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/slab.h>
44 #include <linux/spinlock.h>
45 #include <linux/rwsem.h>
46 #include <linux/string.h>
47 #include <linux/sort.h>
48 #include <linux/kmod.h>
49 #include <linux/delayacct.h>
50 #include <linux/cgroupstats.h>
51 #include <linux/hashtable.h>
52 #include <linux/pid_namespace.h>
53 #include <linux/idr.h>
54 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
55 #include <linux/flex_array.h> /* used in cgroup_attach_task */
56 #include <linux/kthread.h>
57 #include <linux/delay.h>
59 #include <linux/atomic.h>
62 * pidlists linger the following amount before being destroyed. The goal
63 * is avoiding frequent destruction in the middle of consecutive read calls
64 * Expiring in the middle is a performance problem not a correctness one.
65 * 1 sec should be enough.
67 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
69 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
73 * cgroup_tree_mutex nests above cgroup_mutex and protects cftypes, file
74 * creation/removal and hierarchy changing operations including cgroup
75 * creation, removal, css association and controller rebinding. This outer
76 * lock is needed mainly to resolve the circular dependency between kernfs
77 * active ref and cgroup_mutex. cgroup_tree_mutex nests above both.
79 static DEFINE_MUTEX(cgroup_tree_mutex);
82 * cgroup_mutex is the master lock. Any modification to cgroup or its
83 * hierarchy must be performed while holding it.
85 #ifdef CONFIG_PROVE_RCU
86 DEFINE_MUTEX(cgroup_mutex);
87 EXPORT_SYMBOL_GPL(cgroup_mutex); /* only for lockdep */
89 static DEFINE_MUTEX(cgroup_mutex);
93 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
94 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
96 static DEFINE_SPINLOCK(release_agent_path_lock);
98 #define cgroup_assert_mutexes_or_rcu_locked() \
99 rcu_lockdep_assert(rcu_read_lock_held() || \
100 lockdep_is_held(&cgroup_tree_mutex) || \
101 lockdep_is_held(&cgroup_mutex), \
102 "cgroup_[tree_]mutex or RCU read lock required");
105 * cgroup destruction makes heavy use of work items and there can be a lot
106 * of concurrent destructions. Use a separate workqueue so that cgroup
107 * destruction work items don't end up filling up max_active of system_wq
108 * which may lead to deadlock.
110 static struct workqueue_struct *cgroup_destroy_wq;
113 * pidlist destructions need to be flushed on cgroup destruction. Use a
114 * separate workqueue as flush domain.
116 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
118 /* generate an array of cgroup subsystem pointers */
119 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
120 static struct cgroup_subsys *cgroup_subsys[] = {
121 #include <linux/cgroup_subsys.h>
125 /* array of cgroup subsystem names */
126 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
127 static const char *cgroup_subsys_name[] = {
128 #include <linux/cgroup_subsys.h>
133 * The dummy hierarchy, reserved for the subsystems that are otherwise
134 * unattached - it never has more than a single cgroup, and all tasks are
135 * part of that cgroup.
137 static struct cgroupfs_root cgroup_dummy_root;
139 /* dummy_top is a shorthand for the dummy hierarchy's top cgroup */
140 static struct cgroup * const cgroup_dummy_top = &cgroup_dummy_root.top_cgroup;
142 /* The list of hierarchy roots */
144 static LIST_HEAD(cgroup_roots);
145 static int cgroup_root_count;
147 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
148 static DEFINE_IDR(cgroup_hierarchy_idr);
151 * Assign a monotonically increasing serial number to cgroups. It
152 * guarantees cgroups with bigger numbers are newer than those with smaller
153 * numbers. Also, as cgroups are always appended to the parent's
154 * ->children list, it guarantees that sibling cgroups are always sorted in
155 * the ascending serial number order on the list. Protected by
158 static u64 cgroup_serial_nr_next = 1;
160 /* This flag indicates whether tasks in the fork and exit paths should
161 * check for fork/exit handlers to call. This avoids us having to do
162 * extra work in the fork/exit path if none of the subsystems need to
165 static int need_forkexit_callback __read_mostly;
167 static struct cftype cgroup_base_files[];
169 static void cgroup_put(struct cgroup *cgrp);
170 static int rebind_subsystems(struct cgroupfs_root *root,
171 unsigned long added_mask, unsigned removed_mask);
172 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
173 static int cgroup_destroy_locked(struct cgroup *cgrp);
174 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
176 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
179 * cgroup_css - obtain a cgroup's css for the specified subsystem
180 * @cgrp: the cgroup of interest
181 * @ss: the subsystem of interest (%NULL returns the dummy_css)
183 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
184 * function must be called either under cgroup_mutex or rcu_read_lock() and
185 * the caller is responsible for pinning the returned css if it wants to
186 * keep accessing it outside the said locks. This function may return
187 * %NULL if @cgrp doesn't have @subsys_id enabled.
189 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
190 struct cgroup_subsys *ss)
193 return rcu_dereference_check(cgrp->subsys[ss->id],
194 lockdep_is_held(&cgroup_tree_mutex) ||
195 lockdep_is_held(&cgroup_mutex));
197 return &cgrp->dummy_css;
200 /* convenient tests for these bits */
201 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
203 return test_bit(CGRP_DEAD, &cgrp->flags);
206 struct cgroup_subsys_state *seq_css(struct seq_file *seq)
208 struct kernfs_open_file *of = seq->private;
209 struct cgroup *cgrp = of->kn->parent->priv;
210 struct cftype *cft = seq_cft(seq);
213 * This is open and unprotected implementation of cgroup_css().
214 * seq_css() is only called from a kernfs file operation which has
215 * an active reference on the file. Because all the subsystem
216 * files are drained before a css is disassociated with a cgroup,
217 * the matching css from the cgroup's subsys table is guaranteed to
218 * be and stay valid until the enclosing operation is complete.
221 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
223 return &cgrp->dummy_css;
225 EXPORT_SYMBOL_GPL(seq_css);
228 * cgroup_is_descendant - test ancestry
229 * @cgrp: the cgroup to be tested
230 * @ancestor: possible ancestor of @cgrp
232 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
233 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
234 * and @ancestor are accessible.
236 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
239 if (cgrp == ancestor)
246 static int cgroup_is_releasable(const struct cgroup *cgrp)
249 (1 << CGRP_RELEASABLE) |
250 (1 << CGRP_NOTIFY_ON_RELEASE);
251 return (cgrp->flags & bits) == bits;
254 static int notify_on_release(const struct cgroup *cgrp)
256 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
260 * for_each_css - iterate all css's of a cgroup
261 * @css: the iteration cursor
262 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
263 * @cgrp: the target cgroup to iterate css's of
265 * Should be called under cgroup_mutex.
267 #define for_each_css(css, ssid, cgrp) \
268 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
269 if (!((css) = rcu_dereference_check( \
270 (cgrp)->subsys[(ssid)], \
271 lockdep_is_held(&cgroup_tree_mutex) || \
272 lockdep_is_held(&cgroup_mutex)))) { } \
276 * for_each_subsys - iterate all enabled cgroup subsystems
277 * @ss: the iteration cursor
278 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
280 #define for_each_subsys(ss, ssid) \
281 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
282 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
284 /* iterate across the active hierarchies */
285 #define for_each_active_root(root) \
286 list_for_each_entry((root), &cgroup_roots, root_list)
289 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
290 * @cgrp: the cgroup to be checked for liveness
292 * On success, returns true; the mutex should be later unlocked. On
293 * failure returns false with no lock held.
295 static bool cgroup_lock_live_group(struct cgroup *cgrp)
297 mutex_lock(&cgroup_mutex);
298 if (cgroup_is_dead(cgrp)) {
299 mutex_unlock(&cgroup_mutex);
305 /* the list of cgroups eligible for automatic release. Protected by
306 * release_list_lock */
307 static LIST_HEAD(release_list);
308 static DEFINE_RAW_SPINLOCK(release_list_lock);
309 static void cgroup_release_agent(struct work_struct *work);
310 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
311 static void check_for_release(struct cgroup *cgrp);
314 * A cgroup can be associated with multiple css_sets as different tasks may
315 * belong to different cgroups on different hierarchies. In the other
316 * direction, a css_set is naturally associated with multiple cgroups.
317 * This M:N relationship is represented by the following link structure
318 * which exists for each association and allows traversing the associations
321 struct cgrp_cset_link {
322 /* the cgroup and css_set this link associates */
324 struct css_set *cset;
326 /* list of cgrp_cset_links anchored at cgrp->cset_links */
327 struct list_head cset_link;
329 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
330 struct list_head cgrp_link;
333 /* The default css_set - used by init and its children prior to any
334 * hierarchies being mounted. It contains a pointer to the root state
335 * for each subsystem. Also used to anchor the list of css_sets. Not
336 * reference-counted, to improve performance when child cgroups
337 * haven't been created.
340 static struct css_set init_css_set;
341 static struct cgrp_cset_link init_cgrp_cset_link;
344 * css_set_rwsem protects the list of css_set objects, and the chain of
345 * tasks off each css_set.
347 static DECLARE_RWSEM(css_set_rwsem);
348 static int css_set_count;
351 * hash table for cgroup groups. This improves the performance to find
352 * an existing css_set. This hash doesn't (currently) take into
353 * account cgroups in empty hierarchies.
355 #define CSS_SET_HASH_BITS 7
356 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
358 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
360 unsigned long key = 0UL;
361 struct cgroup_subsys *ss;
364 for_each_subsys(ss, i)
365 key += (unsigned long)css[i];
366 key = (key >> 16) ^ key;
371 static void put_css_set_locked(struct css_set *cset, bool taskexit)
373 struct cgrp_cset_link *link, *tmp_link;
375 lockdep_assert_held(&css_set_rwsem);
377 if (!atomic_dec_and_test(&cset->refcount))
380 /* This css_set is dead. unlink it and release cgroup refcounts */
381 hash_del(&cset->hlist);
384 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
385 struct cgroup *cgrp = link->cgrp;
387 list_del(&link->cset_link);
388 list_del(&link->cgrp_link);
390 /* @cgrp can't go away while we're holding css_set_rwsem */
391 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
393 set_bit(CGRP_RELEASABLE, &cgrp->flags);
394 check_for_release(cgrp);
400 kfree_rcu(cset, rcu_head);
403 static void put_css_set(struct css_set *cset, bool taskexit)
406 * Ensure that the refcount doesn't hit zero while any readers
407 * can see it. Similar to atomic_dec_and_lock(), but for an
410 if (atomic_add_unless(&cset->refcount, -1, 1))
413 down_write(&css_set_rwsem);
414 put_css_set_locked(cset, taskexit);
415 up_write(&css_set_rwsem);
419 * refcounted get/put for css_set objects
421 static inline void get_css_set(struct css_set *cset)
423 atomic_inc(&cset->refcount);
427 * compare_css_sets - helper function for find_existing_css_set().
428 * @cset: candidate css_set being tested
429 * @old_cset: existing css_set for a task
430 * @new_cgrp: cgroup that's being entered by the task
431 * @template: desired set of css pointers in css_set (pre-calculated)
433 * Returns true if "cset" matches "old_cset" except for the hierarchy
434 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
436 static bool compare_css_sets(struct css_set *cset,
437 struct css_set *old_cset,
438 struct cgroup *new_cgrp,
439 struct cgroup_subsys_state *template[])
441 struct list_head *l1, *l2;
443 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
444 /* Not all subsystems matched */
449 * Compare cgroup pointers in order to distinguish between
450 * different cgroups in heirarchies with no subsystems. We
451 * could get by with just this check alone (and skip the
452 * memcmp above) but on most setups the memcmp check will
453 * avoid the need for this more expensive check on almost all
457 l1 = &cset->cgrp_links;
458 l2 = &old_cset->cgrp_links;
460 struct cgrp_cset_link *link1, *link2;
461 struct cgroup *cgrp1, *cgrp2;
465 /* See if we reached the end - both lists are equal length. */
466 if (l1 == &cset->cgrp_links) {
467 BUG_ON(l2 != &old_cset->cgrp_links);
470 BUG_ON(l2 == &old_cset->cgrp_links);
472 /* Locate the cgroups associated with these links. */
473 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
474 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
477 /* Hierarchies should be linked in the same order. */
478 BUG_ON(cgrp1->root != cgrp2->root);
481 * If this hierarchy is the hierarchy of the cgroup
482 * that's changing, then we need to check that this
483 * css_set points to the new cgroup; if it's any other
484 * hierarchy, then this css_set should point to the
485 * same cgroup as the old css_set.
487 if (cgrp1->root == new_cgrp->root) {
488 if (cgrp1 != new_cgrp)
499 * find_existing_css_set - init css array and find the matching css_set
500 * @old_cset: the css_set that we're using before the cgroup transition
501 * @cgrp: the cgroup that we're moving into
502 * @template: out param for the new set of csses, should be clear on entry
504 static struct css_set *find_existing_css_set(struct css_set *old_cset,
506 struct cgroup_subsys_state *template[])
508 struct cgroupfs_root *root = cgrp->root;
509 struct cgroup_subsys *ss;
510 struct css_set *cset;
515 * Build the set of subsystem state objects that we want to see in the
516 * new css_set. while subsystems can change globally, the entries here
517 * won't change, so no need for locking.
519 for_each_subsys(ss, i) {
520 if (root->subsys_mask & (1UL << i)) {
521 /* Subsystem is in this hierarchy. So we want
522 * the subsystem state from the new
524 template[i] = cgroup_css(cgrp, ss);
526 /* Subsystem is not in this hierarchy, so we
527 * don't want to change the subsystem state */
528 template[i] = old_cset->subsys[i];
532 key = css_set_hash(template);
533 hash_for_each_possible(css_set_table, cset, hlist, key) {
534 if (!compare_css_sets(cset, old_cset, cgrp, template))
537 /* This css_set matches what we need */
541 /* No existing cgroup group matched */
545 static void free_cgrp_cset_links(struct list_head *links_to_free)
547 struct cgrp_cset_link *link, *tmp_link;
549 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
550 list_del(&link->cset_link);
556 * allocate_cgrp_cset_links - allocate cgrp_cset_links
557 * @count: the number of links to allocate
558 * @tmp_links: list_head the allocated links are put on
560 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
561 * through ->cset_link. Returns 0 on success or -errno.
563 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
565 struct cgrp_cset_link *link;
568 INIT_LIST_HEAD(tmp_links);
570 for (i = 0; i < count; i++) {
571 link = kzalloc(sizeof(*link), GFP_KERNEL);
573 free_cgrp_cset_links(tmp_links);
576 list_add(&link->cset_link, tmp_links);
582 * link_css_set - a helper function to link a css_set to a cgroup
583 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
584 * @cset: the css_set to be linked
585 * @cgrp: the destination cgroup
587 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
590 struct cgrp_cset_link *link;
592 BUG_ON(list_empty(tmp_links));
593 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
596 list_move(&link->cset_link, &cgrp->cset_links);
598 * Always add links to the tail of the list so that the list
599 * is sorted by order of hierarchy creation
601 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
605 * find_css_set - return a new css_set with one cgroup updated
606 * @old_cset: the baseline css_set
607 * @cgrp: the cgroup to be updated
609 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
610 * substituted into the appropriate hierarchy.
612 static struct css_set *find_css_set(struct css_set *old_cset,
615 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
616 struct css_set *cset;
617 struct list_head tmp_links;
618 struct cgrp_cset_link *link;
621 lockdep_assert_held(&cgroup_mutex);
623 /* First see if we already have a cgroup group that matches
625 down_read(&css_set_rwsem);
626 cset = find_existing_css_set(old_cset, cgrp, template);
629 up_read(&css_set_rwsem);
634 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
638 /* Allocate all the cgrp_cset_link objects that we'll need */
639 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
644 atomic_set(&cset->refcount, 1);
645 INIT_LIST_HEAD(&cset->cgrp_links);
646 INIT_LIST_HEAD(&cset->tasks);
647 INIT_HLIST_NODE(&cset->hlist);
649 /* Copy the set of subsystem state objects generated in
650 * find_existing_css_set() */
651 memcpy(cset->subsys, template, sizeof(cset->subsys));
653 down_write(&css_set_rwsem);
654 /* Add reference counts and links from the new css_set. */
655 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
656 struct cgroup *c = link->cgrp;
658 if (c->root == cgrp->root)
660 link_css_set(&tmp_links, cset, c);
663 BUG_ON(!list_empty(&tmp_links));
667 /* Add this cgroup group to the hash table */
668 key = css_set_hash(cset->subsys);
669 hash_add(css_set_table, &cset->hlist, key);
671 up_write(&css_set_rwsem);
676 static struct cgroupfs_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
678 struct cgroup *top_cgrp = kf_root->kn->priv;
680 return top_cgrp->root;
683 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
687 lockdep_assert_held(&cgroup_mutex);
689 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
694 root->hierarchy_id = id;
698 static void cgroup_exit_root_id(struct cgroupfs_root *root)
700 lockdep_assert_held(&cgroup_mutex);
702 if (root->hierarchy_id) {
703 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
704 root->hierarchy_id = 0;
708 static void cgroup_free_root(struct cgroupfs_root *root)
711 /* hierarhcy ID shoulid already have been released */
712 WARN_ON_ONCE(root->hierarchy_id);
714 idr_destroy(&root->cgroup_idr);
719 static void cgroup_destroy_root(struct cgroupfs_root *root)
721 struct cgroup *cgrp = &root->top_cgroup;
722 struct cgrp_cset_link *link, *tmp_link;
724 mutex_lock(&cgroup_tree_mutex);
725 mutex_lock(&cgroup_mutex);
727 BUG_ON(atomic_read(&root->nr_cgrps));
728 BUG_ON(!list_empty(&cgrp->children));
730 /* Rebind all subsystems back to the default hierarchy */
731 WARN_ON(rebind_subsystems(root, 0, root->subsys_mask));
734 * Release all the links from cset_links to this hierarchy's
737 down_write(&css_set_rwsem);
739 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
740 list_del(&link->cset_link);
741 list_del(&link->cgrp_link);
744 up_write(&css_set_rwsem);
746 if (!list_empty(&root->root_list)) {
747 list_del(&root->root_list);
751 cgroup_exit_root_id(root);
753 mutex_unlock(&cgroup_mutex);
754 mutex_unlock(&cgroup_tree_mutex);
756 kernfs_destroy_root(root->kf_root);
757 cgroup_free_root(root);
761 * Return the cgroup for "task" from the given hierarchy. Must be
762 * called with cgroup_mutex and css_set_rwsem held.
764 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
765 struct cgroupfs_root *root)
767 struct css_set *cset;
768 struct cgroup *res = NULL;
770 lockdep_assert_held(&cgroup_mutex);
771 lockdep_assert_held(&css_set_rwsem);
774 * No need to lock the task - since we hold cgroup_mutex the
775 * task can't change groups, so the only thing that can happen
776 * is that it exits and its css is set back to init_css_set.
778 cset = task_css_set(task);
779 if (cset == &init_css_set) {
780 res = &root->top_cgroup;
782 struct cgrp_cset_link *link;
784 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
785 struct cgroup *c = link->cgrp;
787 if (c->root == root) {
799 * There is one global cgroup mutex. We also require taking
800 * task_lock() when dereferencing a task's cgroup subsys pointers.
801 * See "The task_lock() exception", at the end of this comment.
803 * A task must hold cgroup_mutex to modify cgroups.
805 * Any task can increment and decrement the count field without lock.
806 * So in general, code holding cgroup_mutex can't rely on the count
807 * field not changing. However, if the count goes to zero, then only
808 * cgroup_attach_task() can increment it again. Because a count of zero
809 * means that no tasks are currently attached, therefore there is no
810 * way a task attached to that cgroup can fork (the other way to
811 * increment the count). So code holding cgroup_mutex can safely
812 * assume that if the count is zero, it will stay zero. Similarly, if
813 * a task holds cgroup_mutex on a cgroup with zero count, it
814 * knows that the cgroup won't be removed, as cgroup_rmdir()
817 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
818 * (usually) take cgroup_mutex. These are the two most performance
819 * critical pieces of code here. The exception occurs on cgroup_exit(),
820 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
821 * is taken, and if the cgroup count is zero, a usermode call made
822 * to the release agent with the name of the cgroup (path relative to
823 * the root of cgroup file system) as the argument.
825 * A cgroup can only be deleted if both its 'count' of using tasks
826 * is zero, and its list of 'children' cgroups is empty. Since all
827 * tasks in the system use _some_ cgroup, and since there is always at
828 * least one task in the system (init, pid == 1), therefore, top_cgroup
829 * always has either children cgroups and/or using tasks. So we don't
830 * need a special hack to ensure that top_cgroup cannot be deleted.
832 * The task_lock() exception
834 * The need for this exception arises from the action of
835 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
836 * another. It does so using cgroup_mutex, however there are
837 * several performance critical places that need to reference
838 * task->cgroup without the expense of grabbing a system global
839 * mutex. Therefore except as noted below, when dereferencing or, as
840 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
841 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
842 * the task_struct routinely used for such matters.
844 * P.S. One more locking exception. RCU is used to guard the
845 * update of a tasks cgroup pointer by cgroup_attach_task()
848 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
849 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
850 static const struct file_operations proc_cgroupstats_operations;
852 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
855 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
856 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
857 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
858 cft->ss->name, cft->name);
860 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
865 * cgroup_file_mode - deduce file mode of a control file
866 * @cft: the control file in question
868 * returns cft->mode if ->mode is not 0
869 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
870 * returns S_IRUGO if it has only a read handler
871 * returns S_IWUSR if it has only a write hander
873 static umode_t cgroup_file_mode(const struct cftype *cft)
880 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
883 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
890 static void cgroup_free_fn(struct work_struct *work)
892 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
894 atomic_dec(&cgrp->root->nr_cgrps);
895 cgroup_pidlist_destroy_all(cgrp);
899 * We get a ref to the parent, and put the ref when this
900 * cgroup is being freed, so it's guaranteed that the
901 * parent won't be destroyed before its children.
903 cgroup_put(cgrp->parent);
904 kernfs_put(cgrp->kn);
908 * This is top cgroup's refcnt reaching zero, which
909 * indicates that the root should be released.
911 cgroup_destroy_root(cgrp->root);
915 static void cgroup_free_rcu(struct rcu_head *head)
917 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
919 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
920 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
923 static void cgroup_get(struct cgroup *cgrp)
925 WARN_ON_ONCE(cgroup_is_dead(cgrp));
926 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
927 atomic_inc(&cgrp->refcnt);
930 static void cgroup_put(struct cgroup *cgrp)
932 if (!atomic_dec_and_test(&cgrp->refcnt))
934 if (WARN_ON_ONCE(cgrp->parent && !cgroup_is_dead(cgrp)))
938 * XXX: cgrp->id is only used to look up css's. As cgroup and
939 * css's lifetimes will be decoupled, it should be made
940 * per-subsystem and moved to css->id so that lookups are
941 * successful until the target css is released.
943 mutex_lock(&cgroup_mutex);
944 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
945 mutex_unlock(&cgroup_mutex);
948 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
951 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
953 char name[CGROUP_FILE_NAME_MAX];
955 lockdep_assert_held(&cgroup_tree_mutex);
956 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
960 * cgroup_clear_dir - remove subsys files in a cgroup directory
961 * @cgrp: target cgroup
962 * @subsys_mask: mask of the subsystem ids whose files should be removed
964 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
966 struct cgroup_subsys *ss;
969 for_each_subsys(ss, i) {
972 if (!test_bit(i, &subsys_mask))
974 list_for_each_entry(cfts, &ss->cfts, node)
975 cgroup_addrm_files(cgrp, cfts, false);
979 static int rebind_subsystems(struct cgroupfs_root *root,
980 unsigned long added_mask, unsigned removed_mask)
982 struct cgroup *cgrp = &root->top_cgroup;
983 struct cgroup_subsys *ss;
986 lockdep_assert_held(&cgroup_tree_mutex);
987 lockdep_assert_held(&cgroup_mutex);
989 /* Check that any added subsystems are currently free */
990 for_each_subsys(ss, i)
991 if ((added_mask & (1 << i)) && ss->root != &cgroup_dummy_root)
994 ret = cgroup_populate_dir(cgrp, added_mask);
999 * Nothing can fail from this point on. Remove files for the
1000 * removed subsystems and rebind each subsystem.
1002 mutex_unlock(&cgroup_mutex);
1003 cgroup_clear_dir(cgrp, removed_mask);
1004 mutex_lock(&cgroup_mutex);
1006 for_each_subsys(ss, i) {
1007 unsigned long bit = 1UL << i;
1009 if (bit & added_mask) {
1010 /* We're binding this subsystem to this hierarchy */
1011 BUG_ON(cgroup_css(cgrp, ss));
1012 BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
1013 BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
1015 rcu_assign_pointer(cgrp->subsys[i],
1016 cgroup_css(cgroup_dummy_top, ss));
1017 cgroup_css(cgrp, ss)->cgroup = cgrp;
1021 ss->bind(cgroup_css(cgrp, ss));
1023 /* refcount was already taken, and we're keeping it */
1024 root->subsys_mask |= bit;
1025 } else if (bit & removed_mask) {
1026 /* We're removing this subsystem */
1027 BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1028 BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1031 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1033 cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1034 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1036 cgroup_subsys[i]->root = &cgroup_dummy_root;
1037 root->subsys_mask &= ~bit;
1041 kernfs_activate(cgrp->kn);
1045 static int cgroup_show_options(struct seq_file *seq,
1046 struct kernfs_root *kf_root)
1048 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1049 struct cgroup_subsys *ss;
1052 for_each_subsys(ss, ssid)
1053 if (root->subsys_mask & (1 << ssid))
1054 seq_printf(seq, ",%s", ss->name);
1055 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1056 seq_puts(seq, ",sane_behavior");
1057 if (root->flags & CGRP_ROOT_NOPREFIX)
1058 seq_puts(seq, ",noprefix");
1059 if (root->flags & CGRP_ROOT_XATTR)
1060 seq_puts(seq, ",xattr");
1062 spin_lock(&release_agent_path_lock);
1063 if (strlen(root->release_agent_path))
1064 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1065 spin_unlock(&release_agent_path_lock);
1067 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1068 seq_puts(seq, ",clone_children");
1069 if (strlen(root->name))
1070 seq_printf(seq, ",name=%s", root->name);
1074 struct cgroup_sb_opts {
1075 unsigned long subsys_mask;
1076 unsigned long flags;
1077 char *release_agent;
1078 bool cpuset_clone_children;
1080 /* User explicitly requested empty subsystem */
1085 * Convert a hierarchy specifier into a bitmask of subsystems and
1086 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1087 * array. This function takes refcounts on subsystems to be used, unless it
1088 * returns error, in which case no refcounts are taken.
1090 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1092 char *token, *o = data;
1093 bool all_ss = false, one_ss = false;
1094 unsigned long mask = (unsigned long)-1;
1095 struct cgroup_subsys *ss;
1098 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1100 #ifdef CONFIG_CPUSETS
1101 mask = ~(1UL << cpuset_cgrp_id);
1104 memset(opts, 0, sizeof(*opts));
1106 while ((token = strsep(&o, ",")) != NULL) {
1109 if (!strcmp(token, "none")) {
1110 /* Explicitly have no subsystems */
1114 if (!strcmp(token, "all")) {
1115 /* Mutually exclusive option 'all' + subsystem name */
1121 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1122 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1125 if (!strcmp(token, "noprefix")) {
1126 opts->flags |= CGRP_ROOT_NOPREFIX;
1129 if (!strcmp(token, "clone_children")) {
1130 opts->cpuset_clone_children = true;
1133 if (!strcmp(token, "xattr")) {
1134 opts->flags |= CGRP_ROOT_XATTR;
1137 if (!strncmp(token, "release_agent=", 14)) {
1138 /* Specifying two release agents is forbidden */
1139 if (opts->release_agent)
1141 opts->release_agent =
1142 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1143 if (!opts->release_agent)
1147 if (!strncmp(token, "name=", 5)) {
1148 const char *name = token + 5;
1149 /* Can't specify an empty name */
1152 /* Must match [\w.-]+ */
1153 for (i = 0; i < strlen(name); i++) {
1157 if ((c == '.') || (c == '-') || (c == '_'))
1161 /* Specifying two names is forbidden */
1164 opts->name = kstrndup(name,
1165 MAX_CGROUP_ROOT_NAMELEN - 1,
1173 for_each_subsys(ss, i) {
1174 if (strcmp(token, ss->name))
1179 /* Mutually exclusive option 'all' + subsystem name */
1182 set_bit(i, &opts->subsys_mask);
1187 if (i == CGROUP_SUBSYS_COUNT)
1192 * If the 'all' option was specified select all the subsystems,
1193 * otherwise if 'none', 'name=' and a subsystem name options
1194 * were not specified, let's default to 'all'
1196 if (all_ss || (!one_ss && !opts->none && !opts->name))
1197 for_each_subsys(ss, i)
1199 set_bit(i, &opts->subsys_mask);
1201 /* Consistency checks */
1203 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1204 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1206 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1207 opts->cpuset_clone_children || opts->release_agent ||
1209 pr_err("cgroup: sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1215 * Option noprefix was introduced just for backward compatibility
1216 * with the old cpuset, so we allow noprefix only if mounting just
1217 * the cpuset subsystem.
1219 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1223 /* Can't specify "none" and some subsystems */
1224 if (opts->subsys_mask && opts->none)
1228 * We either have to specify by name or by subsystems. (So all
1229 * empty hierarchies must have a name).
1231 if (!opts->subsys_mask && !opts->name)
1237 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1240 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1241 struct cgroup_sb_opts opts;
1242 unsigned long added_mask, removed_mask;
1244 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1245 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1249 mutex_lock(&cgroup_tree_mutex);
1250 mutex_lock(&cgroup_mutex);
1252 /* See what subsystems are wanted */
1253 ret = parse_cgroupfs_options(data, &opts);
1257 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1258 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1259 task_tgid_nr(current), current->comm);
1261 added_mask = opts.subsys_mask & ~root->subsys_mask;
1262 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1264 /* Don't allow flags or name to change at remount */
1265 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1266 (opts.name && strcmp(opts.name, root->name))) {
1267 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1268 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1269 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1274 /* remounting is not allowed for populated hierarchies */
1275 if (!list_empty(&root->top_cgroup.children)) {
1280 ret = rebind_subsystems(root, added_mask, removed_mask);
1284 if (opts.release_agent) {
1285 spin_lock(&release_agent_path_lock);
1286 strcpy(root->release_agent_path, opts.release_agent);
1287 spin_unlock(&release_agent_path_lock);
1290 kfree(opts.release_agent);
1292 mutex_unlock(&cgroup_mutex);
1293 mutex_unlock(&cgroup_tree_mutex);
1298 * To reduce the fork() overhead for systems that are not actually using
1299 * their cgroups capability, we don't maintain the lists running through
1300 * each css_set to its tasks until we see the list actually used - in other
1301 * words after the first mount.
1303 static bool use_task_css_set_links __read_mostly;
1305 static void cgroup_enable_task_cg_lists(void)
1307 struct task_struct *p, *g;
1309 down_write(&css_set_rwsem);
1311 if (use_task_css_set_links)
1314 use_task_css_set_links = true;
1317 * We need tasklist_lock because RCU is not safe against
1318 * while_each_thread(). Besides, a forking task that has passed
1319 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1320 * is not guaranteed to have its child immediately visible in the
1321 * tasklist if we walk through it with RCU.
1323 read_lock(&tasklist_lock);
1324 do_each_thread(g, p) {
1327 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1328 task_css_set(p) != &init_css_set);
1331 * We should check if the process is exiting, otherwise
1332 * it will race with cgroup_exit() in that the list
1333 * entry won't be deleted though the process has exited.
1335 if (!(p->flags & PF_EXITING))
1336 list_add(&p->cg_list, &task_css_set(p)->tasks);
1339 } while_each_thread(g, p);
1340 read_unlock(&tasklist_lock);
1342 up_write(&css_set_rwsem);
1345 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1347 atomic_set(&cgrp->refcnt, 1);
1348 INIT_LIST_HEAD(&cgrp->sibling);
1349 INIT_LIST_HEAD(&cgrp->children);
1350 INIT_LIST_HEAD(&cgrp->cset_links);
1351 INIT_LIST_HEAD(&cgrp->release_list);
1352 INIT_LIST_HEAD(&cgrp->pidlists);
1353 mutex_init(&cgrp->pidlist_mutex);
1354 cgrp->dummy_css.cgroup = cgrp;
1357 static void init_cgroup_root(struct cgroupfs_root *root)
1359 struct cgroup *cgrp = &root->top_cgroup;
1361 INIT_LIST_HEAD(&root->root_list);
1362 atomic_set(&root->nr_cgrps, 1);
1364 init_cgroup_housekeeping(cgrp);
1365 idr_init(&root->cgroup_idr);
1368 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1370 struct cgroupfs_root *root;
1372 if (!opts->subsys_mask && !opts->none)
1373 return ERR_PTR(-EINVAL);
1375 root = kzalloc(sizeof(*root), GFP_KERNEL);
1377 return ERR_PTR(-ENOMEM);
1379 init_cgroup_root(root);
1381 root->flags = opts->flags;
1382 if (opts->release_agent)
1383 strcpy(root->release_agent_path, opts->release_agent);
1385 strcpy(root->name, opts->name);
1386 if (opts->cpuset_clone_children)
1387 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1391 static int cgroup_setup_root(struct cgroupfs_root *root, unsigned long ss_mask)
1393 LIST_HEAD(tmp_links);
1394 struct cgroup *root_cgrp = &root->top_cgroup;
1395 struct css_set *cset;
1398 lockdep_assert_held(&cgroup_tree_mutex);
1399 lockdep_assert_held(&cgroup_mutex);
1401 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1404 root_cgrp->id = ret;
1407 * We're accessing css_set_count without locking css_set_rwsem here,
1408 * but that's OK - it can only be increased by someone holding
1409 * cgroup_lock, and that's us. The worst that can happen is that we
1410 * have some link structures left over
1412 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1416 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1417 ret = cgroup_init_root_id(root, 2, 0);
1421 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1422 KERNFS_ROOT_CREATE_DEACTIVATED,
1424 if (IS_ERR(root->kf_root)) {
1425 ret = PTR_ERR(root->kf_root);
1428 root_cgrp->kn = root->kf_root->kn;
1430 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1434 ret = rebind_subsystems(root, ss_mask, 0);
1439 * There must be no failure case after here, since rebinding takes
1440 * care of subsystems' refcounts, which are explicitly dropped in
1441 * the failure exit path.
1443 list_add(&root->root_list, &cgroup_roots);
1444 cgroup_root_count++;
1447 * Link the top cgroup in this hierarchy into all the css_set
1450 down_write(&css_set_rwsem);
1451 hash_for_each(css_set_table, i, cset, hlist)
1452 link_css_set(&tmp_links, cset, root_cgrp);
1453 up_write(&css_set_rwsem);
1455 BUG_ON(!list_empty(&root_cgrp->children));
1456 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1458 kernfs_activate(root_cgrp->kn);
1463 kernfs_destroy_root(root->kf_root);
1464 root->kf_root = NULL;
1466 cgroup_exit_root_id(root);
1468 free_cgrp_cset_links(&tmp_links);
1472 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1473 int flags, const char *unused_dev_name,
1476 struct cgroupfs_root *root;
1477 struct cgroup_sb_opts opts;
1478 struct dentry *dentry;
1482 * The first time anyone tries to mount a cgroup, enable the list
1483 * linking each css_set to its tasks and fix up all existing tasks.
1485 if (!use_task_css_set_links)
1486 cgroup_enable_task_cg_lists();
1488 mutex_lock(&cgroup_tree_mutex);
1489 mutex_lock(&cgroup_mutex);
1491 /* First find the desired set of subsystems */
1492 ret = parse_cgroupfs_options(data, &opts);
1496 /* look for a matching existing root */
1497 for_each_active_root(root) {
1498 bool name_match = false;
1501 * If we asked for a name then it must match. Also, if
1502 * name matches but sybsys_mask doesn't, we should fail.
1503 * Remember whether name matched.
1506 if (strcmp(opts.name, root->name))
1512 * If we asked for subsystems (or explicitly for no
1513 * subsystems) then they must match.
1515 if ((opts.subsys_mask || opts.none) &&
1516 (opts.subsys_mask != root->subsys_mask)) {
1523 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1524 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1525 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1529 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1534 * A root's lifetime is governed by its top cgroup. Zero
1535 * ref indicate that the root is being destroyed. Wait for
1536 * destruction to complete so that the subsystems are free.
1537 * We can use wait_queue for the wait but this path is
1538 * super cold. Let's just sleep for a bit and retry.
1540 if (!atomic_inc_not_zero(&root->top_cgroup.refcnt)) {
1541 mutex_unlock(&cgroup_mutex);
1542 mutex_unlock(&cgroup_tree_mutex);
1551 /* no such thing, create a new one */
1552 root = cgroup_root_from_opts(&opts);
1554 ret = PTR_ERR(root);
1558 ret = cgroup_setup_root(root, opts.subsys_mask);
1560 cgroup_free_root(root);
1563 mutex_unlock(&cgroup_mutex);
1564 mutex_unlock(&cgroup_tree_mutex);
1566 kfree(opts.release_agent);
1570 return ERR_PTR(ret);
1572 dentry = kernfs_mount(fs_type, flags, root->kf_root);
1574 cgroup_put(&root->top_cgroup);
1578 static void cgroup_kill_sb(struct super_block *sb)
1580 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1581 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1583 cgroup_put(&root->top_cgroup);
1587 static struct file_system_type cgroup_fs_type = {
1589 .mount = cgroup_mount,
1590 .kill_sb = cgroup_kill_sb,
1593 static struct kobject *cgroup_kobj;
1596 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1597 * @task: target task
1598 * @buf: the buffer to write the path into
1599 * @buflen: the length of the buffer
1601 * Determine @task's cgroup on the first (the one with the lowest non-zero
1602 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1603 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1604 * cgroup controller callbacks.
1606 * Return value is the same as kernfs_path().
1608 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1610 struct cgroupfs_root *root;
1611 struct cgroup *cgrp;
1612 int hierarchy_id = 1;
1615 mutex_lock(&cgroup_mutex);
1616 down_read(&css_set_rwsem);
1618 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1621 cgrp = task_cgroup_from_root(task, root);
1622 path = cgroup_path(cgrp, buf, buflen);
1624 /* if no hierarchy exists, everyone is in "/" */
1625 if (strlcpy(buf, "/", buflen) < buflen)
1629 up_read(&css_set_rwsem);
1630 mutex_unlock(&cgroup_mutex);
1633 EXPORT_SYMBOL_GPL(task_cgroup_path);
1636 * Control Group taskset
1638 struct task_and_cgroup {
1639 struct task_struct *task;
1640 struct cgroup *cgrp;
1641 struct css_set *cset;
1644 struct cgroup_taskset {
1645 struct task_and_cgroup single;
1646 struct flex_array *tc_array;
1652 * cgroup_taskset_first - reset taskset and return the first task
1653 * @tset: taskset of interest
1655 * @tset iteration is initialized and the first task is returned.
1657 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1659 if (tset->tc_array) {
1661 return cgroup_taskset_next(tset);
1663 return tset->single.task;
1668 * cgroup_taskset_next - iterate to the next task in taskset
1669 * @tset: taskset of interest
1671 * Return the next task in @tset. Iteration must have been initialized
1672 * with cgroup_taskset_first().
1674 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1676 struct task_and_cgroup *tc;
1678 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1681 tc = flex_array_get(tset->tc_array, tset->idx++);
1686 * cgroup_task_migrate - move a task from one cgroup to another.
1687 * @old_cgrp; the cgroup @tsk is being migrated from
1688 * @tsk: the task being migrated
1689 * @new_cset: the new css_set @tsk is being attached to
1691 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1693 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1694 struct task_struct *tsk,
1695 struct css_set *new_cset)
1697 struct css_set *old_cset;
1699 lockdep_assert_held(&cgroup_mutex);
1700 lockdep_assert_held(&css_set_rwsem);
1703 * We are synchronized through threadgroup_lock() against PF_EXITING
1704 * setting such that we can't race against cgroup_exit() changing the
1705 * css_set to init_css_set and dropping the old one.
1707 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1708 old_cset = task_css_set(tsk);
1711 rcu_assign_pointer(tsk->cgroups, new_cset);
1714 list_move(&tsk->cg_list, &new_cset->tasks);
1717 * We just gained a reference on old_cset by taking it from the
1718 * task. As trading it for new_cset is protected by cgroup_mutex,
1719 * we're safe to drop it here; it will be freed under RCU.
1721 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1722 put_css_set_locked(old_cset, false);
1726 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1727 * @cgrp: the cgroup to attach to
1728 * @leader: the task or the leader of the threadgroup to be attached
1729 * @threadgroup: attach the whole threadgroup?
1731 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1732 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1734 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *leader,
1737 int ret, i, group_size;
1738 struct cgroupfs_root *root = cgrp->root;
1739 struct cgroup_subsys_state *css, *failed_css = NULL;
1740 /* threadgroup list cursor and array */
1741 struct task_struct *task;
1742 struct task_and_cgroup *tc;
1743 struct flex_array *group;
1744 struct cgroup_taskset tset = { };
1747 * step 0: in order to do expensive, possibly blocking operations for
1748 * every thread, we cannot iterate the thread group list, since it needs
1749 * rcu or tasklist locked. instead, build an array of all threads in the
1750 * group - group_rwsem prevents new threads from appearing, and if
1751 * threads exit, this will just be an over-estimate.
1754 group_size = get_nr_threads(leader);
1757 /* flex_array supports very large thread-groups better than kmalloc. */
1758 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1761 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1762 ret = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1764 goto out_free_group_list;
1768 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1769 * already PF_EXITING could be freed from underneath us unless we
1770 * take an rcu_read_lock.
1772 down_read(&css_set_rwsem);
1776 struct task_and_cgroup ent;
1778 /* @task either already exited or can't exit until the end */
1779 if (task->flags & PF_EXITING)
1782 /* as per above, nr_threads may decrease, but not increase. */
1783 BUG_ON(i >= group_size);
1785 ent.cgrp = task_cgroup_from_root(task, root);
1786 /* nothing to do if this task is already in the cgroup */
1787 if (ent.cgrp == cgrp)
1790 * saying GFP_ATOMIC has no effect here because we did prealloc
1791 * earlier, but it's good form to communicate our expectations.
1793 ret = flex_array_put(group, i, &ent, GFP_ATOMIC);
1799 } while_each_thread(leader, task);
1801 up_read(&css_set_rwsem);
1802 /* remember the number of threads in the array for later. */
1804 tset.tc_array = group;
1805 tset.tc_array_len = group_size;
1807 /* methods shouldn't be called if no task is actually migrating */
1810 goto out_free_group_list;
1813 * step 1: check that we can legitimately attach to the cgroup.
1815 for_each_css(css, i, cgrp) {
1816 if (css->ss->can_attach) {
1817 ret = css->ss->can_attach(css, &tset);
1820 goto out_cancel_attach;
1826 * step 2: make sure css_sets exist for all threads to be migrated.
1827 * we use find_css_set, which allocates a new one if necessary.
1829 for (i = 0; i < group_size; i++) {
1830 struct css_set *old_cset;
1832 tc = flex_array_get(group, i);
1833 old_cset = task_css_set(tc->task);
1834 tc->cset = find_css_set(old_cset, cgrp);
1837 goto out_put_css_set_refs;
1842 * step 3: now that we're guaranteed success wrt the css_sets,
1843 * proceed to move all tasks to the new cgroup. There are no
1844 * failure cases after here, so this is the commit point.
1846 down_write(&css_set_rwsem);
1847 for (i = 0; i < group_size; i++) {
1848 tc = flex_array_get(group, i);
1849 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
1851 up_write(&css_set_rwsem);
1852 /* nothing is sensitive to fork() after this point. */
1855 * step 4: do subsystem attach callbacks.
1857 for_each_css(css, i, cgrp)
1858 if (css->ss->attach)
1859 css->ss->attach(css, &tset);
1862 * step 5: success! and cleanup
1865 out_put_css_set_refs:
1867 for (i = 0; i < group_size; i++) {
1868 tc = flex_array_get(group, i);
1871 put_css_set(tc->cset, false);
1876 for_each_css(css, i, cgrp) {
1877 if (css == failed_css)
1879 if (css->ss->cancel_attach)
1880 css->ss->cancel_attach(css, &tset);
1883 out_free_group_list:
1884 flex_array_free(group);
1889 * Find the task_struct of the task to attach by vpid and pass it along to the
1890 * function to attach either it or all tasks in its threadgroup. Will lock
1891 * cgroup_mutex and threadgroup; may take task_lock of task.
1893 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
1895 struct task_struct *tsk;
1896 const struct cred *cred = current_cred(), *tcred;
1899 if (!cgroup_lock_live_group(cgrp))
1905 tsk = find_task_by_vpid(pid);
1909 goto out_unlock_cgroup;
1912 * even if we're attaching all tasks in the thread group, we
1913 * only need to check permissions on one of them.
1915 tcred = __task_cred(tsk);
1916 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
1917 !uid_eq(cred->euid, tcred->uid) &&
1918 !uid_eq(cred->euid, tcred->suid)) {
1921 goto out_unlock_cgroup;
1927 tsk = tsk->group_leader;
1930 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
1931 * trapped in a cpuset, or RT worker may be born in a cgroup
1932 * with no rt_runtime allocated. Just say no.
1934 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
1937 goto out_unlock_cgroup;
1940 get_task_struct(tsk);
1943 threadgroup_lock(tsk);
1945 if (!thread_group_leader(tsk)) {
1947 * a race with de_thread from another thread's exec()
1948 * may strip us of our leadership, if this happens,
1949 * there is no choice but to throw this task away and
1950 * try again; this is
1951 * "double-double-toil-and-trouble-check locking".
1953 threadgroup_unlock(tsk);
1954 put_task_struct(tsk);
1955 goto retry_find_task;
1959 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
1961 threadgroup_unlock(tsk);
1963 put_task_struct(tsk);
1965 mutex_unlock(&cgroup_mutex);
1970 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1971 * @from: attach to all cgroups of a given task
1972 * @tsk: the task to be attached
1974 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1976 struct cgroupfs_root *root;
1979 mutex_lock(&cgroup_mutex);
1980 for_each_active_root(root) {
1981 struct cgroup *from_cgrp;
1983 down_read(&css_set_rwsem);
1984 from_cgrp = task_cgroup_from_root(from, root);
1985 up_read(&css_set_rwsem);
1987 retval = cgroup_attach_task(from_cgrp, tsk, false);
1991 mutex_unlock(&cgroup_mutex);
1995 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
1997 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
1998 struct cftype *cft, u64 pid)
2000 return attach_task_by_pid(css->cgroup, pid, false);
2003 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2004 struct cftype *cft, u64 tgid)
2006 return attach_task_by_pid(css->cgroup, tgid, true);
2009 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2010 struct cftype *cft, const char *buffer)
2012 struct cgroupfs_root *root = css->cgroup->root;
2014 BUILD_BUG_ON(sizeof(root->release_agent_path) < PATH_MAX);
2015 if (!cgroup_lock_live_group(css->cgroup))
2017 spin_lock(&release_agent_path_lock);
2018 strlcpy(root->release_agent_path, buffer,
2019 sizeof(root->release_agent_path));
2020 spin_unlock(&release_agent_path_lock);
2021 mutex_unlock(&cgroup_mutex);
2025 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2027 struct cgroup *cgrp = seq_css(seq)->cgroup;
2029 if (!cgroup_lock_live_group(cgrp))
2031 seq_puts(seq, cgrp->root->release_agent_path);
2032 seq_putc(seq, '\n');
2033 mutex_unlock(&cgroup_mutex);
2037 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2039 struct cgroup *cgrp = seq_css(seq)->cgroup;
2041 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2045 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2046 size_t nbytes, loff_t off)
2048 struct cgroup *cgrp = of->kn->parent->priv;
2049 struct cftype *cft = of->kn->priv;
2050 struct cgroup_subsys_state *css;
2054 * kernfs guarantees that a file isn't deleted with operations in
2055 * flight, which means that the matching css is and stays alive and
2056 * doesn't need to be pinned. The RCU locking is not necessary
2057 * either. It's just for the convenience of using cgroup_css().
2060 css = cgroup_css(cgrp, cft->ss);
2063 if (cft->write_string) {
2064 ret = cft->write_string(css, cft, strstrip(buf));
2065 } else if (cft->write_u64) {
2066 unsigned long long v;
2067 ret = kstrtoull(buf, 0, &v);
2069 ret = cft->write_u64(css, cft, v);
2070 } else if (cft->write_s64) {
2072 ret = kstrtoll(buf, 0, &v);
2074 ret = cft->write_s64(css, cft, v);
2075 } else if (cft->trigger) {
2076 ret = cft->trigger(css, (unsigned int)cft->private);
2081 return ret ?: nbytes;
2084 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2086 return seq_cft(seq)->seq_start(seq, ppos);
2089 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2091 return seq_cft(seq)->seq_next(seq, v, ppos);
2094 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2096 seq_cft(seq)->seq_stop(seq, v);
2099 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2101 struct cftype *cft = seq_cft(m);
2102 struct cgroup_subsys_state *css = seq_css(m);
2105 return cft->seq_show(m, arg);
2108 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2109 else if (cft->read_s64)
2110 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2116 static struct kernfs_ops cgroup_kf_single_ops = {
2117 .atomic_write_len = PAGE_SIZE,
2118 .write = cgroup_file_write,
2119 .seq_show = cgroup_seqfile_show,
2122 static struct kernfs_ops cgroup_kf_ops = {
2123 .atomic_write_len = PAGE_SIZE,
2124 .write = cgroup_file_write,
2125 .seq_start = cgroup_seqfile_start,
2126 .seq_next = cgroup_seqfile_next,
2127 .seq_stop = cgroup_seqfile_stop,
2128 .seq_show = cgroup_seqfile_show,
2132 * cgroup_rename - Only allow simple rename of directories in place.
2134 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2135 const char *new_name_str)
2137 struct cgroup *cgrp = kn->priv;
2140 if (kernfs_type(kn) != KERNFS_DIR)
2142 if (kn->parent != new_parent)
2146 * This isn't a proper migration and its usefulness is very
2147 * limited. Disallow if sane_behavior.
2149 if (cgroup_sane_behavior(cgrp))
2152 mutex_lock(&cgroup_tree_mutex);
2153 mutex_lock(&cgroup_mutex);
2155 ret = kernfs_rename(kn, new_parent, new_name_str);
2157 mutex_unlock(&cgroup_mutex);
2158 mutex_unlock(&cgroup_tree_mutex);
2162 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2164 char name[CGROUP_FILE_NAME_MAX];
2165 struct kernfs_node *kn;
2166 struct lock_class_key *key = NULL;
2168 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2169 key = &cft->lockdep_key;
2171 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2172 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2174 return PTR_ERR_OR_ZERO(kn);
2178 * cgroup_addrm_files - add or remove files to a cgroup directory
2179 * @cgrp: the target cgroup
2180 * @cfts: array of cftypes to be added
2181 * @is_add: whether to add or remove
2183 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2184 * For removals, this function never fails. If addition fails, this
2185 * function doesn't remove files already added. The caller is responsible
2188 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2194 lockdep_assert_held(&cgroup_tree_mutex);
2196 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2197 /* does cft->flags tell us to skip this file on @cgrp? */
2198 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2200 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2202 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2206 ret = cgroup_add_file(cgrp, cft);
2208 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2213 cgroup_rm_file(cgrp, cft);
2219 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2222 struct cgroup_subsys *ss = cfts[0].ss;
2223 struct cgroup *root = &ss->root->top_cgroup;
2224 struct cgroup_subsys_state *css;
2227 lockdep_assert_held(&cgroup_tree_mutex);
2229 /* don't bother if @ss isn't attached */
2230 if (ss->root == &cgroup_dummy_root)
2233 /* add/rm files for all cgroups created before */
2234 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2235 struct cgroup *cgrp = css->cgroup;
2237 if (cgroup_is_dead(cgrp))
2240 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2246 kernfs_activate(root->kn);
2250 static void cgroup_exit_cftypes(struct cftype *cfts)
2254 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2255 /* free copy for custom atomic_write_len, see init_cftypes() */
2256 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2263 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2267 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2268 struct kernfs_ops *kf_ops;
2270 WARN_ON(cft->ss || cft->kf_ops);
2273 kf_ops = &cgroup_kf_ops;
2275 kf_ops = &cgroup_kf_single_ops;
2278 * Ugh... if @cft wants a custom max_write_len, we need to
2279 * make a copy of kf_ops to set its atomic_write_len.
2281 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
2282 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
2284 cgroup_exit_cftypes(cfts);
2287 kf_ops->atomic_write_len = cft->max_write_len;
2290 cft->kf_ops = kf_ops;
2297 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
2299 lockdep_assert_held(&cgroup_tree_mutex);
2301 if (!cfts || !cfts[0].ss)
2304 list_del(&cfts->node);
2305 cgroup_apply_cftypes(cfts, false);
2306 cgroup_exit_cftypes(cfts);
2311 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2312 * @cfts: zero-length name terminated array of cftypes
2314 * Unregister @cfts. Files described by @cfts are removed from all
2315 * existing cgroups and all future cgroups won't have them either. This
2316 * function can be called anytime whether @cfts' subsys is attached or not.
2318 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2321 int cgroup_rm_cftypes(struct cftype *cfts)
2325 mutex_lock(&cgroup_tree_mutex);
2326 ret = cgroup_rm_cftypes_locked(cfts);
2327 mutex_unlock(&cgroup_tree_mutex);
2332 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2333 * @ss: target cgroup subsystem
2334 * @cfts: zero-length name terminated array of cftypes
2336 * Register @cfts to @ss. Files described by @cfts are created for all
2337 * existing cgroups to which @ss is attached and all future cgroups will
2338 * have them too. This function can be called anytime whether @ss is
2341 * Returns 0 on successful registration, -errno on failure. Note that this
2342 * function currently returns 0 as long as @cfts registration is successful
2343 * even if some file creation attempts on existing cgroups fail.
2345 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2349 ret = cgroup_init_cftypes(ss, cfts);
2353 mutex_lock(&cgroup_tree_mutex);
2355 list_add_tail(&cfts->node, &ss->cfts);
2356 ret = cgroup_apply_cftypes(cfts, true);
2358 cgroup_rm_cftypes_locked(cfts);
2360 mutex_unlock(&cgroup_tree_mutex);
2365 * cgroup_task_count - count the number of tasks in a cgroup.
2366 * @cgrp: the cgroup in question
2368 * Return the number of tasks in the cgroup.
2370 static int cgroup_task_count(const struct cgroup *cgrp)
2373 struct cgrp_cset_link *link;
2375 down_read(&css_set_rwsem);
2376 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2377 count += atomic_read(&link->cset->refcount);
2378 up_read(&css_set_rwsem);
2383 * css_next_child - find the next child of a given css
2384 * @pos_css: the current position (%NULL to initiate traversal)
2385 * @parent_css: css whose children to walk
2387 * This function returns the next child of @parent_css and should be called
2388 * under either cgroup_mutex or RCU read lock. The only requirement is
2389 * that @parent_css and @pos_css are accessible. The next sibling is
2390 * guaranteed to be returned regardless of their states.
2392 struct cgroup_subsys_state *
2393 css_next_child(struct cgroup_subsys_state *pos_css,
2394 struct cgroup_subsys_state *parent_css)
2396 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2397 struct cgroup *cgrp = parent_css->cgroup;
2398 struct cgroup *next;
2400 cgroup_assert_mutexes_or_rcu_locked();
2403 * @pos could already have been removed. Once a cgroup is removed,
2404 * its ->sibling.next is no longer updated when its next sibling
2405 * changes. As CGRP_DEAD assertion is serialized and happens
2406 * before the cgroup is taken off the ->sibling list, if we see it
2407 * unasserted, it's guaranteed that the next sibling hasn't
2408 * finished its grace period even if it's already removed, and thus
2409 * safe to dereference from this RCU critical section. If
2410 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2411 * to be visible as %true here.
2413 * If @pos is dead, its next pointer can't be dereferenced;
2414 * however, as each cgroup is given a monotonically increasing
2415 * unique serial number and always appended to the sibling list,
2416 * the next one can be found by walking the parent's children until
2417 * we see a cgroup with higher serial number than @pos's. While
2418 * this path can be slower, it's taken only when either the current
2419 * cgroup is removed or iteration and removal race.
2422 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2423 } else if (likely(!cgroup_is_dead(pos))) {
2424 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2426 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2427 if (next->serial_nr > pos->serial_nr)
2431 if (&next->sibling == &cgrp->children)
2434 return cgroup_css(next, parent_css->ss);
2438 * css_next_descendant_pre - find the next descendant for pre-order walk
2439 * @pos: the current position (%NULL to initiate traversal)
2440 * @root: css whose descendants to walk
2442 * To be used by css_for_each_descendant_pre(). Find the next descendant
2443 * to visit for pre-order traversal of @root's descendants. @root is
2444 * included in the iteration and the first node to be visited.
2446 * While this function requires cgroup_mutex or RCU read locking, it
2447 * doesn't require the whole traversal to be contained in a single critical
2448 * section. This function will return the correct next descendant as long
2449 * as both @pos and @root are accessible and @pos is a descendant of @root.
2451 struct cgroup_subsys_state *
2452 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2453 struct cgroup_subsys_state *root)
2455 struct cgroup_subsys_state *next;
2457 cgroup_assert_mutexes_or_rcu_locked();
2459 /* if first iteration, visit @root */
2463 /* visit the first child if exists */
2464 next = css_next_child(NULL, pos);
2468 /* no child, visit my or the closest ancestor's next sibling */
2469 while (pos != root) {
2470 next = css_next_child(pos, css_parent(pos));
2473 pos = css_parent(pos);
2480 * css_rightmost_descendant - return the rightmost descendant of a css
2481 * @pos: css of interest
2483 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2484 * is returned. This can be used during pre-order traversal to skip
2487 * While this function requires cgroup_mutex or RCU read locking, it
2488 * doesn't require the whole traversal to be contained in a single critical
2489 * section. This function will return the correct rightmost descendant as
2490 * long as @pos is accessible.
2492 struct cgroup_subsys_state *
2493 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2495 struct cgroup_subsys_state *last, *tmp;
2497 cgroup_assert_mutexes_or_rcu_locked();
2501 /* ->prev isn't RCU safe, walk ->next till the end */
2503 css_for_each_child(tmp, last)
2510 static struct cgroup_subsys_state *
2511 css_leftmost_descendant(struct cgroup_subsys_state *pos)
2513 struct cgroup_subsys_state *last;
2517 pos = css_next_child(NULL, pos);
2524 * css_next_descendant_post - find the next descendant for post-order walk
2525 * @pos: the current position (%NULL to initiate traversal)
2526 * @root: css whose descendants to walk
2528 * To be used by css_for_each_descendant_post(). Find the next descendant
2529 * to visit for post-order traversal of @root's descendants. @root is
2530 * included in the iteration and the last node to be visited.
2532 * While this function requires cgroup_mutex or RCU read locking, it
2533 * doesn't require the whole traversal to be contained in a single critical
2534 * section. This function will return the correct next descendant as long
2535 * as both @pos and @cgroup are accessible and @pos is a descendant of
2538 struct cgroup_subsys_state *
2539 css_next_descendant_post(struct cgroup_subsys_state *pos,
2540 struct cgroup_subsys_state *root)
2542 struct cgroup_subsys_state *next;
2544 cgroup_assert_mutexes_or_rcu_locked();
2546 /* if first iteration, visit leftmost descendant which may be @root */
2548 return css_leftmost_descendant(root);
2550 /* if we visited @root, we're done */
2554 /* if there's an unvisited sibling, visit its leftmost descendant */
2555 next = css_next_child(pos, css_parent(pos));
2557 return css_leftmost_descendant(next);
2559 /* no sibling left, visit parent */
2560 return css_parent(pos);
2564 * css_advance_task_iter - advance a task itererator to the next css_set
2565 * @it: the iterator to advance
2567 * Advance @it to the next css_set to walk.
2569 static void css_advance_task_iter(struct css_task_iter *it)
2571 struct list_head *l = it->cset_link;
2572 struct cgrp_cset_link *link;
2573 struct css_set *cset;
2575 /* Advance to the next non-empty css_set */
2578 if (l == &it->origin_css->cgroup->cset_links) {
2579 it->cset_link = NULL;
2582 link = list_entry(l, struct cgrp_cset_link, cset_link);
2584 } while (list_empty(&cset->tasks));
2586 it->task = cset->tasks.next;
2590 * css_task_iter_start - initiate task iteration
2591 * @css: the css to walk tasks of
2592 * @it: the task iterator to use
2594 * Initiate iteration through the tasks of @css. The caller can call
2595 * css_task_iter_next() to walk through the tasks until the function
2596 * returns NULL. On completion of iteration, css_task_iter_end() must be
2599 * Note that this function acquires a lock which is released when the
2600 * iteration finishes. The caller can't sleep while iteration is in
2603 void css_task_iter_start(struct cgroup_subsys_state *css,
2604 struct css_task_iter *it)
2605 __acquires(css_set_rwsem)
2607 /* no one should try to iterate before mounting cgroups */
2608 WARN_ON_ONCE(!use_task_css_set_links);
2610 down_read(&css_set_rwsem);
2612 it->origin_css = css;
2613 it->cset_link = &css->cgroup->cset_links;
2615 css_advance_task_iter(it);
2619 * css_task_iter_next - return the next task for the iterator
2620 * @it: the task iterator being iterated
2622 * The "next" function for task iteration. @it should have been
2623 * initialized via css_task_iter_start(). Returns NULL when the iteration
2626 struct task_struct *css_task_iter_next(struct css_task_iter *it)
2628 struct task_struct *res;
2629 struct list_head *l = it->task;
2630 struct cgrp_cset_link *link;
2632 /* If the iterator cg is NULL, we have no tasks */
2635 res = list_entry(l, struct task_struct, cg_list);
2636 /* Advance iterator to find next entry */
2638 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
2639 if (l == &link->cset->tasks) {
2641 * We reached the end of this task list - move on to the
2642 * next cgrp_cset_link.
2644 css_advance_task_iter(it);
2652 * css_task_iter_end - finish task iteration
2653 * @it: the task iterator to finish
2655 * Finish task iteration started by css_task_iter_start().
2657 void css_task_iter_end(struct css_task_iter *it)
2658 __releases(css_set_rwsem)
2660 up_read(&css_set_rwsem);
2664 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
2665 * @to: cgroup to which the tasks will be moved
2666 * @from: cgroup in which the tasks currently reside
2668 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
2670 struct css_task_iter it;
2671 struct task_struct *task;
2675 css_task_iter_start(&from->dummy_css, &it);
2676 task = css_task_iter_next(&it);
2678 get_task_struct(task);
2679 css_task_iter_end(&it);
2682 mutex_lock(&cgroup_mutex);
2683 ret = cgroup_attach_task(to, task, false);
2684 mutex_unlock(&cgroup_mutex);
2685 put_task_struct(task);
2687 } while (task && !ret);
2693 * Stuff for reading the 'tasks'/'procs' files.
2695 * Reading this file can return large amounts of data if a cgroup has
2696 * *lots* of attached tasks. So it may need several calls to read(),
2697 * but we cannot guarantee that the information we produce is correct
2698 * unless we produce it entirely atomically.
2702 /* which pidlist file are we talking about? */
2703 enum cgroup_filetype {
2709 * A pidlist is a list of pids that virtually represents the contents of one
2710 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
2711 * a pair (one each for procs, tasks) for each pid namespace that's relevant
2714 struct cgroup_pidlist {
2716 * used to find which pidlist is wanted. doesn't change as long as
2717 * this particular list stays in the list.
2719 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
2722 /* how many elements the above list has */
2724 /* each of these stored in a list by its cgroup */
2725 struct list_head links;
2726 /* pointer to the cgroup we belong to, for list removal purposes */
2727 struct cgroup *owner;
2728 /* for delayed destruction */
2729 struct delayed_work destroy_dwork;
2733 * The following two functions "fix" the issue where there are more pids
2734 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2735 * TODO: replace with a kernel-wide solution to this problem
2737 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2738 static void *pidlist_allocate(int count)
2740 if (PIDLIST_TOO_LARGE(count))
2741 return vmalloc(count * sizeof(pid_t));
2743 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2746 static void pidlist_free(void *p)
2748 if (is_vmalloc_addr(p))
2755 * Used to destroy all pidlists lingering waiting for destroy timer. None
2756 * should be left afterwards.
2758 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
2760 struct cgroup_pidlist *l, *tmp_l;
2762 mutex_lock(&cgrp->pidlist_mutex);
2763 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
2764 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
2765 mutex_unlock(&cgrp->pidlist_mutex);
2767 flush_workqueue(cgroup_pidlist_destroy_wq);
2768 BUG_ON(!list_empty(&cgrp->pidlists));
2771 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
2773 struct delayed_work *dwork = to_delayed_work(work);
2774 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
2776 struct cgroup_pidlist *tofree = NULL;
2778 mutex_lock(&l->owner->pidlist_mutex);
2781 * Destroy iff we didn't get queued again. The state won't change
2782 * as destroy_dwork can only be queued while locked.
2784 if (!delayed_work_pending(dwork)) {
2785 list_del(&l->links);
2786 pidlist_free(l->list);
2787 put_pid_ns(l->key.ns);
2791 mutex_unlock(&l->owner->pidlist_mutex);
2796 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2797 * Returns the number of unique elements.
2799 static int pidlist_uniq(pid_t *list, int length)
2804 * we presume the 0th element is unique, so i starts at 1. trivial
2805 * edge cases first; no work needs to be done for either
2807 if (length == 0 || length == 1)
2809 /* src and dest walk down the list; dest counts unique elements */
2810 for (src = 1; src < length; src++) {
2811 /* find next unique element */
2812 while (list[src] == list[src-1]) {
2817 /* dest always points to where the next unique element goes */
2818 list[dest] = list[src];
2826 * The two pid files - task and cgroup.procs - guaranteed that the result
2827 * is sorted, which forced this whole pidlist fiasco. As pid order is
2828 * different per namespace, each namespace needs differently sorted list,
2829 * making it impossible to use, for example, single rbtree of member tasks
2830 * sorted by task pointer. As pidlists can be fairly large, allocating one
2831 * per open file is dangerous, so cgroup had to implement shared pool of
2832 * pidlists keyed by cgroup and namespace.
2834 * All this extra complexity was caused by the original implementation
2835 * committing to an entirely unnecessary property. In the long term, we
2836 * want to do away with it. Explicitly scramble sort order if
2837 * sane_behavior so that no such expectation exists in the new interface.
2839 * Scrambling is done by swapping every two consecutive bits, which is
2840 * non-identity one-to-one mapping which disturbs sort order sufficiently.
2842 static pid_t pid_fry(pid_t pid)
2844 unsigned a = pid & 0x55555555;
2845 unsigned b = pid & 0xAAAAAAAA;
2847 return (a << 1) | (b >> 1);
2850 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
2852 if (cgroup_sane_behavior(cgrp))
2853 return pid_fry(pid);
2858 static int cmppid(const void *a, const void *b)
2860 return *(pid_t *)a - *(pid_t *)b;
2863 static int fried_cmppid(const void *a, const void *b)
2865 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
2868 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
2869 enum cgroup_filetype type)
2871 struct cgroup_pidlist *l;
2872 /* don't need task_nsproxy() if we're looking at ourself */
2873 struct pid_namespace *ns = task_active_pid_ns(current);
2875 lockdep_assert_held(&cgrp->pidlist_mutex);
2877 list_for_each_entry(l, &cgrp->pidlists, links)
2878 if (l->key.type == type && l->key.ns == ns)
2884 * find the appropriate pidlist for our purpose (given procs vs tasks)
2885 * returns with the lock on that pidlist already held, and takes care
2886 * of the use count, or returns NULL with no locks held if we're out of
2889 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
2890 enum cgroup_filetype type)
2892 struct cgroup_pidlist *l;
2894 lockdep_assert_held(&cgrp->pidlist_mutex);
2896 l = cgroup_pidlist_find(cgrp, type);
2900 /* entry not found; create a new one */
2901 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
2905 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
2907 /* don't need task_nsproxy() if we're looking at ourself */
2908 l->key.ns = get_pid_ns(task_active_pid_ns(current));
2910 list_add(&l->links, &cgrp->pidlists);
2915 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
2917 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
2918 struct cgroup_pidlist **lp)
2922 int pid, n = 0; /* used for populating the array */
2923 struct css_task_iter it;
2924 struct task_struct *tsk;
2925 struct cgroup_pidlist *l;
2927 lockdep_assert_held(&cgrp->pidlist_mutex);
2930 * If cgroup gets more users after we read count, we won't have
2931 * enough space - tough. This race is indistinguishable to the
2932 * caller from the case that the additional cgroup users didn't
2933 * show up until sometime later on.
2935 length = cgroup_task_count(cgrp);
2936 array = pidlist_allocate(length);
2939 /* now, populate the array */
2940 css_task_iter_start(&cgrp->dummy_css, &it);
2941 while ((tsk = css_task_iter_next(&it))) {
2942 if (unlikely(n == length))
2944 /* get tgid or pid for procs or tasks file respectively */
2945 if (type == CGROUP_FILE_PROCS)
2946 pid = task_tgid_vnr(tsk);
2948 pid = task_pid_vnr(tsk);
2949 if (pid > 0) /* make sure to only use valid results */
2952 css_task_iter_end(&it);
2954 /* now sort & (if procs) strip out duplicates */
2955 if (cgroup_sane_behavior(cgrp))
2956 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
2958 sort(array, length, sizeof(pid_t), cmppid, NULL);
2959 if (type == CGROUP_FILE_PROCS)
2960 length = pidlist_uniq(array, length);
2962 l = cgroup_pidlist_find_create(cgrp, type);
2964 mutex_unlock(&cgrp->pidlist_mutex);
2965 pidlist_free(array);
2969 /* store array, freeing old if necessary */
2970 pidlist_free(l->list);
2978 * cgroupstats_build - build and fill cgroupstats
2979 * @stats: cgroupstats to fill information into
2980 * @dentry: A dentry entry belonging to the cgroup for which stats have
2983 * Build and fill cgroupstats so that taskstats can export it to user
2986 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2988 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
2989 struct cgroup *cgrp;
2990 struct css_task_iter it;
2991 struct task_struct *tsk;
2993 /* it should be kernfs_node belonging to cgroupfs and is a directory */
2994 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
2995 kernfs_type(kn) != KERNFS_DIR)
2999 * We aren't being called from kernfs and there's no guarantee on
3000 * @kn->priv's validity. For this and css_tryget_from_dir(),
3001 * @kn->priv is RCU safe. Let's do the RCU dancing.
3004 cgrp = rcu_dereference(kn->priv);
3010 css_task_iter_start(&cgrp->dummy_css, &it);
3011 while ((tsk = css_task_iter_next(&it))) {
3012 switch (tsk->state) {
3014 stats->nr_running++;
3016 case TASK_INTERRUPTIBLE:
3017 stats->nr_sleeping++;
3019 case TASK_UNINTERRUPTIBLE:
3020 stats->nr_uninterruptible++;
3023 stats->nr_stopped++;
3026 if (delayacct_is_task_waiting_on_io(tsk))
3027 stats->nr_io_wait++;
3031 css_task_iter_end(&it);
3039 * seq_file methods for the tasks/procs files. The seq_file position is the
3040 * next pid to display; the seq_file iterator is a pointer to the pid
3041 * in the cgroup->l->list array.
3044 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3047 * Initially we receive a position value that corresponds to
3048 * one more than the last pid shown (or 0 on the first call or
3049 * after a seek to the start). Use a binary-search to find the
3050 * next pid to display, if any
3052 struct kernfs_open_file *of = s->private;
3053 struct cgroup *cgrp = seq_css(s)->cgroup;
3054 struct cgroup_pidlist *l;
3055 enum cgroup_filetype type = seq_cft(s)->private;
3056 int index = 0, pid = *pos;
3059 mutex_lock(&cgrp->pidlist_mutex);
3062 * !NULL @of->priv indicates that this isn't the first start()
3063 * after open. If the matching pidlist is around, we can use that.
3064 * Look for it. Note that @of->priv can't be used directly. It
3065 * could already have been destroyed.
3068 of->priv = cgroup_pidlist_find(cgrp, type);
3071 * Either this is the first start() after open or the matching
3072 * pidlist has been destroyed inbetween. Create a new one.
3075 ret = pidlist_array_load(cgrp, type,
3076 (struct cgroup_pidlist **)&of->priv);
3078 return ERR_PTR(ret);
3083 int end = l->length;
3085 while (index < end) {
3086 int mid = (index + end) / 2;
3087 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3090 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3096 /* If we're off the end of the array, we're done */
3097 if (index >= l->length)
3099 /* Update the abstract position to be the actual pid that we found */
3100 iter = l->list + index;
3101 *pos = cgroup_pid_fry(cgrp, *iter);
3105 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3107 struct kernfs_open_file *of = s->private;
3108 struct cgroup_pidlist *l = of->priv;
3111 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3112 CGROUP_PIDLIST_DESTROY_DELAY);
3113 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3116 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3118 struct kernfs_open_file *of = s->private;
3119 struct cgroup_pidlist *l = of->priv;
3121 pid_t *end = l->list + l->length;
3123 * Advance to the next pid in the array. If this goes off the
3130 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3135 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3137 return seq_printf(s, "%d\n", *(int *)v);
3141 * seq_operations functions for iterating on pidlists through seq_file -
3142 * independent of whether it's tasks or procs
3144 static const struct seq_operations cgroup_pidlist_seq_operations = {
3145 .start = cgroup_pidlist_start,
3146 .stop = cgroup_pidlist_stop,
3147 .next = cgroup_pidlist_next,
3148 .show = cgroup_pidlist_show,
3151 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3154 return notify_on_release(css->cgroup);
3157 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3158 struct cftype *cft, u64 val)
3160 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3162 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3164 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3168 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3171 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3174 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3175 struct cftype *cft, u64 val)
3178 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3180 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3184 static struct cftype cgroup_base_files[] = {
3186 .name = "cgroup.procs",
3187 .seq_start = cgroup_pidlist_start,
3188 .seq_next = cgroup_pidlist_next,
3189 .seq_stop = cgroup_pidlist_stop,
3190 .seq_show = cgroup_pidlist_show,
3191 .private = CGROUP_FILE_PROCS,
3192 .write_u64 = cgroup_procs_write,
3193 .mode = S_IRUGO | S_IWUSR,
3196 .name = "cgroup.clone_children",
3197 .flags = CFTYPE_INSANE,
3198 .read_u64 = cgroup_clone_children_read,
3199 .write_u64 = cgroup_clone_children_write,
3202 .name = "cgroup.sane_behavior",
3203 .flags = CFTYPE_ONLY_ON_ROOT,
3204 .seq_show = cgroup_sane_behavior_show,
3208 * Historical crazy stuff. These don't have "cgroup." prefix and
3209 * don't exist if sane_behavior. If you're depending on these, be
3210 * prepared to be burned.
3214 .flags = CFTYPE_INSANE, /* use "procs" instead */
3215 .seq_start = cgroup_pidlist_start,
3216 .seq_next = cgroup_pidlist_next,
3217 .seq_stop = cgroup_pidlist_stop,
3218 .seq_show = cgroup_pidlist_show,
3219 .private = CGROUP_FILE_TASKS,
3220 .write_u64 = cgroup_tasks_write,
3221 .mode = S_IRUGO | S_IWUSR,
3224 .name = "notify_on_release",
3225 .flags = CFTYPE_INSANE,
3226 .read_u64 = cgroup_read_notify_on_release,
3227 .write_u64 = cgroup_write_notify_on_release,
3230 .name = "release_agent",
3231 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3232 .seq_show = cgroup_release_agent_show,
3233 .write_string = cgroup_release_agent_write,
3234 .max_write_len = PATH_MAX - 1,
3240 * cgroup_populate_dir - create subsys files in a cgroup directory
3241 * @cgrp: target cgroup
3242 * @subsys_mask: mask of the subsystem ids whose files should be added
3244 * On failure, no file is added.
3246 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3248 struct cgroup_subsys *ss;
3251 /* process cftsets of each subsystem */
3252 for_each_subsys(ss, i) {
3253 struct cftype *cfts;
3255 if (!test_bit(i, &subsys_mask))
3258 list_for_each_entry(cfts, &ss->cfts, node) {
3259 ret = cgroup_addrm_files(cgrp, cfts, true);
3266 cgroup_clear_dir(cgrp, subsys_mask);
3271 * css destruction is four-stage process.
3273 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3274 * Implemented in kill_css().
3276 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3277 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3278 * by invoking offline_css(). After offlining, the base ref is put.
3279 * Implemented in css_killed_work_fn().
3281 * 3. When the percpu_ref reaches zero, the only possible remaining
3282 * accessors are inside RCU read sections. css_release() schedules the
3285 * 4. After the grace period, the css can be freed. Implemented in
3286 * css_free_work_fn().
3288 * It is actually hairier because both step 2 and 4 require process context
3289 * and thus involve punting to css->destroy_work adding two additional
3290 * steps to the already complex sequence.
3292 static void css_free_work_fn(struct work_struct *work)
3294 struct cgroup_subsys_state *css =
3295 container_of(work, struct cgroup_subsys_state, destroy_work);
3296 struct cgroup *cgrp = css->cgroup;
3299 css_put(css->parent);
3301 css->ss->css_free(css);
3305 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3307 struct cgroup_subsys_state *css =
3308 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3310 INIT_WORK(&css->destroy_work, css_free_work_fn);
3311 queue_work(cgroup_destroy_wq, &css->destroy_work);
3314 static void css_release(struct percpu_ref *ref)
3316 struct cgroup_subsys_state *css =
3317 container_of(ref, struct cgroup_subsys_state, refcnt);
3319 rcu_assign_pointer(css->cgroup->subsys[css->ss->id], NULL);
3320 call_rcu(&css->rcu_head, css_free_rcu_fn);
3323 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3324 struct cgroup *cgrp)
3331 css->parent = cgroup_css(cgrp->parent, ss);
3333 css->flags |= CSS_ROOT;
3335 BUG_ON(cgroup_css(cgrp, ss));
3338 /* invoke ->css_online() on a new CSS and mark it online if successful */
3339 static int online_css(struct cgroup_subsys_state *css)
3341 struct cgroup_subsys *ss = css->ss;
3344 lockdep_assert_held(&cgroup_tree_mutex);
3345 lockdep_assert_held(&cgroup_mutex);
3348 ret = ss->css_online(css);
3350 css->flags |= CSS_ONLINE;
3351 css->cgroup->nr_css++;
3352 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
3357 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
3358 static void offline_css(struct cgroup_subsys_state *css)
3360 struct cgroup_subsys *ss = css->ss;
3362 lockdep_assert_held(&cgroup_tree_mutex);
3363 lockdep_assert_held(&cgroup_mutex);
3365 if (!(css->flags & CSS_ONLINE))
3368 if (ss->css_offline)
3369 ss->css_offline(css);
3371 css->flags &= ~CSS_ONLINE;
3372 css->cgroup->nr_css--;
3373 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
3377 * create_css - create a cgroup_subsys_state
3378 * @cgrp: the cgroup new css will be associated with
3379 * @ss: the subsys of new css
3381 * Create a new css associated with @cgrp - @ss pair. On success, the new
3382 * css is online and installed in @cgrp with all interface files created.
3383 * Returns 0 on success, -errno on failure.
3385 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
3387 struct cgroup *parent = cgrp->parent;
3388 struct cgroup_subsys_state *css;
3391 lockdep_assert_held(&cgroup_mutex);
3393 css = ss->css_alloc(cgroup_css(parent, ss));
3395 return PTR_ERR(css);
3397 err = percpu_ref_init(&css->refcnt, css_release);
3401 init_css(css, ss, cgrp);
3403 err = cgroup_populate_dir(cgrp, 1 << ss->id);
3407 err = online_css(css);
3412 css_get(css->parent);
3414 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
3416 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",
3417 current->comm, current->pid, ss->name);
3418 if (!strcmp(ss->name, "memory"))
3419 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
3420 ss->warned_broken_hierarchy = true;
3426 percpu_ref_cancel_init(&css->refcnt);
3432 * cgroup_create - create a cgroup
3433 * @parent: cgroup that will be parent of the new cgroup
3434 * @name: name of the new cgroup
3435 * @mode: mode to set on new cgroup
3437 static long cgroup_create(struct cgroup *parent, const char *name,
3440 struct cgroup *cgrp;
3441 struct cgroupfs_root *root = parent->root;
3443 struct cgroup_subsys *ss;
3444 struct kernfs_node *kn;
3446 /* allocate the cgroup and its ID, 0 is reserved for the root */
3447 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3451 mutex_lock(&cgroup_tree_mutex);
3454 * Only live parents can have children. Note that the liveliness
3455 * check isn't strictly necessary because cgroup_mkdir() and
3456 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
3457 * anyway so that locking is contained inside cgroup proper and we
3458 * don't get nasty surprises if we ever grow another caller.
3460 if (!cgroup_lock_live_group(parent)) {
3462 goto err_unlock_tree;
3466 * Temporarily set the pointer to NULL, so idr_find() won't return
3467 * a half-baked cgroup.
3469 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
3475 init_cgroup_housekeeping(cgrp);
3477 cgrp->parent = parent;
3478 cgrp->dummy_css.parent = &parent->dummy_css;
3479 cgrp->root = parent->root;
3481 if (notify_on_release(parent))
3482 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3484 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
3485 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
3487 /* create the directory */
3488 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
3496 * This extra ref will be put in cgroup_free_fn() and guarantees
3497 * that @cgrp->kn is always accessible.
3501 cgrp->serial_nr = cgroup_serial_nr_next++;
3503 /* allocation complete, commit to creation */
3504 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
3505 atomic_inc(&root->nr_cgrps);
3509 * @cgrp is now fully operational. If something fails after this
3510 * point, it'll be released via the normal destruction path.
3512 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
3514 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
3518 /* let's create and online css's */
3519 for_each_subsys(ss, ssid) {
3520 if (root->subsys_mask & (1 << ssid)) {
3521 err = create_css(cgrp, ss);
3527 kernfs_activate(kn);
3529 mutex_unlock(&cgroup_mutex);
3530 mutex_unlock(&cgroup_tree_mutex);
3535 idr_remove(&root->cgroup_idr, cgrp->id);
3537 mutex_unlock(&cgroup_mutex);
3539 mutex_unlock(&cgroup_tree_mutex);
3544 cgroup_destroy_locked(cgrp);
3545 mutex_unlock(&cgroup_mutex);
3546 mutex_unlock(&cgroup_tree_mutex);
3550 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3553 struct cgroup *parent = parent_kn->priv;
3555 return cgroup_create(parent, name, mode);
3559 * This is called when the refcnt of a css is confirmed to be killed.
3560 * css_tryget() is now guaranteed to fail.
3562 static void css_killed_work_fn(struct work_struct *work)
3564 struct cgroup_subsys_state *css =
3565 container_of(work, struct cgroup_subsys_state, destroy_work);
3566 struct cgroup *cgrp = css->cgroup;
3568 mutex_lock(&cgroup_tree_mutex);
3569 mutex_lock(&cgroup_mutex);
3572 * css_tryget() is guaranteed to fail now. Tell subsystems to
3573 * initate destruction.
3578 * If @cgrp is marked dead, it's waiting for refs of all css's to
3579 * be disabled before proceeding to the second phase of cgroup
3580 * destruction. If we are the last one, kick it off.
3582 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
3583 cgroup_destroy_css_killed(cgrp);
3585 mutex_unlock(&cgroup_mutex);
3586 mutex_unlock(&cgroup_tree_mutex);
3589 * Put the css refs from kill_css(). Each css holds an extra
3590 * reference to the cgroup's dentry and cgroup removal proceeds
3591 * regardless of css refs. On the last put of each css, whenever
3592 * that may be, the extra dentry ref is put so that dentry
3593 * destruction happens only after all css's are released.
3598 /* css kill confirmation processing requires process context, bounce */
3599 static void css_killed_ref_fn(struct percpu_ref *ref)
3601 struct cgroup_subsys_state *css =
3602 container_of(ref, struct cgroup_subsys_state, refcnt);
3604 INIT_WORK(&css->destroy_work, css_killed_work_fn);
3605 queue_work(cgroup_destroy_wq, &css->destroy_work);
3609 * kill_css - destroy a css
3610 * @css: css to destroy
3612 * This function initiates destruction of @css by removing cgroup interface
3613 * files and putting its base reference. ->css_offline() will be invoked
3614 * asynchronously once css_tryget() is guaranteed to fail and when the
3615 * reference count reaches zero, @css will be released.
3617 static void kill_css(struct cgroup_subsys_state *css)
3620 * This must happen before css is disassociated with its cgroup.
3621 * See seq_css() for details.
3623 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3626 * Killing would put the base ref, but we need to keep it alive
3627 * until after ->css_offline().
3632 * cgroup core guarantees that, by the time ->css_offline() is
3633 * invoked, no new css reference will be given out via
3634 * css_tryget(). We can't simply call percpu_ref_kill() and
3635 * proceed to offlining css's because percpu_ref_kill() doesn't
3636 * guarantee that the ref is seen as killed on all CPUs on return.
3638 * Use percpu_ref_kill_and_confirm() to get notifications as each
3639 * css is confirmed to be seen as killed on all CPUs.
3641 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
3645 * cgroup_destroy_locked - the first stage of cgroup destruction
3646 * @cgrp: cgroup to be destroyed
3648 * css's make use of percpu refcnts whose killing latency shouldn't be
3649 * exposed to userland and are RCU protected. Also, cgroup core needs to
3650 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
3651 * invoked. To satisfy all the requirements, destruction is implemented in
3652 * the following two steps.
3654 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
3655 * userland visible parts and start killing the percpu refcnts of
3656 * css's. Set up so that the next stage will be kicked off once all
3657 * the percpu refcnts are confirmed to be killed.
3659 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
3660 * rest of destruction. Once all cgroup references are gone, the
3661 * cgroup is RCU-freed.
3663 * This function implements s1. After this step, @cgrp is gone as far as
3664 * the userland is concerned and a new cgroup with the same name may be
3665 * created. As cgroup doesn't care about the names internally, this
3666 * doesn't cause any problem.
3668 static int cgroup_destroy_locked(struct cgroup *cgrp)
3669 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
3671 struct cgroup *child;
3672 struct cgroup_subsys_state *css;
3676 lockdep_assert_held(&cgroup_tree_mutex);
3677 lockdep_assert_held(&cgroup_mutex);
3680 * css_set_rwsem synchronizes access to ->cset_links and prevents
3681 * @cgrp from being removed while put_css_set() is in progress.
3683 down_read(&css_set_rwsem);
3684 empty = list_empty(&cgrp->cset_links);
3685 up_read(&css_set_rwsem);
3690 * Make sure there's no live children. We can't test ->children
3691 * emptiness as dead children linger on it while being destroyed;
3692 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
3696 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
3697 empty = cgroup_is_dead(child);
3706 * Initiate massacre of all css's. cgroup_destroy_css_killed()
3707 * will be invoked to perform the rest of destruction once the
3708 * percpu refs of all css's are confirmed to be killed. This
3709 * involves removing the subsystem's files, drop cgroup_mutex.
3711 mutex_unlock(&cgroup_mutex);
3712 for_each_css(css, ssid, cgrp)
3714 mutex_lock(&cgroup_mutex);
3717 * Mark @cgrp dead. This prevents further task migration and child
3718 * creation by disabling cgroup_lock_live_group(). Note that
3719 * CGRP_DEAD assertion is depended upon by css_next_child() to
3720 * resume iteration after dropping RCU read lock. See
3721 * css_next_child() for details.
3723 set_bit(CGRP_DEAD, &cgrp->flags);
3725 /* CGRP_DEAD is set, remove from ->release_list for the last time */
3726 raw_spin_lock(&release_list_lock);
3727 if (!list_empty(&cgrp->release_list))
3728 list_del_init(&cgrp->release_list);
3729 raw_spin_unlock(&release_list_lock);
3732 * If @cgrp has css's attached, the second stage of cgroup
3733 * destruction is kicked off from css_killed_work_fn() after the
3734 * refs of all attached css's are killed. If @cgrp doesn't have
3735 * any css, we kick it off here.
3738 cgroup_destroy_css_killed(cgrp);
3740 /* remove @cgrp directory along with the base files */
3741 mutex_unlock(&cgroup_mutex);
3744 * There are two control paths which try to determine cgroup from
3745 * dentry without going through kernfs - cgroupstats_build() and
3746 * css_tryget_from_dir(). Those are supported by RCU protecting
3747 * clearing of cgrp->kn->priv backpointer, which should happen
3748 * after all files under it have been removed.
3750 kernfs_remove(cgrp->kn); /* @cgrp has an extra ref on its kn */
3751 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
3753 mutex_lock(&cgroup_mutex);
3759 * cgroup_destroy_css_killed - the second step of cgroup destruction
3760 * @work: cgroup->destroy_free_work
3762 * This function is invoked from a work item for a cgroup which is being
3763 * destroyed after all css's are offlined and performs the rest of
3764 * destruction. This is the second step of destruction described in the
3765 * comment above cgroup_destroy_locked().
3767 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
3769 struct cgroup *parent = cgrp->parent;
3771 lockdep_assert_held(&cgroup_tree_mutex);
3772 lockdep_assert_held(&cgroup_mutex);
3774 /* delete this cgroup from parent->children */
3775 list_del_rcu(&cgrp->sibling);
3779 set_bit(CGRP_RELEASABLE, &parent->flags);
3780 check_for_release(parent);
3783 static int cgroup_rmdir(struct kernfs_node *kn)
3785 struct cgroup *cgrp = kn->priv;
3789 * This is self-destruction but @kn can't be removed while this
3790 * callback is in progress. Let's break active protection. Once
3791 * the protection is broken, @cgrp can be destroyed at any point.
3792 * Pin it so that it stays accessible.
3795 kernfs_break_active_protection(kn);
3797 mutex_lock(&cgroup_tree_mutex);
3798 mutex_lock(&cgroup_mutex);
3801 * @cgrp might already have been destroyed while we're trying to
3804 if (!cgroup_is_dead(cgrp))
3805 ret = cgroup_destroy_locked(cgrp);
3807 mutex_unlock(&cgroup_mutex);
3808 mutex_unlock(&cgroup_tree_mutex);
3810 kernfs_unbreak_active_protection(kn);
3815 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
3816 .remount_fs = cgroup_remount,
3817 .show_options = cgroup_show_options,
3818 .mkdir = cgroup_mkdir,
3819 .rmdir = cgroup_rmdir,
3820 .rename = cgroup_rename,
3823 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
3825 struct cgroup_subsys_state *css;
3827 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
3829 mutex_lock(&cgroup_tree_mutex);
3830 mutex_lock(&cgroup_mutex);
3832 INIT_LIST_HEAD(&ss->cfts);
3834 /* Create the top cgroup state for this subsystem */
3835 ss->root = &cgroup_dummy_root;
3836 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
3837 /* We don't handle early failures gracefully */
3838 BUG_ON(IS_ERR(css));
3839 init_css(css, ss, cgroup_dummy_top);
3841 /* Update the init_css_set to contain a subsys
3842 * pointer to this state - since the subsystem is
3843 * newly registered, all tasks and hence the
3844 * init_css_set is in the subsystem's top cgroup. */
3845 init_css_set.subsys[ss->id] = css;
3847 need_forkexit_callback |= ss->fork || ss->exit;
3849 /* At system boot, before all subsystems have been
3850 * registered, no tasks have been forked, so we don't
3851 * need to invoke fork callbacks here. */
3852 BUG_ON(!list_empty(&init_task.tasks));
3854 BUG_ON(online_css(css));
3856 mutex_unlock(&cgroup_mutex);
3857 mutex_unlock(&cgroup_tree_mutex);
3861 * cgroup_init_early - cgroup initialization at system boot
3863 * Initialize cgroups at system boot, and initialize any
3864 * subsystems that request early init.
3866 int __init cgroup_init_early(void)
3868 struct cgroup_subsys *ss;
3871 atomic_set(&init_css_set.refcount, 1);
3872 INIT_LIST_HEAD(&init_css_set.cgrp_links);
3873 INIT_LIST_HEAD(&init_css_set.tasks);
3874 INIT_HLIST_NODE(&init_css_set.hlist);
3876 init_cgroup_root(&cgroup_dummy_root);
3877 cgroup_root_count = 1;
3878 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
3880 init_cgrp_cset_link.cset = &init_css_set;
3881 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
3882 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
3883 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
3885 for_each_subsys(ss, i) {
3886 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
3887 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
3888 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
3890 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
3891 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
3894 ss->name = cgroup_subsys_name[i];
3897 cgroup_init_subsys(ss);
3903 * cgroup_init - cgroup initialization
3905 * Register cgroup filesystem and /proc file, and initialize
3906 * any subsystems that didn't request early init.
3908 int __init cgroup_init(void)
3910 struct cgroup_subsys *ss;
3914 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
3916 for_each_subsys(ss, i) {
3917 if (!ss->early_init)
3918 cgroup_init_subsys(ss);
3921 * cftype registration needs kmalloc and can't be done
3922 * during early_init. Register base cftypes separately.
3924 if (ss->base_cftypes)
3925 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
3928 /* allocate id for the dummy hierarchy */
3929 mutex_lock(&cgroup_mutex);
3931 /* Add init_css_set to the hash table */
3932 key = css_set_hash(init_css_set.subsys);
3933 hash_add(css_set_table, &init_css_set.hlist, key);
3935 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
3937 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
3941 mutex_unlock(&cgroup_mutex);
3943 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
3947 err = register_filesystem(&cgroup_fs_type);
3949 kobject_put(cgroup_kobj);
3953 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
3957 static int __init cgroup_wq_init(void)
3960 * There isn't much point in executing destruction path in
3961 * parallel. Good chunk is serialized with cgroup_mutex anyway.
3963 * XXX: Must be ordered to make sure parent is offlined after
3964 * children. The ordering requirement is for memcg where a
3965 * parent's offline may wait for a child's leading to deadlock. In
3966 * the long term, this should be fixed from memcg side.
3968 * We would prefer to do this in cgroup_init() above, but that
3969 * is called before init_workqueues(): so leave this until after.
3971 cgroup_destroy_wq = alloc_ordered_workqueue("cgroup_destroy", 0);
3972 BUG_ON(!cgroup_destroy_wq);
3975 * Used to destroy pidlists and separate to serve as flush domain.
3976 * Cap @max_active to 1 too.
3978 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
3980 BUG_ON(!cgroup_pidlist_destroy_wq);
3984 core_initcall(cgroup_wq_init);
3987 * proc_cgroup_show()
3988 * - Print task's cgroup paths into seq_file, one line for each hierarchy
3989 * - Used for /proc/<pid>/cgroup.
3990 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
3991 * doesn't really matter if tsk->cgroup changes after we read it,
3992 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
3993 * anyway. No need to check that tsk->cgroup != NULL, thanks to
3994 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
3995 * cgroup to top_cgroup.
3998 /* TODO: Use a proper seq_file iterator */
3999 int proc_cgroup_show(struct seq_file *m, void *v)
4002 struct task_struct *tsk;
4005 struct cgroupfs_root *root;
4008 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4014 tsk = get_pid_task(pid, PIDTYPE_PID);
4020 mutex_lock(&cgroup_mutex);
4021 down_read(&css_set_rwsem);
4023 for_each_active_root(root) {
4024 struct cgroup_subsys *ss;
4025 struct cgroup *cgrp;
4026 int ssid, count = 0;
4028 seq_printf(m, "%d:", root->hierarchy_id);
4029 for_each_subsys(ss, ssid)
4030 if (root->subsys_mask & (1 << ssid))
4031 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4032 if (strlen(root->name))
4033 seq_printf(m, "%sname=%s", count ? "," : "",
4036 cgrp = task_cgroup_from_root(tsk, root);
4037 path = cgroup_path(cgrp, buf, PATH_MAX);
4039 retval = -ENAMETOOLONG;
4047 up_read(&css_set_rwsem);
4048 mutex_unlock(&cgroup_mutex);
4049 put_task_struct(tsk);
4056 /* Display information about each subsystem and each hierarchy */
4057 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4059 struct cgroup_subsys *ss;
4062 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4064 * ideally we don't want subsystems moving around while we do this.
4065 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4066 * subsys/hierarchy state.
4068 mutex_lock(&cgroup_mutex);
4070 for_each_subsys(ss, i)
4071 seq_printf(m, "%s\t%d\t%d\t%d\n",
4072 ss->name, ss->root->hierarchy_id,
4073 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4075 mutex_unlock(&cgroup_mutex);
4079 static int cgroupstats_open(struct inode *inode, struct file *file)
4081 return single_open(file, proc_cgroupstats_show, NULL);
4084 static const struct file_operations proc_cgroupstats_operations = {
4085 .open = cgroupstats_open,
4087 .llseek = seq_lseek,
4088 .release = single_release,
4092 * cgroup_fork - attach newly forked task to its parents cgroup.
4093 * @child: pointer to task_struct of forking parent process.
4095 * Description: A task inherits its parent's cgroup at fork().
4097 * A pointer to the shared css_set was automatically copied in
4098 * fork.c by dup_task_struct(). However, we ignore that copy, since
4099 * it was not made under the protection of RCU or cgroup_mutex, so
4100 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4101 * have already changed current->cgroups, allowing the previously
4102 * referenced cgroup group to be removed and freed.
4104 * At the point that cgroup_fork() is called, 'current' is the parent
4105 * task, and the passed argument 'child' points to the child task.
4107 void cgroup_fork(struct task_struct *child)
4110 get_css_set(task_css_set(current));
4111 child->cgroups = current->cgroups;
4112 task_unlock(current);
4113 INIT_LIST_HEAD(&child->cg_list);
4117 * cgroup_post_fork - called on a new task after adding it to the task list
4118 * @child: the task in question
4120 * Adds the task to the list running through its css_set if necessary and
4121 * call the subsystem fork() callbacks. Has to be after the task is
4122 * visible on the task list in case we race with the first call to
4123 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4126 void cgroup_post_fork(struct task_struct *child)
4128 struct cgroup_subsys *ss;
4132 * use_task_css_set_links is set to 1 before we walk the tasklist
4133 * under the tasklist_lock and we read it here after we added the child
4134 * to the tasklist under the tasklist_lock as well. If the child wasn't
4135 * yet in the tasklist when we walked through it from
4136 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4137 * should be visible now due to the paired locking and barriers implied
4138 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4139 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4142 if (use_task_css_set_links) {
4143 down_write(&css_set_rwsem);
4145 if (list_empty(&child->cg_list))
4146 list_add(&child->cg_list, &task_css_set(child)->tasks);
4148 up_write(&css_set_rwsem);
4152 * Call ss->fork(). This must happen after @child is linked on
4153 * css_set; otherwise, @child might change state between ->fork()
4154 * and addition to css_set.
4156 if (need_forkexit_callback) {
4157 for_each_subsys(ss, i)
4164 * cgroup_exit - detach cgroup from exiting task
4165 * @tsk: pointer to task_struct of exiting process
4166 * @run_callback: run exit callbacks?
4168 * Description: Detach cgroup from @tsk and release it.
4170 * Note that cgroups marked notify_on_release force every task in
4171 * them to take the global cgroup_mutex mutex when exiting.
4172 * This could impact scaling on very large systems. Be reluctant to
4173 * use notify_on_release cgroups where very high task exit scaling
4174 * is required on large systems.
4176 * the_top_cgroup_hack:
4178 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4180 * We call cgroup_exit() while the task is still competent to
4181 * handle notify_on_release(), then leave the task attached to the
4182 * root cgroup in each hierarchy for the remainder of its exit.
4184 * To do this properly, we would increment the reference count on
4185 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4186 * code we would add a second cgroup function call, to drop that
4187 * reference. This would just create an unnecessary hot spot on
4188 * the top_cgroup reference count, to no avail.
4190 * Normally, holding a reference to a cgroup without bumping its
4191 * count is unsafe. The cgroup could go away, or someone could
4192 * attach us to a different cgroup, decrementing the count on
4193 * the first cgroup that we never incremented. But in this case,
4194 * top_cgroup isn't going away, and either task has PF_EXITING set,
4195 * which wards off any cgroup_attach_task() attempts, or task is a failed
4196 * fork, never visible to cgroup_attach_task.
4198 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4200 struct cgroup_subsys *ss;
4201 struct css_set *cset;
4205 * Unlink from the css_set task list if necessary. Optimistically
4206 * check cg_list before taking css_set_rwsem.
4208 if (!list_empty(&tsk->cg_list)) {
4209 down_write(&css_set_rwsem);
4210 if (!list_empty(&tsk->cg_list))
4211 list_del_init(&tsk->cg_list);
4212 up_write(&css_set_rwsem);
4215 /* Reassign the task to the init_css_set. */
4217 cset = task_css_set(tsk);
4218 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4220 if (run_callbacks && need_forkexit_callback) {
4221 /* see cgroup_post_fork() for details */
4222 for_each_subsys(ss, i) {
4224 struct cgroup_subsys_state *old_css = cset->subsys[i];
4225 struct cgroup_subsys_state *css = task_css(tsk, i);
4227 ss->exit(css, old_css, tsk);
4233 put_css_set(cset, true);
4236 static void check_for_release(struct cgroup *cgrp)
4238 if (cgroup_is_releasable(cgrp) &&
4239 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4241 * Control Group is currently removeable. If it's not
4242 * already queued for a userspace notification, queue
4245 int need_schedule_work = 0;
4247 raw_spin_lock(&release_list_lock);
4248 if (!cgroup_is_dead(cgrp) &&
4249 list_empty(&cgrp->release_list)) {
4250 list_add(&cgrp->release_list, &release_list);
4251 need_schedule_work = 1;
4253 raw_spin_unlock(&release_list_lock);
4254 if (need_schedule_work)
4255 schedule_work(&release_agent_work);
4260 * Notify userspace when a cgroup is released, by running the
4261 * configured release agent with the name of the cgroup (path
4262 * relative to the root of cgroup file system) as the argument.
4264 * Most likely, this user command will try to rmdir this cgroup.
4266 * This races with the possibility that some other task will be
4267 * attached to this cgroup before it is removed, or that some other
4268 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4269 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4270 * unused, and this cgroup will be reprieved from its death sentence,
4271 * to continue to serve a useful existence. Next time it's released,
4272 * we will get notified again, if it still has 'notify_on_release' set.
4274 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4275 * means only wait until the task is successfully execve()'d. The
4276 * separate release agent task is forked by call_usermodehelper(),
4277 * then control in this thread returns here, without waiting for the
4278 * release agent task. We don't bother to wait because the caller of
4279 * this routine has no use for the exit status of the release agent
4280 * task, so no sense holding our caller up for that.
4282 static void cgroup_release_agent(struct work_struct *work)
4284 BUG_ON(work != &release_agent_work);
4285 mutex_lock(&cgroup_mutex);
4286 raw_spin_lock(&release_list_lock);
4287 while (!list_empty(&release_list)) {
4288 char *argv[3], *envp[3];
4290 char *pathbuf = NULL, *agentbuf = NULL, *path;
4291 struct cgroup *cgrp = list_entry(release_list.next,
4294 list_del_init(&cgrp->release_list);
4295 raw_spin_unlock(&release_list_lock);
4296 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
4299 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
4302 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4307 argv[i++] = agentbuf;
4312 /* minimal command environment */
4313 envp[i++] = "HOME=/";
4314 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4317 /* Drop the lock while we invoke the usermode helper,
4318 * since the exec could involve hitting disk and hence
4319 * be a slow process */
4320 mutex_unlock(&cgroup_mutex);
4321 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4322 mutex_lock(&cgroup_mutex);
4326 raw_spin_lock(&release_list_lock);
4328 raw_spin_unlock(&release_list_lock);
4329 mutex_unlock(&cgroup_mutex);
4332 static int __init cgroup_disable(char *str)
4334 struct cgroup_subsys *ss;
4338 while ((token = strsep(&str, ",")) != NULL) {
4342 for_each_subsys(ss, i) {
4343 if (!strcmp(token, ss->name)) {
4345 printk(KERN_INFO "Disabling %s control group"
4346 " subsystem\n", ss->name);
4353 __setup("cgroup_disable=", cgroup_disable);
4356 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
4357 * @dentry: directory dentry of interest
4358 * @ss: subsystem of interest
4360 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
4361 * to get the corresponding css and return it. If such css doesn't exist
4362 * or can't be pinned, an ERR_PTR value is returned.
4364 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
4365 struct cgroup_subsys *ss)
4367 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4368 struct cgroup_subsys_state *css = NULL;
4369 struct cgroup *cgrp;
4371 /* is @dentry a cgroup dir? */
4372 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4373 kernfs_type(kn) != KERNFS_DIR)
4374 return ERR_PTR(-EBADF);
4379 * This path doesn't originate from kernfs and @kn could already
4380 * have been or be removed at any point. @kn->priv is RCU
4381 * protected for this access. See destroy_locked() for details.
4383 cgrp = rcu_dereference(kn->priv);
4385 css = cgroup_css(cgrp, ss);
4387 if (!css || !css_tryget(css))
4388 css = ERR_PTR(-ENOENT);
4395 * css_from_id - lookup css by id
4396 * @id: the cgroup id
4397 * @ss: cgroup subsys to be looked into
4399 * Returns the css if there's valid one with @id, otherwise returns NULL.
4400 * Should be called under rcu_read_lock().
4402 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
4404 struct cgroup *cgrp;
4406 cgroup_assert_mutexes_or_rcu_locked();
4408 cgrp = idr_find(&ss->root->cgroup_idr, id);
4410 return cgroup_css(cgrp, ss);
4414 #ifdef CONFIG_CGROUP_DEBUG
4415 static struct cgroup_subsys_state *
4416 debug_css_alloc(struct cgroup_subsys_state *parent_css)
4418 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4421 return ERR_PTR(-ENOMEM);
4426 static void debug_css_free(struct cgroup_subsys_state *css)
4431 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
4434 return cgroup_task_count(css->cgroup);
4437 static u64 current_css_set_read(struct cgroup_subsys_state *css,
4440 return (u64)(unsigned long)current->cgroups;
4443 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
4449 count = atomic_read(&task_css_set(current)->refcount);
4454 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
4456 struct cgrp_cset_link *link;
4457 struct css_set *cset;
4460 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
4464 down_read(&css_set_rwsem);
4466 cset = rcu_dereference(current->cgroups);
4467 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
4468 struct cgroup *c = link->cgrp;
4469 const char *name = "?";
4471 if (c != cgroup_dummy_top) {
4472 cgroup_name(c, name_buf, NAME_MAX + 1);
4476 seq_printf(seq, "Root %d group %s\n",
4477 c->root->hierarchy_id, name);
4480 up_read(&css_set_rwsem);
4485 #define MAX_TASKS_SHOWN_PER_CSS 25
4486 static int cgroup_css_links_read(struct seq_file *seq, void *v)
4488 struct cgroup_subsys_state *css = seq_css(seq);
4489 struct cgrp_cset_link *link;
4491 down_read(&css_set_rwsem);
4492 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
4493 struct css_set *cset = link->cset;
4494 struct task_struct *task;
4496 seq_printf(seq, "css_set %p\n", cset);
4497 list_for_each_entry(task, &cset->tasks, cg_list) {
4498 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4499 seq_puts(seq, " ...\n");
4502 seq_printf(seq, " task %d\n",
4503 task_pid_vnr(task));
4507 up_read(&css_set_rwsem);
4511 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
4513 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4516 static struct cftype debug_files[] = {
4518 .name = "taskcount",
4519 .read_u64 = debug_taskcount_read,
4523 .name = "current_css_set",
4524 .read_u64 = current_css_set_read,
4528 .name = "current_css_set_refcount",
4529 .read_u64 = current_css_set_refcount_read,
4533 .name = "current_css_set_cg_links",
4534 .seq_show = current_css_set_cg_links_read,
4538 .name = "cgroup_css_links",
4539 .seq_show = cgroup_css_links_read,
4543 .name = "releasable",
4544 .read_u64 = releasable_read,
4550 struct cgroup_subsys debug_cgrp_subsys = {
4551 .css_alloc = debug_css_alloc,
4552 .css_free = debug_css_free,
4553 .base_cftypes = debug_files,
4555 #endif /* CONFIG_CGROUP_DEBUG */