 * and @cgroup are accessible and @pos is a descendant of @cgroup.
3204 struct cgroup *cgroup_next_descendant_post(struct cgroup *pos,
3205 struct cgroup *cgroup)
3207 struct cgroup *next;
3209 WARN_ON_ONCE(!rcu_read_lock_held());
3211 /* if first iteration, visit the leftmost descendant */
3213 next = cgroup_leftmost_descendant(cgroup);
3214 return next != cgroup ? next : NULL;
3217 /* if there's an unvisited sibling, visit its leftmost descendant */
3218 next = cgroup_next_sibling(pos);
3220 return cgroup_leftmost_descendant(next);
3222 /* no sibling left, visit parent */
3224 return next != cgroup ? next : NULL;
3226 EXPORT_SYMBOL_GPL(cgroup_next_descendant_post);
3228 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
3229 __acquires(css_set_lock)
3232 * The first time anyone tries to iterate across a cgroup,
3233 * we need to enable the list linking each css_set to its
3234 * tasks, and fix up all existing tasks.
3236 if (!use_task_css_set_links)
3237 cgroup_enable_task_cg_lists();
3239 read_lock(&css_set_lock);
3240 it->cset_link = &cgrp->cset_links;
3241 cgroup_advance_iter(cgrp, it);
3244 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
3245 struct cgroup_iter *it)
3247 struct task_struct *res;
3248 struct list_head *l = it->task;
3249 struct cgrp_cset_link *link;
3251 /* If the iterator cg is NULL, we have no tasks */
3254 res = list_entry(l, struct task_struct, cg_list);
3255 /* Advance iterator to find next entry */
3257 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
3258 if (l == &link->cset->tasks) {
3259 /* We reached the end of this task list - move on to
3260 * the next cg_cgroup_link */
3261 cgroup_advance_iter(cgrp, it);
3268 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
3269 __releases(css_set_lock)
3271 read_unlock(&css_set_lock);
3274 static inline int started_after_time(struct task_struct *t1,
3275 struct timespec *time,
3276 struct task_struct *t2)
3278 int start_diff = timespec_compare(&t1->start_time, time);
3279 if (start_diff > 0) {
3281 } else if (start_diff < 0) {
3285 * Arbitrarily, if two processes started at the same
3286 * time, we'll say that the lower pointer value
3287 * started first. Note that t2 may have exited by now
3288 * so this may not be a valid pointer any longer, but
3289 * that's fine - it still serves to distinguish
3290 * between two tasks started (effectively) simultaneously.
3297 * This function is a callback from heap_insert() and is used to order
3299 * In this case we order the heap in descending task start time.
3301 static inline int started_after(void *p1, void *p2)
3303 struct task_struct *t1 = p1;
3304 struct task_struct *t2 = p2;
3305 return started_after_time(t1, &t2->start_time, t2);
3309 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
3310 * @scan: struct cgroup_scanner containing arguments for the scan
3312 * Arguments include pointers to callback functions test_task() and
3314 * Iterate through all the tasks in a cgroup, calling test_task() for each,
3315 * and if it returns true, call process_task() for it also.
3316 * The test_task pointer may be NULL, meaning always true (select all tasks).
3317 * Effectively duplicates cgroup_iter_{start,next,end}()
3318 * but does not lock css_set_lock for the call to process_task().
3319 * The struct cgroup_scanner may be embedded in any structure of the caller's
3321 * It is guaranteed that process_task() will act on every task that
3322 * is a member of the cgroup for the duration of this call. This
3323 * function may or may not call process_task() for tasks that exit
3324 * or move to a different cgroup during the call, or are forked or
3325 * move into the cgroup during the call.
3327 * Note that test_task() may be called with locks held, and may in some
3328 * situations be called multiple times for the same task, so it should
3330 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
3331 * pre-allocated and will be used for heap operations (and its "gt" member will
3332 * be overwritten), else a temporary heap will be used (allocation of which
3333 * may cause this function to fail).
3335 int cgroup_scan_tasks(struct cgroup_scanner *scan)
3338 struct cgroup_iter it;
3339 struct task_struct *p, *dropped;
3340 /* Never dereference latest_task, since it's not refcounted */
3341 struct task_struct *latest_task = NULL;
3342 struct ptr_heap tmp_heap;
3343 struct ptr_heap *heap;
3344 struct timespec latest_time = { 0, 0 };
3347 /* The caller supplied our heap and pre-allocated its memory */
3349 heap->gt = &started_after;
3351 /* We need to allocate our own heap memory */
3353 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
3355 /* cannot allocate the heap */
3361 * Scan tasks in the cgroup, using the scanner's "test_task" callback
3362 * to determine which are of interest, and using the scanner's
3363 * "process_task" callback to process any of them that need an update.
3364 * Since we don't want to hold any locks during the task updates,
3365 * gather tasks to be processed in a heap structure.
3366 * The heap is sorted by descending task start time.
3367 * If the statically-sized heap fills up, we overflow tasks that
3368 * started later, and in future iterations only consider tasks that
3369 * started after the latest task in the previous pass. This
3370 * guarantees forward progress and that we don't miss any tasks.
3373 cgroup_iter_start(scan->cg, &it);
3374 while ((p = cgroup_iter_next(scan->cg, &it))) {
3376 * Only affect tasks that qualify per the caller's callback,
3377 * if he provided one
3379 if (scan->test_task && !scan->test_task(p, scan))
3382 * Only process tasks that started after the last task
3385 if (!started_after_time(p, &latest_time, latest_task))
3387 dropped = heap_insert(heap, p);
3388 if (dropped == NULL) {
3390 * The new task was inserted; the heap wasn't
3394 } else if (dropped != p) {
3396 * The new task was inserted, and pushed out a
3400 put_task_struct(dropped);
3403 * Else the new task was newer than anything already in
3404 * the heap and wasn't inserted
3407 cgroup_iter_end(scan->cg, &it);
3410 for (i = 0; i < heap->size; i++) {
3411 struct task_struct *q = heap->ptrs[i];
3413 latest_time = q->start_time;
3416 /* Process the task per the caller's callback */
3417 scan->process_task(q, scan);
3421 * If we had to process any tasks at all, scan again
3422 * in case some of them were in the middle of forking
3423 * children that didn't get processed.
3424 * Not the most efficient way to do it, but it avoids
3425 * having to take callback_mutex in the fork path
3429 if (heap == &tmp_heap)
3430 heap_free(&tmp_heap);
3434 static void cgroup_transfer_one_task(struct task_struct *task,
3435 struct cgroup_scanner *scan)
3437 struct cgroup *new_cgroup = scan->data;
3439 mutex_lock(&cgroup_mutex);
3440 cgroup_attach_task(new_cgroup, task, false);
3441 mutex_unlock(&cgroup_mutex);
3445 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3446 * @to: cgroup to which the tasks will be moved
3447 * @from: cgroup in which the tasks currently reside
3449 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3451 struct cgroup_scanner scan;
3454 scan.test_task = NULL; /* select all tasks in cgroup */
3455 scan.process_task = cgroup_transfer_one_task;
3459 return cgroup_scan_tasks(&scan);
3463 * Stuff for reading the 'tasks'/'procs' files.
3465 * Reading this file can return large amounts of data if a cgroup has
3466 * *lots* of attached tasks. So it may need several calls to read(),
3467 * but we cannot guarantee that the information we produce is correct
3468 * unless we produce it entirely atomically.
3472 /* which pidlist file are we talking about? */
3473 enum cgroup_filetype {
3479 * A pidlist is a list of pids that virtually represents the contents of one
3480 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3481 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3484 struct cgroup_pidlist {
3486 * used to find which pidlist is wanted. doesn't change as long as
3487 * this particular list stays in the list.
3489 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3492 /* how many elements the above list has */
3494 /* how many files are using the current array */
3496 /* each of these stored in a list by its cgroup */
3497 struct list_head links;
3498 /* pointer to the cgroup we belong to, for list removal purposes */
3499 struct cgroup *owner;
3500 /* protects the other fields */
3501 struct rw_semaphore mutex;
3505 * The following two functions "fix" the issue where there are more pids
3506 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3507 * TODO: replace with a kernel-wide solution to this problem
3509 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3510 static void *pidlist_allocate(int count)
3512 if (PIDLIST_TOO_LARGE(count))
3513 return vmalloc(count * sizeof(pid_t));
3515 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3517 static void pidlist_free(void *p)
3519 if (is_vmalloc_addr(p))
3526 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3527 * Returns the number of unique elements.
3529 static int pidlist_uniq(pid_t *list, int length)
3534 * we presume the 0th element is unique, so i starts at 1. trivial
3535 * edge cases first; no work needs to be done for either
3537 if (length == 0 || length == 1)
3539 /* src and dest walk down the list; dest counts unique elements */
3540 for (src = 1; src < length; src++) {
3541 /* find next unique element */
3542 while (list[src] == list[src-1]) {
3547 /* dest always points to where the next unique element goes */
3548 list[dest] = list[src];
3555 static int cmppid(const void *a, const void *b)
3557 return *(pid_t *)a - *(pid_t *)b;
3561 * find the appropriate pidlist for our purpose (given procs vs tasks)
3562 * returns with the lock on that pidlist already held, and takes care
3563 * of the use count, or returns NULL with no locks held if we're out of
3566 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3567 enum cgroup_filetype type)
3569 struct cgroup_pidlist *l;
3570 /* don't need task_nsproxy() if we're looking at ourself */
3571 struct pid_namespace *ns = task_active_pid_ns(current);
3574 * We can't drop the pidlist_mutex before taking the l->mutex in case
3575 * the last ref-holder is trying to remove l from the list at the same
3576 * time. Holding the pidlist_mutex precludes somebody taking whichever
3577 * list we find out from under us - compare release_pid_array().
3579 mutex_lock(&cgrp->pidlist_mutex);
3580 list_for_each_entry(l, &cgrp->pidlists, links) {
3581 if (l->key.type == type && l->key.ns == ns) {
3582 /* make sure l doesn't vanish out from under us */
3583 down_write(&l->mutex);
3584 mutex_unlock(&cgrp->pidlist_mutex);
3588 /* entry not found; create a new one */
3589 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3591 mutex_unlock(&cgrp->pidlist_mutex);
3594 init_rwsem(&l->mutex);
3595 down_write(&l->mutex);
3597 l->key.ns = get_pid_ns(ns);
3599 list_add(&l->links, &cgrp->pidlists);
3600 mutex_unlock(&cgrp->pidlist_mutex);
3605 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3607 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3608 struct cgroup_pidlist **lp)
3612 int pid, n = 0; /* used for populating the array */
3613 struct cgroup_iter it;
3614 struct task_struct *tsk;
3615 struct cgroup_pidlist *l;
3618 * If cgroup gets more users after we read count, we won't have
3619 * enough space - tough. This race is indistinguishable to the
3620 * caller from the case that the additional cgroup users didn't
3621 * show up until sometime later on.
3623 length = cgroup_task_count(cgrp);
3624 array = pidlist_allocate(length);
3627 /* now, populate the array */
3628 cgroup_iter_start(cgrp, &it);
3629 while ((tsk = cgroup_iter_next(cgrp, &it))) {
3630 if (unlikely(n == length))
3632 /* get tgid or pid for procs or tasks file respectively */
3633 if (type == CGROUP_FILE_PROCS)
3634 pid = task_tgid_vnr(tsk);
3636 pid = task_pid_vnr(tsk);
3637 if (pid > 0) /* make sure to only use valid results */
3640 cgroup_iter_end(cgrp, &it);
3642 /* now sort & (if procs) strip out duplicates */
3643 sort(array, length, sizeof(pid_t), cmppid, NULL);
3644 if (type == CGROUP_FILE_PROCS)
3645 length = pidlist_uniq(array, length);
3646 l = cgroup_pidlist_find(cgrp, type);
3648 pidlist_free(array);
3651 /* store array, freeing old if necessary - lock already held */
3652 pidlist_free(l->list);
3656 up_write(&l->mutex);
3662 * cgroupstats_build - build and fill cgroupstats
3663 * @stats: cgroupstats to fill information into
3664 * @dentry: A dentry entry belonging to the cgroup for which stats have
3667 * Build and fill cgroupstats so that taskstats can export it to user
3670 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3673 struct cgroup *cgrp;
3674 struct cgroup_iter it;
3675 struct task_struct *tsk;
3678 * Validate dentry by checking the superblock operations,
3679 * and make sure it's a directory.
3681 if (dentry->d_sb->s_op != &cgroup_ops ||
3682 !S_ISDIR(dentry->d_inode->i_mode))
3686 cgrp = dentry->d_fsdata;
3688 cgroup_iter_start(cgrp, &it);
3689 while ((tsk = cgroup_iter_next(cgrp, &it))) {
3690 switch (tsk->state) {
3692 stats->nr_running++;
3694 case TASK_INTERRUPTIBLE:
3695 stats->nr_sleeping++;
3697 case TASK_UNINTERRUPTIBLE:
3698 stats->nr_uninterruptible++;
3701 stats->nr_stopped++;
3704 if (delayacct_is_task_waiting_on_io(tsk))
3705 stats->nr_io_wait++;
3709 cgroup_iter_end(cgrp, &it);
3717 * seq_file methods for the tasks/procs files. The seq_file position is the
3718 * next pid to display; the seq_file iterator is a pointer to the pid
3719 * in the cgroup->l->list array.
3722 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3725 * Initially we receive a position value that corresponds to
3726 * one more than the last pid shown (or 0 on the first call or
3727 * after a seek to the start). Use a binary-search to find the
3728 * next pid to display, if any
3730 struct cgroup_pidlist *l = s->private;
3731 int index = 0, pid = *pos;
3734 down_read(&l->mutex);
3736 int end = l->length;
3738 while (index < end) {
3739 int mid = (index + end) / 2;
3740 if (l->list[mid] == pid) {
3743 } else if (l->list[mid] <= pid)
3749 /* If we're off the end of the array, we're done */
3750 if (index >= l->length)
3752 /* Update the abstract position to be the actual pid that we found */
3753 iter = l->list + index;
3758 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3760 struct cgroup_pidlist *l = s->private;
3764 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3766 struct cgroup_pidlist *l = s->private;
3768 pid_t *end = l->list + l->length;
3770 * Advance to the next pid in the array. If this goes off the
3782 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3784 return seq_printf(s, "%d\n", *(int *)v);
3788 * seq_operations functions for iterating on pidlists through seq_file -
3789 * independent of whether it's tasks or procs
3791 static const struct seq_operations cgroup_pidlist_seq_operations = {
3792 .start = cgroup_pidlist_start,
3793 .stop = cgroup_pidlist_stop,
3794 .next = cgroup_pidlist_next,
3795 .show = cgroup_pidlist_show,
3798 static void cgroup_release_pid_array(struct cgroup_pidlist *l)
3801 * the case where we're the last user of this particular pidlist will
3802 * have us remove it from the cgroup's list, which entails taking the
3803 * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
3804 * pidlist_mutex, we have to take pidlist_mutex first.
3806 mutex_lock(&l->owner->pidlist_mutex);
3807 down_write(&l->mutex);
3808 BUG_ON(!l->use_count);
3809 if (!--l->use_count) {
3810 /* we're the last user if refcount is 0; remove and free */
3811 list_del(&l->links);
3812 mutex_unlock(&l->owner->pidlist_mutex);
3813 pidlist_free(l->list);
3814 put_pid_ns(l->key.ns);
3815 up_write(&l->mutex);
3819 mutex_unlock(&l->owner->pidlist_mutex);
3820 up_write(&l->mutex);
3823 static int cgroup_pidlist_release(struct inode *inode, struct file *file)
3825 struct cgroup_pidlist *l;
3826 if (!(file->f_mode & FMODE_READ))
3829 * the seq_file will only be initialized if the file was opened for
3830 * reading; hence we check if it's not null only in that case.
3832 l = ((struct seq_file *)file->private_data)->private;
3833 cgroup_release_pid_array(l);
3834 return seq_release(inode, file);
3837 static const struct file_operations cgroup_pidlist_operations = {
3839 .llseek = seq_lseek,
3840 .write = cgroup_file_write,
3841 .release = cgroup_pidlist_release,
3845 * The following functions handle opens on a file that displays a pidlist
3846 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
3849 /* helper function for the two below it */
3850 static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
3852 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
3853 struct cgroup_pidlist *l;
3856 /* Nothing to do for write-only files */
3857 if (!(file->f_mode & FMODE_READ))
3860 /* have the array populated */
3861 retval = pidlist_array_load(cgrp, type, &l);
3864 /* configure file information */
3865 file->f_op = &cgroup_pidlist_operations;
3867 retval = seq_open(file, &cgroup_pidlist_seq_operations);
3869 cgroup_release_pid_array(l);
3872 ((struct seq_file *)file->private_data)->private = l;
3875 static int cgroup_tasks_open(struct inode *unused, struct file *file)
3877 return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
3879 static int cgroup_procs_open(struct inode *unused, struct file *file)
3881 return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
3884 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
3887 return notify_on_release(cgrp);
3890 static int cgroup_write_notify_on_release(struct cgroup *cgrp,
3894 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
3896 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3898 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3903 * When dput() is called asynchronously, if umount has been done and
3904 * then deactivate_super() in cgroup_free_fn() kills the superblock,
3905 * there's a small window that vfs will see the root dentry with non-zero
3906 * refcnt and trigger BUG().
3908 * That's why we hold a reference before dput() and drop it right after.
3910 static void cgroup_dput(struct cgroup *cgrp)
3912 struct super_block *sb = cgrp->root->sb;
3914 atomic_inc(&sb->s_active);
3916 deactivate_super(sb);
3920 * Unregister event and free resources.
3922 * Gets called from workqueue.
3924 static void cgroup_event_remove(struct work_struct *work)
3926 struct cgroup_event *event = container_of(work, struct cgroup_event,
3928 struct cgroup *cgrp = event->cgrp;
3930 remove_wait_queue(event->wqh, &event->wait);
3932 event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3934 /* Notify userspace the event is going away. */
3935 eventfd_signal(event->eventfd, 1);
3937 eventfd_ctx_put(event->eventfd);
3943 * Gets called on POLLHUP on eventfd when user closes it.
3945 * Called with wqh->lock held and interrupts disabled.
3947 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3948 int sync, void *key)
3950 struct cgroup_event *event = container_of(wait,
3951 struct cgroup_event, wait);
3952 struct cgroup *cgrp = event->cgrp;
3953 unsigned long flags = (unsigned long)key;
3955 if (flags & POLLHUP) {
3957 * If the event has been detached at cgroup removal, we
3958 * can simply return knowing the other side will cleanup
3961 * We can't race against event freeing since the other
3962 * side will require wqh->lock via remove_wait_queue(),
3965 spin_lock(&cgrp->event_list_lock);
3966 if (!list_empty(&event->list)) {
3967 list_del_init(&event->list);
3969 * We are in atomic context, but cgroup_event_remove()
3970 * may sleep, so we have to call it in workqueue.
3972 schedule_work(&event->remove);
3974 spin_unlock(&cgrp->event_list_lock);
3980 static void cgroup_event_ptable_queue_proc(struct file *file,
3981 wait_queue_head_t *wqh, poll_table *pt)
3983 struct cgroup_event *event = container_of(pt,
3984 struct cgroup_event, pt);
3987 add_wait_queue(wqh, &event->wait);
3991 * Parse input and register new cgroup event handler.
3993 * Input must be in format '<event_fd> <control_fd> <args>'.
3994 * Interpretation of args is defined by control file implementation.
3996 static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3999 struct cgroup_event *event = NULL;
4000 struct cgroup *cgrp_cfile;
4001 unsigned int efd, cfd;
4002 struct file *efile = NULL;
4003 struct file *cfile = NULL;
4007 efd = simple_strtoul(buffer, &endp, 10);
4012 cfd = simple_strtoul(buffer, &endp, 10);
4013 if ((*endp != ' ') && (*endp != '\0'))
4017 event = kzalloc(sizeof(*event), GFP_KERNEL);
4021 INIT_LIST_HEAD(&event->list);
4022 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
4023 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
4024 INIT_WORK(&event->remove, cgroup_event_remove);
4026 efile = eventfd_fget(efd);
4027 if (IS_ERR(efile)) {
4028 ret = PTR_ERR(efile);
4032 event->eventfd = eventfd_ctx_fileget(efile);
4033 if (IS_ERR(event->eventfd)) {
4034 ret = PTR_ERR(event->eventfd);
4044 /* the process need read permission on control file */
4045 /* AV: shouldn't we check that it's been opened for read instead? */
4046 ret = inode_permission(file_inode(cfile), MAY_READ);
4050 event->cft = __file_cft(cfile);
4051 if (IS_ERR(event->cft)) {
4052 ret = PTR_ERR(event->cft);
4057 * The file to be monitored must be in the same cgroup as
4058 * cgroup.event_control is.
4060 cgrp_cfile = __d_cgrp(cfile->f_dentry->d_parent);
4061 if (cgrp_cfile != cgrp) {
4066 if (!event->cft->register_event || !event->cft->unregister_event) {
4071 ret = event->cft->register_event(cgrp, event->cft,
4072 event->eventfd, buffer);
4076 efile->f_op->poll(efile, &event->pt);
4079 * Events should be removed after rmdir of cgroup directory, but before
4080 * destroying subsystem state objects. Let's take reference to cgroup
4081 * directory dentry to do that.
4085 spin_lock(&cgrp->event_list_lock);
4086 list_add(&event->list, &cgrp->event_list);
4087 spin_unlock(&cgrp->event_list_lock);
4098 if (event && event->eventfd && !IS_ERR(event->eventfd))
4099 eventfd_ctx_put(event->eventfd);
4101 if (!IS_ERR_OR_NULL(efile))
4109 static u64 cgroup_clone_children_read(struct cgroup *cgrp,
4112 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4115 static int cgroup_clone_children_write(struct cgroup *cgrp,
4120 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4122 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4126 static struct cftype cgroup_base_files[] = {
4128 .name = "cgroup.procs",
4129 .open = cgroup_procs_open,
4130 .write_u64 = cgroup_procs_write,
4131 .release = cgroup_pidlist_release,
4132 .mode = S_IRUGO | S_IWUSR,
4135 .name = "cgroup.event_control",
4136 .write_string = cgroup_write_event_control,
4140 .name = "cgroup.clone_children",
4141 .flags = CFTYPE_INSANE,
4142 .read_u64 = cgroup_clone_children_read,
4143 .write_u64 = cgroup_clone_children_write,
4146 .name = "cgroup.sane_behavior",
4147 .flags = CFTYPE_ONLY_ON_ROOT,
4148 .read_seq_string = cgroup_sane_behavior_show,
4152 * Historical crazy stuff. These don't have "cgroup." prefix and
4153 * don't exist if sane_behavior. If you're depending on these, be
4154 * prepared to be burned.
4158 .flags = CFTYPE_INSANE, /* use "procs" instead */
4159 .open = cgroup_tasks_open,
4160 .write_u64 = cgroup_tasks_write,
4161 .release = cgroup_pidlist_release,
4162 .mode = S_IRUGO | S_IWUSR,
4165 .name = "notify_on_release",
4166 .flags = CFTYPE_INSANE,
4167 .read_u64 = cgroup_read_notify_on_release,
4168 .write_u64 = cgroup_write_notify_on_release,
4171 .name = "release_agent",
4172 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
4173 .read_seq_string = cgroup_release_agent_show,
4174 .write_string = cgroup_release_agent_write,
4175 .max_write_len = PATH_MAX,
4181 * cgroup_populate_dir - create subsys files in a cgroup directory
4182 * @cgrp: target cgroup
4183 * @subsys_mask: mask of the subsystem ids whose files should be added
4185 * On failure, no file is added.
4187 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
4189 struct cgroup_subsys *ss;
4192 /* process cftsets of each subsystem */
4193 for_each_subsys(ss, i) {
4194 struct cftype_set *set;
4196 if (!test_bit(i, &subsys_mask))
4199 list_for_each_entry(set, &ss->cftsets, node) {
4200 ret = cgroup_addrm_files(cgrp, ss, set->cfts, true);
4206 /* This cgroup is ready now */
4207 for_each_root_subsys(cgrp->root, ss) {
4208 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
4209 struct css_id *id = rcu_dereference_protected(css->id, true);
4212 * Update id->css pointer and make this css visible from
4213 * CSS ID functions. This pointer will be dereferened
4214 * from RCU-read-side without locks.
4217 rcu_assign_pointer(id->css, css);
4222 cgroup_clear_dir(cgrp, subsys_mask);
4226 static void css_dput_fn(struct work_struct *work)
4228 struct cgroup_subsys_state *css =
4229 container_of(work, struct cgroup_subsys_state, dput_work);
4231 cgroup_dput(css->cgroup);
4234 static void css_release(struct percpu_ref *ref)
4236 struct cgroup_subsys_state *css =
4237 container_of(ref, struct cgroup_subsys_state, refcnt);
4239 schedule_work(&css->dput_work);
4242 static void init_cgroup_css(struct cgroup_subsys_state *css,
4243 struct cgroup_subsys *ss,
4244 struct cgroup *cgrp)
4249 if (cgrp == cgroup_dummy_top)
4250 css->flags |= CSS_ROOT;
4251 BUG_ON(cgrp->subsys[ss->subsys_id]);
4252 cgrp->subsys[ss->subsys_id] = css;
4255 * css holds an extra ref to @cgrp->dentry which is put on the last
4256 * css_put(). dput() requires process context, which css_put() may
4257 * be called without. @css->dput_work will be used to invoke
4258 * dput() asynchronously from css_put().
4260 INIT_WORK(&css->dput_work, css_dput_fn);
4263 /* invoke ->post_create() on a new CSS and mark it online if successful */
4264 static int online_css(struct cgroup_subsys *ss, struct cgroup *cgrp)
4268 lockdep_assert_held(&cgroup_mutex);
4271 ret = ss->css_online(cgrp);
4273 cgrp->subsys[ss->subsys_id]->flags |= CSS_ONLINE;
4277 /* if the CSS is online, invoke ->pre_destory() on it and mark it offline */
4278 static void offline_css(struct cgroup_subsys *ss, struct cgroup *cgrp)
4279 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4281 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
4283 lockdep_assert_held(&cgroup_mutex);
4285 if (!(css->flags & CSS_ONLINE))
4288 if (ss->css_offline)
4289 ss->css_offline(cgrp);
4291 cgrp->subsys[ss->subsys_id]->flags &= ~CSS_ONLINE;
4295 * cgroup_create - create a cgroup
4296 * @parent: cgroup that will be parent of the new cgroup
4297 * @dentry: dentry of the new cgroup
4298 * @mode: mode to set on new inode
4300 * Must be called with the mutex on the parent inode held
4302 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
4305 struct cgroup *cgrp;
4306 struct cgroup_name *name;
4307 struct cgroupfs_root *root = parent->root;
4309 struct cgroup_subsys *ss;
4310 struct super_block *sb = root->sb;
4312 /* allocate the cgroup and its ID, 0 is reserved for the root */
4313 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4317 name = cgroup_alloc_name(dentry);
4320 rcu_assign_pointer(cgrp->name, name);
4322 cgrp->id = ida_simple_get(&root->cgroup_ida, 1, 0, GFP_KERNEL);
4327 * Only live parents can have children. Note that the liveliness
4328 * check isn't strictly necessary because cgroup_mkdir() and
4329 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
4330 * anyway so that locking is contained inside cgroup proper and we
4331 * don't get nasty surprises if we ever grow another caller.
4333 if (!cgroup_lock_live_group(parent)) {
4338 /* Grab a reference on the superblock so the hierarchy doesn't
4339 * get deleted on unmount if there are child cgroups. This
4340 * can be done outside cgroup_mutex, since the sb can't
4341 * disappear while someone has an open control file on the
4343 atomic_inc(&sb->s_active);
4345 init_cgroup_housekeeping(cgrp);
4347 dentry->d_fsdata = cgrp;
4348 cgrp->dentry = dentry;
4350 cgrp->parent = parent;
4351 cgrp->root = parent->root;
4353 if (notify_on_release(parent))
4354 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4356 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4357 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4359 for_each_root_subsys(root, ss) {
4360 struct cgroup_subsys_state *css;
4362 css = ss->css_alloc(cgrp);
4368 err = percpu_ref_init(&css->refcnt, css_release);
4372 init_cgroup_css(css, ss, cgrp);
4375 err = alloc_css_id(ss, parent, cgrp);
4382 * Create directory. cgroup_create_file() returns with the new
4383 * directory locked on success so that it can be populated without
4384 * dropping cgroup_mutex.
4386 err = cgroup_create_file(dentry, S_IFDIR | mode, sb);
4389 lockdep_assert_held(&dentry->d_inode->i_mutex);
4391 cgrp->serial_nr = cgroup_serial_nr_next++;
4393 /* allocation complete, commit to creation */
4394 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
4395 root->number_of_cgroups++;
4397 /* each css holds a ref to the cgroup's dentry */
4398 for_each_root_subsys(root, ss)
4401 /* hold a ref to the parent's dentry */
4402 dget(parent->dentry);
4404 /* creation succeeded, notify subsystems */
4405 for_each_root_subsys(root, ss) {
4406 err = online_css(ss, cgrp);
4410 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4412 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",
4413 current->comm, current->pid, ss->name);
4414 if (!strcmp(ss->name, "memory"))
4415 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
4416 ss->warned_broken_hierarchy = true;
4420 err = cgroup_addrm_files(cgrp, NULL, cgroup_base_files, true);
4424 err = cgroup_populate_dir(cgrp, root->subsys_mask);
4428 mutex_unlock(&cgroup_mutex);
4429 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
4434 for_each_root_subsys(root, ss) {
4435 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
4438 percpu_ref_cancel_init(&css->refcnt);
4442 mutex_unlock(&cgroup_mutex);
4443 /* Release the reference count that we took on the superblock */
4444 deactivate_super(sb);
4446 ida_simple_remove(&root->cgroup_ida, cgrp->id);
4448 kfree(rcu_dereference_raw(cgrp->name));
4454 cgroup_destroy_locked(cgrp);
4455 mutex_unlock(&cgroup_mutex);
4456 mutex_unlock(&dentry->d_inode->i_mutex);
4460 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
4462 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
4464 /* the vfs holds inode->i_mutex already */
4465 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
4468 static void cgroup_css_killed(struct cgroup *cgrp)
4470 if (!atomic_dec_and_test(&cgrp->css_kill_cnt))
4473 /* percpu ref's of all css's are killed, kick off the next step */
4474 INIT_WORK(&cgrp->destroy_work, cgroup_offline_fn);
4475 schedule_work(&cgrp->destroy_work);
4478 static void css_ref_killed_fn(struct percpu_ref *ref)
4480 struct cgroup_subsys_state *css =
4481 container_of(ref, struct cgroup_subsys_state, refcnt);
4483 cgroup_css_killed(css->cgroup);
4487 * cgroup_destroy_locked - the first stage of cgroup destruction
4488 * @cgrp: cgroup to be destroyed
4490 * css's make use of percpu refcnts whose killing latency shouldn't be
4491 * exposed to userland and are RCU protected. Also, cgroup core needs to
4492 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
4493 * invoked. To satisfy all the requirements, destruction is implemented in
4494 * the following two steps.
4496 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4497 * userland visible parts and start killing the percpu refcnts of
4498 * css's. Set up so that the next stage will be kicked off once all
4499 * the percpu refcnts are confirmed to be killed.
4501 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4502 * rest of destruction. Once all cgroup references are gone, the
4503 * cgroup is RCU-freed.
4505 * This function implements s1. After this step, @cgrp is gone as far as
4506 * the userland is concerned and a new cgroup with the same name may be
4507 * created. As cgroup doesn't care about the names internally, this
4508 * doesn't cause any problem.
4510 static int cgroup_destroy_locked(struct cgroup *cgrp)
4511 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4513 struct dentry *d = cgrp->dentry;
4514 struct cgroup_event *event, *tmp;
4515 struct cgroup_subsys *ss;
4518 lockdep_assert_held(&d->d_inode->i_mutex);
4519 lockdep_assert_held(&cgroup_mutex);
4522 * css_set_lock synchronizes access to ->cset_links and prevents
4523 * @cgrp from being removed while __put_css_set() is in progress.
4525 read_lock(&css_set_lock);
4526 empty = list_empty(&cgrp->cset_links) && list_empty(&cgrp->children);
4527 read_unlock(&css_set_lock);
4532 * Block new css_tryget() by killing css refcnts. cgroup core
4533 * guarantees that, by the time ->css_offline() is invoked, no new
4534 * css reference will be given out via css_tryget(). We can't
4535 * simply call percpu_ref_kill() and proceed to offlining css's
4536 * because percpu_ref_kill() doesn't guarantee that the ref is seen
4537 * as killed on all CPUs on return.
4539 * Use percpu_ref_kill_and_confirm() to get notifications as each
4540 * css is confirmed to be seen as killed on all CPUs. The
4541 * notification callback keeps track of the number of css's to be
4542 * killed and schedules cgroup_offline_fn() to perform the rest of
4543 * destruction once the percpu refs of all css's are confirmed to
4546 atomic_set(&cgrp->css_kill_cnt, 1);
4547 for_each_root_subsys(cgrp->root, ss) {
4548 struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
4551 * Killing would put the base ref, but we need to keep it
4552 * alive until after ->css_offline.
4554 percpu_ref_get(&css->refcnt);
4556 atomic_inc(&cgrp->css_kill_cnt);
4557 percpu_ref_kill_and_confirm(&css->refcnt, css_ref_killed_fn);
4559 cgroup_css_killed(cgrp);
4562 * Mark @cgrp dead. This prevents further task migration and child
4563 * creation by disabling cgroup_lock_live_group(). Note that
4564 * CGRP_DEAD assertion is depended upon by cgroup_next_sibling() to
4565 * resume iteration after dropping RCU read lock. See
4566 * cgroup_next_sibling() for details.
4568 set_bit(CGRP_DEAD, &cgrp->flags);
4570 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4571 raw_spin_lock(&release_list_lock);
4572 if (!list_empty(&cgrp->release_list))
4573 list_del_init(&cgrp->release_list);
4574 raw_spin_unlock(&release_list_lock);
4577 * Clear and remove @cgrp directory. The removal puts the base ref
4578 * but we aren't quite done with @cgrp yet, so hold onto it.
4580 cgroup_clear_dir(cgrp, cgrp->root->subsys_mask);
4581 cgroup_addrm_files(cgrp, NULL, cgroup_base_files, false);
4583 cgroup_d_remove_dir(d);
4586 * Unregister events and notify userspace.
4587 * Notify userspace about cgroup removing only after rmdir of cgroup
4588 * directory to avoid race between userspace and kernelspace.
4590 spin_lock(&cgrp->event_list_lock);
4591 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
4592 list_del_init(&event->list);
4593 schedule_work(&event->remove);
4595 spin_unlock(&cgrp->event_list_lock);
4601 * cgroup_offline_fn - the second step of cgroup destruction
4602 * @work: cgroup->destroy_free_work
4604 * This function is invoked from a work item for a cgroup which is being
4605 * destroyed after the percpu refcnts of all css's are guaranteed to be
4606 * seen as killed on all CPUs, and performs the rest of destruction. This
4607 * is the second step of destruction described in the comment above
4608 * cgroup_destroy_locked().
4610 static void cgroup_offline_fn(struct work_struct *work)
4612 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
4613 struct cgroup *parent = cgrp->parent;
4614 struct dentry *d = cgrp->dentry;
4615 struct cgroup_subsys *ss;
4617 mutex_lock(&cgroup_mutex);
4620 * css_tryget() is guaranteed to fail now. Tell subsystems to
4621 * initate destruction.
4623 for_each_root_subsys(cgrp->root, ss)
4624 offline_css(ss, cgrp);
4627 * Put the css refs from cgroup_destroy_locked(). Each css holds
4628 * an extra reference to the cgroup's dentry and cgroup removal
4629 * proceeds regardless of css refs. On the last put of each css,
4630 * whenever that may be, the extra dentry ref is put so that dentry
4631 * destruction happens only after all css's are released.
4633 for_each_root_subsys(cgrp->root, ss)
4634 css_put(cgrp->subsys[ss->subsys_id]);
4636 /* delete this cgroup from parent->children */
4637 list_del_rcu(&cgrp->sibling);
4641 set_bit(CGRP_RELEASABLE, &parent->flags);
4642 check_for_release(parent);
4644 mutex_unlock(&cgroup_mutex);
4647 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
4651 mutex_lock(&cgroup_mutex);
4652 ret = cgroup_destroy_locked(dentry->d_fsdata);
4653 mutex_unlock(&cgroup_mutex);
4658 static void __init_or_module cgroup_init_cftsets(struct cgroup_subsys *ss)
4660 INIT_LIST_HEAD(&ss->cftsets);
4663 * base_cftset is embedded in subsys itself, no need to worry about
4666 if (ss->base_cftypes) {
4667 ss->base_cftset.cfts = ss->base_cftypes;
4668 list_add_tail(&ss->base_cftset.node, &ss->cftsets);
4672 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4674 struct cgroup_subsys_state *css;
4676 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4678 mutex_lock(&cgroup_mutex);
4680 /* init base cftset */
4681 cgroup_init_cftsets(ss);
4683 /* Create the top cgroup state for this subsystem */
4684 list_add(&ss->sibling, &cgroup_dummy_root.subsys_list);
4685 ss->root = &cgroup_dummy_root;
4686 css = ss->css_alloc(cgroup_dummy_top);
4687 /* We don't handle early failures gracefully */
4688 BUG_ON(IS_ERR(css));
4689 init_cgroup_css(css, ss, cgroup_dummy_top);
4691 /* Update the init_css_set to contain a subsys
4692 * pointer to this state - since the subsystem is
4693 * newly registered, all tasks and hence the
4694 * init_css_set is in the subsystem's top cgroup. */
4695 init_css_set.subsys[ss->subsys_id] = css;
4697 need_forkexit_callback |= ss->fork || ss->exit;
4699 /* At system boot, before all subsystems have been
4700 * registered, no tasks have been forked, so we don't
4701 * need to invoke fork callbacks here. */
4702 BUG_ON(!list_empty(&init_task.tasks));
4704 BUG_ON(online_css(ss, cgroup_dummy_top));
4706 mutex_unlock(&cgroup_mutex);
4708 /* this function shouldn't be used with modular subsystems, since they
4709 * need to register a subsys_id, among other things */
4714 * cgroup_load_subsys: load and register a modular subsystem at runtime
4715 * @ss: the subsystem to load
4717 * This function should be called in a modular subsystem's initcall. If the
4718 * subsystem is built as a module, it will be assigned a new subsys_id and set
4719 * up for use. If the subsystem is built-in anyway, work is delegated to the
4720 * simpler cgroup_init_subsys.
4722 int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
4724 struct cgroup_subsys_state *css;
4726 struct hlist_node *tmp;
4727 struct css_set *cset;
4730 /* check name and function validity */
4731 if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
4732 ss->css_alloc == NULL || ss->css_free == NULL)
4736 * we don't support callbacks in modular subsystems. this check is
4737 * before the ss->module check for consistency; a subsystem that could
4738 * be a module should still have no callbacks even if the user isn't
4739 * compiling it as one.
4741 if (ss->fork || ss->exit)
4745 * an optionally modular subsystem is built-in: we want to do nothing,
4746 * since cgroup_init_subsys will have already taken care of it.
4748 if (ss->module == NULL) {
4749 /* a sanity check */
4750 BUG_ON(cgroup_subsys[ss->subsys_id] != ss);
4754 /* init base cftset */
4755 cgroup_init_cftsets(ss);
4757 mutex_lock(&cgroup_mutex);
4758 cgroup_subsys[ss->subsys_id] = ss;
4761 * no ss->css_alloc seems to need anything important in the ss
4762 * struct, so this can happen first (i.e. before the dummy root
4765 css = ss->css_alloc(cgroup_dummy_top);
4767 /* failure case - need to deassign the cgroup_subsys[] slot. */
4768 cgroup_subsys[ss->subsys_id] = NULL;
4769 mutex_unlock(&cgroup_mutex);
4770 return PTR_ERR(css);
4773 list_add(&ss->sibling, &cgroup_dummy_root.subsys_list);
4774 ss->root = &cgroup_dummy_root;
4776 /* our new subsystem will be attached to the dummy hierarchy. */
4777 init_cgroup_css(css, ss, cgroup_dummy_top);
4778 /* init_idr must be after init_cgroup_css because it sets css->id. */
4780 ret = cgroup_init_idr(ss, css);
4786 * Now we need to entangle the css into the existing css_sets. unlike
4787 * in cgroup_init_subsys, there are now multiple css_sets, so each one
4788 * will need a new pointer to it; done by iterating the css_set_table.
4789 * furthermore, modifying the existing css_sets will corrupt the hash
4790 * table state, so each changed css_set will need its hash recomputed.
4791 * this is all done under the css_set_lock.
4793 write_lock(&css_set_lock);
4794 hash_for_each_safe(css_set_table, i, tmp, cset, hlist) {
4795 /* skip entries that we already rehashed */
4796 if (cset->subsys[ss->subsys_id])
4798 /* remove existing entry */
4799 hash_del(&cset->hlist);
4801 cset->subsys[ss->subsys_id] = css;
4802 /* recompute hash and restore entry */
4803 key = css_set_hash(cset->subsys);
4804 hash_add(css_set_table, &cset->hlist, key);
4806 write_unlock(&css_set_lock);
4808 ret = online_css(ss, cgroup_dummy_top);
4813 mutex_unlock(&cgroup_mutex);
4817 mutex_unlock(&cgroup_mutex);
4818 /* @ss can't be mounted here as try_module_get() would fail */
4819 cgroup_unload_subsys(ss);
4822 EXPORT_SYMBOL_GPL(cgroup_load_subsys);
4825 * cgroup_unload_subsys: unload a modular subsystem
4826 * @ss: the subsystem to unload
4828 * This function should be called in a modular subsystem's exitcall. When this
4829 * function is invoked, the refcount on the subsystem's module will be 0, so
4830 * the subsystem will not be attached to any hierarchy.
4832 void cgroup_unload_subsys(struct cgroup_subsys *ss)
4834 struct cgrp_cset_link *link;
4836 BUG_ON(ss->module == NULL);
4839 * we shouldn't be called if the subsystem is in use, and the use of
4840 * try_module_get in parse_cgroupfs_options should ensure that it
4841 * doesn't start being used while we're killing it off.
4843 BUG_ON(ss->root != &cgroup_dummy_root);
4845 mutex_lock(&cgroup_mutex);
4847 offline_css(ss, cgroup_dummy_top);
4850 idr_destroy(&ss->idr);
4852 /* deassign the subsys_id */
4853 cgroup_subsys[ss->subsys_id] = NULL;
4855 /* remove subsystem from the dummy root's list of subsystems */
4856 list_del_init(&ss->sibling);
4859 * disentangle the css from all css_sets attached to the dummy
4860 * top. as in loading, we need to pay our respects to the hashtable
4863 write_lock(&css_set_lock);
4864 list_for_each_entry(link, &cgroup_dummy_top->cset_links, cset_link) {
4865 struct css_set *cset = link->cset;
4868 hash_del(&cset->hlist);
4869 cset->subsys[ss->subsys_id] = NULL;
4870 key = css_set_hash(cset->subsys);
4871 hash_add(css_set_table, &cset->hlist, key);
4873 write_unlock(&css_set_lock);
4876 * remove subsystem's css from the cgroup_dummy_top and free it -
4877 * need to free before marking as null because ss->css_free needs
4878 * the cgrp->subsys pointer to find their state. note that this
4879 * also takes care of freeing the css_id.
4881 ss->css_free(cgroup_dummy_top);
4882 cgroup_dummy_top->subsys[ss->subsys_id] = NULL;
4884 mutex_unlock(&cgroup_mutex);
4886 EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
4889 * cgroup_init_early - cgroup initialization at system boot
4891 * Initialize cgroups at system boot, and initialize any
4892 * subsystems that request early init.
4894 int __init cgroup_init_early(void)
4896 struct cgroup_subsys *ss;
4899 atomic_set(&init_css_set.refcount, 1);
4900 INIT_LIST_HEAD(&init_css_set.cgrp_links);
4901 INIT_LIST_HEAD(&init_css_set.tasks);
4902 INIT_HLIST_NODE(&init_css_set.hlist);
4904 init_cgroup_root(&cgroup_dummy_root);
4905 cgroup_root_count = 1;
4906 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4908 init_cgrp_cset_link.cset = &init_css_set;
4909 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4910 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4911 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4913 /* at bootup time, we don't worry about modular subsystems */
4914 for_each_builtin_subsys(ss, i) {
4916 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
4917 BUG_ON(!ss->css_alloc);
4918 BUG_ON(!ss->css_free);
4919 if (ss->subsys_id != i) {
4920 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
4921 ss->name, ss->subsys_id);
4926 cgroup_init_subsys(ss);
4932 * cgroup_init - cgroup initialization
4934 * Register cgroup filesystem and /proc file, and initialize
4935 * any subsystems that didn't request early init.
4937 int __init cgroup_init(void)
4939 struct cgroup_subsys *ss;
4943 err = bdi_init(&cgroup_backing_dev_info);
4947 for_each_builtin_subsys(ss, i) {
4948 if (!ss->early_init)
4949 cgroup_init_subsys(ss);
4951 cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
4954 /* allocate id for the dummy hierarchy */
4955 mutex_lock(&cgroup_mutex);
4956 mutex_lock(&cgroup_root_mutex);
4958 /* Add init_css_set to the hash table */
4959 key = css_set_hash(init_css_set.subsys);
4960 hash_add(css_set_table, &init_css_set.hlist, key);
4962 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4964 mutex_unlock(&cgroup_root_mutex);
4965 mutex_unlock(&cgroup_mutex);
4967 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4973 err = register_filesystem(&cgroup_fs_type);
4975 kobject_put(cgroup_kobj);
4979 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4983 bdi_destroy(&cgroup_backing_dev_info);
4989 * proc_cgroup_show()
4990 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4991 * - Used for /proc/<pid>/cgroup.
4992 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4993 * doesn't really matter if tsk->cgroup changes after we read it,
4994 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4995 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4996 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4997 * cgroup to top_cgroup.
5000 /* TODO: Use a proper seq_file iterator */
5001 int proc_cgroup_show(struct seq_file *m, void *v)
5004 struct task_struct *tsk;
5007 struct cgroupfs_root *root;
5010 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
5016 tsk = get_pid_task(pid, PIDTYPE_PID);
5022 mutex_lock(&cgroup_mutex);
5024 for_each_active_root(root) {
5025 struct cgroup_subsys *ss;
5026 struct cgroup *cgrp;
5029 seq_printf(m, "%d:", root->hierarchy_id);
5030 for_each_root_subsys(root, ss)
5031 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
5032 if (strlen(root->name))
5033 seq_printf(m, "%sname=%s", count ? "," : "",
5036 cgrp = task_cgroup_from_root(tsk, root);
5037 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
5045 mutex_unlock(&cgroup_mutex);
5046 put_task_struct(tsk);
5053 /* Display information about each subsystem and each hierarchy */
5054 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5056 struct cgroup_subsys *ss;
5059 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5061 * ideally we don't want subsystems moving around while we do this.
5062 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5063 * subsys/hierarchy state.
5065 mutex_lock(&cgroup_mutex);
5067 for_each_subsys(ss, i)
5068 seq_printf(m, "%s\t%d\t%d\t%d\n",
5069 ss->name, ss->root->hierarchy_id,
5070 ss->root->number_of_cgroups, !ss->disabled);
5072 mutex_unlock(&cgroup_mutex);
5076 static int cgroupstats_open(struct inode *inode, struct file *file)
5078 return single_open(file, proc_cgroupstats_show, NULL);
5081 static const struct file_operations proc_cgroupstats_operations = {
5082 .open = cgroupstats_open,
5084 .llseek = seq_lseek,
5085 .release = single_release,
5089 * cgroup_fork - attach newly forked task to its parents cgroup.
5090 * @child: pointer to task_struct of forking parent process.
5092 * Description: A task inherits its parent's cgroup at fork().
5094 * A pointer to the shared css_set was automatically copied in
5095 * fork.c by dup_task_struct(). However, we ignore that copy, since
5096 * it was not made under the protection of RCU or cgroup_mutex, so
5097 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
5098 * have already changed current->cgroups, allowing the previously
5099 * referenced cgroup group to be removed and freed.
5101 * At the point that cgroup_fork() is called, 'current' is the parent
5102 * task, and the passed argument 'child' points to the child task.
5104 void cgroup_fork(struct task_struct *child)
5107 get_css_set(task_css_set(current));
5108 child->cgroups = current->cgroups;
5109 task_unlock(current);
5110 INIT_LIST_HEAD(&child->cg_list);
5114 * cgroup_post_fork - called on a new task after adding it to the task list
5115 * @child: the task in question
5117 * Adds the task to the list running through its css_set if necessary and
5118 * call the subsystem fork() callbacks. Has to be after the task is
5119 * visible on the task list in case we race with the first call to
5120 * cgroup_iter_start() - to guarantee that the new task ends up on its
5123 void cgroup_post_fork(struct task_struct *child)
5125 struct cgroup_subsys *ss;
5129 * use_task_css_set_links is set to 1 before we walk the tasklist
5130 * under the tasklist_lock and we read it here after we added the child
5131 * to the tasklist under the tasklist_lock as well. If the child wasn't
5132 * yet in the tasklist when we walked through it from
5133 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
5134 * should be visible now due to the paired locking and barriers implied
5135 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
5136 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
5139 if (use_task_css_set_links) {
5140 write_lock(&css_set_lock);
5142 if (list_empty(&child->cg_list))
5143 list_add(&child->cg_list, &task_css_set(child)->tasks);
5145 write_unlock(&css_set_lock);
5149 * Call ss->fork(). This must happen after @child is linked on
5150 * css_set; otherwise, @child might change state between ->fork()
5151 * and addition to css_set.
5153 if (need_forkexit_callback) {
5155 * fork/exit callbacks are supported only for builtin
5156 * subsystems, and the builtin section of the subsys
5157 * array is immutable, so we don't need to lock the
5158 * subsys array here. On the other hand, modular section
5159 * of the array can be freed at module unload, so we
5162 for_each_builtin_subsys(ss, i)
5169 * cgroup_exit - detach cgroup from exiting task
5170 * @tsk: pointer to task_struct of exiting process
5171 * @run_callback: run exit callbacks?
5173 * Description: Detach cgroup from @tsk and release it.
5175 * Note that cgroups marked notify_on_release force every task in
5176 * them to take the global cgroup_mutex mutex when exiting.
5177 * This could impact scaling on very large systems. Be reluctant to
5178 * use notify_on_release cgroups where very high task exit scaling
5179 * is required on large systems.
5181 * the_top_cgroup_hack:
5183 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
5185 * We call cgroup_exit() while the task is still competent to
5186 * handle notify_on_release(), then leave the task attached to the
5187 * root cgroup in each hierarchy for the remainder of its exit.
5189 * To do this properly, we would increment the reference count on
5190 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
5191 * code we would add a second cgroup function call, to drop that
5192 * reference. This would just create an unnecessary hot spot on
5193 * the top_cgroup reference count, to no avail.
5195 * Normally, holding a reference to a cgroup without bumping its
5196 * count is unsafe. The cgroup could go away, or someone could
5197 * attach us to a different cgroup, decrementing the count on
5198 * the first cgroup that we never incremented. But in this case,
5199 * top_cgroup isn't going away, and either task has PF_EXITING set,
5200 * which wards off any cgroup_attach_task() attempts, or task is a failed
5201 * fork, never visible to cgroup_attach_task.
5203 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
5205 struct cgroup_subsys *ss;
5206 struct css_set *cset;
5210 * Unlink from the css_set task list if necessary.
5211 * Optimistically check cg_list before taking
5214 if (!list_empty(&tsk->cg_list)) {
5215 write_lock(&css_set_lock);
5216 if (!list_empty(&tsk->cg_list))
5217 list_del_init(&tsk->cg_list);
5218 write_unlock(&css_set_lock);
5221 /* Reassign the task to the init_css_set. */
5223 cset = task_css_set(tsk);
5224 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5226 if (run_callbacks && need_forkexit_callback) {
5228 * fork/exit callbacks are supported only for builtin
5229 * subsystems, see cgroup_post_fork() for details.
5231 for_each_builtin_subsys(ss, i) {
5233 struct cgroup *old_cgrp = cset->subsys[i]->cgroup;
5234 struct cgroup *cgrp = task_cgroup(tsk, i);
5236 ss->exit(cgrp, old_cgrp, tsk);
5242 put_css_set_taskexit(cset);
5245 static void check_for_release(struct cgroup *cgrp)
5247 if (cgroup_is_releasable(cgrp) &&
5248 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
5250 * Control Group is currently removeable. If it's not
5251 * already queued for a userspace notification, queue
5254 int need_schedule_work = 0;
5256 raw_spin_lock(&release_list_lock);
5257 if (!cgroup_is_dead(cgrp) &&
5258 list_empty(&cgrp->release_list)) {
5259 list_add(&cgrp->release_list, &release_list);
5260 need_schedule_work = 1;
5262 raw_spin_unlock(&release_list_lock);
5263 if (need_schedule_work)
5264 schedule_work(&release_agent_work);
5269 * Notify userspace when a cgroup is released, by running the
5270 * configured release agent with the name of the cgroup (path
5271 * relative to the root of cgroup file system) as the argument.
5273 * Most likely, this user command will try to rmdir this cgroup.
5275 * This races with the possibility that some other task will be
5276 * attached to this cgroup before it is removed, or that some other
5277 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5278 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5279 * unused, and this cgroup will be reprieved from its death sentence,
5280 * to continue to serve a useful existence. Next time it's released,
5281 * we will get notified again, if it still has 'notify_on_release' set.
5283 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5284 * means only wait until the task is successfully execve()'d. The
5285 * separate release agent task is forked by call_usermodehelper(),
5286 * then control in this thread returns here, without waiting for the
5287 * release agent task. We don't bother to wait because the caller of
5288 * this routine has no use for the exit status of the release agent
5289 * task, so no sense holding our caller up for that.
5291 static void cgroup_release_agent(struct work_struct *work)
5293 BUG_ON(work != &release_agent_work);
5294 mutex_lock(&cgroup_mutex);
5295 raw_spin_lock(&release_list_lock);
5296 while (!list_empty(&release_list)) {
5297 char *argv[3], *envp[3];
5299 char *pathbuf = NULL, *agentbuf = NULL;
5300 struct cgroup *cgrp = list_entry(release_list.next,
5303 list_del_init(&cgrp->release_list);
5304 raw_spin_unlock(&release_list_lock);
5305 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
5308 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
5310 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5315 argv[i++] = agentbuf;
5316 argv[i++] = pathbuf;
5320 /* minimal command environment */
5321 envp[i++] = "HOME=/";
5322 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5325 /* Drop the lock while we invoke the usermode helper,
5326 * since the exec could involve hitting disk and hence
5327 * be a slow process */
5328 mutex_unlock(&cgroup_mutex);
5329 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5330 mutex_lock(&cgroup_mutex);
5334 raw_spin_lock(&release_list_lock);
5336 raw_spin_unlock(&release_list_lock);
5337 mutex_unlock(&cgroup_mutex);
5340 static int __init cgroup_disable(char *str)
5342 struct cgroup_subsys *ss;
5346 while ((token = strsep(&str, ",")) != NULL) {
5351 * cgroup_disable, being at boot time, can't know about
5352 * module subsystems, so we don't worry about them.
5354 for_each_builtin_subsys(ss, i) {
5355 if (!strcmp(token, ss->name)) {
5357 printk(KERN_INFO "Disabling %s control group"
5358 " subsystem\n", ss->name);
5365 __setup("cgroup_disable=", cgroup_disable);
5368 * Functons for CSS ID.
5371 /* to get ID other than 0, this should be called when !cgroup_is_dead() */
5372 unsigned short css_id(struct cgroup_subsys_state *css)
5374 struct css_id *cssid;
5377 * This css_id() can return correct value when somone has refcnt
5378 * on this or this is under rcu_read_lock(). Once css->id is allocated,
5379 * it's unchanged until freed.
5381 cssid = rcu_dereference_raw(css->id);
5387 EXPORT_SYMBOL_GPL(css_id);
5390 * css_is_ancestor - test "root" css is an ancestor of "child"
5391 * @child: the css to be tested.
5392 * @root: the css supporsed to be an ancestor of the child.
5394 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
5395 * this function reads css->id, the caller must hold rcu_read_lock().
5396 * But, considering usual usage, the csses should be valid objects after test.
5397 * Assuming that the caller will do some action to the child if this returns
5398 * returns true, the caller must take "child";s reference count.
5399 * If "child" is valid object and this returns true, "root" is valid, too.
5402 bool css_is_ancestor(struct cgroup_subsys_state *child,
5403 const struct cgroup_subsys_state *root)
5405 struct css_id *child_id;
5406 struct css_id *root_id;
5408 child_id = rcu_dereference(child->id);
5411 root_id = rcu_dereference(root->id);
5414 if (child_id->depth < root_id->depth)
5416 if (child_id->stack[root_id->depth] != root_id->id)
5421 void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
5423 struct css_id *id = rcu_dereference_protected(css->id, true);
5425 /* When this is called before css_id initialization, id can be NULL */
5429 BUG_ON(!ss->use_id);
5431 rcu_assign_pointer(id->css, NULL);
5432 rcu_assign_pointer(css->id, NULL);
5433 spin_lock(&ss->id_lock);
5434 idr_remove(&ss->idr, id->id);
5435 spin_unlock(&ss->id_lock);
5436 kfree_rcu(id, rcu_head);
5438 EXPORT_SYMBOL_GPL(free_css_id);
5441 * This is called by init or create(). Then, calls to this function are
5442 * always serialized (By cgroup_mutex() at create()).
5445 static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
5447 struct css_id *newid;
5450 BUG_ON(!ss->use_id);
5452 size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
5453 newid = kzalloc(size, GFP_KERNEL);
5455 return ERR_PTR(-ENOMEM);
5457 idr_preload(GFP_KERNEL);
5458 spin_lock(&ss->id_lock);
5459 /* Don't use 0. allocates an ID of 1-65535 */
5460 ret = idr_alloc(&ss->idr, newid, 1, CSS_ID_MAX + 1, GFP_NOWAIT);
5461 spin_unlock(&ss->id_lock);
5464 /* Returns error when there are no free spaces for new ID.*/
5469 newid->depth = depth;
5473 return ERR_PTR(ret);
5477 static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
5478 struct cgroup_subsys_state *rootcss)
5480 struct css_id *newid;
5482 spin_lock_init(&ss->id_lock);
5485 newid = get_new_cssid(ss, 0);
5487 return PTR_ERR(newid);
5489 newid->stack[0] = newid->id;
5490 RCU_INIT_POINTER(newid->css, rootcss);
5491 RCU_INIT_POINTER(rootcss->id, newid);
5495 static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
5496 struct cgroup *child)
5498 int subsys_id, i, depth = 0;
5499 struct cgroup_subsys_state *parent_css, *child_css;
5500 struct css_id *child_id, *parent_id;
5502 subsys_id = ss->subsys_id;
5503 parent_css = parent->subsys[subsys_id];
5504 child_css = child->subsys[subsys_id];
5505 parent_id = rcu_dereference_protected(parent_css->id, true);
5506 depth = parent_id->depth + 1;
5508 child_id = get_new_cssid(ss, depth);
5509 if (IS_ERR(child_id))
5510 return PTR_ERR(child_id);
5512 for (i = 0; i < depth; i++)
5513 child_id->stack[i] = parent_id->stack[i];
5514 child_id->stack[depth] = child_id->id;
5516 * child_id->css pointer will be set after this cgroup is available
5517 * see cgroup_populate_dir()
5519 rcu_assign_pointer(child_css->id, child_id);
5525 * css_lookup - lookup css by id
5526 * @ss: cgroup subsys to be looked into.
5529 * Returns pointer to cgroup_subsys_state if there is valid one with id.
5530 * NULL if not. Should be called under rcu_read_lock()
5532 struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
5534 struct css_id *cssid = NULL;
5536 BUG_ON(!ss->use_id);
5537 cssid = idr_find(&ss->idr, id);
5539 if (unlikely(!cssid))
5542 return rcu_dereference(cssid->css);
5544 EXPORT_SYMBOL_GPL(css_lookup);
5547 * get corresponding css from file open on cgroupfs directory
5549 struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id)
5551 struct cgroup *cgrp;
5552 struct inode *inode;
5553 struct cgroup_subsys_state *css;
5555 inode = file_inode(f);
5556 /* check in cgroup filesystem dir */
5557 if (inode->i_op != &cgroup_dir_inode_operations)
5558 return ERR_PTR(-EBADF);
5560 if (id < 0 || id >= CGROUP_SUBSYS_COUNT)
5561 return ERR_PTR(-EINVAL);
5564 cgrp = __d_cgrp(f->f_dentry);
5565 css = cgrp->subsys[id];
5566 return css ? css : ERR_PTR(-ENOENT);
5569 #ifdef CONFIG_CGROUP_DEBUG
5570 static struct cgroup_subsys_state *debug_css_alloc(struct cgroup *cgrp)
5572 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5575 return ERR_PTR(-ENOMEM);
5580 static void debug_css_free(struct cgroup *cgrp)
5582 kfree(cgrp->subsys[debug_subsys_id]);
5585 static u64 debug_taskcount_read(struct cgroup *cgrp, struct cftype *cft)
5587 return cgroup_task_count(cgrp);
5590 static u64 current_css_set_read(struct cgroup *cgrp, struct cftype *cft)
5592 return (u64)(unsigned long)current->cgroups;
5595 static u64 current_css_set_refcount_read(struct cgroup *cgrp,
5601 count = atomic_read(&task_css_set(current)->refcount);
5606 static int current_css_set_cg_links_read(struct cgroup *cgrp,
5608 struct seq_file *seq)
5610 struct cgrp_cset_link *link;
5611 struct css_set *cset;
5613 read_lock(&css_set_lock);
5615 cset = rcu_dereference(current->cgroups);
5616 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5617 struct cgroup *c = link->cgrp;
5621 name = c->dentry->d_name.name;
5624 seq_printf(seq, "Root %d group %s\n",
5625 c->root->hierarchy_id, name);
5628 read_unlock(&css_set_lock);
5632 #define MAX_TASKS_SHOWN_PER_CSS 25
5633 static int cgroup_css_links_read(struct cgroup *cgrp,
5635 struct seq_file *seq)
5637 struct cgrp_cset_link *link;
5639 read_lock(&css_set_lock);
5640 list_for_each_entry(link, &cgrp->cset_links, cset_link) {
5641 struct css_set *cset = link->cset;
5642 struct task_struct *task;
5644 seq_printf(seq, "css_set %p\n", cset);
5645 list_for_each_entry(task, &cset->tasks, cg_list) {
5646 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
5647 seq_puts(seq, " ...\n");
5650 seq_printf(seq, " task %d\n",
5651 task_pid_vnr(task));
5655 read_unlock(&css_set_lock);
5659 static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
5661 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
5664 static struct cftype debug_files[] = {
5666 .name = "taskcount",
5667 .read_u64 = debug_taskcount_read,
5671 .name = "current_css_set",
5672 .read_u64 = current_css_set_read,
5676 .name = "current_css_set_refcount",
5677 .read_u64 = current_css_set_refcount_read,
5681 .name = "current_css_set_cg_links",
5682 .read_seq_string = current_css_set_cg_links_read,
5686 .name = "cgroup_css_links",
5687 .read_seq_string = cgroup_css_links_read,
5691 .name = "releasable",
5692 .read_u64 = releasable_read,
5698 struct cgroup_subsys debug_subsys = {
5700 .css_alloc = debug_css_alloc,
5701 .css_free = debug_css_free,
5702 .subsys_id = debug_subsys_id,
5703 .base_cftypes = debug_files,
5705 #endif /* CONFIG_CGROUP_DEBUG */