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;
211 /* convenient tests for these bits */
212 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
214 return test_bit(CGRP_DEAD, &cgrp->flags);
217 struct cgroup_subsys_state *seq_css(struct seq_file *seq)
219 struct kernfs_open_file *of = seq->private;
220 struct cgroup *cgrp = of->kn->parent->priv;
221 struct cftype *cft = seq_cft(seq);
224 * This is open and unprotected implementation of cgroup_css().
225 * seq_css() is only called from a kernfs file operation which has
226 * an active reference on the file. Because all the subsystem
227 * files are drained before a css is disassociated with a cgroup,
228 * the matching css from the cgroup's subsys table is guaranteed to
229 * be and stay valid until the enclosing operation is complete.
232 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
234 return &cgrp->dummy_css;
236 EXPORT_SYMBOL_GPL(seq_css);
239 * cgroup_is_descendant - test ancestry
240 * @cgrp: the cgroup to be tested
241 * @ancestor: possible ancestor of @cgrp
243 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
244 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
245 * and @ancestor are accessible.
247 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
250 if (cgrp == ancestor)
257 static int cgroup_is_releasable(const struct cgroup *cgrp)
260 (1 << CGRP_RELEASABLE) |
261 (1 << CGRP_NOTIFY_ON_RELEASE);
262 return (cgrp->flags & bits) == bits;
265 static int notify_on_release(const struct cgroup *cgrp)
267 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
271 * for_each_css - iterate all css's of a cgroup
272 * @css: the iteration cursor
273 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
274 * @cgrp: the target cgroup to iterate css's of
276 * Should be called under cgroup_mutex.
278 #define for_each_css(css, ssid, cgrp) \
279 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
280 if (!((css) = rcu_dereference_check( \
281 (cgrp)->subsys[(ssid)], \
282 lockdep_is_held(&cgroup_tree_mutex) || \
283 lockdep_is_held(&cgroup_mutex)))) { } \
287 * for_each_subsys - iterate all enabled cgroup subsystems
288 * @ss: the iteration cursor
289 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
291 #define for_each_subsys(ss, ssid) \
292 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
293 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
295 /* iterate across the hierarchies */
296 #define for_each_root(root) \
297 list_for_each_entry((root), &cgroup_roots, root_list)
300 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
301 * @cgrp: the cgroup to be checked for liveness
303 * On success, returns true; the mutex should be later unlocked. On
304 * failure returns false with no lock held.
306 static bool cgroup_lock_live_group(struct cgroup *cgrp)
308 mutex_lock(&cgroup_mutex);
309 if (cgroup_is_dead(cgrp)) {
310 mutex_unlock(&cgroup_mutex);
316 /* the list of cgroups eligible for automatic release. Protected by
317 * release_list_lock */
318 static LIST_HEAD(release_list);
319 static DEFINE_RAW_SPINLOCK(release_list_lock);
320 static void cgroup_release_agent(struct work_struct *work);
321 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
322 static void check_for_release(struct cgroup *cgrp);
325 * A cgroup can be associated with multiple css_sets as different tasks may
326 * belong to different cgroups on different hierarchies. In the other
327 * direction, a css_set is naturally associated with multiple cgroups.
328 * This M:N relationship is represented by the following link structure
329 * which exists for each association and allows traversing the associations
332 struct cgrp_cset_link {
333 /* the cgroup and css_set this link associates */
335 struct css_set *cset;
337 /* list of cgrp_cset_links anchored at cgrp->cset_links */
338 struct list_head cset_link;
340 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
341 struct list_head cgrp_link;
345 * The default css_set - used by init and its children prior to any
346 * hierarchies being mounted. It contains a pointer to the root state
347 * for each subsystem. Also used to anchor the list of css_sets. Not
348 * reference-counted, to improve performance when child cgroups
349 * haven't been created.
351 static struct css_set init_css_set = {
352 .refcount = ATOMIC_INIT(1),
353 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
354 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
355 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
356 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
357 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
360 static int css_set_count = 1; /* 1 for init_css_set */
363 * hash table for cgroup groups. This improves the performance to find
364 * an existing css_set. This hash doesn't (currently) take into
365 * account cgroups in empty hierarchies.
367 #define CSS_SET_HASH_BITS 7
368 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
370 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
372 unsigned long key = 0UL;
373 struct cgroup_subsys *ss;
376 for_each_subsys(ss, i)
377 key += (unsigned long)css[i];
378 key = (key >> 16) ^ key;
383 static void put_css_set_locked(struct css_set *cset, bool taskexit)
385 struct cgrp_cset_link *link, *tmp_link;
387 lockdep_assert_held(&css_set_rwsem);
389 if (!atomic_dec_and_test(&cset->refcount))
392 /* This css_set is dead. unlink it and release cgroup refcounts */
393 hash_del(&cset->hlist);
396 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
397 struct cgroup *cgrp = link->cgrp;
399 list_del(&link->cset_link);
400 list_del(&link->cgrp_link);
402 /* @cgrp can't go away while we're holding css_set_rwsem */
403 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
405 set_bit(CGRP_RELEASABLE, &cgrp->flags);
406 check_for_release(cgrp);
412 kfree_rcu(cset, rcu_head);
415 static void put_css_set(struct css_set *cset, bool taskexit)
418 * Ensure that the refcount doesn't hit zero while any readers
419 * can see it. Similar to atomic_dec_and_lock(), but for an
422 if (atomic_add_unless(&cset->refcount, -1, 1))
425 down_write(&css_set_rwsem);
426 put_css_set_locked(cset, taskexit);
427 up_write(&css_set_rwsem);
431 * refcounted get/put for css_set objects
433 static inline void get_css_set(struct css_set *cset)
435 atomic_inc(&cset->refcount);
439 * compare_css_sets - helper function for find_existing_css_set().
440 * @cset: candidate css_set being tested
441 * @old_cset: existing css_set for a task
442 * @new_cgrp: cgroup that's being entered by the task
443 * @template: desired set of css pointers in css_set (pre-calculated)
445 * Returns true if "cset" matches "old_cset" except for the hierarchy
446 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
448 static bool compare_css_sets(struct css_set *cset,
449 struct css_set *old_cset,
450 struct cgroup *new_cgrp,
451 struct cgroup_subsys_state *template[])
453 struct list_head *l1, *l2;
455 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
456 /* Not all subsystems matched */
461 * Compare cgroup pointers in order to distinguish between
462 * different cgroups in heirarchies with no subsystems. We
463 * could get by with just this check alone (and skip the
464 * memcmp above) but on most setups the memcmp check will
465 * avoid the need for this more expensive check on almost all
469 l1 = &cset->cgrp_links;
470 l2 = &old_cset->cgrp_links;
472 struct cgrp_cset_link *link1, *link2;
473 struct cgroup *cgrp1, *cgrp2;
477 /* See if we reached the end - both lists are equal length. */
478 if (l1 == &cset->cgrp_links) {
479 BUG_ON(l2 != &old_cset->cgrp_links);
482 BUG_ON(l2 == &old_cset->cgrp_links);
484 /* Locate the cgroups associated with these links. */
485 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
486 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
489 /* Hierarchies should be linked in the same order. */
490 BUG_ON(cgrp1->root != cgrp2->root);
493 * If this hierarchy is the hierarchy of the cgroup
494 * that's changing, then we need to check that this
495 * css_set points to the new cgroup; if it's any other
496 * hierarchy, then this css_set should point to the
497 * same cgroup as the old css_set.
499 if (cgrp1->root == new_cgrp->root) {
500 if (cgrp1 != new_cgrp)
511 * find_existing_css_set - init css array and find the matching css_set
512 * @old_cset: the css_set that we're using before the cgroup transition
513 * @cgrp: the cgroup that we're moving into
514 * @template: out param for the new set of csses, should be clear on entry
516 static struct css_set *find_existing_css_set(struct css_set *old_cset,
518 struct cgroup_subsys_state *template[])
520 struct cgroup_root *root = cgrp->root;
521 struct cgroup_subsys *ss;
522 struct css_set *cset;
527 * Build the set of subsystem state objects that we want to see in the
528 * new css_set. while subsystems can change globally, the entries here
529 * won't change, so no need for locking.
531 for_each_subsys(ss, i) {
532 if (root->cgrp.subsys_mask & (1UL << i)) {
533 /* Subsystem is in this hierarchy. So we want
534 * the subsystem state from the new
536 template[i] = cgroup_css(cgrp, ss);
538 /* Subsystem is not in this hierarchy, so we
539 * don't want to change the subsystem state */
540 template[i] = old_cset->subsys[i];
544 key = css_set_hash(template);
545 hash_for_each_possible(css_set_table, cset, hlist, key) {
546 if (!compare_css_sets(cset, old_cset, cgrp, template))
549 /* This css_set matches what we need */
553 /* No existing cgroup group matched */
557 static void free_cgrp_cset_links(struct list_head *links_to_free)
559 struct cgrp_cset_link *link, *tmp_link;
561 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
562 list_del(&link->cset_link);
568 * allocate_cgrp_cset_links - allocate cgrp_cset_links
569 * @count: the number of links to allocate
570 * @tmp_links: list_head the allocated links are put on
572 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
573 * through ->cset_link. Returns 0 on success or -errno.
575 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
577 struct cgrp_cset_link *link;
580 INIT_LIST_HEAD(tmp_links);
582 for (i = 0; i < count; i++) {
583 link = kzalloc(sizeof(*link), GFP_KERNEL);
585 free_cgrp_cset_links(tmp_links);
588 list_add(&link->cset_link, tmp_links);
594 * link_css_set - a helper function to link a css_set to a cgroup
595 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
596 * @cset: the css_set to be linked
597 * @cgrp: the destination cgroup
599 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
602 struct cgrp_cset_link *link;
604 BUG_ON(list_empty(tmp_links));
605 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
608 list_move(&link->cset_link, &cgrp->cset_links);
610 * Always add links to the tail of the list so that the list
611 * is sorted by order of hierarchy creation
613 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
617 * find_css_set - return a new css_set with one cgroup updated
618 * @old_cset: the baseline css_set
619 * @cgrp: the cgroup to be updated
621 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
622 * substituted into the appropriate hierarchy.
624 static struct css_set *find_css_set(struct css_set *old_cset,
627 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
628 struct css_set *cset;
629 struct list_head tmp_links;
630 struct cgrp_cset_link *link;
633 lockdep_assert_held(&cgroup_mutex);
635 /* First see if we already have a cgroup group that matches
637 down_read(&css_set_rwsem);
638 cset = find_existing_css_set(old_cset, cgrp, template);
641 up_read(&css_set_rwsem);
646 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
650 /* Allocate all the cgrp_cset_link objects that we'll need */
651 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
656 atomic_set(&cset->refcount, 1);
657 INIT_LIST_HEAD(&cset->cgrp_links);
658 INIT_LIST_HEAD(&cset->tasks);
659 INIT_LIST_HEAD(&cset->mg_tasks);
660 INIT_LIST_HEAD(&cset->mg_preload_node);
661 INIT_LIST_HEAD(&cset->mg_node);
662 INIT_HLIST_NODE(&cset->hlist);
664 /* Copy the set of subsystem state objects generated in
665 * find_existing_css_set() */
666 memcpy(cset->subsys, template, sizeof(cset->subsys));
668 down_write(&css_set_rwsem);
669 /* Add reference counts and links from the new css_set. */
670 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
671 struct cgroup *c = link->cgrp;
673 if (c->root == cgrp->root)
675 link_css_set(&tmp_links, cset, c);
678 BUG_ON(!list_empty(&tmp_links));
682 /* Add this cgroup group to the hash table */
683 key = css_set_hash(cset->subsys);
684 hash_add(css_set_table, &cset->hlist, key);
686 up_write(&css_set_rwsem);
691 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
693 struct cgroup *root_cgrp = kf_root->kn->priv;
695 return root_cgrp->root;
698 static int cgroup_init_root_id(struct cgroup_root *root)
702 lockdep_assert_held(&cgroup_mutex);
704 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
708 root->hierarchy_id = id;
712 static void cgroup_exit_root_id(struct cgroup_root *root)
714 lockdep_assert_held(&cgroup_mutex);
716 if (root->hierarchy_id) {
717 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
718 root->hierarchy_id = 0;
722 static void cgroup_free_root(struct cgroup_root *root)
725 /* hierarhcy ID shoulid already have been released */
726 WARN_ON_ONCE(root->hierarchy_id);
728 idr_destroy(&root->cgroup_idr);
733 static void cgroup_destroy_root(struct cgroup_root *root)
735 struct cgroup *cgrp = &root->cgrp;
736 struct cgrp_cset_link *link, *tmp_link;
738 mutex_lock(&cgroup_tree_mutex);
739 mutex_lock(&cgroup_mutex);
741 BUG_ON(atomic_read(&root->nr_cgrps));
742 BUG_ON(!list_empty(&cgrp->children));
744 /* Rebind all subsystems back to the default hierarchy */
745 rebind_subsystems(&cgrp_dfl_root, cgrp->subsys_mask);
748 * Release all the links from cset_links to this hierarchy's
751 down_write(&css_set_rwsem);
753 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
754 list_del(&link->cset_link);
755 list_del(&link->cgrp_link);
758 up_write(&css_set_rwsem);
760 if (!list_empty(&root->root_list)) {
761 list_del(&root->root_list);
765 cgroup_exit_root_id(root);
767 mutex_unlock(&cgroup_mutex);
768 mutex_unlock(&cgroup_tree_mutex);
770 kernfs_destroy_root(root->kf_root);
771 cgroup_free_root(root);
774 /* look up cgroup associated with given css_set on the specified hierarchy */
775 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
776 struct cgroup_root *root)
778 struct cgroup *res = NULL;
780 lockdep_assert_held(&cgroup_mutex);
781 lockdep_assert_held(&css_set_rwsem);
783 if (cset == &init_css_set) {
786 struct cgrp_cset_link *link;
788 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
789 struct cgroup *c = link->cgrp;
791 if (c->root == root) {
803 * Return the cgroup for "task" from the given hierarchy. Must be
804 * called with cgroup_mutex and css_set_rwsem held.
806 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
807 struct cgroup_root *root)
810 * No need to lock the task - since we hold cgroup_mutex the
811 * task can't change groups, so the only thing that can happen
812 * is that it exits and its css is set back to init_css_set.
814 return cset_cgroup_from_root(task_css_set(task), root);
818 * A task must hold cgroup_mutex to modify cgroups.
820 * Any task can increment and decrement the count field without lock.
821 * So in general, code holding cgroup_mutex can't rely on the count
822 * field not changing. However, if the count goes to zero, then only
823 * cgroup_attach_task() can increment it again. Because a count of zero
824 * means that no tasks are currently attached, therefore there is no
825 * way a task attached to that cgroup can fork (the other way to
826 * increment the count). So code holding cgroup_mutex can safely
827 * assume that if the count is zero, it will stay zero. Similarly, if
828 * a task holds cgroup_mutex on a cgroup with zero count, it
829 * knows that the cgroup won't be removed, as cgroup_rmdir()
832 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
833 * (usually) take cgroup_mutex. These are the two most performance
834 * critical pieces of code here. The exception occurs on cgroup_exit(),
835 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
836 * is taken, and if the cgroup count is zero, a usermode call made
837 * to the release agent with the name of the cgroup (path relative to
838 * the root of cgroup file system) as the argument.
840 * A cgroup can only be deleted if both its 'count' of using tasks
841 * is zero, and its list of 'children' cgroups is empty. Since all
842 * tasks in the system use _some_ cgroup, and since there is always at
843 * least one task in the system (init, pid == 1), therefore, root cgroup
844 * always has either children cgroups and/or using tasks. So we don't
845 * need a special hack to ensure that root cgroup cannot be deleted.
847 * P.S. One more locking exception. RCU is used to guard the
848 * update of a tasks cgroup pointer by cgroup_attach_task()
851 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
852 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
853 static const struct file_operations proc_cgroupstats_operations;
855 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
858 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
859 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
860 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
861 cft->ss->name, cft->name);
863 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
868 * cgroup_file_mode - deduce file mode of a control file
869 * @cft: the control file in question
871 * returns cft->mode if ->mode is not 0
872 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
873 * returns S_IRUGO if it has only a read handler
874 * returns S_IWUSR if it has only a write hander
876 static umode_t cgroup_file_mode(const struct cftype *cft)
883 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
886 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
893 static void cgroup_free_fn(struct work_struct *work)
895 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
897 atomic_dec(&cgrp->root->nr_cgrps);
898 cgroup_pidlist_destroy_all(cgrp);
902 * We get a ref to the parent, and put the ref when this
903 * cgroup is being freed, so it's guaranteed that the
904 * parent won't be destroyed before its children.
906 cgroup_put(cgrp->parent);
907 kernfs_put(cgrp->kn);
911 * This is root cgroup's refcnt reaching zero, which
912 * indicates that the root should be released.
914 cgroup_destroy_root(cgrp->root);
918 static void cgroup_free_rcu(struct rcu_head *head)
920 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
922 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
923 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
926 static void cgroup_get(struct cgroup *cgrp)
928 WARN_ON_ONCE(cgroup_is_dead(cgrp));
929 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
930 atomic_inc(&cgrp->refcnt);
933 static void cgroup_put(struct cgroup *cgrp)
935 if (!atomic_dec_and_test(&cgrp->refcnt))
937 if (WARN_ON_ONCE(cgrp->parent && !cgroup_is_dead(cgrp)))
941 * XXX: cgrp->id is only used to look up css's. As cgroup and
942 * css's lifetimes will be decoupled, it should be made
943 * per-subsystem and moved to css->id so that lookups are
944 * successful until the target css is released.
946 mutex_lock(&cgroup_mutex);
947 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
948 mutex_unlock(&cgroup_mutex);
951 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
954 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
956 char name[CGROUP_FILE_NAME_MAX];
958 lockdep_assert_held(&cgroup_tree_mutex);
959 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
963 * cgroup_clear_dir - remove subsys files in a cgroup directory
964 * @cgrp: target cgroup
965 * @subsys_mask: mask of the subsystem ids whose files should be removed
967 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
969 struct cgroup_subsys *ss;
972 for_each_subsys(ss, i) {
975 if (!test_bit(i, &subsys_mask))
977 list_for_each_entry(cfts, &ss->cfts, node)
978 cgroup_addrm_files(cgrp, cfts, false);
982 static int rebind_subsystems(struct cgroup_root *dst_root,
983 unsigned long ss_mask)
985 struct cgroup_subsys *ss;
988 lockdep_assert_held(&cgroup_tree_mutex);
989 lockdep_assert_held(&cgroup_mutex);
991 for_each_subsys(ss, ssid) {
992 if (!(ss_mask & (1 << ssid)))
995 /* if @ss is on the dummy_root, we can always move it */
996 if (ss->root == &cgrp_dfl_root)
999 /* if @ss has non-root cgroups attached to it, can't move */
1000 if (!list_empty(&ss->root->cgrp.children))
1003 /* can't move between two non-dummy roots either */
1004 if (dst_root != &cgrp_dfl_root)
1008 ret = cgroup_populate_dir(&dst_root->cgrp, ss_mask);
1010 if (dst_root != &cgrp_dfl_root)
1014 * Rebinding back to the default root is not allowed to
1015 * fail. Using both default and non-default roots should
1016 * be rare. Moving subsystems back and forth even more so.
1017 * Just warn about it and continue.
1019 if (cgrp_dfl_root_visible) {
1020 pr_warning("cgroup: failed to create files (%d) while rebinding 0x%lx to default root\n",
1022 pr_warning("cgroup: you may retry by moving them to a different hierarchy and unbinding\n");
1027 * Nothing can fail from this point on. Remove files for the
1028 * removed subsystems and rebind each subsystem.
1030 mutex_unlock(&cgroup_mutex);
1031 for_each_subsys(ss, ssid)
1032 if (ss_mask & (1 << ssid))
1033 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1034 mutex_lock(&cgroup_mutex);
1036 for_each_subsys(ss, ssid) {
1037 struct cgroup_root *src_root;
1038 struct cgroup_subsys_state *css;
1040 if (!(ss_mask & (1 << ssid)))
1043 src_root = ss->root;
1044 css = cgroup_css(&src_root->cgrp, ss);
1046 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1048 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1049 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1050 ss->root = dst_root;
1051 css->cgroup = &dst_root->cgrp;
1053 src_root->cgrp.subsys_mask &= ~(1 << ssid);
1054 dst_root->cgrp.subsys_mask |= 1 << ssid;
1060 kernfs_activate(dst_root->cgrp.kn);
1064 static int cgroup_show_options(struct seq_file *seq,
1065 struct kernfs_root *kf_root)
1067 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1068 struct cgroup_subsys *ss;
1071 for_each_subsys(ss, ssid)
1072 if (root->cgrp.subsys_mask & (1 << ssid))
1073 seq_printf(seq, ",%s", ss->name);
1074 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1075 seq_puts(seq, ",sane_behavior");
1076 if (root->flags & CGRP_ROOT_NOPREFIX)
1077 seq_puts(seq, ",noprefix");
1078 if (root->flags & CGRP_ROOT_XATTR)
1079 seq_puts(seq, ",xattr");
1081 spin_lock(&release_agent_path_lock);
1082 if (strlen(root->release_agent_path))
1083 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1084 spin_unlock(&release_agent_path_lock);
1086 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1087 seq_puts(seq, ",clone_children");
1088 if (strlen(root->name))
1089 seq_printf(seq, ",name=%s", root->name);
1093 struct cgroup_sb_opts {
1094 unsigned long subsys_mask;
1095 unsigned long flags;
1096 char *release_agent;
1097 bool cpuset_clone_children;
1099 /* User explicitly requested empty subsystem */
1104 * Convert a hierarchy specifier into a bitmask of subsystems and
1105 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1106 * array. This function takes refcounts on subsystems to be used, unless it
1107 * returns error, in which case no refcounts are taken.
1109 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1111 char *token, *o = data;
1112 bool all_ss = false, one_ss = false;
1113 unsigned long mask = (unsigned long)-1;
1114 struct cgroup_subsys *ss;
1117 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1119 #ifdef CONFIG_CPUSETS
1120 mask = ~(1UL << cpuset_cgrp_id);
1123 memset(opts, 0, sizeof(*opts));
1125 while ((token = strsep(&o, ",")) != NULL) {
1128 if (!strcmp(token, "none")) {
1129 /* Explicitly have no subsystems */
1133 if (!strcmp(token, "all")) {
1134 /* Mutually exclusive option 'all' + subsystem name */
1140 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1141 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1144 if (!strcmp(token, "noprefix")) {
1145 opts->flags |= CGRP_ROOT_NOPREFIX;
1148 if (!strcmp(token, "clone_children")) {
1149 opts->cpuset_clone_children = true;
1152 if (!strcmp(token, "xattr")) {
1153 opts->flags |= CGRP_ROOT_XATTR;
1156 if (!strncmp(token, "release_agent=", 14)) {
1157 /* Specifying two release agents is forbidden */
1158 if (opts->release_agent)
1160 opts->release_agent =
1161 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1162 if (!opts->release_agent)
1166 if (!strncmp(token, "name=", 5)) {
1167 const char *name = token + 5;
1168 /* Can't specify an empty name */
1171 /* Must match [\w.-]+ */
1172 for (i = 0; i < strlen(name); i++) {
1176 if ((c == '.') || (c == '-') || (c == '_'))
1180 /* Specifying two names is forbidden */
1183 opts->name = kstrndup(name,
1184 MAX_CGROUP_ROOT_NAMELEN - 1,
1192 for_each_subsys(ss, i) {
1193 if (strcmp(token, ss->name))
1198 /* Mutually exclusive option 'all' + subsystem name */
1201 set_bit(i, &opts->subsys_mask);
1206 if (i == CGROUP_SUBSYS_COUNT)
1210 /* Consistency checks */
1212 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1213 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1215 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1216 opts->cpuset_clone_children || opts->release_agent ||
1218 pr_err("cgroup: sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1223 * If the 'all' option was specified select all the
1224 * subsystems, otherwise if 'none', 'name=' and a subsystem
1225 * name options were not specified, let's default to 'all'
1227 if (all_ss || (!one_ss && !opts->none && !opts->name))
1228 for_each_subsys(ss, i)
1230 set_bit(i, &opts->subsys_mask);
1233 * We either have to specify by name or by subsystems. (So
1234 * all empty hierarchies must have a name).
1236 if (!opts->subsys_mask && !opts->name)
1241 * Option noprefix was introduced just for backward compatibility
1242 * with the old cpuset, so we allow noprefix only if mounting just
1243 * the cpuset subsystem.
1245 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1249 /* Can't specify "none" and some subsystems */
1250 if (opts->subsys_mask && opts->none)
1256 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1259 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1260 struct cgroup_sb_opts opts;
1261 unsigned long added_mask, removed_mask;
1263 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1264 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1268 mutex_lock(&cgroup_tree_mutex);
1269 mutex_lock(&cgroup_mutex);
1271 /* See what subsystems are wanted */
1272 ret = parse_cgroupfs_options(data, &opts);
1276 if (opts.subsys_mask != root->cgrp.subsys_mask || opts.release_agent)
1277 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1278 task_tgid_nr(current), current->comm);
1280 added_mask = opts.subsys_mask & ~root->cgrp.subsys_mask;
1281 removed_mask = root->cgrp.subsys_mask & ~opts.subsys_mask;
1283 /* Don't allow flags or name to change at remount */
1284 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1285 (opts.name && strcmp(opts.name, root->name))) {
1286 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1287 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1288 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1293 /* remounting is not allowed for populated hierarchies */
1294 if (!list_empty(&root->cgrp.children)) {
1299 ret = rebind_subsystems(root, added_mask);
1303 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1305 if (opts.release_agent) {
1306 spin_lock(&release_agent_path_lock);
1307 strcpy(root->release_agent_path, opts.release_agent);
1308 spin_unlock(&release_agent_path_lock);
1311 kfree(opts.release_agent);
1313 mutex_unlock(&cgroup_mutex);
1314 mutex_unlock(&cgroup_tree_mutex);
1319 * To reduce the fork() overhead for systems that are not actually using
1320 * their cgroups capability, we don't maintain the lists running through
1321 * each css_set to its tasks until we see the list actually used - in other
1322 * words after the first mount.
1324 static bool use_task_css_set_links __read_mostly;
1326 static void cgroup_enable_task_cg_lists(void)
1328 struct task_struct *p, *g;
1330 down_write(&css_set_rwsem);
1332 if (use_task_css_set_links)
1335 use_task_css_set_links = true;
1338 * We need tasklist_lock because RCU is not safe against
1339 * while_each_thread(). Besides, a forking task that has passed
1340 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1341 * is not guaranteed to have its child immediately visible in the
1342 * tasklist if we walk through it with RCU.
1344 read_lock(&tasklist_lock);
1345 do_each_thread(g, p) {
1346 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1347 task_css_set(p) != &init_css_set);
1350 * We should check if the process is exiting, otherwise
1351 * it will race with cgroup_exit() in that the list
1352 * entry won't be deleted though the process has exited.
1353 * Do it while holding siglock so that we don't end up
1354 * racing against cgroup_exit().
1356 spin_lock_irq(&p->sighand->siglock);
1357 if (!(p->flags & PF_EXITING)) {
1358 struct css_set *cset = task_css_set(p);
1360 list_add(&p->cg_list, &cset->tasks);
1363 spin_unlock_irq(&p->sighand->siglock);
1364 } while_each_thread(g, p);
1365 read_unlock(&tasklist_lock);
1367 up_write(&css_set_rwsem);
1370 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1372 atomic_set(&cgrp->refcnt, 1);
1373 INIT_LIST_HEAD(&cgrp->sibling);
1374 INIT_LIST_HEAD(&cgrp->children);
1375 INIT_LIST_HEAD(&cgrp->cset_links);
1376 INIT_LIST_HEAD(&cgrp->release_list);
1377 INIT_LIST_HEAD(&cgrp->pidlists);
1378 mutex_init(&cgrp->pidlist_mutex);
1379 cgrp->dummy_css.cgroup = cgrp;
1382 static void init_cgroup_root(struct cgroup_root *root,
1383 struct cgroup_sb_opts *opts)
1385 struct cgroup *cgrp = &root->cgrp;
1387 INIT_LIST_HEAD(&root->root_list);
1388 atomic_set(&root->nr_cgrps, 1);
1390 init_cgroup_housekeeping(cgrp);
1391 idr_init(&root->cgroup_idr);
1393 root->flags = opts->flags;
1394 if (opts->release_agent)
1395 strcpy(root->release_agent_path, opts->release_agent);
1397 strcpy(root->name, opts->name);
1398 if (opts->cpuset_clone_children)
1399 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1402 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1404 LIST_HEAD(tmp_links);
1405 struct cgroup *root_cgrp = &root->cgrp;
1406 struct css_set *cset;
1409 lockdep_assert_held(&cgroup_tree_mutex);
1410 lockdep_assert_held(&cgroup_mutex);
1412 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1415 root_cgrp->id = ret;
1418 * We're accessing css_set_count without locking css_set_rwsem here,
1419 * but that's OK - it can only be increased by someone holding
1420 * cgroup_lock, and that's us. The worst that can happen is that we
1421 * have some link structures left over
1423 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1427 ret = cgroup_init_root_id(root);
1431 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1432 KERNFS_ROOT_CREATE_DEACTIVATED,
1434 if (IS_ERR(root->kf_root)) {
1435 ret = PTR_ERR(root->kf_root);
1438 root_cgrp->kn = root->kf_root->kn;
1440 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1444 ret = rebind_subsystems(root, ss_mask);
1449 * There must be no failure case after here, since rebinding takes
1450 * care of subsystems' refcounts, which are explicitly dropped in
1451 * the failure exit path.
1453 list_add(&root->root_list, &cgroup_roots);
1454 cgroup_root_count++;
1457 * Link the root cgroup in this hierarchy into all the css_set
1460 down_write(&css_set_rwsem);
1461 hash_for_each(css_set_table, i, cset, hlist)
1462 link_css_set(&tmp_links, cset, root_cgrp);
1463 up_write(&css_set_rwsem);
1465 BUG_ON(!list_empty(&root_cgrp->children));
1466 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1468 kernfs_activate(root_cgrp->kn);
1473 kernfs_destroy_root(root->kf_root);
1474 root->kf_root = NULL;
1476 cgroup_exit_root_id(root);
1478 free_cgrp_cset_links(&tmp_links);
1482 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1483 int flags, const char *unused_dev_name,
1486 struct cgroup_root *root;
1487 struct cgroup_sb_opts opts;
1488 struct dentry *dentry;
1492 * The first time anyone tries to mount a cgroup, enable the list
1493 * linking each css_set to its tasks and fix up all existing tasks.
1495 if (!use_task_css_set_links)
1496 cgroup_enable_task_cg_lists();
1498 mutex_lock(&cgroup_tree_mutex);
1499 mutex_lock(&cgroup_mutex);
1501 /* First find the desired set of subsystems */
1502 ret = parse_cgroupfs_options(data, &opts);
1506 /* look for a matching existing root */
1507 if (!opts.subsys_mask && !opts.none && !opts.name) {
1508 cgrp_dfl_root_visible = true;
1509 root = &cgrp_dfl_root;
1510 cgroup_get(&root->cgrp);
1515 for_each_root(root) {
1516 bool name_match = false;
1518 if (root == &cgrp_dfl_root)
1522 * If we asked for a name then it must match. Also, if
1523 * name matches but sybsys_mask doesn't, we should fail.
1524 * Remember whether name matched.
1527 if (strcmp(opts.name, root->name))
1533 * If we asked for subsystems (or explicitly for no
1534 * subsystems) then they must match.
1536 if ((opts.subsys_mask || opts.none) &&
1537 (opts.subsys_mask != root->cgrp.subsys_mask)) {
1544 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1545 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1546 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1550 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1555 * A root's lifetime is governed by its root cgroup. Zero
1556 * ref indicate that the root is being destroyed. Wait for
1557 * destruction to complete so that the subsystems are free.
1558 * We can use wait_queue for the wait but this path is
1559 * super cold. Let's just sleep for a bit and retry.
1561 if (!atomic_inc_not_zero(&root->cgrp.refcnt)) {
1562 mutex_unlock(&cgroup_mutex);
1563 mutex_unlock(&cgroup_tree_mutex);
1564 kfree(opts.release_agent);
1575 * No such thing, create a new one. name= matching without subsys
1576 * specification is allowed for already existing hierarchies but we
1577 * can't create new one without subsys specification.
1579 if (!opts.subsys_mask && !opts.none) {
1584 root = kzalloc(sizeof(*root), GFP_KERNEL);
1590 init_cgroup_root(root, &opts);
1592 ret = cgroup_setup_root(root, opts.subsys_mask);
1594 cgroup_free_root(root);
1597 mutex_unlock(&cgroup_mutex);
1598 mutex_unlock(&cgroup_tree_mutex);
1600 kfree(opts.release_agent);
1604 return ERR_PTR(ret);
1606 dentry = kernfs_mount(fs_type, flags, root->kf_root);
1608 cgroup_put(&root->cgrp);
1612 static void cgroup_kill_sb(struct super_block *sb)
1614 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1615 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1617 cgroup_put(&root->cgrp);
1621 static struct file_system_type cgroup_fs_type = {
1623 .mount = cgroup_mount,
1624 .kill_sb = cgroup_kill_sb,
1627 static struct kobject *cgroup_kobj;
1630 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1631 * @task: target task
1632 * @buf: the buffer to write the path into
1633 * @buflen: the length of the buffer
1635 * Determine @task's cgroup on the first (the one with the lowest non-zero
1636 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1637 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1638 * cgroup controller callbacks.
1640 * Return value is the same as kernfs_path().
1642 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1644 struct cgroup_root *root;
1645 struct cgroup *cgrp;
1646 int hierarchy_id = 1;
1649 mutex_lock(&cgroup_mutex);
1650 down_read(&css_set_rwsem);
1652 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1655 cgrp = task_cgroup_from_root(task, root);
1656 path = cgroup_path(cgrp, buf, buflen);
1658 /* if no hierarchy exists, everyone is in "/" */
1659 if (strlcpy(buf, "/", buflen) < buflen)
1663 up_read(&css_set_rwsem);
1664 mutex_unlock(&cgroup_mutex);
1667 EXPORT_SYMBOL_GPL(task_cgroup_path);
1669 /* used to track tasks and other necessary states during migration */
1670 struct cgroup_taskset {
1671 /* the src and dst cset list running through cset->mg_node */
1672 struct list_head src_csets;
1673 struct list_head dst_csets;
1676 * Fields for cgroup_taskset_*() iteration.
1678 * Before migration is committed, the target migration tasks are on
1679 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1680 * the csets on ->dst_csets. ->csets point to either ->src_csets
1681 * or ->dst_csets depending on whether migration is committed.
1683 * ->cur_csets and ->cur_task point to the current task position
1686 struct list_head *csets;
1687 struct css_set *cur_cset;
1688 struct task_struct *cur_task;
1692 * cgroup_taskset_first - reset taskset and return the first task
1693 * @tset: taskset of interest
1695 * @tset iteration is initialized and the first task is returned.
1697 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1699 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
1700 tset->cur_task = NULL;
1702 return cgroup_taskset_next(tset);
1706 * cgroup_taskset_next - iterate to the next task in taskset
1707 * @tset: taskset of interest
1709 * Return the next task in @tset. Iteration must have been initialized
1710 * with cgroup_taskset_first().
1712 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1714 struct css_set *cset = tset->cur_cset;
1715 struct task_struct *task = tset->cur_task;
1717 while (&cset->mg_node != tset->csets) {
1719 task = list_first_entry(&cset->mg_tasks,
1720 struct task_struct, cg_list);
1722 task = list_next_entry(task, cg_list);
1724 if (&task->cg_list != &cset->mg_tasks) {
1725 tset->cur_cset = cset;
1726 tset->cur_task = task;
1730 cset = list_next_entry(cset, mg_node);
1738 * cgroup_task_migrate - move a task from one cgroup to another.
1739 * @old_cgrp; the cgroup @tsk is being migrated from
1740 * @tsk: the task being migrated
1741 * @new_cset: the new css_set @tsk is being attached to
1743 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1745 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1746 struct task_struct *tsk,
1747 struct css_set *new_cset)
1749 struct css_set *old_cset;
1751 lockdep_assert_held(&cgroup_mutex);
1752 lockdep_assert_held(&css_set_rwsem);
1755 * We are synchronized through threadgroup_lock() against PF_EXITING
1756 * setting such that we can't race against cgroup_exit() changing the
1757 * css_set to init_css_set and dropping the old one.
1759 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1760 old_cset = task_css_set(tsk);
1762 get_css_set(new_cset);
1763 rcu_assign_pointer(tsk->cgroups, new_cset);
1764 list_move(&tsk->cg_list, &new_cset->mg_tasks);
1767 * We just gained a reference on old_cset by taking it from the
1768 * task. As trading it for new_cset is protected by cgroup_mutex,
1769 * we're safe to drop it here; it will be freed under RCU.
1771 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1772 put_css_set_locked(old_cset, false);
1776 * cgroup_migrate_finish - cleanup after attach
1777 * @preloaded_csets: list of preloaded css_sets
1779 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
1780 * those functions for details.
1782 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
1784 struct css_set *cset, *tmp_cset;
1786 lockdep_assert_held(&cgroup_mutex);
1788 down_write(&css_set_rwsem);
1789 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
1790 cset->mg_src_cgrp = NULL;
1791 cset->mg_dst_cset = NULL;
1792 list_del_init(&cset->mg_preload_node);
1793 put_css_set_locked(cset, false);
1795 up_write(&css_set_rwsem);
1799 * cgroup_migrate_add_src - add a migration source css_set
1800 * @src_cset: the source css_set to add
1801 * @dst_cgrp: the destination cgroup
1802 * @preloaded_csets: list of preloaded css_sets
1804 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
1805 * @src_cset and add it to @preloaded_csets, which should later be cleaned
1806 * up by cgroup_migrate_finish().
1808 * This function may be called without holding threadgroup_lock even if the
1809 * target is a process. Threads may be created and destroyed but as long
1810 * as cgroup_mutex is not dropped, no new css_set can be put into play and
1811 * the preloaded css_sets are guaranteed to cover all migrations.
1813 static void cgroup_migrate_add_src(struct css_set *src_cset,
1814 struct cgroup *dst_cgrp,
1815 struct list_head *preloaded_csets)
1817 struct cgroup *src_cgrp;
1819 lockdep_assert_held(&cgroup_mutex);
1820 lockdep_assert_held(&css_set_rwsem);
1822 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
1824 /* nothing to do if this cset already belongs to the cgroup */
1825 if (src_cgrp == dst_cgrp)
1828 if (!list_empty(&src_cset->mg_preload_node))
1831 WARN_ON(src_cset->mg_src_cgrp);
1832 WARN_ON(!list_empty(&src_cset->mg_tasks));
1833 WARN_ON(!list_empty(&src_cset->mg_node));
1835 src_cset->mg_src_cgrp = src_cgrp;
1836 get_css_set(src_cset);
1837 list_add(&src_cset->mg_preload_node, preloaded_csets);
1841 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
1842 * @dst_cgrp: the destination cgroup
1843 * @preloaded_csets: list of preloaded source css_sets
1845 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
1846 * have been preloaded to @preloaded_csets. This function looks up and
1847 * pins all destination css_sets, links each to its source, and put them on
1850 * This function must be called after cgroup_migrate_add_src() has been
1851 * called on each migration source css_set. After migration is performed
1852 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
1855 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
1856 struct list_head *preloaded_csets)
1859 struct css_set *src_cset;
1861 lockdep_assert_held(&cgroup_mutex);
1863 /* look up the dst cset for each src cset and link it to src */
1864 list_for_each_entry(src_cset, preloaded_csets, mg_preload_node) {
1865 struct css_set *dst_cset;
1867 dst_cset = find_css_set(src_cset, dst_cgrp);
1871 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
1872 src_cset->mg_dst_cset = dst_cset;
1874 if (list_empty(&dst_cset->mg_preload_node))
1875 list_add(&dst_cset->mg_preload_node, &csets);
1877 put_css_set(dst_cset, false);
1880 list_splice(&csets, preloaded_csets);
1883 cgroup_migrate_finish(&csets);
1888 * cgroup_migrate - migrate a process or task to a cgroup
1889 * @cgrp: the destination cgroup
1890 * @leader: the leader of the process or the task to migrate
1891 * @threadgroup: whether @leader points to the whole process or a single task
1893 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
1894 * process, the caller must be holding threadgroup_lock of @leader. The
1895 * caller is also responsible for invoking cgroup_migrate_add_src() and
1896 * cgroup_migrate_prepare_dst() on the targets before invoking this
1897 * function and following up with cgroup_migrate_finish().
1899 * As long as a controller's ->can_attach() doesn't fail, this function is
1900 * guaranteed to succeed. This means that, excluding ->can_attach()
1901 * failure, when migrating multiple targets, the success or failure can be
1902 * decided for all targets by invoking group_migrate_prepare_dst() before
1903 * actually starting migrating.
1905 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
1908 struct cgroup_taskset tset = {
1909 .src_csets = LIST_HEAD_INIT(tset.src_csets),
1910 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
1911 .csets = &tset.src_csets,
1913 struct cgroup_subsys_state *css, *failed_css = NULL;
1914 struct css_set *cset, *tmp_cset;
1915 struct task_struct *task, *tmp_task;
1919 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1920 * already PF_EXITING could be freed from underneath us unless we
1921 * take an rcu_read_lock.
1923 down_write(&css_set_rwsem);
1927 /* @task either already exited or can't exit until the end */
1928 if (task->flags & PF_EXITING)
1931 /* leave @task alone if post_fork() hasn't linked it yet */
1932 if (list_empty(&task->cg_list))
1935 cset = task_css_set(task);
1936 if (!cset->mg_src_cgrp)
1939 list_move(&task->cg_list, &cset->mg_tasks);
1940 list_move(&cset->mg_node, &tset.src_csets);
1941 list_move(&cset->mg_dst_cset->mg_node, &tset.dst_csets);
1945 } while_each_thread(leader, task);
1947 up_write(&css_set_rwsem);
1949 /* methods shouldn't be called if no task is actually migrating */
1950 if (list_empty(&tset.src_csets))
1953 /* check that we can legitimately attach to the cgroup */
1954 for_each_css(css, i, cgrp) {
1955 if (css->ss->can_attach) {
1956 ret = css->ss->can_attach(css, &tset);
1959 goto out_cancel_attach;
1965 * Now that we're guaranteed success, proceed to move all tasks to
1966 * the new cgroup. There are no failure cases after here, so this
1967 * is the commit point.
1969 down_write(&css_set_rwsem);
1970 list_for_each_entry(cset, &tset.src_csets, mg_node) {
1971 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
1972 cgroup_task_migrate(cset->mg_src_cgrp, task,
1975 up_write(&css_set_rwsem);
1978 * Migration is committed, all target tasks are now on dst_csets.
1979 * Nothing is sensitive to fork() after this point. Notify
1980 * controllers that migration is complete.
1982 tset.csets = &tset.dst_csets;
1984 for_each_css(css, i, cgrp)
1985 if (css->ss->attach)
1986 css->ss->attach(css, &tset);
1989 goto out_release_tset;
1992 for_each_css(css, i, cgrp) {
1993 if (css == failed_css)
1995 if (css->ss->cancel_attach)
1996 css->ss->cancel_attach(css, &tset);
1999 down_write(&css_set_rwsem);
2000 list_splice_init(&tset.dst_csets, &tset.src_csets);
2001 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2002 list_splice_init(&cset->mg_tasks, &cset->tasks);
2003 list_del_init(&cset->mg_node);
2005 up_write(&css_set_rwsem);
2010 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2011 * @dst_cgrp: the cgroup to attach to
2012 * @leader: the task or the leader of the threadgroup to be attached
2013 * @threadgroup: attach the whole threadgroup?
2015 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2017 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2018 struct task_struct *leader, bool threadgroup)
2020 LIST_HEAD(preloaded_csets);
2021 struct task_struct *task;
2024 /* look up all src csets */
2025 down_read(&css_set_rwsem);
2029 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2033 } while_each_thread(leader, task);
2035 up_read(&css_set_rwsem);
2037 /* prepare dst csets and commit */
2038 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2040 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2042 cgroup_migrate_finish(&preloaded_csets);
2047 * Find the task_struct of the task to attach by vpid and pass it along to the
2048 * function to attach either it or all tasks in its threadgroup. Will lock
2049 * cgroup_mutex and threadgroup.
2051 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2053 struct task_struct *tsk;
2054 const struct cred *cred = current_cred(), *tcred;
2057 if (!cgroup_lock_live_group(cgrp))
2063 tsk = find_task_by_vpid(pid);
2067 goto out_unlock_cgroup;
2070 * even if we're attaching all tasks in the thread group, we
2071 * only need to check permissions on one of them.
2073 tcred = __task_cred(tsk);
2074 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2075 !uid_eq(cred->euid, tcred->uid) &&
2076 !uid_eq(cred->euid, tcred->suid)) {
2079 goto out_unlock_cgroup;
2085 tsk = tsk->group_leader;
2088 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2089 * trapped in a cpuset, or RT worker may be born in a cgroup
2090 * with no rt_runtime allocated. Just say no.
2092 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2095 goto out_unlock_cgroup;
2098 get_task_struct(tsk);
2101 threadgroup_lock(tsk);
2103 if (!thread_group_leader(tsk)) {
2105 * a race with de_thread from another thread's exec()
2106 * may strip us of our leadership, if this happens,
2107 * there is no choice but to throw this task away and
2108 * try again; this is
2109 * "double-double-toil-and-trouble-check locking".
2111 threadgroup_unlock(tsk);
2112 put_task_struct(tsk);
2113 goto retry_find_task;
2117 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2119 threadgroup_unlock(tsk);
2121 put_task_struct(tsk);
2123 mutex_unlock(&cgroup_mutex);
2128 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2129 * @from: attach to all cgroups of a given task
2130 * @tsk: the task to be attached
2132 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2134 struct cgroup_root *root;
2137 mutex_lock(&cgroup_mutex);
2138 for_each_root(root) {
2139 struct cgroup *from_cgrp;
2141 if (root == &cgrp_dfl_root)
2144 down_read(&css_set_rwsem);
2145 from_cgrp = task_cgroup_from_root(from, root);
2146 up_read(&css_set_rwsem);
2148 retval = cgroup_attach_task(from_cgrp, tsk, false);
2152 mutex_unlock(&cgroup_mutex);
2156 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2158 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2159 struct cftype *cft, u64 pid)
2161 return attach_task_by_pid(css->cgroup, pid, false);
2164 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2165 struct cftype *cft, u64 tgid)
2167 return attach_task_by_pid(css->cgroup, tgid, true);
2170 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2171 struct cftype *cft, char *buffer)
2173 struct cgroup_root *root = css->cgroup->root;
2175 BUILD_BUG_ON(sizeof(root->release_agent_path) < PATH_MAX);
2176 if (!cgroup_lock_live_group(css->cgroup))
2178 spin_lock(&release_agent_path_lock);
2179 strlcpy(root->release_agent_path, buffer,
2180 sizeof(root->release_agent_path));
2181 spin_unlock(&release_agent_path_lock);
2182 mutex_unlock(&cgroup_mutex);
2186 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2188 struct cgroup *cgrp = seq_css(seq)->cgroup;
2190 if (!cgroup_lock_live_group(cgrp))
2192 seq_puts(seq, cgrp->root->release_agent_path);
2193 seq_putc(seq, '\n');
2194 mutex_unlock(&cgroup_mutex);
2198 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2200 struct cgroup *cgrp = seq_css(seq)->cgroup;
2202 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2206 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2207 size_t nbytes, loff_t off)
2209 struct cgroup *cgrp = of->kn->parent->priv;
2210 struct cftype *cft = of->kn->priv;
2211 struct cgroup_subsys_state *css;
2215 * kernfs guarantees that a file isn't deleted with operations in
2216 * flight, which means that the matching css is and stays alive and
2217 * doesn't need to be pinned. The RCU locking is not necessary
2218 * either. It's just for the convenience of using cgroup_css().
2221 css = cgroup_css(cgrp, cft->ss);
2224 if (cft->write_string) {
2225 ret = cft->write_string(css, cft, strstrip(buf));
2226 } else if (cft->write_u64) {
2227 unsigned long long v;
2228 ret = kstrtoull(buf, 0, &v);
2230 ret = cft->write_u64(css, cft, v);
2231 } else if (cft->write_s64) {
2233 ret = kstrtoll(buf, 0, &v);
2235 ret = cft->write_s64(css, cft, v);
2236 } else if (cft->trigger) {
2237 ret = cft->trigger(css, (unsigned int)cft->private);
2242 return ret ?: nbytes;
2245 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2247 return seq_cft(seq)->seq_start(seq, ppos);
2250 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2252 return seq_cft(seq)->seq_next(seq, v, ppos);
2255 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2257 seq_cft(seq)->seq_stop(seq, v);
2260 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2262 struct cftype *cft = seq_cft(m);
2263 struct cgroup_subsys_state *css = seq_css(m);
2266 return cft->seq_show(m, arg);
2269 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2270 else if (cft->read_s64)
2271 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2277 static struct kernfs_ops cgroup_kf_single_ops = {
2278 .atomic_write_len = PAGE_SIZE,
2279 .write = cgroup_file_write,
2280 .seq_show = cgroup_seqfile_show,
2283 static struct kernfs_ops cgroup_kf_ops = {
2284 .atomic_write_len = PAGE_SIZE,
2285 .write = cgroup_file_write,
2286 .seq_start = cgroup_seqfile_start,
2287 .seq_next = cgroup_seqfile_next,
2288 .seq_stop = cgroup_seqfile_stop,
2289 .seq_show = cgroup_seqfile_show,
2293 * cgroup_rename - Only allow simple rename of directories in place.
2295 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2296 const char *new_name_str)
2298 struct cgroup *cgrp = kn->priv;
2301 if (kernfs_type(kn) != KERNFS_DIR)
2303 if (kn->parent != new_parent)
2307 * This isn't a proper migration and its usefulness is very
2308 * limited. Disallow if sane_behavior.
2310 if (cgroup_sane_behavior(cgrp))
2313 mutex_lock(&cgroup_tree_mutex);
2314 mutex_lock(&cgroup_mutex);
2316 ret = kernfs_rename(kn, new_parent, new_name_str);
2318 mutex_unlock(&cgroup_mutex);
2319 mutex_unlock(&cgroup_tree_mutex);
2323 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2325 char name[CGROUP_FILE_NAME_MAX];
2326 struct kernfs_node *kn;
2327 struct lock_class_key *key = NULL;
2329 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2330 key = &cft->lockdep_key;
2332 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2333 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2335 return PTR_ERR_OR_ZERO(kn);
2339 * cgroup_addrm_files - add or remove files to a cgroup directory
2340 * @cgrp: the target cgroup
2341 * @cfts: array of cftypes to be added
2342 * @is_add: whether to add or remove
2344 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2345 * For removals, this function never fails. If addition fails, this
2346 * function doesn't remove files already added. The caller is responsible
2349 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2355 lockdep_assert_held(&cgroup_tree_mutex);
2357 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2358 /* does cft->flags tell us to skip this file on @cgrp? */
2359 if ((cft->flags & CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
2361 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2363 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2365 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2369 ret = cgroup_add_file(cgrp, cft);
2371 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2376 cgroup_rm_file(cgrp, cft);
2382 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2385 struct cgroup_subsys *ss = cfts[0].ss;
2386 struct cgroup *root = &ss->root->cgrp;
2387 struct cgroup_subsys_state *css;
2390 lockdep_assert_held(&cgroup_tree_mutex);
2392 /* don't bother if @ss isn't attached */
2393 if (ss->root == &cgrp_dfl_root)
2396 /* add/rm files for all cgroups created before */
2397 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2398 struct cgroup *cgrp = css->cgroup;
2400 if (cgroup_is_dead(cgrp))
2403 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2409 kernfs_activate(root->kn);
2413 static void cgroup_exit_cftypes(struct cftype *cfts)
2417 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2418 /* free copy for custom atomic_write_len, see init_cftypes() */
2419 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2426 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2430 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2431 struct kernfs_ops *kf_ops;
2433 WARN_ON(cft->ss || cft->kf_ops);
2436 kf_ops = &cgroup_kf_ops;
2438 kf_ops = &cgroup_kf_single_ops;
2441 * Ugh... if @cft wants a custom max_write_len, we need to
2442 * make a copy of kf_ops to set its atomic_write_len.
2444 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
2445 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
2447 cgroup_exit_cftypes(cfts);
2450 kf_ops->atomic_write_len = cft->max_write_len;
2453 cft->kf_ops = kf_ops;
2460 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
2462 lockdep_assert_held(&cgroup_tree_mutex);
2464 if (!cfts || !cfts[0].ss)
2467 list_del(&cfts->node);
2468 cgroup_apply_cftypes(cfts, false);
2469 cgroup_exit_cftypes(cfts);
2474 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2475 * @cfts: zero-length name terminated array of cftypes
2477 * Unregister @cfts. Files described by @cfts are removed from all
2478 * existing cgroups and all future cgroups won't have them either. This
2479 * function can be called anytime whether @cfts' subsys is attached or not.
2481 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2484 int cgroup_rm_cftypes(struct cftype *cfts)
2488 mutex_lock(&cgroup_tree_mutex);
2489 ret = cgroup_rm_cftypes_locked(cfts);
2490 mutex_unlock(&cgroup_tree_mutex);
2495 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2496 * @ss: target cgroup subsystem
2497 * @cfts: zero-length name terminated array of cftypes
2499 * Register @cfts to @ss. Files described by @cfts are created for all
2500 * existing cgroups to which @ss is attached and all future cgroups will
2501 * have them too. This function can be called anytime whether @ss is
2504 * Returns 0 on successful registration, -errno on failure. Note that this
2505 * function currently returns 0 as long as @cfts registration is successful
2506 * even if some file creation attempts on existing cgroups fail.
2508 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2512 if (!cfts || cfts[0].name[0] == '\0')
2515 ret = cgroup_init_cftypes(ss, cfts);
2519 mutex_lock(&cgroup_tree_mutex);
2521 list_add_tail(&cfts->node, &ss->cfts);
2522 ret = cgroup_apply_cftypes(cfts, true);
2524 cgroup_rm_cftypes_locked(cfts);
2526 mutex_unlock(&cgroup_tree_mutex);
2531 * cgroup_task_count - count the number of tasks in a cgroup.
2532 * @cgrp: the cgroup in question
2534 * Return the number of tasks in the cgroup.
2536 static int cgroup_task_count(const struct cgroup *cgrp)
2539 struct cgrp_cset_link *link;
2541 down_read(&css_set_rwsem);
2542 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2543 count += atomic_read(&link->cset->refcount);
2544 up_read(&css_set_rwsem);
2549 * css_next_child - find the next child of a given css
2550 * @pos_css: the current position (%NULL to initiate traversal)
2551 * @parent_css: css whose children to walk
2553 * This function returns the next child of @parent_css and should be called
2554 * under either cgroup_mutex or RCU read lock. The only requirement is
2555 * that @parent_css and @pos_css are accessible. The next sibling is
2556 * guaranteed to be returned regardless of their states.
2558 struct cgroup_subsys_state *
2559 css_next_child(struct cgroup_subsys_state *pos_css,
2560 struct cgroup_subsys_state *parent_css)
2562 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2563 struct cgroup *cgrp = parent_css->cgroup;
2564 struct cgroup *next;
2566 cgroup_assert_mutexes_or_rcu_locked();
2569 * @pos could already have been removed. Once a cgroup is removed,
2570 * its ->sibling.next is no longer updated when its next sibling
2571 * changes. As CGRP_DEAD assertion is serialized and happens
2572 * before the cgroup is taken off the ->sibling list, if we see it
2573 * unasserted, it's guaranteed that the next sibling hasn't
2574 * finished its grace period even if it's already removed, and thus
2575 * safe to dereference from this RCU critical section. If
2576 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2577 * to be visible as %true here.
2579 * If @pos is dead, its next pointer can't be dereferenced;
2580 * however, as each cgroup is given a monotonically increasing
2581 * unique serial number and always appended to the sibling list,
2582 * the next one can be found by walking the parent's children until
2583 * we see a cgroup with higher serial number than @pos's. While
2584 * this path can be slower, it's taken only when either the current
2585 * cgroup is removed or iteration and removal race.
2588 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2589 } else if (likely(!cgroup_is_dead(pos))) {
2590 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2592 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2593 if (next->serial_nr > pos->serial_nr)
2597 if (&next->sibling == &cgrp->children)
2600 return cgroup_css(next, parent_css->ss);
2604 * css_next_descendant_pre - find the next descendant for pre-order walk
2605 * @pos: the current position (%NULL to initiate traversal)
2606 * @root: css whose descendants to walk
2608 * To be used by css_for_each_descendant_pre(). Find the next descendant
2609 * to visit for pre-order traversal of @root's descendants. @root is
2610 * included in the iteration and the first node to be visited.
2612 * While this function requires cgroup_mutex or RCU read locking, it
2613 * doesn't require the whole traversal to be contained in a single critical
2614 * section. This function will return the correct next descendant as long
2615 * as both @pos and @root are accessible and @pos is a descendant of @root.
2617 struct cgroup_subsys_state *
2618 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2619 struct cgroup_subsys_state *root)
2621 struct cgroup_subsys_state *next;
2623 cgroup_assert_mutexes_or_rcu_locked();
2625 /* if first iteration, visit @root */
2629 /* visit the first child if exists */
2630 next = css_next_child(NULL, pos);
2634 /* no child, visit my or the closest ancestor's next sibling */
2635 while (pos != root) {
2636 next = css_next_child(pos, css_parent(pos));
2639 pos = css_parent(pos);
2646 * css_rightmost_descendant - return the rightmost descendant of a css
2647 * @pos: css of interest
2649 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2650 * is returned. This can be used during pre-order traversal to skip
2653 * While this function requires cgroup_mutex or RCU read locking, it
2654 * doesn't require the whole traversal to be contained in a single critical
2655 * section. This function will return the correct rightmost descendant as
2656 * long as @pos is accessible.
2658 struct cgroup_subsys_state *
2659 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2661 struct cgroup_subsys_state *last, *tmp;
2663 cgroup_assert_mutexes_or_rcu_locked();
2667 /* ->prev isn't RCU safe, walk ->next till the end */
2669 css_for_each_child(tmp, last)
2676 static struct cgroup_subsys_state *
2677 css_leftmost_descendant(struct cgroup_subsys_state *pos)
2679 struct cgroup_subsys_state *last;
2683 pos = css_next_child(NULL, pos);
2690 * css_next_descendant_post - find the next descendant for post-order walk
2691 * @pos: the current position (%NULL to initiate traversal)
2692 * @root: css whose descendants to walk
2694 * To be used by css_for_each_descendant_post(). Find the next descendant
2695 * to visit for post-order traversal of @root's descendants. @root is
2696 * included in the iteration and the last node to be visited.
2698 * While this function requires cgroup_mutex or RCU read locking, it
2699 * doesn't require the whole traversal to be contained in a single critical
2700 * section. This function will return the correct next descendant as long
2701 * as both @pos and @cgroup are accessible and @pos is a descendant of
2704 struct cgroup_subsys_state *
2705 css_next_descendant_post(struct cgroup_subsys_state *pos,
2706 struct cgroup_subsys_state *root)
2708 struct cgroup_subsys_state *next;
2710 cgroup_assert_mutexes_or_rcu_locked();
2712 /* if first iteration, visit leftmost descendant which may be @root */
2714 return css_leftmost_descendant(root);
2716 /* if we visited @root, we're done */
2720 /* if there's an unvisited sibling, visit its leftmost descendant */
2721 next = css_next_child(pos, css_parent(pos));
2723 return css_leftmost_descendant(next);
2725 /* no sibling left, visit parent */
2726 return css_parent(pos);
2730 * css_advance_task_iter - advance a task itererator to the next css_set
2731 * @it: the iterator to advance
2733 * Advance @it to the next css_set to walk.
2735 static void css_advance_task_iter(struct css_task_iter *it)
2737 struct list_head *l = it->cset_link;
2738 struct cgrp_cset_link *link;
2739 struct css_set *cset;
2741 /* Advance to the next non-empty css_set */
2744 if (l == &it->origin_css->cgroup->cset_links) {
2745 it->cset_link = NULL;
2748 link = list_entry(l, struct cgrp_cset_link, cset_link);
2750 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
2754 if (!list_empty(&cset->tasks))
2755 it->task = cset->tasks.next;
2757 it->task = cset->mg_tasks.next;
2761 * css_task_iter_start - initiate task iteration
2762 * @css: the css to walk tasks of
2763 * @it: the task iterator to use
2765 * Initiate iteration through the tasks of @css. The caller can call
2766 * css_task_iter_next() to walk through the tasks until the function
2767 * returns NULL. On completion of iteration, css_task_iter_end() must be
2770 * Note that this function acquires a lock which is released when the
2771 * iteration finishes. The caller can't sleep while iteration is in
2774 void css_task_iter_start(struct cgroup_subsys_state *css,
2775 struct css_task_iter *it)
2776 __acquires(css_set_rwsem)
2778 /* no one should try to iterate before mounting cgroups */
2779 WARN_ON_ONCE(!use_task_css_set_links);
2781 down_read(&css_set_rwsem);
2783 it->origin_css = css;
2784 it->cset_link = &css->cgroup->cset_links;
2786 css_advance_task_iter(it);
2790 * css_task_iter_next - return the next task for the iterator
2791 * @it: the task iterator being iterated
2793 * The "next" function for task iteration. @it should have been
2794 * initialized via css_task_iter_start(). Returns NULL when the iteration
2797 struct task_struct *css_task_iter_next(struct css_task_iter *it)
2799 struct task_struct *res;
2800 struct list_head *l = it->task;
2801 struct cgrp_cset_link *link = list_entry(it->cset_link,
2802 struct cgrp_cset_link, cset_link);
2804 /* If the iterator cg is NULL, we have no tasks */
2807 res = list_entry(l, struct task_struct, cg_list);
2810 * Advance iterator to find next entry. cset->tasks is consumed
2811 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
2816 if (l == &link->cset->tasks)
2817 l = link->cset->mg_tasks.next;
2819 if (l == &link->cset->mg_tasks)
2820 css_advance_task_iter(it);
2828 * css_task_iter_end - finish task iteration
2829 * @it: the task iterator to finish
2831 * Finish task iteration started by css_task_iter_start().
2833 void css_task_iter_end(struct css_task_iter *it)
2834 __releases(css_set_rwsem)
2836 up_read(&css_set_rwsem);
2840 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
2841 * @to: cgroup to which the tasks will be moved
2842 * @from: cgroup in which the tasks currently reside
2844 * Locking rules between cgroup_post_fork() and the migration path
2845 * guarantee that, if a task is forking while being migrated, the new child
2846 * is guaranteed to be either visible in the source cgroup after the
2847 * parent's migration is complete or put into the target cgroup. No task
2848 * can slip out of migration through forking.
2850 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
2852 LIST_HEAD(preloaded_csets);
2853 struct cgrp_cset_link *link;
2854 struct css_task_iter it;
2855 struct task_struct *task;
2858 mutex_lock(&cgroup_mutex);
2860 /* all tasks in @from are being moved, all csets are source */
2861 down_read(&css_set_rwsem);
2862 list_for_each_entry(link, &from->cset_links, cset_link)
2863 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
2864 up_read(&css_set_rwsem);
2866 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
2871 * Migrate tasks one-by-one until @form is empty. This fails iff
2872 * ->can_attach() fails.
2875 css_task_iter_start(&from->dummy_css, &it);
2876 task = css_task_iter_next(&it);
2878 get_task_struct(task);
2879 css_task_iter_end(&it);
2882 ret = cgroup_migrate(to, task, false);
2883 put_task_struct(task);
2885 } while (task && !ret);
2887 cgroup_migrate_finish(&preloaded_csets);
2888 mutex_unlock(&cgroup_mutex);
2893 * Stuff for reading the 'tasks'/'procs' files.
2895 * Reading this file can return large amounts of data if a cgroup has
2896 * *lots* of attached tasks. So it may need several calls to read(),
2897 * but we cannot guarantee that the information we produce is correct
2898 * unless we produce it entirely atomically.
2902 /* which pidlist file are we talking about? */
2903 enum cgroup_filetype {
2909 * A pidlist is a list of pids that virtually represents the contents of one
2910 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
2911 * a pair (one each for procs, tasks) for each pid namespace that's relevant
2914 struct cgroup_pidlist {
2916 * used to find which pidlist is wanted. doesn't change as long as
2917 * this particular list stays in the list.
2919 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
2922 /* how many elements the above list has */
2924 /* each of these stored in a list by its cgroup */
2925 struct list_head links;
2926 /* pointer to the cgroup we belong to, for list removal purposes */
2927 struct cgroup *owner;
2928 /* for delayed destruction */
2929 struct delayed_work destroy_dwork;
2933 * The following two functions "fix" the issue where there are more pids
2934 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2935 * TODO: replace with a kernel-wide solution to this problem
2937 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2938 static void *pidlist_allocate(int count)
2940 if (PIDLIST_TOO_LARGE(count))
2941 return vmalloc(count * sizeof(pid_t));
2943 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2946 static void pidlist_free(void *p)
2948 if (is_vmalloc_addr(p))
2955 * Used to destroy all pidlists lingering waiting for destroy timer. None
2956 * should be left afterwards.
2958 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
2960 struct cgroup_pidlist *l, *tmp_l;
2962 mutex_lock(&cgrp->pidlist_mutex);
2963 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
2964 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
2965 mutex_unlock(&cgrp->pidlist_mutex);
2967 flush_workqueue(cgroup_pidlist_destroy_wq);
2968 BUG_ON(!list_empty(&cgrp->pidlists));
2971 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
2973 struct delayed_work *dwork = to_delayed_work(work);
2974 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
2976 struct cgroup_pidlist *tofree = NULL;
2978 mutex_lock(&l->owner->pidlist_mutex);
2981 * Destroy iff we didn't get queued again. The state won't change
2982 * as destroy_dwork can only be queued while locked.
2984 if (!delayed_work_pending(dwork)) {
2985 list_del(&l->links);
2986 pidlist_free(l->list);
2987 put_pid_ns(l->key.ns);
2991 mutex_unlock(&l->owner->pidlist_mutex);
2996 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2997 * Returns the number of unique elements.
2999 static int pidlist_uniq(pid_t *list, int length)
3004 * we presume the 0th element is unique, so i starts at 1. trivial
3005 * edge cases first; no work needs to be done for either
3007 if (length == 0 || length == 1)
3009 /* src and dest walk down the list; dest counts unique elements */
3010 for (src = 1; src < length; src++) {
3011 /* find next unique element */
3012 while (list[src] == list[src-1]) {
3017 /* dest always points to where the next unique element goes */
3018 list[dest] = list[src];
3026 * The two pid files - task and cgroup.procs - guaranteed that the result
3027 * is sorted, which forced this whole pidlist fiasco. As pid order is
3028 * different per namespace, each namespace needs differently sorted list,
3029 * making it impossible to use, for example, single rbtree of member tasks
3030 * sorted by task pointer. As pidlists can be fairly large, allocating one
3031 * per open file is dangerous, so cgroup had to implement shared pool of
3032 * pidlists keyed by cgroup and namespace.
3034 * All this extra complexity was caused by the original implementation
3035 * committing to an entirely unnecessary property. In the long term, we
3036 * want to do away with it. Explicitly scramble sort order if
3037 * sane_behavior so that no such expectation exists in the new interface.
3039 * Scrambling is done by swapping every two consecutive bits, which is
3040 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3042 static pid_t pid_fry(pid_t pid)
3044 unsigned a = pid & 0x55555555;
3045 unsigned b = pid & 0xAAAAAAAA;
3047 return (a << 1) | (b >> 1);
3050 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3052 if (cgroup_sane_behavior(cgrp))
3053 return pid_fry(pid);
3058 static int cmppid(const void *a, const void *b)
3060 return *(pid_t *)a - *(pid_t *)b;
3063 static int fried_cmppid(const void *a, const void *b)
3065 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3068 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3069 enum cgroup_filetype type)
3071 struct cgroup_pidlist *l;
3072 /* don't need task_nsproxy() if we're looking at ourself */
3073 struct pid_namespace *ns = task_active_pid_ns(current);
3075 lockdep_assert_held(&cgrp->pidlist_mutex);
3077 list_for_each_entry(l, &cgrp->pidlists, links)
3078 if (l->key.type == type && l->key.ns == ns)
3084 * find the appropriate pidlist for our purpose (given procs vs tasks)
3085 * returns with the lock on that pidlist already held, and takes care
3086 * of the use count, or returns NULL with no locks held if we're out of
3089 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3090 enum cgroup_filetype type)
3092 struct cgroup_pidlist *l;
3094 lockdep_assert_held(&cgrp->pidlist_mutex);
3096 l = cgroup_pidlist_find(cgrp, type);
3100 /* entry not found; create a new one */
3101 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3105 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3107 /* don't need task_nsproxy() if we're looking at ourself */
3108 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3110 list_add(&l->links, &cgrp->pidlists);
3115 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3117 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3118 struct cgroup_pidlist **lp)
3122 int pid, n = 0; /* used for populating the array */
3123 struct css_task_iter it;
3124 struct task_struct *tsk;
3125 struct cgroup_pidlist *l;
3127 lockdep_assert_held(&cgrp->pidlist_mutex);
3130 * If cgroup gets more users after we read count, we won't have
3131 * enough space - tough. This race is indistinguishable to the
3132 * caller from the case that the additional cgroup users didn't
3133 * show up until sometime later on.
3135 length = cgroup_task_count(cgrp);
3136 array = pidlist_allocate(length);
3139 /* now, populate the array */
3140 css_task_iter_start(&cgrp->dummy_css, &it);
3141 while ((tsk = css_task_iter_next(&it))) {
3142 if (unlikely(n == length))
3144 /* get tgid or pid for procs or tasks file respectively */
3145 if (type == CGROUP_FILE_PROCS)
3146 pid = task_tgid_vnr(tsk);
3148 pid = task_pid_vnr(tsk);
3149 if (pid > 0) /* make sure to only use valid results */
3152 css_task_iter_end(&it);
3154 /* now sort & (if procs) strip out duplicates */
3155 if (cgroup_sane_behavior(cgrp))
3156 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3158 sort(array, length, sizeof(pid_t), cmppid, NULL);
3159 if (type == CGROUP_FILE_PROCS)
3160 length = pidlist_uniq(array, length);
3162 l = cgroup_pidlist_find_create(cgrp, type);
3164 mutex_unlock(&cgrp->pidlist_mutex);
3165 pidlist_free(array);
3169 /* store array, freeing old if necessary */
3170 pidlist_free(l->list);
3178 * cgroupstats_build - build and fill cgroupstats
3179 * @stats: cgroupstats to fill information into
3180 * @dentry: A dentry entry belonging to the cgroup for which stats have
3183 * Build and fill cgroupstats so that taskstats can export it to user
3186 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3188 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3189 struct cgroup *cgrp;
3190 struct css_task_iter it;
3191 struct task_struct *tsk;
3193 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3194 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3195 kernfs_type(kn) != KERNFS_DIR)
3198 mutex_lock(&cgroup_mutex);
3201 * We aren't being called from kernfs and there's no guarantee on
3202 * @kn->priv's validity. For this and css_tryget_from_dir(),
3203 * @kn->priv is RCU safe. Let's do the RCU dancing.
3206 cgrp = rcu_dereference(kn->priv);
3207 if (!cgrp || cgroup_is_dead(cgrp)) {
3209 mutex_unlock(&cgroup_mutex);
3214 css_task_iter_start(&cgrp->dummy_css, &it);
3215 while ((tsk = css_task_iter_next(&it))) {
3216 switch (tsk->state) {
3218 stats->nr_running++;
3220 case TASK_INTERRUPTIBLE:
3221 stats->nr_sleeping++;
3223 case TASK_UNINTERRUPTIBLE:
3224 stats->nr_uninterruptible++;
3227 stats->nr_stopped++;
3230 if (delayacct_is_task_waiting_on_io(tsk))
3231 stats->nr_io_wait++;
3235 css_task_iter_end(&it);
3237 mutex_unlock(&cgroup_mutex);
3243 * seq_file methods for the tasks/procs files. The seq_file position is the
3244 * next pid to display; the seq_file iterator is a pointer to the pid
3245 * in the cgroup->l->list array.
3248 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3251 * Initially we receive a position value that corresponds to
3252 * one more than the last pid shown (or 0 on the first call or
3253 * after a seek to the start). Use a binary-search to find the
3254 * next pid to display, if any
3256 struct kernfs_open_file *of = s->private;
3257 struct cgroup *cgrp = seq_css(s)->cgroup;
3258 struct cgroup_pidlist *l;
3259 enum cgroup_filetype type = seq_cft(s)->private;
3260 int index = 0, pid = *pos;
3263 mutex_lock(&cgrp->pidlist_mutex);
3266 * !NULL @of->priv indicates that this isn't the first start()
3267 * after open. If the matching pidlist is around, we can use that.
3268 * Look for it. Note that @of->priv can't be used directly. It
3269 * could already have been destroyed.
3272 of->priv = cgroup_pidlist_find(cgrp, type);
3275 * Either this is the first start() after open or the matching
3276 * pidlist has been destroyed inbetween. Create a new one.
3279 ret = pidlist_array_load(cgrp, type,
3280 (struct cgroup_pidlist **)&of->priv);
3282 return ERR_PTR(ret);
3287 int end = l->length;
3289 while (index < end) {
3290 int mid = (index + end) / 2;
3291 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3294 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3300 /* If we're off the end of the array, we're done */
3301 if (index >= l->length)
3303 /* Update the abstract position to be the actual pid that we found */
3304 iter = l->list + index;
3305 *pos = cgroup_pid_fry(cgrp, *iter);
3309 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3311 struct kernfs_open_file *of = s->private;
3312 struct cgroup_pidlist *l = of->priv;
3315 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3316 CGROUP_PIDLIST_DESTROY_DELAY);
3317 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3320 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3322 struct kernfs_open_file *of = s->private;
3323 struct cgroup_pidlist *l = of->priv;
3325 pid_t *end = l->list + l->length;
3327 * Advance to the next pid in the array. If this goes off the
3334 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3339 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3341 return seq_printf(s, "%d\n", *(int *)v);
3345 * seq_operations functions for iterating on pidlists through seq_file -
3346 * independent of whether it's tasks or procs
3348 static const struct seq_operations cgroup_pidlist_seq_operations = {
3349 .start = cgroup_pidlist_start,
3350 .stop = cgroup_pidlist_stop,
3351 .next = cgroup_pidlist_next,
3352 .show = cgroup_pidlist_show,
3355 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3358 return notify_on_release(css->cgroup);
3361 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3362 struct cftype *cft, u64 val)
3364 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3366 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3368 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3372 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3375 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3378 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3379 struct cftype *cft, u64 val)
3382 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3384 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3388 static struct cftype cgroup_base_files[] = {
3390 .name = "cgroup.procs",
3391 .seq_start = cgroup_pidlist_start,
3392 .seq_next = cgroup_pidlist_next,
3393 .seq_stop = cgroup_pidlist_stop,
3394 .seq_show = cgroup_pidlist_show,
3395 .private = CGROUP_FILE_PROCS,
3396 .write_u64 = cgroup_procs_write,
3397 .mode = S_IRUGO | S_IWUSR,
3400 .name = "cgroup.clone_children",
3401 .flags = CFTYPE_INSANE,
3402 .read_u64 = cgroup_clone_children_read,
3403 .write_u64 = cgroup_clone_children_write,
3406 .name = "cgroup.sane_behavior",
3407 .flags = CFTYPE_ONLY_ON_ROOT,
3408 .seq_show = cgroup_sane_behavior_show,
3412 * Historical crazy stuff. These don't have "cgroup." prefix and
3413 * don't exist if sane_behavior. If you're depending on these, be
3414 * prepared to be burned.
3418 .flags = CFTYPE_INSANE, /* use "procs" instead */
3419 .seq_start = cgroup_pidlist_start,
3420 .seq_next = cgroup_pidlist_next,
3421 .seq_stop = cgroup_pidlist_stop,
3422 .seq_show = cgroup_pidlist_show,
3423 .private = CGROUP_FILE_TASKS,
3424 .write_u64 = cgroup_tasks_write,
3425 .mode = S_IRUGO | S_IWUSR,
3428 .name = "notify_on_release",
3429 .flags = CFTYPE_INSANE,
3430 .read_u64 = cgroup_read_notify_on_release,
3431 .write_u64 = cgroup_write_notify_on_release,
3434 .name = "release_agent",
3435 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3436 .seq_show = cgroup_release_agent_show,
3437 .write_string = cgroup_release_agent_write,
3438 .max_write_len = PATH_MAX - 1,
3444 * cgroup_populate_dir - create subsys files in a cgroup directory
3445 * @cgrp: target cgroup
3446 * @subsys_mask: mask of the subsystem ids whose files should be added
3448 * On failure, no file is added.
3450 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3452 struct cgroup_subsys *ss;
3455 /* process cftsets of each subsystem */
3456 for_each_subsys(ss, i) {
3457 struct cftype *cfts;
3459 if (!test_bit(i, &subsys_mask))
3462 list_for_each_entry(cfts, &ss->cfts, node) {
3463 ret = cgroup_addrm_files(cgrp, cfts, true);
3470 cgroup_clear_dir(cgrp, subsys_mask);
3475 * css destruction is four-stage process.
3477 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3478 * Implemented in kill_css().
3480 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3481 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3482 * by invoking offline_css(). After offlining, the base ref is put.
3483 * Implemented in css_killed_work_fn().
3485 * 3. When the percpu_ref reaches zero, the only possible remaining
3486 * accessors are inside RCU read sections. css_release() schedules the
3489 * 4. After the grace period, the css can be freed. Implemented in
3490 * css_free_work_fn().
3492 * It is actually hairier because both step 2 and 4 require process context
3493 * and thus involve punting to css->destroy_work adding two additional
3494 * steps to the already complex sequence.
3496 static void css_free_work_fn(struct work_struct *work)
3498 struct cgroup_subsys_state *css =
3499 container_of(work, struct cgroup_subsys_state, destroy_work);
3500 struct cgroup *cgrp = css->cgroup;
3503 css_put(css->parent);
3505 css->ss->css_free(css);
3509 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3511 struct cgroup_subsys_state *css =
3512 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3514 INIT_WORK(&css->destroy_work, css_free_work_fn);
3515 queue_work(cgroup_destroy_wq, &css->destroy_work);
3518 static void css_release(struct percpu_ref *ref)
3520 struct cgroup_subsys_state *css =
3521 container_of(ref, struct cgroup_subsys_state, refcnt);
3523 rcu_assign_pointer(css->cgroup->subsys[css->ss->id], NULL);
3524 call_rcu(&css->rcu_head, css_free_rcu_fn);
3527 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3528 struct cgroup *cgrp)
3535 css->parent = cgroup_css(cgrp->parent, ss);
3537 css->flags |= CSS_ROOT;
3539 BUG_ON(cgroup_css(cgrp, ss));
3542 /* invoke ->css_online() on a new CSS and mark it online if successful */
3543 static int online_css(struct cgroup_subsys_state *css)
3545 struct cgroup_subsys *ss = css->ss;
3548 lockdep_assert_held(&cgroup_tree_mutex);
3549 lockdep_assert_held(&cgroup_mutex);
3552 ret = ss->css_online(css);
3554 css->flags |= CSS_ONLINE;
3555 css->cgroup->nr_css++;
3556 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
3561 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
3562 static void offline_css(struct cgroup_subsys_state *css)
3564 struct cgroup_subsys *ss = css->ss;
3566 lockdep_assert_held(&cgroup_tree_mutex);
3567 lockdep_assert_held(&cgroup_mutex);
3569 if (!(css->flags & CSS_ONLINE))
3572 if (ss->css_offline)
3573 ss->css_offline(css);
3575 css->flags &= ~CSS_ONLINE;
3576 css->cgroup->nr_css--;
3577 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
3581 * create_css - create a cgroup_subsys_state
3582 * @cgrp: the cgroup new css will be associated with
3583 * @ss: the subsys of new css
3585 * Create a new css associated with @cgrp - @ss pair. On success, the new
3586 * css is online and installed in @cgrp with all interface files created.
3587 * Returns 0 on success, -errno on failure.
3589 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
3591 struct cgroup *parent = cgrp->parent;
3592 struct cgroup_subsys_state *css;
3595 lockdep_assert_held(&cgroup_mutex);
3597 css = ss->css_alloc(cgroup_css(parent, ss));
3599 return PTR_ERR(css);
3601 err = percpu_ref_init(&css->refcnt, css_release);
3605 init_css(css, ss, cgrp);
3607 err = cgroup_populate_dir(cgrp, 1 << ss->id);
3611 err = online_css(css);
3616 css_get(css->parent);
3618 cgrp->subsys_mask |= 1 << ss->id;
3620 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
3622 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",
3623 current->comm, current->pid, ss->name);
3624 if (!strcmp(ss->name, "memory"))
3625 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
3626 ss->warned_broken_hierarchy = true;
3632 percpu_ref_cancel_init(&css->refcnt);
3638 * cgroup_create - create a cgroup
3639 * @parent: cgroup that will be parent of the new cgroup
3640 * @name: name of the new cgroup
3641 * @mode: mode to set on new cgroup
3643 static long cgroup_create(struct cgroup *parent, const char *name,
3646 struct cgroup *cgrp;
3647 struct cgroup_root *root = parent->root;
3649 struct cgroup_subsys *ss;
3650 struct kernfs_node *kn;
3653 * XXX: The default hierarchy isn't fully implemented yet. Block
3654 * !root cgroup creation on it for now.
3656 if (root == &cgrp_dfl_root)
3659 /* allocate the cgroup and its ID, 0 is reserved for the root */
3660 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3664 mutex_lock(&cgroup_tree_mutex);
3667 * Only live parents can have children. Note that the liveliness
3668 * check isn't strictly necessary because cgroup_mkdir() and
3669 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
3670 * anyway so that locking is contained inside cgroup proper and we
3671 * don't get nasty surprises if we ever grow another caller.
3673 if (!cgroup_lock_live_group(parent)) {
3675 goto err_unlock_tree;
3679 * Temporarily set the pointer to NULL, so idr_find() won't return
3680 * a half-baked cgroup.
3682 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
3688 init_cgroup_housekeeping(cgrp);
3690 cgrp->parent = parent;
3691 cgrp->dummy_css.parent = &parent->dummy_css;
3692 cgrp->root = parent->root;
3694 if (notify_on_release(parent))
3695 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3697 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
3698 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
3700 /* create the directory */
3701 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
3709 * This extra ref will be put in cgroup_free_fn() and guarantees
3710 * that @cgrp->kn is always accessible.
3714 cgrp->serial_nr = cgroup_serial_nr_next++;
3716 /* allocation complete, commit to creation */
3717 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
3718 atomic_inc(&root->nr_cgrps);
3722 * @cgrp is now fully operational. If something fails after this
3723 * point, it'll be released via the normal destruction path.
3725 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
3727 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
3731 /* let's create and online css's */
3732 for_each_subsys(ss, ssid) {
3733 if (root->cgrp.subsys_mask & (1 << ssid)) {
3734 err = create_css(cgrp, ss);
3740 kernfs_activate(kn);
3742 mutex_unlock(&cgroup_mutex);
3743 mutex_unlock(&cgroup_tree_mutex);
3748 idr_remove(&root->cgroup_idr, cgrp->id);
3750 mutex_unlock(&cgroup_mutex);
3752 mutex_unlock(&cgroup_tree_mutex);
3757 cgroup_destroy_locked(cgrp);
3758 mutex_unlock(&cgroup_mutex);
3759 mutex_unlock(&cgroup_tree_mutex);
3763 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3766 struct cgroup *parent = parent_kn->priv;
3768 return cgroup_create(parent, name, mode);
3772 * This is called when the refcnt of a css is confirmed to be killed.
3773 * css_tryget() is now guaranteed to fail.
3775 static void css_killed_work_fn(struct work_struct *work)
3777 struct cgroup_subsys_state *css =
3778 container_of(work, struct cgroup_subsys_state, destroy_work);
3779 struct cgroup *cgrp = css->cgroup;
3781 mutex_lock(&cgroup_tree_mutex);
3782 mutex_lock(&cgroup_mutex);
3785 * css_tryget() is guaranteed to fail now. Tell subsystems to
3786 * initate destruction.
3791 * If @cgrp is marked dead, it's waiting for refs of all css's to
3792 * be disabled before proceeding to the second phase of cgroup
3793 * destruction. If we are the last one, kick it off.
3795 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
3796 cgroup_destroy_css_killed(cgrp);
3798 mutex_unlock(&cgroup_mutex);
3799 mutex_unlock(&cgroup_tree_mutex);
3802 * Put the css refs from kill_css(). Each css holds an extra
3803 * reference to the cgroup's dentry and cgroup removal proceeds
3804 * regardless of css refs. On the last put of each css, whenever
3805 * that may be, the extra dentry ref is put so that dentry
3806 * destruction happens only after all css's are released.
3811 /* css kill confirmation processing requires process context, bounce */
3812 static void css_killed_ref_fn(struct percpu_ref *ref)
3814 struct cgroup_subsys_state *css =
3815 container_of(ref, struct cgroup_subsys_state, refcnt);
3817 INIT_WORK(&css->destroy_work, css_killed_work_fn);
3818 queue_work(cgroup_destroy_wq, &css->destroy_work);
3821 static void __kill_css(struct cgroup_subsys_state *css)
3823 lockdep_assert_held(&cgroup_tree_mutex);
3826 * This must happen before css is disassociated with its cgroup.
3827 * See seq_css() for details.
3829 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3832 * Killing would put the base ref, but we need to keep it alive
3833 * until after ->css_offline().
3838 * cgroup core guarantees that, by the time ->css_offline() is
3839 * invoked, no new css reference will be given out via
3840 * css_tryget(). We can't simply call percpu_ref_kill() and
3841 * proceed to offlining css's because percpu_ref_kill() doesn't
3842 * guarantee that the ref is seen as killed on all CPUs on return.
3844 * Use percpu_ref_kill_and_confirm() to get notifications as each
3845 * css is confirmed to be seen as killed on all CPUs.
3847 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
3851 * kill_css - destroy a css
3852 * @css: css to destroy
3854 * This function initiates destruction of @css by removing cgroup interface
3855 * files and putting its base reference. ->css_offline() will be invoked
3856 * asynchronously once css_tryget() is guaranteed to fail and when the
3857 * reference count reaches zero, @css will be released.
3859 static void kill_css(struct cgroup_subsys_state *css)
3861 struct cgroup *cgrp = css->cgroup;
3863 lockdep_assert_held(&cgroup_tree_mutex);
3865 /* if already killed, noop */
3866 if (cgrp->subsys_mask & (1 << css->ss->id)) {
3867 cgrp->subsys_mask &= ~(1 << css->ss->id);
3873 * cgroup_destroy_locked - the first stage of cgroup destruction
3874 * @cgrp: cgroup to be destroyed
3876 * css's make use of percpu refcnts whose killing latency shouldn't be
3877 * exposed to userland and are RCU protected. Also, cgroup core needs to
3878 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
3879 * invoked. To satisfy all the requirements, destruction is implemented in
3880 * the following two steps.
3882 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
3883 * userland visible parts and start killing the percpu refcnts of
3884 * css's. Set up so that the next stage will be kicked off once all
3885 * the percpu refcnts are confirmed to be killed.
3887 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
3888 * rest of destruction. Once all cgroup references are gone, the
3889 * cgroup is RCU-freed.
3891 * This function implements s1. After this step, @cgrp is gone as far as
3892 * the userland is concerned and a new cgroup with the same name may be
3893 * created. As cgroup doesn't care about the names internally, this
3894 * doesn't cause any problem.
3896 static int cgroup_destroy_locked(struct cgroup *cgrp)
3897 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
3899 struct cgroup *child;
3900 struct cgroup_subsys_state *css;
3904 lockdep_assert_held(&cgroup_tree_mutex);
3905 lockdep_assert_held(&cgroup_mutex);
3908 * css_set_rwsem synchronizes access to ->cset_links and prevents
3909 * @cgrp from being removed while put_css_set() is in progress.
3911 down_read(&css_set_rwsem);
3912 empty = list_empty(&cgrp->cset_links);
3913 up_read(&css_set_rwsem);
3918 * Make sure there's no live children. We can't test ->children
3919 * emptiness as dead children linger on it while being destroyed;
3920 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
3924 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
3925 empty = cgroup_is_dead(child);
3934 * Mark @cgrp dead. This prevents further task migration and child
3935 * creation by disabling cgroup_lock_live_group(). Note that
3936 * CGRP_DEAD assertion is depended upon by css_next_child() to
3937 * resume iteration after dropping RCU read lock. See
3938 * css_next_child() for details.
3940 set_bit(CGRP_DEAD, &cgrp->flags);
3943 * Initiate massacre of all css's. cgroup_destroy_css_killed()
3944 * will be invoked to perform the rest of destruction once the
3945 * percpu refs of all css's are confirmed to be killed. This
3946 * involves removing the subsystem's files, drop cgroup_mutex.
3948 mutex_unlock(&cgroup_mutex);
3949 for_each_css(css, ssid, cgrp)
3951 mutex_lock(&cgroup_mutex);
3953 /* CGRP_DEAD is set, remove from ->release_list for the last time */
3954 raw_spin_lock(&release_list_lock);
3955 if (!list_empty(&cgrp->release_list))
3956 list_del_init(&cgrp->release_list);
3957 raw_spin_unlock(&release_list_lock);
3960 * If @cgrp has css's attached, the second stage of cgroup
3961 * destruction is kicked off from css_killed_work_fn() after the
3962 * refs of all attached css's are killed. If @cgrp doesn't have
3963 * any css, we kick it off here.
3966 cgroup_destroy_css_killed(cgrp);
3968 /* remove @cgrp directory along with the base files */
3969 mutex_unlock(&cgroup_mutex);
3972 * There are two control paths which try to determine cgroup from
3973 * dentry without going through kernfs - cgroupstats_build() and
3974 * css_tryget_from_dir(). Those are supported by RCU protecting
3975 * clearing of cgrp->kn->priv backpointer, which should happen
3976 * after all files under it have been removed.
3978 kernfs_remove(cgrp->kn); /* @cgrp has an extra ref on its kn */
3979 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
3981 mutex_lock(&cgroup_mutex);
3987 * cgroup_destroy_css_killed - the second step of cgroup destruction
3988 * @work: cgroup->destroy_free_work
3990 * This function is invoked from a work item for a cgroup which is being
3991 * destroyed after all css's are offlined and performs the rest of
3992 * destruction. This is the second step of destruction described in the
3993 * comment above cgroup_destroy_locked().
3995 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
3997 struct cgroup *parent = cgrp->parent;
3999 lockdep_assert_held(&cgroup_tree_mutex);
4000 lockdep_assert_held(&cgroup_mutex);
4002 /* delete this cgroup from parent->children */
4003 list_del_rcu(&cgrp->sibling);
4007 set_bit(CGRP_RELEASABLE, &parent->flags);
4008 check_for_release(parent);
4011 static int cgroup_rmdir(struct kernfs_node *kn)
4013 struct cgroup *cgrp = kn->priv;
4017 * This is self-destruction but @kn can't be removed while this
4018 * callback is in progress. Let's break active protection. Once
4019 * the protection is broken, @cgrp can be destroyed at any point.
4020 * Pin it so that it stays accessible.
4023 kernfs_break_active_protection(kn);
4025 mutex_lock(&cgroup_tree_mutex);
4026 mutex_lock(&cgroup_mutex);
4029 * @cgrp might already have been destroyed while we're trying to
4032 if (!cgroup_is_dead(cgrp))
4033 ret = cgroup_destroy_locked(cgrp);
4035 mutex_unlock(&cgroup_mutex);
4036 mutex_unlock(&cgroup_tree_mutex);
4038 kernfs_unbreak_active_protection(kn);
4043 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4044 .remount_fs = cgroup_remount,
4045 .show_options = cgroup_show_options,
4046 .mkdir = cgroup_mkdir,
4047 .rmdir = cgroup_rmdir,
4048 .rename = cgroup_rename,
4051 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4053 struct cgroup_subsys_state *css;
4055 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4057 mutex_lock(&cgroup_tree_mutex);
4058 mutex_lock(&cgroup_mutex);
4060 INIT_LIST_HEAD(&ss->cfts);
4062 /* Create the root cgroup state for this subsystem */
4063 ss->root = &cgrp_dfl_root;
4064 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4065 /* We don't handle early failures gracefully */
4066 BUG_ON(IS_ERR(css));
4067 init_css(css, ss, &cgrp_dfl_root.cgrp);
4069 /* Update the init_css_set to contain a subsys
4070 * pointer to this state - since the subsystem is
4071 * newly registered, all tasks and hence the
4072 * init_css_set is in the subsystem's root cgroup. */
4073 init_css_set.subsys[ss->id] = css;
4075 need_forkexit_callback |= ss->fork || ss->exit;
4077 /* At system boot, before all subsystems have been
4078 * registered, no tasks have been forked, so we don't
4079 * need to invoke fork callbacks here. */
4080 BUG_ON(!list_empty(&init_task.tasks));
4082 BUG_ON(online_css(css));
4084 cgrp_dfl_root.cgrp.subsys_mask |= 1 << ss->id;
4086 mutex_unlock(&cgroup_mutex);
4087 mutex_unlock(&cgroup_tree_mutex);
4091 * cgroup_init_early - cgroup initialization at system boot
4093 * Initialize cgroups at system boot, and initialize any
4094 * subsystems that request early init.
4096 int __init cgroup_init_early(void)
4098 static struct cgroup_sb_opts __initdata opts =
4099 { .flags = CGRP_ROOT_SANE_BEHAVIOR };
4100 struct cgroup_subsys *ss;
4103 init_cgroup_root(&cgrp_dfl_root, &opts);
4104 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4106 for_each_subsys(ss, i) {
4107 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4108 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4109 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4111 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4112 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4115 ss->name = cgroup_subsys_name[i];
4118 cgroup_init_subsys(ss);
4124 * cgroup_init - cgroup initialization
4126 * Register cgroup filesystem and /proc file, and initialize
4127 * any subsystems that didn't request early init.
4129 int __init cgroup_init(void)
4131 struct cgroup_subsys *ss;
4135 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4137 mutex_lock(&cgroup_tree_mutex);
4138 mutex_lock(&cgroup_mutex);
4140 /* Add init_css_set to the hash table */
4141 key = css_set_hash(init_css_set.subsys);
4142 hash_add(css_set_table, &init_css_set.hlist, key);
4144 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4146 mutex_unlock(&cgroup_mutex);
4147 mutex_unlock(&cgroup_tree_mutex);
4149 for_each_subsys(ss, ssid) {
4150 if (!ss->early_init)
4151 cgroup_init_subsys(ss);
4154 * cftype registration needs kmalloc and can't be done
4155 * during early_init. Register base cftypes separately.
4157 if (ss->base_cftypes)
4158 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4161 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4165 err = register_filesystem(&cgroup_fs_type);
4167 kobject_put(cgroup_kobj);
4171 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4175 static int __init cgroup_wq_init(void)
4178 * There isn't much point in executing destruction path in
4179 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4180 * Use 1 for @max_active.
4182 * We would prefer to do this in cgroup_init() above, but that
4183 * is called before init_workqueues(): so leave this until after.
4185 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
4186 BUG_ON(!cgroup_destroy_wq);
4189 * Used to destroy pidlists and separate to serve as flush domain.
4190 * Cap @max_active to 1 too.
4192 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4194 BUG_ON(!cgroup_pidlist_destroy_wq);
4198 core_initcall(cgroup_wq_init);
4201 * proc_cgroup_show()
4202 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4203 * - Used for /proc/<pid>/cgroup.
4206 /* TODO: Use a proper seq_file iterator */
4207 int proc_cgroup_show(struct seq_file *m, void *v)
4210 struct task_struct *tsk;
4213 struct cgroup_root *root;
4216 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4222 tsk = get_pid_task(pid, PIDTYPE_PID);
4228 mutex_lock(&cgroup_mutex);
4229 down_read(&css_set_rwsem);
4231 for_each_root(root) {
4232 struct cgroup_subsys *ss;
4233 struct cgroup *cgrp;
4234 int ssid, count = 0;
4236 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
4239 seq_printf(m, "%d:", root->hierarchy_id);
4240 for_each_subsys(ss, ssid)
4241 if (root->cgrp.subsys_mask & (1 << ssid))
4242 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4243 if (strlen(root->name))
4244 seq_printf(m, "%sname=%s", count ? "," : "",
4247 cgrp = task_cgroup_from_root(tsk, root);
4248 path = cgroup_path(cgrp, buf, PATH_MAX);
4250 retval = -ENAMETOOLONG;
4258 up_read(&css_set_rwsem);
4259 mutex_unlock(&cgroup_mutex);
4260 put_task_struct(tsk);
4267 /* Display information about each subsystem and each hierarchy */
4268 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4270 struct cgroup_subsys *ss;
4273 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4275 * ideally we don't want subsystems moving around while we do this.
4276 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4277 * subsys/hierarchy state.
4279 mutex_lock(&cgroup_mutex);
4281 for_each_subsys(ss, i)
4282 seq_printf(m, "%s\t%d\t%d\t%d\n",
4283 ss->name, ss->root->hierarchy_id,
4284 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4286 mutex_unlock(&cgroup_mutex);
4290 static int cgroupstats_open(struct inode *inode, struct file *file)
4292 return single_open(file, proc_cgroupstats_show, NULL);
4295 static const struct file_operations proc_cgroupstats_operations = {
4296 .open = cgroupstats_open,
4298 .llseek = seq_lseek,
4299 .release = single_release,
4303 * cgroup_fork - initialize cgroup related fields during copy_process()
4304 * @child: pointer to task_struct of forking parent process.
4306 * A task is associated with the init_css_set until cgroup_post_fork()
4307 * attaches it to the parent's css_set. Empty cg_list indicates that
4308 * @child isn't holding reference to its css_set.
4310 void cgroup_fork(struct task_struct *child)
4312 RCU_INIT_POINTER(child->cgroups, &init_css_set);
4313 INIT_LIST_HEAD(&child->cg_list);
4317 * cgroup_post_fork - called on a new task after adding it to the task list
4318 * @child: the task in question
4320 * Adds the task to the list running through its css_set if necessary and
4321 * call the subsystem fork() callbacks. Has to be after the task is
4322 * visible on the task list in case we race with the first call to
4323 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4326 void cgroup_post_fork(struct task_struct *child)
4328 struct cgroup_subsys *ss;
4332 * This may race against cgroup_enable_task_cg_links(). As that
4333 * function sets use_task_css_set_links before grabbing
4334 * tasklist_lock and we just went through tasklist_lock to add
4335 * @child, it's guaranteed that either we see the set
4336 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
4337 * @child during its iteration.
4339 * If we won the race, @child is associated with %current's
4340 * css_set. Grabbing css_set_rwsem guarantees both that the
4341 * association is stable, and, on completion of the parent's
4342 * migration, @child is visible in the source of migration or
4343 * already in the destination cgroup. This guarantee is necessary
4344 * when implementing operations which need to migrate all tasks of
4345 * a cgroup to another.
4347 * Note that if we lose to cgroup_enable_task_cg_links(), @child
4348 * will remain in init_css_set. This is safe because all tasks are
4349 * in the init_css_set before cg_links is enabled and there's no
4350 * operation which transfers all tasks out of init_css_set.
4352 if (use_task_css_set_links) {
4353 struct css_set *cset;
4355 down_write(&css_set_rwsem);
4356 cset = task_css_set(current);
4357 if (list_empty(&child->cg_list)) {
4358 rcu_assign_pointer(child->cgroups, cset);
4359 list_add(&child->cg_list, &cset->tasks);
4362 up_write(&css_set_rwsem);
4366 * Call ss->fork(). This must happen after @child is linked on
4367 * css_set; otherwise, @child might change state between ->fork()
4368 * and addition to css_set.
4370 if (need_forkexit_callback) {
4371 for_each_subsys(ss, i)
4378 * cgroup_exit - detach cgroup from exiting task
4379 * @tsk: pointer to task_struct of exiting process
4380 * @run_callback: run exit callbacks?
4382 * Description: Detach cgroup from @tsk and release it.
4384 * Note that cgroups marked notify_on_release force every task in
4385 * them to take the global cgroup_mutex mutex when exiting.
4386 * This could impact scaling on very large systems. Be reluctant to
4387 * use notify_on_release cgroups where very high task exit scaling
4388 * is required on large systems.
4390 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
4391 * call cgroup_exit() while the task is still competent to handle
4392 * notify_on_release(), then leave the task attached to the root cgroup in
4393 * each hierarchy for the remainder of its exit. No need to bother with
4394 * init_css_set refcnting. init_css_set never goes away and we can't race
4395 * with migration path - either PF_EXITING is visible to migration path or
4396 * @tsk never got on the tasklist.
4398 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4400 struct cgroup_subsys *ss;
4401 struct css_set *cset;
4402 bool put_cset = false;
4406 * Unlink from @tsk from its css_set. As migration path can't race
4407 * with us, we can check cg_list without grabbing css_set_rwsem.
4409 if (!list_empty(&tsk->cg_list)) {
4410 down_write(&css_set_rwsem);
4411 list_del_init(&tsk->cg_list);
4412 up_write(&css_set_rwsem);
4416 /* Reassign the task to the init_css_set. */
4417 cset = task_css_set(tsk);
4418 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4420 if (run_callbacks && need_forkexit_callback) {
4421 /* see cgroup_post_fork() for details */
4422 for_each_subsys(ss, i) {
4424 struct cgroup_subsys_state *old_css = cset->subsys[i];
4425 struct cgroup_subsys_state *css = task_css(tsk, i);
4427 ss->exit(css, old_css, tsk);
4433 put_css_set(cset, true);
4436 static void check_for_release(struct cgroup *cgrp)
4438 if (cgroup_is_releasable(cgrp) &&
4439 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4441 * Control Group is currently removeable. If it's not
4442 * already queued for a userspace notification, queue
4445 int need_schedule_work = 0;
4447 raw_spin_lock(&release_list_lock);
4448 if (!cgroup_is_dead(cgrp) &&
4449 list_empty(&cgrp->release_list)) {
4450 list_add(&cgrp->release_list, &release_list);
4451 need_schedule_work = 1;
4453 raw_spin_unlock(&release_list_lock);
4454 if (need_schedule_work)
4455 schedule_work(&release_agent_work);
4460 * Notify userspace when a cgroup is released, by running the
4461 * configured release agent with the name of the cgroup (path
4462 * relative to the root of cgroup file system) as the argument.
4464 * Most likely, this user command will try to rmdir this cgroup.
4466 * This races with the possibility that some other task will be
4467 * attached to this cgroup before it is removed, or that some other
4468 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4469 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4470 * unused, and this cgroup will be reprieved from its death sentence,
4471 * to continue to serve a useful existence. Next time it's released,
4472 * we will get notified again, if it still has 'notify_on_release' set.
4474 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4475 * means only wait until the task is successfully execve()'d. The
4476 * separate release agent task is forked by call_usermodehelper(),
4477 * then control in this thread returns here, without waiting for the
4478 * release agent task. We don't bother to wait because the caller of
4479 * this routine has no use for the exit status of the release agent
4480 * task, so no sense holding our caller up for that.
4482 static void cgroup_release_agent(struct work_struct *work)
4484 BUG_ON(work != &release_agent_work);
4485 mutex_lock(&cgroup_mutex);
4486 raw_spin_lock(&release_list_lock);
4487 while (!list_empty(&release_list)) {
4488 char *argv[3], *envp[3];
4490 char *pathbuf = NULL, *agentbuf = NULL, *path;
4491 struct cgroup *cgrp = list_entry(release_list.next,
4494 list_del_init(&cgrp->release_list);
4495 raw_spin_unlock(&release_list_lock);
4496 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
4499 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
4502 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4507 argv[i++] = agentbuf;
4512 /* minimal command environment */
4513 envp[i++] = "HOME=/";
4514 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4517 /* Drop the lock while we invoke the usermode helper,
4518 * since the exec could involve hitting disk and hence
4519 * be a slow process */
4520 mutex_unlock(&cgroup_mutex);
4521 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4522 mutex_lock(&cgroup_mutex);
4526 raw_spin_lock(&release_list_lock);
4528 raw_spin_unlock(&release_list_lock);
4529 mutex_unlock(&cgroup_mutex);
4532 static int __init cgroup_disable(char *str)
4534 struct cgroup_subsys *ss;
4538 while ((token = strsep(&str, ",")) != NULL) {
4542 for_each_subsys(ss, i) {
4543 if (!strcmp(token, ss->name)) {
4545 printk(KERN_INFO "Disabling %s control group"
4546 " subsystem\n", ss->name);
4553 __setup("cgroup_disable=", cgroup_disable);
4556 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
4557 * @dentry: directory dentry of interest
4558 * @ss: subsystem of interest
4560 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
4561 * to get the corresponding css and return it. If such css doesn't exist
4562 * or can't be pinned, an ERR_PTR value is returned.
4564 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
4565 struct cgroup_subsys *ss)
4567 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4568 struct cgroup_subsys_state *css = NULL;
4569 struct cgroup *cgrp;
4571 /* is @dentry a cgroup dir? */
4572 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4573 kernfs_type(kn) != KERNFS_DIR)
4574 return ERR_PTR(-EBADF);
4579 * This path doesn't originate from kernfs and @kn could already
4580 * have been or be removed at any point. @kn->priv is RCU
4581 * protected for this access. See destroy_locked() for details.
4583 cgrp = rcu_dereference(kn->priv);
4585 css = cgroup_css(cgrp, ss);
4587 if (!css || !css_tryget(css))
4588 css = ERR_PTR(-ENOENT);
4595 * css_from_id - lookup css by id
4596 * @id: the cgroup id
4597 * @ss: cgroup subsys to be looked into
4599 * Returns the css if there's valid one with @id, otherwise returns NULL.
4600 * Should be called under rcu_read_lock().
4602 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
4604 struct cgroup *cgrp;
4606 cgroup_assert_mutexes_or_rcu_locked();
4608 cgrp = idr_find(&ss->root->cgroup_idr, id);
4610 return cgroup_css(cgrp, ss);
4614 #ifdef CONFIG_CGROUP_DEBUG
4615 static struct cgroup_subsys_state *
4616 debug_css_alloc(struct cgroup_subsys_state *parent_css)
4618 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4621 return ERR_PTR(-ENOMEM);
4626 static void debug_css_free(struct cgroup_subsys_state *css)
4631 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
4634 return cgroup_task_count(css->cgroup);
4637 static u64 current_css_set_read(struct cgroup_subsys_state *css,
4640 return (u64)(unsigned long)current->cgroups;
4643 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
4649 count = atomic_read(&task_css_set(current)->refcount);
4654 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
4656 struct cgrp_cset_link *link;
4657 struct css_set *cset;
4660 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
4664 down_read(&css_set_rwsem);
4666 cset = rcu_dereference(current->cgroups);
4667 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
4668 struct cgroup *c = link->cgrp;
4670 cgroup_name(c, name_buf, NAME_MAX + 1);
4671 seq_printf(seq, "Root %d group %s\n",
4672 c->root->hierarchy_id, name_buf);
4675 up_read(&css_set_rwsem);
4680 #define MAX_TASKS_SHOWN_PER_CSS 25
4681 static int cgroup_css_links_read(struct seq_file *seq, void *v)
4683 struct cgroup_subsys_state *css = seq_css(seq);
4684 struct cgrp_cset_link *link;
4686 down_read(&css_set_rwsem);
4687 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
4688 struct css_set *cset = link->cset;
4689 struct task_struct *task;
4692 seq_printf(seq, "css_set %p\n", cset);
4694 list_for_each_entry(task, &cset->tasks, cg_list) {
4695 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
4697 seq_printf(seq, " task %d\n", task_pid_vnr(task));
4700 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
4701 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
4703 seq_printf(seq, " task %d\n", task_pid_vnr(task));
4707 seq_puts(seq, " ...\n");
4709 up_read(&css_set_rwsem);
4713 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
4715 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4718 static struct cftype debug_files[] = {
4720 .name = "taskcount",
4721 .read_u64 = debug_taskcount_read,
4725 .name = "current_css_set",
4726 .read_u64 = current_css_set_read,
4730 .name = "current_css_set_refcount",
4731 .read_u64 = current_css_set_refcount_read,
4735 .name = "current_css_set_cg_links",
4736 .seq_show = current_css_set_cg_links_read,
4740 .name = "cgroup_css_links",
4741 .seq_show = cgroup_css_links_read,
4745 .name = "releasable",
4746 .read_u64 = releasable_read,
4752 struct cgroup_subsys debug_cgrp_subsys = {
4753 .css_alloc = debug_css_alloc,
4754 .css_free = debug_css_free,
4755 .base_cftypes = debug_files,
4757 #endif /* CONFIG_CGROUP_DEBUG */