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)
245 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
247 static int cgroup_is_releasable(const struct cgroup *cgrp)
250 (1 << CGRP_RELEASABLE) |
251 (1 << CGRP_NOTIFY_ON_RELEASE);
252 return (cgrp->flags & bits) == bits;
255 static int notify_on_release(const struct cgroup *cgrp)
257 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
261 * for_each_css - iterate all css's of a cgroup
262 * @css: the iteration cursor
263 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
264 * @cgrp: the target cgroup to iterate css's of
266 * Should be called under cgroup_mutex.
268 #define for_each_css(css, ssid, cgrp) \
269 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
270 if (!((css) = rcu_dereference_check( \
271 (cgrp)->subsys[(ssid)], \
272 lockdep_is_held(&cgroup_tree_mutex) || \
273 lockdep_is_held(&cgroup_mutex)))) { } \
277 * for_each_subsys - iterate all enabled cgroup subsystems
278 * @ss: the iteration cursor
279 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
281 #define for_each_subsys(ss, ssid) \
282 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
283 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
285 /* iterate across the active hierarchies */
286 #define for_each_active_root(root) \
287 list_for_each_entry((root), &cgroup_roots, root_list)
290 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
291 * @cgrp: the cgroup to be checked for liveness
293 * On success, returns true; the mutex should be later unlocked. On
294 * failure returns false with no lock held.
296 static bool cgroup_lock_live_group(struct cgroup *cgrp)
298 mutex_lock(&cgroup_mutex);
299 if (cgroup_is_dead(cgrp)) {
300 mutex_unlock(&cgroup_mutex);
306 /* the list of cgroups eligible for automatic release. Protected by
307 * release_list_lock */
308 static LIST_HEAD(release_list);
309 static DEFINE_RAW_SPINLOCK(release_list_lock);
310 static void cgroup_release_agent(struct work_struct *work);
311 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
312 static void check_for_release(struct cgroup *cgrp);
315 * A cgroup can be associated with multiple css_sets as different tasks may
316 * belong to different cgroups on different hierarchies. In the other
317 * direction, a css_set is naturally associated with multiple cgroups.
318 * This M:N relationship is represented by the following link structure
319 * which exists for each association and allows traversing the associations
322 struct cgrp_cset_link {
323 /* the cgroup and css_set this link associates */
325 struct css_set *cset;
327 /* list of cgrp_cset_links anchored at cgrp->cset_links */
328 struct list_head cset_link;
330 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
331 struct list_head cgrp_link;
334 /* The default css_set - used by init and its children prior to any
335 * hierarchies being mounted. It contains a pointer to the root state
336 * for each subsystem. Also used to anchor the list of css_sets. Not
337 * reference-counted, to improve performance when child cgroups
338 * haven't been created.
341 static struct css_set init_css_set;
342 static struct cgrp_cset_link init_cgrp_cset_link;
345 * css_set_rwsem protects the list of css_set objects, and the chain of
346 * tasks off each css_set.
348 static DECLARE_RWSEM(css_set_rwsem);
349 static int css_set_count;
352 * hash table for cgroup groups. This improves the performance to find
353 * an existing css_set. This hash doesn't (currently) take into
354 * account cgroups in empty hierarchies.
356 #define CSS_SET_HASH_BITS 7
357 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
359 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
361 unsigned long key = 0UL;
362 struct cgroup_subsys *ss;
365 for_each_subsys(ss, i)
366 key += (unsigned long)css[i];
367 key = (key >> 16) ^ key;
372 static void __put_css_set(struct css_set *cset, int taskexit)
374 struct cgrp_cset_link *link, *tmp_link;
377 * Ensure that the refcount doesn't hit zero while any readers
378 * can see it. Similar to atomic_dec_and_lock(), but for an
381 if (atomic_add_unless(&cset->refcount, -1, 1))
383 down_write(&css_set_rwsem);
384 if (!atomic_dec_and_test(&cset->refcount)) {
385 up_write(&css_set_rwsem);
389 /* This css_set is dead. unlink it and release cgroup refcounts */
390 hash_del(&cset->hlist);
393 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
394 struct cgroup *cgrp = link->cgrp;
396 list_del(&link->cset_link);
397 list_del(&link->cgrp_link);
399 /* @cgrp can't go away while we're holding css_set_rwsem */
400 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
402 set_bit(CGRP_RELEASABLE, &cgrp->flags);
403 check_for_release(cgrp);
409 up_write(&css_set_rwsem);
410 kfree_rcu(cset, rcu_head);
414 * refcounted get/put for css_set objects
416 static inline void get_css_set(struct css_set *cset)
418 atomic_inc(&cset->refcount);
421 static inline void put_css_set(struct css_set *cset)
423 __put_css_set(cset, 0);
426 static inline void put_css_set_taskexit(struct css_set *cset)
428 __put_css_set(cset, 1);
432 * compare_css_sets - helper function for find_existing_css_set().
433 * @cset: candidate css_set being tested
434 * @old_cset: existing css_set for a task
435 * @new_cgrp: cgroup that's being entered by the task
436 * @template: desired set of css pointers in css_set (pre-calculated)
438 * Returns true if "cset" matches "old_cset" except for the hierarchy
439 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
441 static bool compare_css_sets(struct css_set *cset,
442 struct css_set *old_cset,
443 struct cgroup *new_cgrp,
444 struct cgroup_subsys_state *template[])
446 struct list_head *l1, *l2;
448 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
449 /* Not all subsystems matched */
454 * Compare cgroup pointers in order to distinguish between
455 * different cgroups in heirarchies with no subsystems. We
456 * could get by with just this check alone (and skip the
457 * memcmp above) but on most setups the memcmp check will
458 * avoid the need for this more expensive check on almost all
462 l1 = &cset->cgrp_links;
463 l2 = &old_cset->cgrp_links;
465 struct cgrp_cset_link *link1, *link2;
466 struct cgroup *cgrp1, *cgrp2;
470 /* See if we reached the end - both lists are equal length. */
471 if (l1 == &cset->cgrp_links) {
472 BUG_ON(l2 != &old_cset->cgrp_links);
475 BUG_ON(l2 == &old_cset->cgrp_links);
477 /* Locate the cgroups associated with these links. */
478 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
479 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
482 /* Hierarchies should be linked in the same order. */
483 BUG_ON(cgrp1->root != cgrp2->root);
486 * If this hierarchy is the hierarchy of the cgroup
487 * that's changing, then we need to check that this
488 * css_set points to the new cgroup; if it's any other
489 * hierarchy, then this css_set should point to the
490 * same cgroup as the old css_set.
492 if (cgrp1->root == new_cgrp->root) {
493 if (cgrp1 != new_cgrp)
504 * find_existing_css_set - init css array and find the matching css_set
505 * @old_cset: the css_set that we're using before the cgroup transition
506 * @cgrp: the cgroup that we're moving into
507 * @template: out param for the new set of csses, should be clear on entry
509 static struct css_set *find_existing_css_set(struct css_set *old_cset,
511 struct cgroup_subsys_state *template[])
513 struct cgroupfs_root *root = cgrp->root;
514 struct cgroup_subsys *ss;
515 struct css_set *cset;
520 * Build the set of subsystem state objects that we want to see in the
521 * new css_set. while subsystems can change globally, the entries here
522 * won't change, so no need for locking.
524 for_each_subsys(ss, i) {
525 if (root->subsys_mask & (1UL << i)) {
526 /* Subsystem is in this hierarchy. So we want
527 * the subsystem state from the new
529 template[i] = cgroup_css(cgrp, ss);
531 /* Subsystem is not in this hierarchy, so we
532 * don't want to change the subsystem state */
533 template[i] = old_cset->subsys[i];
537 key = css_set_hash(template);
538 hash_for_each_possible(css_set_table, cset, hlist, key) {
539 if (!compare_css_sets(cset, old_cset, cgrp, template))
542 /* This css_set matches what we need */
546 /* No existing cgroup group matched */
550 static void free_cgrp_cset_links(struct list_head *links_to_free)
552 struct cgrp_cset_link *link, *tmp_link;
554 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
555 list_del(&link->cset_link);
561 * allocate_cgrp_cset_links - allocate cgrp_cset_links
562 * @count: the number of links to allocate
563 * @tmp_links: list_head the allocated links are put on
565 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
566 * through ->cset_link. Returns 0 on success or -errno.
568 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
570 struct cgrp_cset_link *link;
573 INIT_LIST_HEAD(tmp_links);
575 for (i = 0; i < count; i++) {
576 link = kzalloc(sizeof(*link), GFP_KERNEL);
578 free_cgrp_cset_links(tmp_links);
581 list_add(&link->cset_link, tmp_links);
587 * link_css_set - a helper function to link a css_set to a cgroup
588 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
589 * @cset: the css_set to be linked
590 * @cgrp: the destination cgroup
592 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
595 struct cgrp_cset_link *link;
597 BUG_ON(list_empty(tmp_links));
598 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
601 list_move(&link->cset_link, &cgrp->cset_links);
603 * Always add links to the tail of the list so that the list
604 * is sorted by order of hierarchy creation
606 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
610 * find_css_set - return a new css_set with one cgroup updated
611 * @old_cset: the baseline css_set
612 * @cgrp: the cgroup to be updated
614 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
615 * substituted into the appropriate hierarchy.
617 static struct css_set *find_css_set(struct css_set *old_cset,
620 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
621 struct css_set *cset;
622 struct list_head tmp_links;
623 struct cgrp_cset_link *link;
626 lockdep_assert_held(&cgroup_mutex);
628 /* First see if we already have a cgroup group that matches
630 down_read(&css_set_rwsem);
631 cset = find_existing_css_set(old_cset, cgrp, template);
634 up_read(&css_set_rwsem);
639 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
643 /* Allocate all the cgrp_cset_link objects that we'll need */
644 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
649 atomic_set(&cset->refcount, 1);
650 INIT_LIST_HEAD(&cset->cgrp_links);
651 INIT_LIST_HEAD(&cset->tasks);
652 INIT_HLIST_NODE(&cset->hlist);
654 /* Copy the set of subsystem state objects generated in
655 * find_existing_css_set() */
656 memcpy(cset->subsys, template, sizeof(cset->subsys));
658 down_write(&css_set_rwsem);
659 /* Add reference counts and links from the new css_set. */
660 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
661 struct cgroup *c = link->cgrp;
663 if (c->root == cgrp->root)
665 link_css_set(&tmp_links, cset, c);
668 BUG_ON(!list_empty(&tmp_links));
672 /* Add this cgroup group to the hash table */
673 key = css_set_hash(cset->subsys);
674 hash_add(css_set_table, &cset->hlist, key);
676 up_write(&css_set_rwsem);
681 static struct cgroupfs_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
683 struct cgroup *top_cgrp = kf_root->kn->priv;
685 return top_cgrp->root;
688 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
692 lockdep_assert_held(&cgroup_mutex);
694 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
699 root->hierarchy_id = id;
703 static void cgroup_exit_root_id(struct cgroupfs_root *root)
705 lockdep_assert_held(&cgroup_mutex);
707 if (root->hierarchy_id) {
708 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
709 root->hierarchy_id = 0;
713 static void cgroup_free_root(struct cgroupfs_root *root)
716 /* hierarhcy ID shoulid already have been released */
717 WARN_ON_ONCE(root->hierarchy_id);
719 idr_destroy(&root->cgroup_idr);
724 static void cgroup_destroy_root(struct cgroupfs_root *root)
726 struct cgroup *cgrp = &root->top_cgroup;
727 struct cgrp_cset_link *link, *tmp_link;
729 mutex_lock(&cgroup_tree_mutex);
730 mutex_lock(&cgroup_mutex);
732 BUG_ON(atomic_read(&root->nr_cgrps));
733 BUG_ON(!list_empty(&cgrp->children));
735 /* Rebind all subsystems back to the default hierarchy */
736 WARN_ON(rebind_subsystems(root, 0, root->subsys_mask));
739 * Release all the links from cset_links to this hierarchy's
742 down_write(&css_set_rwsem);
744 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
745 list_del(&link->cset_link);
746 list_del(&link->cgrp_link);
749 up_write(&css_set_rwsem);
751 if (!list_empty(&root->root_list)) {
752 list_del(&root->root_list);
756 cgroup_exit_root_id(root);
758 mutex_unlock(&cgroup_mutex);
759 mutex_unlock(&cgroup_tree_mutex);
761 kernfs_destroy_root(root->kf_root);
762 cgroup_free_root(root);
766 * Return the cgroup for "task" from the given hierarchy. Must be
767 * called with cgroup_mutex and css_set_rwsem held.
769 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
770 struct cgroupfs_root *root)
772 struct css_set *cset;
773 struct cgroup *res = NULL;
775 lockdep_assert_held(&cgroup_mutex);
776 lockdep_assert_held(&css_set_rwsem);
779 * No need to lock the task - since we hold cgroup_mutex the
780 * task can't change groups, so the only thing that can happen
781 * is that it exits and its css is set back to init_css_set.
783 cset = task_css_set(task);
784 if (cset == &init_css_set) {
785 res = &root->top_cgroup;
787 struct cgrp_cset_link *link;
789 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
790 struct cgroup *c = link->cgrp;
792 if (c->root == root) {
804 * There is one global cgroup mutex. We also require taking
805 * task_lock() when dereferencing a task's cgroup subsys pointers.
806 * See "The task_lock() exception", at the end of this comment.
808 * A task must hold cgroup_mutex to modify cgroups.
810 * Any task can increment and decrement the count field without lock.
811 * So in general, code holding cgroup_mutex can't rely on the count
812 * field not changing. However, if the count goes to zero, then only
813 * cgroup_attach_task() can increment it again. Because a count of zero
814 * means that no tasks are currently attached, therefore there is no
815 * way a task attached to that cgroup can fork (the other way to
816 * increment the count). So code holding cgroup_mutex can safely
817 * assume that if the count is zero, it will stay zero. Similarly, if
818 * a task holds cgroup_mutex on a cgroup with zero count, it
819 * knows that the cgroup won't be removed, as cgroup_rmdir()
822 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
823 * (usually) take cgroup_mutex. These are the two most performance
824 * critical pieces of code here. The exception occurs on cgroup_exit(),
825 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
826 * is taken, and if the cgroup count is zero, a usermode call made
827 * to the release agent with the name of the cgroup (path relative to
828 * the root of cgroup file system) as the argument.
830 * A cgroup can only be deleted if both its 'count' of using tasks
831 * is zero, and its list of 'children' cgroups is empty. Since all
832 * tasks in the system use _some_ cgroup, and since there is always at
833 * least one task in the system (init, pid == 1), therefore, top_cgroup
834 * always has either children cgroups and/or using tasks. So we don't
835 * need a special hack to ensure that top_cgroup cannot be deleted.
837 * The task_lock() exception
839 * The need for this exception arises from the action of
840 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
841 * another. It does so using cgroup_mutex, however there are
842 * several performance critical places that need to reference
843 * task->cgroup without the expense of grabbing a system global
844 * mutex. Therefore except as noted below, when dereferencing or, as
845 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
846 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
847 * the task_struct routinely used for such matters.
849 * P.S. One more locking exception. RCU is used to guard the
850 * update of a tasks cgroup pointer by cgroup_attach_task()
853 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
854 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
855 static const struct file_operations proc_cgroupstats_operations;
857 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
860 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
861 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
862 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
863 cft->ss->name, cft->name);
865 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
870 * cgroup_file_mode - deduce file mode of a control file
871 * @cft: the control file in question
873 * returns cft->mode if ->mode is not 0
874 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
875 * returns S_IRUGO if it has only a read handler
876 * returns S_IWUSR if it has only a write hander
878 static umode_t cgroup_file_mode(const struct cftype *cft)
885 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
888 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
895 static void cgroup_free_fn(struct work_struct *work)
897 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
899 atomic_dec(&cgrp->root->nr_cgrps);
900 cgroup_pidlist_destroy_all(cgrp);
904 * We get a ref to the parent, and put the ref when this
905 * cgroup is being freed, so it's guaranteed that the
906 * parent won't be destroyed before its children.
908 cgroup_put(cgrp->parent);
909 kernfs_put(cgrp->kn);
913 * This is top cgroup's refcnt reaching zero, which
914 * indicates that the root should be released.
916 cgroup_destroy_root(cgrp->root);
920 static void cgroup_free_rcu(struct rcu_head *head)
922 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
924 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
925 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
928 static void cgroup_get(struct cgroup *cgrp)
930 WARN_ON_ONCE(cgroup_is_dead(cgrp));
931 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
932 atomic_inc(&cgrp->refcnt);
935 static void cgroup_put(struct cgroup *cgrp)
937 if (!atomic_dec_and_test(&cgrp->refcnt))
939 if (WARN_ON_ONCE(cgrp->parent && !cgroup_is_dead(cgrp)))
943 * XXX: cgrp->id is only used to look up css's. As cgroup and
944 * css's lifetimes will be decoupled, it should be made
945 * per-subsystem and moved to css->id so that lookups are
946 * successful until the target css is released.
948 mutex_lock(&cgroup_mutex);
949 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
950 mutex_unlock(&cgroup_mutex);
953 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
956 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
958 char name[CGROUP_FILE_NAME_MAX];
960 lockdep_assert_held(&cgroup_tree_mutex);
961 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
965 * cgroup_clear_dir - remove subsys files in a cgroup directory
966 * @cgrp: target cgroup
967 * @subsys_mask: mask of the subsystem ids whose files should be removed
969 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
971 struct cgroup_subsys *ss;
974 for_each_subsys(ss, i) {
977 if (!test_bit(i, &subsys_mask))
979 list_for_each_entry(cfts, &ss->cfts, node)
980 cgroup_addrm_files(cgrp, cfts, false);
984 static int rebind_subsystems(struct cgroupfs_root *root,
985 unsigned long added_mask, unsigned removed_mask)
987 struct cgroup *cgrp = &root->top_cgroup;
988 struct cgroup_subsys *ss;
991 lockdep_assert_held(&cgroup_tree_mutex);
992 lockdep_assert_held(&cgroup_mutex);
994 /* Check that any added subsystems are currently free */
995 for_each_subsys(ss, i)
996 if ((added_mask & (1 << i)) && ss->root != &cgroup_dummy_root)
999 ret = cgroup_populate_dir(cgrp, added_mask);
1004 * Nothing can fail from this point on. Remove files for the
1005 * removed subsystems and rebind each subsystem.
1007 mutex_unlock(&cgroup_mutex);
1008 cgroup_clear_dir(cgrp, removed_mask);
1009 mutex_lock(&cgroup_mutex);
1011 for_each_subsys(ss, i) {
1012 unsigned long bit = 1UL << i;
1014 if (bit & added_mask) {
1015 /* We're binding this subsystem to this hierarchy */
1016 BUG_ON(cgroup_css(cgrp, ss));
1017 BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
1018 BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
1020 rcu_assign_pointer(cgrp->subsys[i],
1021 cgroup_css(cgroup_dummy_top, ss));
1022 cgroup_css(cgrp, ss)->cgroup = cgrp;
1026 ss->bind(cgroup_css(cgrp, ss));
1028 /* refcount was already taken, and we're keeping it */
1029 root->subsys_mask |= bit;
1030 } else if (bit & removed_mask) {
1031 /* We're removing this subsystem */
1032 BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1033 BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1036 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1038 cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1039 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1041 cgroup_subsys[i]->root = &cgroup_dummy_root;
1042 root->subsys_mask &= ~bit;
1046 kernfs_activate(cgrp->kn);
1050 static int cgroup_show_options(struct seq_file *seq,
1051 struct kernfs_root *kf_root)
1053 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1054 struct cgroup_subsys *ss;
1057 for_each_subsys(ss, ssid)
1058 if (root->subsys_mask & (1 << ssid))
1059 seq_printf(seq, ",%s", ss->name);
1060 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1061 seq_puts(seq, ",sane_behavior");
1062 if (root->flags & CGRP_ROOT_NOPREFIX)
1063 seq_puts(seq, ",noprefix");
1064 if (root->flags & CGRP_ROOT_XATTR)
1065 seq_puts(seq, ",xattr");
1067 spin_lock(&release_agent_path_lock);
1068 if (strlen(root->release_agent_path))
1069 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1070 spin_unlock(&release_agent_path_lock);
1072 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1073 seq_puts(seq, ",clone_children");
1074 if (strlen(root->name))
1075 seq_printf(seq, ",name=%s", root->name);
1079 struct cgroup_sb_opts {
1080 unsigned long subsys_mask;
1081 unsigned long flags;
1082 char *release_agent;
1083 bool cpuset_clone_children;
1085 /* User explicitly requested empty subsystem */
1090 * Convert a hierarchy specifier into a bitmask of subsystems and
1091 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1092 * array. This function takes refcounts on subsystems to be used, unless it
1093 * returns error, in which case no refcounts are taken.
1095 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1097 char *token, *o = data;
1098 bool all_ss = false, one_ss = false;
1099 unsigned long mask = (unsigned long)-1;
1100 struct cgroup_subsys *ss;
1103 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1105 #ifdef CONFIG_CPUSETS
1106 mask = ~(1UL << cpuset_cgrp_id);
1109 memset(opts, 0, sizeof(*opts));
1111 while ((token = strsep(&o, ",")) != NULL) {
1114 if (!strcmp(token, "none")) {
1115 /* Explicitly have no subsystems */
1119 if (!strcmp(token, "all")) {
1120 /* Mutually exclusive option 'all' + subsystem name */
1126 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1127 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1130 if (!strcmp(token, "noprefix")) {
1131 opts->flags |= CGRP_ROOT_NOPREFIX;
1134 if (!strcmp(token, "clone_children")) {
1135 opts->cpuset_clone_children = true;
1138 if (!strcmp(token, "xattr")) {
1139 opts->flags |= CGRP_ROOT_XATTR;
1142 if (!strncmp(token, "release_agent=", 14)) {
1143 /* Specifying two release agents is forbidden */
1144 if (opts->release_agent)
1146 opts->release_agent =
1147 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1148 if (!opts->release_agent)
1152 if (!strncmp(token, "name=", 5)) {
1153 const char *name = token + 5;
1154 /* Can't specify an empty name */
1157 /* Must match [\w.-]+ */
1158 for (i = 0; i < strlen(name); i++) {
1162 if ((c == '.') || (c == '-') || (c == '_'))
1166 /* Specifying two names is forbidden */
1169 opts->name = kstrndup(name,
1170 MAX_CGROUP_ROOT_NAMELEN - 1,
1178 for_each_subsys(ss, i) {
1179 if (strcmp(token, ss->name))
1184 /* Mutually exclusive option 'all' + subsystem name */
1187 set_bit(i, &opts->subsys_mask);
1192 if (i == CGROUP_SUBSYS_COUNT)
1197 * If the 'all' option was specified select all the subsystems,
1198 * otherwise if 'none', 'name=' and a subsystem name options
1199 * were not specified, let's default to 'all'
1201 if (all_ss || (!one_ss && !opts->none && !opts->name))
1202 for_each_subsys(ss, i)
1204 set_bit(i, &opts->subsys_mask);
1206 /* Consistency checks */
1208 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1209 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1211 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1212 opts->cpuset_clone_children || opts->release_agent ||
1214 pr_err("cgroup: sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1220 * Option noprefix was introduced just for backward compatibility
1221 * with the old cpuset, so we allow noprefix only if mounting just
1222 * the cpuset subsystem.
1224 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1228 /* Can't specify "none" and some subsystems */
1229 if (opts->subsys_mask && opts->none)
1233 * We either have to specify by name or by subsystems. (So all
1234 * empty hierarchies must have a name).
1236 if (!opts->subsys_mask && !opts->name)
1242 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1245 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1246 struct cgroup_sb_opts opts;
1247 unsigned long added_mask, removed_mask;
1249 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1250 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1254 mutex_lock(&cgroup_tree_mutex);
1255 mutex_lock(&cgroup_mutex);
1257 /* See what subsystems are wanted */
1258 ret = parse_cgroupfs_options(data, &opts);
1262 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1263 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1264 task_tgid_nr(current), current->comm);
1266 added_mask = opts.subsys_mask & ~root->subsys_mask;
1267 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1269 /* Don't allow flags or name to change at remount */
1270 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1271 (opts.name && strcmp(opts.name, root->name))) {
1272 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1273 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1274 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1279 /* remounting is not allowed for populated hierarchies */
1280 if (!list_empty(&root->top_cgroup.children)) {
1285 ret = rebind_subsystems(root, added_mask, removed_mask);
1289 if (opts.release_agent) {
1290 spin_lock(&release_agent_path_lock);
1291 strcpy(root->release_agent_path, opts.release_agent);
1292 spin_unlock(&release_agent_path_lock);
1295 kfree(opts.release_agent);
1297 mutex_unlock(&cgroup_mutex);
1298 mutex_unlock(&cgroup_tree_mutex);
1303 * To reduce the fork() overhead for systems that are not actually using
1304 * their cgroups capability, we don't maintain the lists running through
1305 * each css_set to its tasks until we see the list actually used - in other
1306 * words after the first mount.
1308 static bool use_task_css_set_links __read_mostly;
1310 static void cgroup_enable_task_cg_lists(void)
1312 struct task_struct *p, *g;
1314 down_write(&css_set_rwsem);
1316 if (use_task_css_set_links)
1319 use_task_css_set_links = true;
1322 * We need tasklist_lock because RCU is not safe against
1323 * while_each_thread(). Besides, a forking task that has passed
1324 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1325 * is not guaranteed to have its child immediately visible in the
1326 * tasklist if we walk through it with RCU.
1328 read_lock(&tasklist_lock);
1329 do_each_thread(g, p) {
1332 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1333 task_css_set(p) != &init_css_set);
1336 * We should check if the process is exiting, otherwise
1337 * it will race with cgroup_exit() in that the list
1338 * entry won't be deleted though the process has exited.
1340 if (!(p->flags & PF_EXITING))
1341 list_add(&p->cg_list, &task_css_set(p)->tasks);
1344 } while_each_thread(g, p);
1345 read_unlock(&tasklist_lock);
1347 up_write(&css_set_rwsem);
1350 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1352 atomic_set(&cgrp->refcnt, 1);
1353 INIT_LIST_HEAD(&cgrp->sibling);
1354 INIT_LIST_HEAD(&cgrp->children);
1355 INIT_LIST_HEAD(&cgrp->cset_links);
1356 INIT_LIST_HEAD(&cgrp->release_list);
1357 INIT_LIST_HEAD(&cgrp->pidlists);
1358 mutex_init(&cgrp->pidlist_mutex);
1359 cgrp->dummy_css.cgroup = cgrp;
1362 static void init_cgroup_root(struct cgroupfs_root *root)
1364 struct cgroup *cgrp = &root->top_cgroup;
1366 INIT_LIST_HEAD(&root->root_list);
1367 atomic_set(&root->nr_cgrps, 1);
1369 init_cgroup_housekeeping(cgrp);
1370 idr_init(&root->cgroup_idr);
1373 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1375 struct cgroupfs_root *root;
1377 if (!opts->subsys_mask && !opts->none)
1378 return ERR_PTR(-EINVAL);
1380 root = kzalloc(sizeof(*root), GFP_KERNEL);
1382 return ERR_PTR(-ENOMEM);
1384 init_cgroup_root(root);
1386 root->flags = opts->flags;
1387 if (opts->release_agent)
1388 strcpy(root->release_agent_path, opts->release_agent);
1390 strcpy(root->name, opts->name);
1391 if (opts->cpuset_clone_children)
1392 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1396 static int cgroup_setup_root(struct cgroupfs_root *root, unsigned long ss_mask)
1398 LIST_HEAD(tmp_links);
1399 struct cgroup *root_cgrp = &root->top_cgroup;
1400 struct css_set *cset;
1403 lockdep_assert_held(&cgroup_tree_mutex);
1404 lockdep_assert_held(&cgroup_mutex);
1406 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1409 root_cgrp->id = ret;
1412 * We're accessing css_set_count without locking css_set_rwsem here,
1413 * but that's OK - it can only be increased by someone holding
1414 * cgroup_lock, and that's us. The worst that can happen is that we
1415 * have some link structures left over
1417 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1421 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1422 ret = cgroup_init_root_id(root, 2, 0);
1426 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1427 KERNFS_ROOT_CREATE_DEACTIVATED,
1429 if (IS_ERR(root->kf_root)) {
1430 ret = PTR_ERR(root->kf_root);
1433 root_cgrp->kn = root->kf_root->kn;
1435 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1439 ret = rebind_subsystems(root, ss_mask, 0);
1444 * There must be no failure case after here, since rebinding takes
1445 * care of subsystems' refcounts, which are explicitly dropped in
1446 * the failure exit path.
1448 list_add(&root->root_list, &cgroup_roots);
1449 cgroup_root_count++;
1452 * Link the top cgroup in this hierarchy into all the css_set
1455 down_write(&css_set_rwsem);
1456 hash_for_each(css_set_table, i, cset, hlist)
1457 link_css_set(&tmp_links, cset, root_cgrp);
1458 up_write(&css_set_rwsem);
1460 BUG_ON(!list_empty(&root_cgrp->children));
1461 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1463 kernfs_activate(root_cgrp->kn);
1468 kernfs_destroy_root(root->kf_root);
1469 root->kf_root = NULL;
1471 cgroup_exit_root_id(root);
1473 free_cgrp_cset_links(&tmp_links);
1477 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1478 int flags, const char *unused_dev_name,
1481 struct cgroupfs_root *root;
1482 struct cgroup_sb_opts opts;
1483 struct dentry *dentry;
1487 * The first time anyone tries to mount a cgroup, enable the list
1488 * linking each css_set to its tasks and fix up all existing tasks.
1490 if (!use_task_css_set_links)
1491 cgroup_enable_task_cg_lists();
1493 mutex_lock(&cgroup_tree_mutex);
1494 mutex_lock(&cgroup_mutex);
1496 /* First find the desired set of subsystems */
1497 ret = parse_cgroupfs_options(data, &opts);
1501 /* look for a matching existing root */
1502 for_each_active_root(root) {
1503 bool name_match = false;
1506 * If we asked for a name then it must match. Also, if
1507 * name matches but sybsys_mask doesn't, we should fail.
1508 * Remember whether name matched.
1511 if (strcmp(opts.name, root->name))
1517 * If we asked for subsystems (or explicitly for no
1518 * subsystems) then they must match.
1520 if ((opts.subsys_mask || opts.none) &&
1521 (opts.subsys_mask != root->subsys_mask)) {
1528 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1529 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1530 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1534 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1539 * A root's lifetime is governed by its top cgroup. Zero
1540 * ref indicate that the root is being destroyed. Wait for
1541 * destruction to complete so that the subsystems are free.
1542 * We can use wait_queue for the wait but this path is
1543 * super cold. Let's just sleep for a bit and retry.
1545 if (!atomic_inc_not_zero(&root->top_cgroup.refcnt)) {
1546 mutex_unlock(&cgroup_mutex);
1547 mutex_unlock(&cgroup_tree_mutex);
1556 /* no such thing, create a new one */
1557 root = cgroup_root_from_opts(&opts);
1559 ret = PTR_ERR(root);
1563 ret = cgroup_setup_root(root, opts.subsys_mask);
1565 cgroup_free_root(root);
1568 mutex_unlock(&cgroup_mutex);
1569 mutex_unlock(&cgroup_tree_mutex);
1571 kfree(opts.release_agent);
1575 return ERR_PTR(ret);
1577 dentry = kernfs_mount(fs_type, flags, root->kf_root);
1579 cgroup_put(&root->top_cgroup);
1583 static void cgroup_kill_sb(struct super_block *sb)
1585 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1586 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1588 cgroup_put(&root->top_cgroup);
1592 static struct file_system_type cgroup_fs_type = {
1594 .mount = cgroup_mount,
1595 .kill_sb = cgroup_kill_sb,
1598 static struct kobject *cgroup_kobj;
1601 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1602 * @task: target task
1603 * @buf: the buffer to write the path into
1604 * @buflen: the length of the buffer
1606 * Determine @task's cgroup on the first (the one with the lowest non-zero
1607 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1608 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1609 * cgroup controller callbacks.
1611 * Return value is the same as kernfs_path().
1613 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1615 struct cgroupfs_root *root;
1616 struct cgroup *cgrp;
1617 int hierarchy_id = 1;
1620 mutex_lock(&cgroup_mutex);
1621 down_read(&css_set_rwsem);
1623 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1626 cgrp = task_cgroup_from_root(task, root);
1627 path = cgroup_path(cgrp, buf, buflen);
1629 /* if no hierarchy exists, everyone is in "/" */
1630 if (strlcpy(buf, "/", buflen) < buflen)
1634 up_read(&css_set_rwsem);
1635 mutex_unlock(&cgroup_mutex);
1638 EXPORT_SYMBOL_GPL(task_cgroup_path);
1641 * Control Group taskset
1643 struct task_and_cgroup {
1644 struct task_struct *task;
1645 struct cgroup *cgrp;
1646 struct css_set *cset;
1649 struct cgroup_taskset {
1650 struct task_and_cgroup single;
1651 struct flex_array *tc_array;
1654 struct cgroup *cur_cgrp;
1658 * cgroup_taskset_first - reset taskset and return the first task
1659 * @tset: taskset of interest
1661 * @tset iteration is initialized and the first task is returned.
1663 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1665 if (tset->tc_array) {
1667 return cgroup_taskset_next(tset);
1669 tset->cur_cgrp = tset->single.cgrp;
1670 return tset->single.task;
1673 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1676 * cgroup_taskset_next - iterate to the next task in taskset
1677 * @tset: taskset of interest
1679 * Return the next task in @tset. Iteration must have been initialized
1680 * with cgroup_taskset_first().
1682 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1684 struct task_and_cgroup *tc;
1686 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1689 tc = flex_array_get(tset->tc_array, tset->idx++);
1690 tset->cur_cgrp = tc->cgrp;
1693 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1696 * cgroup_taskset_cur_css - return the matching css for the current task
1697 * @tset: taskset of interest
1698 * @subsys_id: the ID of the target subsystem
1700 * Return the css for the current (last returned) task of @tset for
1701 * subsystem specified by @subsys_id. This function must be preceded by
1702 * either cgroup_taskset_first() or cgroup_taskset_next().
1704 struct cgroup_subsys_state *cgroup_taskset_cur_css(struct cgroup_taskset *tset,
1707 return cgroup_css(tset->cur_cgrp, cgroup_subsys[subsys_id]);
1709 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_css);
1712 * cgroup_taskset_size - return the number of tasks in taskset
1713 * @tset: taskset of interest
1715 int cgroup_taskset_size(struct cgroup_taskset *tset)
1717 return tset->tc_array ? tset->tc_array_len : 1;
1719 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1723 * cgroup_task_migrate - move a task from one cgroup to another.
1725 * Must be called with cgroup_mutex and threadgroup locked.
1727 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1728 struct task_struct *tsk,
1729 struct css_set *new_cset)
1731 struct css_set *old_cset;
1734 * We are synchronized through threadgroup_lock() against PF_EXITING
1735 * setting such that we can't race against cgroup_exit() changing the
1736 * css_set to init_css_set and dropping the old one.
1738 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1739 old_cset = task_css_set(tsk);
1742 rcu_assign_pointer(tsk->cgroups, new_cset);
1745 down_write(&css_set_rwsem);
1746 list_move(&tsk->cg_list, &new_cset->tasks);
1747 up_write(&css_set_rwsem);
1750 * We just gained a reference on old_cset by taking it from the
1751 * task. As trading it for new_cset is protected by cgroup_mutex,
1752 * we're safe to drop it here; it will be freed under RCU.
1754 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1755 put_css_set(old_cset);
1759 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1760 * @cgrp: the cgroup to attach to
1761 * @tsk: the task or the leader of the threadgroup to be attached
1762 * @threadgroup: attach the whole threadgroup?
1764 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1765 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1767 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1770 int retval, i, group_size;
1771 struct cgroupfs_root *root = cgrp->root;
1772 struct cgroup_subsys_state *css, *failed_css = NULL;
1773 /* threadgroup list cursor and array */
1774 struct task_struct *leader = tsk;
1775 struct task_and_cgroup *tc;
1776 struct flex_array *group;
1777 struct cgroup_taskset tset = { };
1780 * step 0: in order to do expensive, possibly blocking operations for
1781 * every thread, we cannot iterate the thread group list, since it needs
1782 * rcu or tasklist locked. instead, build an array of all threads in the
1783 * group - group_rwsem prevents new threads from appearing, and if
1784 * threads exit, this will just be an over-estimate.
1787 group_size = get_nr_threads(tsk);
1790 /* flex_array supports very large thread-groups better than kmalloc. */
1791 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1794 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1795 retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1797 goto out_free_group_list;
1801 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1802 * already PF_EXITING could be freed from underneath us unless we
1803 * take an rcu_read_lock.
1805 down_read(&css_set_rwsem);
1808 struct task_and_cgroup ent;
1810 /* @tsk either already exited or can't exit until the end */
1811 if (tsk->flags & PF_EXITING)
1814 /* as per above, nr_threads may decrease, but not increase. */
1815 BUG_ON(i >= group_size);
1817 ent.cgrp = task_cgroup_from_root(tsk, root);
1818 /* nothing to do if this task is already in the cgroup */
1819 if (ent.cgrp == cgrp)
1822 * saying GFP_ATOMIC has no effect here because we did prealloc
1823 * earlier, but it's good form to communicate our expectations.
1825 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
1826 BUG_ON(retval != 0);
1831 } while_each_thread(leader, tsk);
1833 up_read(&css_set_rwsem);
1834 /* remember the number of threads in the array for later. */
1836 tset.tc_array = group;
1837 tset.tc_array_len = group_size;
1839 /* methods shouldn't be called if no task is actually migrating */
1842 goto out_free_group_list;
1845 * step 1: check that we can legitimately attach to the cgroup.
1847 for_each_css(css, i, cgrp) {
1848 if (css->ss->can_attach) {
1849 retval = css->ss->can_attach(css, &tset);
1852 goto out_cancel_attach;
1858 * step 2: make sure css_sets exist for all threads to be migrated.
1859 * we use find_css_set, which allocates a new one if necessary.
1861 for (i = 0; i < group_size; i++) {
1862 struct css_set *old_cset;
1864 tc = flex_array_get(group, i);
1865 old_cset = task_css_set(tc->task);
1866 tc->cset = find_css_set(old_cset, cgrp);
1869 goto out_put_css_set_refs;
1874 * step 3: now that we're guaranteed success wrt the css_sets,
1875 * proceed to move all tasks to the new cgroup. There are no
1876 * failure cases after here, so this is the commit point.
1878 for (i = 0; i < group_size; i++) {
1879 tc = flex_array_get(group, i);
1880 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
1882 /* nothing is sensitive to fork() after this point. */
1885 * step 4: do subsystem attach callbacks.
1887 for_each_css(css, i, cgrp)
1888 if (css->ss->attach)
1889 css->ss->attach(css, &tset);
1892 * step 5: success! and cleanup
1895 out_put_css_set_refs:
1897 for (i = 0; i < group_size; i++) {
1898 tc = flex_array_get(group, i);
1901 put_css_set(tc->cset);
1906 for_each_css(css, i, cgrp) {
1907 if (css == failed_css)
1909 if (css->ss->cancel_attach)
1910 css->ss->cancel_attach(css, &tset);
1913 out_free_group_list:
1914 flex_array_free(group);
1919 * Find the task_struct of the task to attach by vpid and pass it along to the
1920 * function to attach either it or all tasks in its threadgroup. Will lock
1921 * cgroup_mutex and threadgroup; may take task_lock of task.
1923 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
1925 struct task_struct *tsk;
1926 const struct cred *cred = current_cred(), *tcred;
1929 if (!cgroup_lock_live_group(cgrp))
1935 tsk = find_task_by_vpid(pid);
1939 goto out_unlock_cgroup;
1942 * even if we're attaching all tasks in the thread group, we
1943 * only need to check permissions on one of them.
1945 tcred = __task_cred(tsk);
1946 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
1947 !uid_eq(cred->euid, tcred->uid) &&
1948 !uid_eq(cred->euid, tcred->suid)) {
1951 goto out_unlock_cgroup;
1957 tsk = tsk->group_leader;
1960 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
1961 * trapped in a cpuset, or RT worker may be born in a cgroup
1962 * with no rt_runtime allocated. Just say no.
1964 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
1967 goto out_unlock_cgroup;
1970 get_task_struct(tsk);
1973 threadgroup_lock(tsk);
1975 if (!thread_group_leader(tsk)) {
1977 * a race with de_thread from another thread's exec()
1978 * may strip us of our leadership, if this happens,
1979 * there is no choice but to throw this task away and
1980 * try again; this is
1981 * "double-double-toil-and-trouble-check locking".
1983 threadgroup_unlock(tsk);
1984 put_task_struct(tsk);
1985 goto retry_find_task;
1989 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
1991 threadgroup_unlock(tsk);
1993 put_task_struct(tsk);
1995 mutex_unlock(&cgroup_mutex);
2000 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2001 * @from: attach to all cgroups of a given task
2002 * @tsk: the task to be attached
2004 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2006 struct cgroupfs_root *root;
2009 mutex_lock(&cgroup_mutex);
2010 for_each_active_root(root) {
2011 struct cgroup *from_cgrp;
2013 down_read(&css_set_rwsem);
2014 from_cgrp = task_cgroup_from_root(from, root);
2015 up_read(&css_set_rwsem);
2017 retval = cgroup_attach_task(from_cgrp, tsk, false);
2021 mutex_unlock(&cgroup_mutex);
2025 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2027 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2028 struct cftype *cft, u64 pid)
2030 return attach_task_by_pid(css->cgroup, pid, false);
2033 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2034 struct cftype *cft, u64 tgid)
2036 return attach_task_by_pid(css->cgroup, tgid, true);
2039 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2040 struct cftype *cft, const char *buffer)
2042 struct cgroupfs_root *root = css->cgroup->root;
2044 BUILD_BUG_ON(sizeof(root->release_agent_path) < PATH_MAX);
2045 if (!cgroup_lock_live_group(css->cgroup))
2047 spin_lock(&release_agent_path_lock);
2048 strlcpy(root->release_agent_path, buffer,
2049 sizeof(root->release_agent_path));
2050 spin_unlock(&release_agent_path_lock);
2051 mutex_unlock(&cgroup_mutex);
2055 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2057 struct cgroup *cgrp = seq_css(seq)->cgroup;
2059 if (!cgroup_lock_live_group(cgrp))
2061 seq_puts(seq, cgrp->root->release_agent_path);
2062 seq_putc(seq, '\n');
2063 mutex_unlock(&cgroup_mutex);
2067 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2069 struct cgroup *cgrp = seq_css(seq)->cgroup;
2071 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2075 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2076 size_t nbytes, loff_t off)
2078 struct cgroup *cgrp = of->kn->parent->priv;
2079 struct cftype *cft = of->kn->priv;
2080 struct cgroup_subsys_state *css;
2084 * kernfs guarantees that a file isn't deleted with operations in
2085 * flight, which means that the matching css is and stays alive and
2086 * doesn't need to be pinned. The RCU locking is not necessary
2087 * either. It's just for the convenience of using cgroup_css().
2090 css = cgroup_css(cgrp, cft->ss);
2093 if (cft->write_string) {
2094 ret = cft->write_string(css, cft, strstrip(buf));
2095 } else if (cft->write_u64) {
2096 unsigned long long v;
2097 ret = kstrtoull(buf, 0, &v);
2099 ret = cft->write_u64(css, cft, v);
2100 } else if (cft->write_s64) {
2102 ret = kstrtoll(buf, 0, &v);
2104 ret = cft->write_s64(css, cft, v);
2105 } else if (cft->trigger) {
2106 ret = cft->trigger(css, (unsigned int)cft->private);
2111 return ret ?: nbytes;
2114 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2116 return seq_cft(seq)->seq_start(seq, ppos);
2119 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2121 return seq_cft(seq)->seq_next(seq, v, ppos);
2124 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2126 seq_cft(seq)->seq_stop(seq, v);
2129 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2131 struct cftype *cft = seq_cft(m);
2132 struct cgroup_subsys_state *css = seq_css(m);
2135 return cft->seq_show(m, arg);
2138 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2139 else if (cft->read_s64)
2140 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2146 static struct kernfs_ops cgroup_kf_single_ops = {
2147 .atomic_write_len = PAGE_SIZE,
2148 .write = cgroup_file_write,
2149 .seq_show = cgroup_seqfile_show,
2152 static struct kernfs_ops cgroup_kf_ops = {
2153 .atomic_write_len = PAGE_SIZE,
2154 .write = cgroup_file_write,
2155 .seq_start = cgroup_seqfile_start,
2156 .seq_next = cgroup_seqfile_next,
2157 .seq_stop = cgroup_seqfile_stop,
2158 .seq_show = cgroup_seqfile_show,
2162 * cgroup_rename - Only allow simple rename of directories in place.
2164 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2165 const char *new_name_str)
2167 struct cgroup *cgrp = kn->priv;
2170 if (kernfs_type(kn) != KERNFS_DIR)
2172 if (kn->parent != new_parent)
2176 * This isn't a proper migration and its usefulness is very
2177 * limited. Disallow if sane_behavior.
2179 if (cgroup_sane_behavior(cgrp))
2182 mutex_lock(&cgroup_tree_mutex);
2183 mutex_lock(&cgroup_mutex);
2185 ret = kernfs_rename(kn, new_parent, new_name_str);
2187 mutex_unlock(&cgroup_mutex);
2188 mutex_unlock(&cgroup_tree_mutex);
2192 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2194 char name[CGROUP_FILE_NAME_MAX];
2195 struct kernfs_node *kn;
2196 struct lock_class_key *key = NULL;
2198 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2199 key = &cft->lockdep_key;
2201 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2202 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2210 * cgroup_addrm_files - add or remove files to a cgroup directory
2211 * @cgrp: the target cgroup
2212 * @cfts: array of cftypes to be added
2213 * @is_add: whether to add or remove
2215 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2216 * For removals, this function never fails. If addition fails, this
2217 * function doesn't remove files already added. The caller is responsible
2220 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2226 lockdep_assert_held(&cgroup_tree_mutex);
2228 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2229 /* does cft->flags tell us to skip this file on @cgrp? */
2230 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2232 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2234 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2238 ret = cgroup_add_file(cgrp, cft);
2240 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2245 cgroup_rm_file(cgrp, cft);
2251 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2254 struct cgroup_subsys *ss = cfts[0].ss;
2255 struct cgroup *root = &ss->root->top_cgroup;
2256 struct cgroup_subsys_state *css;
2259 lockdep_assert_held(&cgroup_tree_mutex);
2261 /* don't bother if @ss isn't attached */
2262 if (ss->root == &cgroup_dummy_root)
2265 /* add/rm files for all cgroups created before */
2266 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2267 struct cgroup *cgrp = css->cgroup;
2269 if (cgroup_is_dead(cgrp))
2272 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2278 kernfs_activate(root->kn);
2282 static void cgroup_exit_cftypes(struct cftype *cfts)
2286 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2287 /* free copy for custom atomic_write_len, see init_cftypes() */
2288 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2295 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2299 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2300 struct kernfs_ops *kf_ops;
2302 WARN_ON(cft->ss || cft->kf_ops);
2305 kf_ops = &cgroup_kf_ops;
2307 kf_ops = &cgroup_kf_single_ops;
2310 * Ugh... if @cft wants a custom max_write_len, we need to
2311 * make a copy of kf_ops to set its atomic_write_len.
2313 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
2314 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
2316 cgroup_exit_cftypes(cfts);
2319 kf_ops->atomic_write_len = cft->max_write_len;
2322 cft->kf_ops = kf_ops;
2329 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
2331 lockdep_assert_held(&cgroup_tree_mutex);
2333 if (!cfts || !cfts[0].ss)
2336 list_del(&cfts->node);
2337 cgroup_apply_cftypes(cfts, false);
2338 cgroup_exit_cftypes(cfts);
2343 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2344 * @cfts: zero-length name terminated array of cftypes
2346 * Unregister @cfts. Files described by @cfts are removed from all
2347 * existing cgroups and all future cgroups won't have them either. This
2348 * function can be called anytime whether @cfts' subsys is attached or not.
2350 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2353 int cgroup_rm_cftypes(struct cftype *cfts)
2357 mutex_lock(&cgroup_tree_mutex);
2358 ret = cgroup_rm_cftypes_locked(cfts);
2359 mutex_unlock(&cgroup_tree_mutex);
2364 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2365 * @ss: target cgroup subsystem
2366 * @cfts: zero-length name terminated array of cftypes
2368 * Register @cfts to @ss. Files described by @cfts are created for all
2369 * existing cgroups to which @ss is attached and all future cgroups will
2370 * have them too. This function can be called anytime whether @ss is
2373 * Returns 0 on successful registration, -errno on failure. Note that this
2374 * function currently returns 0 as long as @cfts registration is successful
2375 * even if some file creation attempts on existing cgroups fail.
2377 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2381 ret = cgroup_init_cftypes(ss, cfts);
2385 mutex_lock(&cgroup_tree_mutex);
2387 list_add_tail(&cfts->node, &ss->cfts);
2388 ret = cgroup_apply_cftypes(cfts, true);
2390 cgroup_rm_cftypes_locked(cfts);
2392 mutex_unlock(&cgroup_tree_mutex);
2395 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2398 * cgroup_task_count - count the number of tasks in a cgroup.
2399 * @cgrp: the cgroup in question
2401 * Return the number of tasks in the cgroup.
2403 static int cgroup_task_count(const struct cgroup *cgrp)
2406 struct cgrp_cset_link *link;
2408 down_read(&css_set_rwsem);
2409 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2410 count += atomic_read(&link->cset->refcount);
2411 up_read(&css_set_rwsem);
2416 * css_next_child - find the next child of a given css
2417 * @pos_css: the current position (%NULL to initiate traversal)
2418 * @parent_css: css whose children to walk
2420 * This function returns the next child of @parent_css and should be called
2421 * under either cgroup_mutex or RCU read lock. The only requirement is
2422 * that @parent_css and @pos_css are accessible. The next sibling is
2423 * guaranteed to be returned regardless of their states.
2425 struct cgroup_subsys_state *
2426 css_next_child(struct cgroup_subsys_state *pos_css,
2427 struct cgroup_subsys_state *parent_css)
2429 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2430 struct cgroup *cgrp = parent_css->cgroup;
2431 struct cgroup *next;
2433 cgroup_assert_mutexes_or_rcu_locked();
2436 * @pos could already have been removed. Once a cgroup is removed,
2437 * its ->sibling.next is no longer updated when its next sibling
2438 * changes. As CGRP_DEAD assertion is serialized and happens
2439 * before the cgroup is taken off the ->sibling list, if we see it
2440 * unasserted, it's guaranteed that the next sibling hasn't
2441 * finished its grace period even if it's already removed, and thus
2442 * safe to dereference from this RCU critical section. If
2443 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2444 * to be visible as %true here.
2446 * If @pos is dead, its next pointer can't be dereferenced;
2447 * however, as each cgroup is given a monotonically increasing
2448 * unique serial number and always appended to the sibling list,
2449 * the next one can be found by walking the parent's children until
2450 * we see a cgroup with higher serial number than @pos's. While
2451 * this path can be slower, it's taken only when either the current
2452 * cgroup is removed or iteration and removal race.
2455 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2456 } else if (likely(!cgroup_is_dead(pos))) {
2457 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2459 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2460 if (next->serial_nr > pos->serial_nr)
2464 if (&next->sibling == &cgrp->children)
2467 return cgroup_css(next, parent_css->ss);
2469 EXPORT_SYMBOL_GPL(css_next_child);
2472 * css_next_descendant_pre - find the next descendant for pre-order walk
2473 * @pos: the current position (%NULL to initiate traversal)
2474 * @root: css whose descendants to walk
2476 * To be used by css_for_each_descendant_pre(). Find the next descendant
2477 * to visit for pre-order traversal of @root's descendants. @root is
2478 * included in the iteration and the first node to be visited.
2480 * While this function requires cgroup_mutex or RCU read locking, it
2481 * doesn't require the whole traversal to be contained in a single critical
2482 * section. This function will return the correct next descendant as long
2483 * as both @pos and @root are accessible and @pos is a descendant of @root.
2485 struct cgroup_subsys_state *
2486 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2487 struct cgroup_subsys_state *root)
2489 struct cgroup_subsys_state *next;
2491 cgroup_assert_mutexes_or_rcu_locked();
2493 /* if first iteration, visit @root */
2497 /* visit the first child if exists */
2498 next = css_next_child(NULL, pos);
2502 /* no child, visit my or the closest ancestor's next sibling */
2503 while (pos != root) {
2504 next = css_next_child(pos, css_parent(pos));
2507 pos = css_parent(pos);
2512 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
2515 * css_rightmost_descendant - return the rightmost descendant of a css
2516 * @pos: css of interest
2518 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2519 * is returned. This can be used during pre-order traversal to skip
2522 * While this function requires cgroup_mutex or RCU read locking, it
2523 * doesn't require the whole traversal to be contained in a single critical
2524 * section. This function will return the correct rightmost descendant as
2525 * long as @pos is accessible.
2527 struct cgroup_subsys_state *
2528 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2530 struct cgroup_subsys_state *last, *tmp;
2532 cgroup_assert_mutexes_or_rcu_locked();
2536 /* ->prev isn't RCU safe, walk ->next till the end */
2538 css_for_each_child(tmp, last)
2544 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
2546 static struct cgroup_subsys_state *
2547 css_leftmost_descendant(struct cgroup_subsys_state *pos)
2549 struct cgroup_subsys_state *last;
2553 pos = css_next_child(NULL, pos);
2560 * css_next_descendant_post - find the next descendant for post-order walk
2561 * @pos: the current position (%NULL to initiate traversal)
2562 * @root: css whose descendants to walk
2564 * To be used by css_for_each_descendant_post(). Find the next descendant
2565 * to visit for post-order traversal of @root's descendants. @root is
2566 * included in the iteration and the last node to be visited.
2568 * While this function requires cgroup_mutex or RCU read locking, it
2569 * doesn't require the whole traversal to be contained in a single critical
2570 * section. This function will return the correct next descendant as long
2571 * as both @pos and @cgroup are accessible and @pos is a descendant of
2574 struct cgroup_subsys_state *
2575 css_next_descendant_post(struct cgroup_subsys_state *pos,
2576 struct cgroup_subsys_state *root)
2578 struct cgroup_subsys_state *next;
2580 cgroup_assert_mutexes_or_rcu_locked();
2582 /* if first iteration, visit leftmost descendant which may be @root */
2584 return css_leftmost_descendant(root);
2586 /* if we visited @root, we're done */
2590 /* if there's an unvisited sibling, visit its leftmost descendant */
2591 next = css_next_child(pos, css_parent(pos));
2593 return css_leftmost_descendant(next);
2595 /* no sibling left, visit parent */
2596 return css_parent(pos);
2598 EXPORT_SYMBOL_GPL(css_next_descendant_post);
2601 * css_advance_task_iter - advance a task itererator to the next css_set
2602 * @it: the iterator to advance
2604 * Advance @it to the next css_set to walk.
2606 static void css_advance_task_iter(struct css_task_iter *it)
2608 struct list_head *l = it->cset_link;
2609 struct cgrp_cset_link *link;
2610 struct css_set *cset;
2612 /* Advance to the next non-empty css_set */
2615 if (l == &it->origin_css->cgroup->cset_links) {
2616 it->cset_link = NULL;
2619 link = list_entry(l, struct cgrp_cset_link, cset_link);
2621 } while (list_empty(&cset->tasks));
2623 it->task = cset->tasks.next;
2627 * css_task_iter_start - initiate task iteration
2628 * @css: the css to walk tasks of
2629 * @it: the task iterator to use
2631 * Initiate iteration through the tasks of @css. The caller can call
2632 * css_task_iter_next() to walk through the tasks until the function
2633 * returns NULL. On completion of iteration, css_task_iter_end() must be
2636 * Note that this function acquires a lock which is released when the
2637 * iteration finishes. The caller can't sleep while iteration is in
2640 void css_task_iter_start(struct cgroup_subsys_state *css,
2641 struct css_task_iter *it)
2642 __acquires(css_set_rwsem)
2644 /* no one should try to iterate before mounting cgroups */
2645 WARN_ON_ONCE(!use_task_css_set_links);
2647 down_read(&css_set_rwsem);
2649 it->origin_css = css;
2650 it->cset_link = &css->cgroup->cset_links;
2652 css_advance_task_iter(it);
2656 * css_task_iter_next - return the next task for the iterator
2657 * @it: the task iterator being iterated
2659 * The "next" function for task iteration. @it should have been
2660 * initialized via css_task_iter_start(). Returns NULL when the iteration
2663 struct task_struct *css_task_iter_next(struct css_task_iter *it)
2665 struct task_struct *res;
2666 struct list_head *l = it->task;
2667 struct cgrp_cset_link *link;
2669 /* If the iterator cg is NULL, we have no tasks */
2672 res = list_entry(l, struct task_struct, cg_list);
2673 /* Advance iterator to find next entry */
2675 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
2676 if (l == &link->cset->tasks) {
2678 * We reached the end of this task list - move on to the
2679 * next cgrp_cset_link.
2681 css_advance_task_iter(it);
2689 * css_task_iter_end - finish task iteration
2690 * @it: the task iterator to finish
2692 * Finish task iteration started by css_task_iter_start().
2694 void css_task_iter_end(struct css_task_iter *it)
2695 __releases(css_set_rwsem)
2697 up_read(&css_set_rwsem);
2700 static inline int started_after_time(struct task_struct *t1,
2701 struct timespec *time,
2702 struct task_struct *t2)
2704 int start_diff = timespec_compare(&t1->start_time, time);
2705 if (start_diff > 0) {
2707 } else if (start_diff < 0) {
2711 * Arbitrarily, if two processes started at the same
2712 * time, we'll say that the lower pointer value
2713 * started first. Note that t2 may have exited by now
2714 * so this may not be a valid pointer any longer, but
2715 * that's fine - it still serves to distinguish
2716 * between two tasks started (effectively) simultaneously.
2723 * This function is a callback from heap_insert() and is used to order
2725 * In this case we order the heap in descending task start time.
2727 static inline int started_after(void *p1, void *p2)
2729 struct task_struct *t1 = p1;
2730 struct task_struct *t2 = p2;
2731 return started_after_time(t1, &t2->start_time, t2);
2735 * css_scan_tasks - iterate though all the tasks in a css
2736 * @css: the css to iterate tasks of
2737 * @test: optional test callback
2738 * @process: process callback
2739 * @data: data passed to @test and @process
2740 * @heap: optional pre-allocated heap used for task iteration
2742 * Iterate through all the tasks in @css, calling @test for each, and if it
2743 * returns %true, call @process for it also.
2745 * @test may be NULL, meaning always true (select all tasks), which
2746 * effectively duplicates css_task_iter_{start,next,end}() but does not
2747 * lock css_set_rwsem for the call to @process.
2749 * It is guaranteed that @process will act on every task that is a member
2750 * of @css for the duration of this call. This function may or may not
2751 * call @process for tasks that exit or move to a different css during the
2752 * call, or are forked or move into the css during the call.
2754 * Note that @test may be called with locks held, and may in some
2755 * situations be called multiple times for the same task, so it should be
2758 * If @heap is non-NULL, a heap has been pre-allocated and will be used for
2759 * heap operations (and its "gt" member will be overwritten), else a
2760 * temporary heap will be used (allocation of which may cause this function
2763 int css_scan_tasks(struct cgroup_subsys_state *css,
2764 bool (*test)(struct task_struct *, void *),
2765 void (*process)(struct task_struct *, void *),
2766 void *data, struct ptr_heap *heap)
2769 struct css_task_iter it;
2770 struct task_struct *p, *dropped;
2771 /* Never dereference latest_task, since it's not refcounted */
2772 struct task_struct *latest_task = NULL;
2773 struct ptr_heap tmp_heap;
2774 struct timespec latest_time = { 0, 0 };
2777 /* The caller supplied our heap and pre-allocated its memory */
2778 heap->gt = &started_after;
2780 /* We need to allocate our own heap memory */
2782 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2784 /* cannot allocate the heap */
2790 * Scan tasks in the css, using the @test callback to determine
2791 * which are of interest, and invoking @process callback on the
2792 * ones which need an update. Since we don't want to hold any
2793 * locks during the task updates, gather tasks to be processed in a
2794 * heap structure. The heap is sorted by descending task start
2795 * time. If the statically-sized heap fills up, we overflow tasks
2796 * that started later, and in future iterations only consider tasks
2797 * that started after the latest task in the previous pass. This
2798 * guarantees forward progress and that we don't miss any tasks.
2801 css_task_iter_start(css, &it);
2802 while ((p = css_task_iter_next(&it))) {
2804 * Only affect tasks that qualify per the caller's callback,
2805 * if he provided one
2807 if (test && !test(p, data))
2810 * Only process tasks that started after the last task
2813 if (!started_after_time(p, &latest_time, latest_task))
2815 dropped = heap_insert(heap, p);
2816 if (dropped == NULL) {
2818 * The new task was inserted; the heap wasn't
2822 } else if (dropped != p) {
2824 * The new task was inserted, and pushed out a
2828 put_task_struct(dropped);
2831 * Else the new task was newer than anything already in
2832 * the heap and wasn't inserted
2835 css_task_iter_end(&it);
2838 for (i = 0; i < heap->size; i++) {
2839 struct task_struct *q = heap->ptrs[i];
2841 latest_time = q->start_time;
2844 /* Process the task per the caller's callback */
2849 * If we had to process any tasks at all, scan again
2850 * in case some of them were in the middle of forking
2851 * children that didn't get processed.
2852 * Not the most efficient way to do it, but it avoids
2853 * having to take callback_mutex in the fork path
2857 if (heap == &tmp_heap)
2858 heap_free(&tmp_heap);
2863 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
2864 * @to: cgroup to which the tasks will be moved
2865 * @from: cgroup in which the tasks currently reside
2867 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
2869 struct css_task_iter it;
2870 struct task_struct *task;
2874 css_task_iter_start(&from->dummy_css, &it);
2875 task = css_task_iter_next(&it);
2877 get_task_struct(task);
2878 css_task_iter_end(&it);
2881 mutex_lock(&cgroup_mutex);
2882 ret = cgroup_attach_task(to, task, false);
2883 mutex_unlock(&cgroup_mutex);
2884 put_task_struct(task);
2886 } while (task && !ret);
2892 * Stuff for reading the 'tasks'/'procs' files.
2894 * Reading this file can return large amounts of data if a cgroup has
2895 * *lots* of attached tasks. So it may need several calls to read(),
2896 * but we cannot guarantee that the information we produce is correct
2897 * unless we produce it entirely atomically.
2901 /* which pidlist file are we talking about? */
2902 enum cgroup_filetype {
2908 * A pidlist is a list of pids that virtually represents the contents of one
2909 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
2910 * a pair (one each for procs, tasks) for each pid namespace that's relevant
2913 struct cgroup_pidlist {
2915 * used to find which pidlist is wanted. doesn't change as long as
2916 * this particular list stays in the list.
2918 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
2921 /* how many elements the above list has */
2923 /* each of these stored in a list by its cgroup */
2924 struct list_head links;
2925 /* pointer to the cgroup we belong to, for list removal purposes */
2926 struct cgroup *owner;
2927 /* for delayed destruction */
2928 struct delayed_work destroy_dwork;
2932 * The following two functions "fix" the issue where there are more pids
2933 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2934 * TODO: replace with a kernel-wide solution to this problem
2936 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2937 static void *pidlist_allocate(int count)
2939 if (PIDLIST_TOO_LARGE(count))
2940 return vmalloc(count * sizeof(pid_t));
2942 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2945 static void pidlist_free(void *p)
2947 if (is_vmalloc_addr(p))
2954 * Used to destroy all pidlists lingering waiting for destroy timer. None
2955 * should be left afterwards.
2957 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
2959 struct cgroup_pidlist *l, *tmp_l;
2961 mutex_lock(&cgrp->pidlist_mutex);
2962 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
2963 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
2964 mutex_unlock(&cgrp->pidlist_mutex);
2966 flush_workqueue(cgroup_pidlist_destroy_wq);
2967 BUG_ON(!list_empty(&cgrp->pidlists));
2970 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
2972 struct delayed_work *dwork = to_delayed_work(work);
2973 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
2975 struct cgroup_pidlist *tofree = NULL;
2977 mutex_lock(&l->owner->pidlist_mutex);
2980 * Destroy iff we didn't get queued again. The state won't change
2981 * as destroy_dwork can only be queued while locked.
2983 if (!delayed_work_pending(dwork)) {
2984 list_del(&l->links);
2985 pidlist_free(l->list);
2986 put_pid_ns(l->key.ns);
2990 mutex_unlock(&l->owner->pidlist_mutex);
2995 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2996 * Returns the number of unique elements.
2998 static int pidlist_uniq(pid_t *list, int length)
3003 * we presume the 0th element is unique, so i starts at 1. trivial
3004 * edge cases first; no work needs to be done for either
3006 if (length == 0 || length == 1)
3008 /* src and dest walk down the list; dest counts unique elements */
3009 for (src = 1; src < length; src++) {
3010 /* find next unique element */
3011 while (list[src] == list[src-1]) {
3016 /* dest always points to where the next unique element goes */
3017 list[dest] = list[src];
3025 * The two pid files - task and cgroup.procs - guaranteed that the result
3026 * is sorted, which forced this whole pidlist fiasco. As pid order is
3027 * different per namespace, each namespace needs differently sorted list,
3028 * making it impossible to use, for example, single rbtree of member tasks
3029 * sorted by task pointer. As pidlists can be fairly large, allocating one
3030 * per open file is dangerous, so cgroup had to implement shared pool of
3031 * pidlists keyed by cgroup and namespace.
3033 * All this extra complexity was caused by the original implementation
3034 * committing to an entirely unnecessary property. In the long term, we
3035 * want to do away with it. Explicitly scramble sort order if
3036 * sane_behavior so that no such expectation exists in the new interface.
3038 * Scrambling is done by swapping every two consecutive bits, which is
3039 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3041 static pid_t pid_fry(pid_t pid)
3043 unsigned a = pid & 0x55555555;
3044 unsigned b = pid & 0xAAAAAAAA;
3046 return (a << 1) | (b >> 1);
3049 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3051 if (cgroup_sane_behavior(cgrp))
3052 return pid_fry(pid);
3057 static int cmppid(const void *a, const void *b)
3059 return *(pid_t *)a - *(pid_t *)b;
3062 static int fried_cmppid(const void *a, const void *b)
3064 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3067 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3068 enum cgroup_filetype type)
3070 struct cgroup_pidlist *l;
3071 /* don't need task_nsproxy() if we're looking at ourself */
3072 struct pid_namespace *ns = task_active_pid_ns(current);
3074 lockdep_assert_held(&cgrp->pidlist_mutex);
3076 list_for_each_entry(l, &cgrp->pidlists, links)
3077 if (l->key.type == type && l->key.ns == ns)
3083 * find the appropriate pidlist for our purpose (given procs vs tasks)
3084 * returns with the lock on that pidlist already held, and takes care
3085 * of the use count, or returns NULL with no locks held if we're out of
3088 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3089 enum cgroup_filetype type)
3091 struct cgroup_pidlist *l;
3093 lockdep_assert_held(&cgrp->pidlist_mutex);
3095 l = cgroup_pidlist_find(cgrp, type);
3099 /* entry not found; create a new one */
3100 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3104 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3106 /* don't need task_nsproxy() if we're looking at ourself */
3107 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3109 list_add(&l->links, &cgrp->pidlists);
3114 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3116 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3117 struct cgroup_pidlist **lp)
3121 int pid, n = 0; /* used for populating the array */
3122 struct css_task_iter it;
3123 struct task_struct *tsk;
3124 struct cgroup_pidlist *l;
3126 lockdep_assert_held(&cgrp->pidlist_mutex);
3129 * If cgroup gets more users after we read count, we won't have
3130 * enough space - tough. This race is indistinguishable to the
3131 * caller from the case that the additional cgroup users didn't
3132 * show up until sometime later on.
3134 length = cgroup_task_count(cgrp);
3135 array = pidlist_allocate(length);
3138 /* now, populate the array */
3139 css_task_iter_start(&cgrp->dummy_css, &it);
3140 while ((tsk = css_task_iter_next(&it))) {
3141 if (unlikely(n == length))
3143 /* get tgid or pid for procs or tasks file respectively */
3144 if (type == CGROUP_FILE_PROCS)
3145 pid = task_tgid_vnr(tsk);
3147 pid = task_pid_vnr(tsk);
3148 if (pid > 0) /* make sure to only use valid results */
3151 css_task_iter_end(&it);
3153 /* now sort & (if procs) strip out duplicates */
3154 if (cgroup_sane_behavior(cgrp))
3155 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3157 sort(array, length, sizeof(pid_t), cmppid, NULL);
3158 if (type == CGROUP_FILE_PROCS)
3159 length = pidlist_uniq(array, length);
3161 l = cgroup_pidlist_find_create(cgrp, type);
3163 mutex_unlock(&cgrp->pidlist_mutex);
3164 pidlist_free(array);
3168 /* store array, freeing old if necessary */
3169 pidlist_free(l->list);
3177 * cgroupstats_build - build and fill cgroupstats
3178 * @stats: cgroupstats to fill information into
3179 * @dentry: A dentry entry belonging to the cgroup for which stats have
3182 * Build and fill cgroupstats so that taskstats can export it to user
3185 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3187 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3188 struct cgroup *cgrp;
3189 struct css_task_iter it;
3190 struct task_struct *tsk;
3192 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3193 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3194 kernfs_type(kn) != KERNFS_DIR)
3198 * We aren't being called from kernfs and there's no guarantee on
3199 * @kn->priv's validity. For this and css_tryget_from_dir(),
3200 * @kn->priv is RCU safe. Let's do the RCU dancing.
3203 cgrp = rcu_dereference(kn->priv);
3209 css_task_iter_start(&cgrp->dummy_css, &it);
3210 while ((tsk = css_task_iter_next(&it))) {
3211 switch (tsk->state) {
3213 stats->nr_running++;
3215 case TASK_INTERRUPTIBLE:
3216 stats->nr_sleeping++;
3218 case TASK_UNINTERRUPTIBLE:
3219 stats->nr_uninterruptible++;
3222 stats->nr_stopped++;
3225 if (delayacct_is_task_waiting_on_io(tsk))
3226 stats->nr_io_wait++;
3230 css_task_iter_end(&it);
3238 * seq_file methods for the tasks/procs files. The seq_file position is the
3239 * next pid to display; the seq_file iterator is a pointer to the pid
3240 * in the cgroup->l->list array.
3243 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3246 * Initially we receive a position value that corresponds to
3247 * one more than the last pid shown (or 0 on the first call or
3248 * after a seek to the start). Use a binary-search to find the
3249 * next pid to display, if any
3251 struct kernfs_open_file *of = s->private;
3252 struct cgroup *cgrp = seq_css(s)->cgroup;
3253 struct cgroup_pidlist *l;
3254 enum cgroup_filetype type = seq_cft(s)->private;
3255 int index = 0, pid = *pos;
3258 mutex_lock(&cgrp->pidlist_mutex);
3261 * !NULL @of->priv indicates that this isn't the first start()
3262 * after open. If the matching pidlist is around, we can use that.
3263 * Look for it. Note that @of->priv can't be used directly. It
3264 * could already have been destroyed.
3267 of->priv = cgroup_pidlist_find(cgrp, type);
3270 * Either this is the first start() after open or the matching
3271 * pidlist has been destroyed inbetween. Create a new one.
3274 ret = pidlist_array_load(cgrp, type,
3275 (struct cgroup_pidlist **)&of->priv);
3277 return ERR_PTR(ret);
3282 int end = l->length;
3284 while (index < end) {
3285 int mid = (index + end) / 2;
3286 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3289 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3295 /* If we're off the end of the array, we're done */
3296 if (index >= l->length)
3298 /* Update the abstract position to be the actual pid that we found */
3299 iter = l->list + index;
3300 *pos = cgroup_pid_fry(cgrp, *iter);
3304 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3306 struct kernfs_open_file *of = s->private;
3307 struct cgroup_pidlist *l = of->priv;
3310 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3311 CGROUP_PIDLIST_DESTROY_DELAY);
3312 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3315 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3317 struct kernfs_open_file *of = s->private;
3318 struct cgroup_pidlist *l = of->priv;
3320 pid_t *end = l->list + l->length;
3322 * Advance to the next pid in the array. If this goes off the
3329 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3334 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3336 return seq_printf(s, "%d\n", *(int *)v);
3340 * seq_operations functions for iterating on pidlists through seq_file -
3341 * independent of whether it's tasks or procs
3343 static const struct seq_operations cgroup_pidlist_seq_operations = {
3344 .start = cgroup_pidlist_start,
3345 .stop = cgroup_pidlist_stop,
3346 .next = cgroup_pidlist_next,
3347 .show = cgroup_pidlist_show,
3350 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3353 return notify_on_release(css->cgroup);
3356 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3357 struct cftype *cft, u64 val)
3359 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3361 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3363 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3367 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3370 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3373 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3374 struct cftype *cft, u64 val)
3377 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3379 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3383 static struct cftype cgroup_base_files[] = {
3385 .name = "cgroup.procs",
3386 .seq_start = cgroup_pidlist_start,
3387 .seq_next = cgroup_pidlist_next,
3388 .seq_stop = cgroup_pidlist_stop,
3389 .seq_show = cgroup_pidlist_show,
3390 .private = CGROUP_FILE_PROCS,
3391 .write_u64 = cgroup_procs_write,
3392 .mode = S_IRUGO | S_IWUSR,
3395 .name = "cgroup.clone_children",
3396 .flags = CFTYPE_INSANE,
3397 .read_u64 = cgroup_clone_children_read,
3398 .write_u64 = cgroup_clone_children_write,
3401 .name = "cgroup.sane_behavior",
3402 .flags = CFTYPE_ONLY_ON_ROOT,
3403 .seq_show = cgroup_sane_behavior_show,
3407 * Historical crazy stuff. These don't have "cgroup." prefix and
3408 * don't exist if sane_behavior. If you're depending on these, be
3409 * prepared to be burned.
3413 .flags = CFTYPE_INSANE, /* use "procs" instead */
3414 .seq_start = cgroup_pidlist_start,
3415 .seq_next = cgroup_pidlist_next,
3416 .seq_stop = cgroup_pidlist_stop,
3417 .seq_show = cgroup_pidlist_show,
3418 .private = CGROUP_FILE_TASKS,
3419 .write_u64 = cgroup_tasks_write,
3420 .mode = S_IRUGO | S_IWUSR,
3423 .name = "notify_on_release",
3424 .flags = CFTYPE_INSANE,
3425 .read_u64 = cgroup_read_notify_on_release,
3426 .write_u64 = cgroup_write_notify_on_release,
3429 .name = "release_agent",
3430 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3431 .seq_show = cgroup_release_agent_show,
3432 .write_string = cgroup_release_agent_write,
3433 .max_write_len = PATH_MAX - 1,
3439 * cgroup_populate_dir - create subsys files in a cgroup directory
3440 * @cgrp: target cgroup
3441 * @subsys_mask: mask of the subsystem ids whose files should be added
3443 * On failure, no file is added.
3445 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3447 struct cgroup_subsys *ss;
3450 /* process cftsets of each subsystem */
3451 for_each_subsys(ss, i) {
3452 struct cftype *cfts;
3454 if (!test_bit(i, &subsys_mask))
3457 list_for_each_entry(cfts, &ss->cfts, node) {
3458 ret = cgroup_addrm_files(cgrp, cfts, true);
3465 cgroup_clear_dir(cgrp, subsys_mask);
3470 * css destruction is four-stage process.
3472 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3473 * Implemented in kill_css().
3475 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3476 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3477 * by invoking offline_css(). After offlining, the base ref is put.
3478 * Implemented in css_killed_work_fn().
3480 * 3. When the percpu_ref reaches zero, the only possible remaining
3481 * accessors are inside RCU read sections. css_release() schedules the
3484 * 4. After the grace period, the css can be freed. Implemented in
3485 * css_free_work_fn().
3487 * It is actually hairier because both step 2 and 4 require process context
3488 * and thus involve punting to css->destroy_work adding two additional
3489 * steps to the already complex sequence.
3491 static void css_free_work_fn(struct work_struct *work)
3493 struct cgroup_subsys_state *css =
3494 container_of(work, struct cgroup_subsys_state, destroy_work);
3495 struct cgroup *cgrp = css->cgroup;
3498 css_put(css->parent);
3500 css->ss->css_free(css);
3504 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3506 struct cgroup_subsys_state *css =
3507 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3509 INIT_WORK(&css->destroy_work, css_free_work_fn);
3510 queue_work(cgroup_destroy_wq, &css->destroy_work);
3513 static void css_release(struct percpu_ref *ref)
3515 struct cgroup_subsys_state *css =
3516 container_of(ref, struct cgroup_subsys_state, refcnt);
3518 rcu_assign_pointer(css->cgroup->subsys[css->ss->id], NULL);
3519 call_rcu(&css->rcu_head, css_free_rcu_fn);
3522 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3523 struct cgroup *cgrp)
3530 css->parent = cgroup_css(cgrp->parent, ss);
3532 css->flags |= CSS_ROOT;
3534 BUG_ON(cgroup_css(cgrp, ss));
3537 /* invoke ->css_online() on a new CSS and mark it online if successful */
3538 static int online_css(struct cgroup_subsys_state *css)
3540 struct cgroup_subsys *ss = css->ss;
3543 lockdep_assert_held(&cgroup_tree_mutex);
3544 lockdep_assert_held(&cgroup_mutex);
3547 ret = ss->css_online(css);
3549 css->flags |= CSS_ONLINE;
3550 css->cgroup->nr_css++;
3551 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
3556 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
3557 static void offline_css(struct cgroup_subsys_state *css)
3559 struct cgroup_subsys *ss = css->ss;
3561 lockdep_assert_held(&cgroup_tree_mutex);
3562 lockdep_assert_held(&cgroup_mutex);
3564 if (!(css->flags & CSS_ONLINE))
3567 if (ss->css_offline)
3568 ss->css_offline(css);
3570 css->flags &= ~CSS_ONLINE;
3571 css->cgroup->nr_css--;
3572 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
3576 * create_css - create a cgroup_subsys_state
3577 * @cgrp: the cgroup new css will be associated with
3578 * @ss: the subsys of new css
3580 * Create a new css associated with @cgrp - @ss pair. On success, the new
3581 * css is online and installed in @cgrp with all interface files created.
3582 * Returns 0 on success, -errno on failure.
3584 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
3586 struct cgroup *parent = cgrp->parent;
3587 struct cgroup_subsys_state *css;
3590 lockdep_assert_held(&cgroup_mutex);
3592 css = ss->css_alloc(cgroup_css(parent, ss));
3594 return PTR_ERR(css);
3596 err = percpu_ref_init(&css->refcnt, css_release);
3600 init_css(css, ss, cgrp);
3602 err = cgroup_populate_dir(cgrp, 1 << ss->id);
3606 err = online_css(css);
3611 css_get(css->parent);
3613 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
3615 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",
3616 current->comm, current->pid, ss->name);
3617 if (!strcmp(ss->name, "memory"))
3618 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
3619 ss->warned_broken_hierarchy = true;
3625 percpu_ref_cancel_init(&css->refcnt);
3631 * cgroup_create - create a cgroup
3632 * @parent: cgroup that will be parent of the new cgroup
3633 * @name: name of the new cgroup
3634 * @mode: mode to set on new cgroup
3636 static long cgroup_create(struct cgroup *parent, const char *name,
3639 struct cgroup *cgrp;
3640 struct cgroupfs_root *root = parent->root;
3642 struct cgroup_subsys *ss;
3643 struct kernfs_node *kn;
3645 /* allocate the cgroup and its ID, 0 is reserved for the root */
3646 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3650 mutex_lock(&cgroup_tree_mutex);
3653 * Only live parents can have children. Note that the liveliness
3654 * check isn't strictly necessary because cgroup_mkdir() and
3655 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
3656 * anyway so that locking is contained inside cgroup proper and we
3657 * don't get nasty surprises if we ever grow another caller.
3659 if (!cgroup_lock_live_group(parent)) {
3661 goto err_unlock_tree;
3665 * Temporarily set the pointer to NULL, so idr_find() won't return
3666 * a half-baked cgroup.
3668 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
3674 init_cgroup_housekeeping(cgrp);
3676 cgrp->parent = parent;
3677 cgrp->dummy_css.parent = &parent->dummy_css;
3678 cgrp->root = parent->root;
3680 if (notify_on_release(parent))
3681 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3683 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
3684 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
3686 /* create the directory */
3687 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
3695 * This extra ref will be put in cgroup_free_fn() and guarantees
3696 * that @cgrp->kn is always accessible.
3700 cgrp->serial_nr = cgroup_serial_nr_next++;
3702 /* allocation complete, commit to creation */
3703 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
3704 atomic_inc(&root->nr_cgrps);
3708 * @cgrp is now fully operational. If something fails after this
3709 * point, it'll be released via the normal destruction path.
3711 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
3713 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
3717 /* let's create and online css's */
3718 for_each_subsys(ss, ssid) {
3719 if (root->subsys_mask & (1 << ssid)) {
3720 err = create_css(cgrp, ss);
3726 kernfs_activate(kn);
3728 mutex_unlock(&cgroup_mutex);
3729 mutex_unlock(&cgroup_tree_mutex);
3734 idr_remove(&root->cgroup_idr, cgrp->id);
3736 mutex_unlock(&cgroup_mutex);
3738 mutex_unlock(&cgroup_tree_mutex);
3743 cgroup_destroy_locked(cgrp);
3744 mutex_unlock(&cgroup_mutex);
3745 mutex_unlock(&cgroup_tree_mutex);
3749 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3752 struct cgroup *parent = parent_kn->priv;
3754 return cgroup_create(parent, name, mode);
3758 * This is called when the refcnt of a css is confirmed to be killed.
3759 * css_tryget() is now guaranteed to fail.
3761 static void css_killed_work_fn(struct work_struct *work)
3763 struct cgroup_subsys_state *css =
3764 container_of(work, struct cgroup_subsys_state, destroy_work);
3765 struct cgroup *cgrp = css->cgroup;
3767 mutex_lock(&cgroup_tree_mutex);
3768 mutex_lock(&cgroup_mutex);
3771 * css_tryget() is guaranteed to fail now. Tell subsystems to
3772 * initate destruction.
3777 * If @cgrp is marked dead, it's waiting for refs of all css's to
3778 * be disabled before proceeding to the second phase of cgroup
3779 * destruction. If we are the last one, kick it off.
3781 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
3782 cgroup_destroy_css_killed(cgrp);
3784 mutex_unlock(&cgroup_mutex);
3785 mutex_unlock(&cgroup_tree_mutex);
3788 * Put the css refs from kill_css(). Each css holds an extra
3789 * reference to the cgroup's dentry and cgroup removal proceeds
3790 * regardless of css refs. On the last put of each css, whenever
3791 * that may be, the extra dentry ref is put so that dentry
3792 * destruction happens only after all css's are released.
3797 /* css kill confirmation processing requires process context, bounce */
3798 static void css_killed_ref_fn(struct percpu_ref *ref)
3800 struct cgroup_subsys_state *css =
3801 container_of(ref, struct cgroup_subsys_state, refcnt);
3803 INIT_WORK(&css->destroy_work, css_killed_work_fn);
3804 queue_work(cgroup_destroy_wq, &css->destroy_work);
3808 * kill_css - destroy a css
3809 * @css: css to destroy
3811 * This function initiates destruction of @css by removing cgroup interface
3812 * files and putting its base reference. ->css_offline() will be invoked
3813 * asynchronously once css_tryget() is guaranteed to fail and when the
3814 * reference count reaches zero, @css will be released.
3816 static void kill_css(struct cgroup_subsys_state *css)
3819 * This must happen before css is disassociated with its cgroup.
3820 * See seq_css() for details.
3822 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3825 * Killing would put the base ref, but we need to keep it alive
3826 * until after ->css_offline().
3831 * cgroup core guarantees that, by the time ->css_offline() is
3832 * invoked, no new css reference will be given out via
3833 * css_tryget(). We can't simply call percpu_ref_kill() and
3834 * proceed to offlining css's because percpu_ref_kill() doesn't
3835 * guarantee that the ref is seen as killed on all CPUs on return.
3837 * Use percpu_ref_kill_and_confirm() to get notifications as each
3838 * css is confirmed to be seen as killed on all CPUs.
3840 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
3844 * cgroup_destroy_locked - the first stage of cgroup destruction
3845 * @cgrp: cgroup to be destroyed
3847 * css's make use of percpu refcnts whose killing latency shouldn't be
3848 * exposed to userland and are RCU protected. Also, cgroup core needs to
3849 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
3850 * invoked. To satisfy all the requirements, destruction is implemented in
3851 * the following two steps.
3853 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
3854 * userland visible parts and start killing the percpu refcnts of
3855 * css's. Set up so that the next stage will be kicked off once all
3856 * the percpu refcnts are confirmed to be killed.
3858 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
3859 * rest of destruction. Once all cgroup references are gone, the
3860 * cgroup is RCU-freed.
3862 * This function implements s1. After this step, @cgrp is gone as far as
3863 * the userland is concerned and a new cgroup with the same name may be
3864 * created. As cgroup doesn't care about the names internally, this
3865 * doesn't cause any problem.
3867 static int cgroup_destroy_locked(struct cgroup *cgrp)
3868 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
3870 struct cgroup *child;
3871 struct cgroup_subsys_state *css;
3875 lockdep_assert_held(&cgroup_tree_mutex);
3876 lockdep_assert_held(&cgroup_mutex);
3879 * css_set_rwsem synchronizes access to ->cset_links and prevents
3880 * @cgrp from being removed while __put_css_set() is in progress.
3882 down_read(&css_set_rwsem);
3883 empty = list_empty(&cgrp->cset_links);
3884 up_read(&css_set_rwsem);
3889 * Make sure there's no live children. We can't test ->children
3890 * emptiness as dead children linger on it while being destroyed;
3891 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
3895 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
3896 empty = cgroup_is_dead(child);
3905 * Initiate massacre of all css's. cgroup_destroy_css_killed()
3906 * will be invoked to perform the rest of destruction once the
3907 * percpu refs of all css's are confirmed to be killed. This
3908 * involves removing the subsystem's files, drop cgroup_mutex.
3910 mutex_unlock(&cgroup_mutex);
3911 for_each_css(css, ssid, cgrp)
3913 mutex_lock(&cgroup_mutex);
3916 * Mark @cgrp dead. This prevents further task migration and child
3917 * creation by disabling cgroup_lock_live_group(). Note that
3918 * CGRP_DEAD assertion is depended upon by css_next_child() to
3919 * resume iteration after dropping RCU read lock. See
3920 * css_next_child() for details.
3922 set_bit(CGRP_DEAD, &cgrp->flags);
3924 /* CGRP_DEAD is set, remove from ->release_list for the last time */
3925 raw_spin_lock(&release_list_lock);
3926 if (!list_empty(&cgrp->release_list))
3927 list_del_init(&cgrp->release_list);
3928 raw_spin_unlock(&release_list_lock);
3931 * If @cgrp has css's attached, the second stage of cgroup
3932 * destruction is kicked off from css_killed_work_fn() after the
3933 * refs of all attached css's are killed. If @cgrp doesn't have
3934 * any css, we kick it off here.
3937 cgroup_destroy_css_killed(cgrp);
3939 /* remove @cgrp directory along with the base files */
3940 mutex_unlock(&cgroup_mutex);
3943 * There are two control paths which try to determine cgroup from
3944 * dentry without going through kernfs - cgroupstats_build() and
3945 * css_tryget_from_dir(). Those are supported by RCU protecting
3946 * clearing of cgrp->kn->priv backpointer, which should happen
3947 * after all files under it have been removed.
3949 kernfs_remove(cgrp->kn); /* @cgrp has an extra ref on its kn */
3950 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
3952 mutex_lock(&cgroup_mutex);
3958 * cgroup_destroy_css_killed - the second step of cgroup destruction
3959 * @work: cgroup->destroy_free_work
3961 * This function is invoked from a work item for a cgroup which is being
3962 * destroyed after all css's are offlined and performs the rest of
3963 * destruction. This is the second step of destruction described in the
3964 * comment above cgroup_destroy_locked().
3966 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
3968 struct cgroup *parent = cgrp->parent;
3970 lockdep_assert_held(&cgroup_tree_mutex);
3971 lockdep_assert_held(&cgroup_mutex);
3973 /* delete this cgroup from parent->children */
3974 list_del_rcu(&cgrp->sibling);
3978 set_bit(CGRP_RELEASABLE, &parent->flags);
3979 check_for_release(parent);
3982 static int cgroup_rmdir(struct kernfs_node *kn)
3984 struct cgroup *cgrp = kn->priv;
3988 * This is self-destruction but @kn can't be removed while this
3989 * callback is in progress. Let's break active protection. Once
3990 * the protection is broken, @cgrp can be destroyed at any point.
3991 * Pin it so that it stays accessible.
3994 kernfs_break_active_protection(kn);
3996 mutex_lock(&cgroup_tree_mutex);
3997 mutex_lock(&cgroup_mutex);
4000 * @cgrp might already have been destroyed while we're trying to
4003 if (!cgroup_is_dead(cgrp))
4004 ret = cgroup_destroy_locked(cgrp);
4006 mutex_unlock(&cgroup_mutex);
4007 mutex_unlock(&cgroup_tree_mutex);
4009 kernfs_unbreak_active_protection(kn);
4014 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4015 .remount_fs = cgroup_remount,
4016 .show_options = cgroup_show_options,
4017 .mkdir = cgroup_mkdir,
4018 .rmdir = cgroup_rmdir,
4019 .rename = cgroup_rename,
4022 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4024 struct cgroup_subsys_state *css;
4026 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4028 mutex_lock(&cgroup_tree_mutex);
4029 mutex_lock(&cgroup_mutex);
4031 INIT_LIST_HEAD(&ss->cfts);
4033 /* Create the top cgroup state for this subsystem */
4034 ss->root = &cgroup_dummy_root;
4035 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4036 /* We don't handle early failures gracefully */
4037 BUG_ON(IS_ERR(css));
4038 init_css(css, ss, cgroup_dummy_top);
4040 /* Update the init_css_set to contain a subsys
4041 * pointer to this state - since the subsystem is
4042 * newly registered, all tasks and hence the
4043 * init_css_set is in the subsystem's top cgroup. */
4044 init_css_set.subsys[ss->id] = css;
4046 need_forkexit_callback |= ss->fork || ss->exit;
4048 /* At system boot, before all subsystems have been
4049 * registered, no tasks have been forked, so we don't
4050 * need to invoke fork callbacks here. */
4051 BUG_ON(!list_empty(&init_task.tasks));
4053 BUG_ON(online_css(css));
4055 mutex_unlock(&cgroup_mutex);
4056 mutex_unlock(&cgroup_tree_mutex);
4060 * cgroup_init_early - cgroup initialization at system boot
4062 * Initialize cgroups at system boot, and initialize any
4063 * subsystems that request early init.
4065 int __init cgroup_init_early(void)
4067 struct cgroup_subsys *ss;
4070 atomic_set(&init_css_set.refcount, 1);
4071 INIT_LIST_HEAD(&init_css_set.cgrp_links);
4072 INIT_LIST_HEAD(&init_css_set.tasks);
4073 INIT_HLIST_NODE(&init_css_set.hlist);
4075 init_cgroup_root(&cgroup_dummy_root);
4076 cgroup_root_count = 1;
4077 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4079 init_cgrp_cset_link.cset = &init_css_set;
4080 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4081 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4082 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4084 for_each_subsys(ss, i) {
4085 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4086 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4087 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4089 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4090 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4093 ss->name = cgroup_subsys_name[i];
4096 cgroup_init_subsys(ss);
4102 * cgroup_init - cgroup initialization
4104 * Register cgroup filesystem and /proc file, and initialize
4105 * any subsystems that didn't request early init.
4107 int __init cgroup_init(void)
4109 struct cgroup_subsys *ss;
4113 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4115 for_each_subsys(ss, i) {
4116 if (!ss->early_init)
4117 cgroup_init_subsys(ss);
4120 * cftype registration needs kmalloc and can't be done
4121 * during early_init. Register base cftypes separately.
4123 if (ss->base_cftypes)
4124 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4127 /* allocate id for the dummy hierarchy */
4128 mutex_lock(&cgroup_mutex);
4130 /* Add init_css_set to the hash table */
4131 key = css_set_hash(init_css_set.subsys);
4132 hash_add(css_set_table, &init_css_set.hlist, key);
4134 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4136 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
4140 mutex_unlock(&cgroup_mutex);
4142 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4146 err = register_filesystem(&cgroup_fs_type);
4148 kobject_put(cgroup_kobj);
4152 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4156 static int __init cgroup_wq_init(void)
4159 * There isn't much point in executing destruction path in
4160 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4162 * XXX: Must be ordered to make sure parent is offlined after
4163 * children. The ordering requirement is for memcg where a
4164 * parent's offline may wait for a child's leading to deadlock. In
4165 * the long term, this should be fixed from memcg side.
4167 * We would prefer to do this in cgroup_init() above, but that
4168 * is called before init_workqueues(): so leave this until after.
4170 cgroup_destroy_wq = alloc_ordered_workqueue("cgroup_destroy", 0);
4171 BUG_ON(!cgroup_destroy_wq);
4174 * Used to destroy pidlists and separate to serve as flush domain.
4175 * Cap @max_active to 1 too.
4177 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4179 BUG_ON(!cgroup_pidlist_destroy_wq);
4183 core_initcall(cgroup_wq_init);
4186 * proc_cgroup_show()
4187 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4188 * - Used for /proc/<pid>/cgroup.
4189 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4190 * doesn't really matter if tsk->cgroup changes after we read it,
4191 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4192 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4193 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4194 * cgroup to top_cgroup.
4197 /* TODO: Use a proper seq_file iterator */
4198 int proc_cgroup_show(struct seq_file *m, void *v)
4201 struct task_struct *tsk;
4204 struct cgroupfs_root *root;
4207 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4213 tsk = get_pid_task(pid, PIDTYPE_PID);
4219 mutex_lock(&cgroup_mutex);
4220 down_read(&css_set_rwsem);
4222 for_each_active_root(root) {
4223 struct cgroup_subsys *ss;
4224 struct cgroup *cgrp;
4225 int ssid, count = 0;
4227 seq_printf(m, "%d:", root->hierarchy_id);
4228 for_each_subsys(ss, ssid)
4229 if (root->subsys_mask & (1 << ssid))
4230 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4231 if (strlen(root->name))
4232 seq_printf(m, "%sname=%s", count ? "," : "",
4235 cgrp = task_cgroup_from_root(tsk, root);
4236 path = cgroup_path(cgrp, buf, PATH_MAX);
4238 retval = -ENAMETOOLONG;
4246 up_read(&css_set_rwsem);
4247 mutex_unlock(&cgroup_mutex);
4248 put_task_struct(tsk);
4255 /* Display information about each subsystem and each hierarchy */
4256 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4258 struct cgroup_subsys *ss;
4261 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4263 * ideally we don't want subsystems moving around while we do this.
4264 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4265 * subsys/hierarchy state.
4267 mutex_lock(&cgroup_mutex);
4269 for_each_subsys(ss, i)
4270 seq_printf(m, "%s\t%d\t%d\t%d\n",
4271 ss->name, ss->root->hierarchy_id,
4272 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4274 mutex_unlock(&cgroup_mutex);
4278 static int cgroupstats_open(struct inode *inode, struct file *file)
4280 return single_open(file, proc_cgroupstats_show, NULL);
4283 static const struct file_operations proc_cgroupstats_operations = {
4284 .open = cgroupstats_open,
4286 .llseek = seq_lseek,
4287 .release = single_release,
4291 * cgroup_fork - attach newly forked task to its parents cgroup.
4292 * @child: pointer to task_struct of forking parent process.
4294 * Description: A task inherits its parent's cgroup at fork().
4296 * A pointer to the shared css_set was automatically copied in
4297 * fork.c by dup_task_struct(). However, we ignore that copy, since
4298 * it was not made under the protection of RCU or cgroup_mutex, so
4299 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4300 * have already changed current->cgroups, allowing the previously
4301 * referenced cgroup group to be removed and freed.
4303 * At the point that cgroup_fork() is called, 'current' is the parent
4304 * task, and the passed argument 'child' points to the child task.
4306 void cgroup_fork(struct task_struct *child)
4309 get_css_set(task_css_set(current));
4310 child->cgroups = current->cgroups;
4311 task_unlock(current);
4312 INIT_LIST_HEAD(&child->cg_list);
4316 * cgroup_post_fork - called on a new task after adding it to the task list
4317 * @child: the task in question
4319 * Adds the task to the list running through its css_set if necessary and
4320 * call the subsystem fork() callbacks. Has to be after the task is
4321 * visible on the task list in case we race with the first call to
4322 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4325 void cgroup_post_fork(struct task_struct *child)
4327 struct cgroup_subsys *ss;
4331 * use_task_css_set_links is set to 1 before we walk the tasklist
4332 * under the tasklist_lock and we read it here after we added the child
4333 * to the tasklist under the tasklist_lock as well. If the child wasn't
4334 * yet in the tasklist when we walked through it from
4335 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4336 * should be visible now due to the paired locking and barriers implied
4337 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4338 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4341 if (use_task_css_set_links) {
4342 down_write(&css_set_rwsem);
4344 if (list_empty(&child->cg_list))
4345 list_add(&child->cg_list, &task_css_set(child)->tasks);
4347 up_write(&css_set_rwsem);
4351 * Call ss->fork(). This must happen after @child is linked on
4352 * css_set; otherwise, @child might change state between ->fork()
4353 * and addition to css_set.
4355 if (need_forkexit_callback) {
4356 for_each_subsys(ss, i)
4363 * cgroup_exit - detach cgroup from exiting task
4364 * @tsk: pointer to task_struct of exiting process
4365 * @run_callback: run exit callbacks?
4367 * Description: Detach cgroup from @tsk and release it.
4369 * Note that cgroups marked notify_on_release force every task in
4370 * them to take the global cgroup_mutex mutex when exiting.
4371 * This could impact scaling on very large systems. Be reluctant to
4372 * use notify_on_release cgroups where very high task exit scaling
4373 * is required on large systems.
4375 * the_top_cgroup_hack:
4377 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4379 * We call cgroup_exit() while the task is still competent to
4380 * handle notify_on_release(), then leave the task attached to the
4381 * root cgroup in each hierarchy for the remainder of its exit.
4383 * To do this properly, we would increment the reference count on
4384 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4385 * code we would add a second cgroup function call, to drop that
4386 * reference. This would just create an unnecessary hot spot on
4387 * the top_cgroup reference count, to no avail.
4389 * Normally, holding a reference to a cgroup without bumping its
4390 * count is unsafe. The cgroup could go away, or someone could
4391 * attach us to a different cgroup, decrementing the count on
4392 * the first cgroup that we never incremented. But in this case,
4393 * top_cgroup isn't going away, and either task has PF_EXITING set,
4394 * which wards off any cgroup_attach_task() attempts, or task is a failed
4395 * fork, never visible to cgroup_attach_task.
4397 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4399 struct cgroup_subsys *ss;
4400 struct css_set *cset;
4404 * Unlink from the css_set task list if necessary. Optimistically
4405 * check cg_list before taking css_set_rwsem.
4407 if (!list_empty(&tsk->cg_list)) {
4408 down_write(&css_set_rwsem);
4409 if (!list_empty(&tsk->cg_list))
4410 list_del_init(&tsk->cg_list);
4411 up_write(&css_set_rwsem);
4414 /* Reassign the task to the init_css_set. */
4416 cset = task_css_set(tsk);
4417 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4419 if (run_callbacks && need_forkexit_callback) {
4420 /* see cgroup_post_fork() for details */
4421 for_each_subsys(ss, i) {
4423 struct cgroup_subsys_state *old_css = cset->subsys[i];
4424 struct cgroup_subsys_state *css = task_css(tsk, i);
4426 ss->exit(css, old_css, tsk);
4432 put_css_set_taskexit(cset);
4435 static void check_for_release(struct cgroup *cgrp)
4437 if (cgroup_is_releasable(cgrp) &&
4438 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4440 * Control Group is currently removeable. If it's not
4441 * already queued for a userspace notification, queue
4444 int need_schedule_work = 0;
4446 raw_spin_lock(&release_list_lock);
4447 if (!cgroup_is_dead(cgrp) &&
4448 list_empty(&cgrp->release_list)) {
4449 list_add(&cgrp->release_list, &release_list);
4450 need_schedule_work = 1;
4452 raw_spin_unlock(&release_list_lock);
4453 if (need_schedule_work)
4454 schedule_work(&release_agent_work);
4459 * Notify userspace when a cgroup is released, by running the
4460 * configured release agent with the name of the cgroup (path
4461 * relative to the root of cgroup file system) as the argument.
4463 * Most likely, this user command will try to rmdir this cgroup.
4465 * This races with the possibility that some other task will be
4466 * attached to this cgroup before it is removed, or that some other
4467 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4468 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4469 * unused, and this cgroup will be reprieved from its death sentence,
4470 * to continue to serve a useful existence. Next time it's released,
4471 * we will get notified again, if it still has 'notify_on_release' set.
4473 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4474 * means only wait until the task is successfully execve()'d. The
4475 * separate release agent task is forked by call_usermodehelper(),
4476 * then control in this thread returns here, without waiting for the
4477 * release agent task. We don't bother to wait because the caller of
4478 * this routine has no use for the exit status of the release agent
4479 * task, so no sense holding our caller up for that.
4481 static void cgroup_release_agent(struct work_struct *work)
4483 BUG_ON(work != &release_agent_work);
4484 mutex_lock(&cgroup_mutex);
4485 raw_spin_lock(&release_list_lock);
4486 while (!list_empty(&release_list)) {
4487 char *argv[3], *envp[3];
4489 char *pathbuf = NULL, *agentbuf = NULL, *path;
4490 struct cgroup *cgrp = list_entry(release_list.next,
4493 list_del_init(&cgrp->release_list);
4494 raw_spin_unlock(&release_list_lock);
4495 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
4498 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
4501 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4506 argv[i++] = agentbuf;
4511 /* minimal command environment */
4512 envp[i++] = "HOME=/";
4513 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4516 /* Drop the lock while we invoke the usermode helper,
4517 * since the exec could involve hitting disk and hence
4518 * be a slow process */
4519 mutex_unlock(&cgroup_mutex);
4520 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4521 mutex_lock(&cgroup_mutex);
4525 raw_spin_lock(&release_list_lock);
4527 raw_spin_unlock(&release_list_lock);
4528 mutex_unlock(&cgroup_mutex);
4531 static int __init cgroup_disable(char *str)
4533 struct cgroup_subsys *ss;
4537 while ((token = strsep(&str, ",")) != NULL) {
4541 for_each_subsys(ss, i) {
4542 if (!strcmp(token, ss->name)) {
4544 printk(KERN_INFO "Disabling %s control group"
4545 " subsystem\n", ss->name);
4552 __setup("cgroup_disable=", cgroup_disable);
4555 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
4556 * @dentry: directory dentry of interest
4557 * @ss: subsystem of interest
4559 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
4560 * to get the corresponding css and return it. If such css doesn't exist
4561 * or can't be pinned, an ERR_PTR value is returned.
4563 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
4564 struct cgroup_subsys *ss)
4566 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4567 struct cgroup_subsys_state *css = NULL;
4568 struct cgroup *cgrp;
4570 /* is @dentry a cgroup dir? */
4571 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4572 kernfs_type(kn) != KERNFS_DIR)
4573 return ERR_PTR(-EBADF);
4578 * This path doesn't originate from kernfs and @kn could already
4579 * have been or be removed at any point. @kn->priv is RCU
4580 * protected for this access. See destroy_locked() for details.
4582 cgrp = rcu_dereference(kn->priv);
4584 css = cgroup_css(cgrp, ss);
4586 if (!css || !css_tryget(css))
4587 css = ERR_PTR(-ENOENT);
4594 * css_from_id - lookup css by id
4595 * @id: the cgroup id
4596 * @ss: cgroup subsys to be looked into
4598 * Returns the css if there's valid one with @id, otherwise returns NULL.
4599 * Should be called under rcu_read_lock().
4601 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
4603 struct cgroup *cgrp;
4605 cgroup_assert_mutexes_or_rcu_locked();
4607 cgrp = idr_find(&ss->root->cgroup_idr, id);
4609 return cgroup_css(cgrp, ss);
4613 #ifdef CONFIG_CGROUP_DEBUG
4614 static struct cgroup_subsys_state *
4615 debug_css_alloc(struct cgroup_subsys_state *parent_css)
4617 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4620 return ERR_PTR(-ENOMEM);
4625 static void debug_css_free(struct cgroup_subsys_state *css)
4630 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
4633 return cgroup_task_count(css->cgroup);
4636 static u64 current_css_set_read(struct cgroup_subsys_state *css,
4639 return (u64)(unsigned long)current->cgroups;
4642 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
4648 count = atomic_read(&task_css_set(current)->refcount);
4653 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
4655 struct cgrp_cset_link *link;
4656 struct css_set *cset;
4659 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
4663 down_read(&css_set_rwsem);
4665 cset = rcu_dereference(current->cgroups);
4666 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
4667 struct cgroup *c = link->cgrp;
4668 const char *name = "?";
4670 if (c != cgroup_dummy_top) {
4671 cgroup_name(c, name_buf, NAME_MAX + 1);
4675 seq_printf(seq, "Root %d group %s\n",
4676 c->root->hierarchy_id, name);
4679 up_read(&css_set_rwsem);
4684 #define MAX_TASKS_SHOWN_PER_CSS 25
4685 static int cgroup_css_links_read(struct seq_file *seq, void *v)
4687 struct cgroup_subsys_state *css = seq_css(seq);
4688 struct cgrp_cset_link *link;
4690 down_read(&css_set_rwsem);
4691 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
4692 struct css_set *cset = link->cset;
4693 struct task_struct *task;
4695 seq_printf(seq, "css_set %p\n", cset);
4696 list_for_each_entry(task, &cset->tasks, cg_list) {
4697 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4698 seq_puts(seq, " ...\n");
4701 seq_printf(seq, " task %d\n",
4702 task_pid_vnr(task));
4706 up_read(&css_set_rwsem);
4710 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
4712 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4715 static struct cftype debug_files[] = {
4717 .name = "taskcount",
4718 .read_u64 = debug_taskcount_read,
4722 .name = "current_css_set",
4723 .read_u64 = current_css_set_read,
4727 .name = "current_css_set_refcount",
4728 .read_u64 = current_css_set_refcount_read,
4732 .name = "current_css_set_cg_links",
4733 .seq_show = current_css_set_cg_links_read,
4737 .name = "cgroup_css_links",
4738 .seq_show = cgroup_css_links_read,
4742 .name = "releasable",
4743 .read_u64 = releasable_read,
4749 struct cgroup_subsys debug_cgrp_subsys = {
4750 .css_alloc = debug_css_alloc,
4751 .css_free = debug_css_free,
4752 .base_cftypes = debug_files,
4754 #endif /* CONFIG_CGROUP_DEBUG */