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/kthread.h>
56 #include <linux/delay.h>
58 #include <linux/atomic.h>
61 * pidlists linger the following amount before being destroyed. The goal
62 * is avoiding frequent destruction in the middle of consecutive read calls
63 * Expiring in the middle is a performance problem not a correctness one.
64 * 1 sec should be enough.
66 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
68 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
72 * cgroup_tree_mutex nests above cgroup_mutex and protects cftypes, file
73 * creation/removal and hierarchy changing operations including cgroup
74 * creation, removal, css association and controller rebinding. This outer
75 * lock is needed mainly to resolve the circular dependency between kernfs
76 * active ref and cgroup_mutex. cgroup_tree_mutex nests above both.
78 static DEFINE_MUTEX(cgroup_tree_mutex);
81 * cgroup_mutex is the master lock. Any modification to cgroup or its
82 * hierarchy must be performed while holding it.
84 * css_set_rwsem protects task->cgroups pointer, the list of css_set
85 * objects, and the chain of tasks off each css_set.
87 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
88 * cgroup.h can use them for lockdep annotations.
90 #ifdef CONFIG_PROVE_RCU
91 DEFINE_MUTEX(cgroup_mutex);
92 DECLARE_RWSEM(css_set_rwsem);
93 EXPORT_SYMBOL_GPL(cgroup_mutex);
94 EXPORT_SYMBOL_GPL(css_set_rwsem);
96 static DEFINE_MUTEX(cgroup_mutex);
97 static DECLARE_RWSEM(css_set_rwsem);
101 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
102 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
104 static DEFINE_SPINLOCK(release_agent_path_lock);
106 #define cgroup_assert_mutexes_or_rcu_locked() \
107 rcu_lockdep_assert(rcu_read_lock_held() || \
108 lockdep_is_held(&cgroup_tree_mutex) || \
109 lockdep_is_held(&cgroup_mutex), \
110 "cgroup_[tree_]mutex or RCU read lock required");
113 * cgroup destruction makes heavy use of work items and there can be a lot
114 * of concurrent destructions. Use a separate workqueue so that cgroup
115 * destruction work items don't end up filling up max_active of system_wq
116 * which may lead to deadlock.
118 static struct workqueue_struct *cgroup_destroy_wq;
121 * pidlist destructions need to be flushed on cgroup destruction. Use a
122 * separate workqueue as flush domain.
124 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
126 /* generate an array of cgroup subsystem pointers */
127 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
128 static struct cgroup_subsys *cgroup_subsys[] = {
129 #include <linux/cgroup_subsys.h>
133 /* array of cgroup subsystem names */
134 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
135 static const char *cgroup_subsys_name[] = {
136 #include <linux/cgroup_subsys.h>
141 * The default hierarchy, reserved for the subsystems that are otherwise
142 * unattached - it never has more than a single cgroup, and all tasks are
143 * part of that cgroup.
145 struct cgroup_root cgrp_dfl_root;
148 * The default hierarchy always exists but is hidden until mounted for the
149 * first time. This is for backward compatibility.
151 static bool cgrp_dfl_root_visible;
153 /* The list of hierarchy roots */
155 static LIST_HEAD(cgroup_roots);
156 static int cgroup_root_count;
158 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
159 static DEFINE_IDR(cgroup_hierarchy_idr);
162 * Assign a monotonically increasing serial number to cgroups. It
163 * guarantees cgroups with bigger numbers are newer than those with smaller
164 * numbers. Also, as cgroups are always appended to the parent's
165 * ->children list, it guarantees that sibling cgroups are always sorted in
166 * the ascending serial number order on the list. Protected by
169 static u64 cgroup_serial_nr_next = 1;
171 /* This flag indicates whether tasks in the fork and exit paths should
172 * check for fork/exit handlers to call. This avoids us having to do
173 * extra work in the fork/exit path if none of the subsystems need to
176 static int need_forkexit_callback __read_mostly;
178 static struct cftype cgroup_base_files[];
180 static void cgroup_put(struct cgroup *cgrp);
181 static int rebind_subsystems(struct cgroup_root *dst_root,
182 unsigned long ss_mask);
183 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
184 static int cgroup_destroy_locked(struct cgroup *cgrp);
185 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
187 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
190 * cgroup_css - obtain a cgroup's css for the specified subsystem
191 * @cgrp: the cgroup of interest
192 * @ss: the subsystem of interest (%NULL returns the dummy_css)
194 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
195 * function must be called either under cgroup_mutex or rcu_read_lock() and
196 * the caller is responsible for pinning the returned css if it wants to
197 * keep accessing it outside the said locks. This function may return
198 * %NULL if @cgrp doesn't have @subsys_id enabled.
200 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
201 struct cgroup_subsys *ss)
204 return rcu_dereference_check(cgrp->subsys[ss->id],
205 lockdep_is_held(&cgroup_tree_mutex) ||
206 lockdep_is_held(&cgroup_mutex));
208 return &cgrp->dummy_css;
212 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
213 * @cgrp: the cgroup of interest
214 * @ss: the subsystem of interest (%NULL returns the dummy_css)
216 * Similar to cgroup_css() but returns the effctive css, which is defined
217 * as the matching css of the nearest ancestor including self which has @ss
218 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
219 * function is guaranteed to return non-NULL css.
221 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
222 struct cgroup_subsys *ss)
224 lockdep_assert_held(&cgroup_mutex);
227 return &cgrp->dummy_css;
229 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
232 while (cgrp->parent &&
233 !(cgrp->parent->child_subsys_mask & (1 << ss->id)))
236 return cgroup_css(cgrp, ss);
239 /* convenient tests for these bits */
240 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
242 return test_bit(CGRP_DEAD, &cgrp->flags);
245 struct cgroup_subsys_state *seq_css(struct seq_file *seq)
247 struct kernfs_open_file *of = seq->private;
248 struct cgroup *cgrp = of->kn->parent->priv;
249 struct cftype *cft = seq_cft(seq);
252 * This is open and unprotected implementation of cgroup_css().
253 * seq_css() is only called from a kernfs file operation which has
254 * an active reference on the file. Because all the subsystem
255 * files are drained before a css is disassociated with a cgroup,
256 * the matching css from the cgroup's subsys table is guaranteed to
257 * be and stay valid until the enclosing operation is complete.
260 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
262 return &cgrp->dummy_css;
264 EXPORT_SYMBOL_GPL(seq_css);
267 * cgroup_is_descendant - test ancestry
268 * @cgrp: the cgroup to be tested
269 * @ancestor: possible ancestor of @cgrp
271 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
272 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
273 * and @ancestor are accessible.
275 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
278 if (cgrp == ancestor)
285 static int cgroup_is_releasable(const struct cgroup *cgrp)
288 (1 << CGRP_RELEASABLE) |
289 (1 << CGRP_NOTIFY_ON_RELEASE);
290 return (cgrp->flags & bits) == bits;
293 static int notify_on_release(const struct cgroup *cgrp)
295 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
299 * for_each_css - iterate all css's of a cgroup
300 * @css: the iteration cursor
301 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
302 * @cgrp: the target cgroup to iterate css's of
304 * Should be called under cgroup_[tree_]mutex.
306 #define for_each_css(css, ssid, cgrp) \
307 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
308 if (!((css) = rcu_dereference_check( \
309 (cgrp)->subsys[(ssid)], \
310 lockdep_is_held(&cgroup_tree_mutex) || \
311 lockdep_is_held(&cgroup_mutex)))) { } \
315 * for_each_e_css - iterate all effective css's of a cgroup
316 * @css: the iteration cursor
317 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
318 * @cgrp: the target cgroup to iterate css's of
320 * Should be called under cgroup_[tree_]mutex.
322 #define for_each_e_css(css, ssid, cgrp) \
323 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
324 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
329 * for_each_subsys - iterate all enabled cgroup subsystems
330 * @ss: the iteration cursor
331 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
333 #define for_each_subsys(ss, ssid) \
334 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
335 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
337 /* iterate across the hierarchies */
338 #define for_each_root(root) \
339 list_for_each_entry((root), &cgroup_roots, root_list)
342 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
343 * @cgrp: the cgroup to be checked for liveness
345 * On success, returns true; the mutex should be later unlocked. On
346 * failure returns false with no lock held.
348 static bool cgroup_lock_live_group(struct cgroup *cgrp)
350 mutex_lock(&cgroup_mutex);
351 if (cgroup_is_dead(cgrp)) {
352 mutex_unlock(&cgroup_mutex);
358 /* the list of cgroups eligible for automatic release. Protected by
359 * release_list_lock */
360 static LIST_HEAD(release_list);
361 static DEFINE_RAW_SPINLOCK(release_list_lock);
362 static void cgroup_release_agent(struct work_struct *work);
363 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
364 static void check_for_release(struct cgroup *cgrp);
367 * A cgroup can be associated with multiple css_sets as different tasks may
368 * belong to different cgroups on different hierarchies. In the other
369 * direction, a css_set is naturally associated with multiple cgroups.
370 * This M:N relationship is represented by the following link structure
371 * which exists for each association and allows traversing the associations
374 struct cgrp_cset_link {
375 /* the cgroup and css_set this link associates */
377 struct css_set *cset;
379 /* list of cgrp_cset_links anchored at cgrp->cset_links */
380 struct list_head cset_link;
382 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
383 struct list_head cgrp_link;
387 * The default css_set - used by init and its children prior to any
388 * hierarchies being mounted. It contains a pointer to the root state
389 * for each subsystem. Also used to anchor the list of css_sets. Not
390 * reference-counted, to improve performance when child cgroups
391 * haven't been created.
393 static struct css_set init_css_set = {
394 .refcount = ATOMIC_INIT(1),
395 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
396 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
397 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
398 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
399 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
402 static int css_set_count = 1; /* 1 for init_css_set */
405 * hash table for cgroup groups. This improves the performance to find
406 * an existing css_set. This hash doesn't (currently) take into
407 * account cgroups in empty hierarchies.
409 #define CSS_SET_HASH_BITS 7
410 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
412 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
414 unsigned long key = 0UL;
415 struct cgroup_subsys *ss;
418 for_each_subsys(ss, i)
419 key += (unsigned long)css[i];
420 key = (key >> 16) ^ key;
425 static void put_css_set_locked(struct css_set *cset, bool taskexit)
427 struct cgrp_cset_link *link, *tmp_link;
428 struct cgroup_subsys *ss;
431 lockdep_assert_held(&css_set_rwsem);
433 if (!atomic_dec_and_test(&cset->refcount))
436 /* This css_set is dead. unlink it and release cgroup refcounts */
437 for_each_subsys(ss, ssid)
438 list_del(&cset->e_cset_node[ssid]);
439 hash_del(&cset->hlist);
442 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
443 struct cgroup *cgrp = link->cgrp;
445 list_del(&link->cset_link);
446 list_del(&link->cgrp_link);
448 /* @cgrp can't go away while we're holding css_set_rwsem */
449 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
451 set_bit(CGRP_RELEASABLE, &cgrp->flags);
452 check_for_release(cgrp);
458 kfree_rcu(cset, rcu_head);
461 static void put_css_set(struct css_set *cset, bool taskexit)
464 * Ensure that the refcount doesn't hit zero while any readers
465 * can see it. Similar to atomic_dec_and_lock(), but for an
468 if (atomic_add_unless(&cset->refcount, -1, 1))
471 down_write(&css_set_rwsem);
472 put_css_set_locked(cset, taskexit);
473 up_write(&css_set_rwsem);
477 * refcounted get/put for css_set objects
479 static inline void get_css_set(struct css_set *cset)
481 atomic_inc(&cset->refcount);
485 * compare_css_sets - helper function for find_existing_css_set().
486 * @cset: candidate css_set being tested
487 * @old_cset: existing css_set for a task
488 * @new_cgrp: cgroup that's being entered by the task
489 * @template: desired set of css pointers in css_set (pre-calculated)
491 * Returns true if "cset" matches "old_cset" except for the hierarchy
492 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
494 static bool compare_css_sets(struct css_set *cset,
495 struct css_set *old_cset,
496 struct cgroup *new_cgrp,
497 struct cgroup_subsys_state *template[])
499 struct list_head *l1, *l2;
502 * On the default hierarchy, there can be csets which are
503 * associated with the same set of cgroups but different csses.
504 * Let's first ensure that csses match.
506 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
510 * Compare cgroup pointers in order to distinguish between
511 * different cgroups in hierarchies. As different cgroups may
512 * share the same effective css, this comparison is always
515 l1 = &cset->cgrp_links;
516 l2 = &old_cset->cgrp_links;
518 struct cgrp_cset_link *link1, *link2;
519 struct cgroup *cgrp1, *cgrp2;
523 /* See if we reached the end - both lists are equal length. */
524 if (l1 == &cset->cgrp_links) {
525 BUG_ON(l2 != &old_cset->cgrp_links);
528 BUG_ON(l2 == &old_cset->cgrp_links);
530 /* Locate the cgroups associated with these links. */
531 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
532 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
535 /* Hierarchies should be linked in the same order. */
536 BUG_ON(cgrp1->root != cgrp2->root);
539 * If this hierarchy is the hierarchy of the cgroup
540 * that's changing, then we need to check that this
541 * css_set points to the new cgroup; if it's any other
542 * hierarchy, then this css_set should point to the
543 * same cgroup as the old css_set.
545 if (cgrp1->root == new_cgrp->root) {
546 if (cgrp1 != new_cgrp)
557 * find_existing_css_set - init css array and find the matching css_set
558 * @old_cset: the css_set that we're using before the cgroup transition
559 * @cgrp: the cgroup that we're moving into
560 * @template: out param for the new set of csses, should be clear on entry
562 static struct css_set *find_existing_css_set(struct css_set *old_cset,
564 struct cgroup_subsys_state *template[])
566 struct cgroup_root *root = cgrp->root;
567 struct cgroup_subsys *ss;
568 struct css_set *cset;
573 * Build the set of subsystem state objects that we want to see in the
574 * new css_set. while subsystems can change globally, the entries here
575 * won't change, so no need for locking.
577 for_each_subsys(ss, i) {
578 if (root->subsys_mask & (1UL << i)) {
580 * @ss is in this hierarchy, so we want the
581 * effective css from @cgrp.
583 template[i] = cgroup_e_css(cgrp, ss);
586 * @ss is not in this hierarchy, so we don't want
589 template[i] = old_cset->subsys[i];
593 key = css_set_hash(template);
594 hash_for_each_possible(css_set_table, cset, hlist, key) {
595 if (!compare_css_sets(cset, old_cset, cgrp, template))
598 /* This css_set matches what we need */
602 /* No existing cgroup group matched */
606 static void free_cgrp_cset_links(struct list_head *links_to_free)
608 struct cgrp_cset_link *link, *tmp_link;
610 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
611 list_del(&link->cset_link);
617 * allocate_cgrp_cset_links - allocate cgrp_cset_links
618 * @count: the number of links to allocate
619 * @tmp_links: list_head the allocated links are put on
621 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
622 * through ->cset_link. Returns 0 on success or -errno.
624 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
626 struct cgrp_cset_link *link;
629 INIT_LIST_HEAD(tmp_links);
631 for (i = 0; i < count; i++) {
632 link = kzalloc(sizeof(*link), GFP_KERNEL);
634 free_cgrp_cset_links(tmp_links);
637 list_add(&link->cset_link, tmp_links);
643 * link_css_set - a helper function to link a css_set to a cgroup
644 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
645 * @cset: the css_set to be linked
646 * @cgrp: the destination cgroup
648 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
651 struct cgrp_cset_link *link;
653 BUG_ON(list_empty(tmp_links));
654 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
657 list_move(&link->cset_link, &cgrp->cset_links);
659 * Always add links to the tail of the list so that the list
660 * is sorted by order of hierarchy creation
662 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
666 * find_css_set - return a new css_set with one cgroup updated
667 * @old_cset: the baseline css_set
668 * @cgrp: the cgroup to be updated
670 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
671 * substituted into the appropriate hierarchy.
673 static struct css_set *find_css_set(struct css_set *old_cset,
676 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
677 struct css_set *cset;
678 struct list_head tmp_links;
679 struct cgrp_cset_link *link;
680 struct cgroup_subsys *ss;
684 lockdep_assert_held(&cgroup_mutex);
686 /* First see if we already have a cgroup group that matches
688 down_read(&css_set_rwsem);
689 cset = find_existing_css_set(old_cset, cgrp, template);
692 up_read(&css_set_rwsem);
697 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
701 /* Allocate all the cgrp_cset_link objects that we'll need */
702 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
707 atomic_set(&cset->refcount, 1);
708 INIT_LIST_HEAD(&cset->cgrp_links);
709 INIT_LIST_HEAD(&cset->tasks);
710 INIT_LIST_HEAD(&cset->mg_tasks);
711 INIT_LIST_HEAD(&cset->mg_preload_node);
712 INIT_LIST_HEAD(&cset->mg_node);
713 INIT_HLIST_NODE(&cset->hlist);
715 /* Copy the set of subsystem state objects generated in
716 * find_existing_css_set() */
717 memcpy(cset->subsys, template, sizeof(cset->subsys));
719 down_write(&css_set_rwsem);
720 /* Add reference counts and links from the new css_set. */
721 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
722 struct cgroup *c = link->cgrp;
724 if (c->root == cgrp->root)
726 link_css_set(&tmp_links, cset, c);
729 BUG_ON(!list_empty(&tmp_links));
733 /* Add @cset to the hash table */
734 key = css_set_hash(cset->subsys);
735 hash_add(css_set_table, &cset->hlist, key);
737 for_each_subsys(ss, ssid)
738 list_add_tail(&cset->e_cset_node[ssid],
739 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
741 up_write(&css_set_rwsem);
746 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
748 struct cgroup *root_cgrp = kf_root->kn->priv;
750 return root_cgrp->root;
753 static int cgroup_init_root_id(struct cgroup_root *root)
757 lockdep_assert_held(&cgroup_mutex);
759 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
763 root->hierarchy_id = id;
767 static void cgroup_exit_root_id(struct cgroup_root *root)
769 lockdep_assert_held(&cgroup_mutex);
771 if (root->hierarchy_id) {
772 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
773 root->hierarchy_id = 0;
777 static void cgroup_free_root(struct cgroup_root *root)
780 /* hierarhcy ID shoulid already have been released */
781 WARN_ON_ONCE(root->hierarchy_id);
783 idr_destroy(&root->cgroup_idr);
788 static void cgroup_destroy_root(struct cgroup_root *root)
790 struct cgroup *cgrp = &root->cgrp;
791 struct cgrp_cset_link *link, *tmp_link;
793 mutex_lock(&cgroup_tree_mutex);
794 mutex_lock(&cgroup_mutex);
796 BUG_ON(atomic_read(&root->nr_cgrps));
797 BUG_ON(!list_empty(&cgrp->children));
799 /* Rebind all subsystems back to the default hierarchy */
800 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
803 * Release all the links from cset_links to this hierarchy's
806 down_write(&css_set_rwsem);
808 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
809 list_del(&link->cset_link);
810 list_del(&link->cgrp_link);
813 up_write(&css_set_rwsem);
815 if (!list_empty(&root->root_list)) {
816 list_del(&root->root_list);
820 cgroup_exit_root_id(root);
822 mutex_unlock(&cgroup_mutex);
823 mutex_unlock(&cgroup_tree_mutex);
825 kernfs_destroy_root(root->kf_root);
826 cgroup_free_root(root);
829 /* look up cgroup associated with given css_set on the specified hierarchy */
830 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
831 struct cgroup_root *root)
833 struct cgroup *res = NULL;
835 lockdep_assert_held(&cgroup_mutex);
836 lockdep_assert_held(&css_set_rwsem);
838 if (cset == &init_css_set) {
841 struct cgrp_cset_link *link;
843 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
844 struct cgroup *c = link->cgrp;
846 if (c->root == root) {
858 * Return the cgroup for "task" from the given hierarchy. Must be
859 * called with cgroup_mutex and css_set_rwsem held.
861 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
862 struct cgroup_root *root)
865 * No need to lock the task - since we hold cgroup_mutex the
866 * task can't change groups, so the only thing that can happen
867 * is that it exits and its css is set back to init_css_set.
869 return cset_cgroup_from_root(task_css_set(task), root);
873 * A task must hold cgroup_mutex to modify cgroups.
875 * Any task can increment and decrement the count field without lock.
876 * So in general, code holding cgroup_mutex can't rely on the count
877 * field not changing. However, if the count goes to zero, then only
878 * cgroup_attach_task() can increment it again. Because a count of zero
879 * means that no tasks are currently attached, therefore there is no
880 * way a task attached to that cgroup can fork (the other way to
881 * increment the count). So code holding cgroup_mutex can safely
882 * assume that if the count is zero, it will stay zero. Similarly, if
883 * a task holds cgroup_mutex on a cgroup with zero count, it
884 * knows that the cgroup won't be removed, as cgroup_rmdir()
887 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
888 * (usually) take cgroup_mutex. These are the two most performance
889 * critical pieces of code here. The exception occurs on cgroup_exit(),
890 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
891 * is taken, and if the cgroup count is zero, a usermode call made
892 * to the release agent with the name of the cgroup (path relative to
893 * the root of cgroup file system) as the argument.
895 * A cgroup can only be deleted if both its 'count' of using tasks
896 * is zero, and its list of 'children' cgroups is empty. Since all
897 * tasks in the system use _some_ cgroup, and since there is always at
898 * least one task in the system (init, pid == 1), therefore, root cgroup
899 * always has either children cgroups and/or using tasks. So we don't
900 * need a special hack to ensure that root cgroup cannot be deleted.
902 * P.S. One more locking exception. RCU is used to guard the
903 * update of a tasks cgroup pointer by cgroup_attach_task()
906 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
907 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
908 static const struct file_operations proc_cgroupstats_operations;
910 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
913 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
914 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
915 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
916 cft->ss->name, cft->name);
918 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
923 * cgroup_file_mode - deduce file mode of a control file
924 * @cft: the control file in question
926 * returns cft->mode if ->mode is not 0
927 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
928 * returns S_IRUGO if it has only a read handler
929 * returns S_IWUSR if it has only a write hander
931 static umode_t cgroup_file_mode(const struct cftype *cft)
938 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
941 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
948 static void cgroup_free_fn(struct work_struct *work)
950 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
952 atomic_dec(&cgrp->root->nr_cgrps);
953 cgroup_pidlist_destroy_all(cgrp);
957 * We get a ref to the parent, and put the ref when this
958 * cgroup is being freed, so it's guaranteed that the
959 * parent won't be destroyed before its children.
961 cgroup_put(cgrp->parent);
962 kernfs_put(cgrp->kn);
966 * This is root cgroup's refcnt reaching zero, which
967 * indicates that the root should be released.
969 cgroup_destroy_root(cgrp->root);
973 static void cgroup_free_rcu(struct rcu_head *head)
975 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
977 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
978 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
981 static void cgroup_get(struct cgroup *cgrp)
983 WARN_ON_ONCE(cgroup_is_dead(cgrp));
984 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
985 atomic_inc(&cgrp->refcnt);
988 static void cgroup_put(struct cgroup *cgrp)
990 if (!atomic_dec_and_test(&cgrp->refcnt))
992 if (WARN_ON_ONCE(cgrp->parent && !cgroup_is_dead(cgrp)))
996 * XXX: cgrp->id is only used to look up css's. As cgroup and
997 * css's lifetimes will be decoupled, it should be made
998 * per-subsystem and moved to css->id so that lookups are
999 * successful until the target css is released.
1001 mutex_lock(&cgroup_mutex);
1002 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
1003 mutex_unlock(&cgroup_mutex);
1006 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
1009 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1011 char name[CGROUP_FILE_NAME_MAX];
1013 lockdep_assert_held(&cgroup_tree_mutex);
1014 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1018 * cgroup_clear_dir - remove subsys files in a cgroup directory
1019 * @cgrp: target cgroup
1020 * @subsys_mask: mask of the subsystem ids whose files should be removed
1022 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
1024 struct cgroup_subsys *ss;
1027 for_each_subsys(ss, i) {
1028 struct cftype *cfts;
1030 if (!test_bit(i, &subsys_mask))
1032 list_for_each_entry(cfts, &ss->cfts, node)
1033 cgroup_addrm_files(cgrp, cfts, false);
1037 static int rebind_subsystems(struct cgroup_root *dst_root,
1038 unsigned long ss_mask)
1040 struct cgroup_subsys *ss;
1043 lockdep_assert_held(&cgroup_tree_mutex);
1044 lockdep_assert_held(&cgroup_mutex);
1046 for_each_subsys(ss, ssid) {
1047 if (!(ss_mask & (1 << ssid)))
1050 /* if @ss is on the dummy_root, we can always move it */
1051 if (ss->root == &cgrp_dfl_root)
1054 /* if @ss has non-root cgroups attached to it, can't move */
1055 if (!list_empty(&ss->root->cgrp.children))
1058 /* can't move between two non-dummy roots either */
1059 if (dst_root != &cgrp_dfl_root)
1063 ret = cgroup_populate_dir(&dst_root->cgrp, ss_mask);
1065 if (dst_root != &cgrp_dfl_root)
1069 * Rebinding back to the default root is not allowed to
1070 * fail. Using both default and non-default roots should
1071 * be rare. Moving subsystems back and forth even more so.
1072 * Just warn about it and continue.
1074 if (cgrp_dfl_root_visible) {
1075 pr_warning("cgroup: failed to create files (%d) while rebinding 0x%lx to default root\n",
1077 pr_warning("cgroup: you may retry by moving them to a different hierarchy and unbinding\n");
1082 * Nothing can fail from this point on. Remove files for the
1083 * removed subsystems and rebind each subsystem.
1085 mutex_unlock(&cgroup_mutex);
1086 for_each_subsys(ss, ssid)
1087 if (ss_mask & (1 << ssid))
1088 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1089 mutex_lock(&cgroup_mutex);
1091 for_each_subsys(ss, ssid) {
1092 struct cgroup_root *src_root;
1093 struct cgroup_subsys_state *css;
1094 struct css_set *cset;
1096 if (!(ss_mask & (1 << ssid)))
1099 src_root = ss->root;
1100 css = cgroup_css(&src_root->cgrp, ss);
1102 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1104 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1105 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1106 ss->root = dst_root;
1107 css->cgroup = &dst_root->cgrp;
1109 down_write(&css_set_rwsem);
1110 hash_for_each(css_set_table, i, cset, hlist)
1111 list_move_tail(&cset->e_cset_node[ss->id],
1112 &dst_root->cgrp.e_csets[ss->id]);
1113 up_write(&css_set_rwsem);
1115 src_root->subsys_mask &= ~(1 << ssid);
1116 src_root->cgrp.child_subsys_mask &= ~(1 << ssid);
1118 dst_root->subsys_mask |= 1 << ssid;
1119 dst_root->cgrp.child_subsys_mask |= 1 << ssid;
1125 kernfs_activate(dst_root->cgrp.kn);
1129 static int cgroup_show_options(struct seq_file *seq,
1130 struct kernfs_root *kf_root)
1132 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1133 struct cgroup_subsys *ss;
1136 for_each_subsys(ss, ssid)
1137 if (root->subsys_mask & (1 << ssid))
1138 seq_printf(seq, ",%s", ss->name);
1139 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1140 seq_puts(seq, ",sane_behavior");
1141 if (root->flags & CGRP_ROOT_NOPREFIX)
1142 seq_puts(seq, ",noprefix");
1143 if (root->flags & CGRP_ROOT_XATTR)
1144 seq_puts(seq, ",xattr");
1146 spin_lock(&release_agent_path_lock);
1147 if (strlen(root->release_agent_path))
1148 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1149 spin_unlock(&release_agent_path_lock);
1151 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1152 seq_puts(seq, ",clone_children");
1153 if (strlen(root->name))
1154 seq_printf(seq, ",name=%s", root->name);
1158 struct cgroup_sb_opts {
1159 unsigned long subsys_mask;
1160 unsigned long flags;
1161 char *release_agent;
1162 bool cpuset_clone_children;
1164 /* User explicitly requested empty subsystem */
1169 * Convert a hierarchy specifier into a bitmask of subsystems and
1170 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1171 * array. This function takes refcounts on subsystems to be used, unless it
1172 * returns error, in which case no refcounts are taken.
1174 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1176 char *token, *o = data;
1177 bool all_ss = false, one_ss = false;
1178 unsigned long mask = (unsigned long)-1;
1179 struct cgroup_subsys *ss;
1182 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1184 #ifdef CONFIG_CPUSETS
1185 mask = ~(1UL << cpuset_cgrp_id);
1188 memset(opts, 0, sizeof(*opts));
1190 while ((token = strsep(&o, ",")) != NULL) {
1193 if (!strcmp(token, "none")) {
1194 /* Explicitly have no subsystems */
1198 if (!strcmp(token, "all")) {
1199 /* Mutually exclusive option 'all' + subsystem name */
1205 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1206 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1209 if (!strcmp(token, "noprefix")) {
1210 opts->flags |= CGRP_ROOT_NOPREFIX;
1213 if (!strcmp(token, "clone_children")) {
1214 opts->cpuset_clone_children = true;
1217 if (!strcmp(token, "xattr")) {
1218 opts->flags |= CGRP_ROOT_XATTR;
1221 if (!strncmp(token, "release_agent=", 14)) {
1222 /* Specifying two release agents is forbidden */
1223 if (opts->release_agent)
1225 opts->release_agent =
1226 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1227 if (!opts->release_agent)
1231 if (!strncmp(token, "name=", 5)) {
1232 const char *name = token + 5;
1233 /* Can't specify an empty name */
1236 /* Must match [\w.-]+ */
1237 for (i = 0; i < strlen(name); i++) {
1241 if ((c == '.') || (c == '-') || (c == '_'))
1245 /* Specifying two names is forbidden */
1248 opts->name = kstrndup(name,
1249 MAX_CGROUP_ROOT_NAMELEN - 1,
1257 for_each_subsys(ss, i) {
1258 if (strcmp(token, ss->name))
1263 /* Mutually exclusive option 'all' + subsystem name */
1266 set_bit(i, &opts->subsys_mask);
1271 if (i == CGROUP_SUBSYS_COUNT)
1275 /* Consistency checks */
1277 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1278 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1280 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1281 opts->cpuset_clone_children || opts->release_agent ||
1283 pr_err("cgroup: sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1288 * If the 'all' option was specified select all the
1289 * subsystems, otherwise if 'none', 'name=' and a subsystem
1290 * name options were not specified, let's default to 'all'
1292 if (all_ss || (!one_ss && !opts->none && !opts->name))
1293 for_each_subsys(ss, i)
1295 set_bit(i, &opts->subsys_mask);
1298 * We either have to specify by name or by subsystems. (So
1299 * all empty hierarchies must have a name).
1301 if (!opts->subsys_mask && !opts->name)
1306 * Option noprefix was introduced just for backward compatibility
1307 * with the old cpuset, so we allow noprefix only if mounting just
1308 * the cpuset subsystem.
1310 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1314 /* Can't specify "none" and some subsystems */
1315 if (opts->subsys_mask && opts->none)
1321 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1324 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1325 struct cgroup_sb_opts opts;
1326 unsigned long added_mask, removed_mask;
1328 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1329 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1333 mutex_lock(&cgroup_tree_mutex);
1334 mutex_lock(&cgroup_mutex);
1336 /* See what subsystems are wanted */
1337 ret = parse_cgroupfs_options(data, &opts);
1341 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1342 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1343 task_tgid_nr(current), current->comm);
1345 added_mask = opts.subsys_mask & ~root->subsys_mask;
1346 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1348 /* Don't allow flags or name to change at remount */
1349 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1350 (opts.name && strcmp(opts.name, root->name))) {
1351 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1352 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1353 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1358 /* remounting is not allowed for populated hierarchies */
1359 if (!list_empty(&root->cgrp.children)) {
1364 ret = rebind_subsystems(root, added_mask);
1368 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1370 if (opts.release_agent) {
1371 spin_lock(&release_agent_path_lock);
1372 strcpy(root->release_agent_path, opts.release_agent);
1373 spin_unlock(&release_agent_path_lock);
1376 kfree(opts.release_agent);
1378 mutex_unlock(&cgroup_mutex);
1379 mutex_unlock(&cgroup_tree_mutex);
1384 * To reduce the fork() overhead for systems that are not actually using
1385 * their cgroups capability, we don't maintain the lists running through
1386 * each css_set to its tasks until we see the list actually used - in other
1387 * words after the first mount.
1389 static bool use_task_css_set_links __read_mostly;
1391 static void cgroup_enable_task_cg_lists(void)
1393 struct task_struct *p, *g;
1395 down_write(&css_set_rwsem);
1397 if (use_task_css_set_links)
1400 use_task_css_set_links = true;
1403 * We need tasklist_lock because RCU is not safe against
1404 * while_each_thread(). Besides, a forking task that has passed
1405 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1406 * is not guaranteed to have its child immediately visible in the
1407 * tasklist if we walk through it with RCU.
1409 read_lock(&tasklist_lock);
1410 do_each_thread(g, p) {
1411 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1412 task_css_set(p) != &init_css_set);
1415 * We should check if the process is exiting, otherwise
1416 * it will race with cgroup_exit() in that the list
1417 * entry won't be deleted though the process has exited.
1418 * Do it while holding siglock so that we don't end up
1419 * racing against cgroup_exit().
1421 spin_lock_irq(&p->sighand->siglock);
1422 if (!(p->flags & PF_EXITING)) {
1423 struct css_set *cset = task_css_set(p);
1425 list_add(&p->cg_list, &cset->tasks);
1428 spin_unlock_irq(&p->sighand->siglock);
1429 } while_each_thread(g, p);
1430 read_unlock(&tasklist_lock);
1432 up_write(&css_set_rwsem);
1435 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1437 struct cgroup_subsys *ss;
1440 atomic_set(&cgrp->refcnt, 1);
1441 INIT_LIST_HEAD(&cgrp->sibling);
1442 INIT_LIST_HEAD(&cgrp->children);
1443 INIT_LIST_HEAD(&cgrp->cset_links);
1444 INIT_LIST_HEAD(&cgrp->release_list);
1445 INIT_LIST_HEAD(&cgrp->pidlists);
1446 mutex_init(&cgrp->pidlist_mutex);
1447 cgrp->dummy_css.cgroup = cgrp;
1449 for_each_subsys(ss, ssid)
1450 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1453 static void init_cgroup_root(struct cgroup_root *root,
1454 struct cgroup_sb_opts *opts)
1456 struct cgroup *cgrp = &root->cgrp;
1458 INIT_LIST_HEAD(&root->root_list);
1459 atomic_set(&root->nr_cgrps, 1);
1461 init_cgroup_housekeeping(cgrp);
1462 idr_init(&root->cgroup_idr);
1464 root->flags = opts->flags;
1465 if (opts->release_agent)
1466 strcpy(root->release_agent_path, opts->release_agent);
1468 strcpy(root->name, opts->name);
1469 if (opts->cpuset_clone_children)
1470 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1473 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1475 LIST_HEAD(tmp_links);
1476 struct cgroup *root_cgrp = &root->cgrp;
1477 struct css_set *cset;
1480 lockdep_assert_held(&cgroup_tree_mutex);
1481 lockdep_assert_held(&cgroup_mutex);
1483 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1486 root_cgrp->id = ret;
1489 * We're accessing css_set_count without locking css_set_rwsem here,
1490 * but that's OK - it can only be increased by someone holding
1491 * cgroup_lock, and that's us. The worst that can happen is that we
1492 * have some link structures left over
1494 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1498 ret = cgroup_init_root_id(root);
1502 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1503 KERNFS_ROOT_CREATE_DEACTIVATED,
1505 if (IS_ERR(root->kf_root)) {
1506 ret = PTR_ERR(root->kf_root);
1509 root_cgrp->kn = root->kf_root->kn;
1511 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1515 ret = rebind_subsystems(root, ss_mask);
1520 * There must be no failure case after here, since rebinding takes
1521 * care of subsystems' refcounts, which are explicitly dropped in
1522 * the failure exit path.
1524 list_add(&root->root_list, &cgroup_roots);
1525 cgroup_root_count++;
1528 * Link the root cgroup in this hierarchy into all the css_set
1531 down_write(&css_set_rwsem);
1532 hash_for_each(css_set_table, i, cset, hlist)
1533 link_css_set(&tmp_links, cset, root_cgrp);
1534 up_write(&css_set_rwsem);
1536 BUG_ON(!list_empty(&root_cgrp->children));
1537 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1539 kernfs_activate(root_cgrp->kn);
1544 kernfs_destroy_root(root->kf_root);
1545 root->kf_root = NULL;
1547 cgroup_exit_root_id(root);
1549 free_cgrp_cset_links(&tmp_links);
1553 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1554 int flags, const char *unused_dev_name,
1557 struct cgroup_root *root;
1558 struct cgroup_sb_opts opts;
1559 struct dentry *dentry;
1564 * The first time anyone tries to mount a cgroup, enable the list
1565 * linking each css_set to its tasks and fix up all existing tasks.
1567 if (!use_task_css_set_links)
1568 cgroup_enable_task_cg_lists();
1570 mutex_lock(&cgroup_tree_mutex);
1571 mutex_lock(&cgroup_mutex);
1573 /* First find the desired set of subsystems */
1574 ret = parse_cgroupfs_options(data, &opts);
1578 /* look for a matching existing root */
1579 if (!opts.subsys_mask && !opts.none && !opts.name) {
1580 cgrp_dfl_root_visible = true;
1581 root = &cgrp_dfl_root;
1582 cgroup_get(&root->cgrp);
1587 for_each_root(root) {
1588 bool name_match = false;
1590 if (root == &cgrp_dfl_root)
1594 * If we asked for a name then it must match. Also, if
1595 * name matches but sybsys_mask doesn't, we should fail.
1596 * Remember whether name matched.
1599 if (strcmp(opts.name, root->name))
1605 * If we asked for subsystems (or explicitly for no
1606 * subsystems) then they must match.
1608 if ((opts.subsys_mask || opts.none) &&
1609 (opts.subsys_mask != root->subsys_mask)) {
1616 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1617 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1618 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1622 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1627 * A root's lifetime is governed by its root cgroup. Zero
1628 * ref indicate that the root is being destroyed. Wait for
1629 * destruction to complete so that the subsystems are free.
1630 * We can use wait_queue for the wait but this path is
1631 * super cold. Let's just sleep for a bit and retry.
1633 if (!atomic_inc_not_zero(&root->cgrp.refcnt)) {
1634 mutex_unlock(&cgroup_mutex);
1635 mutex_unlock(&cgroup_tree_mutex);
1637 mutex_lock(&cgroup_tree_mutex);
1638 mutex_lock(&cgroup_mutex);
1647 * No such thing, create a new one. name= matching without subsys
1648 * specification is allowed for already existing hierarchies but we
1649 * can't create new one without subsys specification.
1651 if (!opts.subsys_mask && !opts.none) {
1656 root = kzalloc(sizeof(*root), GFP_KERNEL);
1662 init_cgroup_root(root, &opts);
1664 ret = cgroup_setup_root(root, opts.subsys_mask);
1666 cgroup_free_root(root);
1669 mutex_unlock(&cgroup_mutex);
1670 mutex_unlock(&cgroup_tree_mutex);
1672 kfree(opts.release_agent);
1676 return ERR_PTR(ret);
1678 dentry = kernfs_mount(fs_type, flags, root->kf_root, &new_sb);
1679 if (IS_ERR(dentry) || !new_sb)
1680 cgroup_put(&root->cgrp);
1684 static void cgroup_kill_sb(struct super_block *sb)
1686 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1687 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1689 cgroup_put(&root->cgrp);
1693 static struct file_system_type cgroup_fs_type = {
1695 .mount = cgroup_mount,
1696 .kill_sb = cgroup_kill_sb,
1699 static struct kobject *cgroup_kobj;
1702 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1703 * @task: target task
1704 * @buf: the buffer to write the path into
1705 * @buflen: the length of the buffer
1707 * Determine @task's cgroup on the first (the one with the lowest non-zero
1708 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1709 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1710 * cgroup controller callbacks.
1712 * Return value is the same as kernfs_path().
1714 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1716 struct cgroup_root *root;
1717 struct cgroup *cgrp;
1718 int hierarchy_id = 1;
1721 mutex_lock(&cgroup_mutex);
1722 down_read(&css_set_rwsem);
1724 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1727 cgrp = task_cgroup_from_root(task, root);
1728 path = cgroup_path(cgrp, buf, buflen);
1730 /* if no hierarchy exists, everyone is in "/" */
1731 if (strlcpy(buf, "/", buflen) < buflen)
1735 up_read(&css_set_rwsem);
1736 mutex_unlock(&cgroup_mutex);
1739 EXPORT_SYMBOL_GPL(task_cgroup_path);
1741 /* used to track tasks and other necessary states during migration */
1742 struct cgroup_taskset {
1743 /* the src and dst cset list running through cset->mg_node */
1744 struct list_head src_csets;
1745 struct list_head dst_csets;
1748 * Fields for cgroup_taskset_*() iteration.
1750 * Before migration is committed, the target migration tasks are on
1751 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1752 * the csets on ->dst_csets. ->csets point to either ->src_csets
1753 * or ->dst_csets depending on whether migration is committed.
1755 * ->cur_csets and ->cur_task point to the current task position
1758 struct list_head *csets;
1759 struct css_set *cur_cset;
1760 struct task_struct *cur_task;
1764 * cgroup_taskset_first - reset taskset and return the first task
1765 * @tset: taskset of interest
1767 * @tset iteration is initialized and the first task is returned.
1769 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1771 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
1772 tset->cur_task = NULL;
1774 return cgroup_taskset_next(tset);
1778 * cgroup_taskset_next - iterate to the next task in taskset
1779 * @tset: taskset of interest
1781 * Return the next task in @tset. Iteration must have been initialized
1782 * with cgroup_taskset_first().
1784 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1786 struct css_set *cset = tset->cur_cset;
1787 struct task_struct *task = tset->cur_task;
1789 while (&cset->mg_node != tset->csets) {
1791 task = list_first_entry(&cset->mg_tasks,
1792 struct task_struct, cg_list);
1794 task = list_next_entry(task, cg_list);
1796 if (&task->cg_list != &cset->mg_tasks) {
1797 tset->cur_cset = cset;
1798 tset->cur_task = task;
1802 cset = list_next_entry(cset, mg_node);
1810 * cgroup_task_migrate - move a task from one cgroup to another.
1811 * @old_cgrp; the cgroup @tsk is being migrated from
1812 * @tsk: the task being migrated
1813 * @new_cset: the new css_set @tsk is being attached to
1815 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1817 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1818 struct task_struct *tsk,
1819 struct css_set *new_cset)
1821 struct css_set *old_cset;
1823 lockdep_assert_held(&cgroup_mutex);
1824 lockdep_assert_held(&css_set_rwsem);
1827 * We are synchronized through threadgroup_lock() against PF_EXITING
1828 * setting such that we can't race against cgroup_exit() changing the
1829 * css_set to init_css_set and dropping the old one.
1831 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1832 old_cset = task_css_set(tsk);
1834 get_css_set(new_cset);
1835 rcu_assign_pointer(tsk->cgroups, new_cset);
1838 * Use move_tail so that cgroup_taskset_first() still returns the
1839 * leader after migration. This works because cgroup_migrate()
1840 * ensures that the dst_cset of the leader is the first on the
1841 * tset's dst_csets list.
1843 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
1846 * We just gained a reference on old_cset by taking it from the
1847 * task. As trading it for new_cset is protected by cgroup_mutex,
1848 * we're safe to drop it here; it will be freed under RCU.
1850 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1851 put_css_set_locked(old_cset, false);
1855 * cgroup_migrate_finish - cleanup after attach
1856 * @preloaded_csets: list of preloaded css_sets
1858 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
1859 * those functions for details.
1861 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
1863 struct css_set *cset, *tmp_cset;
1865 lockdep_assert_held(&cgroup_mutex);
1867 down_write(&css_set_rwsem);
1868 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
1869 cset->mg_src_cgrp = NULL;
1870 cset->mg_dst_cset = NULL;
1871 list_del_init(&cset->mg_preload_node);
1872 put_css_set_locked(cset, false);
1874 up_write(&css_set_rwsem);
1878 * cgroup_migrate_add_src - add a migration source css_set
1879 * @src_cset: the source css_set to add
1880 * @dst_cgrp: the destination cgroup
1881 * @preloaded_csets: list of preloaded css_sets
1883 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
1884 * @src_cset and add it to @preloaded_csets, which should later be cleaned
1885 * up by cgroup_migrate_finish().
1887 * This function may be called without holding threadgroup_lock even if the
1888 * target is a process. Threads may be created and destroyed but as long
1889 * as cgroup_mutex is not dropped, no new css_set can be put into play and
1890 * the preloaded css_sets are guaranteed to cover all migrations.
1892 static void cgroup_migrate_add_src(struct css_set *src_cset,
1893 struct cgroup *dst_cgrp,
1894 struct list_head *preloaded_csets)
1896 struct cgroup *src_cgrp;
1898 lockdep_assert_held(&cgroup_mutex);
1899 lockdep_assert_held(&css_set_rwsem);
1901 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
1903 /* nothing to do if this cset already belongs to the cgroup */
1904 if (src_cgrp == dst_cgrp)
1907 if (!list_empty(&src_cset->mg_preload_node))
1910 WARN_ON(src_cset->mg_src_cgrp);
1911 WARN_ON(!list_empty(&src_cset->mg_tasks));
1912 WARN_ON(!list_empty(&src_cset->mg_node));
1914 src_cset->mg_src_cgrp = src_cgrp;
1915 get_css_set(src_cset);
1916 list_add(&src_cset->mg_preload_node, preloaded_csets);
1920 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
1921 * @dst_cgrp: the destination cgroup
1922 * @preloaded_csets: list of preloaded source css_sets
1924 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
1925 * have been preloaded to @preloaded_csets. This function looks up and
1926 * pins all destination css_sets, links each to its source, and put them on
1929 * This function must be called after cgroup_migrate_add_src() has been
1930 * called on each migration source css_set. After migration is performed
1931 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
1934 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
1935 struct list_head *preloaded_csets)
1938 struct css_set *src_cset;
1940 lockdep_assert_held(&cgroup_mutex);
1942 /* look up the dst cset for each src cset and link it to src */
1943 list_for_each_entry(src_cset, preloaded_csets, mg_preload_node) {
1944 struct css_set *dst_cset;
1946 dst_cset = find_css_set(src_cset, dst_cgrp);
1950 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
1951 src_cset->mg_dst_cset = dst_cset;
1953 if (list_empty(&dst_cset->mg_preload_node))
1954 list_add(&dst_cset->mg_preload_node, &csets);
1956 put_css_set(dst_cset, false);
1959 list_splice(&csets, preloaded_csets);
1962 cgroup_migrate_finish(&csets);
1967 * cgroup_migrate - migrate a process or task to a cgroup
1968 * @cgrp: the destination cgroup
1969 * @leader: the leader of the process or the task to migrate
1970 * @threadgroup: whether @leader points to the whole process or a single task
1972 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
1973 * process, the caller must be holding threadgroup_lock of @leader. The
1974 * caller is also responsible for invoking cgroup_migrate_add_src() and
1975 * cgroup_migrate_prepare_dst() on the targets before invoking this
1976 * function and following up with cgroup_migrate_finish().
1978 * As long as a controller's ->can_attach() doesn't fail, this function is
1979 * guaranteed to succeed. This means that, excluding ->can_attach()
1980 * failure, when migrating multiple targets, the success or failure can be
1981 * decided for all targets by invoking group_migrate_prepare_dst() before
1982 * actually starting migrating.
1984 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
1987 struct cgroup_taskset tset = {
1988 .src_csets = LIST_HEAD_INIT(tset.src_csets),
1989 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
1990 .csets = &tset.src_csets,
1992 struct cgroup_subsys_state *css, *failed_css = NULL;
1993 struct css_set *cset, *tmp_cset;
1994 struct task_struct *task, *tmp_task;
1998 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1999 * already PF_EXITING could be freed from underneath us unless we
2000 * take an rcu_read_lock.
2002 down_write(&css_set_rwsem);
2006 /* @task either already exited or can't exit until the end */
2007 if (task->flags & PF_EXITING)
2010 /* leave @task alone if post_fork() hasn't linked it yet */
2011 if (list_empty(&task->cg_list))
2014 cset = task_css_set(task);
2015 if (!cset->mg_src_cgrp)
2019 * cgroup_taskset_first() must always return the leader.
2020 * Take care to avoid disturbing the ordering.
2022 list_move_tail(&task->cg_list, &cset->mg_tasks);
2023 if (list_empty(&cset->mg_node))
2024 list_add_tail(&cset->mg_node, &tset.src_csets);
2025 if (list_empty(&cset->mg_dst_cset->mg_node))
2026 list_move_tail(&cset->mg_dst_cset->mg_node,
2031 } while_each_thread(leader, task);
2033 up_write(&css_set_rwsem);
2035 /* methods shouldn't be called if no task is actually migrating */
2036 if (list_empty(&tset.src_csets))
2039 /* check that we can legitimately attach to the cgroup */
2040 for_each_e_css(css, i, cgrp) {
2041 if (css->ss->can_attach) {
2042 ret = css->ss->can_attach(css, &tset);
2045 goto out_cancel_attach;
2051 * Now that we're guaranteed success, proceed to move all tasks to
2052 * the new cgroup. There are no failure cases after here, so this
2053 * is the commit point.
2055 down_write(&css_set_rwsem);
2056 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2057 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2058 cgroup_task_migrate(cset->mg_src_cgrp, task,
2061 up_write(&css_set_rwsem);
2064 * Migration is committed, all target tasks are now on dst_csets.
2065 * Nothing is sensitive to fork() after this point. Notify
2066 * controllers that migration is complete.
2068 tset.csets = &tset.dst_csets;
2070 for_each_e_css(css, i, cgrp)
2071 if (css->ss->attach)
2072 css->ss->attach(css, &tset);
2075 goto out_release_tset;
2078 for_each_e_css(css, i, cgrp) {
2079 if (css == failed_css)
2081 if (css->ss->cancel_attach)
2082 css->ss->cancel_attach(css, &tset);
2085 down_write(&css_set_rwsem);
2086 list_splice_init(&tset.dst_csets, &tset.src_csets);
2087 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2088 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2089 list_del_init(&cset->mg_node);
2091 up_write(&css_set_rwsem);
2096 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2097 * @dst_cgrp: the cgroup to attach to
2098 * @leader: the task or the leader of the threadgroup to be attached
2099 * @threadgroup: attach the whole threadgroup?
2101 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2103 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2104 struct task_struct *leader, bool threadgroup)
2106 LIST_HEAD(preloaded_csets);
2107 struct task_struct *task;
2110 /* look up all src csets */
2111 down_read(&css_set_rwsem);
2115 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2119 } while_each_thread(leader, task);
2121 up_read(&css_set_rwsem);
2123 /* prepare dst csets and commit */
2124 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2126 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2128 cgroup_migrate_finish(&preloaded_csets);
2133 * Find the task_struct of the task to attach by vpid and pass it along to the
2134 * function to attach either it or all tasks in its threadgroup. Will lock
2135 * cgroup_mutex and threadgroup.
2137 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2139 struct task_struct *tsk;
2140 const struct cred *cred = current_cred(), *tcred;
2143 if (!cgroup_lock_live_group(cgrp))
2149 tsk = find_task_by_vpid(pid);
2153 goto out_unlock_cgroup;
2156 * even if we're attaching all tasks in the thread group, we
2157 * only need to check permissions on one of them.
2159 tcred = __task_cred(tsk);
2160 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2161 !uid_eq(cred->euid, tcred->uid) &&
2162 !uid_eq(cred->euid, tcred->suid)) {
2165 goto out_unlock_cgroup;
2171 tsk = tsk->group_leader;
2174 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2175 * trapped in a cpuset, or RT worker may be born in a cgroup
2176 * with no rt_runtime allocated. Just say no.
2178 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2181 goto out_unlock_cgroup;
2184 get_task_struct(tsk);
2187 threadgroup_lock(tsk);
2189 if (!thread_group_leader(tsk)) {
2191 * a race with de_thread from another thread's exec()
2192 * may strip us of our leadership, if this happens,
2193 * there is no choice but to throw this task away and
2194 * try again; this is
2195 * "double-double-toil-and-trouble-check locking".
2197 threadgroup_unlock(tsk);
2198 put_task_struct(tsk);
2199 goto retry_find_task;
2203 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2205 threadgroup_unlock(tsk);
2207 put_task_struct(tsk);
2209 mutex_unlock(&cgroup_mutex);
2214 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2215 * @from: attach to all cgroups of a given task
2216 * @tsk: the task to be attached
2218 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2220 struct cgroup_root *root;
2223 mutex_lock(&cgroup_mutex);
2224 for_each_root(root) {
2225 struct cgroup *from_cgrp;
2227 if (root == &cgrp_dfl_root)
2230 down_read(&css_set_rwsem);
2231 from_cgrp = task_cgroup_from_root(from, root);
2232 up_read(&css_set_rwsem);
2234 retval = cgroup_attach_task(from_cgrp, tsk, false);
2238 mutex_unlock(&cgroup_mutex);
2242 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2244 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2245 struct cftype *cft, u64 pid)
2247 return attach_task_by_pid(css->cgroup, pid, false);
2250 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2251 struct cftype *cft, u64 tgid)
2253 return attach_task_by_pid(css->cgroup, tgid, true);
2256 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2257 struct cftype *cft, char *buffer)
2259 struct cgroup_root *root = css->cgroup->root;
2261 BUILD_BUG_ON(sizeof(root->release_agent_path) < PATH_MAX);
2262 if (!cgroup_lock_live_group(css->cgroup))
2264 spin_lock(&release_agent_path_lock);
2265 strlcpy(root->release_agent_path, buffer,
2266 sizeof(root->release_agent_path));
2267 spin_unlock(&release_agent_path_lock);
2268 mutex_unlock(&cgroup_mutex);
2272 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2274 struct cgroup *cgrp = seq_css(seq)->cgroup;
2276 if (!cgroup_lock_live_group(cgrp))
2278 seq_puts(seq, cgrp->root->release_agent_path);
2279 seq_putc(seq, '\n');
2280 mutex_unlock(&cgroup_mutex);
2284 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2286 struct cgroup *cgrp = seq_css(seq)->cgroup;
2288 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2292 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2293 size_t nbytes, loff_t off)
2295 struct cgroup *cgrp = of->kn->parent->priv;
2296 struct cftype *cft = of->kn->priv;
2297 struct cgroup_subsys_state *css;
2301 * kernfs guarantees that a file isn't deleted with operations in
2302 * flight, which means that the matching css is and stays alive and
2303 * doesn't need to be pinned. The RCU locking is not necessary
2304 * either. It's just for the convenience of using cgroup_css().
2307 css = cgroup_css(cgrp, cft->ss);
2310 if (cft->write_string) {
2311 ret = cft->write_string(css, cft, strstrip(buf));
2312 } else if (cft->write_u64) {
2313 unsigned long long v;
2314 ret = kstrtoull(buf, 0, &v);
2316 ret = cft->write_u64(css, cft, v);
2317 } else if (cft->write_s64) {
2319 ret = kstrtoll(buf, 0, &v);
2321 ret = cft->write_s64(css, cft, v);
2322 } else if (cft->trigger) {
2323 ret = cft->trigger(css, (unsigned int)cft->private);
2328 return ret ?: nbytes;
2331 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2333 return seq_cft(seq)->seq_start(seq, ppos);
2336 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2338 return seq_cft(seq)->seq_next(seq, v, ppos);
2341 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2343 seq_cft(seq)->seq_stop(seq, v);
2346 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2348 struct cftype *cft = seq_cft(m);
2349 struct cgroup_subsys_state *css = seq_css(m);
2352 return cft->seq_show(m, arg);
2355 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2356 else if (cft->read_s64)
2357 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2363 static struct kernfs_ops cgroup_kf_single_ops = {
2364 .atomic_write_len = PAGE_SIZE,
2365 .write = cgroup_file_write,
2366 .seq_show = cgroup_seqfile_show,
2369 static struct kernfs_ops cgroup_kf_ops = {
2370 .atomic_write_len = PAGE_SIZE,
2371 .write = cgroup_file_write,
2372 .seq_start = cgroup_seqfile_start,
2373 .seq_next = cgroup_seqfile_next,
2374 .seq_stop = cgroup_seqfile_stop,
2375 .seq_show = cgroup_seqfile_show,
2379 * cgroup_rename - Only allow simple rename of directories in place.
2381 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2382 const char *new_name_str)
2384 struct cgroup *cgrp = kn->priv;
2387 if (kernfs_type(kn) != KERNFS_DIR)
2389 if (kn->parent != new_parent)
2393 * This isn't a proper migration and its usefulness is very
2394 * limited. Disallow if sane_behavior.
2396 if (cgroup_sane_behavior(cgrp))
2400 * We're gonna grab cgroup_tree_mutex which nests outside kernfs
2401 * active_ref. kernfs_rename() doesn't require active_ref
2402 * protection. Break them before grabbing cgroup_tree_mutex.
2404 kernfs_break_active_protection(new_parent);
2405 kernfs_break_active_protection(kn);
2407 mutex_lock(&cgroup_tree_mutex);
2408 mutex_lock(&cgroup_mutex);
2410 ret = kernfs_rename(kn, new_parent, new_name_str);
2412 mutex_unlock(&cgroup_mutex);
2413 mutex_unlock(&cgroup_tree_mutex);
2415 kernfs_unbreak_active_protection(kn);
2416 kernfs_unbreak_active_protection(new_parent);
2420 /* set uid and gid of cgroup dirs and files to that of the creator */
2421 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
2423 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
2424 .ia_uid = current_fsuid(),
2425 .ia_gid = current_fsgid(), };
2427 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
2428 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
2431 return kernfs_setattr(kn, &iattr);
2434 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2436 char name[CGROUP_FILE_NAME_MAX];
2437 struct kernfs_node *kn;
2438 struct lock_class_key *key = NULL;
2441 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2442 key = &cft->lockdep_key;
2444 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2445 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2450 ret = cgroup_kn_set_ugid(kn);
2457 * cgroup_addrm_files - add or remove files to a cgroup directory
2458 * @cgrp: the target cgroup
2459 * @cfts: array of cftypes to be added
2460 * @is_add: whether to add or remove
2462 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2463 * For removals, this function never fails. If addition fails, this
2464 * function doesn't remove files already added. The caller is responsible
2467 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2473 lockdep_assert_held(&cgroup_tree_mutex);
2475 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2476 /* does cft->flags tell us to skip this file on @cgrp? */
2477 if ((cft->flags & CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
2479 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2481 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2483 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2487 ret = cgroup_add_file(cgrp, cft);
2489 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2494 cgroup_rm_file(cgrp, cft);
2500 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2503 struct cgroup_subsys *ss = cfts[0].ss;
2504 struct cgroup *root = &ss->root->cgrp;
2505 struct cgroup_subsys_state *css;
2508 lockdep_assert_held(&cgroup_tree_mutex);
2510 /* add/rm files for all cgroups created before */
2511 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2512 struct cgroup *cgrp = css->cgroup;
2514 if (cgroup_is_dead(cgrp))
2517 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2523 kernfs_activate(root->kn);
2527 static void cgroup_exit_cftypes(struct cftype *cfts)
2531 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2532 /* free copy for custom atomic_write_len, see init_cftypes() */
2533 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2540 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2544 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2545 struct kernfs_ops *kf_ops;
2547 WARN_ON(cft->ss || cft->kf_ops);
2550 kf_ops = &cgroup_kf_ops;
2552 kf_ops = &cgroup_kf_single_ops;
2555 * Ugh... if @cft wants a custom max_write_len, we need to
2556 * make a copy of kf_ops to set its atomic_write_len.
2558 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
2559 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
2561 cgroup_exit_cftypes(cfts);
2564 kf_ops->atomic_write_len = cft->max_write_len;
2567 cft->kf_ops = kf_ops;
2574 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
2576 lockdep_assert_held(&cgroup_tree_mutex);
2578 if (!cfts || !cfts[0].ss)
2581 list_del(&cfts->node);
2582 cgroup_apply_cftypes(cfts, false);
2583 cgroup_exit_cftypes(cfts);
2588 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2589 * @cfts: zero-length name terminated array of cftypes
2591 * Unregister @cfts. Files described by @cfts are removed from all
2592 * existing cgroups and all future cgroups won't have them either. This
2593 * function can be called anytime whether @cfts' subsys is attached or not.
2595 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2598 int cgroup_rm_cftypes(struct cftype *cfts)
2602 mutex_lock(&cgroup_tree_mutex);
2603 ret = cgroup_rm_cftypes_locked(cfts);
2604 mutex_unlock(&cgroup_tree_mutex);
2609 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2610 * @ss: target cgroup subsystem
2611 * @cfts: zero-length name terminated array of cftypes
2613 * Register @cfts to @ss. Files described by @cfts are created for all
2614 * existing cgroups to which @ss is attached and all future cgroups will
2615 * have them too. This function can be called anytime whether @ss is
2618 * Returns 0 on successful registration, -errno on failure. Note that this
2619 * function currently returns 0 as long as @cfts registration is successful
2620 * even if some file creation attempts on existing cgroups fail.
2622 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2626 if (!cfts || cfts[0].name[0] == '\0')
2629 ret = cgroup_init_cftypes(ss, cfts);
2633 mutex_lock(&cgroup_tree_mutex);
2635 list_add_tail(&cfts->node, &ss->cfts);
2636 ret = cgroup_apply_cftypes(cfts, true);
2638 cgroup_rm_cftypes_locked(cfts);
2640 mutex_unlock(&cgroup_tree_mutex);
2645 * cgroup_task_count - count the number of tasks in a cgroup.
2646 * @cgrp: the cgroup in question
2648 * Return the number of tasks in the cgroup.
2650 static int cgroup_task_count(const struct cgroup *cgrp)
2653 struct cgrp_cset_link *link;
2655 down_read(&css_set_rwsem);
2656 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2657 count += atomic_read(&link->cset->refcount);
2658 up_read(&css_set_rwsem);
2663 * css_next_child - find the next child of a given css
2664 * @pos_css: the current position (%NULL to initiate traversal)
2665 * @parent_css: css whose children to walk
2667 * This function returns the next child of @parent_css and should be called
2668 * under either cgroup_mutex or RCU read lock. The only requirement is
2669 * that @parent_css and @pos_css are accessible. The next sibling is
2670 * guaranteed to be returned regardless of their states.
2672 struct cgroup_subsys_state *
2673 css_next_child(struct cgroup_subsys_state *pos_css,
2674 struct cgroup_subsys_state *parent_css)
2676 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2677 struct cgroup *cgrp = parent_css->cgroup;
2678 struct cgroup *next;
2680 cgroup_assert_mutexes_or_rcu_locked();
2683 * @pos could already have been removed. Once a cgroup is removed,
2684 * its ->sibling.next is no longer updated when its next sibling
2685 * changes. As CGRP_DEAD assertion is serialized and happens
2686 * before the cgroup is taken off the ->sibling list, if we see it
2687 * unasserted, it's guaranteed that the next sibling hasn't
2688 * finished its grace period even if it's already removed, and thus
2689 * safe to dereference from this RCU critical section. If
2690 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2691 * to be visible as %true here.
2693 * If @pos is dead, its next pointer can't be dereferenced;
2694 * however, as each cgroup is given a monotonically increasing
2695 * unique serial number and always appended to the sibling list,
2696 * the next one can be found by walking the parent's children until
2697 * we see a cgroup with higher serial number than @pos's. While
2698 * this path can be slower, it's taken only when either the current
2699 * cgroup is removed or iteration and removal race.
2702 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2703 } else if (likely(!cgroup_is_dead(pos))) {
2704 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2706 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2707 if (next->serial_nr > pos->serial_nr)
2712 * @next, if not pointing to the head, can be dereferenced and is
2713 * the next sibling; however, it might have @ss disabled. If so,
2714 * fast-forward to the next enabled one.
2716 while (&next->sibling != &cgrp->children) {
2717 struct cgroup_subsys_state *next_css = cgroup_css(next, parent_css->ss);
2721 next = list_entry_rcu(next->sibling.next, struct cgroup, sibling);
2727 * css_next_descendant_pre - find the next descendant for pre-order walk
2728 * @pos: the current position (%NULL to initiate traversal)
2729 * @root: css whose descendants to walk
2731 * To be used by css_for_each_descendant_pre(). Find the next descendant
2732 * to visit for pre-order traversal of @root's descendants. @root is
2733 * included in the iteration and the first node to be visited.
2735 * While this function requires cgroup_mutex or RCU read locking, it
2736 * doesn't require the whole traversal to be contained in a single critical
2737 * section. This function will return the correct next descendant as long
2738 * as both @pos and @root are accessible and @pos is a descendant of @root.
2740 struct cgroup_subsys_state *
2741 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2742 struct cgroup_subsys_state *root)
2744 struct cgroup_subsys_state *next;
2746 cgroup_assert_mutexes_or_rcu_locked();
2748 /* if first iteration, visit @root */
2752 /* visit the first child if exists */
2753 next = css_next_child(NULL, pos);
2757 /* no child, visit my or the closest ancestor's next sibling */
2758 while (pos != root) {
2759 next = css_next_child(pos, css_parent(pos));
2762 pos = css_parent(pos);
2769 * css_rightmost_descendant - return the rightmost descendant of a css
2770 * @pos: css of interest
2772 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2773 * is returned. This can be used during pre-order traversal to skip
2776 * While this function requires cgroup_mutex or RCU read locking, it
2777 * doesn't require the whole traversal to be contained in a single critical
2778 * section. This function will return the correct rightmost descendant as
2779 * long as @pos is accessible.
2781 struct cgroup_subsys_state *
2782 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2784 struct cgroup_subsys_state *last, *tmp;
2786 cgroup_assert_mutexes_or_rcu_locked();
2790 /* ->prev isn't RCU safe, walk ->next till the end */
2792 css_for_each_child(tmp, last)
2799 static struct cgroup_subsys_state *
2800 css_leftmost_descendant(struct cgroup_subsys_state *pos)
2802 struct cgroup_subsys_state *last;
2806 pos = css_next_child(NULL, pos);
2813 * css_next_descendant_post - find the next descendant for post-order walk
2814 * @pos: the current position (%NULL to initiate traversal)
2815 * @root: css whose descendants to walk
2817 * To be used by css_for_each_descendant_post(). Find the next descendant
2818 * to visit for post-order traversal of @root's descendants. @root is
2819 * included in the iteration and the last node to be visited.
2821 * While this function requires cgroup_mutex or RCU read locking, it
2822 * doesn't require the whole traversal to be contained in a single critical
2823 * section. This function will return the correct next descendant as long
2824 * as both @pos and @cgroup are accessible and @pos is a descendant of
2827 struct cgroup_subsys_state *
2828 css_next_descendant_post(struct cgroup_subsys_state *pos,
2829 struct cgroup_subsys_state *root)
2831 struct cgroup_subsys_state *next;
2833 cgroup_assert_mutexes_or_rcu_locked();
2835 /* if first iteration, visit leftmost descendant which may be @root */
2837 return css_leftmost_descendant(root);
2839 /* if we visited @root, we're done */
2843 /* if there's an unvisited sibling, visit its leftmost descendant */
2844 next = css_next_child(pos, css_parent(pos));
2846 return css_leftmost_descendant(next);
2848 /* no sibling left, visit parent */
2849 return css_parent(pos);
2853 * css_advance_task_iter - advance a task itererator to the next css_set
2854 * @it: the iterator to advance
2856 * Advance @it to the next css_set to walk.
2858 static void css_advance_task_iter(struct css_task_iter *it)
2860 struct list_head *l = it->cset_link;
2861 struct cgrp_cset_link *link;
2862 struct css_set *cset;
2864 /* Advance to the next non-empty css_set */
2867 if (l == &it->origin_css->cgroup->cset_links) {
2868 it->cset_link = NULL;
2871 link = list_entry(l, struct cgrp_cset_link, cset_link);
2873 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
2877 if (!list_empty(&cset->tasks))
2878 it->task = cset->tasks.next;
2880 it->task = cset->mg_tasks.next;
2884 * css_task_iter_start - initiate task iteration
2885 * @css: the css to walk tasks of
2886 * @it: the task iterator to use
2888 * Initiate iteration through the tasks of @css. The caller can call
2889 * css_task_iter_next() to walk through the tasks until the function
2890 * returns NULL. On completion of iteration, css_task_iter_end() must be
2893 * Note that this function acquires a lock which is released when the
2894 * iteration finishes. The caller can't sleep while iteration is in
2897 void css_task_iter_start(struct cgroup_subsys_state *css,
2898 struct css_task_iter *it)
2899 __acquires(css_set_rwsem)
2901 /* no one should try to iterate before mounting cgroups */
2902 WARN_ON_ONCE(!use_task_css_set_links);
2904 down_read(&css_set_rwsem);
2906 it->origin_css = css;
2907 it->cset_link = &css->cgroup->cset_links;
2909 css_advance_task_iter(it);
2913 * css_task_iter_next - return the next task for the iterator
2914 * @it: the task iterator being iterated
2916 * The "next" function for task iteration. @it should have been
2917 * initialized via css_task_iter_start(). Returns NULL when the iteration
2920 struct task_struct *css_task_iter_next(struct css_task_iter *it)
2922 struct task_struct *res;
2923 struct list_head *l = it->task;
2924 struct cgrp_cset_link *link = list_entry(it->cset_link,
2925 struct cgrp_cset_link, cset_link);
2927 /* If the iterator cg is NULL, we have no tasks */
2930 res = list_entry(l, struct task_struct, cg_list);
2933 * Advance iterator to find next entry. cset->tasks is consumed
2934 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
2939 if (l == &link->cset->tasks)
2940 l = link->cset->mg_tasks.next;
2942 if (l == &link->cset->mg_tasks)
2943 css_advance_task_iter(it);
2951 * css_task_iter_end - finish task iteration
2952 * @it: the task iterator to finish
2954 * Finish task iteration started by css_task_iter_start().
2956 void css_task_iter_end(struct css_task_iter *it)
2957 __releases(css_set_rwsem)
2959 up_read(&css_set_rwsem);
2963 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
2964 * @to: cgroup to which the tasks will be moved
2965 * @from: cgroup in which the tasks currently reside
2967 * Locking rules between cgroup_post_fork() and the migration path
2968 * guarantee that, if a task is forking while being migrated, the new child
2969 * is guaranteed to be either visible in the source cgroup after the
2970 * parent's migration is complete or put into the target cgroup. No task
2971 * can slip out of migration through forking.
2973 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
2975 LIST_HEAD(preloaded_csets);
2976 struct cgrp_cset_link *link;
2977 struct css_task_iter it;
2978 struct task_struct *task;
2981 mutex_lock(&cgroup_mutex);
2983 /* all tasks in @from are being moved, all csets are source */
2984 down_read(&css_set_rwsem);
2985 list_for_each_entry(link, &from->cset_links, cset_link)
2986 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
2987 up_read(&css_set_rwsem);
2989 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
2994 * Migrate tasks one-by-one until @form is empty. This fails iff
2995 * ->can_attach() fails.
2998 css_task_iter_start(&from->dummy_css, &it);
2999 task = css_task_iter_next(&it);
3001 get_task_struct(task);
3002 css_task_iter_end(&it);
3005 ret = cgroup_migrate(to, task, false);
3006 put_task_struct(task);
3008 } while (task && !ret);
3010 cgroup_migrate_finish(&preloaded_csets);
3011 mutex_unlock(&cgroup_mutex);
3016 * Stuff for reading the 'tasks'/'procs' files.
3018 * Reading this file can return large amounts of data if a cgroup has
3019 * *lots* of attached tasks. So it may need several calls to read(),
3020 * but we cannot guarantee that the information we produce is correct
3021 * unless we produce it entirely atomically.
3025 /* which pidlist file are we talking about? */
3026 enum cgroup_filetype {
3032 * A pidlist is a list of pids that virtually represents the contents of one
3033 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3034 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3037 struct cgroup_pidlist {
3039 * used to find which pidlist is wanted. doesn't change as long as
3040 * this particular list stays in the list.
3042 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3045 /* how many elements the above list has */
3047 /* each of these stored in a list by its cgroup */
3048 struct list_head links;
3049 /* pointer to the cgroup we belong to, for list removal purposes */
3050 struct cgroup *owner;
3051 /* for delayed destruction */
3052 struct delayed_work destroy_dwork;
3056 * The following two functions "fix" the issue where there are more pids
3057 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3058 * TODO: replace with a kernel-wide solution to this problem
3060 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3061 static void *pidlist_allocate(int count)
3063 if (PIDLIST_TOO_LARGE(count))
3064 return vmalloc(count * sizeof(pid_t));
3066 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3069 static void pidlist_free(void *p)
3071 if (is_vmalloc_addr(p))
3078 * Used to destroy all pidlists lingering waiting for destroy timer. None
3079 * should be left afterwards.
3081 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3083 struct cgroup_pidlist *l, *tmp_l;
3085 mutex_lock(&cgrp->pidlist_mutex);
3086 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3087 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3088 mutex_unlock(&cgrp->pidlist_mutex);
3090 flush_workqueue(cgroup_pidlist_destroy_wq);
3091 BUG_ON(!list_empty(&cgrp->pidlists));
3094 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3096 struct delayed_work *dwork = to_delayed_work(work);
3097 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3099 struct cgroup_pidlist *tofree = NULL;
3101 mutex_lock(&l->owner->pidlist_mutex);
3104 * Destroy iff we didn't get queued again. The state won't change
3105 * as destroy_dwork can only be queued while locked.
3107 if (!delayed_work_pending(dwork)) {
3108 list_del(&l->links);
3109 pidlist_free(l->list);
3110 put_pid_ns(l->key.ns);
3114 mutex_unlock(&l->owner->pidlist_mutex);
3119 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3120 * Returns the number of unique elements.
3122 static int pidlist_uniq(pid_t *list, int length)
3127 * we presume the 0th element is unique, so i starts at 1. trivial
3128 * edge cases first; no work needs to be done for either
3130 if (length == 0 || length == 1)
3132 /* src and dest walk down the list; dest counts unique elements */
3133 for (src = 1; src < length; src++) {
3134 /* find next unique element */
3135 while (list[src] == list[src-1]) {
3140 /* dest always points to where the next unique element goes */
3141 list[dest] = list[src];
3149 * The two pid files - task and cgroup.procs - guaranteed that the result
3150 * is sorted, which forced this whole pidlist fiasco. As pid order is
3151 * different per namespace, each namespace needs differently sorted list,
3152 * making it impossible to use, for example, single rbtree of member tasks
3153 * sorted by task pointer. As pidlists can be fairly large, allocating one
3154 * per open file is dangerous, so cgroup had to implement shared pool of
3155 * pidlists keyed by cgroup and namespace.
3157 * All this extra complexity was caused by the original implementation
3158 * committing to an entirely unnecessary property. In the long term, we
3159 * want to do away with it. Explicitly scramble sort order if
3160 * sane_behavior so that no such expectation exists in the new interface.
3162 * Scrambling is done by swapping every two consecutive bits, which is
3163 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3165 static pid_t pid_fry(pid_t pid)
3167 unsigned a = pid & 0x55555555;
3168 unsigned b = pid & 0xAAAAAAAA;
3170 return (a << 1) | (b >> 1);
3173 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3175 if (cgroup_sane_behavior(cgrp))
3176 return pid_fry(pid);
3181 static int cmppid(const void *a, const void *b)
3183 return *(pid_t *)a - *(pid_t *)b;
3186 static int fried_cmppid(const void *a, const void *b)
3188 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3191 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3192 enum cgroup_filetype type)
3194 struct cgroup_pidlist *l;
3195 /* don't need task_nsproxy() if we're looking at ourself */
3196 struct pid_namespace *ns = task_active_pid_ns(current);
3198 lockdep_assert_held(&cgrp->pidlist_mutex);
3200 list_for_each_entry(l, &cgrp->pidlists, links)
3201 if (l->key.type == type && l->key.ns == ns)
3207 * find the appropriate pidlist for our purpose (given procs vs tasks)
3208 * returns with the lock on that pidlist already held, and takes care
3209 * of the use count, or returns NULL with no locks held if we're out of
3212 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3213 enum cgroup_filetype type)
3215 struct cgroup_pidlist *l;
3217 lockdep_assert_held(&cgrp->pidlist_mutex);
3219 l = cgroup_pidlist_find(cgrp, type);
3223 /* entry not found; create a new one */
3224 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3228 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3230 /* don't need task_nsproxy() if we're looking at ourself */
3231 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3233 list_add(&l->links, &cgrp->pidlists);
3238 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3240 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3241 struct cgroup_pidlist **lp)
3245 int pid, n = 0; /* used for populating the array */
3246 struct css_task_iter it;
3247 struct task_struct *tsk;
3248 struct cgroup_pidlist *l;
3250 lockdep_assert_held(&cgrp->pidlist_mutex);
3253 * If cgroup gets more users after we read count, we won't have
3254 * enough space - tough. This race is indistinguishable to the
3255 * caller from the case that the additional cgroup users didn't
3256 * show up until sometime later on.
3258 length = cgroup_task_count(cgrp);
3259 array = pidlist_allocate(length);
3262 /* now, populate the array */
3263 css_task_iter_start(&cgrp->dummy_css, &it);
3264 while ((tsk = css_task_iter_next(&it))) {
3265 if (unlikely(n == length))
3267 /* get tgid or pid for procs or tasks file respectively */
3268 if (type == CGROUP_FILE_PROCS)
3269 pid = task_tgid_vnr(tsk);
3271 pid = task_pid_vnr(tsk);
3272 if (pid > 0) /* make sure to only use valid results */
3275 css_task_iter_end(&it);
3277 /* now sort & (if procs) strip out duplicates */
3278 if (cgroup_sane_behavior(cgrp))
3279 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3281 sort(array, length, sizeof(pid_t), cmppid, NULL);
3282 if (type == CGROUP_FILE_PROCS)
3283 length = pidlist_uniq(array, length);
3285 l = cgroup_pidlist_find_create(cgrp, type);
3287 mutex_unlock(&cgrp->pidlist_mutex);
3288 pidlist_free(array);
3292 /* store array, freeing old if necessary */
3293 pidlist_free(l->list);
3301 * cgroupstats_build - build and fill cgroupstats
3302 * @stats: cgroupstats to fill information into
3303 * @dentry: A dentry entry belonging to the cgroup for which stats have
3306 * Build and fill cgroupstats so that taskstats can export it to user
3309 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3311 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3312 struct cgroup *cgrp;
3313 struct css_task_iter it;
3314 struct task_struct *tsk;
3316 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3317 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3318 kernfs_type(kn) != KERNFS_DIR)
3321 mutex_lock(&cgroup_mutex);
3324 * We aren't being called from kernfs and there's no guarantee on
3325 * @kn->priv's validity. For this and css_tryget_from_dir(),
3326 * @kn->priv is RCU safe. Let's do the RCU dancing.
3329 cgrp = rcu_dereference(kn->priv);
3330 if (!cgrp || cgroup_is_dead(cgrp)) {
3332 mutex_unlock(&cgroup_mutex);
3337 css_task_iter_start(&cgrp->dummy_css, &it);
3338 while ((tsk = css_task_iter_next(&it))) {
3339 switch (tsk->state) {
3341 stats->nr_running++;
3343 case TASK_INTERRUPTIBLE:
3344 stats->nr_sleeping++;
3346 case TASK_UNINTERRUPTIBLE:
3347 stats->nr_uninterruptible++;
3350 stats->nr_stopped++;
3353 if (delayacct_is_task_waiting_on_io(tsk))
3354 stats->nr_io_wait++;
3358 css_task_iter_end(&it);
3360 mutex_unlock(&cgroup_mutex);
3366 * seq_file methods for the tasks/procs files. The seq_file position is the
3367 * next pid to display; the seq_file iterator is a pointer to the pid
3368 * in the cgroup->l->list array.
3371 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3374 * Initially we receive a position value that corresponds to
3375 * one more than the last pid shown (or 0 on the first call or
3376 * after a seek to the start). Use a binary-search to find the
3377 * next pid to display, if any
3379 struct kernfs_open_file *of = s->private;
3380 struct cgroup *cgrp = seq_css(s)->cgroup;
3381 struct cgroup_pidlist *l;
3382 enum cgroup_filetype type = seq_cft(s)->private;
3383 int index = 0, pid = *pos;
3386 mutex_lock(&cgrp->pidlist_mutex);
3389 * !NULL @of->priv indicates that this isn't the first start()
3390 * after open. If the matching pidlist is around, we can use that.
3391 * Look for it. Note that @of->priv can't be used directly. It
3392 * could already have been destroyed.
3395 of->priv = cgroup_pidlist_find(cgrp, type);
3398 * Either this is the first start() after open or the matching
3399 * pidlist has been destroyed inbetween. Create a new one.
3402 ret = pidlist_array_load(cgrp, type,
3403 (struct cgroup_pidlist **)&of->priv);
3405 return ERR_PTR(ret);
3410 int end = l->length;
3412 while (index < end) {
3413 int mid = (index + end) / 2;
3414 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3417 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3423 /* If we're off the end of the array, we're done */
3424 if (index >= l->length)
3426 /* Update the abstract position to be the actual pid that we found */
3427 iter = l->list + index;
3428 *pos = cgroup_pid_fry(cgrp, *iter);
3432 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3434 struct kernfs_open_file *of = s->private;
3435 struct cgroup_pidlist *l = of->priv;
3438 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3439 CGROUP_PIDLIST_DESTROY_DELAY);
3440 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3443 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3445 struct kernfs_open_file *of = s->private;
3446 struct cgroup_pidlist *l = of->priv;
3448 pid_t *end = l->list + l->length;
3450 * Advance to the next pid in the array. If this goes off the
3457 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3462 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3464 return seq_printf(s, "%d\n", *(int *)v);
3468 * seq_operations functions for iterating on pidlists through seq_file -
3469 * independent of whether it's tasks or procs
3471 static const struct seq_operations cgroup_pidlist_seq_operations = {
3472 .start = cgroup_pidlist_start,
3473 .stop = cgroup_pidlist_stop,
3474 .next = cgroup_pidlist_next,
3475 .show = cgroup_pidlist_show,
3478 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3481 return notify_on_release(css->cgroup);
3484 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3485 struct cftype *cft, u64 val)
3487 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3489 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3491 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3495 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3498 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3501 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3502 struct cftype *cft, u64 val)
3505 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3507 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3511 static struct cftype cgroup_base_files[] = {
3513 .name = "cgroup.procs",
3514 .seq_start = cgroup_pidlist_start,
3515 .seq_next = cgroup_pidlist_next,
3516 .seq_stop = cgroup_pidlist_stop,
3517 .seq_show = cgroup_pidlist_show,
3518 .private = CGROUP_FILE_PROCS,
3519 .write_u64 = cgroup_procs_write,
3520 .mode = S_IRUGO | S_IWUSR,
3523 .name = "cgroup.clone_children",
3524 .flags = CFTYPE_INSANE,
3525 .read_u64 = cgroup_clone_children_read,
3526 .write_u64 = cgroup_clone_children_write,
3529 .name = "cgroup.sane_behavior",
3530 .flags = CFTYPE_ONLY_ON_ROOT,
3531 .seq_show = cgroup_sane_behavior_show,
3535 * Historical crazy stuff. These don't have "cgroup." prefix and
3536 * don't exist if sane_behavior. If you're depending on these, be
3537 * prepared to be burned.
3541 .flags = CFTYPE_INSANE, /* use "procs" instead */
3542 .seq_start = cgroup_pidlist_start,
3543 .seq_next = cgroup_pidlist_next,
3544 .seq_stop = cgroup_pidlist_stop,
3545 .seq_show = cgroup_pidlist_show,
3546 .private = CGROUP_FILE_TASKS,
3547 .write_u64 = cgroup_tasks_write,
3548 .mode = S_IRUGO | S_IWUSR,
3551 .name = "notify_on_release",
3552 .flags = CFTYPE_INSANE,
3553 .read_u64 = cgroup_read_notify_on_release,
3554 .write_u64 = cgroup_write_notify_on_release,
3557 .name = "release_agent",
3558 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3559 .seq_show = cgroup_release_agent_show,
3560 .write_string = cgroup_release_agent_write,
3561 .max_write_len = PATH_MAX - 1,
3567 * cgroup_populate_dir - create subsys files in a cgroup directory
3568 * @cgrp: target cgroup
3569 * @subsys_mask: mask of the subsystem ids whose files should be added
3571 * On failure, no file is added.
3573 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3575 struct cgroup_subsys *ss;
3578 /* process cftsets of each subsystem */
3579 for_each_subsys(ss, i) {
3580 struct cftype *cfts;
3582 if (!test_bit(i, &subsys_mask))
3585 list_for_each_entry(cfts, &ss->cfts, node) {
3586 ret = cgroup_addrm_files(cgrp, cfts, true);
3593 cgroup_clear_dir(cgrp, subsys_mask);
3598 * css destruction is four-stage process.
3600 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3601 * Implemented in kill_css().
3603 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3604 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3605 * by invoking offline_css(). After offlining, the base ref is put.
3606 * Implemented in css_killed_work_fn().
3608 * 3. When the percpu_ref reaches zero, the only possible remaining
3609 * accessors are inside RCU read sections. css_release() schedules the
3612 * 4. After the grace period, the css can be freed. Implemented in
3613 * css_free_work_fn().
3615 * It is actually hairier because both step 2 and 4 require process context
3616 * and thus involve punting to css->destroy_work adding two additional
3617 * steps to the already complex sequence.
3619 static void css_free_work_fn(struct work_struct *work)
3621 struct cgroup_subsys_state *css =
3622 container_of(work, struct cgroup_subsys_state, destroy_work);
3623 struct cgroup *cgrp = css->cgroup;
3626 css_put(css->parent);
3628 css->ss->css_free(css);
3632 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3634 struct cgroup_subsys_state *css =
3635 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3637 INIT_WORK(&css->destroy_work, css_free_work_fn);
3638 queue_work(cgroup_destroy_wq, &css->destroy_work);
3641 static void css_release(struct percpu_ref *ref)
3643 struct cgroup_subsys_state *css =
3644 container_of(ref, struct cgroup_subsys_state, refcnt);
3646 RCU_INIT_POINTER(css->cgroup->subsys[css->ss->id], NULL);
3647 call_rcu(&css->rcu_head, css_free_rcu_fn);
3650 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3651 struct cgroup *cgrp)
3658 css->parent = cgroup_css(cgrp->parent, ss);
3660 css->flags |= CSS_ROOT;
3662 BUG_ON(cgroup_css(cgrp, ss));
3665 /* invoke ->css_online() on a new CSS and mark it online if successful */
3666 static int online_css(struct cgroup_subsys_state *css)
3668 struct cgroup_subsys *ss = css->ss;
3671 lockdep_assert_held(&cgroup_tree_mutex);
3672 lockdep_assert_held(&cgroup_mutex);
3675 ret = ss->css_online(css);
3677 css->flags |= CSS_ONLINE;
3678 css->cgroup->nr_css++;
3679 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
3684 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
3685 static void offline_css(struct cgroup_subsys_state *css)
3687 struct cgroup_subsys *ss = css->ss;
3689 lockdep_assert_held(&cgroup_tree_mutex);
3690 lockdep_assert_held(&cgroup_mutex);
3692 if (!(css->flags & CSS_ONLINE))
3695 if (ss->css_offline)
3696 ss->css_offline(css);
3698 css->flags &= ~CSS_ONLINE;
3699 css->cgroup->nr_css--;
3700 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
3704 * create_css - create a cgroup_subsys_state
3705 * @cgrp: the cgroup new css will be associated with
3706 * @ss: the subsys of new css
3708 * Create a new css associated with @cgrp - @ss pair. On success, the new
3709 * css is online and installed in @cgrp with all interface files created.
3710 * Returns 0 on success, -errno on failure.
3712 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
3714 struct cgroup *parent = cgrp->parent;
3715 struct cgroup_subsys_state *css;
3718 lockdep_assert_held(&cgroup_mutex);
3720 css = ss->css_alloc(cgroup_css(parent, ss));
3722 return PTR_ERR(css);
3724 err = percpu_ref_init(&css->refcnt, css_release);
3728 init_css(css, ss, cgrp);
3730 err = cgroup_populate_dir(cgrp, 1 << ss->id);
3732 goto err_free_percpu_ref;
3734 err = online_css(css);
3739 css_get(css->parent);
3741 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
3743 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",
3744 current->comm, current->pid, ss->name);
3745 if (!strcmp(ss->name, "memory"))
3746 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
3747 ss->warned_broken_hierarchy = true;
3753 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3754 err_free_percpu_ref:
3755 percpu_ref_cancel_init(&css->refcnt);
3762 * cgroup_create - create a cgroup
3763 * @parent: cgroup that will be parent of the new cgroup
3764 * @name: name of the new cgroup
3765 * @mode: mode to set on new cgroup
3767 static long cgroup_create(struct cgroup *parent, const char *name,
3770 struct cgroup *cgrp;
3771 struct cgroup_root *root = parent->root;
3773 struct cgroup_subsys *ss;
3774 struct kernfs_node *kn;
3777 * XXX: The default hierarchy isn't fully implemented yet. Block
3778 * !root cgroup creation on it for now.
3780 if (root == &cgrp_dfl_root)
3783 /* allocate the cgroup and its ID, 0 is reserved for the root */
3784 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3788 mutex_lock(&cgroup_tree_mutex);
3791 * Only live parents can have children. Note that the liveliness
3792 * check isn't strictly necessary because cgroup_mkdir() and
3793 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
3794 * anyway so that locking is contained inside cgroup proper and we
3795 * don't get nasty surprises if we ever grow another caller.
3797 if (!cgroup_lock_live_group(parent)) {
3799 goto err_unlock_tree;
3803 * Temporarily set the pointer to NULL, so idr_find() won't return
3804 * a half-baked cgroup.
3806 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
3812 init_cgroup_housekeeping(cgrp);
3814 cgrp->parent = parent;
3815 cgrp->dummy_css.parent = &parent->dummy_css;
3816 cgrp->root = parent->root;
3818 if (notify_on_release(parent))
3819 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3821 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
3822 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
3824 /* create the directory */
3825 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
3833 * This extra ref will be put in cgroup_free_fn() and guarantees
3834 * that @cgrp->kn is always accessible.
3838 cgrp->serial_nr = cgroup_serial_nr_next++;
3840 /* allocation complete, commit to creation */
3841 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
3842 atomic_inc(&root->nr_cgrps);
3846 * @cgrp is now fully operational. If something fails after this
3847 * point, it'll be released via the normal destruction path.
3849 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
3851 err = cgroup_kn_set_ugid(kn);
3855 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
3859 /* let's create and online css's */
3860 for_each_subsys(ss, ssid) {
3861 if (parent->child_subsys_mask & (1 << ssid)) {
3862 err = create_css(cgrp, ss);
3868 cgrp->child_subsys_mask = parent->child_subsys_mask;
3870 kernfs_activate(kn);
3872 mutex_unlock(&cgroup_mutex);
3873 mutex_unlock(&cgroup_tree_mutex);
3878 idr_remove(&root->cgroup_idr, cgrp->id);
3880 mutex_unlock(&cgroup_mutex);
3882 mutex_unlock(&cgroup_tree_mutex);
3887 cgroup_destroy_locked(cgrp);
3888 mutex_unlock(&cgroup_mutex);
3889 mutex_unlock(&cgroup_tree_mutex);
3893 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3896 struct cgroup *parent = parent_kn->priv;
3900 * cgroup_create() grabs cgroup_tree_mutex which nests outside
3901 * kernfs active_ref and cgroup_create() already synchronizes
3902 * properly against removal through cgroup_lock_live_group().
3903 * Break it before calling cgroup_create().
3906 kernfs_break_active_protection(parent_kn);
3908 ret = cgroup_create(parent, name, mode);
3910 kernfs_unbreak_active_protection(parent_kn);
3916 * This is called when the refcnt of a css is confirmed to be killed.
3917 * css_tryget() is now guaranteed to fail.
3919 static void css_killed_work_fn(struct work_struct *work)
3921 struct cgroup_subsys_state *css =
3922 container_of(work, struct cgroup_subsys_state, destroy_work);
3923 struct cgroup *cgrp = css->cgroup;
3925 mutex_lock(&cgroup_tree_mutex);
3926 mutex_lock(&cgroup_mutex);
3929 * css_tryget() is guaranteed to fail now. Tell subsystems to
3930 * initate destruction.
3935 * If @cgrp is marked dead, it's waiting for refs of all css's to
3936 * be disabled before proceeding to the second phase of cgroup
3937 * destruction. If we are the last one, kick it off.
3939 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
3940 cgroup_destroy_css_killed(cgrp);
3942 mutex_unlock(&cgroup_mutex);
3943 mutex_unlock(&cgroup_tree_mutex);
3946 * Put the css refs from kill_css(). Each css holds an extra
3947 * reference to the cgroup's dentry and cgroup removal proceeds
3948 * regardless of css refs. On the last put of each css, whenever
3949 * that may be, the extra dentry ref is put so that dentry
3950 * destruction happens only after all css's are released.
3955 /* css kill confirmation processing requires process context, bounce */
3956 static void css_killed_ref_fn(struct percpu_ref *ref)
3958 struct cgroup_subsys_state *css =
3959 container_of(ref, struct cgroup_subsys_state, refcnt);
3961 INIT_WORK(&css->destroy_work, css_killed_work_fn);
3962 queue_work(cgroup_destroy_wq, &css->destroy_work);
3966 * kill_css - destroy a css
3967 * @css: css to destroy
3969 * This function initiates destruction of @css by removing cgroup interface
3970 * files and putting its base reference. ->css_offline() will be invoked
3971 * asynchronously once css_tryget() is guaranteed to fail and when the
3972 * reference count reaches zero, @css will be released.
3974 static void kill_css(struct cgroup_subsys_state *css)
3976 lockdep_assert_held(&cgroup_tree_mutex);
3979 * This must happen before css is disassociated with its cgroup.
3980 * See seq_css() for details.
3982 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3985 * Killing would put the base ref, but we need to keep it alive
3986 * until after ->css_offline().
3991 * cgroup core guarantees that, by the time ->css_offline() is
3992 * invoked, no new css reference will be given out via
3993 * css_tryget(). We can't simply call percpu_ref_kill() and
3994 * proceed to offlining css's because percpu_ref_kill() doesn't
3995 * guarantee that the ref is seen as killed on all CPUs on return.
3997 * Use percpu_ref_kill_and_confirm() to get notifications as each
3998 * css is confirmed to be seen as killed on all CPUs.
4000 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4004 * cgroup_destroy_locked - the first stage of cgroup destruction
4005 * @cgrp: cgroup to be destroyed
4007 * css's make use of percpu refcnts whose killing latency shouldn't be
4008 * exposed to userland and are RCU protected. Also, cgroup core needs to
4009 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
4010 * invoked. To satisfy all the requirements, destruction is implemented in
4011 * the following two steps.
4013 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4014 * userland visible parts and start killing the percpu refcnts of
4015 * css's. Set up so that the next stage will be kicked off once all
4016 * the percpu refcnts are confirmed to be killed.
4018 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4019 * rest of destruction. Once all cgroup references are gone, the
4020 * cgroup is RCU-freed.
4022 * This function implements s1. After this step, @cgrp is gone as far as
4023 * the userland is concerned and a new cgroup with the same name may be
4024 * created. As cgroup doesn't care about the names internally, this
4025 * doesn't cause any problem.
4027 static int cgroup_destroy_locked(struct cgroup *cgrp)
4028 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4030 struct cgroup *child;
4031 struct cgroup_subsys_state *css;
4035 lockdep_assert_held(&cgroup_tree_mutex);
4036 lockdep_assert_held(&cgroup_mutex);
4039 * css_set_rwsem synchronizes access to ->cset_links and prevents
4040 * @cgrp from being removed while put_css_set() is in progress.
4042 down_read(&css_set_rwsem);
4043 empty = list_empty(&cgrp->cset_links);
4044 up_read(&css_set_rwsem);
4049 * Make sure there's no live children. We can't test ->children
4050 * emptiness as dead children linger on it while being destroyed;
4051 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
4055 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
4056 empty = cgroup_is_dead(child);
4065 * Mark @cgrp dead. This prevents further task migration and child
4066 * creation by disabling cgroup_lock_live_group(). Note that
4067 * CGRP_DEAD assertion is depended upon by css_next_child() to
4068 * resume iteration after dropping RCU read lock. See
4069 * css_next_child() for details.
4071 set_bit(CGRP_DEAD, &cgrp->flags);
4074 * Initiate massacre of all css's. cgroup_destroy_css_killed()
4075 * will be invoked to perform the rest of destruction once the
4076 * percpu refs of all css's are confirmed to be killed. This
4077 * involves removing the subsystem's files, drop cgroup_mutex.
4079 mutex_unlock(&cgroup_mutex);
4080 for_each_css(css, ssid, cgrp)
4082 mutex_lock(&cgroup_mutex);
4084 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4085 raw_spin_lock(&release_list_lock);
4086 if (!list_empty(&cgrp->release_list))
4087 list_del_init(&cgrp->release_list);
4088 raw_spin_unlock(&release_list_lock);
4091 * If @cgrp has css's attached, the second stage of cgroup
4092 * destruction is kicked off from css_killed_work_fn() after the
4093 * refs of all attached css's are killed. If @cgrp doesn't have
4094 * any css, we kick it off here.
4097 cgroup_destroy_css_killed(cgrp);
4099 /* remove @cgrp directory along with the base files */
4100 mutex_unlock(&cgroup_mutex);
4103 * There are two control paths which try to determine cgroup from
4104 * dentry without going through kernfs - cgroupstats_build() and
4105 * css_tryget_from_dir(). Those are supported by RCU protecting
4106 * clearing of cgrp->kn->priv backpointer, which should happen
4107 * after all files under it have been removed.
4109 kernfs_remove(cgrp->kn); /* @cgrp has an extra ref on its kn */
4110 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4112 mutex_lock(&cgroup_mutex);
4118 * cgroup_destroy_css_killed - the second step of cgroup destruction
4119 * @work: cgroup->destroy_free_work
4121 * This function is invoked from a work item for a cgroup which is being
4122 * destroyed after all css's are offlined and performs the rest of
4123 * destruction. This is the second step of destruction described in the
4124 * comment above cgroup_destroy_locked().
4126 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
4128 struct cgroup *parent = cgrp->parent;
4130 lockdep_assert_held(&cgroup_tree_mutex);
4131 lockdep_assert_held(&cgroup_mutex);
4133 /* delete this cgroup from parent->children */
4134 list_del_rcu(&cgrp->sibling);
4138 set_bit(CGRP_RELEASABLE, &parent->flags);
4139 check_for_release(parent);
4142 static int cgroup_rmdir(struct kernfs_node *kn)
4144 struct cgroup *cgrp = kn->priv;
4148 * This is self-destruction but @kn can't be removed while this
4149 * callback is in progress. Let's break active protection. Once
4150 * the protection is broken, @cgrp can be destroyed at any point.
4151 * Pin it so that it stays accessible.
4154 kernfs_break_active_protection(kn);
4156 mutex_lock(&cgroup_tree_mutex);
4157 mutex_lock(&cgroup_mutex);
4160 * @cgrp might already have been destroyed while we're trying to
4163 if (!cgroup_is_dead(cgrp))
4164 ret = cgroup_destroy_locked(cgrp);
4166 mutex_unlock(&cgroup_mutex);
4167 mutex_unlock(&cgroup_tree_mutex);
4169 kernfs_unbreak_active_protection(kn);
4174 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4175 .remount_fs = cgroup_remount,
4176 .show_options = cgroup_show_options,
4177 .mkdir = cgroup_mkdir,
4178 .rmdir = cgroup_rmdir,
4179 .rename = cgroup_rename,
4182 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4184 struct cgroup_subsys_state *css;
4186 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4188 mutex_lock(&cgroup_tree_mutex);
4189 mutex_lock(&cgroup_mutex);
4191 INIT_LIST_HEAD(&ss->cfts);
4193 /* Create the root cgroup state for this subsystem */
4194 ss->root = &cgrp_dfl_root;
4195 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4196 /* We don't handle early failures gracefully */
4197 BUG_ON(IS_ERR(css));
4198 init_css(css, ss, &cgrp_dfl_root.cgrp);
4200 /* Update the init_css_set to contain a subsys
4201 * pointer to this state - since the subsystem is
4202 * newly registered, all tasks and hence the
4203 * init_css_set is in the subsystem's root cgroup. */
4204 init_css_set.subsys[ss->id] = css;
4206 need_forkexit_callback |= ss->fork || ss->exit;
4208 /* At system boot, before all subsystems have been
4209 * registered, no tasks have been forked, so we don't
4210 * need to invoke fork callbacks here. */
4211 BUG_ON(!list_empty(&init_task.tasks));
4213 BUG_ON(online_css(css));
4215 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
4217 mutex_unlock(&cgroup_mutex);
4218 mutex_unlock(&cgroup_tree_mutex);
4222 * cgroup_init_early - cgroup initialization at system boot
4224 * Initialize cgroups at system boot, and initialize any
4225 * subsystems that request early init.
4227 int __init cgroup_init_early(void)
4229 static struct cgroup_sb_opts __initdata opts =
4230 { .flags = CGRP_ROOT_SANE_BEHAVIOR };
4231 struct cgroup_subsys *ss;
4234 init_cgroup_root(&cgrp_dfl_root, &opts);
4235 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4237 for_each_subsys(ss, i) {
4238 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4239 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4240 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4242 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4243 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4246 ss->name = cgroup_subsys_name[i];
4249 cgroup_init_subsys(ss);
4255 * cgroup_init - cgroup initialization
4257 * Register cgroup filesystem and /proc file, and initialize
4258 * any subsystems that didn't request early init.
4260 int __init cgroup_init(void)
4262 struct cgroup_subsys *ss;
4266 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4268 mutex_lock(&cgroup_tree_mutex);
4269 mutex_lock(&cgroup_mutex);
4271 /* Add init_css_set to the hash table */
4272 key = css_set_hash(init_css_set.subsys);
4273 hash_add(css_set_table, &init_css_set.hlist, key);
4275 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4277 mutex_unlock(&cgroup_mutex);
4278 mutex_unlock(&cgroup_tree_mutex);
4280 for_each_subsys(ss, ssid) {
4281 if (!ss->early_init)
4282 cgroup_init_subsys(ss);
4284 list_add_tail(&init_css_set.e_cset_node[ssid],
4285 &cgrp_dfl_root.cgrp.e_csets[ssid]);
4288 * cftype registration needs kmalloc and can't be done
4289 * during early_init. Register base cftypes separately.
4291 if (ss->base_cftypes)
4292 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4295 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4299 err = register_filesystem(&cgroup_fs_type);
4301 kobject_put(cgroup_kobj);
4305 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4309 static int __init cgroup_wq_init(void)
4312 * There isn't much point in executing destruction path in
4313 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4314 * Use 1 for @max_active.
4316 * We would prefer to do this in cgroup_init() above, but that
4317 * is called before init_workqueues(): so leave this until after.
4319 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
4320 BUG_ON(!cgroup_destroy_wq);
4323 * Used to destroy pidlists and separate to serve as flush domain.
4324 * Cap @max_active to 1 too.
4326 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4328 BUG_ON(!cgroup_pidlist_destroy_wq);
4332 core_initcall(cgroup_wq_init);
4335 * proc_cgroup_show()
4336 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4337 * - Used for /proc/<pid>/cgroup.
4340 /* TODO: Use a proper seq_file iterator */
4341 int proc_cgroup_show(struct seq_file *m, void *v)
4344 struct task_struct *tsk;
4347 struct cgroup_root *root;
4350 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4356 tsk = get_pid_task(pid, PIDTYPE_PID);
4362 mutex_lock(&cgroup_mutex);
4363 down_read(&css_set_rwsem);
4365 for_each_root(root) {
4366 struct cgroup_subsys *ss;
4367 struct cgroup *cgrp;
4368 int ssid, count = 0;
4370 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
4373 seq_printf(m, "%d:", root->hierarchy_id);
4374 for_each_subsys(ss, ssid)
4375 if (root->subsys_mask & (1 << ssid))
4376 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4377 if (strlen(root->name))
4378 seq_printf(m, "%sname=%s", count ? "," : "",
4381 cgrp = task_cgroup_from_root(tsk, root);
4382 path = cgroup_path(cgrp, buf, PATH_MAX);
4384 retval = -ENAMETOOLONG;
4392 up_read(&css_set_rwsem);
4393 mutex_unlock(&cgroup_mutex);
4394 put_task_struct(tsk);
4401 /* Display information about each subsystem and each hierarchy */
4402 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4404 struct cgroup_subsys *ss;
4407 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4409 * ideally we don't want subsystems moving around while we do this.
4410 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4411 * subsys/hierarchy state.
4413 mutex_lock(&cgroup_mutex);
4415 for_each_subsys(ss, i)
4416 seq_printf(m, "%s\t%d\t%d\t%d\n",
4417 ss->name, ss->root->hierarchy_id,
4418 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4420 mutex_unlock(&cgroup_mutex);
4424 static int cgroupstats_open(struct inode *inode, struct file *file)
4426 return single_open(file, proc_cgroupstats_show, NULL);
4429 static const struct file_operations proc_cgroupstats_operations = {
4430 .open = cgroupstats_open,
4432 .llseek = seq_lseek,
4433 .release = single_release,
4437 * cgroup_fork - initialize cgroup related fields during copy_process()
4438 * @child: pointer to task_struct of forking parent process.
4440 * A task is associated with the init_css_set until cgroup_post_fork()
4441 * attaches it to the parent's css_set. Empty cg_list indicates that
4442 * @child isn't holding reference to its css_set.
4444 void cgroup_fork(struct task_struct *child)
4446 RCU_INIT_POINTER(child->cgroups, &init_css_set);
4447 INIT_LIST_HEAD(&child->cg_list);
4451 * cgroup_post_fork - called on a new task after adding it to the task list
4452 * @child: the task in question
4454 * Adds the task to the list running through its css_set if necessary and
4455 * call the subsystem fork() callbacks. Has to be after the task is
4456 * visible on the task list in case we race with the first call to
4457 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4460 void cgroup_post_fork(struct task_struct *child)
4462 struct cgroup_subsys *ss;
4466 * This may race against cgroup_enable_task_cg_links(). As that
4467 * function sets use_task_css_set_links before grabbing
4468 * tasklist_lock and we just went through tasklist_lock to add
4469 * @child, it's guaranteed that either we see the set
4470 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
4471 * @child during its iteration.
4473 * If we won the race, @child is associated with %current's
4474 * css_set. Grabbing css_set_rwsem guarantees both that the
4475 * association is stable, and, on completion of the parent's
4476 * migration, @child is visible in the source of migration or
4477 * already in the destination cgroup. This guarantee is necessary
4478 * when implementing operations which need to migrate all tasks of
4479 * a cgroup to another.
4481 * Note that if we lose to cgroup_enable_task_cg_links(), @child
4482 * will remain in init_css_set. This is safe because all tasks are
4483 * in the init_css_set before cg_links is enabled and there's no
4484 * operation which transfers all tasks out of init_css_set.
4486 if (use_task_css_set_links) {
4487 struct css_set *cset;
4489 down_write(&css_set_rwsem);
4490 cset = task_css_set(current);
4491 if (list_empty(&child->cg_list)) {
4492 rcu_assign_pointer(child->cgroups, cset);
4493 list_add(&child->cg_list, &cset->tasks);
4496 up_write(&css_set_rwsem);
4500 * Call ss->fork(). This must happen after @child is linked on
4501 * css_set; otherwise, @child might change state between ->fork()
4502 * and addition to css_set.
4504 if (need_forkexit_callback) {
4505 for_each_subsys(ss, i)
4512 * cgroup_exit - detach cgroup from exiting task
4513 * @tsk: pointer to task_struct of exiting process
4515 * Description: Detach cgroup from @tsk and release it.
4517 * Note that cgroups marked notify_on_release force every task in
4518 * them to take the global cgroup_mutex mutex when exiting.
4519 * This could impact scaling on very large systems. Be reluctant to
4520 * use notify_on_release cgroups where very high task exit scaling
4521 * is required on large systems.
4523 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
4524 * call cgroup_exit() while the task is still competent to handle
4525 * notify_on_release(), then leave the task attached to the root cgroup in
4526 * each hierarchy for the remainder of its exit. No need to bother with
4527 * init_css_set refcnting. init_css_set never goes away and we can't race
4528 * with migration path - PF_EXITING is visible to migration path.
4530 void cgroup_exit(struct task_struct *tsk)
4532 struct cgroup_subsys *ss;
4533 struct css_set *cset;
4534 bool put_cset = false;
4538 * Unlink from @tsk from its css_set. As migration path can't race
4539 * with us, we can check cg_list without grabbing css_set_rwsem.
4541 if (!list_empty(&tsk->cg_list)) {
4542 down_write(&css_set_rwsem);
4543 list_del_init(&tsk->cg_list);
4544 up_write(&css_set_rwsem);
4548 /* Reassign the task to the init_css_set. */
4549 cset = task_css_set(tsk);
4550 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4552 if (need_forkexit_callback) {
4553 /* see cgroup_post_fork() for details */
4554 for_each_subsys(ss, i) {
4556 struct cgroup_subsys_state *old_css = cset->subsys[i];
4557 struct cgroup_subsys_state *css = task_css(tsk, i);
4559 ss->exit(css, old_css, tsk);
4565 put_css_set(cset, true);
4568 static void check_for_release(struct cgroup *cgrp)
4570 if (cgroup_is_releasable(cgrp) &&
4571 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4573 * Control Group is currently removeable. If it's not
4574 * already queued for a userspace notification, queue
4577 int need_schedule_work = 0;
4579 raw_spin_lock(&release_list_lock);
4580 if (!cgroup_is_dead(cgrp) &&
4581 list_empty(&cgrp->release_list)) {
4582 list_add(&cgrp->release_list, &release_list);
4583 need_schedule_work = 1;
4585 raw_spin_unlock(&release_list_lock);
4586 if (need_schedule_work)
4587 schedule_work(&release_agent_work);
4592 * Notify userspace when a cgroup is released, by running the
4593 * configured release agent with the name of the cgroup (path
4594 * relative to the root of cgroup file system) as the argument.
4596 * Most likely, this user command will try to rmdir this cgroup.
4598 * This races with the possibility that some other task will be
4599 * attached to this cgroup before it is removed, or that some other
4600 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4601 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4602 * unused, and this cgroup will be reprieved from its death sentence,
4603 * to continue to serve a useful existence. Next time it's released,
4604 * we will get notified again, if it still has 'notify_on_release' set.
4606 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4607 * means only wait until the task is successfully execve()'d. The
4608 * separate release agent task is forked by call_usermodehelper(),
4609 * then control in this thread returns here, without waiting for the
4610 * release agent task. We don't bother to wait because the caller of
4611 * this routine has no use for the exit status of the release agent
4612 * task, so no sense holding our caller up for that.
4614 static void cgroup_release_agent(struct work_struct *work)
4616 BUG_ON(work != &release_agent_work);
4617 mutex_lock(&cgroup_mutex);
4618 raw_spin_lock(&release_list_lock);
4619 while (!list_empty(&release_list)) {
4620 char *argv[3], *envp[3];
4622 char *pathbuf = NULL, *agentbuf = NULL, *path;
4623 struct cgroup *cgrp = list_entry(release_list.next,
4626 list_del_init(&cgrp->release_list);
4627 raw_spin_unlock(&release_list_lock);
4628 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
4631 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
4634 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4639 argv[i++] = agentbuf;
4644 /* minimal command environment */
4645 envp[i++] = "HOME=/";
4646 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4649 /* Drop the lock while we invoke the usermode helper,
4650 * since the exec could involve hitting disk and hence
4651 * be a slow process */
4652 mutex_unlock(&cgroup_mutex);
4653 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4654 mutex_lock(&cgroup_mutex);
4658 raw_spin_lock(&release_list_lock);
4660 raw_spin_unlock(&release_list_lock);
4661 mutex_unlock(&cgroup_mutex);
4664 static int __init cgroup_disable(char *str)
4666 struct cgroup_subsys *ss;
4670 while ((token = strsep(&str, ",")) != NULL) {
4674 for_each_subsys(ss, i) {
4675 if (!strcmp(token, ss->name)) {
4677 printk(KERN_INFO "Disabling %s control group"
4678 " subsystem\n", ss->name);
4685 __setup("cgroup_disable=", cgroup_disable);
4688 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
4689 * @dentry: directory dentry of interest
4690 * @ss: subsystem of interest
4692 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
4693 * to get the corresponding css and return it. If such css doesn't exist
4694 * or can't be pinned, an ERR_PTR value is returned.
4696 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
4697 struct cgroup_subsys *ss)
4699 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4700 struct cgroup_subsys_state *css = NULL;
4701 struct cgroup *cgrp;
4703 /* is @dentry a cgroup dir? */
4704 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4705 kernfs_type(kn) != KERNFS_DIR)
4706 return ERR_PTR(-EBADF);
4711 * This path doesn't originate from kernfs and @kn could already
4712 * have been or be removed at any point. @kn->priv is RCU
4713 * protected for this access. See destroy_locked() for details.
4715 cgrp = rcu_dereference(kn->priv);
4717 css = cgroup_css(cgrp, ss);
4719 if (!css || !css_tryget(css))
4720 css = ERR_PTR(-ENOENT);
4727 * css_from_id - lookup css by id
4728 * @id: the cgroup id
4729 * @ss: cgroup subsys to be looked into
4731 * Returns the css if there's valid one with @id, otherwise returns NULL.
4732 * Should be called under rcu_read_lock().
4734 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
4736 struct cgroup *cgrp;
4738 cgroup_assert_mutexes_or_rcu_locked();
4740 cgrp = idr_find(&ss->root->cgroup_idr, id);
4742 return cgroup_css(cgrp, ss);
4746 #ifdef CONFIG_CGROUP_DEBUG
4747 static struct cgroup_subsys_state *
4748 debug_css_alloc(struct cgroup_subsys_state *parent_css)
4750 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4753 return ERR_PTR(-ENOMEM);
4758 static void debug_css_free(struct cgroup_subsys_state *css)
4763 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
4766 return cgroup_task_count(css->cgroup);
4769 static u64 current_css_set_read(struct cgroup_subsys_state *css,
4772 return (u64)(unsigned long)current->cgroups;
4775 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
4781 count = atomic_read(&task_css_set(current)->refcount);
4786 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
4788 struct cgrp_cset_link *link;
4789 struct css_set *cset;
4792 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
4796 down_read(&css_set_rwsem);
4798 cset = rcu_dereference(current->cgroups);
4799 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
4800 struct cgroup *c = link->cgrp;
4802 cgroup_name(c, name_buf, NAME_MAX + 1);
4803 seq_printf(seq, "Root %d group %s\n",
4804 c->root->hierarchy_id, name_buf);
4807 up_read(&css_set_rwsem);
4812 #define MAX_TASKS_SHOWN_PER_CSS 25
4813 static int cgroup_css_links_read(struct seq_file *seq, void *v)
4815 struct cgroup_subsys_state *css = seq_css(seq);
4816 struct cgrp_cset_link *link;
4818 down_read(&css_set_rwsem);
4819 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
4820 struct css_set *cset = link->cset;
4821 struct task_struct *task;
4824 seq_printf(seq, "css_set %p\n", cset);
4826 list_for_each_entry(task, &cset->tasks, cg_list) {
4827 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
4829 seq_printf(seq, " task %d\n", task_pid_vnr(task));
4832 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
4833 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
4835 seq_printf(seq, " task %d\n", task_pid_vnr(task));
4839 seq_puts(seq, " ...\n");
4841 up_read(&css_set_rwsem);
4845 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
4847 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4850 static struct cftype debug_files[] = {
4852 .name = "taskcount",
4853 .read_u64 = debug_taskcount_read,
4857 .name = "current_css_set",
4858 .read_u64 = current_css_set_read,
4862 .name = "current_css_set_refcount",
4863 .read_u64 = current_css_set_refcount_read,
4867 .name = "current_css_set_cg_links",
4868 .seq_show = current_css_set_cg_links_read,
4872 .name = "cgroup_css_links",
4873 .seq_show = cgroup_css_links_read,
4877 .name = "releasable",
4878 .read_u64 = releasable_read,
4884 struct cgroup_subsys debug_cgrp_subsys = {
4885 .css_alloc = debug_css_alloc,
4886 .css_free = debug_css_free,
4887 .base_cftypes = debug_files,
4889 #endif /* CONFIG_CGROUP_DEBUG */